diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_0/index_tensor([3341, 3341])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_0/index_tensor([3341, 3341])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4268de0bb8a000096b7f20bec6c5fea88834b0d9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_0/index_tensor([3341, 3341])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from a slightly upward angle, the camera moves forward at an incredibly slow pace (slow dolly in), leading the viewer deep into a dense autumn forest. Warm afternoon sunlight penetrates the dense branches of the towering trees, creating distinct, dust-filled golden shafts of light and creating a striking Tyndall effect. Countless golden autumn leaves, their dry edges slightly curled, drift gracefully, twirling and swaying from above like a silent golden raindrop. Each fallen leaf, illuminated as it passes through the shafts of light, becomes almost translucent, its edges shimmering with a soft halo. A thick blanket of crimson and orange leaves blankets the ground, creating a rich gradation of color against the golden hues of the air. The image's intensely saturated colors, interweaving rich golds, fiery oranges, and deep reds, create a powerful documentary feel, reminiscent of a Steve McCurry photograph, capturing a natural moment of tranquility, grandeur, and vitality.", + "seed": 10098, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_0/index_3341_seed_10098_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024951100209727885, + "vlm_logprob": -0.024951100209727885, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.024951100209727885, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04672999307513237 + ], + "opposite_logprob_tokens": [ + -3.2967300415039062 + ], + "logprob_reward": -0.04672999307513237, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04672999307513237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11266077309846878, + 0.0464080274105072, + 0.09451179206371307, + 0.11040807515382767, + 0.12474792450666428, + 0.05771229788661003, + 0.02215728908777237, + 0.20511525869369507, + 0.06915449351072311, + 0.15712398290634155 + ], + "credit_map": [ + 0.013832814060151577, + 0.016774684190750122, + 0.013977471739053726, + 0.015070421621203423, + 0.013184423558413982, + 0.012806051410734653, + 0.01513682957738638, + 0.011878089047968388, + 0.0050655812956392765, + 0.0057451860047876835, + 0.00773854972794652, + 0.005697614047676325, + 0.005465685855597258, + 0.006316186394542456, + 0.005990451201796532, + 0.004388775676488876, + 0.005196222569793463, + 0.006565392017364502, + 0.008309095166623592, + 0.03159140795469284, + 0.02117975614964962, + 0.006238812115043402, + 0.009719958528876305, + 0.005711158737540245, + 0.00845053605735302, + 0.00609190808609128, + 0.009600757621228695, + 0.04294205829501152, + 0.0069461483508348465, + 0.008620656095445156, + 0.016696594655513763, + 0.011059429496526718, + 0.021102985367178917, + 0.016149111092090607, + 0.009144018404185772, + 0.011224404908716679, + 0.021329782903194427, + 0.020459743216633797, + 0.012065914459526539, + 0.01327198464423418, + 0.004083415027707815, + 0.004115744028240442, + 0.0032411878928542137, + 0.0050557455979287624, + 0.01383159402757883, + 0.0200650617480278, + 0.0034937995951622725, + 0.003825756488367915, + 0.0032995566725730896, + 0.0027373230550438166, + 0.0030236602760851383, + 0.0024544137995690107, + 0.003550443798303604, + 0.002985740313306451, + 0.0020550175104290247, + 0.0020511343609541655, + 0.009473900310695171, + 0.019824055954813957, + 0.01383111160248518, + 0.013010404072701931, + 0.061822641640901566, + 0.015754669904708862, + 0.05947033688426018, + 0.011928153224289417, + 0.00533720850944519, + 0.00537020293995738, + 0.005629390478134155, + 0.006436546333134174, + 0.012260198593139648, + 0.006078947335481644, + 0.02219342440366745, + 0.005848579108715057, + 0.01251688040792942, + 0.03126278892159462, + 0.013199309818446636, + 0.021341491490602493, + 0.02452162094414234, + 0.016979562118649483, + 0.014382495544850826, + 0.022919846698641777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.04672999307513237 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497122183442116 + ], + "opposite_logprob_tokens": [ + -2.929971218109131 + ], + "logprob_reward": -0.05497122183442116, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497122183442116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10866283625364304, + 0.017982807010412216, + 0.16010050475597382, + 0.07652811706066132, + 0.1131247878074646, + 0.08877459168434143, + 0.06348400563001633, + 0.28152284026145935, + 0.039844196289777756, + 0.04997534677386284 + ], + "credit_map": [ + 0.012120019644498825, + 0.012266554869711399, + 0.012986596673727036, + 0.012785431928932667, + 0.009915081784129143, + 0.012822875753045082, + 0.02334340289235115, + 0.012422878295183182, + 0.0016299004200845957, + 0.0019018035382032394, + 0.0020359039772301912, + 0.002106626983731985, + 0.0025446556974202394, + 0.003605663776397705, + 0.002522143069654703, + 0.0016361080342903733, + 0.007356016896665096, + 0.010884410701692104, + 0.01187711302191019, + 0.058766528964042664, + 0.036362115293741226, + 0.00928628258407116, + 0.016940375789999962, + 0.008627647534012794, + 0.004407361149787903, + 0.00435757776722312, + 0.006981379818171263, + 0.02812812104821205, + 0.009021284990012646, + 0.005998614709824324, + 0.013573890551924706, + 0.004059887491166592, + 0.018473487347364426, + 0.01479596272110939, + 0.008871584199368954, + 0.010844693519175053, + 0.017989877611398697, + 0.018320128321647644, + 0.012011328712105751, + 0.011817727237939835, + 0.007978430017828941, + 0.009843995794653893, + 0.005082629155367613, + 0.008901211433112621, + 0.018330607563257217, + 0.027431029826402664, + 0.00454967375844717, + 0.0066570108756423, + 0.013553902506828308, + 0.006381031591445208, + 0.010245227254927158, + 0.008452607318758965, + 0.006693257950246334, + 0.00674782507121563, + 0.006425974424928427, + 0.004984172992408276, + 0.011575832031667233, + 0.026463376358151436, + 0.020978564396500587, + 0.019374946132302284, + 0.07711637020111084, + 0.021581849083304405, + 0.0855167806148529, + 0.018915098160505295, + 0.002553449245169759, + 0.004294144455343485, + 0.004004147835075855, + 0.004002848640084267, + 0.00605427660048008, + 0.004279546905308962, + 0.01016488578170538, + 0.004490898456424475, + 0.0037881883326917887, + 0.004249586258083582, + 0.006400267593562603, + 0.00547205051407218, + 0.00584369944408536, + 0.009476122446358204, + 0.005666032899171114, + 0.009079397656023502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.05497122183442116 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022060361225157976 + ], + "opposite_logprob_tokens": [ + -6.127205848693848 + ], + "logprob_reward": -0.0022060361225157976, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022060361225157976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11375167220830917, + 0.07623924314975739, + 0.11547087877988815, + 0.18339505791664124, + 0.10717776417732239, + 0.054219041019678116, + 0.04113585501909256, + 0.16771477460861206, + 0.08096947520971298, + 0.05992627516388893 + ], + "credit_map": [ + 0.011723862029612064, + 0.013653982430696487, + 0.01752321980893612, + 0.013537717051804066, + 0.010234392248094082, + 0.014194284565746784, + 0.019650505855679512, + 0.013233709149062634, + 0.009478151798248291, + 0.010090159252285957, + 0.012944425456225872, + 0.009173182770609856, + 0.0083263348788023, + 0.011177140288054943, + 0.008169655688107014, + 0.00688019348308444, + 0.008942198008298874, + 0.008893808349967003, + 0.009786216542124748, + 0.0378975048661232, + 0.01423660758882761, + 0.00937500037252903, + 0.017763324081897736, + 0.0085762282833457, + 0.013305827043950558, + 0.01652229204773903, + 0.018312394618988037, + 0.06214483827352524, + 0.015907512977719307, + 0.0142599418759346, + 0.029923921450972557, + 0.013018314726650715, + 0.015056313946843147, + 0.010832434520125389, + 0.00794921163469553, + 0.011719265021383762, + 0.022810721769928932, + 0.01624629646539688, + 0.014205720275640488, + 0.008357795886695385, + 0.004074967000633478, + 0.007046507205814123, + 0.003839152166619897, + 0.006435783579945564, + 0.009981085546314716, + 0.0128336101770401, + 0.005556068383157253, + 0.004451864864677191, + 0.005258476827293634, + 0.005787942558526993, + 0.009228021837770939, + 0.004896189086139202, + 0.0044641038402915, + 0.004869319964200258, + 0.0035603714641183615, + 0.0030714261811226606, + 0.007889931090176105, + 0.017610255628824234, + 0.015707045793533325, + 0.011819160543382168, + 0.0446755476295948, + 0.013960130512714386, + 0.045065440237522125, + 0.010987251065671444, + 0.004790885373950005, + 0.006689540576189756, + 0.006482013501226902, + 0.00651430943980813, + 0.018262675032019615, + 0.007115968968719244, + 0.024687103927135468, + 0.006426978390663862, + 0.005743234883993864, + 0.013214911334216595, + 0.004392866510897875, + 0.008657216094434261, + 0.008548442274332047, + 0.008221554569900036, + 0.006047307979315519, + 0.005100745242089033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.0022060361225157976 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781289845705032 + ], + "opposite_logprob_tokens": [ + -4.634781360626221 + ], + "logprob_reward": -0.009781289845705032, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781289845705032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15182587504386902, + 0.08491263538599014, + 0.1310659646987915, + 0.14039519429206848, + 0.03874285891652107, + 0.09515039622783661, + 0.03438473492860794, + 0.07309664040803909, + 0.06421902775764465, + 0.18620668351650238 + ], + "credit_map": [ + 0.010604850947856903, + 0.014465661719441414, + 0.016324853524565697, + 0.016680732369422913, + 0.016867319121956825, + 0.030633050948381424, + 0.03054802119731903, + 0.01570137031376362, + 0.009553874842822552, + 0.009714053943753242, + 0.01539718359708786, + 0.011752132326364517, + 0.008463723585009575, + 0.01269453950226307, + 0.009278565645217896, + 0.008058559149503708, + 0.008997563272714615, + 0.010529031977057457, + 0.008158045820891857, + 0.031118715181946754, + 0.025572147220373154, + 0.011172731406986713, + 0.026049233973026276, + 0.009468485601246357, + 0.008178187534213066, + 0.015318458899855614, + 0.011091766878962517, + 0.04063694179058075, + 0.029953207820653915, + 0.013660269789397717, + 0.0156707800924778, + 0.005885574966669083, + 0.006225043907761574, + 0.0035203516017645597, + 0.0033338002394884825, + 0.004729290492832661, + 0.007811929099261761, + 0.005462165921926498, + 0.003827345324680209, + 0.003832931397482753, + 0.00594166899099946, + 0.010543854907155037, + 0.007424104493111372, + 0.008939090184867382, + 0.018745966255664825, + 0.029465973377227783, + 0.008693069219589233, + 0.005396659951657057, + 0.004453076981008053, + 0.005436387378722429, + 0.005681750830262899, + 0.0040978421457111835, + 0.004291568882763386, + 0.004502185620367527, + 0.0035947191063314676, + 0.0023272016551345587, + 0.0069486522115767, + 0.009659316390752792, + 0.013468200340867043, + 0.00931005273014307, + 0.004966295324265957, + 0.008280525915324688, + 0.009362123906612396, + 0.011101468466222286, + 0.0036703948862850666, + 0.008121860213577747, + 0.00539031159132719, + 0.005057507194578648, + 0.01228354312479496, + 0.007139112800359726, + 0.013523285277187824, + 0.009033008478581905, + 0.014856739901006222, + 0.04415364935994148, + 0.017171842977404594, + 0.025149177759885788, + 0.031311653554439545, + 0.017298182472586632, + 0.016331378370523453, + 0.019934045150876045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.009781289845705032 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011066960170865059 + ], + "opposite_logprob_tokens": [ + -4.511066913604736 + ], + "logprob_reward": -0.011066960170865059, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011066960170865059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14602325856685638, + 0.1839429885149002, + 0.03768794983625412, + 0.04211505129933357, + 0.07810413837432861, + 0.1199861466884613, + 0.03602223098278046, + 0.11079056560993195, + 0.021986251696944237, + 0.22334139049053192 + ], + "credit_map": [ + 0.014636612497270107, + 0.0164759811013937, + 0.02028922736644745, + 0.019597891718149185, + 0.019171442836523056, + 0.028362302109599113, + 0.013783780857920647, + 0.013706028461456299, + 0.019704878330230713, + 0.018102506175637245, + 0.03343142941594124, + 0.023972170427441597, + 0.02559729292988777, + 0.024852389469742775, + 0.019832350313663483, + 0.01844996213912964, + 0.0034960322082042694, + 0.004976354539394379, + 0.004064210690557957, + 0.010168529115617275, + 0.0029129066970199347, + 0.0047551533207297325, + 0.00482935318723321, + 0.002485408913344145, + 0.003973727580159903, + 0.007460537366569042, + 0.003652419662103057, + 0.007270487491041422, + 0.007322824094444513, + 0.004945805761963129, + 0.005074948072433472, + 0.002414301736280322, + 0.010851099155843258, + 0.009687196463346481, + 0.006764642428606749, + 0.011396147310733795, + 0.009586537256836891, + 0.011253952980041504, + 0.010362760163843632, + 0.008201805874705315, + 0.006233636289834976, + 0.006448383443057537, + 0.00652123661711812, + 0.006924187298864126, + 0.03182460740208626, + 0.04583638906478882, + 0.010210351087152958, + 0.005987349897623062, + 0.0043563214130699635, + 0.007353150751441717, + 0.006341957952827215, + 0.004613506607711315, + 0.0035605451557785273, + 0.004824228119105101, + 0.002846884774044156, + 0.002125635975971818, + 0.006327638868242502, + 0.009668240323662758, + 0.01118751335889101, + 0.008392061106860638, + 0.028767596930265427, + 0.008968763053417206, + 0.0301812794059515, + 0.0072974758222699165, + 0.0018332216423004866, + 0.0024825746659189463, + 0.0034288347233086824, + 0.003380783135071397, + 0.0019303804729133844, + 0.002694162540137768, + 0.003932830411940813, + 0.002303462475538254, + 0.01797194592654705, + 0.05752979591488838, + 0.021428897976875305, + 0.030659249052405357, + 0.041178520768880844, + 0.019919969141483307, + 0.016491012647747993, + 0.01816200651228428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.011066960170865059 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly upward angle, the camera moves forward at an incredibly slow pace (slow dolly in), leading the viewer deep into a dense autumn forest. Warm afternoon sunlight penetrates the dense branches of the towering trees, creating distinct, dust-filled golden shafts of light and creating a striking Tyndall effect. Countless golden autumn leaves, their dry edges slightly curled, drift gracefully, twirling and swaying from above like a silent golden raindrop. Each fallen leaf, illuminated as it passes through the shafts of light, becomes almost translucent, its edges shimmering with a soft halo. A thick blanket of crimson and orange leaves blankets the ground, creating a rich gradation of color against the golden hues of the air. The image's intensely saturated colors, interweaving rich golds, fiery oranges, and deep reds, create a powerful documentary feel, reminiscent of a Steve McCurry photograph, capturing a natural moment of tranquility, grandeur, and vitality.", + "seed": 10099, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_0/index_3341_seed_10099_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026333845942281185, + "vlm_logprob": -0.026333845942281185, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.026333845942281185, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04783928394317627 + ], + "opposite_logprob_tokens": [ + -3.2978391647338867 + ], + "logprob_reward": -0.04783928394317627, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04783928394317627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1337784379720688, + 0.11811588704586029, + 0.3257383406162262, + 0.04844546690583229, + 0.08160781860351562, + 0.10905913263559341, + 0.03400854021310806, + 0.034816253930330276, + 0.0984995886683464, + 0.015930568799376488 + ], + "credit_map": [ + 0.009551557712256908, + 0.01839574985206127, + 0.03473973274230957, + 0.021058162674307823, + 0.008248100988566875, + 0.011497928760945797, + 0.015168304555118084, + 0.015118892304599285, + 0.00938466191291809, + 0.03156602382659912, + 0.00955513771623373, + 0.013150856830179691, + 0.013528184033930302, + 0.008259477093815804, + 0.012020191177725792, + 0.020651347935199738, + 0.01853317581117153, + 0.016033876687288284, + 0.12396781146526337, + 0.036286305636167526, + 0.00964797381311655, + 0.012217612005770206, + 0.014291316270828247, + 0.09476024657487869, + 0.005183312576264143, + 0.005237047094851732, + 0.005267047788947821, + 0.006169967353343964, + 0.004713759757578373, + 0.006828132085502148, + 0.010076765902340412, + 0.0049694376066327095, + 0.00706242211163044, + 0.007500752806663513, + 0.011576212011277676, + 0.012159356847405434, + 0.01248263381421566, + 0.009143328294157982, + 0.011610402725636959, + 0.010072712786495686, + 0.010853186249732971, + 0.01407216489315033, + 0.02061409316956997, + 0.017990361899137497, + 0.008485457859933376, + 0.010374320670962334, + 0.013022863306105137, + 0.0136466845870018, + 0.003258597571402788, + 0.004825479816645384, + 0.004146674182265997, + 0.0044423495419323444, + 0.003612611908465624, + 0.0031092918943613768, + 0.004338729660958052, + 0.006274806801229715, + 0.0028125415556132793, + 0.0073408870957791805, + 0.0027086168993264437, + 0.0035770919639617205, + 0.004366929177194834, + 0.0070713278837502, + 0.003930925857275724, + 0.003007933497428894, + 0.004159953910857439, + 0.02539583295583725, + 0.005093705840408802, + 0.00571045745164156, + 0.006613587494939566, + 0.003542575053870678, + 0.03331586718559265, + 0.014667610637843609, + 0.0017381289508193731, + 0.002211106475442648, + 0.0017090560868382454, + 0.0016880310140550137, + 0.0017439998919144273, + 0.0032513246405869722, + 0.001584324287250638, + 0.002004598965868354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.04783928394317627 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498047545552254 + ], + "opposite_logprob_tokens": [ + -2.929980516433716 + ], + "logprob_reward": -0.05498047545552254, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498047545552254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.185402512550354, + 0.09903635084629059, + 0.09636227786540985, + 0.057167865335941315, + 0.08267044275999069, + 0.07551063597202301, + 0.2212124913930893, + 0.05763304978609085, + 0.06308471411466599, + 0.06191962584853172 + ], + "credit_map": [ + 0.017682043835520744, + 0.015935514122247696, + 0.028766974806785583, + 0.028973588719964027, + 0.01873938925564289, + 0.017705610021948814, + 0.03306237980723381, + 0.024537034332752228, + 0.008533847518265247, + 0.01746995560824871, + 0.014817656949162483, + 0.015910614281892776, + 0.008970668539404869, + 0.008241385221481323, + 0.01638692617416382, + 0.008705311454832554, + 0.004944884683936834, + 0.028283093124628067, + 0.013795720413327217, + 0.010949256829917431, + 0.009587761014699936, + 0.006072666961699724, + 0.007691611535847187, + 0.01503728237003088, + 0.0036385883577167988, + 0.006259771063923836, + 0.007519144099205732, + 0.006821951828896999, + 0.005830313544720411, + 0.004643926862627268, + 0.015273364260792732, + 0.007180809509009123, + 0.006482921075075865, + 0.016035769134759903, + 0.012502812780439854, + 0.010359924286603928, + 0.008321104571223259, + 0.009458528831601143, + 0.011503074318170547, + 0.00800631009042263, + 0.007359552662819624, + 0.01038367860019207, + 0.011644392274320126, + 0.013432424515485764, + 0.0059732855297625065, + 0.009653053246438503, + 0.008994781412184238, + 0.008069475181400776, + 0.013961199671030045, + 0.04155326634645462, + 0.061644915491342545, + 0.020668426528573036, + 0.017905278131365776, + 0.014961321838200092, + 0.0343446210026741, + 0.01617347076535225, + 0.005228819325566292, + 0.007303901948034763, + 0.005966956727206707, + 0.008049782365560532, + 0.005449789110571146, + 0.0111834816634655, + 0.007678956259042025, + 0.006771370768547058, + 0.004860627930611372, + 0.0037591694854199886, + 0.009380300529301167, + 0.012510444968938828, + 0.0038140593096613884, + 0.01003845315426588, + 0.008253411389887333, + 0.010468252934515476, + 0.004364491440355778, + 0.01589399203658104, + 0.003026893362402916, + 0.004042916465550661, + 0.006754256784915924, + 0.007551839109510183, + 0.004873341880738735, + 0.015411898493766785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.05498047545552254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019502682844176888 + ], + "opposite_logprob_tokens": [ + -6.251950263977051 + ], + "logprob_reward": -0.0019502682844176888, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019502682844176888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15822164714336395, + 0.16472391784191132, + 0.11945591866970062, + 0.04813631251454353, + 0.07893552631139755, + 0.03516099601984024, + 0.13999556005001068, + 0.06928771734237671, + 0.1202782467007637, + 0.06580419093370438 + ], + "credit_map": [ + 0.020370133221149445, + 0.013531320728361607, + 0.019094372168183327, + 0.02107767015695572, + 0.015959767624735832, + 0.023062698543071747, + 0.02688598446547985, + 0.018239697441458702, + 0.01146462932229042, + 0.05349453538656235, + 0.008329344913363457, + 0.012784674763679504, + 0.021338796243071556, + 0.016000865027308464, + 0.01043263915926218, + 0.030878430232405663, + 0.009122205898165703, + 0.009898620657622814, + 0.02940528094768524, + 0.01325213722884655, + 0.013025716878473759, + 0.009200098924338818, + 0.011710765771567822, + 0.023841088637709618, + 0.004530658945441246, + 0.006225085351616144, + 0.006942144129425287, + 0.0061811585910618305, + 0.008877714164555073, + 0.006066915113478899, + 0.005800134502351284, + 0.0035125012509524822, + 0.007639565039426088, + 0.017749782651662827, + 0.010263236239552498, + 0.00788635853677988, + 0.008527994155883789, + 0.013655345886945724, + 0.007115015294402838, + 0.006098233629018068, + 0.005080313887447119, + 0.005072481464594603, + 0.004040089901536703, + 0.004883402027189732, + 0.004140696953982115, + 0.003710914868861437, + 0.004020512569695711, + 0.004212585743516684, + 0.011015171185135841, + 0.01150485035032034, + 0.0253689493983984, + 0.012686816975474358, + 0.009714349173009396, + 0.011324441060423851, + 0.036249127238988876, + 0.02213185466825962, + 0.007212448865175247, + 0.011548183858394623, + 0.005721119698137045, + 0.005879625678062439, + 0.0071142856031656265, + 0.01824425905942917, + 0.006923486012965441, + 0.0066443076357245445, + 0.00488132843747735, + 0.030035898089408875, + 0.005034859757870436, + 0.006465156562626362, + 0.008252641186118126, + 0.005529937334358692, + 0.04073318839073181, + 0.0193452350795269, + 0.004208823200315237, + 0.019416656345129013, + 0.0028708926402032375, + 0.0040601021610200405, + 0.006967486813664436, + 0.005301787983626127, + 0.005896163638681173, + 0.017082279548048973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0019502682844176888 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287491872906685 + ], + "opposite_logprob_tokens": [ + -3.773287534713745 + ], + "logprob_reward": -0.023287491872906685, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287491872906685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12345805019140244, + 0.06109002232551575, + 0.13249947130680084, + 0.036485034972429276, + 0.04144297167658806, + 0.0604160875082016, + 0.3086121380329132, + 0.030946502462029457, + 0.14865665137767792, + 0.0563930943608284 + ], + "credit_map": [ + 0.009879873134195805, + 0.013369654305279255, + 0.025689495727419853, + 0.020939603447914124, + 0.008299087174236774, + 0.013710450381040573, + 0.018269019201397896, + 0.013300872407853603, + 0.004272673279047012, + 0.013459895737469196, + 0.008102752268314362, + 0.007224262226372957, + 0.007074065040796995, + 0.006181600037962198, + 0.005079991649836302, + 0.009694780223071575, + 0.009758721105754375, + 0.0062676384113729, + 0.0423751026391983, + 0.016073262318968773, + 0.007175324950367212, + 0.008373240940272808, + 0.010646323673427105, + 0.031829848885536194, + 0.005424323957413435, + 0.0041534146293997765, + 0.005596885923296213, + 0.00454731797799468, + 0.0031976313330233097, + 0.003159071085974574, + 0.006286622956395149, + 0.004119767341762781, + 0.0032210247591137886, + 0.0034707526210695505, + 0.00527389720082283, + 0.005696830339729786, + 0.005270826630294323, + 0.005621823947876692, + 0.0071655148640275, + 0.005722303409129381, + 0.010012314654886723, + 0.0067240213975310326, + 0.010551770217716694, + 0.010801691561937332, + 0.006110028829425573, + 0.004754459485411644, + 0.006345351692289114, + 0.005116453394293785, + 0.01830441504716873, + 0.06207976117730141, + 0.07650283724069595, + 0.020448734983801842, + 0.025409797206521034, + 0.020661376416683197, + 0.057677771896123886, + 0.027527417987585068, + 0.0023190458305180073, + 0.007296948693692684, + 0.0019296938553452492, + 0.002858312800526619, + 0.0034565236419439316, + 0.0053549353033304214, + 0.004857839085161686, + 0.002873202320188284, + 0.006052650511264801, + 0.036563001573085785, + 0.008204936981201172, + 0.009806309826672077, + 0.00972509104758501, + 0.006553619168698788, + 0.04923298954963684, + 0.022518057376146317, + 0.006689675152301788, + 0.014371714554727077, + 0.004328818991780281, + 0.004456022288650274, + 0.005733729340136051, + 0.00411443505436182, + 0.004423138219863176, + 0.012275558896362782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 2 + ], + "selected_reward": -0.023287491872906685 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036117101553827524 + ], + "opposite_logprob_tokens": [ + -5.6286115646362305 + ], + "logprob_reward": -0.0036117101553827524, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036117101553827524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06527245789766312, + 0.08456359058618546, + 0.16885076463222504, + 0.043885257095098495, + 0.07092232257127762, + 0.04223855584859848, + 0.12074162065982819, + 0.02619314007461071, + 0.31261131167411804, + 0.06472104787826538 + ], + "credit_map": [ + 0.007483982481062412, + 0.00550494808703661, + 0.011045759543776512, + 0.011161243543028831, + 0.00385719514451921, + 0.0071052792482078075, + 0.011714695021510124, + 0.007399349473416805, + 0.00787745974957943, + 0.025122730061411858, + 0.005332173779606819, + 0.008944079279899597, + 0.00961670558899641, + 0.007090486586093903, + 0.004961367696523666, + 0.015618576668202877, + 0.015993930399417877, + 0.03453925624489784, + 0.048793211579322815, + 0.021079329773783684, + 0.01405283808708191, + 0.010731284506618977, + 0.011529451236128807, + 0.012131452560424805, + 0.00749381585046649, + 0.0056314608082175255, + 0.0056548104621469975, + 0.005743676330894232, + 0.00457307044416666, + 0.004528717137873173, + 0.005164985544979572, + 0.005094716791063547, + 0.005142476875334978, + 0.010487458668649197, + 0.011153860948979855, + 0.010865971446037292, + 0.007319949567317963, + 0.007607020437717438, + 0.008799035102128983, + 0.00954654160887003, + 0.00275393296033144, + 0.006533342879265547, + 0.00947309285402298, + 0.008004427887499332, + 0.003536895150318742, + 0.0035656560212373734, + 0.003822414670139551, + 0.004548791795969009, + 0.0071376715786755085, + 0.020546656101942062, + 0.03151928633451462, + 0.008189601823687553, + 0.0087650241330266, + 0.00706755556166172, + 0.026913460344076157, + 0.010602354072034359, + 0.0020833085291087627, + 0.005346677266061306, + 0.002595398109406233, + 0.002539019100368023, + 0.003022118704393506, + 0.005060029681771994, + 0.003136896761134267, + 0.0024096881970763206, + 0.013396943919360638, + 0.07630214095115662, + 0.013326093554496765, + 0.017616407945752144, + 0.021793602034449577, + 0.01630539633333683, + 0.1049356460571289, + 0.048935070633888245, + 0.006716218311339617, + 0.014156881719827652, + 0.004996968898922205, + 0.006267174147069454, + 0.006764752324670553, + 0.005171370226889849, + 0.006820645183324814, + 0.013827032409608364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 6 + ], + "selected_reward": -0.0036117101553827524 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly upward angle, the camera moves forward at an incredibly slow pace (slow dolly in), leading the viewer deep into a dense autumn forest. Warm afternoon sunlight penetrates the dense branches of the towering trees, creating distinct, dust-filled golden shafts of light and creating a striking Tyndall effect. Countless golden autumn leaves, their dry edges slightly curled, drift gracefully, twirling and swaying from above like a silent golden raindrop. Each fallen leaf, illuminated as it passes through the shafts of light, becomes almost translucent, its edges shimmering with a soft halo. A thick blanket of crimson and orange leaves blankets the ground, creating a rich gradation of color against the golden hues of the air. The image's intensely saturated colors, interweaving rich golds, fiery oranges, and deep reds, create a powerful documentary feel, reminiscent of a Steve McCurry photograph, capturing a natural moment of tranquility, grandeur, and vitality.", + "seed": 10100, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_0/index_3341_seed_10100_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.030204129847697915, + "vlm_logprob": -0.030204129847697915, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.030204129847697915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07440740615129471 + ], + "opposite_logprob_tokens": [ + -2.8244073390960693 + ], + "logprob_reward": -0.07440740615129471, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07440740615129471, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30306118726730347, + 0.11936280131340027, + 0.1094011664390564, + 0.024489875882864, + 0.07740464061498642, + 0.08039262890815735, + 0.13684847950935364, + 0.07796712964773178, + 0.05108880251646042, + 0.019983265548944473 + ], + "credit_map": [ + 0.02576022781431675, + 0.04328780248761177, + 0.053071070462465286, + 0.040423888713121414, + 0.06521879881620407, + 0.03136692941188812, + 0.022523852065205574, + 0.021408624947071075, + 0.011311658658087254, + 0.01220342144370079, + 0.015600461512804031, + 0.015048348344862461, + 0.01721995510160923, + 0.008049244992434978, + 0.028315844014286995, + 0.011613872833549976, + 0.018944235518574715, + 0.013762816786766052, + 0.011634326539933681, + 0.014813097193837166, + 0.010863285511732101, + 0.012004828080534935, + 0.01535153016448021, + 0.012027041986584663, + 0.0034897103905677795, + 0.0030283189844340086, + 0.0020174558740109205, + 0.0026439325883984566, + 0.003905203193426132, + 0.003567429492250085, + 0.0034587050322443247, + 0.0023791200947016478, + 0.011859845370054245, + 0.014287433587014675, + 0.008001149632036686, + 0.008447262458503246, + 0.007673037238419056, + 0.010876355692744255, + 0.008164018392562866, + 0.008095532655715942, + 0.007363176438957453, + 0.017687590792775154, + 0.007042617537081242, + 0.007120110094547272, + 0.009017731063067913, + 0.01153669785708189, + 0.012189195491373539, + 0.008435510098934174, + 0.007581211160868406, + 0.03050878643989563, + 0.016983523964881897, + 0.04424380511045456, + 0.014079388231039047, + 0.004912126809358597, + 0.01252674125134945, + 0.006012883968651295, + 0.008230860345065594, + 0.006824627052992582, + 0.010255823843181133, + 0.01330174133181572, + 0.007994894869625568, + 0.009130148217082024, + 0.014363919384777546, + 0.007865113206207752, + 0.00438894284889102, + 0.005787785165011883, + 0.006623987574130297, + 0.006216185633093119, + 0.004773998167365789, + 0.006869378499686718, + 0.01164316013455391, + 0.004785364959388971, + 0.0016836361028254032, + 0.002775036497041583, + 0.0028196643106639385, + 0.00262205908074975, + 0.0021405029110610485, + 0.0031362513545900583, + 0.002391664544120431, + 0.002414449816569686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.07440740615129471 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206587329506874 + ], + "opposite_logprob_tokens": [ + -2.812065839767456 + ], + "logprob_reward": -0.06206587329506874, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206587329506874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23982524871826172, + 0.08961919695138931, + 0.04900675266981125, + 0.04572497308254242, + 0.06632564961910248, + 0.08917418867349625, + 0.18182553350925446, + 0.08197944611310959, + 0.0828339010477066, + 0.07368510216474533 + ], + "credit_map": [ + 0.03501761332154274, + 0.029751945286989212, + 0.0435638390481472, + 0.035795681178569794, + 0.034781426191329956, + 0.026755142956972122, + 0.015585425309836864, + 0.018574180081486702, + 0.0080764414742589, + 0.010083790868520737, + 0.009979651309549809, + 0.009590272791683674, + 0.011781904846429825, + 0.006688589695841074, + 0.025974009186029434, + 0.007444542832672596, + 0.007250388618558645, + 0.006231122184544802, + 0.005419249180704355, + 0.0069304886274039745, + 0.005727014038711786, + 0.006218135356903076, + 0.0057060024701058865, + 0.0055243512615561485, + 0.005086994264274836, + 0.005211812909692526, + 0.005150847602635622, + 0.006524182856082916, + 0.006554027553647757, + 0.006763096898794174, + 0.004889188800007105, + 0.005544822663068771, + 0.006948862690478563, + 0.008836916647851467, + 0.008504806086421013, + 0.010761091485619545, + 0.006583282724022865, + 0.008843245916068554, + 0.007530965376645327, + 0.008316480554640293, + 0.005914954002946615, + 0.007144721690565348, + 0.01870124228298664, + 0.017316754907369614, + 0.0072934129275381565, + 0.0070673441514372826, + 0.01894049346446991, + 0.0067952703684568405, + 0.009333476424217224, + 0.038963016122579575, + 0.024468813091516495, + 0.06299715489149094, + 0.015750641003251076, + 0.006636326666921377, + 0.014236065559089184, + 0.009440034627914429, + 0.0057491110637784, + 0.006969796493649483, + 0.01147870346903801, + 0.0168943889439106, + 0.013905077241361141, + 0.005867413245141506, + 0.014515970833599567, + 0.006598988082259893, + 0.006630647927522659, + 0.015167941339313984, + 0.007495007943361998, + 0.01324574276804924, + 0.01765628159046173, + 0.00789874792098999, + 0.00920187495648861, + 0.005537657532840967, + 0.00794488936662674, + 0.011569754220545292, + 0.0065429010428488255, + 0.008867727592587471, + 0.011135510168969631, + 0.012025217525660992, + 0.007050890475511551, + 0.00854821503162384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.06206587329506874 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019520529313012958 + ], + "opposite_logprob_tokens": [ + -6.251952171325684 + ], + "logprob_reward": -0.0019520529313012958, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019520529313012958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21065697073936462, + 0.08828289061784744, + 0.10377763956785202, + 0.03788430988788605, + 0.037685684859752655, + 0.07039356976747513, + 0.06426919251680374, + 0.27340155839920044, + 0.07869704812765121, + 0.034951113164424896 + ], + "credit_map": [ + 0.03937745839357376, + 0.027819082140922546, + 0.020680200308561325, + 0.01579725556075573, + 0.030848048627376556, + 0.030389748513698578, + 0.01901857741177082, + 0.02672662027180195, + 0.009506900794804096, + 0.009767935611307621, + 0.013294198550283909, + 0.013600152917206287, + 0.00816600676625967, + 0.010507943108677864, + 0.015442830510437489, + 0.007996920496225357, + 0.019030742347240448, + 0.016008533537387848, + 0.0107707055285573, + 0.011412711814045906, + 0.012911083176732063, + 0.01374001707881689, + 0.00814884714782238, + 0.01175500638782978, + 0.003631813218817115, + 0.005754290148615837, + 0.003163909073919058, + 0.003636147826910019, + 0.00666995532810688, + 0.005373354069888592, + 0.005187069531530142, + 0.004467769060283899, + 0.004588560201227665, + 0.008282586000859737, + 0.004063888918608427, + 0.00366509729065001, + 0.004381611943244934, + 0.0044769286178052425, + 0.0041904388926923275, + 0.00403657415881753, + 0.0059676580131053925, + 0.005736576393246651, + 0.011943248100578785, + 0.010264365002512932, + 0.005600222386419773, + 0.00728012016043067, + 0.01675471104681492, + 0.006846672855317593, + 0.005371234845370054, + 0.007532350718975067, + 0.011391965672373772, + 0.00816429778933525, + 0.009661572054028511, + 0.003411556826904416, + 0.011650266125798225, + 0.007085954304784536, + 0.031607117503881454, + 0.025859300047159195, + 0.03336728364229202, + 0.036885716021060944, + 0.04419054836034775, + 0.030420541763305664, + 0.045072104781866074, + 0.025998974218964577, + 0.009672232903540134, + 0.00893761869519949, + 0.00831224862486124, + 0.008140298537909985, + 0.007138729561120272, + 0.01308518834412098, + 0.014976582489907742, + 0.008434154093265533, + 0.0035568138118833303, + 0.003648418001830578, + 0.0031189480796456337, + 0.005446146707981825, + 0.003720649052411318, + 0.005294310860335827, + 0.005303272046148777, + 0.004862555768340826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.0019520529313012958 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780935011804104 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780935011804104, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780935011804104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13095006346702576, + 0.10666389018297195, + 0.10372024774551392, + 0.05285360664129257, + 0.04140310734510422, + 0.10329439491033554, + 0.131033256649971, + 0.2535025179386139, + 0.056505732238292694, + 0.02007320150732994 + ], + "credit_map": [ + 0.012432015500962734, + 0.018785616382956505, + 0.015344290994107723, + 0.015041130594909191, + 0.02491668425500393, + 0.018526015803217888, + 0.013559706509113312, + 0.0123446024954319, + 0.010601812042295933, + 0.011271667666733265, + 0.016657613217830658, + 0.016932828351855278, + 0.019567221403121948, + 0.007021722383797169, + 0.013328167609870434, + 0.011282864958047867, + 0.017316700890660286, + 0.012399940751492977, + 0.011045298539102077, + 0.014446071349084377, + 0.010595952160656452, + 0.010593698360025883, + 0.013940333388745785, + 0.013382252305746078, + 0.007744868751615286, + 0.005817714612931013, + 0.005026274360716343, + 0.0061219073832035065, + 0.006916519720107317, + 0.009528431110084057, + 0.007748288102447987, + 0.003949605859816074, + 0.004105866886675358, + 0.010057743638753891, + 0.003790666349232197, + 0.005121124442666769, + 0.004672389477491379, + 0.0064910100772976875, + 0.004099675919860601, + 0.0030646317172795534, + 0.006798082962632179, + 0.009083922021090984, + 0.020733201876282692, + 0.018668077886104584, + 0.008275446482002735, + 0.007941381074488163, + 0.022901052609086037, + 0.008893223479390144, + 0.0134034538641572, + 0.007981712929904461, + 0.017875434830784798, + 0.02840578369796276, + 0.013195395469665527, + 0.011291305534541607, + 0.02570272423326969, + 0.013177454471588135, + 0.013866019435226917, + 0.013948299922049046, + 0.047601159662008286, + 0.03653043508529663, + 0.014416772872209549, + 0.016058387234807014, + 0.0789322704076767, + 0.03214917331933975, + 0.005181999411433935, + 0.007547449320554733, + 0.006358450278639793, + 0.008497032336890697, + 0.00564308138564229, + 0.0075566936284303665, + 0.01059062872081995, + 0.00513039855286479, + 0.002237909007817507, + 0.002183410804718733, + 0.0021873959340155125, + 0.003129997057840228, + 0.0016521512297913432, + 0.0034237890504300594, + 0.0023987507447600365, + 0.0028597970958799124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.009780935011804104 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028143818490207195 + ], + "opposite_logprob_tokens": [ + -5.877814292907715 + ], + "logprob_reward": -0.0028143818490207195, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028143818490207195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19980937242507935, + 0.06414156407117844, + 0.07354337722063065, + 0.05052033066749573, + 0.07093127071857452, + 0.09486238658428192, + 0.121128611266613, + 0.182226300239563, + 0.047791145741939545, + 0.09504565596580505 + ], + "credit_map": [ + 0.03258337453007698, + 0.021179981529712677, + 0.030465686693787575, + 0.03578079864382744, + 0.017743706703186035, + 0.026774834841489792, + 0.015129501931369305, + 0.020151499658823013, + 0.00569938076660037, + 0.006348636467009783, + 0.00783588457852602, + 0.008324957452714443, + 0.006606937851756811, + 0.005453875288367271, + 0.01538107916712761, + 0.008490804582834244, + 0.012890742160379887, + 0.01156123261898756, + 0.00828269962221384, + 0.006891326978802681, + 0.009112090803682804, + 0.007781736087054014, + 0.011268548667430878, + 0.005755001679062843, + 0.004735201597213745, + 0.003622975666075945, + 0.007384285796433687, + 0.008916795253753662, + 0.004345952067524195, + 0.006140347104519606, + 0.010030926205217838, + 0.005343849770724773, + 0.007504723966121674, + 0.010037804953753948, + 0.010082683525979519, + 0.012172630988061428, + 0.008657561615109444, + 0.0065604704432189465, + 0.009202114306390285, + 0.006713285110890865, + 0.007479035295546055, + 0.007107342127710581, + 0.019495027139782906, + 0.015699123963713646, + 0.007107770070433617, + 0.007921295240521431, + 0.020523933693766594, + 0.009528854861855507, + 0.012337329797446728, + 0.011455219238996506, + 0.015382662415504456, + 0.01696190796792507, + 0.020155232399702072, + 0.013155949302017689, + 0.020586183294653893, + 0.011094117537140846, + 0.012251642532646656, + 0.013960550539195538, + 0.028780726715922356, + 0.02897688001394272, + 0.020634042099118233, + 0.014836185611784458, + 0.043631549924612045, + 0.019154710695147514, + 0.003042937023565173, + 0.006693615578114986, + 0.004793897736817598, + 0.00850176066160202, + 0.008603785187005997, + 0.004337021615356207, + 0.008056369610130787, + 0.0037617580965161324, + 0.008483619429171085, + 0.013187579810619354, + 0.010469100438058376, + 0.013634885661303997, + 0.015686657279729843, + 0.008122610859572887, + 0.01618935912847519, + 0.009271840564906597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0028143818490207195 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly upward angle, the camera moves forward at an incredibly slow pace (slow dolly in), leading the viewer deep into a dense autumn forest. Warm afternoon sunlight penetrates the dense branches of the towering trees, creating distinct, dust-filled golden shafts of light and creating a striking Tyndall effect. Countless golden autumn leaves, their dry edges slightly curled, drift gracefully, twirling and swaying from above like a silent golden raindrop. Each fallen leaf, illuminated as it passes through the shafts of light, becomes almost translucent, its edges shimmering with a soft halo. A thick blanket of crimson and orange leaves blankets the ground, creating a rich gradation of color against the golden hues of the air. The image's intensely saturated colors, interweaving rich golds, fiery oranges, and deep reds, create a powerful documentary feel, reminiscent of a Steve McCurry photograph, capturing a natural moment of tranquility, grandeur, and vitality.", + "seed": 10101, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_0/index_3341_seed_10101_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03846655241213739, + "vlm_logprob": -0.03846655241213739, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03846655241213739, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09920508414506912 + ], + "opposite_logprob_tokens": [ + -2.5992050170898438 + ], + "logprob_reward": -0.09920508414506912, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09920508414506912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20408368110656738, + 0.03470434248447418, + 0.0704590380191803, + 0.05021142214536667, + 0.16948182880878448, + 0.04790124297142029, + 0.19174955785274506, + 0.05151214078068733, + 0.06864657998085022, + 0.111250139772892 + ], + "credit_map": [ + 0.025732778012752533, + 0.056181441992521286, + 0.025364480912685394, + 0.026080096140503883, + 0.018144095316529274, + 0.01418086513876915, + 0.019244810566306114, + 0.01915510557591915, + 0.00473737483844161, + 0.00676508154720068, + 0.0032563654240220785, + 0.0044489456340670586, + 0.0036357841454446316, + 0.002277332590892911, + 0.005336997099220753, + 0.004246462136507034, + 0.004295848775655031, + 0.00919503066688776, + 0.0069642034359276295, + 0.008440058678388596, + 0.00933814886957407, + 0.008417167700827122, + 0.01435045525431633, + 0.009458120912313461, + 0.0039520543068647385, + 0.0050395093858242035, + 0.00424572266638279, + 0.00610246229916811, + 0.004129846580326557, + 0.004462437704205513, + 0.012353545054793358, + 0.009925847873091698, + 0.011166803538799286, + 0.024392204359173775, + 0.013927619904279709, + 0.03338408097624779, + 0.008264976553618908, + 0.006065824534744024, + 0.04753848910331726, + 0.024741826578974724, + 0.005542562808841467, + 0.004451114218682051, + 0.009044268168509007, + 0.010287013836205006, + 0.004053258802741766, + 0.0028405843768268824, + 0.007002023048698902, + 0.00468041654676199, + 0.014685820788145065, + 0.039426710456609726, + 0.025928085669875145, + 0.019175032153725624, + 0.015621640719473362, + 0.012322703376412392, + 0.014465686865150928, + 0.05012388899922371, + 0.004217020235955715, + 0.0056454031728208065, + 0.004419171251356602, + 0.0053957547061145306, + 0.006328057497739792, + 0.008030726574361324, + 0.007655775174498558, + 0.009820232167840004, + 0.004385247360914946, + 0.005696483422070742, + 0.003677366767078638, + 0.006131470203399658, + 0.0038632217328995466, + 0.007495007012039423, + 0.022428229451179504, + 0.014969556592404842, + 0.00685905059799552, + 0.004455411806702614, + 0.0350772887468338, + 0.005800078622996807, + 0.0038052592426538467, + 0.007896591909229755, + 0.007943360134959221, + 0.03941310569643974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.09920508414506912 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208357587456703 + ], + "opposite_logprob_tokens": [ + -2.8120834827423096 + ], + "logprob_reward": -0.06208357587456703, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208357587456703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09156117588281631, + 0.020035166293382645, + 0.02442777156829834, + 0.027401814237236977, + 0.5964903831481934, + 0.050537191331386566, + 0.04354431852698326, + 0.01961728185415268, + 0.03440673276782036, + 0.09197816997766495 + ], + "credit_map": [ + 0.009991040453314781, + 0.023913223296403885, + 0.00883788987994194, + 0.014120655134320259, + 0.008684087544679642, + 0.006716697942465544, + 0.010609864257276058, + 0.008687720634043217, + 0.0015287385322153568, + 0.004076298791915178, + 0.0015085074119269848, + 0.0018435567617416382, + 0.0023277734871953726, + 0.0031036282889544964, + 0.0028528112452477217, + 0.0027938522398471832, + 0.0022405015770345926, + 0.0033851810730993748, + 0.0031914685387164354, + 0.003522465005517006, + 0.0021696100011467934, + 0.0022248239256441593, + 0.004647061228752136, + 0.003046660218387842, + 0.0017556879902258515, + 0.005953609012067318, + 0.0017807686235755682, + 0.00297150993719697, + 0.0038859827909618616, + 0.0028383582830429077, + 0.004251585341989994, + 0.0039643109776079655, + 0.03730390965938568, + 0.0772574245929718, + 0.04507029801607132, + 0.11661382019519806, + 0.03378593176603317, + 0.025731386616826057, + 0.17409706115722656, + 0.08663057535886765, + 0.005175169091671705, + 0.0057250200770795345, + 0.00542893772944808, + 0.011271525174379349, + 0.005624655168503523, + 0.003798415418714285, + 0.007647163234651089, + 0.005866305902600288, + 0.0033157230354845524, + 0.007446033880114555, + 0.004810358863323927, + 0.004423671402037144, + 0.00326728168874979, + 0.0025950323324650526, + 0.009814230725169182, + 0.007871986366808414, + 0.0024039773270487785, + 0.003382429713383317, + 0.0015633752336725593, + 0.0019206530414521694, + 0.0027040820568799973, + 0.0026436776388436556, + 0.002676683943718672, + 0.0023224044125527143, + 0.0024518133141100407, + 0.0031802684534341097, + 0.0022998955100774765, + 0.0035466698464006186, + 0.003094316925853491, + 0.007382551673799753, + 0.008592672646045685, + 0.0038585448637604713, + 0.005318510811775923, + 0.003236661432310939, + 0.02692723460495472, + 0.004615855403244495, + 0.0028609836008399725, + 0.006750980392098427, + 0.00852844025939703, + 0.03373949974775314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.06208357587456703 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002503001829609275 + ], + "opposite_logprob_tokens": [ + -6.002502918243408 + ], + "logprob_reward": -0.002503001829609275, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002503001829609275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19769778847694397, + 0.06061101332306862, + 0.03963737189769745, + 0.04275653511285782, + 0.3707258105278015, + 0.02411886490881443, + 0.04487256705760956, + 0.03985476493835449, + 0.06894847750663757, + 0.11077680438756943 + ], + "credit_map": [ + 0.021115418523550034, + 0.052779823541641235, + 0.018878081813454628, + 0.02684018202126026, + 0.021688014268875122, + 0.018800171092152596, + 0.02114645577967167, + 0.016449635848402977, + 0.006116305477917194, + 0.008912230841815472, + 0.006387350615113974, + 0.0068729110062122345, + 0.0060307327657938, + 0.00849492009729147, + 0.011229131370782852, + 0.006567426957190037, + 0.0032046695705503225, + 0.004548144526779652, + 0.00621060561388731, + 0.005135932005941868, + 0.0037316582165658474, + 0.0036922041326761246, + 0.006676072254776955, + 0.006438087671995163, + 0.0022370812948793173, + 0.010972326621413231, + 0.0024260857608169317, + 0.0028586839325726032, + 0.00380146992392838, + 0.001825365936383605, + 0.013283169828355312, + 0.005352351348847151, + 0.02341262623667717, + 0.04492141306400299, + 0.02808261290192604, + 0.07400406897068024, + 0.023382052779197693, + 0.017519459128379822, + 0.10447345674037933, + 0.05493013560771942, + 0.002682881662622094, + 0.0032571209594607353, + 0.0025799379218369722, + 0.004229389131069183, + 0.0030111370142549276, + 0.002198370872065425, + 0.0028154035098850727, + 0.0033446233719587326, + 0.003063908778131008, + 0.008894885890185833, + 0.003547587199136615, + 0.00459928996860981, + 0.0037922034971415997, + 0.003079688875004649, + 0.009515495039522648, + 0.008379506878554821, + 0.002983041573315859, + 0.00612467247992754, + 0.0029978917445987463, + 0.0033257727045565844, + 0.004092673771083355, + 0.004705036990344524, + 0.007261967286467552, + 0.008363711647689342, + 0.004500845447182655, + 0.0045180185697972775, + 0.0035937896464020014, + 0.0037723092827945948, + 0.0043391031213104725, + 0.007174085825681686, + 0.02394402027130127, + 0.017106298357248306, + 0.006147427950054407, + 0.005802663508802652, + 0.02971564792096615, + 0.00858971755951643, + 0.005265549290925264, + 0.005760884843766689, + 0.01299014501273632, + 0.03650477156043053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.002503001829609275 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291219025850296 + ], + "opposite_logprob_tokens": [ + -3.7732911109924316 + ], + "logprob_reward": -0.023291219025850296, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291219025850296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11021021753549576, + 0.083334781229496, + 0.03895711898803711, + 0.0909724310040474, + 0.07725109159946442, + 0.029941877350211143, + 0.16554972529411316, + 0.03444381058216095, + 0.031181324273347855, + 0.338157594203949 + ], + "credit_map": [ + 0.009113829582929611, + 0.019295988604426384, + 0.015571637079119682, + 0.02017124742269516, + 0.009822993539273739, + 0.010220245458185673, + 0.01453474909067154, + 0.011479525826871395, + 0.008649391122162342, + 0.014298868365585804, + 0.010756607167422771, + 0.016257308423519135, + 0.006310074590146542, + 0.006382893770933151, + 0.010421130806207657, + 0.010258506052196026, + 0.0028765834867954254, + 0.004207049962133169, + 0.003487939480692148, + 0.006063054781407118, + 0.005524344276636839, + 0.005559149198234081, + 0.0064607467502355576, + 0.004778252448886633, + 0.004390666726976633, + 0.01042978186160326, + 0.004837207496166229, + 0.0104795778170228, + 0.006307796109467745, + 0.004137021489441395, + 0.03372364863753319, + 0.01666673831641674, + 0.0050966874696314335, + 0.011723997071385384, + 0.00605378532782197, + 0.014863036572933197, + 0.005439909175038338, + 0.0031603879760950804, + 0.01868082582950592, + 0.0122324638068676, + 0.0030107321217656136, + 0.00444630766287446, + 0.004736039787530899, + 0.004859484266489744, + 0.0035594156943261623, + 0.0021762712858617306, + 0.0038589320611208677, + 0.0032946954015642405, + 0.012129099108278751, + 0.03538982942700386, + 0.01701878383755684, + 0.017539536580443382, + 0.013919099234044552, + 0.009563357569277287, + 0.02510734833776951, + 0.034882672131061554, + 0.0031109778210520744, + 0.006271190941333771, + 0.0038872724398970604, + 0.006217842921614647, + 0.0028994090389460325, + 0.003249334404245019, + 0.00527305668219924, + 0.0035347254015505314, + 0.0023344873916357756, + 0.0037777279503643513, + 0.0025318663101643324, + 0.0032862466759979725, + 0.0028587235137820244, + 0.005063993390649557, + 0.007457105442881584, + 0.003871174529194832, + 0.019905192777514458, + 0.010072334669530392, + 0.10599610209465027, + 0.022210756316781044, + 0.010133525356650352, + 0.01877664215862751, + 0.029636012390255928, + 0.12142704427242279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.023291219025850296 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly (slow dolly in) into the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are distinct golden shafts of light with a visible Tyndall effect present in the dense autumn forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless golden autumn leaves with dry, curled edges seen drifting and twirling from above?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the ground covered by a thick blanket of crimson and orange leaves?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249881185591221 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005249881185591221, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249881185591221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049505505710840225, + 0.05443549528717995, + 0.035800326615571976, + 0.030764151364564896, + 0.12933284044265747, + 0.021743841469287872, + 0.04612298309803009, + 0.06275289505720139, + 0.056186493486166, + 0.5133554935455322 + ], + "credit_map": [ + 0.004725018050521612, + 0.006065841764211655, + 0.00742806913331151, + 0.008848057128489017, + 0.004740975331515074, + 0.005339895840734243, + 0.007179420441389084, + 0.005178227089345455, + 0.0039050746709108353, + 0.009245920926332474, + 0.007255274336785078, + 0.009420367889106274, + 0.003628453705459833, + 0.0050003319047391415, + 0.00848880410194397, + 0.007491263095289469, + 0.0025457695592194796, + 0.0037060189060866833, + 0.004390689078718424, + 0.004611121024936438, + 0.004031405318528414, + 0.0038605467416346073, + 0.007047941442579031, + 0.005606826860457659, + 0.0015752117615193129, + 0.006258740555495024, + 0.0015574664575979114, + 0.002854734193533659, + 0.0020828803535550833, + 0.00126518658362329, + 0.010323584079742432, + 0.004846345633268356, + 0.008842960000038147, + 0.01559901237487793, + 0.009881560690701008, + 0.024762684479355812, + 0.007462484762072563, + 0.006956067401915789, + 0.037588078528642654, + 0.01823998987674713, + 0.0027245241217315197, + 0.002399459481239319, + 0.0034897581208497286, + 0.003475803416222334, + 0.0017342290375381708, + 0.001391021185554564, + 0.003970164339989424, + 0.0025588811840862036, + 0.002651941729709506, + 0.013161160983145237, + 0.004136629402637482, + 0.0037289918400347233, + 0.003754274919629097, + 0.0021533675026148558, + 0.0024547509383410215, + 0.0140818627551198, + 0.006434121634811163, + 0.011457297019660473, + 0.006446272134780884, + 0.010848128236830235, + 0.005741282366216183, + 0.005253646057099104, + 0.009805863723158836, + 0.006766276899725199, + 0.0026955208741128445, + 0.004110697191208601, + 0.002990410663187504, + 0.004370937589555979, + 0.002560296794399619, + 0.005566440988332033, + 0.01992030255496502, + 0.013971887528896332, + 0.02945283241569996, + 0.01634790375828743, + 0.16096463799476624, + 0.034283630549907684, + 0.012946955859661102, + 0.026003915816545486, + 0.04567977041006088, + 0.18767580389976501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 7 + ], + "selected_reward": -0.005249881185591221 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_1/index_tensor([7920, 7920])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_1/index_tensor([7920, 7920])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8617210fdcdcd47436bf4fd82d508c62eaa2bf42 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_1/index_tensor([7920, 7920])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle shot, the camera slowly panning left, follows a group of heavily armored cavalrymen as they patrol the crest of a desolate wall constructed from massive, rough, black rock. The image is rendered in stunning high-contrast black and white, with a pure, bottomless black sky against the backdrop, rendering the cavalrymen's silhouettes as sharp as siloes. The cavalrymen are clad in full suits of intricately carved Gothic plate armor, their closed helmets revealing only narrow eye slits, obscuring their faces. Their majestic warhorses, similarly clad in heavy metal armor, trot in measured, ponderous steps. A sharp, cold light, from an unknown source, shines from the side, reflecting a dazzling, almost liquid-metal gleam off the edges of their armor and casting long, twisting black shadows behind them and across the weathered, cracked wall. This silent column advances with a steady, oppressive rhythm, spears pointed skyward, creating an atmosphere of austere, timeless, and threatening tranquility.", + "seed": 8832, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_1/index_7920_seed_8832_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.21927751321345568, + "vlm_logprob": -0.21927751321345568, + "vlm_yes_no_margin": 2.0999999940395355, + "vlm_token_credit": -0.21927751321345568, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34619882702827454 + ], + "opposite_logprob_tokens": [ + -1.3461987972259521 + ], + "logprob_reward": -0.34619882702827454, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.34619882702827454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07410062104463577, + 0.02311953343451023, + 0.036805279552936554, + 0.04669683054089546, + 0.037201277911663055, + 0.02561141364276409, + 0.013612638227641582, + 0.03017711080610752, + 0.6996297836303711, + 0.01304551586508751 + ], + "credit_map": [ + 0.011506333947181702, + 0.016384880989789963, + 0.005198757164180279, + 0.005136904306709766, + 0.008105006068944931, + 0.010828468948602676, + 0.007356876507401466, + 0.009583395905792713, + 0.004135939292609692, + 0.004296538420021534, + 0.00184551568236202, + 0.002415713854134083, + 0.0029857202898710966, + 0.003066206816583872, + 0.002159722615033388, + 0.002214177045971155, + 0.011924466118216515, + 0.008984829299151897, + 0.002612171694636345, + 0.004340134095400572, + 0.0035015540197491646, + 0.002403401304036379, + 0.001282919431105256, + 0.0017558049876242876, + 0.005024301819503307, + 0.007528463378548622, + 0.002991299144923687, + 0.002229641191661358, + 0.017033012583851814, + 0.006839212030172348, + 0.0028595197945833206, + 0.002191382460296154, + 0.007447284180670977, + 0.009016952477395535, + 0.004209588281810284, + 0.002312578959390521, + 0.006082328036427498, + 0.0037951464764773846, + 0.0029094512574374676, + 0.0014279481256380677, + 0.004493587650358677, + 0.003914779983460903, + 0.0063315508887171745, + 0.0016669140895828605, + 0.0032298897858709097, + 0.002161384094506502, + 0.001233514049090445, + 0.002579794265329838, + 0.0013872382696717978, + 0.0018027509795501828, + 0.0013491195859387517, + 0.0017956503434106708, + 0.001709208358079195, + 0.0018675007158890367, + 0.0015169335529208183, + 0.0021842368878424168, + 0.0028258650563657284, + 0.003614886198192835, + 0.0031131587456911802, + 0.005692204926162958, + 0.0028021165635436773, + 0.006121678277850151, + 0.001902943942695856, + 0.0041042594239115715, + 0.248224139213562, + 0.08000850677490234, + 0.17334546148777008, + 0.05397850647568703, + 0.022537153214216232, + 0.07020308822393417, + 0.016256505623459816, + 0.03507640212774277, + 0.0016957480693235993, + 0.0017920262180268764, + 0.0015549106756225228, + 0.0018858405528590083, + 0.0017221749294549227, + 0.0016889410326257348, + 0.0014166516484692693, + 0.0012892233207821846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.34619882702827454 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743451178073883 + ], + "opposite_logprob_tokens": [ + -0.9743450880050659 + ], + "logprob_reward": -0.4743451178073883, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.4743451178073883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027703918516635895, + 0.034157298505306244, + 0.10076581686735153, + 0.11761488765478134, + 0.12030751258134842, + 0.05790213122963905, + 0.011193233542144299, + 0.050333283841609955, + 0.4679507315158844, + 0.012071228586137295 + ], + "credit_map": [ + 0.003553972113877535, + 0.005934527609497309, + 0.0024954841937869787, + 0.0034694806672632694, + 0.002916456898674369, + 0.0041033863089978695, + 0.002924124011769891, + 0.002306488575413823, + 0.004403343889862299, + 0.004083998501300812, + 0.0026830818969756365, + 0.002649500500410795, + 0.005141566041857004, + 0.010936756618320942, + 0.001620957744307816, + 0.002638091566041112, + 0.035146672278642654, + 0.027673309668898582, + 0.006004049442708492, + 0.011662494391202927, + 0.008187023922801018, + 0.005113197956234217, + 0.003357738722115755, + 0.0036213239654898643, + 0.01307209674268961, + 0.01914050616323948, + 0.009009327739477158, + 0.00537912966683507, + 0.04248148575425148, + 0.015805847942829132, + 0.007387103978544474, + 0.005339386872947216, + 0.027705447748303413, + 0.029547277837991714, + 0.009175886400043964, + 0.008502872660756111, + 0.01442265510559082, + 0.017433861270546913, + 0.007329538464546204, + 0.006189972627907991, + 0.014943672344088554, + 0.0050581516698002815, + 0.019078822806477547, + 0.0032639377750456333, + 0.004415142349898815, + 0.002622800413519144, + 0.002264835638925433, + 0.006254769861698151, + 0.0014071034966036677, + 0.001575769274495542, + 0.0009920154698193073, + 0.001028961967676878, + 0.0016014489810913801, + 0.0027666043024510145, + 0.0007445030496455729, + 0.0010768277570605278, + 0.005188484210520983, + 0.0065682195127010345, + 0.004724583588540554, + 0.00597709184512496, + 0.00825495645403862, + 0.009927784092724323, + 0.0034349141642451286, + 0.0062572527676820755, + 0.15631937980651855, + 0.055654626339673996, + 0.13558737933635712, + 0.03652874752879143, + 0.013333304785192013, + 0.03625081107020378, + 0.011110816150903702, + 0.023165641352534294, + 0.0014403972309082747, + 0.0014530947664752603, + 0.001873775850981474, + 0.0013356863055378199, + 0.001577515504322946, + 0.0017383161466568708, + 0.0013277768157422543, + 0.0013246653834357858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 3 + ], + "selected_reward": -0.4743451178073883 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705078721046448 + ], + "opposite_logprob_tokens": [ + -2.1270508766174316 + ], + "logprob_reward": -0.12705078721046448, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705078721046448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0719035267829895, + 0.04549445956945419, + 0.09835144132375717, + 0.028401941061019897, + 0.06658802181482315, + 0.5657485723495483, + 0.022026414051651955, + 0.031230546534061432, + 0.04044381156563759, + 0.029811158776283264 + ], + "credit_map": [ + 0.0076288082636892796, + 0.010685370303690434, + 0.007010842673480511, + 0.00811290554702282, + 0.011675684712827206, + 0.013535278849303722, + 0.007139240391552448, + 0.006115402560681105, + 0.007736283354461193, + 0.007707521319389343, + 0.003167887683957815, + 0.005194464698433876, + 0.004977167584002018, + 0.009239548817276955, + 0.003881074720993638, + 0.0035905169788748026, + 0.024134933948516846, + 0.032879412174224854, + 0.0045286207459867, + 0.009662101976573467, + 0.008878416381776333, + 0.009410594590008259, + 0.0038042718078941107, + 0.005053102970123291, + 0.004752560053020716, + 0.004751506727188826, + 0.003769549774006009, + 0.0014351067366078496, + 0.004584160167723894, + 0.0054831127636134624, + 0.0018484005704522133, + 0.0017775465967133641, + 0.015473316423594952, + 0.017542360350489616, + 0.005190849769860506, + 0.004227144178003073, + 0.008009635843336582, + 0.008412934839725494, + 0.0037858965806663036, + 0.003945901058614254, + 0.18116091191768646, + 0.05069083347916603, + 0.2159963697195053, + 0.030036725103855133, + 0.032825373113155365, + 0.017780443653464317, + 0.01826637051999569, + 0.018991617485880852, + 0.0031398278661072254, + 0.003404913702979684, + 0.00218172837048769, + 0.0021575281862169504, + 0.0033641369082033634, + 0.004016245249658823, + 0.001726148184388876, + 0.0020358869805932045, + 0.0029447150882333517, + 0.004197556991130114, + 0.003588838502764702, + 0.0038714820984750986, + 0.0054009403102099895, + 0.0064103864133358, + 0.0023944491986185312, + 0.0024221809580922127, + 0.013566969893872738, + 0.004301757551729679, + 0.0031883888877928257, + 0.003257714444771409, + 0.002122800564393401, + 0.01044554729014635, + 0.0011901543475687504, + 0.002370485570281744, + 0.004084675572812557, + 0.0053901406936347485, + 0.00342003651894629, + 0.003651063423603773, + 0.003507922636345029, + 0.003524130443111062, + 0.002596878446638584, + 0.0036363149993121624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.12705078721046448 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954147316515446 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.005954147316515446, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954147316515446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08417206257581711, + 0.04811856150627136, + 0.20239421725273132, + 0.0313735269010067, + 0.07499478757381439, + 0.3953898847103119, + 0.019600089639425278, + 0.020661117509007454, + 0.10054483264684677, + 0.022750932723283768 + ], + "credit_map": [ + 0.012728366069495678, + 0.012591837905347347, + 0.0076166559010744095, + 0.008268634788691998, + 0.011998296715319157, + 0.01403319276869297, + 0.007472173310816288, + 0.009462903253734112, + 0.010406448505818844, + 0.00590687757357955, + 0.0045028990134596825, + 0.004148982465267181, + 0.006275280378758907, + 0.009635169059038162, + 0.0035812044516205788, + 0.0036616993602365255, + 0.06840857118368149, + 0.0585140623152256, + 0.013001698069274426, + 0.02105695754289627, + 0.01604280434548855, + 0.009824305772781372, + 0.006050421856343746, + 0.009495393373072147, + 0.0044929590076208115, + 0.004479182884097099, + 0.00387186580337584, + 0.0017741407500579953, + 0.0085871247574687, + 0.004320304840803146, + 0.0022695190273225307, + 0.0015784279676154256, + 0.017736388370394707, + 0.016821762546896935, + 0.0062751974910497665, + 0.0050372774712741375, + 0.009204191155731678, + 0.009918301366269588, + 0.005004337523132563, + 0.0049973358400166035, + 0.12154649943113327, + 0.034946221858263016, + 0.15100017189979553, + 0.020728442817926407, + 0.025774946436285973, + 0.01426539197564125, + 0.011066362261772156, + 0.01606183871626854, + 0.0038860293570905924, + 0.002957149874418974, + 0.0019334432436153293, + 0.0019412989495322108, + 0.002424747683107853, + 0.0029814932495355606, + 0.0015698879724368453, + 0.0019060381455346942, + 0.0021230208221822977, + 0.0033416005317121744, + 0.0022341704461723566, + 0.0027600345201790333, + 0.0028672039043158293, + 0.002750858198851347, + 0.0016719321720302105, + 0.0029122973792254925, + 0.03515297546982765, + 0.01103524025529623, + 0.026789845898747444, + 0.007491036783903837, + 0.002845743438228965, + 0.009891826659440994, + 0.002185693010687828, + 0.005152471829205751, + 0.003671039827167988, + 0.0034858970902860165, + 0.002280202927067876, + 0.003088350174948573, + 0.002613468561321497, + 0.002451819833368063, + 0.002513008425012231, + 0.0026471468154340982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 8 + ], + "selected_reward": -0.005954147316515446 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283868670463562 + ], + "opposite_logprob_tokens": [ + -2.017838716506958 + ], + "logprob_reward": -0.14283868670463562, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283868670463562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1097986027598381, + 0.040014758706092834, + 0.05940975993871689, + 0.040424540638923645, + 0.24972711503505707, + 0.1586902141571045, + 0.03132762014865875, + 0.02826746739447117, + 0.25688788294792175, + 0.02545202150940895 + ], + "credit_map": [ + 0.020665287971496582, + 0.03402518853545189, + 0.0063280682079494, + 0.006675645709037781, + 0.013294167816638947, + 0.014057104475796223, + 0.006818690802901983, + 0.007934458553791046, + 0.007094093598425388, + 0.006914315279573202, + 0.004289248492568731, + 0.005225586239248514, + 0.005304115824401379, + 0.004984925966709852, + 0.0032796929590404034, + 0.002922785934060812, + 0.015504059381783009, + 0.01191631332039833, + 0.0048656100407242775, + 0.008161254227161407, + 0.005234369542449713, + 0.007455389015376568, + 0.0028182566165924072, + 0.0034545143134891987, + 0.005284230224788189, + 0.005526923108845949, + 0.0025347606278955936, + 0.0021800759714096785, + 0.01452222466468811, + 0.005379699636250734, + 0.0032188817858695984, + 0.0017777489265426993, + 0.07462310045957565, + 0.08661481738090515, + 0.012694073840975761, + 0.00994663406163454, + 0.02520207315683365, + 0.023928169161081314, + 0.009487390518188477, + 0.0072308690287172794, + 0.048653099685907364, + 0.014015266671776772, + 0.05924252048134804, + 0.008249911479651928, + 0.009982554242014885, + 0.006136524025350809, + 0.004497349262237549, + 0.007913014851510525, + 0.004677632357925177, + 0.005997026339173317, + 0.004285633098334074, + 0.0020790614653378725, + 0.003913013730198145, + 0.005554595030844212, + 0.0021634437143802643, + 0.0026572176720947027, + 0.0028359771240502596, + 0.0032508536241948605, + 0.00263619190081954, + 0.0022803465835750103, + 0.00560900429263711, + 0.006929263938218355, + 0.002388973720371723, + 0.002336860168725252, + 0.08870477229356766, + 0.02590690553188324, + 0.02897706627845764, + 0.02070206217467785, + 0.010941791348159313, + 0.058773528784513474, + 0.008158711716532707, + 0.014723047614097595, + 0.0037944454234093428, + 0.005960326641798019, + 0.0028860673774033785, + 0.0029548206366598606, + 0.002517121843993664, + 0.0028404025360941887, + 0.0018037743866443634, + 0.002695065224543214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.14283868670463562 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, the camera slowly panning left, follows a group of heavily armored cavalrymen as they patrol the crest of a desolate wall constructed from massive, rough, black rock. The image is rendered in stunning high-contrast black and white, with a pure, bottomless black sky against the backdrop, rendering the cavalrymen's silhouettes as sharp as siloes. The cavalrymen are clad in full suits of intricately carved Gothic plate armor, their closed helmets revealing only narrow eye slits, obscuring their faces. Their majestic warhorses, similarly clad in heavy metal armor, trot in measured, ponderous steps. A sharp, cold light, from an unknown source, shines from the side, reflecting a dazzling, almost liquid-metal gleam off the edges of their armor and casting long, twisting black shadows behind them and across the weathered, cracked wall. This silent column advances with a steady, oppressive rhythm, spears pointed skyward, creating an atmosphere of austere, timeless, and threatening tranquility.", + "seed": 8833, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_1/index_7920_seed_8833_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.22576575502753257, + "vlm_logprob": -0.22576575502753257, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.22576575502753257, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4443906247615814 + ], + "opposite_logprob_tokens": [ + -1.1943906545639038 + ], + "logprob_reward": -0.4443906247615814, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.4443906247615814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33459073305130005, + 0.09818358719348907, + 0.017758043482899666, + 0.03541557490825653, + 0.037888046354055405, + 0.034171875566244125, + 0.027666941285133362, + 0.048516884446144104, + 0.33948957920074463, + 0.026318710297346115 + ], + "credit_map": [ + 0.02670915052294731, + 0.017970962449908257, + 0.06079983711242676, + 0.04921993613243103, + 0.017072182148694992, + 0.05056573078036308, + 0.057334501296281815, + 0.054918449372053146, + 0.006614090874791145, + 0.006844591815024614, + 0.020865119993686676, + 0.023467792198061943, + 0.004594684112817049, + 0.00814746506512165, + 0.015384778380393982, + 0.012265064753592014, + 0.0014054171042516828, + 0.0013616277137771249, + 0.0033786154817789793, + 0.0027474630624055862, + 0.0012273240135982633, + 0.0012940242886543274, + 0.0037339716218411922, + 0.002609601244330406, + 0.002805281663313508, + 0.002320503583177924, + 0.011127430014312267, + 0.002832251600921154, + 0.0013183873379603028, + 0.0011547190370038152, + 0.011334678158164024, + 0.0025223244447261095, + 0.0035553928464651108, + 0.0038202491123229265, + 0.0050855278968811035, + 0.004615572281181812, + 0.0036961669102311134, + 0.004703777842223644, + 0.007605362683534622, + 0.004805995617061853, + 0.002612548880279064, + 0.0021856052335351706, + 0.005473732482641935, + 0.008236747235059738, + 0.0025424996856600046, + 0.0041393195278942585, + 0.005068118218332529, + 0.003913304302841425, + 0.0017725799698382616, + 0.0015295639168471098, + 0.003741663880646229, + 0.0056508732959628105, + 0.002992055146023631, + 0.0013558119535446167, + 0.007845903746783733, + 0.002778490772470832, + 0.004038439132273197, + 0.003888959065079689, + 0.01164044439792633, + 0.007150786463171244, + 0.0037623709067702293, + 0.003815506584942341, + 0.007904542610049248, + 0.006315839011222124, + 0.010820579715073109, + 0.01891290210187435, + 0.1305244266986847, + 0.022341284900903702, + 0.0037023050244897604, + 0.015529884025454521, + 0.01800585724413395, + 0.11965232342481613, + 0.0037341860588639975, + 0.002122653881087899, + 0.004184175282716751, + 0.004410567693412304, + 0.0020198600832372904, + 0.0028046267107129097, + 0.003278000745922327, + 0.0037646396085619926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.4443906247615814 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762038230895996 + ], + "opposite_logprob_tokens": [ + -0.8262038230895996 + ], + "logprob_reward": -0.5762038230895996, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762038230895996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14668938517570496, + 0.16529899835586548, + 0.015383940190076828, + 0.0224783793091774, + 0.038595497608184814, + 0.03997374325990677, + 0.02319672517478466, + 0.05639243870973587, + 0.47846272587776184, + 0.01352820172905922 + ], + "credit_map": [ + 0.018061447888612747, + 0.014445710927248001, + 0.027020860463380814, + 0.033371634781360626, + 0.005893915891647339, + 0.014232655987143517, + 0.019401127472519875, + 0.014262013137340546, + 0.007930685766041279, + 0.005869381129741669, + 0.059257932007312775, + 0.06598857045173645, + 0.0039587621577084064, + 0.00748779671266675, + 0.010122451931238174, + 0.004683392588049173, + 0.0018859979463741183, + 0.0009643398807384074, + 0.0028886571526527405, + 0.00227069272659719, + 0.0011989203048869967, + 0.0010888807009905577, + 0.0034375516697764397, + 0.001648897654376924, + 0.0031916543375700712, + 0.0034177356865257025, + 0.003325119847431779, + 0.0029710796661674976, + 0.0021385622676461935, + 0.001361017464660108, + 0.0022026821970939636, + 0.003870524698868394, + 0.0030596060678362846, + 0.0034977723844349384, + 0.004727055784314871, + 0.004702184349298477, + 0.004749045707285404, + 0.003945597913116217, + 0.009031221270561218, + 0.004883007146418095, + 0.003632779698818922, + 0.0033118482679128647, + 0.005981961730867624, + 0.006651152390986681, + 0.003765243338420987, + 0.006365790963172913, + 0.005247202701866627, + 0.005017762538045645, + 0.0019263788126409054, + 0.0013741138391196728, + 0.0034244656562805176, + 0.003849853528663516, + 0.003351660445332527, + 0.001594783621840179, + 0.0048599704168736935, + 0.0028154957108199596, + 0.0042456937953829765, + 0.003880866803228855, + 0.014728664420545101, + 0.011497671715915203, + 0.003779154969379306, + 0.0025609228760004044, + 0.00860253069549799, + 0.007096921559423208, + 0.015417652204632759, + 0.026195697486400604, + 0.18311052024364471, + 0.03165178745985031, + 0.005800686776638031, + 0.021848639473319054, + 0.025272037833929062, + 0.1691656857728958, + 0.0012249047867953777, + 0.0009358101524412632, + 0.00207190727815032, + 0.001850017113611102, + 0.0018231349531561136, + 0.0017231814563274384, + 0.002025998430326581, + 0.0018732459284365177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.5762038230895996 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058931440114975 + ], + "opposite_logprob_tokens": [ + -3.8955893516540527 + ], + "logprob_reward": -0.02058931440114975, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058931440114975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17264653742313385, + 0.411200612783432, + 0.034324098378419876, + 0.03678317368030548, + 0.09080085158348083, + 0.025284457951784134, + 0.010528847575187683, + 0.09073945134878159, + 0.1104237288236618, + 0.017268259078264236 + ], + "credit_map": [ + 0.021192219108343124, + 0.02149139903485775, + 0.026144815608859062, + 0.028169572353363037, + 0.009430116973817348, + 0.024560471996665, + 0.018652454018592834, + 0.023005491122603416, + 0.018460597842931747, + 0.011641261167824268, + 0.1540287286043167, + 0.1680571585893631, + 0.00849893968552351, + 0.0165292639285326, + 0.02400902658700943, + 0.009975656867027283, + 0.0031375230755656958, + 0.001973724225535989, + 0.007226335350424051, + 0.005247479770332575, + 0.0023496572393924, + 0.002141959499567747, + 0.006968734785914421, + 0.0052786823362112045, + 0.0027356718201190233, + 0.002742933575063944, + 0.010642016306519508, + 0.0031108318362385035, + 0.0014283417258411646, + 0.0014800489880144596, + 0.01109995599836111, + 0.0035433764569461346, + 0.00968694407492876, + 0.007759865373373032, + 0.013089862652122974, + 0.009969189763069153, + 0.009673257358372211, + 0.007845254614949226, + 0.021607816219329834, + 0.01116865873336792, + 0.0027640529442578554, + 0.0018921553855761886, + 0.0041302843019366264, + 0.003999416716396809, + 0.002011250238865614, + 0.0016968912677839398, + 0.0052466909401118755, + 0.0035437194164842367, + 0.0009344079880975187, + 0.0006552248378284276, + 0.0017544860020279884, + 0.001924830605275929, + 0.0009335739305242896, + 0.0009214462479576468, + 0.0017954516224563122, + 0.0016094257589429617, + 0.00834045559167862, + 0.007673167157918215, + 0.020907221361994743, + 0.015088528394699097, + 0.005933416076004505, + 0.006447800900787115, + 0.015134942717850208, + 0.011213917285203934, + 0.003601987147703767, + 0.005774734541773796, + 0.04295153170824051, + 0.007669929414987564, + 0.0017024051630869508, + 0.005118071101605892, + 0.006805327720940113, + 0.03679974004626274, + 0.0015900685684755445, + 0.0015018995618447661, + 0.003317736554890871, + 0.00247954111546278, + 0.001990473596379161, + 0.001968476688489318, + 0.0021452349610626698, + 0.00227482826448977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.02058931440114975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646313101053238 + ], + "opposite_logprob_tokens": [ + -4.758646488189697 + ], + "logprob_reward": -0.008646313101053238, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646313101053238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14394797384738922, + 0.11396443098783493, + 0.0314258374273777, + 0.15145736932754517, + 0.04653612896800041, + 0.021410303190350533, + 0.018126914277672768, + 0.262626588344574, + 0.1962202936410904, + 0.014284096658229828 + ], + "credit_map": [ + 0.014239939860999584, + 0.010138829238712788, + 0.029682951048016548, + 0.037602499127388, + 0.005809983238577843, + 0.01251661404967308, + 0.022341091185808182, + 0.011616077274084091, + 0.007945685647428036, + 0.007829339243471622, + 0.029173780232667923, + 0.03272193297743797, + 0.0054484764114022255, + 0.008581188507378101, + 0.012470098212361336, + 0.009793933480978012, + 0.0024875507224351168, + 0.0016977782361209393, + 0.0059051308780908585, + 0.004964892286807299, + 0.0014548796461895108, + 0.0023411603178828955, + 0.006827507168054581, + 0.005746939685195684, + 0.007671814877539873, + 0.007032568566501141, + 0.03759673610329628, + 0.01626775786280632, + 0.006444733589887619, + 0.005761385429650545, + 0.05059746652841568, + 0.020084915682673454, + 0.00696314824745059, + 0.004987847991287708, + 0.007175830192863941, + 0.0056764199398458, + 0.003291185712441802, + 0.005051040090620518, + 0.006446730345487595, + 0.0069439224898815155, + 0.0014637132408097386, + 0.0017788081895560026, + 0.004029334522783756, + 0.0025652991607785225, + 0.0019122763769701123, + 0.002371747512370348, + 0.0029622872825711966, + 0.004326835740357637, + 0.0014766758540645242, + 0.001253862283192575, + 0.0022418322041630745, + 0.0033729986753314734, + 0.002074866322800517, + 0.0012084826594218612, + 0.004491770174354315, + 0.002006425289437175, + 0.02554509975016117, + 0.022594910115003586, + 0.04501901566982269, + 0.03617130219936371, + 0.022358881309628487, + 0.025179611518979073, + 0.05360192805528641, + 0.032155826687812805, + 0.006148551125079393, + 0.010681027546525002, + 0.07621795684099197, + 0.013384384103119373, + 0.002456999383866787, + 0.008999631740152836, + 0.011126809753477573, + 0.06720493733882904, + 0.0013652470661327243, + 0.0010474117007106543, + 0.00307448860257864, + 0.002179670613259077, + 0.001117947162128985, + 0.0017390017164871097, + 0.002017377410084009, + 0.0017429522704333067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.008646313101053238 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899869978427887 + ], + "opposite_logprob_tokens": [ + -2.5789988040924072 + ], + "logprob_reward": -0.07899869978427887, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899869978427887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0979643389582634, + 0.4253090023994446, + 0.05756894499063492, + 0.059794872999191284, + 0.060691703110933304, + 0.026643522083759308, + 0.02032564952969551, + 0.07966745644807816, + 0.14086240530014038, + 0.031172046437859535 + ], + "credit_map": [ + 0.009012156166136265, + 0.00861306767910719, + 0.015332923270761967, + 0.01248808391392231, + 0.008826956152915955, + 0.014273556880652905, + 0.014739676378667355, + 0.014677934348583221, + 0.01899007521569729, + 0.011898496188223362, + 0.15922138094902039, + 0.17468895018100739, + 0.008671974763274193, + 0.017321711406111717, + 0.024687984958291054, + 0.009828457608819008, + 0.005961326882243156, + 0.003323532408103347, + 0.01071332860738039, + 0.009851625189185143, + 0.0030301082879304886, + 0.003806301159784198, + 0.011355002410709858, + 0.009527723304927349, + 0.0037129134871065617, + 0.0025698915123939514, + 0.019330229610204697, + 0.005022051278501749, + 0.0018871311331167817, + 0.0013776462292298675, + 0.020441187545657158, + 0.005453827790915966, + 0.0057175979018211365, + 0.005553890019655228, + 0.007916240021586418, + 0.005953134037554264, + 0.007988065481185913, + 0.006950969807803631, + 0.01278991624712944, + 0.007821897976100445, + 0.002676689065992832, + 0.0019209423335269094, + 0.0036774855107069016, + 0.005905358586460352, + 0.0021084423642605543, + 0.0023342231288552284, + 0.004043059889227152, + 0.003977324813604355, + 0.001621859846636653, + 0.0011556172976270318, + 0.002835693536326289, + 0.0039050616323947906, + 0.0019645753782242537, + 0.001215711236000061, + 0.005011884029954672, + 0.0026152480859309435, + 0.00668872706592083, + 0.007503120228648186, + 0.014045311138033867, + 0.011629697866737843, + 0.006724239327013493, + 0.006925967056304216, + 0.01660669781267643, + 0.009543705731630325, + 0.004610978066921234, + 0.00740148825570941, + 0.05395219475030899, + 0.010155241936445236, + 0.0022476776503026485, + 0.006542566232383251, + 0.00815111305564642, + 0.04780115932226181, + 0.002833109814673662, + 0.0018962558824568987, + 0.0065987384878098965, + 0.004587553907185793, + 0.0023028117138892412, + 0.0038605693262070417, + 0.004247388802468777, + 0.004845621529966593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.07899869978427887 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, the camera slowly panning left, follows a group of heavily armored cavalrymen as they patrol the crest of a desolate wall constructed from massive, rough, black rock. The image is rendered in stunning high-contrast black and white, with a pure, bottomless black sky against the backdrop, rendering the cavalrymen's silhouettes as sharp as siloes. The cavalrymen are clad in full suits of intricately carved Gothic plate armor, their closed helmets revealing only narrow eye slits, obscuring their faces. Their majestic warhorses, similarly clad in heavy metal armor, trot in measured, ponderous steps. A sharp, cold light, from an unknown source, shines from the side, reflecting a dazzling, almost liquid-metal gleam off the edges of their armor and casting long, twisting black shadows behind them and across the weathered, cracked wall. This silent column advances with a steady, oppressive rhythm, spears pointed skyward, creating an atmosphere of austere, timeless, and threatening tranquility.", + "seed": 8834, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_1/index_7920_seed_8834_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10715165557339787, + "vlm_logprob": -0.10715165557339787, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.10715165557339787, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08393396437168121 + ], + "opposite_logprob_tokens": [ + -2.9589340686798096 + ], + "logprob_reward": -0.08393396437168121, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.08393396437168121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12926284968852997, + 0.10820190608501434, + 0.09924370050430298, + 0.1112985908985138, + 0.10831769555807114, + 0.22520634531974792, + 0.058709874749183655, + 0.06476568430662155, + 0.058437179774045944, + 0.0365561917424202 + ], + "credit_map": [ + 0.019460782408714294, + 0.021783486008644104, + 0.015548941679298878, + 0.017839187756180763, + 0.010632283054292202, + 0.007324212230741978, + 0.014484348706901073, + 0.02218961901962757, + 0.022577136754989624, + 0.016421757638454437, + 0.016786741092801094, + 0.013783277943730354, + 0.0083187036216259, + 0.008114284835755825, + 0.0053425319492816925, + 0.016857467591762543, + 0.020141983404755592, + 0.01413647085428238, + 0.017344873398542404, + 0.011377108283340931, + 0.006564267445355654, + 0.006270641461014748, + 0.01237056776881218, + 0.011037786491215229, + 0.023359162732958794, + 0.0177516620606184, + 0.015446938574314117, + 0.015862971544265747, + 0.0063596065156161785, + 0.008052943274378777, + 0.01235496811568737, + 0.01211034320294857, + 0.01644744910299778, + 0.023239677771925926, + 0.017303816974163055, + 0.008538752794265747, + 0.00968123134225607, + 0.007048164494335651, + 0.013504633679986, + 0.012553973123431206, + 0.0708632618188858, + 0.03556778281927109, + 0.01247276272624731, + 0.008803630247712135, + 0.012324905954301357, + 0.011261998675763607, + 0.013498097658157349, + 0.060413893312215805, + 0.012102242559194565, + 0.009276327677071095, + 0.008139223791658878, + 0.005255958531051874, + 0.006409480236470699, + 0.006208522245287895, + 0.00454657943919301, + 0.006771544925868511, + 0.01014044601470232, + 0.00911784078925848, + 0.010509959422051907, + 0.00845436379313469, + 0.005592053290456533, + 0.005378447473049164, + 0.00903761200606823, + 0.006534954532980919, + 0.010926409624516964, + 0.014800531789660454, + 0.009566057473421097, + 0.005667956545948982, + 0.005097458139061928, + 0.003040535608306527, + 0.00438131857663393, + 0.004956910386681557, + 0.004646527115255594, + 0.005166548769921064, + 0.006417074706405401, + 0.003261866280809045, + 0.00340348482131958, + 0.0034899727907031775, + 0.004374281968921423, + 0.005796431563794613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.08393396437168121 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34868693351745605 + ], + "opposite_logprob_tokens": [ + -1.223686933517456 + ], + "logprob_reward": -0.34868693351745605, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34868693351745605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10821636766195297, + 0.08995264768600464, + 0.08119718730449677, + 0.07564762979745865, + 0.10235824435949326, + 0.0808686763048172, + 0.041153114289045334, + 0.07722295075654984, + 0.11512233316898346, + 0.22826088964939117 + ], + "credit_map": [ + 0.01730826124548912, + 0.026836281642317772, + 0.01870816759765148, + 0.009696055203676224, + 0.010511435568332672, + 0.005681077018380165, + 0.009166182950139046, + 0.010308909229934216, + 0.020069260150194168, + 0.015519869513809681, + 0.009648905135691166, + 0.007141446694731712, + 0.010528317652642727, + 0.00742711964994669, + 0.0070228250697255135, + 0.012594903819262981, + 0.017096953466534615, + 0.013766737654805183, + 0.012613953091204166, + 0.011290996335446835, + 0.006843790877610445, + 0.005693906918168068, + 0.008273274637758732, + 0.005617580842226744, + 0.00801435112953186, + 0.008891630917787552, + 0.013899276033043861, + 0.015457501634955406, + 0.005247891414910555, + 0.006490413099527359, + 0.007657798007130623, + 0.009988765232264996, + 0.01508336327970028, + 0.018827907741069794, + 0.014975331723690033, + 0.011214815080165863, + 0.007569372188299894, + 0.008208107203245163, + 0.010766459628939629, + 0.01571287401020527, + 0.013633708469569683, + 0.011477784253656864, + 0.017335649579763412, + 0.006782398093491793, + 0.0061872550286352634, + 0.0042119151912629604, + 0.007125063333660364, + 0.014114901423454285, + 0.0066028316505253315, + 0.006518340203911066, + 0.005278824362903833, + 0.005147150252014399, + 0.003613216569647193, + 0.004668208304792643, + 0.004861382767558098, + 0.004463157616555691, + 0.018979743123054504, + 0.01240602321922779, + 0.015717431902885437, + 0.005544097628444433, + 0.00527442479506135, + 0.004476224072277546, + 0.00927272904664278, + 0.005552276503294706, + 0.02067706733942032, + 0.024956578388810158, + 0.011864563450217247, + 0.013779242523014545, + 0.009391488507390022, + 0.008526098914444447, + 0.01396428607404232, + 0.011963016353547573, + 0.030749278143048286, + 0.02795051969587803, + 0.029568150639533997, + 0.023163560777902603, + 0.01938430406153202, + 0.024554233998060226, + 0.03071816824376583, + 0.04217265918850899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.34868693351745605 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026354555040597916 + ], + "opposite_logprob_tokens": [ + -3.6513545513153076 + ], + "logprob_reward": -0.026354555040597916, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026354555040597916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23122283816337585, + 0.09310884773731232, + 0.12200026959180832, + 0.1318940371274948, + 0.07119789719581604, + 0.11085151880979538, + 0.02577761933207512, + 0.07453463971614838, + 0.026645779609680176, + 0.11276658624410629 + ], + "credit_map": [ + 0.03465253487229347, + 0.03178375959396362, + 0.026184801012277603, + 0.016614623367786407, + 0.03651462122797966, + 0.023961052298545837, + 0.019213104620575905, + 0.04229835048317909, + 0.018797028809785843, + 0.014101126231253147, + 0.009165971539914608, + 0.012859554961323738, + 0.00811452604830265, + 0.007047569844871759, + 0.00489240325987339, + 0.01813066564500332, + 0.024669546633958817, + 0.01980091817677021, + 0.02667846716940403, + 0.01533580757677555, + 0.007746079005300999, + 0.006845727562904358, + 0.011695485562086105, + 0.009228236973285675, + 0.02510010078549385, + 0.01929590106010437, + 0.017747223377227783, + 0.015174266882240772, + 0.007094548549503088, + 0.012912645004689693, + 0.020064907148480415, + 0.014504453167319298, + 0.009252405725419521, + 0.013319604098796844, + 0.007689288351684809, + 0.008034346625208855, + 0.005774879362434149, + 0.007357196882367134, + 0.010038713924586773, + 0.009731465019285679, + 0.018285665661096573, + 0.017554525285959244, + 0.012613550759851933, + 0.009680678136646748, + 0.0191548652946949, + 0.00812169536948204, + 0.01077539287507534, + 0.014665148220956326, + 0.0042244973592460155, + 0.0031292391940951347, + 0.005319107789546251, + 0.0025657766964286566, + 0.0030048510525375605, + 0.002093501389026642, + 0.0017896973295137286, + 0.0036509474739432335, + 0.012211058288812637, + 0.010885053314268589, + 0.014339704997837543, + 0.009354611858725548, + 0.0052414266392588615, + 0.005830937996506691, + 0.010524171404540539, + 0.006147671025246382, + 0.004317922983318567, + 0.0048727779649198055, + 0.004377767443656921, + 0.003435782389715314, + 0.0022034491412341595, + 0.0015766860451549292, + 0.0031211930327117443, + 0.0027402006089687347, + 0.0161497313529253, + 0.014762029983103275, + 0.0229922067373991, + 0.009504478424787521, + 0.008332347497344017, + 0.008568649180233479, + 0.013812156394124031, + 0.01864498294889927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.026354555040597916 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745780352503061 + ], + "opposite_logprob_tokens": [ + -5.0067458152771 + ], + "logprob_reward": -0.006745780352503061, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745780352503061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13342046737670898, + 0.24272283911705017, + 0.09738075733184814, + 0.07699425518512726, + 0.08283054828643799, + 0.06689572334289551, + 0.09987535327672958, + 0.09162739664316177, + 0.049790047109127045, + 0.05846261605620384 + ], + "credit_map": [ + 0.02295020967721939, + 0.019375678151845932, + 0.01664097234606743, + 0.01582942344248295, + 0.019135767593979836, + 0.009621364064514637, + 0.014747627079486847, + 0.015119434334337711, + 0.0503949299454689, + 0.043111830949783325, + 0.025793908163905144, + 0.03340822458267212, + 0.018041787669062614, + 0.016303187236189842, + 0.010812350548803806, + 0.04485661908984184, + 0.013932275585830212, + 0.014075576327741146, + 0.014638980850577354, + 0.018776768818497658, + 0.007707627955824137, + 0.0068472567945718765, + 0.013846704736351967, + 0.007555567659437656, + 0.010996496304869652, + 0.01052800752222538, + 0.010962781496345997, + 0.011306759901344776, + 0.006590630859136581, + 0.009749798104166985, + 0.009517653845250607, + 0.007342128548771143, + 0.012622789479792118, + 0.018979648128151894, + 0.019264433532953262, + 0.007451892364770174, + 0.005863310769200325, + 0.0070166573859751225, + 0.005234618671238422, + 0.0063972012139856815, + 0.01793435774743557, + 0.009188489057123661, + 0.007225402630865574, + 0.004464294295758009, + 0.0038067027926445007, + 0.003551680129021406, + 0.003797655925154686, + 0.016927136108279228, + 0.020629435777664185, + 0.015915757045149803, + 0.01738453470170498, + 0.011297236196696758, + 0.007968120276927948, + 0.007538651581853628, + 0.007718862034380436, + 0.011422764509916306, + 0.01951335184276104, + 0.014185169711709023, + 0.02030411921441555, + 0.007287425454705954, + 0.006511245388537645, + 0.007188147399574518, + 0.006268457509577274, + 0.010369482450187206, + 0.00993738416582346, + 0.012148784473538399, + 0.006497248541563749, + 0.005986180156469345, + 0.004027387592941523, + 0.002267827745527029, + 0.004343504551798105, + 0.004581727087497711, + 0.0083187660202384, + 0.011237643659114838, + 0.010902917943894863, + 0.004452764987945557, + 0.005131027661263943, + 0.004820375703275204, + 0.005741121247410774, + 0.00785799603909254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.006745780352503061 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003704458475113 + ], + "opposite_logprob_tokens": [ + -2.6950371265411377 + ], + "logprob_reward": -0.07003704458475113, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003704458475113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26650503277778625, + 0.09928083419799805, + 0.10382788628339767, + 0.04945743456482887, + 0.08446213603019714, + 0.13482290506362915, + 0.08539699763059616, + 0.03297016769647598, + 0.035780854523181915, + 0.1074957549571991 + ], + "credit_map": [ + 0.04204986244440079, + 0.05798470228910446, + 0.042877741158008575, + 0.035826001316308975, + 0.025198975577950478, + 0.012759806588292122, + 0.024965619668364525, + 0.024842338636517525, + 0.01892749033868313, + 0.012450492940843105, + 0.01264819037169218, + 0.015092342160642147, + 0.007960952818393707, + 0.006738739088177681, + 0.00472284946590662, + 0.020739775151014328, + 0.021686822175979614, + 0.016218017786741257, + 0.020638233050704002, + 0.01495805662125349, + 0.006824242416769266, + 0.005987429525703192, + 0.00943518616259098, + 0.00807990413159132, + 0.007764422334730625, + 0.0070988330990076065, + 0.005596652161329985, + 0.006379891186952591, + 0.004036104306578636, + 0.006097341887652874, + 0.007338377181440592, + 0.005145808681845665, + 0.012331461533904076, + 0.018238862976431847, + 0.015465809963643551, + 0.008112108334898949, + 0.006291494704782963, + 0.008407983928918839, + 0.00790373608469963, + 0.007710679899901152, + 0.027473272755742073, + 0.017118316143751144, + 0.014835220761597157, + 0.013032986782491207, + 0.013585775159299374, + 0.0075637283734977245, + 0.011883817613124847, + 0.029329797253012657, + 0.0102314418181777, + 0.011138154193758965, + 0.015087982639670372, + 0.008464490994811058, + 0.011826494708657265, + 0.009621787816286087, + 0.008428188040852547, + 0.01059845369309187, + 0.00443368311971426, + 0.004685368854552507, + 0.007316866423934698, + 0.003423519665375352, + 0.002069707028567791, + 0.0031269676983356476, + 0.004537766799330711, + 0.003376287641003728, + 0.006393346469849348, + 0.00786751788109541, + 0.0051222192123532295, + 0.0033214762806892395, + 0.002832340309396386, + 0.0021217160392552614, + 0.0035776938311755657, + 0.004544543102383614, + 0.013509840704500675, + 0.010622430592775345, + 0.01746833510696888, + 0.008727713488042355, + 0.009237267076969147, + 0.011388956569135189, + 0.01739225909113884, + 0.01914896070957184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.07003704458475113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, the camera slowly panning left, follows a group of heavily armored cavalrymen as they patrol the crest of a desolate wall constructed from massive, rough, black rock. The image is rendered in stunning high-contrast black and white, with a pure, bottomless black sky against the backdrop, rendering the cavalrymen's silhouettes as sharp as siloes. The cavalrymen are clad in full suits of intricately carved Gothic plate armor, their closed helmets revealing only narrow eye slits, obscuring their faces. Their majestic warhorses, similarly clad in heavy metal armor, trot in measured, ponderous steps. A sharp, cold light, from an unknown source, shines from the side, reflecting a dazzling, almost liquid-metal gleam off the edges of their armor and casting long, twisting black shadows behind them and across the weathered, cracked wall. This silent column advances with a steady, oppressive rhythm, spears pointed skyward, creating an atmosphere of austere, timeless, and threatening tranquility.", + "seed": 8835, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_1/index_7920_seed_8835_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.20851145647466182, + "vlm_logprob": -0.20851145647466182, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.20851145647466182, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4319169223308563 + ], + "opposite_logprob_tokens": [ + -1.1819169521331787 + ], + "logprob_reward": -0.4319169223308563, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.4319169223308563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3118761479854584, + 0.07848179340362549, + 0.06702941656112671, + 0.08550572395324707, + 0.036232300102710724, + 0.2722944915294647, + 0.03128471225500107, + 0.02786009944975376, + 0.04093465581536293, + 0.048500653356313705 + ], + "credit_map": [ + 0.06805054843425751, + 0.021408429369330406, + 0.024729415774345398, + 0.0413752943277359, + 0.023437708616256714, + 0.013361845165491104, + 0.09957908093929291, + 0.019933832809329033, + 0.004092229064553976, + 0.0040601096116006374, + 0.02630857191979885, + 0.024162009358406067, + 0.003695375518873334, + 0.004563646856695414, + 0.00590911228209734, + 0.005690744146704674, + 0.007179803214967251, + 0.009083556942641735, + 0.008254525251686573, + 0.0063456082716584206, + 0.006448067259043455, + 0.009404233656823635, + 0.009325872175395489, + 0.010987752117216587, + 0.010152077302336693, + 0.006920687388628721, + 0.01113065518438816, + 0.005955138243734837, + 0.007622353732585907, + 0.018646612763404846, + 0.014011915773153305, + 0.011066295206546783, + 0.0038636811077594757, + 0.002780139446258545, + 0.006766349542886019, + 0.005763895343989134, + 0.0038037863560020924, + 0.0037481177132576704, + 0.00398646853864193, + 0.005519863683730364, + 0.02995181456208229, + 0.020200828090310097, + 0.03105318918824196, + 0.030171481892466545, + 0.022217730060219765, + 0.08690080046653748, + 0.01963825896382332, + 0.03216040879487991, + 0.003352319123223424, + 0.0032644751481711864, + 0.00465962290763855, + 0.005585567560046911, + 0.0020888454746454954, + 0.004174685105681419, + 0.004174885340034962, + 0.003984312992542982, + 0.0027290259022265673, + 0.002814995124936104, + 0.0033195349387824535, + 0.004218722693622112, + 0.0023500153329223394, + 0.004103110171854496, + 0.004233697894960642, + 0.004090995993465185, + 0.005852276459336281, + 0.004051171708852053, + 0.009810715913772583, + 0.006048606708645821, + 0.0036697951145470142, + 0.0029338584281504154, + 0.003150560660287738, + 0.005417673382908106, + 0.006577543448656797, + 0.0033620416652411222, + 0.004230729769915342, + 0.005917027126997709, + 0.004250289406627417, + 0.004291513003408909, + 0.00830924417823553, + 0.011562271043658257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.4319169223308563 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233879089355469 + ], + "opposite_logprob_tokens": [ + -0.8983879089355469 + ], + "logprob_reward": -0.5233879089355469, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233879089355469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3378388285636902, + 0.16867828369140625, + 0.05624382942914963, + 0.0327145978808403, + 0.02847149595618248, + 0.07019995152950287, + 0.03046402335166931, + 0.028359366580843925, + 0.04121454805135727, + 0.20581501722335815 + ], + "credit_map": [ + 0.07816652953624725, + 0.021540135145187378, + 0.014941922388970852, + 0.053665801882743835, + 0.023200474679470062, + 0.010920328088104725, + 0.11795440316200256, + 0.01744924671947956, + 0.008905311115086079, + 0.008942030370235443, + 0.057845618575811386, + 0.05684648081660271, + 0.00732627185061574, + 0.006656123790889978, + 0.011920149438083172, + 0.010236305184662342, + 0.005661238916218281, + 0.006826637778431177, + 0.009737346321344376, + 0.004814489744603634, + 0.0075580887496471405, + 0.00864136777818203, + 0.0049377488903701305, + 0.008066913112998009, + 0.0032161602284759283, + 0.0027718967758119106, + 0.003832271322607994, + 0.004632466938346624, + 0.0027666206005960703, + 0.006142430007457733, + 0.004180431365966797, + 0.005172320641577244, + 0.003197338664904237, + 0.0028390127699822187, + 0.005058521404862404, + 0.0041097188368439674, + 0.003681878326460719, + 0.0028210217133164406, + 0.0031220915261656046, + 0.0036419122479856014, + 0.006457314360886812, + 0.006030683871358633, + 0.010808269493281841, + 0.009293272159993649, + 0.008011847734451294, + 0.011869510635733604, + 0.008972755633294582, + 0.008756299503147602, + 0.003560297656804323, + 0.004235510248690844, + 0.003265531500801444, + 0.005271616391837597, + 0.0028594760224223137, + 0.0044096712954342365, + 0.0040353997610509396, + 0.0028265209402889013, + 0.0031271828338503838, + 0.0027029633056372404, + 0.004169427324086428, + 0.0027682329528033733, + 0.0022703856229782104, + 0.004269481636583805, + 0.004023031331598759, + 0.005028663203120232, + 0.004493825137615204, + 0.003174252575263381, + 0.009371050633490086, + 0.004792527761310339, + 0.005140034947544336, + 0.003846011823043227, + 0.0029532741755247116, + 0.0074435691349208355, + 0.017215847969055176, + 0.011413000524044037, + 0.033729437738657, + 0.025916676968336105, + 0.017351174727082253, + 0.014855756424367428, + 0.0402800627052784, + 0.04505307972431183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.5233879089355469 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347007602453232 + ], + "opposite_logprob_tokens": [ + -3.6513469219207764 + ], + "logprob_reward": -0.026347007602453232, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347007602453232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37166914343833923, + 0.11617777496576309, + 0.05108790472149849, + 0.05958949774503708, + 0.15301179885864258, + 0.07727856934070587, + 0.04892558231949806, + 0.04056760296225548, + 0.02707202546298504, + 0.05462012439966202 + ], + "credit_map": [ + 0.08608301728963852, + 0.021109314635396004, + 0.020224839448928833, + 0.04622861370444298, + 0.02867160551249981, + 0.01840771548449993, + 0.12628401815891266, + 0.0246600154787302, + 0.006795179098844528, + 0.008319391869008541, + 0.018111441284418106, + 0.04174411669373512, + 0.006264565512537956, + 0.008388872258365154, + 0.013277044519782066, + 0.01327715627849102, + 0.007126957178115845, + 0.0068142167292535305, + 0.007403863128274679, + 0.004477646667510271, + 0.0060409861616790295, + 0.008439361117780209, + 0.0042715976014733315, + 0.006513278000056744, + 0.007391633000224829, + 0.0052373637445271015, + 0.007645214907824993, + 0.00435354420915246, + 0.005956721492111683, + 0.01322184782475233, + 0.008394774049520493, + 0.007388398516923189, + 0.015373867005109787, + 0.008550520054996014, + 0.04604407399892807, + 0.016952311620116234, + 0.011054505594074726, + 0.009980819188058376, + 0.011761333793401718, + 0.03329436108469963, + 0.00860090646892786, + 0.0075474511831998825, + 0.007310197222977877, + 0.010428572073578835, + 0.00847405381500721, + 0.015025732107460499, + 0.008321774192154408, + 0.011569879949092865, + 0.005765280220657587, + 0.005993107799440622, + 0.0069567980244755745, + 0.0075477478094398975, + 0.004000561311841011, + 0.0057993861846625805, + 0.00720784580335021, + 0.005654857028275728, + 0.004364801105111837, + 0.004396639298647642, + 0.0048852358013391495, + 0.0044906060211360455, + 0.005521162413060665, + 0.006227987818419933, + 0.004677347838878632, + 0.006003823131322861, + 0.003697779728099704, + 0.0030233978759497404, + 0.005196357145905495, + 0.003241835394874215, + 0.0029251209925860167, + 0.0028966229874640703, + 0.002323893364518881, + 0.0037670163437724113, + 0.01030756440013647, + 0.00269286148250103, + 0.01142493449151516, + 0.004864430520683527, + 0.0036278720945119858, + 0.0029900912195444107, + 0.010688855312764645, + 0.008023513481020927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.026347007602453232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952725186944008 + ], + "opposite_logprob_tokens": [ + -5.130952835083008 + ], + "logprob_reward": -0.005952725186944008, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952725186944008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16098712384700775, + 0.09603933244943619, + 0.04176974296569824, + 0.06961549818515778, + 0.06149224191904068, + 0.07184430956840515, + 0.3669727146625519, + 0.03197512403130531, + 0.025908920913934708, + 0.07339496165513992 + ], + "credit_map": [ + 0.03262360766530037, + 0.026817098259925842, + 0.015458120964467525, + 0.013621915131807327, + 0.017597245052456856, + 0.014905394986271858, + 0.022712863981723785, + 0.017250875011086464, + 0.0076621342450380325, + 0.008557063527405262, + 0.006413829047232866, + 0.026570996269583702, + 0.007622217293828726, + 0.009757639840245247, + 0.013550465926527977, + 0.01590498350560665, + 0.004068158101290464, + 0.006068456452339888, + 0.007245004177093506, + 0.0032755346037447453, + 0.003972885198891163, + 0.006212840788066387, + 0.00549187371507287, + 0.005434988997876644, + 0.00866318866610527, + 0.006414063740521669, + 0.008751814253628254, + 0.00596559839323163, + 0.006808007601648569, + 0.014759736135601997, + 0.008359489031136036, + 0.009893594309687614, + 0.007695557549595833, + 0.00406583771109581, + 0.009188885800540447, + 0.006940233055502176, + 0.005140781868249178, + 0.005430800840258598, + 0.005423601716756821, + 0.017606543377041817, + 0.008130289614200592, + 0.006168999243527651, + 0.008887791074812412, + 0.010207169689238071, + 0.006862937472760677, + 0.018489902839064598, + 0.0055820029228925705, + 0.007515220902860165, + 0.04483270272612572, + 0.03469497337937355, + 0.05615106225013733, + 0.052021369338035583, + 0.02398688718676567, + 0.039524901658296585, + 0.054049260914325714, + 0.06171157583594322, + 0.0032851072028279305, + 0.0034716834779828787, + 0.004232095554471016, + 0.004467076621949673, + 0.0039151571691036224, + 0.004608005750924349, + 0.0040351920761168, + 0.003960806410759687, + 0.00352609739638865, + 0.0022259356919676065, + 0.00504646822810173, + 0.003382185474038124, + 0.0022672819904983044, + 0.0029118694365024567, + 0.0020817446056753397, + 0.004467339254915714, + 0.0105800312012434, + 0.0038963938131928444, + 0.015865152701735497, + 0.008471588604152203, + 0.005172779317945242, + 0.006155119277536869, + 0.016158677637577057, + 0.0070952181704342365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.005952725186944008 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly pan to the left while following the group?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the subjects heavily armored cavalrymen patrolling a crest of a wall made of massive, rough, black rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the image rendered in high-contrast black and white with a pure, bottomless black sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the cavalrymen wear intricately carved Gothic plate armor with closed helmets featuring narrow eye slits?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549527183175087 + ], + "opposite_logprob_tokens": [ + -2.929952621459961 + ], + "logprob_reward": -0.0549527183175087, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549527183175087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2704337537288666, + 0.19427049160003662, + 0.08987635374069214, + 0.11128830909729004, + 0.08788013458251953, + 0.03942011669278145, + 0.03923703357577324, + 0.023918308317661285, + 0.0421454943716526, + 0.10153000056743622 + ], + "credit_map": [ + 0.06414887309074402, + 0.02213270217180252, + 0.01879376359283924, + 0.02712637558579445, + 0.022945892065763474, + 0.013281440362334251, + 0.0810769721865654, + 0.02092772349715233, + 0.01149554830044508, + 0.010326341725885868, + 0.03308042138814926, + 0.07633452862501144, + 0.008230074308812618, + 0.008460097014904022, + 0.020983370020985603, + 0.025360107421875, + 0.011521887965500355, + 0.010071727447211742, + 0.011024224571883678, + 0.00979540403932333, + 0.010380899533629417, + 0.010228031314909458, + 0.011492619290947914, + 0.015361555851995945, + 0.012317842803895473, + 0.008536497130990028, + 0.015319920144975185, + 0.012489238753914833, + 0.009960806928575039, + 0.022589614614844322, + 0.014124130830168724, + 0.015950262546539307, + 0.008351375348865986, + 0.00897914171218872, + 0.011352185159921646, + 0.012408952228724957, + 0.008331808261573315, + 0.009097045287489891, + 0.008764099329710007, + 0.020595531910657883, + 0.0038491548039019108, + 0.004043739289045334, + 0.003910999279469252, + 0.004764501471072435, + 0.004774916917085648, + 0.008433306589722633, + 0.004877178929746151, + 0.004766321275383234, + 0.005310241598635912, + 0.003950370941311121, + 0.0055930581875145435, + 0.00519614527001977, + 0.0028558755293488503, + 0.0041243755258619785, + 0.005305226426571608, + 0.006901742424815893, + 0.0022601555101573467, + 0.002566057723015547, + 0.0033259487245231867, + 0.004405213985592127, + 0.0025916241575032473, + 0.002402965212240815, + 0.0036642923951148987, + 0.002702051540836692, + 0.00634809723123908, + 0.004745684564113617, + 0.00478626461699605, + 0.006954417563974857, + 0.004252671264111996, + 0.005497022531926632, + 0.0031595875043421984, + 0.006401748396456242, + 0.016267092898488045, + 0.006070374045521021, + 0.013373267836868763, + 0.012292123399674892, + 0.008318858221173286, + 0.007178203202784061, + 0.022862065583467484, + 0.015168005600571632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0549527183175087 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_10/index_tensor([6864, 6864])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_10/index_tensor([6864, 6864])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f5c6e0736ad75761ca16da091fea1294cdfdac4e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_10/index_tensor([6864, 6864])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static, high-angle shot captures a stack of ancient official documents resting on a dark, polished wood surface, the entire scene rendered in ultra-high definition. Thick, cream-colored parchment is covered in dense black calligraphy, and several documents are stamped with fragile crimson wax seals with ornate crests. A single dramatic light source from the side casts deep shadows, accentuating the crisp texture of the paper and the raised relief of the seals, creating a tense and mysterious atmosphere. The transformation begins subtly as the paper fibers visibly darken and soften, the intricate text begins to bleed and flow, and the documents inexorably begin to liquefy. The solid wax seals melt and lose their shape, their crimson swirls reflecting the single light source on their smooth surface with pristine clarity as the entire wax seal gradually collapses into a viscous, ink-stained black puddle, slowly swirling with black eddies.", + "seed": 5043, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_10/index_6864_seed_5043_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01326109046349302, + "vlm_logprob": -0.01326109046349302, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.01326109046349302, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006931426469236612 + ], + "opposite_logprob_tokens": [ + -5.506931304931641 + ], + "logprob_reward": -0.006931426469236612, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006931426469236612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2439052164554596, + 0.10305210202932358, + 0.03001507930457592, + 0.18924392759799957, + 0.08098182082176208, + 0.08490658551454544, + 0.01853056624531746, + 0.07032730430364609, + 0.1011594906449318, + 0.0778779685497284 + ], + "credit_map": [ + 0.034358665347099304, + 0.04429980739951134, + 0.049545902758836746, + 0.022549834102392197, + 0.026154523715376854, + 0.0157051682472229, + 0.017996584996581078, + 0.03329472243785858, + 0.013841183856129646, + 0.024200471118092537, + 0.016879238188266754, + 0.010047735646367073, + 0.009143466129899025, + 0.006113175768405199, + 0.00999610684812069, + 0.01283072680234909, + 0.0037528558168560266, + 0.0035178628750145435, + 0.006034649908542633, + 0.00405924255028367, + 0.002991182031109929, + 0.002256047446280718, + 0.0035761662293225527, + 0.0038270708173513412, + 0.02614699676632881, + 0.030743369832634926, + 0.02759450115263462, + 0.023693086579442024, + 0.025631623342633247, + 0.01707283779978752, + 0.018543671816587448, + 0.019817838445305824, + 0.01003293413668871, + 0.009029320441186428, + 0.01401438843458891, + 0.012293304316699505, + 0.008225062862038612, + 0.006670080590993166, + 0.01336398720741272, + 0.007352737244218588, + 0.009846835397183895, + 0.013719388283789158, + 0.018818100914359093, + 0.010320322588086128, + 0.007237964775413275, + 0.007473511155694723, + 0.006932814605534077, + 0.010557647794485092, + 0.002694798167794943, + 0.00225273915566504, + 0.0018599223112687469, + 0.002769030397757888, + 0.0017746194498613477, + 0.0013619940727949142, + 0.003248324152082205, + 0.002569137839600444, + 0.008507844991981983, + 0.01007213443517685, + 0.012518856674432755, + 0.010922817513346672, + 0.005942058749496937, + 0.004539895337074995, + 0.011412154883146286, + 0.006411542650312185, + 0.0156293585896492, + 0.010136275552213192, + 0.014002895914018154, + 0.010917549021542072, + 0.010001631453633308, + 0.008740684948861599, + 0.012416431680321693, + 0.019314661622047424, + 0.009150388650596142, + 0.008500790223479271, + 0.009914088062942028, + 0.008827531710267067, + 0.006520451977849007, + 0.007669663522392511, + 0.013521424494683743, + 0.013773629441857338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.006931426469236612 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009334497735835612 + ], + "opposite_logprob_tokens": [ + -7.000933647155762 + ], + "logprob_reward": -0.0009334497735835612, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009334497735835612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13506415486335754, + 0.0531325601041317, + 0.0659167468547821, + 0.10296884924173355, + 0.10539893805980682, + 0.08844124525785446, + 0.08491683006286621, + 0.1511954665184021, + 0.08124098181724548, + 0.13172417879104614 + ], + "credit_map": [ + 0.015530548058450222, + 0.015524660237133503, + 0.023006819188594818, + 0.021714407950639725, + 0.013271484524011612, + 0.009230820462107658, + 0.018209848552942276, + 0.018575558438897133, + 0.005901313852518797, + 0.008568646386265755, + 0.008319735527038574, + 0.00736226886510849, + 0.005730839911848307, + 0.004788022954016924, + 0.0060464260168373585, + 0.006415307056158781, + 0.008817148394882679, + 0.008317619562149048, + 0.010023640468716621, + 0.006959425285458565, + 0.008727014996111393, + 0.005162058863788843, + 0.010680477134883404, + 0.0072293635457754135, + 0.013411395251750946, + 0.012992296367883682, + 0.015900269150733948, + 0.014945627190172672, + 0.013266129419207573, + 0.01099411305040121, + 0.01106638740748167, + 0.010392633266746998, + 0.012704742141067982, + 0.012197146192193031, + 0.011745198629796505, + 0.021940847858786583, + 0.01196879893541336, + 0.008057777769863605, + 0.019210632890462875, + 0.007573793176561594, + 0.010126974433660507, + 0.011751720681786537, + 0.012252289801836014, + 0.017080074176192284, + 0.007472230587154627, + 0.006234297528862953, + 0.016592856496572495, + 0.006930807139724493, + 0.008073939010500908, + 0.009854345582425594, + 0.009657165966928005, + 0.013988076709210873, + 0.007942483760416508, + 0.006801977753639221, + 0.019815796986222267, + 0.008783047087490559, + 0.01862667128443718, + 0.017596419900655746, + 0.03395933285355568, + 0.017489254474639893, + 0.01027989573776722, + 0.010811390355229378, + 0.026252374053001404, + 0.01618013344705105, + 0.013449637219309807, + 0.006924862507730722, + 0.00635753246024251, + 0.007779029663652182, + 0.00784621387720108, + 0.008308072574436665, + 0.01192796416580677, + 0.01864766702055931, + 0.014567303471267223, + 0.013890675269067287, + 0.023072119802236557, + 0.016253063455224037, + 0.01912570931017399, + 0.014547549188137054, + 0.01843300275504589, + 0.011834753677248955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0009334497735835612 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043073106557130814 + ], + "opposite_logprob_tokens": [ + -3.1680731773376465 + ], + "logprob_reward": -0.043073106557130814, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043073106557130814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07529736310243607, + 0.059528689831495285, + 0.029035493731498718, + 0.1242109015583992, + 0.07940616458654404, + 0.23856161534786224, + 0.033787764608860016, + 0.15436981618404388, + 0.06677795201539993, + 0.1390242725610733 + ], + "credit_map": [ + 0.010010948404669762, + 0.016062218695878983, + 0.014065105468034744, + 0.008067642338573933, + 0.009922110475599766, + 0.004122719168663025, + 0.00649828277528286, + 0.006548325065523386, + 0.008074788376688957, + 0.012527287937700748, + 0.009587020613253117, + 0.0061835614033043385, + 0.006528641562908888, + 0.0036257256288081408, + 0.005079847760498524, + 0.007921812124550343, + 0.003512674942612648, + 0.004746331367641687, + 0.004374661482870579, + 0.0036831004545092583, + 0.0030687330290675163, + 0.002257511019706726, + 0.004289495293051004, + 0.0031029852107167244, + 0.01787356473505497, + 0.015722664073109627, + 0.01815904676914215, + 0.01593216322362423, + 0.016223451122641563, + 0.011277031153440475, + 0.013444457203149796, + 0.015578514896333218, + 0.010759148746728897, + 0.00830986350774765, + 0.009928165934979916, + 0.012173347175121307, + 0.0089182173833251, + 0.005944904871284962, + 0.012728563509881496, + 0.01064394973218441, + 0.026936551555991173, + 0.03788306191563606, + 0.04609725996851921, + 0.029952023178339005, + 0.020499972626566887, + 0.022962262853980064, + 0.0213613398373127, + 0.0328691229224205, + 0.004793222527951002, + 0.003922284115105867, + 0.0044184233993291855, + 0.00626334547996521, + 0.002418635180220008, + 0.0020083619747310877, + 0.006448070053011179, + 0.003515419317409396, + 0.02043066918849945, + 0.026267170906066895, + 0.02713732421398163, + 0.015363750047981739, + 0.018451374024152756, + 0.013175991363823414, + 0.01589636132121086, + 0.017647160217165947, + 0.00922713615000248, + 0.0067676519975066185, + 0.012357393279671669, + 0.008067754097282887, + 0.005837358999997377, + 0.004862017463892698, + 0.008840464055538177, + 0.010818175971508026, + 0.02048998326063156, + 0.012659715488553047, + 0.017679663375020027, + 0.008791951462626457, + 0.010866343975067139, + 0.016061531379818916, + 0.018991267308592796, + 0.03348379582166672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.043073106557130814 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545241974294186 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545241974294186, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545241974294186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10324042290449142, + 0.06751115620136261, + 0.05756153166294098, + 0.1324620395898819, + 0.1123650074005127, + 0.06948252022266388, + 0.058600813150405884, + 0.2566995620727539, + 0.08344552665948868, + 0.05863139405846596 + ], + "credit_map": [ + 0.014104151166975498, + 0.020334480330348015, + 0.019232027232646942, + 0.012514936737716198, + 0.010633885860443115, + 0.007763626053929329, + 0.008047906681895256, + 0.01060941256582737, + 0.008333319798111916, + 0.011470826342701912, + 0.009662453085184097, + 0.009067717008292675, + 0.006062145344913006, + 0.006428498774766922, + 0.006854189094156027, + 0.009632011875510216, + 0.007481881882995367, + 0.009952888824045658, + 0.008189651183784008, + 0.007592581212520599, + 0.005441733170300722, + 0.005128319375216961, + 0.005541158374398947, + 0.00823331531137228, + 0.017823299393057823, + 0.020644687116146088, + 0.020720375701785088, + 0.017174821346998215, + 0.01735157147049904, + 0.012683019042015076, + 0.012503848411142826, + 0.013560401275753975, + 0.013174942694604397, + 0.013123070821166039, + 0.011764921247959137, + 0.02447528764605522, + 0.012726041488349438, + 0.00792333111166954, + 0.021291323006153107, + 0.007886093109846115, + 0.007809487171471119, + 0.0075001888908445835, + 0.006741466466337442, + 0.011887239292263985, + 0.006097760517150164, + 0.005528055131435394, + 0.015875546261668205, + 0.008042772300541401, + 0.0059121036902070045, + 0.007270168513059616, + 0.007897953502833843, + 0.009154186584055424, + 0.005610062275081873, + 0.004375368356704712, + 0.011952330358326435, + 0.006428636610507965, + 0.030867328867316246, + 0.020840050652623177, + 0.04571567103266716, + 0.03478425741195679, + 0.026180952787399292, + 0.020743388682603836, + 0.05557866021990776, + 0.021989239379763603, + 0.014764432795345783, + 0.0076078432612121105, + 0.009944066405296326, + 0.007002345286309719, + 0.007654325570911169, + 0.00839044339954853, + 0.011458913795650005, + 0.016623150557279587, + 0.009076721966266632, + 0.0055535039864480495, + 0.009925877675414085, + 0.004547921475023031, + 0.006042107008397579, + 0.006125969812273979, + 0.00801071897149086, + 0.009348574094474316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 4 + ], + "selected_reward": -0.012545241974294186 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002822227543219924 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.002822227543219924, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002822227543219924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08493973314762115, + 0.04763643071055412, + 0.02788531221449375, + 0.05334549397230148, + 0.09111054986715317, + 0.09553122520446777, + 0.10535311698913574, + 0.2564271092414856, + 0.12256909161806107, + 0.1152019277215004 + ], + "credit_map": [ + 0.009927675127983093, + 0.00899483822286129, + 0.009541552513837814, + 0.014844859950244427, + 0.009732655249536037, + 0.006714521441608667, + 0.015439046546816826, + 0.009744585491716862, + 0.005753094796091318, + 0.006664182525128126, + 0.0077115329913794994, + 0.004970434587448835, + 0.004652140196412802, + 0.004525607917457819, + 0.00733586261048913, + 0.006023573223501444, + 0.0032955200877040625, + 0.0028234513010829687, + 0.004167257808148861, + 0.003717355662956834, + 0.004119057673960924, + 0.0031887490767985582, + 0.003551982343196869, + 0.0030219396576285362, + 0.007159130647778511, + 0.007436052896082401, + 0.008612202480435371, + 0.008305401541292667, + 0.005725301802158356, + 0.005223738495260477, + 0.006138124968856573, + 0.004745543003082275, + 0.01030605286359787, + 0.010429694317281246, + 0.00925368256866932, + 0.01792200282216072, + 0.010577910579741001, + 0.0055534024722874165, + 0.019524430856108665, + 0.0075433747842907906, + 0.010538173839449883, + 0.011184162460267544, + 0.00872141495347023, + 0.017508601769804955, + 0.00935319159179926, + 0.006902626715600491, + 0.020995017141103745, + 0.010328039526939392, + 0.01126874703913927, + 0.009980512782931328, + 0.013276799581944942, + 0.017894087359309196, + 0.008677822537720203, + 0.007293464615941048, + 0.0231016892939806, + 0.013859993778169155, + 0.023934856057167053, + 0.022577056661248207, + 0.04387719929218292, + 0.037539370357990265, + 0.020913248881697655, + 0.017725402489304543, + 0.05671926587820053, + 0.03314070776104927, + 0.01689315214753151, + 0.013116907328367233, + 0.013481353409588337, + 0.016156746074557304, + 0.014031367376446724, + 0.012620940804481506, + 0.01897323504090309, + 0.017295390367507935, + 0.012721940875053406, + 0.011410214006900787, + 0.022813109681010246, + 0.013400224037468433, + 0.01544483844190836, + 0.012773857451975346, + 0.015409817919135094, + 0.011227928102016449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.002822227543219924 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static, high-angle shot captures a stack of ancient official documents resting on a dark, polished wood surface, the entire scene rendered in ultra-high definition. Thick, cream-colored parchment is covered in dense black calligraphy, and several documents are stamped with fragile crimson wax seals with ornate crests. A single dramatic light source from the side casts deep shadows, accentuating the crisp texture of the paper and the raised relief of the seals, creating a tense and mysterious atmosphere. The transformation begins subtly as the paper fibers visibly darken and soften, the intricate text begins to bleed and flow, and the documents inexorably begin to liquefy. The solid wax seals melt and lose their shape, their crimson swirls reflecting the single light source on their smooth surface with pristine clarity as the entire wax seal gradually collapses into a viscous, ink-stained black puddle, slowly swirling with black eddies.", + "seed": 5044, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_10/index_6864_seed_5044_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01664188797585666, + "vlm_logprob": -0.01664188797585666, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.01664188797585666, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020417165011167526 + ], + "opposite_logprob_tokens": [ + -4.020417213439941 + ], + "logprob_reward": -0.020417165011167526, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.020417165011167526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21387380361557007, + 0.11452758312225342, + 0.03657860681414604, + 0.06286019831895828, + 0.04629667475819588, + 0.056211918592453, + 0.07306420058012009, + 0.21730633080005646, + 0.05215846374630928, + 0.12712222337722778 + ], + "credit_map": [ + 0.02329348959028721, + 0.02746826969087124, + 0.014598081819713116, + 0.015155378729104996, + 0.03652644529938698, + 0.031624894589185715, + 0.026129625737667084, + 0.03907760977745056, + 0.015594416297972202, + 0.017898354679346085, + 0.008787818253040314, + 0.008205612190067768, + 0.013894003815948963, + 0.017795462161302567, + 0.015530967153608799, + 0.01682095415890217, + 0.006348532158881426, + 0.004139489494264126, + 0.003052512416616082, + 0.0029467965941876173, + 0.004262432921677828, + 0.005818706005811691, + 0.004700229503214359, + 0.005309905856847763, + 0.008570085279643536, + 0.008000894449651241, + 0.0052278172224760056, + 0.006747333332896233, + 0.008871819823980331, + 0.00954500213265419, + 0.0073465583845973015, + 0.008550687693059444, + 0.004612084943801165, + 0.005648218095302582, + 0.0031757892575114965, + 0.002996220486238599, + 0.01100996509194374, + 0.007877939380705357, + 0.004167548846453428, + 0.00680890865623951, + 0.005209227558225393, + 0.00879124365746975, + 0.004787044133991003, + 0.0035449450369924307, + 0.013820909895002842, + 0.007437258027493954, + 0.006434083450585604, + 0.006187207996845245, + 0.007566747255623341, + 0.012558013200759888, + 0.006818059831857681, + 0.005156753584742546, + 0.010843895375728607, + 0.014105776324868202, + 0.007445414550602436, + 0.00856954138725996, + 0.018601492047309875, + 0.04789334163069725, + 0.03337542712688446, + 0.017250869423151016, + 0.033692434430122375, + 0.0281074196100235, + 0.014654197730123997, + 0.023731140419840813, + 0.005803844425827265, + 0.00857347808778286, + 0.005715012550354004, + 0.00422031432390213, + 0.006645034532994032, + 0.006424862891435623, + 0.005407459568232298, + 0.009368455968797207, + 0.016918763518333435, + 0.022074375301599503, + 0.008456459268927574, + 0.00478318752720952, + 0.007305034436285496, + 0.050969477742910385, + 0.00667302031069994, + 0.009941897355020046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.020417165011167526 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005279527045786381 + ], + "opposite_logprob_tokens": [ + -5.255279541015625 + ], + "logprob_reward": -0.005279527045786381, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005279527045786381, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13609497249126434, + 0.15010955929756165, + 0.06971479952335358, + 0.1791590303182602, + 0.05777769535779953, + 0.05852093920111656, + 0.035766761749982834, + 0.11187703907489777, + 0.11652911454439163, + 0.08445005863904953 + ], + "credit_map": [ + 0.013786662369966507, + 0.015403460711240768, + 0.00929282233119011, + 0.009310085326433182, + 0.03009558469057083, + 0.021922491490840912, + 0.016206197440624237, + 0.020077677443623543, + 0.013344942592084408, + 0.025552432984113693, + 0.011117021553218365, + 0.010464975610375404, + 0.03519825264811516, + 0.02863480895757675, + 0.012210127897560596, + 0.013587015680968761, + 0.007454133592545986, + 0.009839095175266266, + 0.004661925602704287, + 0.004207038786262274, + 0.015943706035614014, + 0.009875995106995106, + 0.009081711992621422, + 0.00865120068192482, + 0.014656558632850647, + 0.029963674023747444, + 0.009416068904101849, + 0.012707004323601723, + 0.05607324838638306, + 0.024546638131141663, + 0.016556890681385994, + 0.015238966792821884, + 0.006062442436814308, + 0.008131880313158035, + 0.004569154232740402, + 0.0031268929596990347, + 0.01667102240025997, + 0.0076241218484938145, + 0.006145952735096216, + 0.005446232855319977, + 0.0050710272043943405, + 0.007478424347937107, + 0.004191066138446331, + 0.0028525469824671745, + 0.022310445085167885, + 0.007326739374548197, + 0.004647630266845226, + 0.004643065854907036, + 0.0032159541733562946, + 0.004761332646012306, + 0.004660881590098143, + 0.0023859611246734858, + 0.008446749299764633, + 0.004608175717294216, + 0.003702517133206129, + 0.003985189367085695, + 0.010390684939920902, + 0.015234195627272129, + 0.014037053100764751, + 0.005263164173811674, + 0.0266363937407732, + 0.01931663602590561, + 0.00904579646885395, + 0.01195311825722456, + 0.014075072482228279, + 0.015438178554177284, + 0.014673284254968166, + 0.009647357277572155, + 0.01091666892170906, + 0.022135598585009575, + 0.010939705185592175, + 0.018703266978263855, + 0.007425815798342228, + 0.01772361248731613, + 0.008086543530225754, + 0.00543991569429636, + 0.00658860569819808, + 0.024575769901275635, + 0.00650767982006073, + 0.008102118968963623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.005279527045786381 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371965512633324 + ], + "opposite_logprob_tokens": [ + -3.408719539642334 + ], + "logprob_reward": -0.03371965512633324, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371965512633324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17514736950397491, + 0.07706501334905624, + 0.03365490585565567, + 0.060729313641786575, + 0.05187777429819107, + 0.046612467616796494, + 0.06742419302463531, + 0.21315494179725647, + 0.1885240226984024, + 0.08581002056598663 + ], + "credit_map": [ + 0.02005789801478386, + 0.021667078137397766, + 0.012594147585332394, + 0.014278237707912922, + 0.029680462554097176, + 0.026007378473877907, + 0.021767597645521164, + 0.02909456565976143, + 0.010372176766395569, + 0.01288494374603033, + 0.005138860549777746, + 0.004800027702003717, + 0.01013745367527008, + 0.012886793352663517, + 0.010573632083833218, + 0.010271121747791767, + 0.005487429443746805, + 0.00305872387252748, + 0.0027205722872167826, + 0.0021562310867011547, + 0.004868256393820047, + 0.004888980183750391, + 0.006289086304605007, + 0.004185623489320278, + 0.005479611922055483, + 0.008299247361719608, + 0.0035188726615160704, + 0.006182487588375807, + 0.018047567456960678, + 0.008001324720680714, + 0.005592056550085545, + 0.005608138628304005, + 0.006331596523523331, + 0.00707028154283762, + 0.003034407738596201, + 0.0032841351348906755, + 0.011884215287864208, + 0.008128457702696323, + 0.003575647948309779, + 0.008569028228521347, + 0.006658083293586969, + 0.010881079360842705, + 0.002947448752820492, + 0.0037787638138979673, + 0.006368402391672134, + 0.005738787818700075, + 0.006689932197332382, + 0.003549967659637332, + 0.0069602192379534245, + 0.008701220154762268, + 0.007668613456189632, + 0.004984476137906313, + 0.009753018617630005, + 0.013267283327877522, + 0.006384555250406265, + 0.009704798460006714, + 0.023340942338109016, + 0.0259273499250412, + 0.025651054456830025, + 0.013290176168084145, + 0.04476911574602127, + 0.03878588601946831, + 0.0175433661788702, + 0.023847028613090515, + 0.021205134689807892, + 0.028661100193858147, + 0.02306695654988289, + 0.013783127069473267, + 0.031154710799455643, + 0.03840095177292824, + 0.013077954761683941, + 0.019174063578248024, + 0.008696265518665314, + 0.019268114119768143, + 0.007500160951167345, + 0.0048372638411819935, + 0.009524030610918999, + 0.016990305855870247, + 0.008666342124342918, + 0.010327527299523354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.03371965512633324 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020593751221895218 + ], + "opposite_logprob_tokens": [ + -3.8955936431884766 + ], + "logprob_reward": -0.020593751221895218, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020593751221895218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15104639530181885, + 0.0442507378757, + 0.03912395238876343, + 0.04780307039618492, + 0.053873684257268906, + 0.04900383949279785, + 0.06316336244344711, + 0.30468985438346863, + 0.11722967773675919, + 0.12981542944908142 + ], + "credit_map": [ + 0.01544085331261158, + 0.01770438626408577, + 0.008788150735199451, + 0.011167080141603947, + 0.0393625907599926, + 0.02288060635328293, + 0.01619294472038746, + 0.01950978673994541, + 0.0077422866597771645, + 0.007552324794232845, + 0.002997869858518243, + 0.003950406331568956, + 0.006569046061486006, + 0.005917542148381472, + 0.004811974707990885, + 0.004709291271865368, + 0.007715189829468727, + 0.0039856163784861565, + 0.003754835808649659, + 0.003161650849506259, + 0.004430885426700115, + 0.005810344126075506, + 0.006224519573152065, + 0.004040910396724939, + 0.004964836407452822, + 0.006314368452876806, + 0.002937064506113529, + 0.0038519827648997307, + 0.012575062923133373, + 0.0069975098595023155, + 0.005753983277827501, + 0.00440826453268528, + 0.004670530091971159, + 0.007347509264945984, + 0.0038317798171192408, + 0.0030929245986044407, + 0.016674000769853592, + 0.007958412170410156, + 0.005646364763379097, + 0.004652164876461029, + 0.003783775959163904, + 0.009091207757592201, + 0.005564677529036999, + 0.003019806230440736, + 0.011200275272130966, + 0.0064977542497217655, + 0.005473917815834284, + 0.004372426308691502, + 0.007274622097611427, + 0.00661617424339056, + 0.009070740081369877, + 0.0044303229078650475, + 0.008642112836241722, + 0.01151884626597166, + 0.004889042116701603, + 0.010721509344875813, + 0.02094285935163498, + 0.05913756042718887, + 0.027286594733595848, + 0.02255120314657688, + 0.0695156455039978, + 0.05830748751759529, + 0.022850647568702698, + 0.024097859859466553, + 0.014384768903255463, + 0.01874295435845852, + 0.013733230531215668, + 0.008076056838035583, + 0.010431695729494095, + 0.021450495347380638, + 0.010914530605077744, + 0.01949596218764782, + 0.013375034555792809, + 0.025424832478165627, + 0.01114710420370102, + 0.006981728598475456, + 0.011349356733262539, + 0.04045480862259865, + 0.008455961011350155, + 0.012626617215573788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.020593751221895218 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031993414741009474 + ], + "opposite_logprob_tokens": [ + -5.753199577331543 + ], + "logprob_reward": -0.0031993414741009474, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031993414741009474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13843929767608643, + 0.03726857900619507, + 0.032744016498327255, + 0.04700854420661926, + 0.04589734598994255, + 0.05371205508708954, + 0.06195991858839989, + 0.3510824739933014, + 0.15630468726158142, + 0.07558304816484451 + ], + "credit_map": [ + 0.011518435552716255, + 0.016419151797890663, + 0.008859303779900074, + 0.009923620149493217, + 0.036867085844278336, + 0.020085781812667847, + 0.01554662361741066, + 0.01921929232776165, + 0.004427846055477858, + 0.0061569856479763985, + 0.002653986681252718, + 0.0027790903113782406, + 0.00748031260445714, + 0.005501091480255127, + 0.004706988576799631, + 0.0035622783470898867, + 0.003478966187685728, + 0.003749219235032797, + 0.002418575109913945, + 0.002326754154637456, + 0.006823791656643152, + 0.004688485525548458, + 0.005239545833319426, + 0.004018679261207581, + 0.005474028177559376, + 0.007795287761837244, + 0.002782432595267892, + 0.0035223609302192926, + 0.009796211495995522, + 0.0069533889181911945, + 0.005892147310078144, + 0.004792685620486736, + 0.004311075434088707, + 0.005970736034214497, + 0.003971070051193237, + 0.002838932443410158, + 0.013188125565648079, + 0.005792008712887764, + 0.005310763139277697, + 0.004514635540544987, + 0.005371519830077887, + 0.006235755514353514, + 0.00543154077604413, + 0.003452551318332553, + 0.018779881298542023, + 0.00700030941516161, + 0.0038391256239265203, + 0.003601371543481946, + 0.006522504147142172, + 0.006599946413189173, + 0.008347032591700554, + 0.0048244670033454895, + 0.010136806406080723, + 0.012413310818374157, + 0.005069790408015251, + 0.008046063594520092, + 0.02475089766085148, + 0.06378377228975296, + 0.0316031239926815, + 0.02454036846756935, + 0.07753258943557739, + 0.06759372353553772, + 0.028525274246931076, + 0.03275274857878685, + 0.017446978017687798, + 0.026719143614172935, + 0.020192371681332588, + 0.009815292432904243, + 0.016161635518074036, + 0.02954719588160515, + 0.012596682645380497, + 0.0238253865391016, + 0.006742538418620825, + 0.013786237686872482, + 0.007329493761062622, + 0.005770338699221611, + 0.00904808659106493, + 0.01842685230076313, + 0.006083218846470118, + 0.008396285586059093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.0031993414741009474 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static, high-angle shot captures a stack of ancient official documents resting on a dark, polished wood surface, the entire scene rendered in ultra-high definition. Thick, cream-colored parchment is covered in dense black calligraphy, and several documents are stamped with fragile crimson wax seals with ornate crests. A single dramatic light source from the side casts deep shadows, accentuating the crisp texture of the paper and the raised relief of the seals, creating a tense and mysterious atmosphere. The transformation begins subtly as the paper fibers visibly darken and soften, the intricate text begins to bleed and flow, and the documents inexorably begin to liquefy. The solid wax seals melt and lose their shape, their crimson swirls reflecting the single light source on their smooth surface with pristine clarity as the entire wax seal gradually collapses into a viscous, ink-stained black puddle, slowly swirling with black eddies.", + "seed": 5045, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_10/index_6864_seed_5045_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13572216322645544, + "vlm_logprob": -0.13572216322645544, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.13572216322645544, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47723978757858276 + ], + "opposite_logprob_tokens": [ + -0.9772397875785828 + ], + "logprob_reward": -0.47723978757858276, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47723978757858276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11103519052267075, + 0.0733347237110138, + 0.03363465145230293, + 0.057062093168497086, + 0.06144857034087181, + 0.026881378144025803, + 0.17660091817378998, + 0.1588909924030304, + 0.2104025036096573, + 0.09070895612239838 + ], + "credit_map": [ + 0.010867835022509098, + 0.015196372754871845, + 0.02511097677052021, + 0.015238260850310326, + 0.009958243928849697, + 0.008724621497094631, + 0.013265087269246578, + 0.012673801742494106, + 0.008334686048328876, + 0.010263647884130478, + 0.016838030889630318, + 0.009335607290267944, + 0.005971509497612715, + 0.006752180866897106, + 0.007829449139535427, + 0.00800961535423994, + 0.003978040535002947, + 0.0041872719302773476, + 0.003907243255525827, + 0.005964066367596388, + 0.003989247605204582, + 0.0028743362054228783, + 0.005305891390889883, + 0.0034285567235201597, + 0.004360583610832691, + 0.004856965038925409, + 0.013772579841315746, + 0.009755722247064114, + 0.005010593682527542, + 0.003970712423324585, + 0.006893788930028677, + 0.008441150188446045, + 0.006169300060719252, + 0.006693475414067507, + 0.011767641641199589, + 0.01306920126080513, + 0.006070534698665142, + 0.003561869030818343, + 0.009038281626999378, + 0.005078275687992573, + 0.0027702990919351578, + 0.004850985482335091, + 0.003304832149296999, + 0.004030571784824133, + 0.004099118057638407, + 0.0020761762280017138, + 0.0028181299567222595, + 0.002931267721578479, + 0.019646156579256058, + 0.015621593222022057, + 0.03920067474246025, + 0.04892423376441002, + 0.008952385745942593, + 0.010677466168999672, + 0.02125033736228943, + 0.012328088283538818, + 0.017286386340856552, + 0.019531013444066048, + 0.015449905768036842, + 0.025209058076143265, + 0.03304610773921013, + 0.0122812669724226, + 0.020954739302396774, + 0.015132540836930275, + 0.02946045994758606, + 0.024324815720319748, + 0.03868694603443146, + 0.02831856906414032, + 0.02660752646625042, + 0.017748311161994934, + 0.026414072141051292, + 0.0188418198376894, + 0.008080721832811832, + 0.006794197019189596, + 0.00601513497531414, + 0.008866915479302406, + 0.007102433126419783, + 0.004442212637513876, + 0.031011080369353294, + 0.018396275117993355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.47723978757858276 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00766869867220521 + ], + "opposite_logprob_tokens": [ + -4.882668495178223 + ], + "logprob_reward": -0.00766869867220521, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00766869867220521, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08082260191440582, + 0.10006959736347198, + 0.06226474791765213, + 0.05055227503180504, + 0.057580187916755676, + 0.03775689750909805, + 0.11233102530241013, + 0.12083160132169724, + 0.1449665129184723, + 0.23282454907894135 + ], + "credit_map": [ + 0.0069707296788692474, + 0.009985002689063549, + 0.01369617972522974, + 0.013278857804834843, + 0.009384018369019032, + 0.006999887526035309, + 0.011329728178679943, + 0.009178202599287033, + 0.010134962387382984, + 0.013382398523390293, + 0.024260198697447777, + 0.012769355438649654, + 0.008749924600124359, + 0.009060915559530258, + 0.010479683056473732, + 0.01123217586427927, + 0.004934923257678747, + 0.007156909443438053, + 0.009308185428380966, + 0.006399034988135099, + 0.008543388918042183, + 0.005687950178980827, + 0.01223279070109129, + 0.008001567795872688, + 0.0036773979663848877, + 0.005386113189160824, + 0.011094371788203716, + 0.005549473688006401, + 0.005960267502814531, + 0.003846127539873123, + 0.007134797517210245, + 0.00790372770279646, + 0.005974473897367716, + 0.007371495943516493, + 0.008828861638903618, + 0.010897304862737656, + 0.005931180901825428, + 0.003702538786455989, + 0.009573820047080517, + 0.005300514865666628, + 0.0031335237435996532, + 0.005302289966493845, + 0.0042684911750257015, + 0.0056047700345516205, + 0.005361168645322323, + 0.005724912043660879, + 0.005205383989959955, + 0.003156360238790512, + 0.01260040421038866, + 0.010028747841715813, + 0.0234146099537611, + 0.02543802000582218, + 0.007627138402312994, + 0.011586609296500683, + 0.011976249516010284, + 0.009659255854785442, + 0.01355940755456686, + 0.013556824997067451, + 0.01057467982172966, + 0.01996796764433384, + 0.023825623095035553, + 0.011486324481666088, + 0.016729529947042465, + 0.011131259612739086, + 0.02375924214720726, + 0.016236500814557076, + 0.016076816245913506, + 0.02217131108045578, + 0.013589178211987019, + 0.014386534690856934, + 0.02110879495739937, + 0.017638150602579117, + 0.01877991482615471, + 0.01153306383639574, + 0.014055393636226654, + 0.021219635382294655, + 0.01689736731350422, + 0.011052776128053665, + 0.0868140161037445, + 0.05247238650918007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.00766869867220521 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283227920532227 + ], + "opposite_logprob_tokens": [ + -2.0178322792053223 + ], + "logprob_reward": -0.14283227920532227, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283227920532227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028676308691501617, + 0.016333455219864845, + 0.03013507090508938, + 0.02441146783530712, + 0.0265052393078804, + 0.0227076206356287, + 0.2220655232667923, + 0.1031218096613884, + 0.1516331285238266, + 0.37441036105155945 + ], + "credit_map": [ + 0.0045475163497030735, + 0.004099265206605196, + 0.003690240904688835, + 0.002535759238526225, + 0.00340228620916605, + 0.0034882016479969025, + 0.0035957570653408766, + 0.003317283932119608, + 0.001847703941166401, + 0.00253389123827219, + 0.0032584366854280233, + 0.0019949213601648808, + 0.0017081840196624398, + 0.0013166754506528378, + 0.002089576330035925, + 0.0015840664273127913, + 0.0029157723765820265, + 0.003444293513894081, + 0.0026921711396425962, + 0.004632958211004734, + 0.0041039129719138145, + 0.00278597604483366, + 0.006135785020887852, + 0.0034242032561451197, + 0.0021560429595410824, + 0.0025200669188052416, + 0.003652154002338648, + 0.003197034355252981, + 0.0044749341905117035, + 0.001753299729898572, + 0.00394402164965868, + 0.0027139142621308565, + 0.0024558387231081724, + 0.0028538478072732687, + 0.004672190640121698, + 0.005545059684664011, + 0.0028957652393728495, + 0.0017444990808144212, + 0.0037808052729815245, + 0.0025572343729436398, + 0.0021486161276698112, + 0.0029541703406721354, + 0.0025815244298428297, + 0.0029949862509965897, + 0.003678795648738742, + 0.0023391873110085726, + 0.0027299041394144297, + 0.0032804368529468775, + 0.02487415075302124, + 0.02171320654451847, + 0.046790242195129395, + 0.06505754590034485, + 0.012705509550869465, + 0.008402151055634022, + 0.027334989979863167, + 0.01518772728741169, + 0.009928110055625439, + 0.013397596776485443, + 0.009833628311753273, + 0.01437329687178135, + 0.021951641887426376, + 0.008502334356307983, + 0.014985887333750725, + 0.01014931034296751, + 0.03404279053211212, + 0.012589773163199425, + 0.030181527137756348, + 0.02023950219154358, + 0.011174088343977928, + 0.009915932081639767, + 0.019476477056741714, + 0.014013045467436314, + 0.030428094789385796, + 0.016580382362008095, + 0.023162420839071274, + 0.034214090555906296, + 0.01498120091855526, + 0.01823665387928486, + 0.15141227841377258, + 0.0853952169418335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.14283227920532227 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867067560553551 + ], + "opposite_logprob_tokens": [ + -3.048670768737793 + ], + "logprob_reward": -0.04867067560553551, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867067560553551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04910198971629143, + 0.03850317746400833, + 0.04201933369040489, + 0.05826059728860855, + 0.04531489312648773, + 0.05389673635363579, + 0.17129157483577728, + 0.06283260136842728, + 0.11393793672323227, + 0.36484116315841675 + ], + "credit_map": [ + 0.005152266472578049, + 0.008063524030148983, + 0.008345507085323334, + 0.005717415362596512, + 0.005211977753788233, + 0.0036716200411319733, + 0.007207736372947693, + 0.005731943063437939, + 0.003161005675792694, + 0.004671165719628334, + 0.0072707245126366615, + 0.005216867662966251, + 0.004950277507305145, + 0.002888947259634733, + 0.005772527307271957, + 0.0045716650784015656, + 0.0030588917434215546, + 0.004315326455980539, + 0.006188590079545975, + 0.0035297933500260115, + 0.006968510337173939, + 0.004496898967772722, + 0.007785722613334656, + 0.005675605498254299, + 0.004421481397002935, + 0.004385093227028847, + 0.016194360330700874, + 0.007973413914442062, + 0.005252768751233816, + 0.003395593725144863, + 0.006758073810487986, + 0.009879814460873604, + 0.004669609479606152, + 0.007789954077452421, + 0.005435258150100708, + 0.007720248308032751, + 0.006050736643373966, + 0.003575572045519948, + 0.0054659913294017315, + 0.004607527982443571, + 0.004950330127030611, + 0.005746778100728989, + 0.006625625304877758, + 0.007431469392031431, + 0.007689400110393763, + 0.006517701782286167, + 0.00833907164633274, + 0.006596364546567202, + 0.020737120881676674, + 0.017141222953796387, + 0.03732434660196304, + 0.0466792918741703, + 0.009681936353445053, + 0.009844524785876274, + 0.018835516646504402, + 0.011047625914216042, + 0.005388334859162569, + 0.00947044137865305, + 0.006075989920645952, + 0.009285244159400463, + 0.012177892960608006, + 0.0062148612923920155, + 0.008591191843152046, + 0.005628649611026049, + 0.02272086776793003, + 0.01107838936150074, + 0.014393914490938187, + 0.0163545124232769, + 0.010598625987768173, + 0.006969851907342672, + 0.018963217735290527, + 0.01285856869071722, + 0.030749687924981117, + 0.017535246908664703, + 0.021559325978159904, + 0.03525526449084282, + 0.016955725848674774, + 0.017834918573498726, + 0.14345820248126984, + 0.08149278908967972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.04867067560553551 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002199375070631504 + ], + "opposite_logprob_tokens": [ + -6.127199172973633 + ], + "logprob_reward": -0.002199375070631504, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002199375070631504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03512473776936531, + 0.028116708621382713, + 0.015430117957293987, + 0.0528181716799736, + 0.017484061419963837, + 0.02881615422666073, + 0.4735182225704193, + 0.10531622171401978, + 0.057187192142009735, + 0.18618841469287872 + ], + "credit_map": [ + 0.00406715227290988, + 0.003850375534966588, + 0.00602746894583106, + 0.004040861502289772, + 0.0048058549873530865, + 0.0036967613268643618, + 0.005053814500570297, + 0.0035824517253786325, + 0.002251288155093789, + 0.003501483239233494, + 0.006948127411305904, + 0.0037769183982163668, + 0.002379530342295766, + 0.003039055038243532, + 0.002940606325864792, + 0.003279702505096793, + 0.0015559825114905834, + 0.0014916048385202885, + 0.0030070908833295107, + 0.0012106562498956919, + 0.0014365683309733868, + 0.0018216698663309216, + 0.002031168667599559, + 0.002875378355383873, + 0.0038335113786160946, + 0.002966345055028796, + 0.01753370091319084, + 0.006598769221454859, + 0.0026359164621680975, + 0.002370689297094941, + 0.004568265751004219, + 0.012310977093875408, + 0.0015048790955916047, + 0.0018457247642800212, + 0.0023424101527780294, + 0.002889467403292656, + 0.0022015769500285387, + 0.0017080936813727021, + 0.003055631648749113, + 0.0019362788880243897, + 0.002677241340279579, + 0.0030167961958795786, + 0.003182226326316595, + 0.003247506683692336, + 0.003629646496847272, + 0.0032991517800837755, + 0.004679960198700428, + 0.005083624739199877, + 0.05581054463982582, + 0.04393228888511658, + 0.10120292752981186, + 0.12509743869304657, + 0.02982218563556671, + 0.023228876292705536, + 0.05654351785778999, + 0.037880461663007736, + 0.010536099784076214, + 0.011842361651360989, + 0.010119341313838959, + 0.01581520028412342, + 0.024147065356373787, + 0.009758666157722473, + 0.013602828606963158, + 0.009494679048657417, + 0.009544571861624718, + 0.004831964150071144, + 0.006709551904350519, + 0.008243605494499207, + 0.007645872887223959, + 0.0036899857223033905, + 0.009517325088381767, + 0.007004321552813053, + 0.016856949776411057, + 0.009467343799769878, + 0.01106734573841095, + 0.02076285146176815, + 0.014197181910276413, + 0.008694253861904144, + 0.06811816990375519, + 0.03702433779835701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.002199375070631504 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static, high-angle shot captures a stack of ancient official documents resting on a dark, polished wood surface, the entire scene rendered in ultra-high definition. Thick, cream-colored parchment is covered in dense black calligraphy, and several documents are stamped with fragile crimson wax seals with ornate crests. A single dramatic light source from the side casts deep shadows, accentuating the crisp texture of the paper and the raised relief of the seals, creating a tense and mysterious atmosphere. The transformation begins subtly as the paper fibers visibly darken and soften, the intricate text begins to bleed and flow, and the documents inexorably begin to liquefy. The solid wax seals melt and lose their shape, their crimson swirls reflecting the single light source on their smooth surface with pristine clarity as the entire wax seal gradually collapses into a viscous, ink-stained black puddle, slowly swirling with black eddies.", + "seed": 5046, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_10/index_6864_seed_5046_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03942266176454723, + "vlm_logprob": -0.03942266176454723, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.03942266176454723, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011492596939206123 + ], + "opposite_logprob_tokens": [ + -4.761492729187012 + ], + "logprob_reward": -0.011492596939206123, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011492596939206123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10684613883495331, + 0.06932993233203888, + 0.11386366188526154, + 0.0690704807639122, + 0.054290927946567535, + 0.3018547296524048, + 0.12507286667823792, + 0.052423544228076935, + 0.07478056848049164, + 0.032467152923345566 + ], + "credit_map": [ + 0.012763160280883312, + 0.013122114352881908, + 0.016912171617150307, + 0.008471902459859848, + 0.013305885717272758, + 0.018744347617030144, + 0.008949263021349907, + 0.014577287249267101, + 0.006997574586421251, + 0.010263197124004364, + 0.005892311688512564, + 0.005953642539680004, + 0.010514971800148487, + 0.014258158393204212, + 0.008752919733524323, + 0.006697154138237238, + 0.012570085003972054, + 0.017565928399562836, + 0.010482451878488064, + 0.01059349812567234, + 0.017000889405608177, + 0.021717224270105362, + 0.013813752681016922, + 0.010119831189513206, + 0.008581741712987423, + 0.011274359188973904, + 0.005841162521392107, + 0.005839671473950148, + 0.010832755826413631, + 0.012483720667660236, + 0.008201679214835167, + 0.006015397142618895, + 0.005952675826847553, + 0.008407684974372387, + 0.006069499999284744, + 0.003758193226531148, + 0.007240092847496271, + 0.013940314762294292, + 0.005251290276646614, + 0.003671176265925169, + 0.08060817420482635, + 0.0729108527302742, + 0.023258697241544724, + 0.0186252873390913, + 0.023581862449645996, + 0.03731689974665642, + 0.019275616854429245, + 0.0262773260474205, + 0.02118278667330742, + 0.032821446657180786, + 0.013300632126629353, + 0.006751145236194134, + 0.01273173838853836, + 0.021419189870357513, + 0.008903663605451584, + 0.007962259463965893, + 0.00975407287478447, + 0.01074556540697813, + 0.005557287018746138, + 0.005885890685021877, + 0.004939600359648466, + 0.00686244759708643, + 0.004688900429755449, + 0.003989777062088251, + 0.012719978578388691, + 0.015126165933907032, + 0.010442844592034817, + 0.005291156005114317, + 0.008443964645266533, + 0.008877333253622055, + 0.0073185330256819725, + 0.006560586392879486, + 0.005143520422279835, + 0.004060899838805199, + 0.0034647141583263874, + 0.0038788120727986097, + 0.003632504027336836, + 0.006297538056969643, + 0.0032912129536271095, + 0.0026979525573551655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.011492596939206123 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032204927410930395 + ], + "opposite_logprob_tokens": [ + -5.753220558166504 + ], + "logprob_reward": -0.0032204927410930395, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032204927410930395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2699456512928009, + 0.13019762933254242, + 0.11127923429012299, + 0.050547804683446884, + 0.04707210138440132, + 0.11030900478363037, + 0.058182474225759506, + 0.10300647467374802, + 0.05108492076396942, + 0.06837475299835205 + ], + "credit_map": [ + 0.04889330267906189, + 0.05014511197805405, + 0.02846520021557808, + 0.02104790508747101, + 0.0320618599653244, + 0.03463837876915932, + 0.021977819502353668, + 0.032716054469347, + 0.019000543281435966, + 0.014254511334002018, + 0.010862207040190697, + 0.01591895893216133, + 0.023436646908521652, + 0.013964761979877949, + 0.017286267131567, + 0.015473715960979462, + 0.010425982996821404, + 0.016706274822354317, + 0.0106973797082901, + 0.010436387732625008, + 0.014861037023365498, + 0.022405754774808884, + 0.015697943046689034, + 0.010048472322523594, + 0.007195705082267523, + 0.007698973175138235, + 0.004572879057377577, + 0.004520096816122532, + 0.008310492150485516, + 0.009148290380835533, + 0.0042997095733881, + 0.004801654256880283, + 0.004349530208855867, + 0.007925326935946941, + 0.004143909085541964, + 0.0034712746273726225, + 0.005954388994723558, + 0.01273287646472454, + 0.004585993941873312, + 0.00390879949554801, + 0.03940087929368019, + 0.03619307652115822, + 0.005917095113545656, + 0.004437768366187811, + 0.0062638139352202415, + 0.007704425603151321, + 0.0053159925155341625, + 0.005075947847217321, + 0.009449745528399944, + 0.014544235542416573, + 0.005511037539690733, + 0.003546401858329773, + 0.006045809015631676, + 0.011154635809361935, + 0.0042730895802378654, + 0.003657514462247491, + 0.01728496514260769, + 0.028242947533726692, + 0.010936337523162365, + 0.0072047095745801926, + 0.009787357412278652, + 0.012629499658942223, + 0.007420070469379425, + 0.009500580839812756, + 0.008194426074624062, + 0.005678162444382906, + 0.004277725704014301, + 0.006679479964077473, + 0.006252576597034931, + 0.006909746676683426, + 0.008392374031245708, + 0.004700427874922752, + 0.011789378710091114, + 0.009769116528332233, + 0.006742363795638084, + 0.008805188350379467, + 0.00794997252523899, + 0.010369107127189636, + 0.006783205550163984, + 0.006166419014334679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0032204927410930395 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905887603759766 + ], + "opposite_logprob_tokens": [ + -2.4640588760375977 + ], + "logprob_reward": -0.08905887603759766, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905887603759766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15434284508228302, + 0.06604312360286713, + 0.06019452214241028, + 0.044725414365530014, + 0.031316883862018585, + 0.28975969552993774, + 0.06653420627117157, + 0.03275256231427193, + 0.09509485214948654, + 0.15923583507537842 + ], + "credit_map": [ + 0.021480154246091843, + 0.020169908180832863, + 0.01543435174971819, + 0.012338892556726933, + 0.023186076432466507, + 0.028433561325073242, + 0.013775436207652092, + 0.01952447183430195, + 0.006541547831147909, + 0.010810463689267635, + 0.006364154629409313, + 0.0049689463339746, + 0.009418143890798092, + 0.014378986321389675, + 0.00673138489946723, + 0.006829503457993269, + 0.006101260427385569, + 0.009690573439002037, + 0.005756570026278496, + 0.005324808415025473, + 0.008948576636612415, + 0.012312037870287895, + 0.00669280206784606, + 0.005367901176214218, + 0.00607414823025465, + 0.007393833715468645, + 0.005213632248342037, + 0.003613897133618593, + 0.0056607043370604515, + 0.008120420388877392, + 0.004079002887010574, + 0.0045697810128331184, + 0.0034338573459535837, + 0.004704163409769535, + 0.0035877502523362637, + 0.002314562676474452, + 0.003979224245995283, + 0.007481549400836229, + 0.0033752298913896084, + 0.002440548734739423, + 0.07946161180734634, + 0.06994590908288956, + 0.021518554538488388, + 0.017576873302459717, + 0.023276004940271378, + 0.0354427732527256, + 0.017522631213068962, + 0.02501535229384899, + 0.011796985752880573, + 0.01566203311085701, + 0.006818551104515791, + 0.003881856333464384, + 0.006903121713548899, + 0.01243443600833416, + 0.004549434408545494, + 0.004487795289605856, + 0.005570255219936371, + 0.006279746536165476, + 0.0032032399903982878, + 0.0031094623263925314, + 0.003172346856445074, + 0.0033800348173826933, + 0.004700393415987492, + 0.0033370857127010822, + 0.01545771211385727, + 0.0221832487732172, + 0.012777465395629406, + 0.0082520991563797, + 0.009590830653905869, + 0.011918582953512669, + 0.007389664184302092, + 0.007525255903601646, + 0.02269027940928936, + 0.023024490103125572, + 0.02729061245918274, + 0.014738311059772968, + 0.014122998341917992, + 0.028306011110544205, + 0.011776427738368511, + 0.017286717891693115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.08905887603759766 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004115730524063 + ], + "opposite_logprob_tokens": [ + -2.6950411796569824 + ], + "logprob_reward": -0.07004115730524063, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004115730524063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21436707675457, + 0.09387224167585373, + 0.06419187039136887, + 0.04073130339384079, + 0.032448433339595795, + 0.1353527456521988, + 0.10312765836715698, + 0.07010187208652496, + 0.0990237221121788, + 0.14678309857845306 + ], + "credit_map": [ + 0.042950354516506195, + 0.04986833035945892, + 0.021761294454336166, + 0.014546794816851616, + 0.021358679980039597, + 0.022463208064436913, + 0.015634650364518166, + 0.025783751159906387, + 0.012188189662992954, + 0.0124546242877841, + 0.008124890737235546, + 0.009448991157114506, + 0.014675160869956017, + 0.012696761637926102, + 0.011233313009142876, + 0.013050306588411331, + 0.006323331501334906, + 0.008932962082326412, + 0.004959865938872099, + 0.006640732754021883, + 0.010518756695091724, + 0.00997216533869505, + 0.011439369060099125, + 0.005404681898653507, + 0.006198877934366465, + 0.004931286908686161, + 0.0034861143212765455, + 0.004485578741878271, + 0.007154152728617191, + 0.005149576347321272, + 0.004597936756908894, + 0.004727782681584358, + 0.003183523193001747, + 0.004971652757376432, + 0.004758110735565424, + 0.002840447472408414, + 0.0045496514067053795, + 0.005554070230573416, + 0.004106918815523386, + 0.0024840596597641706, + 0.045572634786367416, + 0.04130103439092636, + 0.007958974689245224, + 0.006971961818635464, + 0.009391414932906628, + 0.008727671578526497, + 0.008468744345009327, + 0.006960309110581875, + 0.02546047233045101, + 0.023074369877576828, + 0.007859406061470509, + 0.006324816029518843, + 0.008951300755143166, + 0.009615834802389145, + 0.01403491385281086, + 0.007806537207216024, + 0.012756546027958393, + 0.010256433859467506, + 0.004417459946125746, + 0.010169473476707935, + 0.007177697028964758, + 0.007971840910613537, + 0.010615689679980278, + 0.00673672603443265, + 0.017437225207686424, + 0.0102995615452528, + 0.006848384626209736, + 0.01427067443728447, + 0.011355542577803135, + 0.013418355025351048, + 0.016059480607509613, + 0.009334493428468704, + 0.023181850090622902, + 0.020924046635627747, + 0.020794032141566277, + 0.02210523933172226, + 0.013442909345030785, + 0.013519383035600185, + 0.018037719652056694, + 0.014777912758290768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.07004115730524063 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a stack of ancient official documents on a dark polished wood surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the documents made of thick cream-colored parchment with dense black calligraphy and crimson wax seals featuring ornate crests?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the lighting provided by a single dramatic source from the side that casts deep shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the paper fibers visibly darken and soften while the text bleeds and flows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the wax seals melt and collapse into a viscous ink-stained black puddle with swirling eddies?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023300185799598694 + ], + "opposite_logprob_tokens": [ + -3.7733001708984375 + ], + "logprob_reward": -0.023300185799598694, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023300185799598694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17848855257034302, + 0.08979715406894684, + 0.04140009358525276, + 0.0385405533015728, + 0.032458674162626266, + 0.09117526561021805, + 0.07228494435548782, + 0.0821789875626564, + 0.1558099240064621, + 0.21786583960056305 + ], + "credit_map": [ + 0.02218415029346943, + 0.02240469492971897, + 0.022622182965278625, + 0.016506865620613098, + 0.024830374866724014, + 0.020900443196296692, + 0.02315315417945385, + 0.025886710733175278, + 0.008463531732559204, + 0.012372232042253017, + 0.006785320118069649, + 0.008937353268265724, + 0.014148939400911331, + 0.014351172372698784, + 0.01611495390534401, + 0.008623656816780567, + 0.005569752771407366, + 0.006084640510380268, + 0.004231737926602364, + 0.004500229377299547, + 0.007532165851444006, + 0.00447181286290288, + 0.004786277189850807, + 0.004223475232720375, + 0.004180080257356167, + 0.005555570125579834, + 0.0026792995631694794, + 0.004201341420412064, + 0.006256020627915859, + 0.005138739477843046, + 0.006583308335393667, + 0.003946198616176844, + 0.004261645954102278, + 0.004708882886916399, + 0.0037539005279541016, + 0.002718632575124502, + 0.005051944404840469, + 0.005210778210312128, + 0.0036699874326586723, + 0.0030829040333628654, + 0.02433430776000023, + 0.021099010482430458, + 0.006019494496285915, + 0.005273821298032999, + 0.01042909361422062, + 0.009452112019062042, + 0.008399032056331635, + 0.006168402265757322, + 0.015593698248267174, + 0.01531832292675972, + 0.006525373086333275, + 0.004715804476290941, + 0.00625991728156805, + 0.009306871332228184, + 0.00962915364652872, + 0.004935805220156908, + 0.014801164157688618, + 0.011546527035534382, + 0.0055212643928825855, + 0.011330129578709602, + 0.010362522676587105, + 0.009395330213010311, + 0.012927026487886906, + 0.006295025814324617, + 0.026579998433589935, + 0.01532384566962719, + 0.011999816633760929, + 0.021952275186777115, + 0.01917293854057789, + 0.017441049218177795, + 0.028064189478754997, + 0.015275809913873672, + 0.03833915293216705, + 0.02654406987130642, + 0.022777246311306953, + 0.036184560507535934, + 0.023074673488736153, + 0.019574787467718124, + 0.029317574575543404, + 0.022053789347410202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.023300185799598694 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_11/index_tensor([6414, 6414])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_11/index_tensor([6414, 6414])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..44ecab535e7d916aa254c51ac6932aabb4f70630 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_11/index_tensor([6414, 6414])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, dynamic tracking shot follows a black Maybach as it smoothly and powerfully navigates the rainy streets of a bustling city. The luxury sedan's mirror-smooth obsidian paint is dampened by the rain, yet its iconic chrome grille and the prominent Double M logo on the hood still gleam with a cool, stern sheen in the dim light. Like a flowing black mirror, the car's body perfectly captures and reflects the vibrant red, blue, purple, and green neon signs blazing from the towering buildings on either side. These vibrant shadows stretch, twist, and intertwine across the wet surface, creating abstract, psychedelic, and dynamic light paintings that shift constantly as the vehicle moves forward. Dense raindrops are clearly visible in the bright beams of the car's LED headlights, while the wheels' tread on the slippery asphalt creates delicate splashes that blend seamlessly with the neon reflections on the ground. The entire scene is full of luxury, mystery and a cyberpunk sense of the future, creating a strong atmosphere of walking alone and elegantly in a cold city.", + "seed": 17500, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_11/index_6414_seed_17500_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021200180728919805, + "vlm_logprob": -0.021200180728919805, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.021200180728919805, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05383073166012764 + ], + "opposite_logprob_tokens": [ + -3.053830623626709 + ], + "logprob_reward": -0.05383073166012764, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05383073166012764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1959623098373413, + 0.045446086674928665, + 0.08623576909303665, + 0.058273568749427795, + 0.07125180959701538, + 0.07292333245277405, + 0.14074359834194183, + 0.04572073742747307, + 0.17655597627162933, + 0.10688682645559311 + ], + "credit_map": [ + 0.01818336546421051, + 0.026792606338858604, + 0.020907044410705566, + 0.05745694413781166, + 0.017793109640479088, + 0.02302362397313118, + 0.01524949911981821, + 0.01655612140893936, + 0.004921566229313612, + 0.006410506553947926, + 0.004580384120345116, + 0.009026397950947285, + 0.004206161946058273, + 0.006890538148581982, + 0.004949007648974657, + 0.004461522214114666, + 0.007700844202190638, + 0.007848633453249931, + 0.012750205583870411, + 0.012955774553120136, + 0.0063165053725242615, + 0.01639322191476822, + 0.013029328547418118, + 0.009241254068911076, + 0.004386864602565765, + 0.0059140706434845924, + 0.006924052257090807, + 0.011088380590081215, + 0.005093712825328112, + 0.006972149945795536, + 0.005884340964257717, + 0.012010002508759499, + 0.010636674240231514, + 0.005989732686430216, + 0.007962954230606556, + 0.010621513240039349, + 0.0070942724123597145, + 0.014046506956219673, + 0.007939746603369713, + 0.006960410624742508, + 0.0071386005729436874, + 0.007587247993797064, + 0.0100832125172019, + 0.011890879832208157, + 0.01020723581314087, + 0.006093156524002552, + 0.010236866772174835, + 0.009686139412224293, + 0.006215265952050686, + 0.010112411342561245, + 0.011423185467720032, + 0.04773489385843277, + 0.005999693647027016, + 0.008848048746585846, + 0.038937047123909, + 0.011473053134977818, + 0.004079004283994436, + 0.003334330627694726, + 0.009296711534261703, + 0.0058545744977891445, + 0.002428328385576606, + 0.002792613347992301, + 0.00950478482991457, + 0.008430391550064087, + 0.008353158831596375, + 0.010593978688120842, + 0.01538475975394249, + 0.032981567084789276, + 0.018016977235674858, + 0.013320812955498695, + 0.032217126339673996, + 0.04568759351968765, + 0.006344727240502834, + 0.006864646915346384, + 0.006067168433219194, + 0.010472379624843597, + 0.027063192799687386, + 0.01885462924838066, + 0.017334509640932083, + 0.01388557069003582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.05383073166012764 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006759160198271275 + ], + "opposite_logprob_tokens": [ + -5.006759166717529 + ], + "logprob_reward": -0.006759160198271275, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006759160198271275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1127648577094078, + 0.1139429435133934, + 0.07628423720598221, + 0.11423303186893463, + 0.05165719985961914, + 0.11080150306224823, + 0.12513495981693268, + 0.10460842400789261, + 0.09404030442237854, + 0.09653258323669434 + ], + "credit_map": [ + 0.016612274572253227, + 0.012566786259412766, + 0.017164569348096848, + 0.012572631239891052, + 0.013709892518818378, + 0.013924778439104557, + 0.01020185835659504, + 0.016012053936719894, + 0.00871907826513052, + 0.0107044642791152, + 0.0074676936492323875, + 0.03479832410812378, + 0.006298540625721216, + 0.033275388181209564, + 0.006693175528198481, + 0.0059862746857106686, + 0.010665325447916985, + 0.008702549152076244, + 0.013110739178955555, + 0.010946903377771378, + 0.0050426144152879715, + 0.007347341626882553, + 0.006937637459486723, + 0.013531128875911236, + 0.009334859438240528, + 0.0108861792832613, + 0.015541527420282364, + 0.019168280065059662, + 0.009176349267363548, + 0.010707755573093891, + 0.01359505020081997, + 0.02582302689552307, + 0.0066506993025541306, + 0.00685537327080965, + 0.005851959343999624, + 0.009163769893348217, + 0.004784618504345417, + 0.005493907257914543, + 0.007120990660041571, + 0.0057358830235898495, + 0.010405728593468666, + 0.015691185370087624, + 0.014454973861575127, + 0.017607489600777626, + 0.010978791862726212, + 0.009094351902604103, + 0.01816079206764698, + 0.014408188872039318, + 0.00842174794524908, + 0.013056308031082153, + 0.01341528631746769, + 0.03084613010287285, + 0.005930120591074228, + 0.008015243336558342, + 0.028525318950414658, + 0.01692480593919754, + 0.0085718659684062, + 0.009662995114922523, + 0.02825896255671978, + 0.014760391786694527, + 0.00605454808101058, + 0.005007662810385227, + 0.018609067425131798, + 0.013682929798960686, + 0.0069787949323654175, + 0.008870377205312252, + 0.008864641189575195, + 0.015391084365546703, + 0.009226562455296516, + 0.010607521049678326, + 0.01601075939834118, + 0.018090561032295227, + 0.007982415147125721, + 0.005432317033410072, + 0.011035575531423092, + 0.014566504396498203, + 0.008448644541203976, + 0.01615004613995552, + 0.01754329912364483, + 0.015373777598142624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.006759160198271275 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015216212486848235 + ], + "opposite_logprob_tokens": [ + -6.501521587371826 + ], + "logprob_reward": -0.0015216212486848235, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015216212486848235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16972756385803223, + 0.16098888218402863, + 0.13604947924613953, + 0.05760793015360832, + 0.03460909426212311, + 0.20900458097457886, + 0.04074385389685631, + 0.07621605694293976, + 0.04470444843173027, + 0.0703481063246727 + ], + "credit_map": [ + 0.016521699726581573, + 0.03658869117498398, + 0.013493762351572514, + 0.03317822143435478, + 0.018664587289094925, + 0.021194087341427803, + 0.012663866393268108, + 0.017422648146748543, + 0.009450793266296387, + 0.015690766274929047, + 0.009032880887389183, + 0.05246468260884285, + 0.009103176183998585, + 0.048903681337833405, + 0.008837594650685787, + 0.007505310233682394, + 0.015506110154092312, + 0.012981591746211052, + 0.034887831658124924, + 0.014551252126693726, + 0.012948909774422646, + 0.025272207334637642, + 0.008840382099151611, + 0.01106120552867651, + 0.006315902341157198, + 0.006428681313991547, + 0.008710881695151329, + 0.009481426328420639, + 0.004957548808306456, + 0.009335843846201897, + 0.005434707272797823, + 0.006942938547581434, + 0.0033158648293465376, + 0.003182264044880867, + 0.004761220421642065, + 0.005471223033964634, + 0.002772958716377616, + 0.007419387809932232, + 0.004089890047907829, + 0.003596284193918109, + 0.023822898045182228, + 0.027839280664920807, + 0.029211679473519325, + 0.03591994568705559, + 0.020928768441081047, + 0.029814770445227623, + 0.02510436438024044, + 0.016362857073545456, + 0.004305173177272081, + 0.004105682019144297, + 0.004471616353839636, + 0.00730183394625783, + 0.002977831056341529, + 0.0029742633923888206, + 0.007655052002519369, + 0.00695240031927824, + 0.006832063663750887, + 0.0061954474076628685, + 0.012465071864426136, + 0.008521795272827148, + 0.006298562046140432, + 0.006644422188401222, + 0.015897458419203758, + 0.013361236080527306, + 0.005435611587017775, + 0.0059962198138237, + 0.005148933734744787, + 0.005036333575844765, + 0.006262848153710365, + 0.0037764664739370346, + 0.006332992110401392, + 0.006715043447911739, + 0.008132388815283775, + 0.005625683348625898, + 0.00811269972473383, + 0.01015437487512827, + 0.007794203702360392, + 0.007000981830060482, + 0.015038331970572472, + 0.008489441126585007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.0015216212486848235 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307778552174568 + ], + "opposite_logprob_tokens": [ + -3.1680777072906494 + ], + "logprob_reward": -0.04307778552174568, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307778552174568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13778868317604065, + 0.07281192392110825, + 0.09949648380279541, + 0.30532488226890564, + 0.13991950452327728, + 0.05177036300301552, + 0.022023377940058708, + 0.04243749752640724, + 0.04339059814810753, + 0.08503670990467072 + ], + "credit_map": [ + 0.014126739464700222, + 0.01669790968298912, + 0.01564846746623516, + 0.03333715349435806, + 0.015764327719807625, + 0.017893586307764053, + 0.01083349622786045, + 0.013486984185874462, + 0.007028976920992136, + 0.00868435949087143, + 0.00543544627726078, + 0.017875831574201584, + 0.006399445701390505, + 0.01577862538397312, + 0.005271236877888441, + 0.006338001228868961, + 0.008707295171916485, + 0.006804863456636667, + 0.03407337889075279, + 0.012725955806672573, + 0.006467482075095177, + 0.020647335797548294, + 0.003790207440033555, + 0.006279969122260809, + 0.02843972109258175, + 0.02176983840763569, + 0.034902364015579224, + 0.05367860198020935, + 0.021661052480340004, + 0.0349966362118721, + 0.0222983006387949, + 0.08757837116718292, + 0.023134421557188034, + 0.014863173477351665, + 0.016776103526353836, + 0.02620275504887104, + 0.008910265751183033, + 0.014226309023797512, + 0.026479879394173622, + 0.009326587431132793, + 0.005212696734815836, + 0.005007418803870678, + 0.0064114793203771114, + 0.009923762641847134, + 0.004725635517388582, + 0.005660680588334799, + 0.008157297037541866, + 0.006671390030533075, + 0.0016505162930116057, + 0.002756531583145261, + 0.0027910105418413877, + 0.004173167981207371, + 0.0018596878508105874, + 0.0016600987873971462, + 0.003400994697585702, + 0.0037313709035515785, + 0.0033057956025004387, + 0.003530532820150256, + 0.008509060367941856, + 0.0046713813208043575, + 0.002430511638522148, + 0.0031525148078799248, + 0.009465429000556469, + 0.007372268941253424, + 0.00421808660030365, + 0.0048117320984601974, + 0.004315062426030636, + 0.00651276670396328, + 0.004846181254833937, + 0.0032468342687934637, + 0.006701200269162655, + 0.008738730102777481, + 0.009216184727847576, + 0.0065880087204277515, + 0.007746527902781963, + 0.01026098895817995, + 0.010651842691004276, + 0.015918618068099022, + 0.0116164181381464, + 0.013038119301199913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.04307778552174568 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008116050157696009 + ], + "opposite_logprob_tokens": [ + -7.125811576843262 + ], + "logprob_reward": -0.0008116050157696009, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008116050157696009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21323096752166748, + 0.16165786981582642, + 0.13349756598472595, + 0.08930488675832748, + 0.0802859216928482, + 0.06738269329071045, + 0.07640433311462402, + 0.05614852160215378, + 0.06083577126264572, + 0.061251457780599594 + ], + "credit_map": [ + 0.019400309771299362, + 0.04374457895755768, + 0.020306045189499855, + 0.06984615325927734, + 0.015561950393021107, + 0.013730966486036777, + 0.01563810184597969, + 0.01500286627560854, + 0.01182354986667633, + 0.013750906102359295, + 0.012866741046309471, + 0.041784897446632385, + 0.012856532819569111, + 0.04273810610175133, + 0.01076399628072977, + 0.015073152258992195, + 0.012980863451957703, + 0.014564678072929382, + 0.03548729047179222, + 0.01606474444270134, + 0.010379867628216743, + 0.02337815798819065, + 0.008669734001159668, + 0.011972233653068542, + 0.009921358898282051, + 0.00835970975458622, + 0.012723485939204693, + 0.01622500643134117, + 0.006792695727199316, + 0.014374745078384876, + 0.008043505251407623, + 0.012864388525485992, + 0.016284581273794174, + 0.007617500144988298, + 0.010208438150584698, + 0.014018530026078224, + 0.006150996312499046, + 0.008158408105373383, + 0.012203739024698734, + 0.005643740296363831, + 0.007993380539119244, + 0.005582266952842474, + 0.009503616020083427, + 0.014121703803539276, + 0.00521851284429431, + 0.011939211748540401, + 0.007890447974205017, + 0.005133553873747587, + 0.004930438939481974, + 0.006369307171553373, + 0.008404804393649101, + 0.020317865535616875, + 0.004221299663186073, + 0.00456273416057229, + 0.01756007969379425, + 0.010037804022431374, + 0.005006680730730295, + 0.004376139957457781, + 0.01425514928996563, + 0.007130054756999016, + 0.0031282745767384768, + 0.003467712551355362, + 0.010914243757724762, + 0.007870271801948547, + 0.005121677182614803, + 0.0062934928573668, + 0.005614895839244127, + 0.010392743162810802, + 0.006898862309753895, + 0.005657559726387262, + 0.010691597126424313, + 0.01016494631767273, + 0.006071660667657852, + 0.005321152973920107, + 0.00797811709344387, + 0.006888325326144695, + 0.00596759095788002, + 0.010518820956349373, + 0.01042660791426897, + 0.008079185150563717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0008116050157696009 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows a black Maybach as it smoothly and powerfully navigates the rainy streets of a bustling city. The luxury sedan's mirror-smooth obsidian paint is dampened by the rain, yet its iconic chrome grille and the prominent Double M logo on the hood still gleam with a cool, stern sheen in the dim light. Like a flowing black mirror, the car's body perfectly captures and reflects the vibrant red, blue, purple, and green neon signs blazing from the towering buildings on either side. These vibrant shadows stretch, twist, and intertwine across the wet surface, creating abstract, psychedelic, and dynamic light paintings that shift constantly as the vehicle moves forward. Dense raindrops are clearly visible in the bright beams of the car's LED headlights, while the wheels' tread on the slippery asphalt creates delicate splashes that blend seamlessly with the neon reflections on the ground. The entire scene is full of luxury, mystery and a cyberpunk sense of the future, creating a strong atmosphere of walking alone and elegantly in a cold city.", + "seed": 17501, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_11/index_6414_seed_17501_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024389683781191707, + "vlm_logprob": -0.024389683781191707, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.024389683781191707, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04900875687599182 + ], + "opposite_logprob_tokens": [ + -3.299008846282959 + ], + "logprob_reward": -0.04900875687599182, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04900875687599182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08640571683645248, + 0.6513480544090271, + 0.05101674795150757, + 0.06444843858480453, + 0.0364556685090065, + 0.014868773519992828, + 0.026880871504545212, + 0.02246774546802044, + 0.032891031354665756, + 0.013217017985880375 + ], + "credit_map": [ + 0.008668346330523491, + 0.007987621240317822, + 0.01093421969562769, + 0.011821641586720943, + 0.011477933265268803, + 0.01966750994324684, + 0.009032178670167923, + 0.0068162577226758, + 0.06253796070814133, + 0.05379807949066162, + 0.03252298757433891, + 0.28771206736564636, + 0.029939552769064903, + 0.031096074730157852, + 0.0353754460811615, + 0.1183658167719841, + 0.00595541438087821, + 0.006959476508200169, + 0.005770721938461065, + 0.007067978382110596, + 0.007579044904559851, + 0.0037337287794798613, + 0.007951881736516953, + 0.005998500157147646, + 0.005737852305173874, + 0.007366274017840624, + 0.0051162149757146835, + 0.010537425056099892, + 0.00899837352335453, + 0.013716465793550014, + 0.006301043555140495, + 0.006674784701317549, + 0.0023267848882824183, + 0.0027011113706976175, + 0.007663915399461985, + 0.005500707309693098, + 0.004739407915621996, + 0.005626414902508259, + 0.0046126884408295155, + 0.003284635953605175, + 0.0012130712857469916, + 0.0013558637583628297, + 0.002287294017150998, + 0.0020998776890337467, + 0.0017078127712011337, + 0.0033254208974540234, + 0.0014097655657678843, + 0.0014696663711220026, + 0.002294072648510337, + 0.0018992454279214144, + 0.0032664681784808636, + 0.005975570995360613, + 0.0022904989309608936, + 0.005666324868798256, + 0.003796797711402178, + 0.001691889250651002, + 0.0024466863833367825, + 0.002333730226382613, + 0.004068565554916859, + 0.003349010366946459, + 0.002167885657399893, + 0.002591522876173258, + 0.0035107885487377644, + 0.001999554457142949, + 0.003893765388056636, + 0.0024524121545255184, + 0.005492701195180416, + 0.0063612619414925575, + 0.002498846733942628, + 0.003484665649011731, + 0.004069934133440256, + 0.004637444857507944, + 0.0010451494017615914, + 0.0010538537753745914, + 0.0024770174641162157, + 0.002897790400311351, + 0.001585400546900928, + 0.0016840845346450806, + 0.001183371408842504, + 0.0012903489405289292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.04900875687599182 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018221953883767128 + ], + "opposite_logprob_tokens": [ + -4.018221855163574 + ], + "logprob_reward": -0.018221953883767128, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018221953883767128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19629289209842682, + 0.2126210778951645, + 0.06050847843289375, + 0.07445498555898666, + 0.08979318290948868, + 0.08328357338905334, + 0.05737371742725372, + 0.06512482464313507, + 0.11571886390447617, + 0.04482841119170189 + ], + "credit_map": [ + 0.01800442859530449, + 0.022808611392974854, + 0.03643225505948067, + 0.018589260056614876, + 0.028404943645000458, + 0.033688124269247055, + 0.018210185691714287, + 0.020155081525444984, + 0.034598320722579956, + 0.015279210172593594, + 0.012945864349603653, + 0.052790552377700806, + 0.008413557894527912, + 0.01334108505398035, + 0.009703933261334896, + 0.06554855406284332, + 0.009559297934174538, + 0.003851254004985094, + 0.007083413656800985, + 0.013441324234008789, + 0.003715408267453313, + 0.004916645586490631, + 0.00963356252759695, + 0.00830757338553667, + 0.007288894150406122, + 0.009460214525461197, + 0.006196225993335247, + 0.016641391441226006, + 0.0063838837668299675, + 0.011698631569743156, + 0.008809735998511314, + 0.007976002059876919, + 0.008968373760581017, + 0.009083865210413933, + 0.015651339665055275, + 0.01127469353377819, + 0.01419252809137106, + 0.00990999024361372, + 0.014236162416636944, + 0.006476226262748241, + 0.017677010968327522, + 0.008593058213591576, + 0.01059118565171957, + 0.010538938455283642, + 0.007592684589326382, + 0.014378737658262253, + 0.007102695293724537, + 0.006809265352785587, + 0.005836132448166609, + 0.004220019094645977, + 0.007460320834070444, + 0.010536237619817257, + 0.005251659546047449, + 0.012266517616808414, + 0.008411770686507225, + 0.0033910595811903477, + 0.005921870470046997, + 0.006971574388444424, + 0.01206365879625082, + 0.009449639357626438, + 0.00579220661893487, + 0.013520697131752968, + 0.006802643183618784, + 0.004602532368153334, + 0.012603354640305042, + 0.005825050640851259, + 0.01898781582713127, + 0.012341871857643127, + 0.005241442937403917, + 0.008671825751662254, + 0.018999440595507622, + 0.03304806724190712, + 0.004601103253662586, + 0.0037631848827004433, + 0.008785486221313477, + 0.010110486298799515, + 0.005059743765741587, + 0.004598652012646198, + 0.003439987078309059, + 0.004469767212867737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.018221953883767128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004105354659259319 + ], + "opposite_logprob_tokens": [ + -5.504105567932129 + ], + "logprob_reward": -0.004105354659259319, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004105354659259319, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07626145333051682, + 0.06765756756067276, + 0.04581253230571747, + 0.07125061005353928, + 0.5470898747444153, + 0.035391952842473984, + 0.04710995405912399, + 0.029903259128332138, + 0.042336802929639816, + 0.03718595951795578 + ], + "credit_map": [ + 0.007327912375330925, + 0.008843915536999702, + 0.014028039760887623, + 0.007033723406493664, + 0.012123276479542255, + 0.013933333568274975, + 0.006339132785797119, + 0.006632122676819563, + 0.011220748536288738, + 0.008048505522310734, + 0.005649501457810402, + 0.007102541625499725, + 0.006456038914620876, + 0.009233029559254646, + 0.006471442524343729, + 0.013475759886205196, + 0.0067710671573877335, + 0.0030029406771063805, + 0.0052855731919407845, + 0.008050238713622093, + 0.0032549817115068436, + 0.006700732279568911, + 0.005406183190643787, + 0.0073408158496022224, + 0.006770665757358074, + 0.00852867029607296, + 0.0072085061110556126, + 0.01155839953571558, + 0.010069573298096657, + 0.012835459783673286, + 0.007190486416220665, + 0.007088843733072281, + 0.055904705077409744, + 0.05403205007314682, + 0.0955938845872879, + 0.05353199690580368, + 0.06891737878322601, + 0.08193228393793106, + 0.08517233282327652, + 0.052005279809236526, + 0.00404029805213213, + 0.002818615874275565, + 0.004033383913338184, + 0.006251515354961157, + 0.0057846237905323505, + 0.006834153551608324, + 0.0030734108295291662, + 0.002555948682129383, + 0.004079823382198811, + 0.003887060796841979, + 0.007476880680769682, + 0.008845247328281403, + 0.006012658588588238, + 0.007855093106627464, + 0.006457698531448841, + 0.0024954876862466335, + 0.003289335174486041, + 0.002674642950296402, + 0.005600349046289921, + 0.0045802961103618145, + 0.00401003984734416, + 0.004777225665748119, + 0.0025907158851623535, + 0.002380654215812683, + 0.004751671105623245, + 0.002509997459128499, + 0.006390477996319532, + 0.004053641110658646, + 0.003536375006660819, + 0.007112227380275726, + 0.005835849791765213, + 0.008146560750901699, + 0.003144490532577038, + 0.00264907069504261, + 0.007168410811573267, + 0.00788677018135786, + 0.0065883733332157135, + 0.004103620536625385, + 0.0029241014271974564, + 0.002721122233197093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.004105354659259319 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867362603545189 + ], + "opposite_logprob_tokens": [ + -3.048673629760742 + ], + "logprob_reward": -0.04867362603545189, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867362603545189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21783259510993958, + 0.4156268239021301, + 0.03632023185491562, + 0.07634653896093369, + 0.07302802801132202, + 0.050960395485162735, + 0.054769597947597504, + 0.04355337843298912, + 0.020817125216126442, + 0.01074532326310873 + ], + "credit_map": [ + 0.01835320144891739, + 0.019699515774846077, + 0.045428480952978134, + 0.019687006250023842, + 0.02848994918167591, + 0.04965723305940628, + 0.01671055145561695, + 0.019806668162345886, + 0.05792629346251488, + 0.03520172834396362, + 0.025154944509267807, + 0.17630721628665924, + 0.02217230759561062, + 0.024214908480644226, + 0.025535695254802704, + 0.049113739281892776, + 0.003251831280067563, + 0.004525492433458567, + 0.003635240253061056, + 0.00561253959313035, + 0.004884972237050533, + 0.004327514674514532, + 0.005157564301043749, + 0.004925076384097338, + 0.007832174189388752, + 0.008497114293277264, + 0.00647676270455122, + 0.013204598799347878, + 0.010295207612216473, + 0.01678788848221302, + 0.006379313766956329, + 0.006873477715998888, + 0.009231300093233585, + 0.007243408355861902, + 0.01141006126999855, + 0.007937345653772354, + 0.010102725587785244, + 0.007524579297751188, + 0.014052611775696278, + 0.0055259959772229195, + 0.005388098768889904, + 0.005106788128614426, + 0.006420946680009365, + 0.006517533212900162, + 0.006610484328120947, + 0.012119661085307598, + 0.004384325351566076, + 0.004412556998431683, + 0.0038256736006587744, + 0.0032973706256598234, + 0.013724224641919136, + 0.016275152564048767, + 0.005567499902099371, + 0.003779549617320299, + 0.006240911316126585, + 0.0020592135842889547, + 0.0038537303917109966, + 0.004254830535501242, + 0.007814744487404823, + 0.0076892320066690445, + 0.003860881319269538, + 0.008101146668195724, + 0.004434901289641857, + 0.003543913597241044, + 0.0022952777799218893, + 0.0012117375154048204, + 0.003528149565681815, + 0.003921560477465391, + 0.001628266996704042, + 0.0034866556525230408, + 0.002142357174307108, + 0.002603120869025588, + 0.001254315604455769, + 0.0006932643009349704, + 0.0017342526698485017, + 0.0015447684563696384, + 0.0018683316884562373, + 0.0016368229407817125, + 0.0009124359930865467, + 0.0011011315509676933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.04867362603545189 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019387274514883757 + ], + "opposite_logprob_tokens": [ + -6.251938819885254 + ], + "logprob_reward": -0.0019387274514883757, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019387274514883757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18852701783180237, + 0.32537007331848145, + 0.04855496063828468, + 0.10942436009645462, + 0.05696037784218788, + 0.08325477689504623, + 0.059274569153785706, + 0.04520094022154808, + 0.04055894538760185, + 0.042873963713645935 + ], + "credit_map": [ + 0.015316356904804707, + 0.02057805098593235, + 0.03721465915441513, + 0.017310917377471924, + 0.028179578483104706, + 0.0330086275935173, + 0.01929035782814026, + 0.017628485336899757, + 0.02671823278069496, + 0.025245260447263718, + 0.016751255840063095, + 0.13514971733093262, + 0.017828166484832764, + 0.022150995209813118, + 0.01849442347884178, + 0.06303206086158752, + 0.007051281630992889, + 0.004866247996687889, + 0.004556193016469479, + 0.007258569356054068, + 0.004604610148817301, + 0.008962617255747318, + 0.0061017172411084175, + 0.005153728183358908, + 0.010782666504383087, + 0.012743045575916767, + 0.01024119183421135, + 0.020289462059736252, + 0.014723013155162334, + 0.019552581012248993, + 0.010171890258789062, + 0.010920518077909946, + 0.0076115187257528305, + 0.004861941561102867, + 0.012620383873581886, + 0.009448167867958546, + 0.010132357478141785, + 0.0039025729056447744, + 0.005328802391886711, + 0.0030546411871910095, + 0.007172673474997282, + 0.006926884409040213, + 0.00975732784718275, + 0.014955178834497929, + 0.014843539334833622, + 0.01597752794623375, + 0.007161314599215984, + 0.0064603416249156, + 0.00656634196639061, + 0.004871570039540529, + 0.011188223026692867, + 0.012273513711988926, + 0.005829882808029652, + 0.008469479158520699, + 0.007078573107719421, + 0.0029969916213303804, + 0.003812979906797409, + 0.004834229592233896, + 0.0077918400056660175, + 0.009023887105286121, + 0.004106373991817236, + 0.0070265959948301315, + 0.005210559815168381, + 0.003394481260329485, + 0.004409787245094776, + 0.002717222087085247, + 0.006322970613837242, + 0.006090588867664337, + 0.0027256286703050137, + 0.006155852228403091, + 0.004858728032559156, + 0.007278170902282, + 0.0037175193428993225, + 0.0038016997277736664, + 0.007452312391251326, + 0.006492475513368845, + 0.007822045125067234, + 0.006515973247587681, + 0.003409485099837184, + 0.0036624588537961245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0019387274514883757 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows a black Maybach as it smoothly and powerfully navigates the rainy streets of a bustling city. The luxury sedan's mirror-smooth obsidian paint is dampened by the rain, yet its iconic chrome grille and the prominent Double M logo on the hood still gleam with a cool, stern sheen in the dim light. Like a flowing black mirror, the car's body perfectly captures and reflects the vibrant red, blue, purple, and green neon signs blazing from the towering buildings on either side. These vibrant shadows stretch, twist, and intertwine across the wet surface, creating abstract, psychedelic, and dynamic light paintings that shift constantly as the vehicle moves forward. Dense raindrops are clearly visible in the bright beams of the car's LED headlights, while the wheels' tread on the slippery asphalt creates delicate splashes that blend seamlessly with the neon reflections on the ground. The entire scene is full of luxury, mystery and a cyberpunk sense of the future, creating a strong atmosphere of walking alone and elegantly in a cold city.", + "seed": 17502, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_11/index_6414_seed_17502_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02103442270308733, + "vlm_logprob": -0.02103442270308733, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.02103442270308733, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04569554328918457 + ], + "opposite_logprob_tokens": [ + -3.2956955432891846 + ], + "logprob_reward": -0.04569554328918457, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04569554328918457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03905188664793968, + 0.02278795652091503, + 0.566694438457489, + 0.021438928321003914, + 0.03645208850502968, + 0.07403679192066193, + 0.0714879110455513, + 0.07801319658756256, + 0.031022341921925545, + 0.05901450291275978 + ], + "credit_map": [ + 0.004036351572722197, + 0.00889624934643507, + 0.003495589131489396, + 0.006701063364744186, + 0.005644125398248434, + 0.0024869574699550867, + 0.004743789788335562, + 0.0030477612745016813, + 0.0017489073798060417, + 0.003514095675200224, + 0.0033618470188230276, + 0.0030601657927036285, + 0.0019481146009638906, + 0.0025302665308117867, + 0.0048366389237344265, + 0.0017879202496260405, + 0.020151562988758087, + 0.0116964653134346, + 0.01438212115317583, + 0.02983286790549755, + 0.03170791268348694, + 0.23661011457443237, + 0.18699057400226593, + 0.03532281890511513, + 0.0018989737145602703, + 0.002100015291944146, + 0.0021593731362372637, + 0.00362145877443254, + 0.0026623101439327, + 0.0016557829221710563, + 0.0038882701192051172, + 0.0034527459647506475, + 0.0025899000465869904, + 0.0028218000661581755, + 0.0034226616844534874, + 0.005134036298841238, + 0.0025016034487634897, + 0.0027322082314640284, + 0.00889661069959402, + 0.00835326872766018, + 0.005030143074691296, + 0.009295150637626648, + 0.010320373810827732, + 0.012875053100287914, + 0.009563345462083817, + 0.003520174650475383, + 0.012273775413632393, + 0.0111587755382061, + 0.002802527043968439, + 0.0033192704431712627, + 0.006881535518914461, + 0.011766286566853523, + 0.0033651920966804028, + 0.0022393562830984592, + 0.023959344252943993, + 0.017154403030872345, + 0.007268871646374464, + 0.006265267264097929, + 0.017496567219495773, + 0.013882034458220005, + 0.006291519850492477, + 0.006888280156999826, + 0.011791463941335678, + 0.008129196241497993, + 0.0032081466633826494, + 0.00328906811773777, + 0.007355559151619673, + 0.0032168403267860413, + 0.0017880165250971913, + 0.002826193580403924, + 0.0036672763526439667, + 0.005671242717653513, + 0.007192215416580439, + 0.0036602080799639225, + 0.011941391043365002, + 0.00835712905973196, + 0.004301613196730614, + 0.005603860132396221, + 0.009034201502799988, + 0.008923881687223911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 5 + ], + "selected_reward": -0.04569554328918457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012566785328090191 + ], + "opposite_logprob_tokens": [ + -4.387566566467285 + ], + "logprob_reward": -0.012566785328090191, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012566785328090191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.064371757209301, + 0.046910859644412994, + 0.34044939279556274, + 0.049902353435754776, + 0.11114770174026489, + 0.08136279135942459, + 0.07061109691858292, + 0.16257603466510773, + 0.024944132193922997, + 0.04772388935089111 + ], + "credit_map": [ + 0.0042737070471048355, + 0.006404622923582792, + 0.0032551195472478867, + 0.01251437421888113, + 0.004768273327499628, + 0.008709581568837166, + 0.01780172623693943, + 0.006644352804869413, + 0.004489080049097538, + 0.0060369595885276794, + 0.006406498607248068, + 0.008209563791751862, + 0.00462726503610611, + 0.005674164742231369, + 0.007966091856360435, + 0.0035012350417673588, + 0.01252853125333786, + 0.007130681537091732, + 0.008625376969575882, + 0.017669880762696266, + 0.019299669191241264, + 0.14196595549583435, + 0.11196677386760712, + 0.02126254141330719, + 0.003910692874342203, + 0.005287016276270151, + 0.005707836244255304, + 0.007302779238671064, + 0.0065344562754035, + 0.005151411984115839, + 0.008867820724844933, + 0.007140338886529207, + 0.00442555733025074, + 0.0037096114829182625, + 0.0039909309707582, + 0.007039863616228104, + 0.004999419208616018, + 0.004359751474112272, + 0.04301748424768448, + 0.03960508108139038, + 0.007710236590355635, + 0.008197711780667305, + 0.012002580799162388, + 0.014303888194262981, + 0.007743236608803272, + 0.007357532158493996, + 0.013781675137579441, + 0.010265928693115711, + 0.003820614656433463, + 0.005539846606552601, + 0.010925219394266605, + 0.014552028849720955, + 0.004804669413715601, + 0.004094832576811314, + 0.016173895448446274, + 0.010699989274144173, + 0.013056175783276558, + 0.014270061627030373, + 0.03633979707956314, + 0.027515681460499763, + 0.012916356325149536, + 0.013315425254404545, + 0.026117801666259766, + 0.01904473640024662, + 0.0023231450468301773, + 0.0027593455743044615, + 0.0036515756510198116, + 0.003262103069573641, + 0.0020404367242008448, + 0.0032473301980644464, + 0.0033472890499979258, + 0.004312906879931688, + 0.004735494032502174, + 0.002812843769788742, + 0.010112262330949306, + 0.006006075534969568, + 0.0035419773776084185, + 0.005489991046488285, + 0.007552598137408495, + 0.007472650613635778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.012566785328090191 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024996723514050245 + ], + "opposite_logprob_tokens": [ + -6.002499580383301 + ], + "logprob_reward": -0.0024996723514050245, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024996723514050245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05444970726966858, + 0.17458504438400269, + 0.12893536686897278, + 0.04107322543859482, + 0.2742341458797455, + 0.08220858871936798, + 0.07886727154254913, + 0.06617940962314606, + 0.025696659460663795, + 0.07377054542303085 + ], + "credit_map": [ + 0.003263036487624049, + 0.006133888382464647, + 0.003930108156055212, + 0.009300691075623035, + 0.006475085858255625, + 0.005830495618283749, + 0.013935813680291176, + 0.00558058638125658, + 0.014470033347606659, + 0.020660318434238434, + 0.0135442353785038, + 0.018805861473083496, + 0.02304881438612938, + 0.03391752019524574, + 0.03932269290089607, + 0.010815581306815147, + 0.004656659904867411, + 0.0030501345172524452, + 0.0029698226135224104, + 0.0069275968708097935, + 0.007098272908478975, + 0.05326061695814133, + 0.042787037789821625, + 0.008185231126844883, + 0.003175242803990841, + 0.0025135523173958063, + 0.002961411839351058, + 0.0025101604405790567, + 0.0062053389847278595, + 0.011703351512551308, + 0.006390776950865984, + 0.005613391753286123, + 0.010857687331736088, + 0.008318061009049416, + 0.008894487284123898, + 0.02394985966384411, + 0.011182165704667568, + 0.010666510090231895, + 0.10537748038768768, + 0.0949879139661789, + 0.007967629469931126, + 0.007623425219208002, + 0.013619322329759598, + 0.01766373962163925, + 0.0055337511003017426, + 0.004624516237527132, + 0.014665503986179829, + 0.010510701686143875, + 0.0037512341514229774, + 0.003162182169035077, + 0.01025024987757206, + 0.013917766511440277, + 0.0041821859776973724, + 0.004372922703623772, + 0.02376958727836609, + 0.015461143106222153, + 0.005309258121997118, + 0.005735975224524736, + 0.014725790359079838, + 0.01195543259382248, + 0.00558193726465106, + 0.005864394828677177, + 0.01023496687412262, + 0.006771658547222614, + 0.0020214379765093327, + 0.002572538796812296, + 0.0033696049358695745, + 0.002419925993308425, + 0.0023604249581694603, + 0.005398923996835947, + 0.003000727156177163, + 0.004553075879812241, + 0.005649273749440908, + 0.0049495575949549675, + 0.012489531189203262, + 0.007090872153639793, + 0.005885269958525896, + 0.008255964145064354, + 0.017846867442131042, + 0.011603211984038353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.0024996723514050245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043075501918792725 + ], + "opposite_logprob_tokens": [ + -3.1680755615234375 + ], + "logprob_reward": -0.043075501918792725, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043075501918792725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08714262396097183, + 0.07000216841697693, + 0.14679646492004395, + 0.04416653886437416, + 0.4154859781265259, + 0.028504198417067528, + 0.029397884383797646, + 0.055581703782081604, + 0.01960272155702114, + 0.10331971198320389 + ], + "credit_map": [ + 0.005895111244171858, + 0.007107991725206375, + 0.004401060286909342, + 0.00979145709425211, + 0.010707790032029152, + 0.02182416431605816, + 0.018811363726854324, + 0.00860367901623249, + 0.005834873765707016, + 0.00858574453741312, + 0.007336675189435482, + 0.010680456645786762, + 0.007177476305514574, + 0.011546672321856022, + 0.013279524631798267, + 0.005560746882110834, + 0.005779405124485493, + 0.0035521266981959343, + 0.004032924771308899, + 0.007913379929959774, + 0.008603270165622234, + 0.059814777225255966, + 0.04764885455369949, + 0.009451723657548428, + 0.005780285689979792, + 0.003189682960510254, + 0.0038629891350865364, + 0.0039132460951805115, + 0.005592531990259886, + 0.009324274957180023, + 0.006283348426222801, + 0.006220181938260794, + 0.013953511603176594, + 0.01118510589003563, + 0.011159129440784454, + 0.031010445207357407, + 0.014143438078463078, + 0.011128098703920841, + 0.16812407970428467, + 0.15478219091892242, + 0.0036817786749452353, + 0.0021886632312089205, + 0.0032818641047924757, + 0.004966314882040024, + 0.001977855572476983, + 0.0020260282326489687, + 0.004742735996842384, + 0.00563895795494318, + 0.002069726586341858, + 0.002607300877571106, + 0.0032448528800159693, + 0.004508661106228828, + 0.003739846171811223, + 0.002527636708691716, + 0.006543076131492853, + 0.004156784154474735, + 0.005353497341275215, + 0.004759563133120537, + 0.013316464610397816, + 0.010090784169733524, + 0.005232301540672779, + 0.00311834248714149, + 0.007718835957348347, + 0.005991914309561253, + 0.0020152446813881397, + 0.001694201142527163, + 0.0022774520330131054, + 0.0017255598213523626, + 0.0018977968720719218, + 0.004241724032908678, + 0.001964620780199766, + 0.003786121029406786, + 0.007525300141423941, + 0.005606594029814005, + 0.01764627732336521, + 0.010827678255736828, + 0.006527838297188282, + 0.010093091987073421, + 0.027454467490315437, + 0.017638469114899635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.043075501918792725 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001334610627964139 + ], + "opposite_logprob_tokens": [ + -6.6263346672058105 + ], + "logprob_reward": -0.001334610627964139, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001334610627964139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08006888628005981, + 0.07033929973840714, + 0.3215033710002899, + 0.04888569191098213, + 0.10618175566196442, + 0.04902821034193039, + 0.07229644805192947, + 0.09486444294452667, + 0.07491546124219894, + 0.08191638439893723 + ], + "credit_map": [ + 0.005310262553393841, + 0.00720244599506259, + 0.004362873267382383, + 0.009026744402945042, + 0.00998664740473032, + 0.01639244332909584, + 0.020143238827586174, + 0.007644226308912039, + 0.006674054078757763, + 0.0074131423607468605, + 0.007414636202156544, + 0.010735530406236649, + 0.008266197517514229, + 0.011406890116631985, + 0.013284838758409023, + 0.005144007038325071, + 0.013212478719651699, + 0.010024930350482464, + 0.010485196486115456, + 0.01726372353732586, + 0.020062485709786415, + 0.12866432964801788, + 0.10157507658004761, + 0.020215163007378578, + 0.0036541977897286415, + 0.0041179414838552475, + 0.004304895177483559, + 0.00677046924829483, + 0.006133069284260273, + 0.008294569328427315, + 0.008690855465829372, + 0.006919694598764181, + 0.004958886653184891, + 0.004044069442898035, + 0.004287579096853733, + 0.011456525884568691, + 0.005631467793136835, + 0.004046388436108828, + 0.03805377706885338, + 0.03370305523276329, + 0.004257979802787304, + 0.004611742217093706, + 0.008689110167324543, + 0.007848373614251614, + 0.003126762807369232, + 0.003410330740734935, + 0.008019015192985535, + 0.009064893238246441, + 0.0049235825426876545, + 0.004201695788651705, + 0.006942618638277054, + 0.010926580987870693, + 0.002750545274466276, + 0.0038952042814344168, + 0.021794965490698814, + 0.01686125434935093, + 0.008167523890733719, + 0.007103035692125559, + 0.02175581082701683, + 0.017480764538049698, + 0.006703217048197985, + 0.007980089634656906, + 0.014607276767492294, + 0.011066720820963383, + 0.005871026311069727, + 0.007693292107433081, + 0.010416842065751553, + 0.007783933077007532, + 0.0075538610108196735, + 0.015878837555646896, + 0.006926856003701687, + 0.012790813110768795, + 0.007114582695066929, + 0.0063422066159546375, + 0.014782289043068886, + 0.008378260768949986, + 0.006298251450061798, + 0.009025830775499344, + 0.017049528658390045, + 0.012925435788929462 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.001334610627964139 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, dynamic tracking shot follows a black Maybach as it smoothly and powerfully navigates the rainy streets of a bustling city. The luxury sedan's mirror-smooth obsidian paint is dampened by the rain, yet its iconic chrome grille and the prominent Double M logo on the hood still gleam with a cool, stern sheen in the dim light. Like a flowing black mirror, the car's body perfectly captures and reflects the vibrant red, blue, purple, and green neon signs blazing from the towering buildings on either side. These vibrant shadows stretch, twist, and intertwine across the wet surface, creating abstract, psychedelic, and dynamic light paintings that shift constantly as the vehicle moves forward. Dense raindrops are clearly visible in the bright beams of the car's LED headlights, while the wheels' tread on the slippery asphalt creates delicate splashes that blend seamlessly with the neon reflections on the ground. The entire scene is full of luxury, mystery and a cyberpunk sense of the future, creating a strong atmosphere of walking alone and elegantly in a cold city.", + "seed": 17503, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_11/index_6414_seed_17503_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02735862717963755, + "vlm_logprob": -0.02735862717963755, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.02735862717963755, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.049044057726860046 + ], + "opposite_logprob_tokens": [ + -3.299044132232666 + ], + "logprob_reward": -0.049044057726860046, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.049044057726860046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21277329325675964, + 0.08644438534975052, + 0.10648033767938614, + 0.11929678916931152, + 0.04015897959470749, + 0.1269548386335373, + 0.1023402288556099, + 0.11614729464054108, + 0.0422603040933609, + 0.04714352265000343 + ], + "credit_map": [ + 0.013889280147850513, + 0.058742705732584, + 0.05623912066221237, + 0.015777001157402992, + 0.017686640843749046, + 0.011078004725277424, + 0.013640593737363815, + 0.02571997418999672, + 0.008650382049381733, + 0.011044146493077278, + 0.011851281858980656, + 0.008971747010946274, + 0.020803086459636688, + 0.00829519797116518, + 0.006867915391921997, + 0.009960639290511608, + 0.021397802978754044, + 0.011021628975868225, + 0.017272379249334335, + 0.009744485840201378, + 0.01081435289233923, + 0.008595281280577183, + 0.010449170134961605, + 0.01718524657189846, + 0.018711095675826073, + 0.019699811935424805, + 0.015396793372929096, + 0.012637392617762089, + 0.012595891952514648, + 0.018065186217427254, + 0.009441150352358818, + 0.012749476358294487, + 0.004435121547430754, + 0.005421530455350876, + 0.00583291333168745, + 0.004007081501185894, + 0.00834172498434782, + 0.003936811350286007, + 0.004079134669154882, + 0.004104666877537966, + 0.012979275546967983, + 0.013859454542398453, + 0.016626056283712387, + 0.0100706210359931, + 0.018597647547721863, + 0.02059563249349594, + 0.01266662497073412, + 0.021559538319706917, + 0.006437196861952543, + 0.010054936632514, + 0.012034216895699501, + 0.01354399137198925, + 0.016512669622898102, + 0.012185378931462765, + 0.015617323108017445, + 0.015954522415995598, + 0.0075086490251123905, + 0.005379802547395229, + 0.006527119316160679, + 0.045719586312770844, + 0.008580158464610577, + 0.0035158346872776747, + 0.0327334962785244, + 0.006182653363794088, + 0.005583338905125856, + 0.004676221404224634, + 0.0051712882705032825, + 0.004943517502397299, + 0.006353394128382206, + 0.005224842578172684, + 0.004954186733812094, + 0.005353515036404133, + 0.005861287470906973, + 0.006364903878420591, + 0.005449859891086817, + 0.00706811947748065, + 0.006236782297492027, + 0.008879531174898148, + 0.00379757396876812, + 0.0034854670520871878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.049044057726860046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016109542921185493 + ], + "opposite_logprob_tokens": [ + -4.141109466552734 + ], + "logprob_reward": -0.016109542921185493, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016109542921185493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2598610818386078, + 0.10926155000925064, + 0.0605694055557251, + 0.19498394429683685, + 0.06268184632062912, + 0.08635857701301575, + 0.04794861003756523, + 0.02058504708111286, + 0.05141877755522728, + 0.10633119940757751 + ], + "credit_map": [ + 0.01995132863521576, + 0.09133879840373993, + 0.06164053454995155, + 0.015540056861937046, + 0.03152129426598549, + 0.011258093640208244, + 0.01606227271258831, + 0.012548700906336308, + 0.013687900267541409, + 0.025401590391993523, + 0.010527023114264011, + 0.008221208117902279, + 0.02878858521580696, + 0.006283246912062168, + 0.009119421243667603, + 0.007232580799609423, + 0.009295550175011158, + 0.006157746072858572, + 0.013000975362956524, + 0.006913486402481794, + 0.006999585777521133, + 0.005181991960853338, + 0.006621244363486767, + 0.006398826837539673, + 0.023810559883713722, + 0.037758857011795044, + 0.05041615292429924, + 0.021202104166150093, + 0.015874342992901802, + 0.014226049184799194, + 0.01373277697712183, + 0.01796310395002365, + 0.009908861480653286, + 0.008385553024709225, + 0.009225331246852875, + 0.008099159225821495, + 0.007595093920826912, + 0.00472443038597703, + 0.006442502606660128, + 0.008300917223095894, + 0.007667800411581993, + 0.017135169357061386, + 0.006942101288586855, + 0.00794352125376463, + 0.013554076664149761, + 0.012235593050718307, + 0.008619297295808792, + 0.012261022813618183, + 0.003203907050192356, + 0.003522280603647232, + 0.00405778968706727, + 0.00291782571002841, + 0.014009821228682995, + 0.004920518025755882, + 0.010405244305729866, + 0.004911220632493496, + 0.00217406521551311, + 0.002735262271016836, + 0.0035197839606553316, + 0.0028098176699131727, + 0.002077934332191944, + 0.003275071270763874, + 0.0018621948547661304, + 0.0021309165749698877, + 0.007166577503085136, + 0.005875145550817251, + 0.00724567798897624, + 0.007112816441804171, + 0.007340169511735439, + 0.006739321164786816, + 0.005141158122569323, + 0.004797912668436766, + 0.014003663323819637, + 0.01262608077377081, + 0.016306374222040176, + 0.017495181411504745, + 0.015816068276762962, + 0.009476037696003914, + 0.011767098680138588, + 0.008840690366923809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.016109542921185493 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006751700770109892 + ], + "opposite_logprob_tokens": [ + -5.006751537322998 + ], + "logprob_reward": -0.006751700770109892, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751700770109892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3637627065181732, + 0.1302308440208435, + 0.04626630246639252, + 0.05396629869937897, + 0.07682798057794571, + 0.059795308858156204, + 0.05456080287694931, + 0.09293661266565323, + 0.030368831008672714, + 0.09128426760435104 + ], + "credit_map": [ + 0.033742960542440414, + 0.07841578125953674, + 0.12660938501358032, + 0.017860708758234978, + 0.04031936079263687, + 0.02296151965856552, + 0.021378109231591225, + 0.022474879398941994, + 0.014991121366620064, + 0.01275066751986742, + 0.012912520207464695, + 0.011074035428464413, + 0.024092623963952065, + 0.020289286971092224, + 0.012075327336788177, + 0.022045275196433067, + 0.005005426239222288, + 0.011226377449929714, + 0.005610391963273287, + 0.00515395263209939, + 0.006701281294226646, + 0.003771793795749545, + 0.004789493978023529, + 0.004007586743682623, + 0.004907074850052595, + 0.008887425996363163, + 0.006004313472658396, + 0.009649412706494331, + 0.0046843234449625015, + 0.005012612324208021, + 0.005307555664330721, + 0.009513582102954388, + 0.01896403729915619, + 0.008714959025382996, + 0.013890880160033703, + 0.007044182159006596, + 0.006926137488335371, + 0.008977864868938923, + 0.005121319089084864, + 0.00718859676271677, + 0.006604338530451059, + 0.013868764042854309, + 0.005950345657765865, + 0.005879245698451996, + 0.009805519133806229, + 0.008957845158874989, + 0.004915423691272736, + 0.0038138297386467457, + 0.004862959496676922, + 0.005700535606592894, + 0.008368309587240219, + 0.004721084143966436, + 0.00724419578909874, + 0.011613390408456326, + 0.007631228771060705, + 0.004419101867824793, + 0.005730284843593836, + 0.004734428133815527, + 0.005242752842605114, + 0.03531930223107338, + 0.006257680710405111, + 0.0034924778155982494, + 0.026835337281227112, + 0.005324346479028463, + 0.005169990472495556, + 0.0026620521675795317, + 0.0034107198007404804, + 0.0035931372549384832, + 0.004124420695006847, + 0.006029562093317509, + 0.0023358704056590796, + 0.0030430792830884457, + 0.013000152073800564, + 0.009937097318470478, + 0.010980084538459778, + 0.016547342762351036, + 0.01343152392655611, + 0.010298616252839565, + 0.010843220166862011, + 0.0062462384812533855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.006751700770109892 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207024306058884 + ], + "opposite_logprob_tokens": [ + -2.81207013130188 + ], + "logprob_reward": -0.06207024306058884, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207024306058884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2237606793642044, + 0.11038980633020401, + 0.1282302439212799, + 0.06220117583870888, + 0.06553106755018234, + 0.12861010432243347, + 0.027589650824666023, + 0.08371302485466003, + 0.054781027138233185, + 0.11519329249858856 + ], + "credit_map": [ + 0.02059393748641014, + 0.05298704653978348, + 0.047946806997060776, + 0.013100935146212578, + 0.025271261110901833, + 0.014287755824625492, + 0.016512706875801086, + 0.033060215413570404, + 0.013989896513521671, + 0.014337986707687378, + 0.013031006790697575, + 0.010852217674255371, + 0.012727595865726471, + 0.01567511446774006, + 0.010476135648787022, + 0.019299842417240143, + 0.020347321406006813, + 0.01603652536869049, + 0.028759567067027092, + 0.014010971412062645, + 0.015170605853199959, + 0.00906581524759531, + 0.011346453800797462, + 0.013492973521351814, + 0.009193962439894676, + 0.008907741867005825, + 0.00782602559775114, + 0.006986134219914675, + 0.007885871455073357, + 0.006321157794445753, + 0.006381009705364704, + 0.0086992671713233, + 0.012582815252244473, + 0.006640314124524593, + 0.00802238006144762, + 0.008586710318922997, + 0.0069470638409256935, + 0.007627538405358791, + 0.0068585691042244434, + 0.008265668526291847, + 0.01004006713628769, + 0.025396935641765594, + 0.010484433732926846, + 0.010939231142401695, + 0.019454844295978546, + 0.026099201291799545, + 0.011938686482608318, + 0.01425669714808464, + 0.004047457594424486, + 0.0023542018607258797, + 0.0037340063136070967, + 0.0025973229203373194, + 0.005623657256364822, + 0.003757015336304903, + 0.0031052527483552694, + 0.0023707349319010973, + 0.005615762434899807, + 0.004387137480080128, + 0.005115262232720852, + 0.031310953199863434, + 0.0056771100498735905, + 0.004376509692519903, + 0.022476736456155777, + 0.004753548186272383, + 0.005855618976056576, + 0.005914295092225075, + 0.005606482271105051, + 0.008788149803876877, + 0.008288590237498283, + 0.00943798664957285, + 0.005023432429879904, + 0.00586647167801857, + 0.020760787650942802, + 0.01182533334940672, + 0.014391407370567322, + 0.014373105950653553, + 0.017916014418005943, + 0.014469309709966183, + 0.012405846267938614, + 0.009051479399204254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.06207024306058884 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a black Maybach luxury sedan?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera executing a low-angle dynamic tracking shot following the car?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vibrant neon signs in red, blue, purple, and green colors visible on the surrounding buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dense raindrops clearly visible in the beams of the car's LED headlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene depict a bustling city with wet streets and a cyberpunk atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817591419443488 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.002817591419443488, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817591419443488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.272144615650177, + 0.05809156224131584, + 0.07237548381090164, + 0.07860877364873886, + 0.10893788933753967, + 0.0720401257276535, + 0.05527298152446747, + 0.11941581964492798, + 0.048092179000377655, + 0.11502066254615784 + ], + "credit_map": [ + 0.022187601774930954, + 0.09259750694036484, + 0.07259801030158997, + 0.01246056891977787, + 0.03014337830245495, + 0.012318686582148075, + 0.01402370072901249, + 0.015815138816833496, + 0.010580467991530895, + 0.007446622475981712, + 0.007116577588021755, + 0.0074197337962687016, + 0.006519998423755169, + 0.004555811174213886, + 0.008938211016356945, + 0.005514135118573904, + 0.013218572363257408, + 0.010617726482450962, + 0.014895211905241013, + 0.006375613156706095, + 0.006147629115730524, + 0.005309952422976494, + 0.0070074331015348434, + 0.008803347125649452, + 0.011910498142242432, + 0.01686105877161026, + 0.011430724523961544, + 0.009606823325157166, + 0.007304878905415535, + 0.008934061042964458, + 0.0055188704282045364, + 0.0070418547838926315, + 0.025557586923241615, + 0.011958584189414978, + 0.022440176457166672, + 0.01008602324873209, + 0.0071074459701776505, + 0.01575888693332672, + 0.007254047319293022, + 0.008775129914283752, + 0.007485211361199617, + 0.01485582534223795, + 0.007963517680764198, + 0.0068377358838915825, + 0.01115454826503992, + 0.010418283753097057, + 0.006637498736381531, + 0.00668750237673521, + 0.00507697369903326, + 0.004792073275893927, + 0.006827523931860924, + 0.005127389449626207, + 0.01282060518860817, + 0.005660885013639927, + 0.009197193197906017, + 0.005770339164882898, + 0.006964964792132378, + 0.006410912610590458, + 0.0069129918701946735, + 0.04573985934257507, + 0.008314577862620354, + 0.0063390908762812614, + 0.03259150683879852, + 0.006141917780041695, + 0.006697480566799641, + 0.006731423083692789, + 0.006349187344312668, + 0.0055387308821082115, + 0.005827515851706266, + 0.010175380855798721, + 0.002806910779327154, + 0.003965552896261215, + 0.023383744060993195, + 0.013275954872369766, + 0.012972453609108925, + 0.016699666157364845, + 0.016087545081973076, + 0.012440943159162998, + 0.013086862862110138, + 0.007073494140058756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.002817591419443488 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_12/index_tensor([6106, 6106])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_12/index_tensor([6106, 6106])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..79f777f370453c437abb1d099793f5f413c4d0c0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_12/index_tensor([6106, 6106])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up static shot captures a profoundly intimate moment, tightly focused on the profile of a person's face. The scene is illuminated by a single, warm, soft light source, like a pillowcase lamp, casting gentle shadows and highlighting the delicate texture of their skin. A perfectly formed tear falls onto their cheek, shimmering, its surface reflecting the warm light. The soft, golden muzzle of a dog gently intrudes from the edge of the frame. The camera maintains focus as the dog's soft, pink tongue extends with infinite care, making delicate contact with the person's skin. Carefully and tenderly, the dog licks away the lonely tear with slow, soothing movements, silently offering comfort. After a brief moment, the dog's muzzle rests against the person's cheek, its soft fur a tangible comfort. The shallow depth of field creates an atmosphere of deep tranquility and emotional resonance, filled with unconditional love and quiet support, blurring the background into a warm, fuzzy pattern.", + "seed": 1782, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_12/index_6106_seed_1782_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06410903440555557, + "vlm_logprob": -0.06410903440555557, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.06410903440555557, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007259420119225979 + ], + "opposite_logprob_tokens": [ + -5.632259368896484 + ], + "logprob_reward": -0.007259420119225979, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.007259420119225979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15951108932495117, + 0.0871720016002655, + 0.028997618705034256, + 0.019381893798708916, + 0.06925490498542786, + 0.19604241847991943, + 0.08720650523900986, + 0.13054172694683075, + 0.11215856671333313, + 0.10973332077264786 + ], + "credit_map": [ + 0.01358040701597929, + 0.01983676478266716, + 0.022120820358395576, + 0.01983828656375408, + 0.02301352471113205, + 0.01367135439068079, + 0.015440231189131737, + 0.03200969845056534, + 0.0046250903978943825, + 0.006272868253290653, + 0.028879109770059586, + 0.03102612867951393, + 0.0026777596212923527, + 0.0037482245825231075, + 0.0047767250798642635, + 0.0051660980097949505, + 0.002348761074244976, + 0.003629556857049465, + 0.0059817214496433735, + 0.0036582278553396463, + 0.0038363677449524403, + 0.003415360813960433, + 0.0022954284213483334, + 0.0038321942556649446, + 0.002235585357993841, + 0.00237627443857491, + 0.0024681829381734133, + 0.0015473896637558937, + 0.0027390734758228064, + 0.0026310845278203487, + 0.002531911013647914, + 0.002852391917258501, + 0.006072798743844032, + 0.006671628449112177, + 0.009858494624495506, + 0.011084592901170254, + 0.010170415043830872, + 0.010124778375029564, + 0.006542093586176634, + 0.008730098605155945, + 0.01835675723850727, + 0.01565217785537243, + 0.019587533548474312, + 0.020079558715224266, + 0.04822460189461708, + 0.02300865389406681, + 0.02746150828897953, + 0.02367161028087139, + 0.007473623845726252, + 0.00860644318163395, + 0.012014462612569332, + 0.01091516949236393, + 0.018355682492256165, + 0.008258206769824028, + 0.00936497375369072, + 0.012217946350574493, + 0.012094371020793915, + 0.014219813980162144, + 0.016520392149686813, + 0.019080061465501785, + 0.022694429382681847, + 0.01660454086959362, + 0.01140263769775629, + 0.017925476655364037, + 0.018056606873869896, + 0.008241557516157627, + 0.02727787010371685, + 0.011888528242707253, + 0.008891390636563301, + 0.01778322644531727, + 0.00744341965764761, + 0.0125759681686759, + 0.014105971902608871, + 0.00726265599951148, + 0.023238714784383774, + 0.012105048634111881, + 0.010062094777822495, + 0.015755334869027138, + 0.013133760541677475, + 0.014069733209908009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.007259420119225979 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03375457599759102 + ], + "opposite_logprob_tokens": [ + -3.408754587173462 + ], + "logprob_reward": -0.03375457599759102, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03375457599759102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22827836871147156, + 0.06473851948976517, + 0.04384283348917961, + 0.06359256058931351, + 0.20407482981681824, + 0.09759379178285599, + 0.08491641283035278, + 0.052333731204271317, + 0.09524263441562653, + 0.06538639962673187 + ], + "credit_map": [ + 0.013235273770987988, + 0.020381731912493706, + 0.059486523270606995, + 0.0784115195274353, + 0.016368206590414047, + 0.00929050613194704, + 0.011811940930783749, + 0.019292647019028664, + 0.004631630145013332, + 0.004217954818159342, + 0.018076319247484207, + 0.015808260068297386, + 0.0040420000441372395, + 0.004352456424385309, + 0.008303028531372547, + 0.005306870676577091, + 0.005988177843391895, + 0.0046136584132909775, + 0.005666933488100767, + 0.006649976596236229, + 0.005007558036595583, + 0.003949080128222704, + 0.0073333620093762875, + 0.004634082783013582, + 0.0070722224190831184, + 0.012398949824273586, + 0.011360850185155869, + 0.005750827956944704, + 0.007839180529117584, + 0.005080726929008961, + 0.006848841905593872, + 0.007240950595587492, + 0.021260343492031097, + 0.023244312033057213, + 0.03137054294347763, + 0.03390580415725708, + 0.025753259658813477, + 0.025281351059675217, + 0.0188923180103302, + 0.02436686120927334, + 0.00900461059063673, + 0.009937742725014687, + 0.010008922778069973, + 0.014622990041971207, + 0.018049543723464012, + 0.011204469949007034, + 0.010803449898958206, + 0.01396204810589552, + 0.00743382703512907, + 0.009409014135599136, + 0.01131235621869564, + 0.011278639547526836, + 0.016327856108546257, + 0.0088468287140131, + 0.008972033858299255, + 0.011335845105350018, + 0.005763496272265911, + 0.007586236577481031, + 0.007222857791930437, + 0.005931443069130182, + 0.006794922985136509, + 0.006838664878159761, + 0.004409739747643471, + 0.007786362431943417, + 0.010482264682650566, + 0.010663525201380253, + 0.016577675938606262, + 0.008634100668132305, + 0.00887517724186182, + 0.01670362614095211, + 0.011900858953595161, + 0.011405388824641705, + 0.006606206763535738, + 0.007990862242877483, + 0.0090660834684968, + 0.011541185900568962, + 0.008061235770583153, + 0.007222534157335758, + 0.007078896276652813, + 0.00781939085572958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.03375457599759102 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2522323727607727 + ], + "opposite_logprob_tokens": [ + -1.502232313156128 + ], + "logprob_reward": -0.2522323727607727, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2522323727607727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17454804480075836, + 0.09836526960134506, + 0.04040177911520004, + 0.0372931994497776, + 0.4029475748538971, + 0.05230448395013809, + 0.039604634046554565, + 0.030663181096315384, + 0.06091615930199623, + 0.06295560300350189 + ], + "credit_map": [ + 0.011812127195298672, + 0.017906300723552704, + 0.027290714904665947, + 0.035467155277729034, + 0.019744090735912323, + 0.015314683318138123, + 0.014244173653423786, + 0.03276880085468292, + 0.005280195269733667, + 0.00475797476246953, + 0.032202064990997314, + 0.027665363624691963, + 0.005116922780871391, + 0.005717731546610594, + 0.010438043624162674, + 0.007186978589743376, + 0.0040662456303834915, + 0.0046732607297599316, + 0.004890813492238522, + 0.005703812465071678, + 0.0055353292264044285, + 0.00408273795619607, + 0.00562467472627759, + 0.005824901629239321, + 0.003215629607439041, + 0.0037685378920286894, + 0.0058298916555941105, + 0.0035672318190336227, + 0.005644051358103752, + 0.004081503488123417, + 0.0067170774564146996, + 0.004469276871532202, + 0.03684159368276596, + 0.04018951207399368, + 0.05521133542060852, + 0.060490112751722336, + 0.06326577067375183, + 0.05084452033042908, + 0.04792772978544235, + 0.04817701876163483, + 0.004466473124921322, + 0.005574566312134266, + 0.00519315293058753, + 0.00900234840810299, + 0.009538521990180016, + 0.00651613948866725, + 0.005387565121054649, + 0.006625716108828783, + 0.003630471182987094, + 0.004886428825557232, + 0.004109937697649002, + 0.007833712734282017, + 0.0058929529041051865, + 0.0038348317611962557, + 0.0041828821413218975, + 0.005233414936810732, + 0.0030063323210924864, + 0.0030055693350732327, + 0.003995157312601805, + 0.00404849648475647, + 0.005669152829796076, + 0.003082763869315386, + 0.0030912463553249836, + 0.004764464683830738, + 0.007393489591777325, + 0.00706758676096797, + 0.012072496116161346, + 0.00716518284752965, + 0.0063261305913329124, + 0.008205092512071133, + 0.006565394811332226, + 0.006120785605162382, + 0.008250127546489239, + 0.00504617253318429, + 0.011995172128081322, + 0.008813404478132725, + 0.005476190708577633, + 0.008049252443015575, + 0.00716816633939743, + 0.008157119154930115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.2522323727607727 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636999823153019 + ], + "opposite_logprob_tokens": [ + -3.651370048522949 + ], + "logprob_reward": -0.02636999823153019, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636999823153019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1884283572435379, + 0.04687793180346489, + 0.01890569180250168, + 0.062466345727443695, + 0.1804354190826416, + 0.12889158725738525, + 0.10423349589109421, + 0.07935000211000443, + 0.09889953583478928, + 0.09151168167591095 + ], + "credit_map": [ + 0.01087761390954256, + 0.01607491448521614, + 0.0512690469622612, + 0.06437553465366364, + 0.012062144465744495, + 0.007763326168060303, + 0.009848983958363533, + 0.01615678146481514, + 0.0032205272000283003, + 0.004723347257822752, + 0.00925581343472004, + 0.00699826143682003, + 0.004520227666944265, + 0.0031949172262102365, + 0.009217587299644947, + 0.005747254937887192, + 0.0018822320271283388, + 0.0026475051417946815, + 0.0019462215714156628, + 0.0033566909842193127, + 0.002487550489604473, + 0.002009691670536995, + 0.002282525645568967, + 0.002293272875249386, + 0.005332887638360262, + 0.007429417688399553, + 0.009350775741040707, + 0.006200436502695084, + 0.00748490123078227, + 0.006374831777065992, + 0.011604375205934048, + 0.008688717149198055, + 0.015604383312165737, + 0.018693262711167336, + 0.02617036923766136, + 0.026105521246790886, + 0.029519641771912575, + 0.023783210664987564, + 0.018462562933564186, + 0.022096475586295128, + 0.011310496367514133, + 0.011435525491833687, + 0.013608324341475964, + 0.021297484636306763, + 0.02455396018922329, + 0.014115746133029461, + 0.01508317980915308, + 0.017486874014139175, + 0.010214468464255333, + 0.011102544143795967, + 0.014251909218728542, + 0.0162412878125906, + 0.01917661726474762, + 0.009897775016725063, + 0.011923860758543015, + 0.011425039730966091, + 0.008825687691569328, + 0.0077456398867070675, + 0.011652977205812931, + 0.010479181073606014, + 0.014139504171907902, + 0.008423051796853542, + 0.006209234241396189, + 0.011874726973474026, + 0.010903538204729557, + 0.013957138173282146, + 0.014560231007635593, + 0.012078682892024517, + 0.010780011303722858, + 0.01412882748991251, + 0.013659751042723656, + 0.008831351064145565, + 0.009094190783798695, + 0.008984167128801346, + 0.010542485862970352, + 0.011618627235293388, + 0.013402413576841354, + 0.011278231628239155, + 0.013734325766563416, + 0.012857241556048393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.02636999823153019 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009288049186579883 + ], + "opposite_logprob_tokens": [ + -7.00092887878418 + ], + "logprob_reward": -0.0009288049186579883, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009288049186579883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19233602285385132, + 0.24492648243904114, + 0.024563942104578018, + 0.051974210888147354, + 0.07179941236972809, + 0.07993999868631363, + 0.067352294921875, + 0.08677057176828384, + 0.07261434197425842, + 0.10772272199392319 + ], + "credit_map": [ + 0.010834939777851105, + 0.010723709128797054, + 0.05619512498378754, + 0.057537149637937546, + 0.015906237065792084, + 0.0095409220084548, + 0.009649465791881084, + 0.02194845862686634, + 0.010614012368023396, + 0.029153281822800636, + 0.06895823031663895, + 0.09451301395893097, + 0.009719044901430607, + 0.009157839231193066, + 0.012642541900277138, + 0.010168502107262611, + 0.0028054267168045044, + 0.004011895973235369, + 0.004056669306010008, + 0.00391016760841012, + 0.0030081027653068304, + 0.0020915039349347353, + 0.0022738766856491566, + 0.002406297717243433, + 0.005472387652844191, + 0.006001061759889126, + 0.007534210570156574, + 0.004761688876897097, + 0.006531181279569864, + 0.005278296303004026, + 0.007683379575610161, + 0.008712004870176315, + 0.006742984987795353, + 0.0070974756963551044, + 0.009744767099618912, + 0.00847572274506092, + 0.012438813224434853, + 0.009391282685101032, + 0.006890761200338602, + 0.01101760845631361, + 0.009342855773866177, + 0.006996354088187218, + 0.010249941609799862, + 0.010049676522612572, + 0.013096625916659832, + 0.013377362862229347, + 0.00903412327170372, + 0.007793056312948465, + 0.006877790205180645, + 0.006970307789742947, + 0.009545138105750084, + 0.008563250303268433, + 0.012398771941661835, + 0.006408250890672207, + 0.007741862442344427, + 0.008846919052302837, + 0.007797396741807461, + 0.010369377210736275, + 0.011633767746388912, + 0.01230192556977272, + 0.01464702095836401, + 0.011125349439680576, + 0.007008018437772989, + 0.011887715198099613, + 0.011785280890762806, + 0.005802124738693237, + 0.0159404706209898, + 0.007721948903053999, + 0.005869386252015829, + 0.011325378902256489, + 0.005320672411471605, + 0.008849076926708221, + 0.013639479875564575, + 0.006794302258640528, + 0.023811671882867813, + 0.012628278695046902, + 0.008141852915287018, + 0.016407808288931847, + 0.014196611009538174, + 0.012102717533707619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0009288049186579883 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures a profoundly intimate moment, tightly focused on the profile of a person's face. The scene is illuminated by a single, warm, soft light source, like a pillowcase lamp, casting gentle shadows and highlighting the delicate texture of their skin. A perfectly formed tear falls onto their cheek, shimmering, its surface reflecting the warm light. The soft, golden muzzle of a dog gently intrudes from the edge of the frame. The camera maintains focus as the dog's soft, pink tongue extends with infinite care, making delicate contact with the person's skin. Carefully and tenderly, the dog licks away the lonely tear with slow, soothing movements, silently offering comfort. After a brief moment, the dog's muzzle rests against the person's cheek, its soft fur a tangible comfort. The shallow depth of field creates an atmosphere of deep tranquility and emotional resonance, filled with unconditional love and quiet support, blurring the background into a warm, fuzzy pattern.", + "seed": 1783, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_12/index_6106_seed_1783_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15768131674267352, + "vlm_logprob": -0.15768131674267352, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.15768131674267352, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005690090823918581 + ], + "opposite_logprob_tokens": [ + -6.880690097808838 + ], + "logprob_reward": -0.005690090823918581, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.005690090823918581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13148967921733856, + 0.20819202065467834, + 0.05531974509358406, + 0.08959857374429703, + 0.12386766821146011, + 0.08356981724500656, + 0.1262822151184082, + 0.06752242147922516, + 0.056994400918483734, + 0.05716348811984062 + ], + "credit_map": [ + 0.017042288556694984, + 0.011663474142551422, + 0.0320851132273674, + 0.023246703669428825, + 0.0069482652470469475, + 0.015669653192162514, + 0.010760579258203506, + 0.014073595404624939, + 0.013558930717408657, + 0.009391906671226025, + 0.01995135098695755, + 0.015009610913693905, + 0.023355629295110703, + 0.05402720347046852, + 0.016813449561595917, + 0.05608394369482994, + 0.0054870229214429855, + 0.005557999946177006, + 0.010715598240494728, + 0.006777072325348854, + 0.005995568819344044, + 0.00803006999194622, + 0.007167801260948181, + 0.005588609259575605, + 0.010362801142036915, + 0.009313761256635189, + 0.017846575006842613, + 0.007818292826414108, + 0.015456153079867363, + 0.011997843161225319, + 0.008208147250115871, + 0.0085949981585145, + 0.012181528843939304, + 0.014173638075590134, + 0.019909624010324478, + 0.01701144501566887, + 0.01415290031582117, + 0.015561675652861595, + 0.018210941925644875, + 0.012665899470448494, + 0.008932206779718399, + 0.012144147418439388, + 0.013243359513580799, + 0.008953982032835484, + 0.011794837191700935, + 0.009915702976286411, + 0.011549395509064198, + 0.007036179769784212, + 0.01022431068122387, + 0.012410788796842098, + 0.03231850638985634, + 0.014565160498023033, + 0.019740179181098938, + 0.016654122620821, + 0.0114959217607975, + 0.008873219601809978, + 0.006434903014451265, + 0.005520741455256939, + 0.014403804205358028, + 0.007815093733370304, + 0.007381063420325518, + 0.009355326183140278, + 0.01127045601606369, + 0.005341032054275274, + 0.006138146389275789, + 0.0052426885813474655, + 0.009227204136550426, + 0.007771333213895559, + 0.005168576259166002, + 0.007140521891415119, + 0.008842971175909042, + 0.0074629588052630424, + 0.007445373572409153, + 0.00859665684401989, + 0.005970868282020092, + 0.007984692230820656, + 0.0073521677404642105, + 0.006968130823224783, + 0.006672677583992481, + 0.006172920111566782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.005690090823918581 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07010417431592941 + ], + "opposite_logprob_tokens": [ + -2.6951041221618652 + ], + "logprob_reward": -0.07010417431592941, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07010417431592941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030410656705498695, + 0.07672137022018433, + 0.04120466858148575, + 0.16511201858520508, + 0.06488150358200073, + 0.030651263892650604, + 0.03722045198082924, + 0.24510760605335236, + 0.031192131340503693, + 0.277498334646225 + ], + "credit_map": [ + 0.003940011374652386, + 0.003252206137403846, + 0.00506174098700285, + 0.004812669474631548, + 0.002841012319549918, + 0.003526038955897093, + 0.0035209094639867544, + 0.003456069389358163, + 0.005637789610773325, + 0.0052589657716453075, + 0.008908075280487537, + 0.009401489980518818, + 0.012177862226963043, + 0.014242851175367832, + 0.004968852736055851, + 0.01612548530101776, + 0.0036420461256057024, + 0.003330581821501255, + 0.010680343955755234, + 0.006168747320771217, + 0.0035838487092405558, + 0.0064142667688429356, + 0.0026985956355929375, + 0.004686235450208187, + 0.0192252267152071, + 0.015824003145098686, + 0.03346284478902817, + 0.017286187037825584, + 0.02446630969643593, + 0.02175154909491539, + 0.0173311997205019, + 0.01576470024883747, + 0.006067454349249601, + 0.007573304232209921, + 0.008873366750776768, + 0.008855261839926243, + 0.00856570340692997, + 0.007696634624153376, + 0.010103079490363598, + 0.007146690972149372, + 0.003590015461668372, + 0.004263716284185648, + 0.0031948471441864967, + 0.0033835540525615215, + 0.004112110007554293, + 0.00353088416159153, + 0.005643793847411871, + 0.002932341769337654, + 0.0026154760271310806, + 0.0022603038232773542, + 0.009105238132178783, + 0.0059642367996275425, + 0.0037060470785945654, + 0.005451778881251812, + 0.006184882018715143, + 0.0019324898021295667, + 0.02318284660577774, + 0.024089034646749496, + 0.05552047863602638, + 0.02343064360320568, + 0.024563509970903397, + 0.03368173539638519, + 0.04226711392402649, + 0.018372220918536186, + 0.003942615818232298, + 0.0026194765232503414, + 0.004871778655797243, + 0.0046636564657092094, + 0.00209550722502172, + 0.004351504612714052, + 0.005056583788245916, + 0.003591009881347418, + 0.038895200937986374, + 0.03881818428635597, + 0.030908901244401932, + 0.04726264625787735, + 0.03355898708105087, + 0.03193415701389313, + 0.03217332065105438, + 0.02394692599773407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.07010417431592941 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.693480908870697 + ], + "opposite_logprob_tokens": [ + -0.693480908870697 + ], + "logprob_reward": -0.693480908870697, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.693480908870697, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09373076260089874, + 0.3249340057373047, + 0.051597628742456436, + 0.0160344447940588, + 0.30962643027305603, + 0.01938994787633419, + 0.03503947705030441, + 0.054070670157670975, + 0.04096357524394989, + 0.054613083600997925 + ], + "credit_map": [ + 0.007637820206582546, + 0.006951917428523302, + 0.01959318481385708, + 0.014414803124964237, + 0.005974502302706242, + 0.016305487602949142, + 0.00780255114659667, + 0.015050491318106651, + 0.021127616986632347, + 0.02587803825736046, + 0.03929976746439934, + 0.03024912066757679, + 0.032020147889852524, + 0.05657200515270233, + 0.061771687120199203, + 0.0580156072974205, + 0.005032154731452465, + 0.004627170506864786, + 0.01007913239300251, + 0.007437196094542742, + 0.005822071339935064, + 0.008026833645999432, + 0.005657606292515993, + 0.0049154614098370075, + 0.0014046969590708613, + 0.0020357558969408274, + 0.0018716028425842524, + 0.002218721667304635, + 0.0022320416755974293, + 0.0023341812193393707, + 0.001942256698384881, + 0.0019951865542680025, + 0.02696707844734192, + 0.03009369783103466, + 0.049980439245700836, + 0.047188401222229004, + 0.03606792166829109, + 0.039652422070503235, + 0.0486336350440979, + 0.0310427937656641, + 0.0019151612650603056, + 0.0018853219226002693, + 0.002594274003058672, + 0.0026772208511829376, + 0.003138558007776737, + 0.0027846505399793386, + 0.002408906351774931, + 0.001985854934900999, + 0.0031134216114878654, + 0.0036350367590785027, + 0.007621164433658123, + 0.00465736910700798, + 0.004848866257816553, + 0.004587836097925901, + 0.0043627070263028145, + 0.0022130776196718216, + 0.005400395020842552, + 0.00568403908982873, + 0.012337223626673222, + 0.005285629536956549, + 0.005608867853879929, + 0.007274090778082609, + 0.008705050684511662, + 0.0037753733340650797, + 0.005028717219829559, + 0.0037319008260965347, + 0.0059206425212323666, + 0.005974920466542244, + 0.0027963819447904825, + 0.005622390192002058, + 0.006914107128977776, + 0.004974514711648226, + 0.005488562863320112, + 0.006043389905244112, + 0.006220484152436256, + 0.008081615902483463, + 0.006034032441675663, + 0.006976528093218803, + 0.007848287932574749, + 0.007920186035335064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.693480908870697 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820462942123413 + ], + "opposite_logprob_tokens": [ + -4.018204689025879 + ], + "logprob_reward": -0.01820462942123413, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01820462942123413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05970188230276108, + 0.1757877618074417, + 0.03759868070483208, + 0.049357034265995026, + 0.41007548570632935, + 0.07044553011655807, + 0.03637554869055748, + 0.05285707861185074, + 0.05851028114557266, + 0.04929068684577942 + ], + "credit_map": [ + 0.004434201400727034, + 0.0047788359224796295, + 0.013531596399843693, + 0.011409126222133636, + 0.005161216948181391, + 0.00784480944275856, + 0.00583190843462944, + 0.006710182875394821, + 0.012888353317975998, + 0.014836202375590801, + 0.02226470224559307, + 0.017886733636260033, + 0.016921276226639748, + 0.027347980067133904, + 0.0352533683180809, + 0.028389157727360725, + 0.0029281466268002987, + 0.0023770793341100216, + 0.011659882962703705, + 0.004620097577571869, + 0.0028114186134189367, + 0.006103444844484329, + 0.0030824376735836267, + 0.0040161702781915665, + 0.00491902744397521, + 0.004693543538451195, + 0.009330556727945805, + 0.00494325952604413, + 0.008531966246664524, + 0.006568498909473419, + 0.005139696877449751, + 0.0052304863929748535, + 0.038393471390008926, + 0.04365098103880882, + 0.06540259718894958, + 0.0589057095348835, + 0.045865267515182495, + 0.05178888887166977, + 0.06272010505199432, + 0.04334846884012222, + 0.006679318845272064, + 0.005020991433411837, + 0.009629075415432453, + 0.011304290033876896, + 0.007673918269574642, + 0.008851214312016964, + 0.01339805405586958, + 0.007888665422797203, + 0.0033492373768240213, + 0.004000874236226082, + 0.01036981400102377, + 0.0035497695207595825, + 0.0050825085490942, + 0.004596459213644266, + 0.0033845913130789995, + 0.0020422975067049265, + 0.00514270318672061, + 0.004372537601739168, + 0.011027863249182701, + 0.004358427133411169, + 0.003951879218220711, + 0.007181850727647543, + 0.011462493799626827, + 0.005359326489269733, + 0.007087105419486761, + 0.004771828651428223, + 0.009313498623669147, + 0.009396077133715153, + 0.004318231716752052, + 0.007533848285675049, + 0.009702904149889946, + 0.0063867876306176186, + 0.004662840161472559, + 0.004000641405582428, + 0.006214646156877279, + 0.004969088360667229, + 0.00538446381688118, + 0.007160603534430265, + 0.007353971246629953, + 0.009544426575303078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 5 + ], + "selected_reward": -0.01820462942123413 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009267802815884352 + ], + "opposite_logprob_tokens": [ + -7.000926971435547 + ], + "logprob_reward": -0.0009267802815884352, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009267802815884352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08690278232097626, + 0.11366179585456848, + 0.04245775565505028, + 0.0553349032998085, + 0.2855706810951233, + 0.05280357226729393, + 0.027696797624230385, + 0.08213850110769272, + 0.018763594329357147, + 0.23466956615447998 + ], + "credit_map": [ + 0.00977314356714487, + 0.006154220085591078, + 0.022349240258336067, + 0.014699951745569706, + 0.004642381798475981, + 0.012639900669455528, + 0.005194986239075661, + 0.011448956094682217, + 0.0068593379110097885, + 0.007501350715756416, + 0.010924797505140305, + 0.007362936157733202, + 0.009509679861366749, + 0.028965970501303673, + 0.01313001848757267, + 0.029407702386379242, + 0.003631730331107974, + 0.004451458808034658, + 0.007830148562788963, + 0.005387717392295599, + 0.005115689244121313, + 0.00644666375592351, + 0.0051736594177782536, + 0.004420687910169363, + 0.0055689262226223946, + 0.004211376886814833, + 0.012527583166956902, + 0.006955543998628855, + 0.005940411239862442, + 0.008293697610497475, + 0.005690595135092735, + 0.006146770901978016, + 0.02663157507777214, + 0.03356461226940155, + 0.04192587733268738, + 0.03979077190160751, + 0.04015609622001648, + 0.035736773163080215, + 0.04026874899864197, + 0.027496229857206345, + 0.00784076377749443, + 0.010924355126917362, + 0.004608185961842537, + 0.003869209671393037, + 0.007292595691978931, + 0.007078868802636862, + 0.006846302188932896, + 0.004343290347605944, + 0.002812845166772604, + 0.002463693730533123, + 0.0067889876663684845, + 0.0027861963026225567, + 0.003404104383662343, + 0.003919078968465328, + 0.003616179572418332, + 0.0019057143945246935, + 0.008409842848777771, + 0.005645040422677994, + 0.01733904890716076, + 0.0075815599411726, + 0.005805295892059803, + 0.011218245141208172, + 0.016973724588751793, + 0.0091657480224967, + 0.001696380670182407, + 0.0015174351865425706, + 0.0024428488686680794, + 0.002779455156996846, + 0.0015939543955028057, + 0.0023735163267701864, + 0.0031801718287169933, + 0.003179832361638546, + 0.02632369101047516, + 0.028994357213377953, + 0.025833718478679657, + 0.04145379737019539, + 0.02769363671541214, + 0.027186550199985504, + 0.027874663472175598, + 0.029309159144759178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.0009267802815884352 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures a profoundly intimate moment, tightly focused on the profile of a person's face. The scene is illuminated by a single, warm, soft light source, like a pillowcase lamp, casting gentle shadows and highlighting the delicate texture of their skin. A perfectly formed tear falls onto their cheek, shimmering, its surface reflecting the warm light. The soft, golden muzzle of a dog gently intrudes from the edge of the frame. The camera maintains focus as the dog's soft, pink tongue extends with infinite care, making delicate contact with the person's skin. Carefully and tenderly, the dog licks away the lonely tear with slow, soothing movements, silently offering comfort. After a brief moment, the dog's muzzle rests against the person's cheek, its soft fur a tangible comfort. The shallow depth of field creates an atmosphere of deep tranquility and emotional resonance, filled with unconditional love and quiet support, blurring the background into a warm, fuzzy pattern.", + "seed": 1784, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_12/index_6106_seed_1784_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09029536014422775, + "vlm_logprob": -0.09029536014422775, + "vlm_yes_no_margin": 4.125000011920929, + "vlm_token_credit": -0.09029536014422775, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010150950402021408 + ], + "opposite_logprob_tokens": [ + -6.010150909423828 + ], + "logprob_reward": -0.010150950402021408, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.010150950402021408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.092757448554039, + 0.04454616829752922, + 0.10379939526319504, + 0.06879767775535583, + 0.14232969284057617, + 0.0455915592610836, + 0.14630334079265594, + 0.1475919634103775, + 0.16077953577041626, + 0.04750315845012665 + ], + "credit_map": [ + 0.009710179641842842, + 0.007268059067428112, + 0.010516883805394173, + 0.010482962243258953, + 0.010817294009029865, + 0.012923166155815125, + 0.021310623735189438, + 0.009728276170790195, + 0.006428633816540241, + 0.0044148163869977, + 0.008378727361559868, + 0.006226545199751854, + 0.005140322260558605, + 0.00428759353235364, + 0.004647275432944298, + 0.00502225523814559, + 0.011269974522292614, + 0.012773114256560802, + 0.014395413920283318, + 0.010993058793246746, + 0.011245804838836193, + 0.01635706052184105, + 0.014063338749110699, + 0.012701635248959064, + 0.007775019854307175, + 0.007692329119890928, + 0.00816059298813343, + 0.00983502808958292, + 0.005745823495090008, + 0.010790062136948109, + 0.011917068623006344, + 0.006881757639348507, + 0.013973112218081951, + 0.01187884621322155, + 0.018199242651462555, + 0.017664410173892975, + 0.01962057128548622, + 0.016260121017694473, + 0.02413266710937023, + 0.020600728690624237, + 0.00632095942273736, + 0.0027392094489187002, + 0.006657173857092857, + 0.003308376530185342, + 0.005536507815122604, + 0.004642842803150415, + 0.0064810398034751415, + 0.009905449114739895, + 0.016410019248723984, + 0.013468349352478981, + 0.021413851529359818, + 0.02262771688401699, + 0.013034282252192497, + 0.014487716369330883, + 0.027322081848978996, + 0.01753932237625122, + 0.018850868567824364, + 0.010280042886734009, + 0.02192695066332817, + 0.013835594058036804, + 0.01662285067141056, + 0.01490089576691389, + 0.017190849408507347, + 0.033983904868364334, + 0.016342874616384506, + 0.01571827009320259, + 0.017644667997956276, + 0.012903157621622086, + 0.017520274966955185, + 0.03704296052455902, + 0.020405462011694908, + 0.023201867938041687, + 0.0044616288505494595, + 0.006812614854425192, + 0.00527994753792882, + 0.007673180196434259, + 0.006789804436266422, + 0.007677470333874226, + 0.004915202502161264, + 0.00389331066980958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.010150950402021408 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0381646566092968 + ], + "opposite_logprob_tokens": [ + -3.2881646156311035 + ], + "logprob_reward": -0.0381646566092968, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0381646566092968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13910429179668427, + 0.024481160566210747, + 0.09219592064619064, + 0.16865821182727814, + 0.08802345395088196, + 0.06188983470201492, + 0.22922350466251373, + 0.09278874844312668, + 0.07765752822160721, + 0.025977373123168945 + ], + "credit_map": [ + 0.013244700618088245, + 0.009199737571179867, + 0.014099222607910633, + 0.015874532982707024, + 0.017378725111484528, + 0.019820697605609894, + 0.020847419276833534, + 0.028639251366257668, + 0.004069629590958357, + 0.0027033486403524876, + 0.0045821587555110455, + 0.0031582615338265896, + 0.002862695138901472, + 0.0020169117487967014, + 0.0024123694747686386, + 0.002675785683095455, + 0.010161202400922775, + 0.008840980939567089, + 0.011860107071697712, + 0.0110031608492136, + 0.007847259752452374, + 0.014616556465625763, + 0.014696383848786354, + 0.013170274905860424, + 0.016723085194826126, + 0.011278104037046432, + 0.023573584854602814, + 0.017886225134134293, + 0.01576630212366581, + 0.01578798145055771, + 0.03425600007176399, + 0.033386919647455215, + 0.0072281030006706715, + 0.009351721033453941, + 0.012069190852344036, + 0.011125754565000534, + 0.014057379215955734, + 0.00942443497478962, + 0.014101582579314709, + 0.010665295645594597, + 0.009576814249157906, + 0.00413518724963069, + 0.008815730921924114, + 0.004108527675271034, + 0.0063814870081841946, + 0.0073784166015684605, + 0.009174156002700329, + 0.012319514527916908, + 0.0213753804564476, + 0.018971819430589676, + 0.035056911408901215, + 0.03275887668132782, + 0.01846294477581978, + 0.022271640598773956, + 0.049514152109622955, + 0.03081177920103073, + 0.010036476887762547, + 0.007850819267332554, + 0.013037877157330513, + 0.010079366154968739, + 0.011057859286665916, + 0.010372206568717957, + 0.010814736597239971, + 0.01953941024839878, + 0.006869385484606028, + 0.00958593375980854, + 0.008405571803450584, + 0.011676321737468243, + 0.00723494915291667, + 0.012692702934145927, + 0.007549024187028408, + 0.013643643818795681, + 0.0029823018703609705, + 0.0036530042998492718, + 0.0033869680482894182, + 0.003520111320540309, + 0.003136203857138753, + 0.004187810234725475, + 0.0030316768679767847, + 0.0020792968571186066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.0381646566092968 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38722938299179077 + ], + "opposite_logprob_tokens": [ + -1.1372294425964355 + ], + "logprob_reward": -0.38722938299179077, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38722938299179077, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04378838092088699, + 0.03987404331564903, + 0.18149949610233307, + 0.1652769297361374, + 0.10719593614339828, + 0.03521193936467171, + 0.35289016366004944, + 0.022618189454078674, + 0.04171453416347504, + 0.009930385276675224 + ], + "credit_map": [ + 0.004205675330013037, + 0.0033813579939305782, + 0.004216019529849291, + 0.004234204534441233, + 0.004205711185932159, + 0.007013634312897921, + 0.01103358343243599, + 0.005498194135725498, + 0.006256171502172947, + 0.003483421402052045, + 0.0054954770021140575, + 0.005140689667314291, + 0.004861272405833006, + 0.003761769738048315, + 0.005382629111409187, + 0.005492611788213253, + 0.016046851873397827, + 0.022762464359402657, + 0.024004671722650528, + 0.02170214243233204, + 0.018301982432603836, + 0.02849714457988739, + 0.02814340405166149, + 0.022040823474526405, + 0.016949407756328583, + 0.008456706069409847, + 0.02324630320072174, + 0.01608566753566265, + 0.014775650575757027, + 0.017028631642460823, + 0.03714239224791527, + 0.03159216791391373, + 0.010573714040219784, + 0.007299876306205988, + 0.013035775162279606, + 0.023636627942323685, + 0.009582608938217163, + 0.011504736728966236, + 0.01880273036658764, + 0.01275987084954977, + 0.005012615118175745, + 0.002888469723984599, + 0.00501517066732049, + 0.002577058505266905, + 0.003265051171183586, + 0.0038888277485966682, + 0.005582399200648069, + 0.006982346531003714, + 0.031146446242928505, + 0.028426039963960648, + 0.05419421195983887, + 0.0496244914829731, + 0.026753250509500504, + 0.03564722463488579, + 0.07827888429164886, + 0.048819638788700104, + 0.0030165391508489847, + 0.0018211941933259368, + 0.0031795098911970854, + 0.0025495230220258236, + 0.0022713742218911648, + 0.002311163581907749, + 0.0030256989412009716, + 0.004443186800926924, + 0.0070109195075929165, + 0.0025489777326583862, + 0.009229172952473164, + 0.0035402567591518164, + 0.003324866760522127, + 0.0036193695850670338, + 0.00681170541793108, + 0.005629267077893019, + 0.0012279656948521733, + 0.0017281895270571113, + 0.001668868470005691, + 0.0014622187009081244, + 0.0009482099558226764, + 0.0012111240066587925, + 0.0010146816493943334, + 0.0006691272719763219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.38722938299179077 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208933338522911 + ], + "opposite_logprob_tokens": [ + -4.264208793640137 + ], + "logprob_reward": -0.014208933338522911, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208933338522911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07450911402702332, + 0.04114573821425438, + 0.13002122938632965, + 0.19999246299266815, + 0.07533948868513107, + 0.029585404321551323, + 0.30167871713638306, + 0.06307712197303772, + 0.06756885349750519, + 0.01708187162876129 + ], + "credit_map": [ + 0.006991284899413586, + 0.005488561931997538, + 0.0066437809728085995, + 0.008424283005297184, + 0.007182478904724121, + 0.012074710801243782, + 0.01722424291074276, + 0.010479778051376343, + 0.007056253496557474, + 0.004212322644889355, + 0.005928957834839821, + 0.004937829449772835, + 0.005287067499011755, + 0.003890847321599722, + 0.004435957875102758, + 0.005396503955125809, + 0.011930776759982109, + 0.016179881989955902, + 0.016805488616228104, + 0.017175162211060524, + 0.012036710977554321, + 0.020480573177337646, + 0.020745446905493736, + 0.01466719526797533, + 0.018607309088110924, + 0.01442203763872385, + 0.028112556785345078, + 0.021341973915696144, + 0.018648499622941017, + 0.018543977290391922, + 0.04225565120577812, + 0.03806047514081001, + 0.007282008416950703, + 0.0069140722043812275, + 0.009977241046726704, + 0.014652949757874012, + 0.007756777573376894, + 0.0076980008743703365, + 0.010681653395295143, + 0.010376797057688236, + 0.0035660427529364824, + 0.002322710119187832, + 0.004487686324864626, + 0.0025111366994678974, + 0.0032564520370215178, + 0.003765669884160161, + 0.004751971457153559, + 0.0049237352795898914, + 0.0259538684040308, + 0.024350792169570923, + 0.045720770955085754, + 0.04394462704658508, + 0.022261900827288628, + 0.03097512200474739, + 0.06757066398859024, + 0.04090096056461334, + 0.00793108157813549, + 0.006212917622178793, + 0.008788593113422394, + 0.007660229690372944, + 0.005805689841508865, + 0.006984022445976734, + 0.010207261890172958, + 0.009487328119575977, + 0.0066002910025417805, + 0.004368468653410673, + 0.006954474840313196, + 0.005054039414972067, + 0.007300221361219883, + 0.016642818227410316, + 0.008632201701402664, + 0.012016342021524906, + 0.002390534384176135, + 0.0028248983435332775, + 0.003236893331632018, + 0.002563467714935541, + 0.0015380604891106486, + 0.0018401797860860825, + 0.0015220127534121275, + 0.0011658258736133575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.014208933338522911 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017228773795068264 + ], + "opposite_logprob_tokens": [ + -6.376722812652588 + ], + "logprob_reward": -0.0017228773795068264, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017228773795068264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13300853967666626, + 0.052593715488910675, + 0.054020218551158905, + 0.07115720957517624, + 0.04736189916729927, + 0.08820269256830215, + 0.22222279012203217, + 0.2532307207584381, + 0.04670492187142372, + 0.03149732947349548 + ], + "credit_map": [ + 0.015634233132004738, + 0.014755811542272568, + 0.015764037147164345, + 0.017894664779305458, + 0.015067492611706257, + 0.020972054451704025, + 0.01934877410531044, + 0.013571475632488728, + 0.006196306552737951, + 0.006712558213621378, + 0.006874277722090483, + 0.0081833740696311, + 0.009364288300275803, + 0.004775332286953926, + 0.005015794187784195, + 0.0054717836901545525, + 0.006217069458216429, + 0.01007303036749363, + 0.006761417258530855, + 0.006244598422199488, + 0.005679833237081766, + 0.006734330207109451, + 0.006614175625145435, + 0.005695766303688288, + 0.007151113357394934, + 0.007984446361660957, + 0.009234284982085228, + 0.006631137803196907, + 0.007146306801587343, + 0.006345612928271294, + 0.013029176741838455, + 0.013635126873850822, + 0.005783469416201115, + 0.003058897564187646, + 0.006502059753984213, + 0.008109427988529205, + 0.004635498393326998, + 0.005157087929546833, + 0.007344482000917196, + 0.0067709749564528465, + 0.01463864091783762, + 0.006995318923145533, + 0.015013066120445728, + 0.009618650190532207, + 0.008838867768645287, + 0.009309599176049232, + 0.011936802417039871, + 0.011851739138364792, + 0.019932152703404427, + 0.018669843673706055, + 0.033082906156778336, + 0.03195127099752426, + 0.018224403262138367, + 0.023242691531777382, + 0.04798726737499237, + 0.029132254421710968, + 0.04160667210817337, + 0.020558204501867294, + 0.03936280310153961, + 0.02683056704699993, + 0.025732774287462234, + 0.025703750550746918, + 0.03934156522154808, + 0.03409437835216522, + 0.006113206967711449, + 0.003036036854609847, + 0.00627332367002964, + 0.0038577287923544645, + 0.005429851822555065, + 0.010593678802251816, + 0.005418465007096529, + 0.005982629023492336, + 0.004692028276622295, + 0.004090428352355957, + 0.004662694875150919, + 0.004154639318585396, + 0.0029288569930940866, + 0.004199425224214792, + 0.0040159872733056545, + 0.002753268927335739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.0017228773795068264 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures a profoundly intimate moment, tightly focused on the profile of a person's face. The scene is illuminated by a single, warm, soft light source, like a pillowcase lamp, casting gentle shadows and highlighting the delicate texture of their skin. A perfectly formed tear falls onto their cheek, shimmering, its surface reflecting the warm light. The soft, golden muzzle of a dog gently intrudes from the edge of the frame. The camera maintains focus as the dog's soft, pink tongue extends with infinite care, making delicate contact with the person's skin. Carefully and tenderly, the dog licks away the lonely tear with slow, soothing movements, silently offering comfort. After a brief moment, the dog's muzzle rests against the person's cheek, its soft fur a tangible comfort. The shallow depth of field creates an atmosphere of deep tranquility and emotional resonance, filled with unconditional love and quiet support, blurring the background into a warm, fuzzy pattern.", + "seed": 1785, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_12/index_6106_seed_1785_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04839227085467428, + "vlm_logprob": -0.04839227085467428, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.04839227085467428, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005461770575493574 + ], + "opposite_logprob_tokens": [ + -7.130461692810059 + ], + "logprob_reward": -0.005461770575493574, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.005461770575493574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2651541531085968, + 0.03523217514157295, + 0.188792422413826, + 0.10851480066776276, + 0.11422107368707657, + 0.08352179080247879, + 0.06124266982078552, + 0.0348430834710598, + 0.05546605959534645, + 0.053011707961559296 + ], + "credit_map": [ + 0.032536186277866364, + 0.010202587582170963, + 0.10189288854598999, + 0.026909980922937393, + 0.010029567405581474, + 0.055129144340753555, + 0.01133404579013586, + 0.017119770869612694, + 0.005458048079162836, + 0.0036232052370905876, + 0.006147478707134724, + 0.004707419779151678, + 0.0031653984915465117, + 0.0036545104812830687, + 0.004757598973810673, + 0.003718517953529954, + 0.008963107131421566, + 0.009644504636526108, + 0.007850905880331993, + 0.00842328555881977, + 0.006599289830774069, + 0.0658111646771431, + 0.012288072146475315, + 0.06921209394931793, + 0.005856659729033709, + 0.008249812759459019, + 0.008251182734966278, + 0.006856435444205999, + 0.006145521067082882, + 0.032170552760362625, + 0.0076860771514475346, + 0.033298563212156296, + 0.008827440440654755, + 0.004930822644382715, + 0.0073358286172151566, + 0.016276050359010696, + 0.005003426223993301, + 0.03045317903161049, + 0.011104349046945572, + 0.03028997965157032, + 0.007105718366801739, + 0.008474531583487988, + 0.007526678964495659, + 0.010196668095886707, + 0.008076253347098827, + 0.01300740148872137, + 0.014482383616268635, + 0.014652160927653313, + 0.007005313411355019, + 0.008304238319396973, + 0.0069113983772695065, + 0.01023238617926836, + 0.006129010114818811, + 0.008039457723498344, + 0.0080445921048522, + 0.006576272193342447, + 0.0030984911136329174, + 0.0032325750216841698, + 0.004610675852745771, + 0.004653037525713444, + 0.003902970114722848, + 0.0051922062411904335, + 0.005232058931142092, + 0.0049210721626877785, + 0.00622927350923419, + 0.003815671196207404, + 0.007235165219753981, + 0.007715174462646246, + 0.004314271733164787, + 0.0069681694731116295, + 0.01118821743875742, + 0.008000113070011139, + 0.00755520723760128, + 0.003709458513185382, + 0.00926094688475132, + 0.008753571659326553, + 0.004905936773866415, + 0.005326707381755114, + 0.008411699905991554, + 0.005088180303573608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.005461770575493574 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048712801188230515 + ], + "opposite_logprob_tokens": [ + -3.048712730407715 + ], + "logprob_reward": -0.048712801188230515, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048712801188230515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20850718021392822, + 0.11484236270189285, + 0.035313863307237625, + 0.09576994180679321, + 0.1467178463935852, + 0.0733456239104271, + 0.02229408733546734, + 0.027575872838497162, + 0.11419839411973953, + 0.1614348441362381 + ], + "credit_map": [ + 0.024388061836361885, + 0.009030092507600784, + 0.08332881331443787, + 0.02166728302836418, + 0.009523704648017883, + 0.026864491403102875, + 0.009523428045213223, + 0.0241813026368618, + 0.006710892543196678, + 0.004020092077553272, + 0.008515785448253155, + 0.007300146389752626, + 0.004336450714617968, + 0.03856378793716431, + 0.005650672130286694, + 0.0397445410490036, + 0.00385632854886353, + 0.004328719340264797, + 0.004758009221404791, + 0.007050930522382259, + 0.0025117830373346806, + 0.0036521514412015676, + 0.005447853356599808, + 0.0037080864422023296, + 0.0035489387810230255, + 0.0048109968192875385, + 0.0027868892066180706, + 0.00360321463085711, + 0.0038921902887523174, + 0.03554321825504303, + 0.005017289891839027, + 0.03656720742583275, + 0.013809927739202976, + 0.015281395986676216, + 0.010392310097813606, + 0.019521629437804222, + 0.019209543243050575, + 0.021800575777888298, + 0.027732228860259056, + 0.018970230594277382, + 0.00749473599717021, + 0.005097902845591307, + 0.0100626889616251, + 0.020628631114959717, + 0.004962394014000893, + 0.010110857896506786, + 0.006559201516211033, + 0.008429217152297497, + 0.001961104106158018, + 0.0022583717945963144, + 0.002181495539844036, + 0.0035659538116306067, + 0.0016101151704788208, + 0.0037055187858641148, + 0.003416376654058695, + 0.0035951510071754456, + 0.003012432949617505, + 0.0024489513598382473, + 0.0036245291121304035, + 0.007154598832130432, + 0.0024754623882472515, + 0.003486044006422162, + 0.002704028273001313, + 0.0026698263827711344, + 0.01753677986562252, + 0.013269385322928429, + 0.01988169364631176, + 0.0150342071428895, + 0.01135939545929432, + 0.011966800317168236, + 0.013562814332544804, + 0.01158731710165739, + 0.018574003130197525, + 0.012930830009281635, + 0.03842544928193092, + 0.026790672913193703, + 0.014476366341114044, + 0.01495488453656435, + 0.02059643715620041, + 0.01468619890511036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.048712801188230515 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17999251186847687 + ], + "opposite_logprob_tokens": [ + -1.8049925565719604 + ], + "logprob_reward": -0.17999251186847687, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17999251186847687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06505975872278214, + 0.13085117936134338, + 0.3462044596672058, + 0.13624955713748932, + 0.1432706117630005, + 0.024839460849761963, + 0.06344535201787949, + 0.01566627249121666, + 0.06549805402755737, + 0.008915337733924389 + ], + "credit_map": [ + 0.007096452638506889, + 0.0043816897086799145, + 0.01743168756365776, + 0.005264636129140854, + 0.0038984434213489294, + 0.006226621568202972, + 0.005103301256895065, + 0.015656927600502968, + 0.006039600353688002, + 0.00366308749653399, + 0.006316520739346743, + 0.004454372450709343, + 0.004904743749648333, + 0.0498785562813282, + 0.005421592388302088, + 0.050172705203294754, + 0.018284635618329048, + 0.0157497338950634, + 0.01835990697145462, + 0.03465856984257698, + 0.010916014201939106, + 0.10981890559196472, + 0.019259950146079063, + 0.11915672570466995, + 0.007182870525866747, + 0.004422140307724476, + 0.00837758183479309, + 0.010373190976679325, + 0.005116346292197704, + 0.04607456922531128, + 0.006170535460114479, + 0.04853232577443123, + 0.013293090276420116, + 0.014344301074743271, + 0.011246511712670326, + 0.02086837962269783, + 0.01559516042470932, + 0.02495657280087471, + 0.022284893319010735, + 0.02068168856203556, + 0.002540747867897153, + 0.0032693054527044296, + 0.002388473367318511, + 0.00580265186727047, + 0.0017225509509444237, + 0.003338033799082041, + 0.0031751315109431744, + 0.0026025648694485426, + 0.007155857514590025, + 0.007604338228702545, + 0.009100086987018585, + 0.010771828703582287, + 0.0049665155820548534, + 0.008627514354884624, + 0.005445257294923067, + 0.009773951023817062, + 0.0014049876481294632, + 0.0017721160547807813, + 0.0011759651824831963, + 0.002142044249922037, + 0.0016390628879889846, + 0.002162326592952013, + 0.0035084704868495464, + 0.001861300435848534, + 0.006553817540407181, + 0.003679480869323015, + 0.010199466720223427, + 0.00901862233877182, + 0.0051625678315758705, + 0.00839195679873228, + 0.0129243778064847, + 0.009567763656377792, + 0.0014665366616100073, + 0.0007292091613635421, + 0.0010523836826905608, + 0.0013815171550959349, + 0.0008488523890264332, + 0.0010050761047750711, + 0.0015363275306299329, + 0.0008954353979788721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.17999251186847687 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746016908437014 + ], + "opposite_logprob_tokens": [ + -5.0067458152771 + ], + "logprob_reward": -0.006746016908437014, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746016908437014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09826195985078812, + 0.11840866506099701, + 0.19160325825214386, + 0.23704229295253754, + 0.15243345499038696, + 0.08554385602474213, + 0.03491255268454552, + 0.02925129048526287, + 0.03180304914712906, + 0.020739590749144554 + ], + "credit_map": [ + 0.011490249074995518, + 0.006733681075274944, + 0.02368222177028656, + 0.015502536669373512, + 0.005097392946481705, + 0.01760546676814556, + 0.007654219400137663, + 0.01049620658159256, + 0.009297197684645653, + 0.005738009233027697, + 0.008602602407336235, + 0.008831534534692764, + 0.004620400723069906, + 0.03530534729361534, + 0.00778057798743248, + 0.038232993334531784, + 0.015014148317277431, + 0.01853148825466633, + 0.017832474783062935, + 0.031548310071229935, + 0.008934316225349903, + 0.035956233739852905, + 0.021819068118929863, + 0.0419672429561615, + 0.008540516719222069, + 0.0072377887554466724, + 0.010516563430428505, + 0.013645526953041553, + 0.006422674283385277, + 0.08770475536584854, + 0.010893771424889565, + 0.09208071231842041, + 0.012873792089521885, + 0.012236028909683228, + 0.009668122977018356, + 0.02307196334004402, + 0.016377855092287064, + 0.025397054851055145, + 0.02769581973552704, + 0.02511281706392765, + 0.007927698083221912, + 0.006908804643899202, + 0.010624686256051064, + 0.024464191868901253, + 0.005162502638995647, + 0.011811585165560246, + 0.010099085979163647, + 0.00854531116783619, + 0.003188610542565584, + 0.0030000233091413975, + 0.003421030705794692, + 0.006869171746075153, + 0.004101657308638096, + 0.004710260778665543, + 0.004811486694961786, + 0.004810316953808069, + 0.003608373925089836, + 0.003350191516801715, + 0.003599656280130148, + 0.0055697024799883366, + 0.0034810220822691917, + 0.003169106086716056, + 0.00386128225363791, + 0.0026119579561054707, + 0.004008255898952484, + 0.0022531100548803806, + 0.00623988127335906, + 0.004797494504600763, + 0.0020724197383970022, + 0.00345528032630682, + 0.00510363420471549, + 0.0038729747757315636, + 0.0026770061813294888, + 0.0011688296217471361, + 0.004197791684418917, + 0.0033594449050724506, + 0.001705923117697239, + 0.0017366621177643538, + 0.003571384586393833, + 0.0023225501645356417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 4 + ], + "selected_reward": -0.006746016908437014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up focused on the profile of a person's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene illuminated by a single warm soft light source like a pillowcase lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a perfectly formed tear fall onto the person's cheek and shimmer?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a dog with a soft golden muzzle extend its pink tongue to lick the tear from the person's cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background blurred into a warm fuzzy pattern due to shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010482537327334285 + ], + "opposite_logprob_tokens": [ + -6.8760480880737305 + ], + "logprob_reward": -0.0010482537327334285, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010482537327334285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18664784729480743, + 0.4411613643169403, + 0.0756053552031517, + 0.05197308585047722, + 0.09104122966527939, + 0.01187955867499113, + 0.024780355393886566, + 0.02179960533976555, + 0.034709349274635315, + 0.06040218099951744 + ], + "credit_map": [ + 0.02331511490046978, + 0.006483468227088451, + 0.07867828011512756, + 0.016555286943912506, + 0.007196648512035608, + 0.02708527445793152, + 0.007517020683735609, + 0.019816765561699867, + 0.031113717705011368, + 0.017120977863669395, + 0.03249938786029816, + 0.028350599110126495, + 0.014947463758289814, + 0.14380910992622375, + 0.019709179177880287, + 0.1536109894514084, + 0.004043715540319681, + 0.0022726207971572876, + 0.006276554893702269, + 0.0067839003168046474, + 0.002106074243783951, + 0.024489687755703926, + 0.003582748118788004, + 0.026050051674246788, + 0.0040511018596589565, + 0.007381292525678873, + 0.003562517464160919, + 0.00551901338621974, + 0.005300787277519703, + 0.009020217694342136, + 0.007928300648927689, + 0.009209861047565937, + 0.00816070381551981, + 0.008582532405853271, + 0.008316533640027046, + 0.011347122490406036, + 0.009450577199459076, + 0.016989463940262794, + 0.013244890607893467, + 0.014949418604373932, + 0.0015794601058587432, + 0.0009907211642712355, + 0.001731300144456327, + 0.0019209134625270963, + 0.0008338512852787971, + 0.0016849657986313105, + 0.0017341545317322016, + 0.0014041936956346035, + 0.0026955148205161095, + 0.0030158692970871925, + 0.002382447477430105, + 0.003927233163267374, + 0.002607493894174695, + 0.00308545446023345, + 0.003697093343362212, + 0.003369250101968646, + 0.0017741679912433028, + 0.00211285799741745, + 0.001916083274409175, + 0.0026226360350847244, + 0.0025949322152882814, + 0.0033656705636531115, + 0.0040590898133814335, + 0.003354169661179185, + 0.005189184099435806, + 0.0032419227063655853, + 0.004844607785344124, + 0.004415265284478664, + 0.0030480078421533108, + 0.004304428119212389, + 0.0052291518077254295, + 0.004436783958226442, + 0.009451616555452347, + 0.003698802087455988, + 0.012257061898708344, + 0.01065012440085411, + 0.004543585702776909, + 0.006921801716089249, + 0.005928793456405401, + 0.0069503989070653915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0010482537327334285 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_13/index_tensor([1654, 1654])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_13/index_tensor([1654, 1654])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8112c911d17a74c72f51204fa39a21bf4c1ae815 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_13/index_tensor([1654, 1654])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme macro shot, taken from above and slightly tilted, focuses on an ancient, circular cork mat resting on the dark, polished surface of a wooden table. The coaster's surface is an ancient landscape: pitted, dry, and faded to a dusty tan, its edges frayed, revealing a network of fine, pre-existing stress lines. A warm light source from the side casts long, dramatic shadows across the coaster's porous texture and illuminates the rich, deep grain of the wood beneath, creating a focused and intimate atmosphere. Triggered by subtle, imperceptible vibrations, a new, hairline fracture suddenly appears near the center, then inexorably spreads outward in a jagged, branching pattern. With a delicate, almost silent finale, the structural integrity collapses; the coaster begins to crumble, first shedding tiny particles of cork like dust, then as larger, more brittle fragments separate and collapse onto the table, exposing the lighter, more granular material within and leaving behind a small, chaotic pile of debris.", + "seed": 2809, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_13/index_1654_seed_2809_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.059651678800582884, + "vlm_logprob": -0.059651678800582884, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.059651678800582884, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07174745202064514 + ], + "opposite_logprob_tokens": [ + -3.0717475414276123 + ], + "logprob_reward": -0.07174745202064514, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.07174745202064514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41650310158729553, + 0.11224432289600372, + 0.05032825097441673, + 0.06702720373868942, + 0.024395441636443138, + 0.0526350736618042, + 0.04098999500274658, + 0.0530591681599617, + 0.12714765965938568, + 0.05566973239183426 + ], + "credit_map": [ + 0.05601263791322708, + 0.07576034218072891, + 0.026638507843017578, + 0.035137809813022614, + 0.0516674630343914, + 0.0707731544971466, + 0.05201113224029541, + 0.04850209131836891, + 0.01372356154024601, + 0.02367560565471649, + 0.008250518701970577, + 0.011292275972664356, + 0.01456599123775959, + 0.01650768518447876, + 0.010826938785612583, + 0.013401749543845654, + 0.007595989387482405, + 0.008478600531816483, + 0.003571439068764448, + 0.0040807113982737064, + 0.005665353499352932, + 0.009993511252105236, + 0.005208384245634079, + 0.00573426578193903, + 0.0076587204821407795, + 0.008471333421766758, + 0.003720765234902501, + 0.0038507592398673296, + 0.008952466771006584, + 0.017940249294042587, + 0.007135143503546715, + 0.009297769516706467, + 0.003010349115356803, + 0.004389834590256214, + 0.0017899201484397054, + 0.0018988600932061672, + 0.004176237620413303, + 0.004221773240715265, + 0.002178657567128539, + 0.0027298100758343935, + 0.008612003177404404, + 0.005695853848010302, + 0.00345215224660933, + 0.004642139654606581, + 0.005748342722654343, + 0.012479813769459724, + 0.0056897797621786594, + 0.006314993370324373, + 0.005402392242103815, + 0.005629973020404577, + 0.002976004034280777, + 0.004004090558737516, + 0.005180267617106438, + 0.007654281333088875, + 0.005518157500773668, + 0.004624831024557352, + 0.006245368625968695, + 0.009281895123422146, + 0.0034122427459806204, + 0.004958096891641617, + 0.006105297245085239, + 0.011270766146481037, + 0.005343115422874689, + 0.00644238805398345, + 0.017442427575588226, + 0.01481543481349945, + 0.011517943814396858, + 0.015693511813879013, + 0.012386461719870567, + 0.02292606793344021, + 0.0184752456843853, + 0.013890587724745274, + 0.008915675804018974, + 0.007639140821993351, + 0.004354630131274462, + 0.004985857289284468, + 0.004873176105320454, + 0.01394384540617466, + 0.0038742434699088335, + 0.007083169650286436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.07174745202064514 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076733119785785675 + ], + "opposite_logprob_tokens": [ + -4.882673263549805 + ], + "logprob_reward": -0.0076733119785785675, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076733119785785675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.243104487657547, + 0.09231185913085938, + 0.06857425719499588, + 0.09161102026700974, + 0.10822542011737823, + 0.1237049400806427, + 0.05082245543599129, + 0.0911160260438919, + 0.05065811797976494, + 0.07987137138843536 + ], + "credit_map": [ + 0.026879772543907166, + 0.042225293815135956, + 0.015797177329659462, + 0.01821666955947876, + 0.032866138964891434, + 0.04835107922554016, + 0.029592949897050858, + 0.029175398871302605, + 0.010495134629309177, + 0.018469450995326042, + 0.005946140270680189, + 0.007720610126852989, + 0.013489321805536747, + 0.017753619700670242, + 0.007923326455056667, + 0.010514256544411182, + 0.008114966563880444, + 0.011584039777517319, + 0.004419233649969101, + 0.005757136736065149, + 0.008872868493199348, + 0.01557070855051279, + 0.005424998234957457, + 0.008830303326249123, + 0.01255600992590189, + 0.01017281785607338, + 0.006888235919177532, + 0.0059539321810007095, + 0.012739080004394054, + 0.0209590345621109, + 0.009691358543932438, + 0.012650558724999428, + 0.012776296585798264, + 0.008726302534341812, + 0.008734308183193207, + 0.007683495059609413, + 0.013355704955756664, + 0.023178694769740105, + 0.021196000277996063, + 0.01257462240755558, + 0.014540334232151508, + 0.01523454487323761, + 0.010334823280572891, + 0.009886665269732475, + 0.018716685473918915, + 0.02402876317501068, + 0.015403695404529572, + 0.015559432096779346, + 0.006944010965526104, + 0.005818664561957121, + 0.004205815959721804, + 0.006329387426376343, + 0.00530361570417881, + 0.011661597527563572, + 0.004419220611453056, + 0.006140146870166063, + 0.011813363060355186, + 0.012361159548163414, + 0.008286588825285435, + 0.012333506718277931, + 0.00840755645185709, + 0.01819946989417076, + 0.008731521666049957, + 0.01098286360502243, + 0.005309807136654854, + 0.008811825886368752, + 0.004415343049913645, + 0.0043950434774160385, + 0.005284242331981659, + 0.009270064532756805, + 0.006950695998966694, + 0.006221091840416193, + 0.010511422529816628, + 0.015330839902162552, + 0.0062876977026462555, + 0.007657337002456188, + 0.009670136496424675, + 0.014677037484943867, + 0.0073408824391663074, + 0.00839601457118988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.0076733119785785675 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009807731956243515 + ], + "opposite_logprob_tokens": [ + -4.634807586669922 + ], + "logprob_reward": -0.009807731956243515, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009807731956243515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1544884890317917, + 0.09295371174812317, + 0.06487320363521576, + 0.06477121263742447, + 0.09504181146621704, + 0.0855671614408493, + 0.06104214861989021, + 0.18332092463970184, + 0.07039158791303635, + 0.12754976749420166 + ], + "credit_map": [ + 0.01912444643676281, + 0.019596103578805923, + 0.010517636314034462, + 0.008290380239486694, + 0.015687651932239532, + 0.041546471416950226, + 0.020958539098501205, + 0.018767250701785088, + 0.008542530238628387, + 0.01997455023229122, + 0.007413865067064762, + 0.009703394025564194, + 0.014210994355380535, + 0.015089855529367924, + 0.00688189035281539, + 0.011136624030768871, + 0.007841480895876884, + 0.009210404008626938, + 0.004340298939496279, + 0.00587780587375164, + 0.008626949973404408, + 0.01623907871544361, + 0.005551455542445183, + 0.007185733411461115, + 0.006830113008618355, + 0.00921218004077673, + 0.005346228368580341, + 0.006404554937034845, + 0.008456519804894924, + 0.014944039285182953, + 0.005588810425251722, + 0.007988766767084599, + 0.012431084178388119, + 0.015339304693043232, + 0.0049566300585865974, + 0.007245592772960663, + 0.01640879549086094, + 0.01901468262076378, + 0.008582239039242268, + 0.01106348354369402, + 0.009572128765285015, + 0.014018314890563488, + 0.00653798459097743, + 0.007600519806146622, + 0.016343601047992706, + 0.014619278721511364, + 0.007776389364153147, + 0.009098946116864681, + 0.006242149043828249, + 0.00994175672531128, + 0.005249363370239735, + 0.007377737667411566, + 0.009137002751231194, + 0.009924965910613537, + 0.0058151232078671455, + 0.0073540532030165195, + 0.02634524367749691, + 0.029194150120019913, + 0.016396887600421906, + 0.02155468985438347, + 0.01986439898610115, + 0.029943712055683136, + 0.01807692088186741, + 0.021944917738437653, + 0.008424791507422924, + 0.010156490840017796, + 0.006517982576042414, + 0.012317276559770107, + 0.008305472321808338, + 0.009057339280843735, + 0.008638001047074795, + 0.006974236108362675, + 0.020307976752519608, + 0.023023750633001328, + 0.01090088952332735, + 0.01271812617778778, + 0.013936339877545834, + 0.02143244259059429, + 0.012773649767041206, + 0.012456588447093964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.009807731956243515 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865773022174835 + ], + "opposite_logprob_tokens": [ + -3.0486576557159424 + ], + "logprob_reward": -0.04865773022174835, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865773022174835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0689268410205841, + 0.0672987699508667, + 0.08398313820362091, + 0.07780724763870239, + 0.0764586329460144, + 0.09667591005563736, + 0.08681017905473709, + 0.26004454493522644, + 0.029934315010905266, + 0.15206046402454376 + ], + "credit_map": [ + 0.008930733427405357, + 0.013738133944571018, + 0.004333869554102421, + 0.004705081228166819, + 0.007952006533741951, + 0.013781801797449589, + 0.007819726131856441, + 0.007665478624403477, + 0.006704217754304409, + 0.012990946881473064, + 0.00544422771781683, + 0.006066207308322191, + 0.007911447435617447, + 0.014089571312069893, + 0.006347379647195339, + 0.00774476258084178, + 0.008393186144530773, + 0.012573124840855598, + 0.005136496387422085, + 0.00792300421744585, + 0.012014488689601421, + 0.02046164683997631, + 0.0060313087888062, + 0.011449873447418213, + 0.007992579601705074, + 0.011803031899034977, + 0.005500095430761576, + 0.008492067456245422, + 0.012624239549040794, + 0.014871308580040932, + 0.006675415206700563, + 0.009848508052527905, + 0.009148531593382359, + 0.013986660167574883, + 0.004153455141931772, + 0.006810015067458153, + 0.011301761493086815, + 0.013960851356387138, + 0.008491597138345242, + 0.008605748414993286, + 0.011014536023139954, + 0.014067534357309341, + 0.007813198491930962, + 0.009559914469718933, + 0.01809880882501602, + 0.01658010110259056, + 0.009362312965095043, + 0.010179492644965649, + 0.008301977068185806, + 0.014857631176710129, + 0.006905173882842064, + 0.010766522027552128, + 0.014296905137598515, + 0.01340114139020443, + 0.008028639480471611, + 0.010252179577946663, + 0.03438692167401314, + 0.05026315897703171, + 0.020620377734303474, + 0.028248414397239685, + 0.03204898536205292, + 0.03934725001454353, + 0.02407142147421837, + 0.031057992950081825, + 0.003694826504215598, + 0.005476685706526041, + 0.002403094433248043, + 0.003965042065829039, + 0.003435293911024928, + 0.003511622780933976, + 0.004204014781862497, + 0.003243733663111925, + 0.0227526668459177, + 0.026278745383024216, + 0.01318923570215702, + 0.014654546976089478, + 0.015627281740307808, + 0.027077680453658104, + 0.01582290045917034, + 0.0166573878377676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.04865773022174835 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037216782569885 + ], + "opposite_logprob_tokens": [ + -1.9103721380233765 + ], + "logprob_reward": -0.16037216782569885, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037216782569885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09904058277606964, + 0.043638963252305984, + 0.09383499622344971, + 0.10962782800197601, + 0.1072327122092247, + 0.13396768271923065, + 0.07904402911663055, + 0.17614023387432098, + 0.04971698671579361, + 0.10775593668222427 + ], + "credit_map": [ + 0.01097376924008131, + 0.018614189699292183, + 0.0072394320741295815, + 0.008258963003754616, + 0.013446890749037266, + 0.01987294852733612, + 0.008533046580851078, + 0.012101340107619762, + 0.004497359972447157, + 0.007715381681919098, + 0.0034189776051789522, + 0.004543475341051817, + 0.005376945249736309, + 0.00884915329515934, + 0.004283975809812546, + 0.00495369266718626, + 0.008772308006882668, + 0.015382696874439716, + 0.0055094328708946705, + 0.008594208396971226, + 0.010978613048791885, + 0.026293165981769562, + 0.00640889722853899, + 0.011895674280822277, + 0.01220623031258583, + 0.016218673437833786, + 0.00731635931879282, + 0.01029589120298624, + 0.016432547941803932, + 0.02482892945408821, + 0.008685748092830181, + 0.013643446378409863, + 0.014390982687473297, + 0.01725822314620018, + 0.006275741383433342, + 0.009077693335711956, + 0.01754133775830269, + 0.019709458574652672, + 0.011402929201722145, + 0.011576350778341293, + 0.014775849878787994, + 0.023742493242025375, + 0.009593411348760128, + 0.011614497750997543, + 0.025909125804901123, + 0.022509926930069923, + 0.01167603675276041, + 0.014146342873573303, + 0.008656596764922142, + 0.010875808075070381, + 0.0074825887568295, + 0.01035262644290924, + 0.011406960897147655, + 0.013812514021992683, + 0.007395757827907801, + 0.009061177261173725, + 0.024997398257255554, + 0.024391042068600655, + 0.017696645110845566, + 0.022606199607253075, + 0.016832726076245308, + 0.03103168122470379, + 0.018135590478777885, + 0.020448952913284302, + 0.006999589968472719, + 0.007571346126496792, + 0.004066860303282738, + 0.006587149575352669, + 0.0052062212489545345, + 0.008652071468532085, + 0.005137105938047171, + 0.005496643483638763, + 0.0178766380995512, + 0.01902981474995613, + 0.008158446289598942, + 0.009863934479653835, + 0.012664367444813251, + 0.01727948896586895, + 0.011136203072965145, + 0.011747042648494244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 3 + ], + "selected_reward": -0.16037216782569885 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot, taken from above and slightly tilted, focuses on an ancient, circular cork mat resting on the dark, polished surface of a wooden table. The coaster's surface is an ancient landscape: pitted, dry, and faded to a dusty tan, its edges frayed, revealing a network of fine, pre-existing stress lines. A warm light source from the side casts long, dramatic shadows across the coaster's porous texture and illuminates the rich, deep grain of the wood beneath, creating a focused and intimate atmosphere. Triggered by subtle, imperceptible vibrations, a new, hairline fracture suddenly appears near the center, then inexorably spreads outward in a jagged, branching pattern. With a delicate, almost silent finale, the structural integrity collapses; the coaster begins to crumble, first shedding tiny particles of cork like dust, then as larger, more brittle fragments separate and collapse onto the table, exposing the lighter, more granular material within and leaving behind a small, chaotic pile of debris.", + "seed": 2810, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_13/index_1654_seed_2810_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.027118058316409586, + "vlm_logprob": -0.027118058316409586, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.027118058316409586, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09247203171253204 + ], + "opposite_logprob_tokens": [ + -2.9674720764160156 + ], + "logprob_reward": -0.09247203171253204, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.09247203171253204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26172196865081787, + 0.04288271442055702, + 0.027093417942523956, + 0.03280516713857651, + 0.03422378748655319, + 0.023632293567061424, + 0.09659029543399811, + 0.19375672936439514, + 0.1588791459798813, + 0.12841443717479706 + ], + "credit_map": [ + 0.0277202520519495, + 0.035128410905599594, + 0.037844352424144745, + 0.036210715770721436, + 0.03741654381155968, + 0.04272773116827011, + 0.027947157621383667, + 0.016726799309253693, + 0.0038845743983983994, + 0.004582779947668314, + 0.006371759343892336, + 0.006870198994874954, + 0.008134766481816769, + 0.006243470124900341, + 0.004216284491121769, + 0.0025788822676986456, + 0.0024077058769762516, + 0.0030698347836732864, + 0.0024171993136405945, + 0.00460424367338419, + 0.004893188830465078, + 0.005256951786577702, + 0.0030241915956139565, + 0.0014201027806848288, + 0.003322635078802705, + 0.0035613635554909706, + 0.004361437633633614, + 0.005339682102203369, + 0.004812083672732115, + 0.006228853017091751, + 0.003141441149637103, + 0.002037670696154237, + 0.002562299370765686, + 0.00519188866019249, + 0.005172016564756632, + 0.005894049536436796, + 0.004900350235402584, + 0.004321984946727753, + 0.00407128082588315, + 0.0021099180448800325, + 0.0022986396215856075, + 0.0023595457896590233, + 0.0041258372366428375, + 0.0038766420911997557, + 0.0031153096351772547, + 0.0036028753966093063, + 0.0024699342902749777, + 0.0017835113685578108, + 0.011198876425623894, + 0.013296574354171753, + 0.013541250489652157, + 0.009858274832367897, + 0.010575477033853531, + 0.013929898850619793, + 0.007623099721968174, + 0.016566837206482887, + 0.018009930849075317, + 0.023497477173805237, + 0.030604690313339233, + 0.023005403578281403, + 0.03259086236357689, + 0.02447817474603653, + 0.02120635099709034, + 0.020363837480545044, + 0.018084168434143066, + 0.016726966947317123, + 0.02505224384367466, + 0.022594615817070007, + 0.017097076401114464, + 0.027110682800412178, + 0.015504068695008755, + 0.016709327697753906, + 0.014693065546452999, + 0.015316922217607498, + 0.022571569308638573, + 0.019944103434681892, + 0.01204713899642229, + 0.021201714873313904, + 0.008533955551683903, + 0.014105974696576595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.09247203171253204 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014263346791267395 + ], + "opposite_logprob_tokens": [ + -4.264263153076172 + ], + "logprob_reward": -0.014263346791267395, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014263346791267395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12827277183532715, + 0.020578132942318916, + 0.02576065994799137, + 0.017049362882971764, + 0.027176043018698692, + 0.027861544862389565, + 0.09966076910495758, + 0.13146789371967316, + 0.2368084192276001, + 0.28536441922187805 + ], + "credit_map": [ + 0.011154201813042164, + 0.01318206824362278, + 0.018541106954216957, + 0.023391757160425186, + 0.019205918535590172, + 0.01859048753976822, + 0.017844051122665405, + 0.006363171152770519, + 0.0016454228898510337, + 0.0022663644049316645, + 0.0026733786799013615, + 0.0037624756805598736, + 0.003491471754387021, + 0.002757082926109433, + 0.002473928965628147, + 0.00150800752453506, + 0.002139236079528928, + 0.0037745104636996984, + 0.003566477680578828, + 0.0051680151373147964, + 0.003965535666793585, + 0.0029431588482111692, + 0.002796596148982644, + 0.001407130155712366, + 0.0018037565751001239, + 0.0023219455033540726, + 0.0023373193107545376, + 0.0016710014315322042, + 0.0030660468619316816, + 0.003126006806269288, + 0.0013626273721456528, + 0.0013606584398075938, + 0.0021048591006547213, + 0.003675294341519475, + 0.00319368252530694, + 0.005872864276170731, + 0.004407069645822048, + 0.003484374610707164, + 0.0027913162484765053, + 0.0016465820372104645, + 0.002649945905432105, + 0.002271810779348016, + 0.003605770878493786, + 0.004294079728424549, + 0.003227588487789035, + 0.0050203693099319935, + 0.004806801676750183, + 0.00198517763055861, + 0.009312584064900875, + 0.016125915572047234, + 0.012740956619381905, + 0.010234532877802849, + 0.011859792284667492, + 0.014633768238127232, + 0.007876915857195854, + 0.016876302659511566, + 0.012554975226521492, + 0.014921118505299091, + 0.020139113068580627, + 0.01637776754796505, + 0.02288614585995674, + 0.015892812982201576, + 0.015803145244717598, + 0.012892809696495533, + 0.028249094262719154, + 0.024280479177832603, + 0.03298352286219597, + 0.03629894554615021, + 0.025424087420105934, + 0.03914095461368561, + 0.022097952663898468, + 0.028333377093076706, + 0.03499419614672661, + 0.031848639249801636, + 0.04059119150042534, + 0.03609072044491768, + 0.026273373514413834, + 0.06179382652044296, + 0.015998465940356255, + 0.03777402639389038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.014263346791267395 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007657223381102085 + ], + "opposite_logprob_tokens": [ + -4.882657051086426 + ], + "logprob_reward": -0.007657223381102085, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007657223381102085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11729317158460617, + 0.061697348952293396, + 0.029992420226335526, + 0.03090069442987442, + 0.03852711245417595, + 0.0556519478559494, + 0.08969949185848236, + 0.16935569047927856, + 0.14181384444236755, + 0.26506826281547546 + ], + "credit_map": [ + 0.01086563803255558, + 0.016722261905670166, + 0.019209137186408043, + 0.016494641080498695, + 0.018657242879271507, + 0.015200842171907425, + 0.01226679515093565, + 0.007876617833971977, + 0.004565054085105658, + 0.008185619488358498, + 0.006630266085267067, + 0.017765101045370102, + 0.006860213354229927, + 0.007356217596679926, + 0.007522636093199253, + 0.002812240505591035, + 0.0024945009499788284, + 0.0031682513654232025, + 0.003472935175523162, + 0.0058828298933804035, + 0.005349895916879177, + 0.004379612859338522, + 0.0037274446804076433, + 0.0015169491525739431, + 0.0028418570291250944, + 0.00338992802426219, + 0.003455467987805605, + 0.005095352418720722, + 0.005805161315947771, + 0.004738946910947561, + 0.0039055885281413794, + 0.0016683897702023387, + 0.0033031045459210873, + 0.004930190276354551, + 0.005770550575107336, + 0.007378921378403902, + 0.006691794842481613, + 0.004834041465073824, + 0.0037553024012595415, + 0.0018632084829732776, + 0.005796294659376144, + 0.006648007780313492, + 0.009563952684402466, + 0.009297613985836506, + 0.008074886165559292, + 0.006890569813549519, + 0.0052178422920405865, + 0.00416278000921011, + 0.008513555862009525, + 0.013102618046104908, + 0.010935669764876366, + 0.011313899420201778, + 0.011155746877193451, + 0.013565503992140293, + 0.008429001085460186, + 0.012683500535786152, + 0.015953468158841133, + 0.01892533153295517, + 0.025764597579836845, + 0.02137896977365017, + 0.028530962765216827, + 0.021211739629507065, + 0.019927185028791428, + 0.017663441598415375, + 0.01582510955631733, + 0.015084759332239628, + 0.021623758599162102, + 0.017900729551911354, + 0.01629733294248581, + 0.026087390258908272, + 0.013981876894831657, + 0.015012887306511402, + 0.03333793208003044, + 0.029644053429365158, + 0.03527824580669403, + 0.029778923839330673, + 0.02656356804072857, + 0.060424137860536575, + 0.014431090094149113, + 0.035610318183898926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.007657223381102085 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012548303231596947 + ], + "opposite_logprob_tokens": [ + -4.387548446655273 + ], + "logprob_reward": -0.012548303231596947, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012548303231596947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11163399368524551, + 0.037419598549604416, + 0.016608161851763725, + 0.0130162313580513, + 0.022027404978871346, + 0.03007384017109871, + 0.05173388868570328, + 0.11186838150024414, + 0.32861199975013733, + 0.2770063877105713 + ], + "credit_map": [ + 0.007830402813851833, + 0.010541134513914585, + 0.010635079815983772, + 0.02583911642432213, + 0.016054468229413033, + 0.015737488865852356, + 0.018869487568736076, + 0.006126821041107178, + 0.0039805760607123375, + 0.005457030143588781, + 0.004648301750421524, + 0.005811089649796486, + 0.005018420051783323, + 0.006164395250380039, + 0.003557107876986265, + 0.0027826810255646706, + 0.0011232405668124557, + 0.0016881892224773765, + 0.0018273000605404377, + 0.0036602644249796867, + 0.0021234082523733377, + 0.002662386279553175, + 0.0023337658494710922, + 0.0011896088253706694, + 0.0009798002429306507, + 0.001362362876534462, + 0.0012418165570124984, + 0.0019820183515548706, + 0.002054600277915597, + 0.0025086139794439077, + 0.0017914575291797519, + 0.0010955629404634237, + 0.0019172308966517448, + 0.0032758484594523907, + 0.0026829622220247984, + 0.0034980305936187506, + 0.0037538460455834866, + 0.0030290139839053154, + 0.002170752501115203, + 0.0016997229540720582, + 0.002490281593054533, + 0.002764966804534197, + 0.003266961546614766, + 0.005080574657768011, + 0.003800679463893175, + 0.004220493137836456, + 0.006133330520242453, + 0.0023165561724454165, + 0.0049215941689908504, + 0.00622839480638504, + 0.008203172124922276, + 0.008054181933403015, + 0.006266728509217501, + 0.006811941042542458, + 0.006467238534241915, + 0.004780645482242107, + 0.010148867964744568, + 0.018380679190158844, + 0.013193301856517792, + 0.011472652666270733, + 0.01428502332419157, + 0.020240671932697296, + 0.012738092802464962, + 0.01140910480171442, + 0.037791963666677475, + 0.03324200212955475, + 0.046984393149614334, + 0.04501725360751152, + 0.038451097905635834, + 0.05626079440116882, + 0.03147397190332413, + 0.03939056769013405, + 0.034528184682130814, + 0.030890433117747307, + 0.0403946191072464, + 0.03457222133874893, + 0.02484721504151821, + 0.060714468359947205, + 0.015310671180486679, + 0.03574863448739052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.012548303231596947 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008649386465549469 + ], + "opposite_logprob_tokens": [ + -4.7586493492126465 + ], + "logprob_reward": -0.008649386465549469, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008649386465549469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06253689527511597, + 0.02489382214844227, + 0.018981531262397766, + 0.03224795311689377, + 0.02173362672328949, + 0.03262970969080925, + 0.038083307445049286, + 0.05405787006020546, + 0.1743127405643463, + 0.5405226349830627 + ], + "credit_map": [ + 0.00557295186445117, + 0.007629698608070612, + 0.006241213530302048, + 0.014059984125196934, + 0.01059092115610838, + 0.007395810913294554, + 0.007395272608846426, + 0.003651039209216833, + 0.002793093677610159, + 0.003046478144824505, + 0.0030704380478709936, + 0.003456966020166874, + 0.003661060007289052, + 0.004935304168611765, + 0.002379262587055564, + 0.0015512208919972181, + 0.0019286767346784472, + 0.0023208509664982557, + 0.0027397891972213984, + 0.0026874979957938194, + 0.003213417949154973, + 0.0028904967475682497, + 0.00198151171207428, + 0.0012192904250696301, + 0.0025213619228452444, + 0.0026816357858479023, + 0.0041474816389381886, + 0.0055989776737987995, + 0.004855584353208542, + 0.005521796643733978, + 0.004957596305757761, + 0.001963516231626272, + 0.001949966885149479, + 0.0028507730457931757, + 0.0035636192187666893, + 0.0029462925158441067, + 0.003667996497824788, + 0.002704458311200142, + 0.0026762858033180237, + 0.0013742349110543728, + 0.002783973468467593, + 0.003586649661883712, + 0.004977676551789045, + 0.00579354353249073, + 0.004016377031803131, + 0.004107464104890823, + 0.005046673584729433, + 0.0023173505906015635, + 0.003763169050216675, + 0.005818525794893503, + 0.004386281594634056, + 0.00413212226703763, + 0.004403770435601473, + 0.006360380444675684, + 0.0029432172887027264, + 0.0062758405692875385, + 0.00437278114259243, + 0.005316818132996559, + 0.008812051266431808, + 0.0075012207962572575, + 0.005192694254219532, + 0.009179149754345417, + 0.009095202200114727, + 0.004587953444570303, + 0.020334718748927116, + 0.018038254231214523, + 0.032840557396411896, + 0.02547214925289154, + 0.019286438822746277, + 0.01834709569811821, + 0.01941152848303318, + 0.02058199606835842, + 0.06744223833084106, + 0.0611313097178936, + 0.07877876609563828, + 0.06706808507442474, + 0.047244567424058914, + 0.11709495633840561, + 0.029104191809892654, + 0.0726584866642952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.008649386465549469 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot, taken from above and slightly tilted, focuses on an ancient, circular cork mat resting on the dark, polished surface of a wooden table. The coaster's surface is an ancient landscape: pitted, dry, and faded to a dusty tan, its edges frayed, revealing a network of fine, pre-existing stress lines. A warm light source from the side casts long, dramatic shadows across the coaster's porous texture and illuminates the rich, deep grain of the wood beneath, creating a focused and intimate atmosphere. Triggered by subtle, imperceptible vibrations, a new, hairline fracture suddenly appears near the center, then inexorably spreads outward in a jagged, branching pattern. With a delicate, almost silent finale, the structural integrity collapses; the coaster begins to crumble, first shedding tiny particles of cork like dust, then as larger, more brittle fragments separate and collapse onto the table, exposing the lighter, more granular material within and leaving behind a small, chaotic pile of debris.", + "seed": 2811, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_13/index_1654_seed_2811_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.015177038684487342, + "vlm_logprob": -0.015177038684487342, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.015177038684487342, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02366441860795021 + ], + "opposite_logprob_tokens": [ + -4.523664474487305 + ], + "logprob_reward": -0.02366441860795021, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.02366441860795021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0811072438955307, + 0.05250415578484535, + 0.0924147516489029, + 0.016920140013098717, + 0.03697171062231064, + 0.031675927340984344, + 0.037463922053575516, + 0.5265145301818848, + 0.024589844048023224, + 0.09983772784471512 + ], + "credit_map": [ + 0.004980935249477625, + 0.02798617072403431, + 0.003981183283030987, + 0.006001036148518324, + 0.028236957266926765, + 0.004543545190244913, + 0.0032931219320744276, + 0.002084301318973303, + 0.0029927007853984833, + 0.018703579902648926, + 0.002602833788841963, + 0.002720013726502657, + 0.019308263435959816, + 0.002907684538513422, + 0.0017029097070917487, + 0.0015661735087633133, + 0.009820116683840752, + 0.019808907061815262, + 0.011490992270410061, + 0.016015347093343735, + 0.009807669557631016, + 0.011212675832211971, + 0.00906730629503727, + 0.005191745702177286, + 0.0025334178935736418, + 0.001970042707398534, + 0.0019534765742719173, + 0.0027943772729486227, + 0.00413524592295289, + 0.0013544549001380801, + 0.0010609272867441177, + 0.0011181979207322001, + 0.006951582618057728, + 0.006010564975440502, + 0.003969681449234486, + 0.006445801816880703, + 0.003280755365267396, + 0.003882292192429304, + 0.004174205474555492, + 0.0022568251006305218, + 0.005757246166467667, + 0.004796986468136311, + 0.005641005001962185, + 0.0024785534478724003, + 0.0032869703136384487, + 0.004367894493043423, + 0.002367319306358695, + 0.002979954704642296, + 0.010324445553123951, + 0.002946451073512435, + 0.00868411734700203, + 0.0033360482193529606, + 0.003079602960497141, + 0.005078804679214954, + 0.001976924715563655, + 0.002037529367953539, + 0.16637030243873596, + 0.030856488272547722, + 0.18725447356700897, + 0.023592578247189522, + 0.020045563578605652, + 0.054685503244400024, + 0.014762704260647297, + 0.02894696034491062, + 0.0037990070413798094, + 0.0016614582855254412, + 0.003668309887871146, + 0.0015961744356900454, + 0.002105916617438197, + 0.00893621426075697, + 0.001214352436363697, + 0.001608413876965642, + 0.03164305165410042, + 0.005903004202991724, + 0.003806761233136058, + 0.0035644755698740482, + 0.00719305407255888, + 0.040948864072561264, + 0.003185478737577796, + 0.0035930483136326075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.02366441860795021 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007664558012038469 + ], + "opposite_logprob_tokens": [ + -4.882664680480957 + ], + "logprob_reward": -0.007664558012038469, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007664558012038469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05416185036301613, + 0.10154950618743896, + 0.10581691563129425, + 0.07598070800304413, + 0.044726401567459106, + 0.07282449305057526, + 0.19986538589000702, + 0.15010268986225128, + 0.0580858439207077, + 0.13688617944717407 + ], + "credit_map": [ + 0.004937784746289253, + 0.009975495748221874, + 0.0059333909302949905, + 0.007874124683439732, + 0.008233021013438702, + 0.008174553513526917, + 0.004910367541015148, + 0.004123110324144363, + 0.004749168176203966, + 0.03520428389310837, + 0.00446633156388998, + 0.00898387935012579, + 0.032957594841718674, + 0.008871767669916153, + 0.0038478465285152197, + 0.002468629740178585, + 0.013586996123194695, + 0.01547145750373602, + 0.014516938477754593, + 0.019132675603032112, + 0.010602050460875034, + 0.013180816546082497, + 0.011515950784087181, + 0.007810032460838556, + 0.007105569820851088, + 0.011282001622021198, + 0.010738692246377468, + 0.012863780371844769, + 0.01592567376792431, + 0.009023941121995449, + 0.004466647282242775, + 0.004574407357722521, + 0.006164830643683672, + 0.009283198975026608, + 0.005468099378049374, + 0.004096298012882471, + 0.0076017542742192745, + 0.0053448425605893135, + 0.0037211570888757706, + 0.0030462180729955435, + 0.014353615231812, + 0.011152520775794983, + 0.01338969450443983, + 0.006591565907001495, + 0.007084304932504892, + 0.009015579707920551, + 0.0065241362899541855, + 0.004713067319244146, + 0.08076900243759155, + 0.013227700255811214, + 0.060329996049404144, + 0.010670092888176441, + 0.006855356507003307, + 0.014905085787177086, + 0.005131445825099945, + 0.007976717315614223, + 0.04814530164003372, + 0.008966974914073944, + 0.05294598266482353, + 0.006734697613865137, + 0.005710475146770477, + 0.014653920195996761, + 0.004391955677419901, + 0.008553383871912956, + 0.01504920981824398, + 0.003964571747928858, + 0.004626768175512552, + 0.0036198419984430075, + 0.004422888625413179, + 0.01922944374382496, + 0.0034411936067044735, + 0.0037319243419915438, + 0.04186028987169266, + 0.008316772058606148, + 0.004489675164222717, + 0.0039418418891727924, + 0.010915945284068584, + 0.05926559865474701, + 0.0036701750941574574, + 0.004425886552780867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.007664558012038469 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00221055606380105 + ], + "opposite_logprob_tokens": [ + -6.12721061706543 + ], + "logprob_reward": -0.00221055606380105, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.00221055606380105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09943010658025742, + 0.07473617047071457, + 0.062457289546728134, + 0.11051063984632492, + 0.045228492468595505, + 0.01962227188050747, + 0.3016144931316376, + 0.07066281139850616, + 0.14134426414966583, + 0.07439353317022324 + ], + "credit_map": [ + 0.006845482625067234, + 0.02778851054608822, + 0.008714335970580578, + 0.008822351694107056, + 0.02534552849829197, + 0.006509195547550917, + 0.009444211609661579, + 0.005960479378700256, + 0.0034152385778725147, + 0.02526680938899517, + 0.0033929296769201756, + 0.005957866087555885, + 0.023781578987836838, + 0.0053026131354272366, + 0.004376636352390051, + 0.0032424910459667444, + 0.00542149692773819, + 0.009753991849720478, + 0.0060805645771324635, + 0.012072835117578506, + 0.009915893897414207, + 0.009138588793575764, + 0.006390080787241459, + 0.0036838322412222624, + 0.010825090110301971, + 0.012519275769591331, + 0.018324710428714752, + 0.02343636564910412, + 0.018174724653363228, + 0.011941829696297646, + 0.009496555663645267, + 0.005792074371129274, + 0.01459714025259018, + 0.003945722244679928, + 0.011550135910511017, + 0.005252415779978037, + 0.001907146186567843, + 0.002849821001291275, + 0.003114151768386364, + 0.0020119540859013796, + 0.00272063328884542, + 0.0024691352155059576, + 0.0022670591715723276, + 0.002232593484222889, + 0.001668918295763433, + 0.003806228516623378, + 0.002094516297802329, + 0.0023631853982806206, + 0.12423566728830338, + 0.018557287752628326, + 0.0921953096985817, + 0.015474919229745865, + 0.010864765383303165, + 0.0226436834782362, + 0.007101879455149174, + 0.010540967807173729, + 0.026885325089097023, + 0.004916633479297161, + 0.02214149571955204, + 0.0032606234308332205, + 0.002356903627514839, + 0.0053225732408463955, + 0.0022717807441949844, + 0.003507467685267329, + 0.04300941526889801, + 0.006483388599008322, + 0.009878951124846935, + 0.003630966879427433, + 0.007969372905790806, + 0.05841498076915741, + 0.004283745773136616, + 0.007673410698771477, + 0.021047793328762054, + 0.00481421547010541, + 0.003788937348872423, + 0.001884100609458983, + 0.006372630596160889, + 0.03167621046304703, + 0.002062985673546791, + 0.0027466448955237865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.00221055606380105 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539238668978214 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539238668978214, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539238668978214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12540008127689362, + 0.04493848979473114, + 0.10764177143573761, + 0.08736258000135422, + 0.04817572608590126, + 0.06005827337503433, + 0.2968699634075165, + 0.11059783399105072, + 0.04363415390253067, + 0.07532109320163727 + ], + "credit_map": [ + 0.005733075086027384, + 0.04262585565447807, + 0.005777581594884396, + 0.006265466101467609, + 0.04407449811697006, + 0.012744872830808163, + 0.0050189560279250145, + 0.0031597884371876717, + 0.002282587578520179, + 0.014357393607497215, + 0.0020157182589173317, + 0.0035527804866433144, + 0.01413778681308031, + 0.00453156465664506, + 0.002330393297597766, + 0.0017302683554589748, + 0.011303241364657879, + 0.017038311809301376, + 0.011612212285399437, + 0.017202502116560936, + 0.012684059329330921, + 0.01812111772596836, + 0.012056157924234867, + 0.007624169811606407, + 0.006738937925547361, + 0.01153580006211996, + 0.016244951635599136, + 0.012612201273441315, + 0.02273678034543991, + 0.007789607625454664, + 0.005816596560180187, + 0.003887710627168417, + 0.013237831182777882, + 0.005500017665326595, + 0.012075567618012428, + 0.00495348684489727, + 0.0030136536806821823, + 0.003360319882631302, + 0.00366591545753181, + 0.0023689381778240204, + 0.01852540485560894, + 0.0073381755501031876, + 0.012168873101472855, + 0.005004710052162409, + 0.006097408942878246, + 0.0050196293741464615, + 0.003359697526320815, + 0.002544374903663993, + 0.11829520761966705, + 0.020141350105404854, + 0.0867466852068901, + 0.015110032632946968, + 0.011911685578525066, + 0.024830061942338943, + 0.008427692577242851, + 0.011407238431274891, + 0.02629193477332592, + 0.0069598183035850525, + 0.03761087357997894, + 0.005214255768805742, + 0.005127464886754751, + 0.020130669698119164, + 0.0028847919311374426, + 0.006378031801432371, + 0.011064665392041206, + 0.0028058334719389677, + 0.0035035316832363605, + 0.002538684057071805, + 0.002714870497584343, + 0.0157316904515028, + 0.0027254177257418633, + 0.0025494610890746117, + 0.023538712412118912, + 0.004792592953890562, + 0.002429555170238018, + 0.0022385832853615284, + 0.005191240459680557, + 0.03219664841890335, + 0.0022708165924996138, + 0.002662949962541461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.012539238668978214 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980642206966877 + ], + "opposite_logprob_tokens": [ + -3.52980637550354 + ], + "logprob_reward": -0.02980642206966877, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980642206966877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04990247264504433, + 0.0427035316824913, + 0.05713653936982155, + 0.03901853412389755, + 0.057383958250284195, + 0.020043302327394485, + 0.3524552881717682, + 0.06753093004226685, + 0.07994969934225082, + 0.23387576639652252 + ], + "credit_map": [ + 0.0028904632199555635, + 0.015223140828311443, + 0.003441700479015708, + 0.004010885953903198, + 0.014387443661689758, + 0.0027409191243350506, + 0.003672653576359153, + 0.00353526440449059, + 0.0028410921804606915, + 0.012181621044874191, + 0.00295160966925323, + 0.00537763861939311, + 0.010218299925327301, + 0.00406221067532897, + 0.0028382427990436554, + 0.0022328165359795094, + 0.006936695892363787, + 0.010563050396740437, + 0.006168480031192303, + 0.008776790462434292, + 0.0054008848965168, + 0.007739186752587557, + 0.007494315039366484, + 0.004057129845023155, + 0.003981170244514942, + 0.004957740660756826, + 0.006487031001597643, + 0.00702415918931365, + 0.006716630421578884, + 0.004826719872653484, + 0.0030428224708884954, + 0.001982262125238776, + 0.021266093477606773, + 0.004875565879046917, + 0.01826673187315464, + 0.002226051641628146, + 0.0030789196025580168, + 0.0032876748591661453, + 0.0016846093349158764, + 0.0026983090210705996, + 0.0025793462991714478, + 0.002565921749919653, + 0.00324182934127748, + 0.0032653396483510733, + 0.0020772404968738556, + 0.0030456313397735357, + 0.001674198661930859, + 0.0015937953721731901, + 0.1406940370798111, + 0.022291060537099838, + 0.10861143469810486, + 0.020384933799505234, + 0.012704664841294289, + 0.025126392021775246, + 0.009931333363056183, + 0.01271142903715372, + 0.013271136209368706, + 0.004674208350479603, + 0.021816767752170563, + 0.0031843604519963264, + 0.003996465355157852, + 0.014707625843584538, + 0.0020017027854919434, + 0.0038786630611866713, + 0.02406037598848343, + 0.005104209296405315, + 0.005381749011576176, + 0.0023465559352189302, + 0.004598564468324184, + 0.03138233348727226, + 0.002835356630384922, + 0.0042405566200613976, + 0.07534432411193848, + 0.013768147677183151, + 0.007528423797339201, + 0.006485966965556145, + 0.0159844271838665, + 0.1014479398727417, + 0.006007458083331585, + 0.0073090833611786366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.02980642206966877 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot, taken from above and slightly tilted, focuses on an ancient, circular cork mat resting on the dark, polished surface of a wooden table. The coaster's surface is an ancient landscape: pitted, dry, and faded to a dusty tan, its edges frayed, revealing a network of fine, pre-existing stress lines. A warm light source from the side casts long, dramatic shadows across the coaster's porous texture and illuminates the rich, deep grain of the wood beneath, creating a focused and intimate atmosphere. Triggered by subtle, imperceptible vibrations, a new, hairline fracture suddenly appears near the center, then inexorably spreads outward in a jagged, branching pattern. With a delicate, almost silent finale, the structural integrity collapses; the coaster begins to crumble, first shedding tiny particles of cork like dust, then as larger, more brittle fragments separate and collapse onto the table, exposing the lighter, more granular material within and leaving behind a small, chaotic pile of debris.", + "seed": 2812, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_13/index_1654_seed_2812_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01369408001191914, + "vlm_logprob": -0.01369408001191914, + "vlm_yes_no_margin": 4.799999952316284, + "vlm_token_credit": -0.01369408001191914, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038763679563999176 + ], + "opposite_logprob_tokens": [ + -4.038763523101807 + ], + "logprob_reward": -0.038763679563999176, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.038763679563999176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15295757353305817, + 0.038566190749406815, + 0.024256883189082146, + 0.1475912481546402, + 0.04744129255414009, + 0.13308937847614288, + 0.1689712107181549, + 0.11288107931613922, + 0.12119609117507935, + 0.05304904282093048 + ], + "credit_map": [ + 0.0186611358076334, + 0.036591943353414536, + 0.011314352042973042, + 0.02012128196656704, + 0.020976457744836807, + 0.019970549270510674, + 0.009473893791437149, + 0.015847962349653244, + 0.003886550199240446, + 0.007635992020368576, + 0.0029404035303741693, + 0.004877009894698858, + 0.006879797205328941, + 0.0053488523699343204, + 0.002557234140112996, + 0.004440348129719496, + 0.0027204949874430895, + 0.005960679613053799, + 0.001456894795410335, + 0.0023050035815685987, + 0.0033581191673874855, + 0.0034104716032743454, + 0.003277685260400176, + 0.0017675335984677076, + 0.01285592932254076, + 0.032361023128032684, + 0.010587927885353565, + 0.01687490940093994, + 0.02988981455564499, + 0.019696684554219246, + 0.010701549239456654, + 0.014623398892581463, + 0.005841477774083614, + 0.00646649394184351, + 0.0035876266192644835, + 0.004428259562700987, + 0.006629199720919132, + 0.009983940050005913, + 0.005844041705131531, + 0.004660254344344139, + 0.011386673897504807, + 0.014335032552480698, + 0.00899036880582571, + 0.014257659204304218, + 0.018159424886107445, + 0.0339449904859066, + 0.02040298841893673, + 0.011612238362431526, + 0.01655367948114872, + 0.016978146508336067, + 0.016312167048454285, + 0.014936957508325577, + 0.021326610818505287, + 0.0316283144056797, + 0.03005164861679077, + 0.021183671429753304, + 0.012219130992889404, + 0.019207265228033066, + 0.023251108825206757, + 0.010500296019017696, + 0.011284041218459606, + 0.01671338453888893, + 0.009517326951026917, + 0.01018852461129427, + 0.011961414478719234, + 0.015777738764882088, + 0.00884830392897129, + 0.010182374157011509, + 0.031445588916540146, + 0.015122080221772194, + 0.010759150609374046, + 0.01709943637251854, + 0.005738809704780579, + 0.005852728616446257, + 0.004201135132461786, + 0.00506018428131938, + 0.012129444628953934, + 0.008498028852045536, + 0.004268235061317682, + 0.00730047095566988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.038763679563999176 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00678153894841671 + ], + "opposite_logprob_tokens": [ + -5.006781578063965 + ], + "logprob_reward": -0.00678153894841671, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00678153894841671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11832230538129807, + 0.03598400950431824, + 0.04793855920433998, + 0.04265974462032318, + 0.10365885496139526, + 0.07656776905059814, + 0.07875613123178482, + 0.3252792954444885, + 0.11546410620212555, + 0.05536920949816704 + ], + "credit_map": [ + 0.014255725778639317, + 0.027879929170012474, + 0.008100599981844425, + 0.018095267936587334, + 0.016433965414762497, + 0.01647125370800495, + 0.008653409779071808, + 0.008432159200310707, + 0.002974498551338911, + 0.008191872388124466, + 0.002130748238414526, + 0.004550114739686251, + 0.008044173941016197, + 0.005323863122612238, + 0.0027444548904895782, + 0.0020242836326360703, + 0.005160005763173103, + 0.012139949947595596, + 0.004355832003057003, + 0.004896144382655621, + 0.007194939069449902, + 0.006010987795889378, + 0.003668525256216526, + 0.0045121763832867146, + 0.005140712950378656, + 0.00829313974827528, + 0.003642686875537038, + 0.004874135833233595, + 0.006013897247612476, + 0.006545263808220625, + 0.004375489894300699, + 0.003774420591071248, + 0.007278279401361942, + 0.017488572746515274, + 0.005193398334085941, + 0.010048463009297848, + 0.022291630506515503, + 0.026832064613699913, + 0.007581264246255159, + 0.0069451769813895226, + 0.0071666063740849495, + 0.00962473638355732, + 0.005208942573517561, + 0.004593890625983477, + 0.010076303966343403, + 0.021773071959614754, + 0.008857819251716137, + 0.009266400709748268, + 0.009520789608359337, + 0.009832934476435184, + 0.009417537599802017, + 0.008746232837438583, + 0.00884335022419691, + 0.010007573291659355, + 0.01382659375667572, + 0.008561118505895138, + 0.03325415775179863, + 0.05353366211056709, + 0.10966404527425766, + 0.02434324100613594, + 0.016391662880778313, + 0.05177382752299309, + 0.02061324007809162, + 0.015705455094575882, + 0.008860464207828045, + 0.012812224216759205, + 0.007749633397907019, + 0.008424906060099602, + 0.03724570572376251, + 0.014254637062549591, + 0.007228147238492966, + 0.018888387829065323, + 0.005824224557727575, + 0.004993500653654337, + 0.004855688661336899, + 0.004170053638517857, + 0.01564430072903633, + 0.008654721081256866, + 0.0039250231347978115, + 0.007301699370145798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.00678153894841671 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003633802989497781 + ], + "opposite_logprob_tokens": [ + -5.628633975982666 + ], + "logprob_reward": -0.003633802989497781, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003633802989497781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08362842351198196, + 0.07812914997339249, + 0.08334232866764069, + 0.11113232374191284, + 0.10919415205717087, + 0.1323290467262268, + 0.11794613301753998, + 0.1482085883617401, + 0.08969307690858841, + 0.04639673978090286 + ], + "credit_map": [ + 0.00869328249245882, + 0.02327772043645382, + 0.006586712319403887, + 0.008148721419274807, + 0.011411645449697971, + 0.012264366261661053, + 0.0064010354690253735, + 0.006844949442893267, + 0.008191084489226341, + 0.02065945789217949, + 0.004139145836234093, + 0.011008558794856071, + 0.013818830251693726, + 0.011351922526955605, + 0.004481826443225145, + 0.004478334449231625, + 0.009206132031977177, + 0.023757822811603546, + 0.005327408201992512, + 0.007889418862760067, + 0.012371545657515526, + 0.01146015990525484, + 0.00619497150182724, + 0.007134868763387203, + 0.009464023634791374, + 0.030162330716848373, + 0.008030852302908897, + 0.015768516808748245, + 0.013950735330581665, + 0.013299491256475449, + 0.011283504776656628, + 0.009172878228127956, + 0.008152377791702747, + 0.021589215844869614, + 0.0047590299509465694, + 0.007663091644644737, + 0.027042319998145103, + 0.02683103270828724, + 0.007299806922674179, + 0.005857289303094149, + 0.014079143293201923, + 0.014306462369859219, + 0.011060540564358234, + 0.014493334107100964, + 0.011651773937046528, + 0.01938485912978649, + 0.029636718332767487, + 0.017716236412525177, + 0.009235001169145107, + 0.012457597069442272, + 0.010533824563026428, + 0.009867074899375439, + 0.012224121019244194, + 0.03155108541250229, + 0.017447318881750107, + 0.014630124904215336, + 0.019790975376963615, + 0.016918236389756203, + 0.035347454249858856, + 0.01291778776794672, + 0.010789148509502411, + 0.024532483890652657, + 0.00993372779339552, + 0.01797880046069622, + 0.010250166058540344, + 0.017025625333189964, + 0.007859153673052788, + 0.012428436428308487, + 0.010128289461135864, + 0.013871514238417149, + 0.008062914945185184, + 0.010066990740597248, + 0.005925226025283337, + 0.004420220386236906, + 0.006276135798543692, + 0.006005018018186092, + 0.005194570869207382, + 0.006622391287237406, + 0.006303364876657724, + 0.005649816710501909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.003633802989497781 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012545124627649784 + ], + "opposite_logprob_tokens": [ + -4.387545108795166 + ], + "logprob_reward": -0.012545124627649784, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012545124627649784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08665931224822998, + 0.03550891950726509, + 0.05814323201775551, + 0.21134403347969055, + 0.40932905673980713, + 0.01782826893031597, + 0.05502039194107056, + 0.050776705145835876, + 0.035062506794929504, + 0.04032750800251961 + ], + "credit_map": [ + 0.007761165965348482, + 0.016583532094955444, + 0.0077527835965156555, + 0.010849094949662685, + 0.013616328127682209, + 0.01266338862478733, + 0.008771143853664398, + 0.008661876432597637, + 0.0033980943262577057, + 0.006989813409745693, + 0.0026939017698168755, + 0.005057822447270155, + 0.005204153712838888, + 0.006344834808260202, + 0.002939628902822733, + 0.0028806712944060564, + 0.004236845765262842, + 0.01504835020750761, + 0.00395817868411541, + 0.007817619480192661, + 0.00901913270354271, + 0.0075958226807415485, + 0.0050906711257994175, + 0.005376613698899746, + 0.01698181964457035, + 0.058774203062057495, + 0.013456880114972591, + 0.02945115603506565, + 0.03500314801931381, + 0.02727300114929676, + 0.0161538515239954, + 0.014249988831579685, + 0.03134780749678612, + 0.08246418088674545, + 0.020246392115950584, + 0.03304698318243027, + 0.0846569761633873, + 0.10637471079826355, + 0.02721351943910122, + 0.023978477343916893, + 0.00206339405849576, + 0.002135604154318571, + 0.0017860057996585965, + 0.0014726965455338359, + 0.0021149127278476954, + 0.002313764998689294, + 0.0028468554373830557, + 0.003095034509897232, + 0.006053738761693239, + 0.005226843990385532, + 0.005642345175147057, + 0.005879114847630262, + 0.005203859880566597, + 0.010147185996174812, + 0.008871225640177727, + 0.00799607951194048, + 0.003485222812741995, + 0.015051281079649925, + 0.004361021798104048, + 0.003666982054710388, + 0.005442299880087376, + 0.01092516165226698, + 0.00345431175082922, + 0.0043904283083975315, + 0.004346305038779974, + 0.004627193324267864, + 0.0027455980889499187, + 0.0041070980951189995, + 0.007267517503350973, + 0.00560452276840806, + 0.0030204791110008955, + 0.003343794262036681, + 0.004887876100838184, + 0.00469201197847724, + 0.005381121765822172, + 0.005400898400694132, + 0.006195633206516504, + 0.00602327473461628, + 0.004121646750718355, + 0.0036250471603125334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.012545124627649784 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view of an ancient, circular cork mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned above and slightly tilted looking down at the mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cork mat appear pitted, dry, faded to dusty tan, with frayed edges and stress lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a hairline fracture appear near the center and spread outward in a jagged, branching pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the coaster crumble, shedding particles and fragments to form a chaotic pile of debris?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746253930032253 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746253930032253, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746253930032253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09887406975030899, + 0.038174428045749664, + 0.04278364032506943, + 0.05884154513478279, + 0.24748919904232025, + 0.0561821311712265, + 0.18415489792823792, + 0.06267551332712173, + 0.09474366903305054, + 0.1160808727145195 + ], + "credit_map": [ + 0.011987174861133099, + 0.022687381133437157, + 0.007075970992445946, + 0.012692744843661785, + 0.014857127331197262, + 0.014179606921970844, + 0.007061045616865158, + 0.008333012461662292, + 0.004275683779269457, + 0.008657471276819706, + 0.002916199155151844, + 0.002587223658338189, + 0.006433059927076101, + 0.0070752850733697414, + 0.0023386161774396896, + 0.0038908880669623613, + 0.0036121681332588196, + 0.009836370125412941, + 0.0027613607235252857, + 0.006034399848431349, + 0.005685931071639061, + 0.006853907369077206, + 0.004754035733640194, + 0.003245467087253928, + 0.005932842381298542, + 0.014814387075603008, + 0.004285476170480251, + 0.008937772363424301, + 0.007969352416694164, + 0.0077393739484250546, + 0.005546554923057556, + 0.003615787485614419, + 0.0193121749907732, + 0.05020035430788994, + 0.010455013252794743, + 0.01849786937236786, + 0.05324669927358627, + 0.0674326941370964, + 0.016090188175439835, + 0.012254199013113976, + 0.007397978100925684, + 0.006148881744593382, + 0.004950377158820629, + 0.005046680569648743, + 0.006031551863998175, + 0.009352924302220345, + 0.010598843917250633, + 0.006654895842075348, + 0.021804634481668472, + 0.025326769798994064, + 0.024515660479664803, + 0.017869193106889725, + 0.017118016257882118, + 0.030085066333413124, + 0.02422409877181053, + 0.023211464285850525, + 0.007361514959484339, + 0.010100464336574078, + 0.005087144672870636, + 0.005486978217959404, + 0.0057730525732040405, + 0.014061978086829185, + 0.005202997475862503, + 0.009601389057934284, + 0.011321107856929302, + 0.009633183479309082, + 0.005713522434234619, + 0.00682569295167923, + 0.026694651693105698, + 0.01107114925980568, + 0.0062005650252103806, + 0.01728379912674427, + 0.013242867775261402, + 0.011719836853444576, + 0.014958566054701805, + 0.016411347314715385, + 0.013849523849785328, + 0.01699184440076351, + 0.015790563076734543, + 0.01311632338911295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.006746253930032253 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_14/index_tensor([4754, 4754])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_14/index_tensor([4754, 4754])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..92414cd0a77cb96e2269165a7e95607a0e65d44d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_14/index_tensor([4754, 4754])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Captured in a distinctive style of grainy, desaturated vintage amateur footage, a tight, static close-up focuses on a massive rock lobster. The creature inhabits a dark, damp shoreline composed of smooth black pebbles that gleam in the faint, cool light of a distant moon. Its carapace, a complex mosaic of dark, mottled browns and grays, perfectly mimics the texture of damp, volcanic rock, making it almost blend in with its surroundings. The creature's body remains almost completely inert, a solid, motionless mass. The only signs of life come from its two exceptionally long, segmented tentacles, which twitch erratically, their smooth, dark surfaces catching the silvery moonlight and shimmering with light. In the foreground, the foamy white edges of gentle waves periodically lap gently against the creature's rocky shell, creating a scene of profound silence and the mystery of nocturnal life, as if documenting a rare, secret moment in nature.", + "seed": 27320, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_14/index_4754_seed_27320_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.026142569631338118, + "vlm_logprob": -0.026142569631338118, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.026142569631338118, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03651348501443863 + ], + "opposite_logprob_tokens": [ + -3.536513566970825 + ], + "logprob_reward": -0.03651348501443863, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03651348501443863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10511153936386108, + 0.040827129036188126, + 0.018494050949811935, + 0.04282975569367409, + 0.11429508030414581, + 0.4053865373134613, + 0.02091081254184246, + 0.04537490755319595, + 0.10751309990882874, + 0.09925708919763565 + ], + "credit_map": [ + 0.00806872732937336, + 0.006314456928521395, + 0.007830465212464333, + 0.005986685398966074, + 0.008516916073858738, + 0.033990900963544846, + 0.005493590142577887, + 0.02890978567302227, + 0.007994902320206165, + 0.00317202671431005, + 0.003434407524764538, + 0.007154594175517559, + 0.0067525217309594154, + 0.0033917089458554983, + 0.005394974257797003, + 0.0035319875460118055, + 0.0020472609903663397, + 0.001305923331528902, + 0.0020504449494183064, + 0.00360878580249846, + 0.0025308302138000727, + 0.0025776822585612535, + 0.0027788446750491858, + 0.0015942767495289445, + 0.003864138387143612, + 0.0028728183824568987, + 0.0035663896705955267, + 0.0039841062389314175, + 0.008793625980615616, + 0.00827949307858944, + 0.005213391967117786, + 0.00625578872859478, + 0.008955701254308224, + 0.009558172896504402, + 0.007540729828178883, + 0.013854209333658218, + 0.023105043917894363, + 0.010249851271510124, + 0.023559806868433952, + 0.0174715518951416, + 0.03268647938966751, + 0.014745359309017658, + 0.012446565553545952, + 0.08306761085987091, + 0.022515779361128807, + 0.09916257858276367, + 0.12207597494125366, + 0.018686188384890556, + 0.0020657265558838844, + 0.0028468884993344545, + 0.001912504667416215, + 0.0019059838959947228, + 0.005651543848216534, + 0.003012064378708601, + 0.001872836728580296, + 0.0016432609409093857, + 0.00484677916392684, + 0.0034817794803529978, + 0.008066301234066486, + 0.003649079939350486, + 0.005937537178397179, + 0.008657512255012989, + 0.005188603885471821, + 0.0055473060347139835, + 0.004578825552016497, + 0.0039683436043560505, + 0.004675962030887604, + 0.012431995011866093, + 0.008611592464148998, + 0.026188742369413376, + 0.014114895835518837, + 0.03294273093342781, + 0.017511101439595222, + 0.0061227865517139435, + 0.00476542953401804, + 0.01365758664906025, + 0.025537123903632164, + 0.005990046542137861, + 0.021367935463786125, + 0.004305067006498575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.03651348501443863 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372080624103546 + ], + "opposite_logprob_tokens": [ + -3.4087207317352295 + ], + "logprob_reward": -0.03372080624103546, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372080624103546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047544024884700775, + 0.0830514207482338, + 0.03402921184897423, + 0.12617194652557373, + 0.11040560156106949, + 0.08192604780197144, + 0.23530717194080353, + 0.03602880612015724, + 0.0462358258664608, + 0.1992999166250229 + ], + "credit_map": [ + 0.00637512793764472, + 0.007133967243134975, + 0.00742738414555788, + 0.004582853522151709, + 0.0043028281070292, + 0.008860249072313309, + 0.00453819427639246, + 0.004323423374444246, + 0.004283941350877285, + 0.004021783359348774, + 0.0038289015647023916, + 0.022316427901387215, + 0.007086088880896568, + 0.0029517770744860172, + 0.03301817923784256, + 0.00554431788623333, + 0.002769667189568281, + 0.002380188787356019, + 0.0028767865151166916, + 0.005859825294464827, + 0.008382228203117847, + 0.002024475485086441, + 0.008066749200224876, + 0.0016692891949787736, + 0.014763358049094677, + 0.007847004570066929, + 0.012723660096526146, + 0.018998373299837112, + 0.02414713427424431, + 0.008256766945123672, + 0.01792686991393566, + 0.02150879055261612, + 0.008654030971229076, + 0.008358974941074848, + 0.006896091625094414, + 0.012211178429424763, + 0.018773512914776802, + 0.008685347624123096, + 0.02145572565495968, + 0.025370730087161064, + 0.006026640068739653, + 0.0038046177942305803, + 0.0064160325564444065, + 0.015410283580422401, + 0.008313653990626335, + 0.013515067286789417, + 0.022736256942152977, + 0.005703488830476999, + 0.023039763793349266, + 0.020306715741753578, + 0.03794313967227936, + 0.02227785810828209, + 0.02942214347422123, + 0.04632926359772682, + 0.02443506568670273, + 0.0315532349050045, + 0.0030551045201718807, + 0.0028015137650072575, + 0.005437877494841814, + 0.003353693289682269, + 0.004792772699147463, + 0.00784830842167139, + 0.004439044743776321, + 0.004300491418689489, + 0.002430038060992956, + 0.0025074761360883713, + 0.0018470677314326167, + 0.00516433035954833, + 0.004131268709897995, + 0.011774985119700432, + 0.00543143879622221, + 0.012949223630130291, + 0.019612345844507217, + 0.011613396927714348, + 0.012517343275249004, + 0.03205058351159096, + 0.02885497733950615, + 0.008480689488351345, + 0.06051560118794441, + 0.0256549920886755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.03372080624103546 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978577509522438 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.00978577509522438, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978577509522438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07000397145748138, + 0.08272397518157959, + 0.02799229882657528, + 0.07393673062324524, + 0.17372345924377441, + 0.2595530152320862, + 0.027707280591130257, + 0.049379877746105194, + 0.13202516734600067, + 0.10295425355434418 + ], + "credit_map": [ + 0.009460674598813057, + 0.014775779098272324, + 0.005237363278865814, + 0.006703874561935663, + 0.01070654857903719, + 0.008686820976436138, + 0.006522837560623884, + 0.007910079322755337, + 0.01504805963486433, + 0.004274676088243723, + 0.0037695341743528843, + 0.015725573524832726, + 0.013059929944574833, + 0.004554718732833862, + 0.022374708205461502, + 0.0039167762733995914, + 0.0035650634672492743, + 0.0018166714580729604, + 0.0021161814220249653, + 0.004261129070073366, + 0.006950446870177984, + 0.002762432908639312, + 0.003946981392800808, + 0.002573394449427724, + 0.0069191246293485165, + 0.00748275313526392, + 0.007701859809458256, + 0.009026961401104927, + 0.009028635919094086, + 0.00410598237067461, + 0.015949556604027748, + 0.013721860945224762, + 0.012304795905947685, + 0.013973431661725044, + 0.011592499911785126, + 0.01907058246433735, + 0.0348702147603035, + 0.015724709257483482, + 0.03304007649421692, + 0.03314715251326561, + 0.02274589240550995, + 0.009703275747597218, + 0.00922897458076477, + 0.04867711663246155, + 0.015245985239744186, + 0.06780703365802765, + 0.0743645504117012, + 0.011780175380408764, + 0.0030593324918299913, + 0.00262093567289412, + 0.0037045341450721025, + 0.0033004607539623976, + 0.0031561548821628094, + 0.004412497393786907, + 0.0034477203153073788, + 0.004005644470453262, + 0.003775904420763254, + 0.003532075323164463, + 0.0063215093687176704, + 0.0040623703971505165, + 0.008023339323699474, + 0.011801032349467278, + 0.005994791630655527, + 0.005868850275874138, + 0.005726498086005449, + 0.004410127177834511, + 0.00543103227391839, + 0.015434496104717255, + 0.011155184358358383, + 0.03256116062402725, + 0.017076965421438217, + 0.04022970795631409, + 0.018909268081188202, + 0.007426760159432888, + 0.005554146133363247, + 0.012069257907569408, + 0.02943415939807892, + 0.007224869914352894, + 0.017100753262639046, + 0.005235044285655022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.00978577509522438 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306454211473465 + ], + "opposite_logprob_tokens": [ + -3.168064594268799 + ], + "logprob_reward": -0.04306454211473465, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306454211473465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0560334213078022, + 0.041006751358509064, + 0.01644807681441307, + 0.10562905669212341, + 0.09781385213136673, + 0.2413414865732193, + 0.019177434965968132, + 0.021639397367835045, + 0.27373015880584717, + 0.12718039751052856 + ], + "credit_map": [ + 0.007068411912769079, + 0.0107569620013237, + 0.004551668651401997, + 0.004255735781043768, + 0.00667981943115592, + 0.00930785108357668, + 0.004536512307822704, + 0.00887646246701479, + 0.008342201821506023, + 0.0022893075365573168, + 0.0023520642425864935, + 0.007302744314074516, + 0.010517964139580727, + 0.002162946853786707, + 0.006191456224769354, + 0.0018480661092326045, + 0.002338780090212822, + 0.001362161710858345, + 0.0016173070762306452, + 0.0028884659986943007, + 0.003743566572666168, + 0.0014806631952524185, + 0.001951931626535952, + 0.00106519996188581, + 0.008646426722407341, + 0.010521295480430126, + 0.007034578826278448, + 0.011852155439555645, + 0.010071578435599804, + 0.01292126439511776, + 0.021434694528579712, + 0.023147063329815865, + 0.008414724841713905, + 0.010316727682948112, + 0.006847730837762356, + 0.010929069481790066, + 0.020710235461592674, + 0.009406592696905136, + 0.01797950640320778, + 0.013209265656769276, + 0.021291032433509827, + 0.008832755498588085, + 0.006514095701277256, + 0.047341443598270416, + 0.013970880769193172, + 0.06343159079551697, + 0.0693383440375328, + 0.01062134001404047, + 0.001743882312439382, + 0.0022412408143281937, + 0.002405626466497779, + 0.002091784728690982, + 0.0032554923091083765, + 0.0019355886615812778, + 0.0030530374497175217, + 0.002450782572850585, + 0.0024380204267799854, + 0.002088758163154125, + 0.0027055866084992886, + 0.0023673155810683966, + 0.004589268937706947, + 0.0029980952385812998, + 0.002141991164535284, + 0.0023103600833564997, + 0.011384681798517704, + 0.009240366518497467, + 0.011829173192381859, + 0.031225048005580902, + 0.02315293252468109, + 0.06845320016145706, + 0.034993670880794525, + 0.08345110714435577, + 0.023677904158830643, + 0.009560092352330685, + 0.0074472250416874886, + 0.013625497929751873, + 0.037533629685640335, + 0.008953960612416267, + 0.018344219774007797, + 0.008037874475121498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 9 + ], + "selected_reward": -0.04306454211473465 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628239691257477 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628239691257477, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628239691257477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033680886030197144, + 0.09217190742492676, + 0.017984522506594658, + 0.03492388129234314, + 0.10137681663036346, + 0.055998265743255615, + 0.039926063269376755, + 0.029218649491667747, + 0.19200226664543152, + 0.4027167558670044 + ], + "credit_map": [ + 0.003140936838462949, + 0.003234944539144635, + 0.0034174020402133465, + 0.004232344683259726, + 0.004636224824935198, + 0.007872380316257477, + 0.00299439555965364, + 0.004152260255068541, + 0.006250630132853985, + 0.005110402125865221, + 0.004922937601804733, + 0.020536435768008232, + 0.009058616124093533, + 0.003675475250929594, + 0.03585454449057579, + 0.006762863136827946, + 0.0020403587259352207, + 0.0012269691796973348, + 0.0019482444040477276, + 0.00254752766340971, + 0.004408240783959627, + 0.0013553111348301172, + 0.0026424915995448828, + 0.0018153804121538997, + 0.002514428459107876, + 0.002308593364432454, + 0.003539340104907751, + 0.0035688329953700304, + 0.007983693853020668, + 0.007604163605719805, + 0.004251737147569656, + 0.0031530936248600483, + 0.00867171585559845, + 0.009407375939190388, + 0.006152146961539984, + 0.011229471303522587, + 0.019695403054356575, + 0.009522826410830021, + 0.019198214635252953, + 0.017499662935733795, + 0.003994186408817768, + 0.0027725049294531345, + 0.004958780482411385, + 0.012397662736475468, + 0.004149558022618294, + 0.00637855427339673, + 0.017395639792084694, + 0.003951378632336855, + 0.003956301603466272, + 0.003418457228690386, + 0.005849352106451988, + 0.004765270743519068, + 0.006816423032432795, + 0.0057756491005420685, + 0.004741867538541555, + 0.004602743778377771, + 0.0029337008018046618, + 0.0025596492923796177, + 0.0038394813891500235, + 0.0033803735859692097, + 0.0045723761431872845, + 0.005567324813455343, + 0.004220612812787294, + 0.002145132515579462, + 0.008136422373354435, + 0.008568922989070415, + 0.010330085642635822, + 0.023403039202094078, + 0.014952384866774082, + 0.04362199828028679, + 0.025050848722457886, + 0.05793857201933861, + 0.03777947649359703, + 0.017519710585474968, + 0.022533996030688286, + 0.07630790024995804, + 0.04202183336019516, + 0.014813118614256382, + 0.14348070323467255, + 0.048260003328323364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.007628239691257477 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured in a distinctive style of grainy, desaturated vintage amateur footage, a tight, static close-up focuses on a massive rock lobster. The creature inhabits a dark, damp shoreline composed of smooth black pebbles that gleam in the faint, cool light of a distant moon. Its carapace, a complex mosaic of dark, mottled browns and grays, perfectly mimics the texture of damp, volcanic rock, making it almost blend in with its surroundings. The creature's body remains almost completely inert, a solid, motionless mass. The only signs of life come from its two exceptionally long, segmented tentacles, which twitch erratically, their smooth, dark surfaces catching the silvery moonlight and shimmering with light. In the foreground, the foamy white edges of gentle waves periodically lap gently against the creature's rocky shell, creating a scene of profound silence and the mystery of nocturnal life, as if documenting a rare, secret moment in nature.", + "seed": 27321, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_14/index_4754_seed_27321_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.041618645563721655, + "vlm_logprob": -0.041618645563721655, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.041618645563721655, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01927502639591694 + ], + "opposite_logprob_tokens": [ + -4.394275188446045 + ], + "logprob_reward": -0.01927502639591694, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01927502639591694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15782175958156586, + 0.3075030446052551, + 0.09926266223192215, + 0.03153776749968529, + 0.03746405243873596, + 0.024613942950963974, + 0.20124346017837524, + 0.08632910251617432, + 0.029015202075242996, + 0.02520904131233692 + ], + "credit_map": [ + 0.014322648756206036, + 0.015339811332523823, + 0.009916615672409534, + 0.01291433721780777, + 0.03855901584029198, + 0.03180542588233948, + 0.017762256786227226, + 0.017201652750372887, + 0.025577323511242867, + 0.032371751964092255, + 0.05030859634280205, + 0.03410390764474869, + 0.05553363263607025, + 0.043311040848493576, + 0.035062648355960846, + 0.0312341321259737, + 0.006400522775948048, + 0.0165080688893795, + 0.006425971630960703, + 0.012235604226589203, + 0.01709521748125553, + 0.007163827307522297, + 0.015269051305949688, + 0.01816440187394619, + 0.0038996199145913124, + 0.0035399720072746277, + 0.00393713777884841, + 0.004184338264167309, + 0.0042279851622879505, + 0.003834727918729186, + 0.0035492554306983948, + 0.004364730324596167, + 0.005337262526154518, + 0.004324492532759905, + 0.004074853379279375, + 0.0034717668313533068, + 0.007977401837706566, + 0.0051914541982114315, + 0.004178802948445082, + 0.0029080193489789963, + 0.0020895679481327534, + 0.002105539198964834, + 0.0019217045046389103, + 0.0030468148179352283, + 0.002349559683352709, + 0.0029459241777658463, + 0.005517175421118736, + 0.004637659061700106, + 0.023529144003987312, + 0.024638906121253967, + 0.010411429218947887, + 0.02184067852795124, + 0.04444735497236252, + 0.030265122652053833, + 0.022068995982408524, + 0.024041825905442238, + 0.009812192991375923, + 0.00845533050596714, + 0.004231561906635761, + 0.010073305107653141, + 0.014991335570812225, + 0.015400024130940437, + 0.009524974972009659, + 0.013840378262102604, + 0.00258046668022871, + 0.003774027107283473, + 0.002419149037450552, + 0.00612865062430501, + 0.0026074349880218506, + 0.0037301932461559772, + 0.0037048414815217257, + 0.00407043844461441, + 0.0017722150078043342, + 0.00291500985622406, + 0.0024097836576402187, + 0.003252160968258977, + 0.002772409003227949, + 0.0043615540489554405, + 0.003338593989610672, + 0.004387315828353167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.01927502639591694 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704038619995117 + ], + "opposite_logprob_tokens": [ + -2.127040386199951 + ], + "logprob_reward": -0.12704038619995117, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704038619995117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3461471199989319, + 0.2354421317577362, + 0.1371980756521225, + 0.05571410059928894, + 0.033483315259218216, + 0.025238212198019028, + 0.02203899249434471, + 0.0207087229937315, + 0.08323552459478378, + 0.04079379886388779 + ], + "credit_map": [ + 0.0327574647963047, + 0.036533813923597336, + 0.02232949621975422, + 0.03098100982606411, + 0.05401865392923355, + 0.07997792214155197, + 0.03794478625059128, + 0.05160398781299591, + 0.01794307306408882, + 0.02818743884563446, + 0.0306971687823534, + 0.024848340079188347, + 0.04595622047781944, + 0.032300762832164764, + 0.0276215560734272, + 0.027887573465704918, + 0.010927851311862469, + 0.024513771757483482, + 0.009779723361134529, + 0.02200087159872055, + 0.02787255309522152, + 0.009983716532588005, + 0.017316220328211784, + 0.014803371392190456, + 0.00801831390708685, + 0.00569114601239562, + 0.00807549711316824, + 0.0074601685628294945, + 0.008792509324848652, + 0.006764418911188841, + 0.00530152628198266, + 0.00561052281409502, + 0.0034069097600877285, + 0.0027621611952781677, + 0.003972571808844805, + 0.0040357899852097034, + 0.003936354536563158, + 0.006245497148483992, + 0.004848660435527563, + 0.0042753699235618114, + 0.0024237241595983505, + 0.0022847310174256563, + 0.004645171109586954, + 0.004193130414932966, + 0.00246296520344913, + 0.0024275928735733032, + 0.0038500060327351093, + 0.002950893249362707, + 0.0023264442570507526, + 0.00241886917501688, + 0.002322288928553462, + 0.0026105090510100126, + 0.0035697282291948795, + 0.003336807945743203, + 0.002759650582447648, + 0.002694695023819804, + 0.002076335484161973, + 0.0026237706188112497, + 0.0022050337865948677, + 0.0036694840528070927, + 0.002298100385814905, + 0.003122176043689251, + 0.002337539102882147, + 0.0023762856144458055, + 0.0068119484931230545, + 0.011139550246298313, + 0.007149064913392067, + 0.020038295537233353, + 0.010752744041383266, + 0.0073070228099823, + 0.010216946713626385, + 0.009819956496357918, + 0.004890412092208862, + 0.005372708663344383, + 0.0032628204207867384, + 0.006686689332127571, + 0.00578529667109251, + 0.0049642231315374374, + 0.005259510595351458, + 0.004572139121592045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.12704038619995117 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076280031353235245 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.0076280031353235245, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076280031353235245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17970788478851318, + 0.12021980434656143, + 0.09188006073236465, + 0.03311660885810852, + 0.08110885322093964, + 0.03615556284785271, + 0.26335662603378296, + 0.08461453020572662, + 0.0713462233543396, + 0.0384938083589077 + ], + "credit_map": [ + 0.019379427656531334, + 0.019195562228560448, + 0.014117284677922726, + 0.014398636296391487, + 0.02288186363875866, + 0.0373392254114151, + 0.02136475406587124, + 0.031031154096126556, + 0.01037751603871584, + 0.011854976415634155, + 0.01796635240316391, + 0.012129830196499825, + 0.021648183465003967, + 0.016874920576810837, + 0.014630780555307865, + 0.014737245626747608, + 0.012130779214203358, + 0.019460884854197502, + 0.006259731948375702, + 0.012208613567054272, + 0.014578742906451225, + 0.007457204163074493, + 0.009547886438667774, + 0.010236220434308052, + 0.0048877582885324955, + 0.00480175344273448, + 0.003013553563505411, + 0.003816131502389908, + 0.004541303962469101, + 0.0038679083809256554, + 0.0037887103389948606, + 0.004399488680064678, + 0.010366530157625675, + 0.009152366779744625, + 0.008958075195550919, + 0.006384213920682669, + 0.015805918723344803, + 0.013485489413142204, + 0.008675296790897846, + 0.008280948735773563, + 0.003157422412186861, + 0.0033309024292975664, + 0.002946599153801799, + 0.004862723872065544, + 0.004566340707242489, + 0.006899118423461914, + 0.005150515120476484, + 0.005241940263658762, + 0.030737774446606636, + 0.030741849914193153, + 0.013534813188016415, + 0.02907549776136875, + 0.05831084027886391, + 0.03946733847260475, + 0.02859901450574398, + 0.03288949280977249, + 0.008502931334078312, + 0.009512341581285, + 0.008808545768260956, + 0.014694910496473312, + 0.012928941287100315, + 0.009687492623925209, + 0.00963808037340641, + 0.010841288603842258, + 0.007991495542228222, + 0.01359353307634592, + 0.0062581324018538, + 0.00944426003843546, + 0.010223046876490116, + 0.006090313661843538, + 0.007374114822596312, + 0.010371324606239796, + 0.002995725953951478, + 0.0049257283098995686, + 0.0029788310639560223, + 0.008445151150226593, + 0.0050435555167496204, + 0.003768225433304906, + 0.004553740378469229, + 0.005782848224043846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0076280031353235245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606742851436138 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.01606742851436138, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606742851436138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24099107086658478, + 0.12007934600114822, + 0.09268568456172943, + 0.054997269064188004, + 0.04573196917772293, + 0.036067359149456024, + 0.06003430858254433, + 0.08285219967365265, + 0.13310804963111877, + 0.13345275819301605 + ], + "credit_map": [ + 0.028021449223160744, + 0.03029278665781021, + 0.015076635405421257, + 0.016811545938253403, + 0.0400637648999691, + 0.056564781814813614, + 0.02642999030649662, + 0.027730092406272888, + 0.010376349091529846, + 0.011842458508908749, + 0.019490564242005348, + 0.018260134384036064, + 0.010927575640380383, + 0.012757619842886925, + 0.016184410080313683, + 0.020240236073732376, + 0.011346859857439995, + 0.011369929648935795, + 0.007206462323665619, + 0.01229487918317318, + 0.015581185929477215, + 0.012231576256453991, + 0.010748526081442833, + 0.011906259693205357, + 0.005508386064320803, + 0.007023766171187162, + 0.007686302997171879, + 0.007983210496604443, + 0.009221685118973255, + 0.005701574496924877, + 0.006779517978429794, + 0.005092828068882227, + 0.003811922622844577, + 0.005292909685522318, + 0.007512417621910572, + 0.0077664717100560665, + 0.004872725810855627, + 0.006311099510639906, + 0.005069649312645197, + 0.0050947698764503, + 0.0026623327285051346, + 0.0025088400579988956, + 0.0060943723656237125, + 0.004911441821604967, + 0.0035360921174287796, + 0.004359617829322815, + 0.00509257847443223, + 0.006902080494910479, + 0.006220280658453703, + 0.006850629113614559, + 0.004291127901524305, + 0.009636425413191319, + 0.011831074021756649, + 0.007932967506349087, + 0.0062868716195225716, + 0.006984931882470846, + 0.009230679832398891, + 0.011585963889956474, + 0.006160861346870661, + 0.009404010139405727, + 0.01563737541437149, + 0.012381132692098618, + 0.008853921666741371, + 0.009598255157470703, + 0.012137852609157562, + 0.01808265782892704, + 0.009470585733652115, + 0.024555204436182976, + 0.01667364127933979, + 0.02001262828707695, + 0.014911863021552563, + 0.017263617366552353, + 0.013389559462666512, + 0.012628355994820595, + 0.00981899630278349, + 0.023321466520428658, + 0.014874533750116825, + 0.019943540915846825, + 0.01533584762364626, + 0.024140462279319763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.01606742851436138 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808238357305527 + ], + "opposite_logprob_tokens": [ + -3.2880823612213135 + ], + "logprob_reward": -0.03808238357305527, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808238357305527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.315016508102417, + 0.118081234395504, + 0.12057796120643616, + 0.02000347152352333, + 0.11292187869548798, + 0.026523306965827942, + 0.03356124833226204, + 0.04873406141996384, + 0.10760067403316498, + 0.09697964787483215 + ], + "credit_map": [ + 0.02492842823266983, + 0.02519981376826763, + 0.018355460837483406, + 0.022069202736020088, + 0.06429136544466019, + 0.08028501272201538, + 0.03598804399371147, + 0.04389917850494385, + 0.008878021501004696, + 0.01676821894943714, + 0.007299044169485569, + 0.010655282065272331, + 0.03042767010629177, + 0.015441604889929295, + 0.014561626128852367, + 0.01404976099729538, + 0.010546463541686535, + 0.016914425417780876, + 0.012178273871541023, + 0.024101633578538895, + 0.013432004489004612, + 0.008612743578851223, + 0.016774646937847137, + 0.01801777258515358, + 0.0014061769470572472, + 0.001978108426555991, + 0.003989442251622677, + 0.0036302488297224045, + 0.002239690627902746, + 0.0021694861352443695, + 0.002345575951039791, + 0.0022447421215474606, + 0.008346770890057087, + 0.008999335579574108, + 0.016462692990899086, + 0.012295786291360855, + 0.016168693080544472, + 0.022328438237309456, + 0.013975873589515686, + 0.014344295486807823, + 0.002584921894595027, + 0.0021527709905058146, + 0.003244989085942507, + 0.0046984124928712845, + 0.0035573027562350035, + 0.002626793459057808, + 0.004027789458632469, + 0.003630325896665454, + 0.003224254585802555, + 0.0036372740287333727, + 0.0030306135304272175, + 0.007505965884774923, + 0.004565178416669369, + 0.00391616253182292, + 0.004028100986033678, + 0.003653700463473797, + 0.004563204012811184, + 0.0053146472200751305, + 0.0068986560218036175, + 0.00711811613291502, + 0.0066428580321371555, + 0.008167185820639133, + 0.004663914442062378, + 0.005365480203181505, + 0.011799540370702744, + 0.015713121742010117, + 0.008070182986557484, + 0.01378101110458374, + 0.014444725587964058, + 0.010552131570875645, + 0.012708153575658798, + 0.02053181454539299, + 0.006182115059345961, + 0.010605244897305965, + 0.01231054775416851, + 0.021662302315235138, + 0.01137149054557085, + 0.008504658006131649, + 0.012557592242956161, + 0.013785704039037228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.03808238357305527 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured in a distinctive style of grainy, desaturated vintage amateur footage, a tight, static close-up focuses on a massive rock lobster. The creature inhabits a dark, damp shoreline composed of smooth black pebbles that gleam in the faint, cool light of a distant moon. Its carapace, a complex mosaic of dark, mottled browns and grays, perfectly mimics the texture of damp, volcanic rock, making it almost blend in with its surroundings. The creature's body remains almost completely inert, a solid, motionless mass. The only signs of life come from its two exceptionally long, segmented tentacles, which twitch erratically, their smooth, dark surfaces catching the silvery moonlight and shimmering with light. In the foreground, the foamy white edges of gentle waves periodically lap gently against the creature's rocky shell, creating a scene of profound silence and the mystery of nocturnal life, as if documenting a rare, secret moment in nature.", + "seed": 27322, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_14/index_4754_seed_27322_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019874509144574403, + "vlm_logprob": -0.019874509144574403, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.019874509144574403, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020140310749411583 + ], + "opposite_logprob_tokens": [ + -4.270140171051025 + ], + "logprob_reward": -0.020140310749411583, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.020140310749411583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29560214281082153, + 0.040505409240722656, + 0.02748408354818821, + 0.030699944123625755, + 0.04264550283551216, + 0.1810368448495865, + 0.034687288105487823, + 0.17827647924423218, + 0.12248370051383972, + 0.04657852277159691 + ], + "credit_map": [ + 0.029516372829675674, + 0.03212481364607811, + 0.0326349139213562, + 0.039092764258384705, + 0.046315282583236694, + 0.049315158277750015, + 0.03843400254845619, + 0.028168894350528717, + 0.00403597392141819, + 0.0050682611763477325, + 0.005224531516432762, + 0.0067695798352360725, + 0.005782200023531914, + 0.005245922598987818, + 0.0037662137765437365, + 0.004612732212990522, + 0.00339326960965991, + 0.003813056042417884, + 0.002884769346565008, + 0.004510789178311825, + 0.004181855358183384, + 0.004144138656556606, + 0.0024820349644869566, + 0.0020741745829582214, + 0.0032872396986931562, + 0.0035653135273605585, + 0.00442116754129529, + 0.004109598696231842, + 0.003114194143563509, + 0.003816653974354267, + 0.00520439213141799, + 0.0031813893001526594, + 0.004462519660592079, + 0.003923418000340462, + 0.003794632386416197, + 0.00463821180164814, + 0.011276490986347198, + 0.008763357065618038, + 0.003655507229268551, + 0.0021313710603863, + 0.020325133576989174, + 0.010611409321427345, + 0.02176549844443798, + 0.011325330473482609, + 0.038517531007528305, + 0.04094930365681648, + 0.02248702198266983, + 0.015055645257234573, + 0.003842533566057682, + 0.004348244983702898, + 0.00341789866797626, + 0.0036644015926867723, + 0.008172618225216866, + 0.006339648738503456, + 0.0023711719550192356, + 0.0025307736359536648, + 0.020694512873888016, + 0.01323385164141655, + 0.02242632396519184, + 0.017596017569303513, + 0.03808548301458359, + 0.04426087066531181, + 0.011258689686655998, + 0.010720746591687202, + 0.01562102697789669, + 0.016940046101808548, + 0.023140044882893562, + 0.013076850213110447, + 0.013519122265279293, + 0.01176369097083807, + 0.01417197473347187, + 0.014250965788960457, + 0.00357068027369678, + 0.004806954879313707, + 0.0025528566911816597, + 0.002633070107549429, + 0.013700377196073532, + 0.014117863029241562, + 0.002886205678805709, + 0.002310524694621563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.020140310749411583 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004110222216695547 + ], + "opposite_logprob_tokens": [ + -5.504110336303711 + ], + "logprob_reward": -0.004110222216695547, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004110222216695547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23141083121299744, + 0.07455696165561676, + 0.11010507494211197, + 0.05518705025315285, + 0.14249172806739807, + 0.11409397423267365, + 0.07913779467344284, + 0.11444228887557983, + 0.04009908065199852, + 0.03847526013851166 + ], + "credit_map": [ + 0.027122830972075462, + 0.03581961989402771, + 0.028001587837934494, + 0.0323333740234375, + 0.022645151242613792, + 0.042459797114133835, + 0.02376003935933113, + 0.019268423318862915, + 0.008793638087809086, + 0.007759700529277325, + 0.008760744705796242, + 0.010483866557478905, + 0.008160363882780075, + 0.014168553985655308, + 0.008781874552369118, + 0.007648221682757139, + 0.013141636736690998, + 0.015527567826211452, + 0.01420607790350914, + 0.018130207434296608, + 0.012959453277289867, + 0.016519764438271523, + 0.010059342719614506, + 0.0095610236749053, + 0.0071642412804067135, + 0.010098867118358612, + 0.007044572848826647, + 0.009178269654512405, + 0.007222806103527546, + 0.005527549423277378, + 0.004346473608165979, + 0.004604271147400141, + 0.02095567248761654, + 0.01085685845464468, + 0.018142223358154297, + 0.01688709668815136, + 0.022072864696383476, + 0.0358906090259552, + 0.009222681634128094, + 0.008463717065751553, + 0.011392506770789623, + 0.007061568088829517, + 0.012149926275014877, + 0.0063159591518342495, + 0.03465573862195015, + 0.031794480979442596, + 0.005632701795548201, + 0.005091091617941856, + 0.0076260934583842754, + 0.005455037113279104, + 0.006589439697563648, + 0.005557388998568058, + 0.022972047328948975, + 0.020903633907437325, + 0.004513130988925695, + 0.005521018523722887, + 0.016970783472061157, + 0.02107001654803753, + 0.01667984388768673, + 0.013028101064264774, + 0.01275864988565445, + 0.01582852005958557, + 0.010633363388478756, + 0.007473015692085028, + 0.005452908109873533, + 0.006477318238466978, + 0.0056987362913787365, + 0.003919563256204128, + 0.004919864237308502, + 0.0036278823390603065, + 0.005521669983863831, + 0.004481139127165079, + 0.00444490322843194, + 0.003646932076662779, + 0.0041176751255989075, + 0.0026497214566916227, + 0.0075071752071380615, + 0.006120686884969473, + 0.0036456433590501547, + 0.006342517677694559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.004110222216695547 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031929248943924904 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.0031929248943924904, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031929248943924904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19114069640636444, + 0.0876215398311615, + 0.07789087295532227, + 0.05984874069690704, + 0.07491235435009003, + 0.12205339968204498, + 0.034635432064533234, + 0.053157202899456024, + 0.07795802503824234, + 0.22078174352645874 + ], + "credit_map": [ + 0.02010301686823368, + 0.020755138248205185, + 0.024876758456230164, + 0.022642897441983223, + 0.022886158898472786, + 0.0328349843621254, + 0.018405316397547722, + 0.02863643690943718, + 0.00996403582394123, + 0.009600825607776642, + 0.009155426174402237, + 0.011252435855567455, + 0.011749804019927979, + 0.014025728218257427, + 0.011163255199790001, + 0.010710031725466251, + 0.008619026280939579, + 0.006732509005814791, + 0.009114695712924004, + 0.00977806281298399, + 0.012775585055351257, + 0.011435545980930328, + 0.009351585060358047, + 0.010083873756229877, + 0.005772237200289965, + 0.005859595723450184, + 0.008463937789201736, + 0.007804405875504017, + 0.0048041921108961105, + 0.008146263659000397, + 0.010950281284749508, + 0.008047831244766712, + 0.012403014115989208, + 0.005559599958360195, + 0.010765167884528637, + 0.010003701783716679, + 0.00706166448071599, + 0.018252743408083916, + 0.005281044170260429, + 0.005585429258644581, + 0.02039537951350212, + 0.009089604951441288, + 0.016025498509407043, + 0.010236330330371857, + 0.015530253760516644, + 0.01814473792910576, + 0.01953076757490635, + 0.013100830838084221, + 0.003762197680771351, + 0.004085495136678219, + 0.0037895601708441973, + 0.0058491285890340805, + 0.0037282032426446676, + 0.003430766286328435, + 0.006748809479176998, + 0.003241277299821377, + 0.006109788082540035, + 0.004380523227155209, + 0.006177427247166634, + 0.005862934980541468, + 0.010009551420807838, + 0.01267240010201931, + 0.004308031406253576, + 0.0036365545820444822, + 0.010608500801026821, + 0.0077808694913983345, + 0.018613768741488457, + 0.009228680282831192, + 0.009302297607064247, + 0.006874976679682732, + 0.007449019700288773, + 0.00809992104768753, + 0.021099645644426346, + 0.017104998230934143, + 0.011298678815364838, + 0.010461054742336273, + 0.0685412734746933, + 0.05921885371208191, + 0.012314267456531525, + 0.020742977038025856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.0031929248943924904 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023291567340493202 + ], + "opposite_logprob_tokens": [ + -3.77329158782959 + ], + "logprob_reward": -0.023291567340493202, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023291567340493202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18144093453884125, + 0.06762384623289108, + 0.05123274028301239, + 0.09422314167022705, + 0.1560056507587433, + 0.14607322216033936, + 0.1101444736123085, + 0.09486590325832367, + 0.03977138176560402, + 0.05861877650022507 + ], + "credit_map": [ + 0.021255191415548325, + 0.029520591720938683, + 0.022047430276870728, + 0.023440750315785408, + 0.022820008918642998, + 0.0317620225250721, + 0.0162124652415514, + 0.014382475055754185, + 0.0078943632543087, + 0.010358672589063644, + 0.007386163808405399, + 0.009000019170343876, + 0.008648131042718887, + 0.00765937939286232, + 0.008923999033868313, + 0.007753119338303804, + 0.0063271415419876575, + 0.008803103119134903, + 0.006102303508669138, + 0.0065246159210801125, + 0.0061639598570764065, + 0.007473611738532782, + 0.004890690557658672, + 0.00494731217622757, + 0.012186363339424133, + 0.011753218248486519, + 0.010865675285458565, + 0.012364986352622509, + 0.011821208521723747, + 0.018016712740063667, + 0.009759705513715744, + 0.007455266080796719, + 0.016097959131002426, + 0.019274068996310234, + 0.01949336938560009, + 0.03739958629012108, + 0.02112760953605175, + 0.021288540214300156, + 0.013316398486495018, + 0.008008106611669064, + 0.013056735508143902, + 0.01376565732061863, + 0.015423111617565155, + 0.012842020951211452, + 0.03240653872489929, + 0.03717360645532608, + 0.01264178566634655, + 0.008763782680034637, + 0.011600697413086891, + 0.010107191279530525, + 0.013780945912003517, + 0.006625497248023748, + 0.025478864088654518, + 0.022855674847960472, + 0.011759649030864239, + 0.007935947738587856, + 0.011834478937089443, + 0.0217208843678236, + 0.010729794390499592, + 0.010703894309699535, + 0.013038862496614456, + 0.007850478403270245, + 0.010000350885093212, + 0.008987161330878735, + 0.004361810628324747, + 0.00745840510353446, + 0.0045791021548211575, + 0.0040773143991827965, + 0.004701931495219469, + 0.004183050710707903, + 0.006292504724115133, + 0.0041172620840370655, + 0.006476970389485359, + 0.006757774390280247, + 0.0065587367862463, + 0.009297805838286877, + 0.006692799273878336, + 0.007477385923266411, + 0.008186648599803448, + 0.0071706571616232395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.023291567340493202 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048637520521879196 + ], + "opposite_logprob_tokens": [ + -3.048637628555298 + ], + "logprob_reward": -0.048637520521879196, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048637520521879196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08101704716682434, + 0.018744638189673424, + 0.030746551230549812, + 0.02938760817050934, + 0.04626668617129326, + 0.06103433296084404, + 0.04755733907222748, + 0.08932403475046158, + 0.04952634498476982, + 0.5463954210281372 + ], + "credit_map": [ + 0.007243855390697718, + 0.010072607547044754, + 0.009548992849886417, + 0.009832469746470451, + 0.009452220052480698, + 0.00904840137809515, + 0.011577078141272068, + 0.014241427183151245, + 0.0016632089391350746, + 0.0021116517018526793, + 0.002199030015617609, + 0.0022503735963255167, + 0.002329269191250205, + 0.0029152908828109503, + 0.0028729226905852556, + 0.002402891404926777, + 0.004387680906802416, + 0.0030798916704952717, + 0.0033169814851135015, + 0.0037238015793263912, + 0.003619373543187976, + 0.006523072719573975, + 0.003374297171831131, + 0.002721452619880438, + 0.002823702059686184, + 0.003223107196390629, + 0.003447694005444646, + 0.0034690438769757748, + 0.0031248771119862795, + 0.0028952881693840027, + 0.00480855256319046, + 0.0055953445844352245, + 0.005214643198996782, + 0.0038068625144660473, + 0.004733640234917402, + 0.00777130201458931, + 0.009522070176899433, + 0.009402699768543243, + 0.002933683106675744, + 0.002881784923374653, + 0.006277115549892187, + 0.005062072072178125, + 0.005952119827270508, + 0.00412383396178484, + 0.01718348264694214, + 0.015140078030526638, + 0.0035435822792351246, + 0.003752047661691904, + 0.006106005981564522, + 0.004429375287145376, + 0.006286599673330784, + 0.007188729476183653, + 0.005596191622316837, + 0.006315821316093206, + 0.004984875209629536, + 0.006649738643318415, + 0.012546992860734463, + 0.006783445365726948, + 0.02275577001273632, + 0.012787408195436, + 0.008618286810815334, + 0.009095014072954655, + 0.006813076324760914, + 0.009924041107296944, + 0.005934278946369886, + 0.007249726913869381, + 0.009419183246791363, + 0.0059113884344697, + 0.00613964581862092, + 0.004496556706726551, + 0.004981785546988249, + 0.00539377611130476, + 0.049378372728824615, + 0.03522995486855507, + 0.026245716959238052, + 0.025299834087491035, + 0.17832250893115997, + 0.15534450113773346, + 0.029279086738824844, + 0.04729544743895531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.048637520521879196 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Captured in a distinctive style of grainy, desaturated vintage amateur footage, a tight, static close-up focuses on a massive rock lobster. The creature inhabits a dark, damp shoreline composed of smooth black pebbles that gleam in the faint, cool light of a distant moon. Its carapace, a complex mosaic of dark, mottled browns and grays, perfectly mimics the texture of damp, volcanic rock, making it almost blend in with its surroundings. The creature's body remains almost completely inert, a solid, motionless mass. The only signs of life come from its two exceptionally long, segmented tentacles, which twitch erratically, their smooth, dark surfaces catching the silvery moonlight and shimmering with light. In the foreground, the foamy white edges of gentle waves periodically lap gently against the creature's rocky shell, creating a scene of profound silence and the mystery of nocturnal life, as if documenting a rare, secret moment in nature.", + "seed": 27323, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_14/index_4754_seed_27323_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.027454517036676406, + "vlm_logprob": -0.027454517036676406, + "vlm_yes_no_margin": 3.6499999523162843, + "vlm_token_credit": -0.027454517036676406, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024131974205374718 + ], + "opposite_logprob_tokens": [ + -4.024131774902344 + ], + "logprob_reward": -0.024131974205374718, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.024131974205374718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5622907280921936, + 0.032856088131666183, + 0.04079705849289894, + 0.03960903361439705, + 0.04306282848119736, + 0.028087710961699486, + 0.07677303999662399, + 0.09216295927762985, + 0.03881479799747467, + 0.045545794069767 + ], + "credit_map": [ + 0.03464563563466072, + 0.023404719308018684, + 0.11835918575525284, + 0.21999816596508026, + 0.017850136384367943, + 0.06906115263700485, + 0.054913897067308426, + 0.02405782975256443, + 0.004464792087674141, + 0.0036904800217598677, + 0.0025557454209774733, + 0.0041616009548306465, + 0.002971607493236661, + 0.0074635581113398075, + 0.003761312225833535, + 0.0037869939114898443, + 0.004976495169103146, + 0.0053381845355033875, + 0.007999195717275143, + 0.004362249281257391, + 0.003506663255393505, + 0.0033266290556639433, + 0.007183663547039032, + 0.0041039809584617615, + 0.00495615229010582, + 0.006641581654548645, + 0.004074848257005215, + 0.0043942551128566265, + 0.003962323069572449, + 0.0036365652922540903, + 0.006634569261223078, + 0.00530873890966177, + 0.005715768318623304, + 0.00565158948302269, + 0.006494343280792236, + 0.0048156362026929855, + 0.005656114779412746, + 0.003606547834351659, + 0.005496692378073931, + 0.005626136902719736, + 0.003590171691030264, + 0.003451999742537737, + 0.004690976347774267, + 0.0038217538967728615, + 0.0026494222693145275, + 0.0023451701272279024, + 0.0035649880301207304, + 0.0039732325822114944, + 0.009662573225796223, + 0.01029160339385271, + 0.0076128095388412476, + 0.009823298081755638, + 0.007366690319031477, + 0.014887085184454918, + 0.009586704894900322, + 0.0075422837398946285, + 0.017790528014302254, + 0.00815915409475565, + 0.01079110149294138, + 0.026486366987228394, + 0.003495546290650964, + 0.0076200817711651325, + 0.008818913251161575, + 0.009001274593174458, + 0.004219627007842064, + 0.011653963476419449, + 0.003933654632419348, + 0.00462397001683712, + 0.003243295708671212, + 0.005133373197168112, + 0.0032593158539384604, + 0.0027475981041789055, + 0.004700704012066126, + 0.010551582090556622, + 0.007285817060619593, + 0.0054423194378614426, + 0.006165030412375927, + 0.00521455192938447, + 0.003261479316279292, + 0.0029243158642202616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.024131974205374718 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636036090552807 + ], + "opposite_logprob_tokens": [ + -3.651360273361206 + ], + "logprob_reward": -0.02636036090552807, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636036090552807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6844744086265564, + 0.026443351060152054, + 0.03716828674077988, + 0.07742355018854141, + 0.02577524073421955, + 0.052538905292749405, + 0.01789647527039051, + 0.026356566697359085, + 0.027941497042775154, + 0.02398170903325081 + ], + "credit_map": [ + 0.10947144031524658, + 0.032502129673957825, + 0.06379131972789764, + 0.14334116876125336, + 0.02219078317284584, + 0.23077067732810974, + 0.03649752587080002, + 0.04590930789709091, + 0.0029628390911966562, + 0.005210395436733961, + 0.002244829898700118, + 0.003182761836796999, + 0.004460470750927925, + 0.003077856032177806, + 0.003367087570950389, + 0.001937109394930303, + 0.0028044467326253653, + 0.0038863648660480976, + 0.00399519270285964, + 0.0049301558174192905, + 0.005293841008096933, + 0.0034325108863413334, + 0.008041687309741974, + 0.004784091375768185, + 0.008261576294898987, + 0.009201455861330032, + 0.009094612672924995, + 0.008355052210390568, + 0.00819909293204546, + 0.006837029010057449, + 0.01360497809946537, + 0.01386975310742855, + 0.003504079068079591, + 0.0030986310448497534, + 0.0022468550596386194, + 0.002536274027079344, + 0.004424425307661295, + 0.0023345118388533592, + 0.004363411106169224, + 0.003267053049057722, + 0.005273169372230768, + 0.0053286743350327015, + 0.009597500786185265, + 0.0070695471949875355, + 0.005047043319791555, + 0.004430348519235849, + 0.008249136619269848, + 0.007543486542999744, + 0.0020313935820013285, + 0.0016811690293252468, + 0.001730500953271985, + 0.003249227534979582, + 0.0014434950426220894, + 0.0033476660028100014, + 0.0025495009031146765, + 0.0018635229207575321, + 0.004081607796251774, + 0.0027489429339766502, + 0.003374022664502263, + 0.00450565479695797, + 0.0024094770196825266, + 0.002708866260945797, + 0.0035428686533123255, + 0.0029851258732378483, + 0.0032628923654556274, + 0.007958103902637959, + 0.003308850107714534, + 0.004195400979369879, + 0.0021698337513953447, + 0.0027278668712824583, + 0.0024013060610741377, + 0.0019172413740307093, + 0.002054982353001833, + 0.006288548000156879, + 0.002648123074322939, + 0.0022200928069651127, + 0.004975591786205769, + 0.001975229475647211, + 0.0023920664098113775, + 0.0014270750107243657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.02636036090552807 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288540542125702 + ], + "opposite_logprob_tokens": [ + -3.7732884883880615 + ], + "logprob_reward": -0.023288540542125702, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288540542125702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4323957562446594, + 0.08235213160514832, + 0.06711390614509583, + 0.05416680499911308, + 0.032531287521123886, + 0.023992734029889107, + 0.08476252108812332, + 0.03089250437915325, + 0.11193833500146866, + 0.07985405623912811 + ], + "credit_map": [ + 0.08572982996702194, + 0.026311248540878296, + 0.024852313101291656, + 0.054898448288440704, + 0.017040954902768135, + 0.1665915548801422, + 0.020935235545039177, + 0.03603619337081909, + 0.013666315004229546, + 0.01256047748029232, + 0.0068441759794950485, + 0.009529022499918938, + 0.006951469928026199, + 0.01864282786846161, + 0.00676838681101799, + 0.007389452308416367, + 0.008358735591173172, + 0.008569344878196716, + 0.009453155100345612, + 0.00869743525981903, + 0.005582520738244057, + 0.006532964296638966, + 0.007220595143735409, + 0.012699153274297714, + 0.007750887423753738, + 0.009562907740473747, + 0.007506130263209343, + 0.007505013141781092, + 0.003991885110735893, + 0.006303341127932072, + 0.006420499179512262, + 0.005126142408698797, + 0.00364677794277668, + 0.004112496506422758, + 0.0034160485956817865, + 0.0035761715844273567, + 0.004538228735327721, + 0.002550639444962144, + 0.005286863539367914, + 0.005404064431786537, + 0.0035407880786806345, + 0.003430162090808153, + 0.0030830518808215857, + 0.0029114061035215855, + 0.003013885347172618, + 0.0021562944166362286, + 0.002636085031554103, + 0.0032210603822022676, + 0.010561194270849228, + 0.01114736869931221, + 0.008102272637188435, + 0.010058939456939697, + 0.009373879060149193, + 0.015964271500706673, + 0.011434369720518589, + 0.008120226673781872, + 0.003239967627450824, + 0.0030466970056295395, + 0.004805253818631172, + 0.005945472978055477, + 0.0025508636608719826, + 0.0032693534158170223, + 0.004491503816097975, + 0.0035433913581073284, + 0.012068393640220165, + 0.0367945171892643, + 0.009702397510409355, + 0.01383394654840231, + 0.0081355394795537, + 0.01063716784119606, + 0.011731811799108982, + 0.009034550748765469, + 0.008220062591135502, + 0.010725274682044983, + 0.013583063147962093, + 0.013230444863438606, + 0.008869539014995098, + 0.00973832793533802, + 0.009493366815149784, + 0.005993977189064026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.023288540542125702 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029801562428474426 + ], + "opposite_logprob_tokens": [ + -3.529801607131958 + ], + "logprob_reward": -0.029801562428474426, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801562428474426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5308466553688049, + 0.030162464827299118, + 0.07754826545715332, + 0.038443613797426224, + 0.04554387181997299, + 0.03660432621836662, + 0.029262015596032143, + 0.07893107831478119, + 0.07001753151416779, + 0.06264021247625351 + ], + "credit_map": [ + 0.11123695969581604, + 0.049159832298755646, + 0.041899047791957855, + 0.03880658373236656, + 0.022275060415267944, + 0.19603024423122406, + 0.028256312012672424, + 0.04318259656429291, + 0.0033314714673906565, + 0.0028169811703264713, + 0.0028898699674755335, + 0.0031815911643207073, + 0.0038192044012248516, + 0.004349112510681152, + 0.005502028856426477, + 0.0042722029611468315, + 0.0058590322732925415, + 0.008554728701710701, + 0.01739411987364292, + 0.006710521876811981, + 0.009008929133415222, + 0.007344626355916262, + 0.010434061288833618, + 0.012242237105965614, + 0.0051625813357532024, + 0.007098495960235596, + 0.004682162310928106, + 0.004696345422416925, + 0.0034035074058920145, + 0.003814074909314513, + 0.004891794174909592, + 0.004694649018347263, + 0.005824462976306677, + 0.0073469956405460835, + 0.006734792143106461, + 0.005750872194766998, + 0.005162940826267004, + 0.0034365945030003786, + 0.005069350823760033, + 0.006217858288437128, + 0.004816629458218813, + 0.004649358335882425, + 0.005469420924782753, + 0.0049325223080813885, + 0.0029931955505162477, + 0.002723369747400284, + 0.004836755804717541, + 0.006183071993291378, + 0.003361994167789817, + 0.004459671676158905, + 0.004463596269488335, + 0.0036698582116514444, + 0.0026528015732765198, + 0.003010871820151806, + 0.00395916448906064, + 0.0036840566899627447, + 0.015596643090248108, + 0.006955633871257305, + 0.008763682097196579, + 0.021909834817051888, + 0.0032036155462265015, + 0.007354904897511005, + 0.007568302098661661, + 0.007578454911708832, + 0.008783675730228424, + 0.019547075033187866, + 0.005013942252844572, + 0.009230546653270721, + 0.005447653122246265, + 0.008798353374004364, + 0.007713885046541691, + 0.005482391454279423, + 0.007501855492591858, + 0.01372518204152584, + 0.007195682264864445, + 0.005843225866556168, + 0.009868512861430645, + 0.007981084287166595, + 0.005578773561865091, + 0.004945891443639994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.029801562428474426 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video style grainy, desaturated, and resembling vintage amateur footage?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot a tight, static close-up focused on a massive rock lobster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a dark, damp shoreline with smooth black pebbles under faint cool moonlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lobster's body almost completely inert and motionless, with only its two long tentacles twitching erratically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do foamy white waves periodically lap against the creature's shell in the foreground?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369014710187912 + ], + "opposite_logprob_tokens": [ + -3.4086902141571045 + ], + "logprob_reward": -0.03369014710187912, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369014710187912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28565165400505066, + 0.20574572682380676, + 0.08575345575809479, + 0.03147215023636818, + 0.021306732669472694, + 0.019376855343580246, + 0.047861747443675995, + 0.05910049006342888, + 0.18406392633914948, + 0.05966726318001747 + ], + "credit_map": [ + 0.021469300612807274, + 0.034385863691568375, + 0.03951585665345192, + 0.06574662029743195, + 0.013817773200571537, + 0.071718730032444, + 0.019295984879136086, + 0.019701503217220306, + 0.0362650491297245, + 0.026669010519981384, + 0.02550768479704857, + 0.0386791005730629, + 0.015399528667330742, + 0.021369513124227524, + 0.020462850108742714, + 0.021392997354269028, + 0.007964322343468666, + 0.012963431887328625, + 0.01481474470347166, + 0.00793285109102726, + 0.008657334372401237, + 0.008373304270207882, + 0.009556765668094158, + 0.015490700490772724, + 0.0029440701473504305, + 0.0029667180497199297, + 0.004632598254829645, + 0.003361821873113513, + 0.00333445705473423, + 0.0035001172218471766, + 0.004723881371319294, + 0.0060084848664700985, + 0.0022085991222411394, + 0.0025318663101643324, + 0.0022129721473902464, + 0.002423742087557912, + 0.002407067222520709, + 0.001077277702279389, + 0.0047396388836205006, + 0.003705569775775075, + 0.002402249025180936, + 0.0021976057905703783, + 0.0031118954066187143, + 0.0025155507028102875, + 0.002836675848811865, + 0.0014222452882677317, + 0.002522258087992668, + 0.002368374727666378, + 0.00451652193441987, + 0.005686589516699314, + 0.0049375081434845924, + 0.007170782890170813, + 0.00456851813942194, + 0.004201959818601608, + 0.009335624054074287, + 0.0074442424811422825, + 0.011007847264409065, + 0.005881509277969599, + 0.006998191587626934, + 0.013946360908448696, + 0.003049369901418686, + 0.005920263007283211, + 0.00620254036039114, + 0.00609440915286541, + 0.024143146350979805, + 0.0731780007481575, + 0.011374540627002716, + 0.023776769638061523, + 0.014470509253442287, + 0.014473512768745422, + 0.009900618344545364, + 0.01274683978408575, + 0.0068328604102134705, + 0.009074859321117401, + 0.009352410212159157, + 0.008952646516263485, + 0.006270367186516523, + 0.009134260006248951, + 0.005240549799054861, + 0.004809309262782335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.03369014710187912 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_15/index_tensor([6605, 6605])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_15/index_tensor([6605, 6605])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..dc0e9f945741990d83a7f05cdce9dffa44a741ec --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_15/index_tensor([6605, 6605])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a slightly upward angle, the camera pans steadily to the right, tracking a chase unfolding in microscopic detail. Deep in a sun-dappled forest, a butterfly, its wings radiating a soft, mysterious blue glow, relentlessly pursues a slowly falling cherry blossom petal. Its wings are covered in velvety black scales, delicate veins visible, while the edges are outlined with a brilliant blue from its inner glow. Each graceful, powerful flap leaves a fleeting, dreamlike streak of light in the air. Its target, the delicate pale pink cherry blossom petal, its edges slightly curled from dehydration, swirls gracefully in the golden shaft of sunlight streaming through the dense treetops. Carried by an invisible breeze, it drifts in an unpredictable path towards the dark green moss-covered ground. The butterfly precisely adjusts its flight path, swooping, circling, always maintaining a delicate distance from the petal, as if performing a silent dance. In the background, the ancient tree trunks and damp soil become blurred due to the depth of field, and only the fragmented and ever-changing spots of light are clearly visible. The blue glow of the butterfly itself is intertwined with the golden sunlight, rendering the entire picture with a tranquil, warm and vibrant magical atmosphere.", + "seed": 20966, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_15/index_6605_seed_20966_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04159388807602227, + "vlm_logprob": -0.04159388807602227, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.04159388807602227, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16526736319065094 + ], + "opposite_logprob_tokens": [ + -1.9152673482894897 + ], + "logprob_reward": -0.16526736319065094, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16526736319065094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2381395697593689, + 0.08395636826753616, + 0.04972544312477112, + 0.05748278275132179, + 0.30672508478164673, + 0.04239682853221893, + 0.02368081733584404, + 0.017660638317465782, + 0.13560613989830017, + 0.04462633281946182 + ], + "credit_map": [ + 0.014197790063917637, + 0.064171202480793, + 0.009217874146997929, + 0.02220085635781288, + 0.019934648647904396, + 0.017775969579815865, + 0.026508191600441933, + 0.06413304060697556, + 0.009029398672282696, + 0.010998432524502277, + 0.006743338890373707, + 0.006140568293631077, + 0.01580210030078888, + 0.007002541329711676, + 0.018669001758098602, + 0.00957097765058279, + 0.004378908313810825, + 0.008673157542943954, + 0.005089423153549433, + 0.006149437744170427, + 0.007143630180507898, + 0.004319339524954557, + 0.006440080236643553, + 0.007531463634222746, + 0.009121122770011425, + 0.004267659969627857, + 0.010252848267555237, + 0.010512101463973522, + 0.006059242412447929, + 0.004912926349788904, + 0.003961554262787104, + 0.00839532446116209, + 0.11187466233968735, + 0.0435318797826767, + 0.02300693653523922, + 0.020026909187436104, + 0.012930003926157951, + 0.011639999225735664, + 0.012078193947672844, + 0.07163650542497635, + 0.004639318212866783, + 0.004631379619240761, + 0.00911527220159769, + 0.009572232142090797, + 0.0028316229581832886, + 0.0034877436701208353, + 0.0025540811475366354, + 0.005565176252275705, + 0.0024469816125929356, + 0.0031788081396371126, + 0.005441928282380104, + 0.004554106388241053, + 0.0018860743148252368, + 0.0018113719997927547, + 0.0018940065056085587, + 0.0024675375316292048, + 0.0019843666814267635, + 0.0025235291104763746, + 0.0028903025668114424, + 0.003055630950257182, + 0.0019338381243869662, + 0.0015605303924530745, + 0.0018235233146697283, + 0.0018889147322624922, + 0.010048981755971909, + 0.02287082001566887, + 0.017264418303966522, + 0.022188134491443634, + 0.014698456972837448, + 0.02705119550228119, + 0.011066729202866554, + 0.010417385958135128, + 0.005165748298168182, + 0.007980328053236008, + 0.005499965976923704, + 0.006031801924109459, + 0.008512846194207668, + 0.004941741470247507, + 0.003655587090179324, + 0.0028383086901158094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.16526736319065094 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004659034311771393 + ], + "opposite_logprob_tokens": [ + -5.379659175872803 + ], + "logprob_reward": -0.004659034311771393, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004659034311771393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0413355715572834, + 0.049894217401742935, + 0.06267084926366806, + 0.03839828819036484, + 0.48177334666252136, + 0.016079913824796677, + 0.03649481385946274, + 0.15545843541622162, + 0.07638849318027496, + 0.041506052017211914 + ], + "credit_map": [ + 0.004241510760039091, + 0.004674175288528204, + 0.004007834475487471, + 0.0044367676600813866, + 0.003767359536141157, + 0.005249009933322668, + 0.005827914923429489, + 0.009130993857979774, + 0.006502776872366667, + 0.00759562011808157, + 0.006139596924185753, + 0.005208984948694706, + 0.006667538080364466, + 0.006496491841971874, + 0.005663150455802679, + 0.0056200590915977955, + 0.006030071061104536, + 0.0062577868811786175, + 0.007070065941661596, + 0.007179529406130314, + 0.0077829258516430855, + 0.014785511419177055, + 0.006413991097360849, + 0.007150964345782995, + 0.004561290610581636, + 0.00284442026168108, + 0.0044119092635810375, + 0.0038614424411207438, + 0.005959395784884691, + 0.004047792870551348, + 0.005080917850136757, + 0.007631117012351751, + 0.19880346953868866, + 0.06222419813275337, + 0.030250204727053642, + 0.019073303788900375, + 0.01597103476524353, + 0.015572724863886833, + 0.019292842596769333, + 0.12058559060096741, + 0.0019271980272606015, + 0.0019340269500389695, + 0.0022248660679906607, + 0.0023837967310100794, + 0.002530729630962014, + 0.001556248520500958, + 0.0018934415420517325, + 0.0016296064713969827, + 0.004019336309283972, + 0.004381381440907717, + 0.0031566780526190996, + 0.003736044280230999, + 0.007938065566122532, + 0.0048373667523264885, + 0.005370238795876503, + 0.00305570219643414, + 0.014716524630784988, + 0.020561078563332558, + 0.013226664625108242, + 0.015620391815900803, + 0.03658775985240936, + 0.01579955965280533, + 0.028380494564771652, + 0.010565956123173237, + 0.005223780404776335, + 0.011717651970684528, + 0.00428880425170064, + 0.006676698103547096, + 0.012020597234368324, + 0.01162153109908104, + 0.008511586114764214, + 0.016327839344739914, + 0.0035325868520885706, + 0.005303907208144665, + 0.005139301531016827, + 0.005790072958916426, + 0.006141323130577803, + 0.005144856404513121, + 0.006041935179382563, + 0.004412065260112286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.004659034311771393 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004642541520297527 + ], + "opposite_logprob_tokens": [ + -5.379642486572266 + ], + "logprob_reward": -0.004642541520297527, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004642541520297527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07317493110895157, + 0.07420402765274048, + 0.10136588662862778, + 0.0378730371594429, + 0.19995316863059998, + 0.043609339743852615, + 0.0423128567636013, + 0.051851317286491394, + 0.281261682510376, + 0.0943937599658966 + ], + "credit_map": [ + 0.00565356807783246, + 0.022851623594760895, + 0.00517376558855176, + 0.005093942396342754, + 0.007722557056695223, + 0.007623889483511448, + 0.00547825638204813, + 0.013577334582805634, + 0.009807473979890347, + 0.00936738308519125, + 0.010982470586895943, + 0.006792735308408737, + 0.006724036298692226, + 0.009668096899986267, + 0.011763871647417545, + 0.009097957983613014, + 0.008589938282966614, + 0.011539936065673828, + 0.012594151310622692, + 0.009312734007835388, + 0.009961227886378765, + 0.027765894308686256, + 0.009021388366818428, + 0.01258061733096838, + 0.0029138405807316303, + 0.003072542604058981, + 0.004218636080622673, + 0.004290516022592783, + 0.007155732251703739, + 0.005362283438444138, + 0.005746784154325724, + 0.005112703889608383, + 0.05690484493970871, + 0.022318467497825623, + 0.015780437737703323, + 0.0135110542178154, + 0.02287271060049534, + 0.010708524845540524, + 0.020723244175314903, + 0.03713388741016388, + 0.009242355823516846, + 0.00891068298369646, + 0.005074258893728256, + 0.005546921864151955, + 0.004579458385705948, + 0.0036768519785255194, + 0.0028337915427982807, + 0.0037450173404067755, + 0.0049904026091098785, + 0.006028028205037117, + 0.004779244773089886, + 0.005483362823724747, + 0.007928309962153435, + 0.004261752590537071, + 0.005293869413435459, + 0.003547884989529848, + 0.004424542188644409, + 0.0069051701575517654, + 0.005457769613713026, + 0.006336405407637358, + 0.009165268391370773, + 0.00511665316298604, + 0.009944243356585503, + 0.00450127013027668, + 0.027860520407557487, + 0.03490481898188591, + 0.016297094523906708, + 0.019786793738603592, + 0.034421157091856, + 0.05919787660241127, + 0.028389375656843185, + 0.060404058545827866, + 0.006136907264590263, + 0.01282226387411356, + 0.005223983433097601, + 0.007969798520207405, + 0.014398837462067604, + 0.01635149121284485, + 0.01698399893939495, + 0.014506480656564236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.004642541520297527 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036145609337836504 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.0036145609337836504, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036145609337836504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13676023483276367, + 0.1037936881184578, + 0.09045429527759552, + 0.06961361318826675, + 0.30733251571655273, + 0.06323307752609253, + 0.05663434788584709, + 0.016868555918335915, + 0.09575904160737991, + 0.05955057591199875 + ], + "credit_map": [ + 0.008980421349406242, + 0.0486333966255188, + 0.005074343644082546, + 0.00988850463181734, + 0.01420887466520071, + 0.008676527068018913, + 0.009098234586417675, + 0.032199926674366, + 0.011305387131869793, + 0.016439173370599747, + 0.00825224258005619, + 0.009221733547747135, + 0.020632797852158546, + 0.010478670708835125, + 0.011936062946915627, + 0.015527630224823952, + 0.010229541920125484, + 0.011988123878836632, + 0.009907743893563747, + 0.012068238109350204, + 0.015790024772286415, + 0.007717909291386604, + 0.01061946526169777, + 0.012133254669606686, + 0.009134959429502487, + 0.005720663350075483, + 0.00833628699183464, + 0.006955382414162159, + 0.012866988778114319, + 0.006118251010775566, + 0.00903888139873743, + 0.011442204937338829, + 0.12181173264980316, + 0.039676934480667114, + 0.017765624448657036, + 0.013578544370830059, + 0.014008572325110435, + 0.01145438477396965, + 0.013196368701756, + 0.07584036886692047, + 0.00786974374204874, + 0.006348423659801483, + 0.007671574596315622, + 0.006195545196533203, + 0.01434299536049366, + 0.007156091742217541, + 0.00739253219217062, + 0.006256178021430969, + 0.005186392460018396, + 0.00838997308164835, + 0.007561650592833757, + 0.007562847808003426, + 0.01081406231969595, + 0.006517508067190647, + 0.005732529330998659, + 0.00486938888207078, + 0.001560789067298174, + 0.00232527288608253, + 0.0028185283299535513, + 0.0031793967355042696, + 0.0020285965874791145, + 0.001644571777433157, + 0.0019531622529029846, + 0.0013582399114966393, + 0.007769839372485876, + 0.014698229730129242, + 0.008421012200415134, + 0.007901282049715519, + 0.01654249243438244, + 0.018195636570453644, + 0.01162331085652113, + 0.010607247240841389, + 0.00428213644772768, + 0.008131164126098156, + 0.006866438314318657, + 0.007984542287886143, + 0.008251385763287544, + 0.008188353851437569, + 0.010817065834999084, + 0.005029491614550352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0036145609337836504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029785940423607826 + ], + "opposite_logprob_tokens": [ + -3.5297858715057373 + ], + "logprob_reward": -0.029785940423607826, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029785940423607826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051834579557180405, + 0.2229626625776291, + 0.15420660376548767, + 0.04075949266552925, + 0.1015712171792984, + 0.030681518837809563, + 0.23610270023345947, + 0.030087728053331375, + 0.08693049848079681, + 0.04486295208334923 + ], + "credit_map": [ + 0.006438237149268389, + 0.009635991416871548, + 0.004213210195302963, + 0.005772111006081104, + 0.0043023694306612015, + 0.005105033982545137, + 0.00701509602367878, + 0.009352535009384155, + 0.028362493962049484, + 0.0226436760276556, + 0.01247564796358347, + 0.012307015247642994, + 0.07084257155656815, + 0.019604891538619995, + 0.028480306267738342, + 0.028246084228157997, + 0.012037969194352627, + 0.018787076696753502, + 0.013551398180425167, + 0.023149719461798668, + 0.026591943576931953, + 0.019142769277095795, + 0.021822556853294373, + 0.019123176112771034, + 0.0037056561559438705, + 0.0034466548822820187, + 0.005072779953479767, + 0.0061549656093120575, + 0.005444228183478117, + 0.003914837259799242, + 0.005904835183173418, + 0.007115540094673634, + 0.013299664482474327, + 0.01048307679593563, + 0.011020400561392307, + 0.012371884658932686, + 0.01810494065284729, + 0.009798092767596245, + 0.014454809948801994, + 0.012038358487188816, + 0.0030187834054231644, + 0.004247902426868677, + 0.00391496904194355, + 0.00406006770208478, + 0.004876234568655491, + 0.0033930449280887842, + 0.0037025208584964275, + 0.0034679975360631943, + 0.019878843799233437, + 0.03314173221588135, + 0.030399411916732788, + 0.04149135947227478, + 0.04193499684333801, + 0.0223504938185215, + 0.02488045021891594, + 0.02202543616294861, + 0.0028454929124563932, + 0.004324811976402998, + 0.0034064785577356815, + 0.0035476358607411385, + 0.005706415977329016, + 0.0026667583733797073, + 0.0054422952234745026, + 0.002147842664271593, + 0.007263957057148218, + 0.009272552095353603, + 0.008498129434883595, + 0.011077681556344032, + 0.01275611761957407, + 0.018707256764173508, + 0.010085786692798138, + 0.009269019588828087, + 0.005071392748504877, + 0.00897224061191082, + 0.005071217659860849, + 0.006306824274361134, + 0.006297476124018431, + 0.004361365921795368, + 0.0049845436587929726, + 0.003797898069024086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.029785940423607826 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward angle, the camera pans steadily to the right, tracking a chase unfolding in microscopic detail. Deep in a sun-dappled forest, a butterfly, its wings radiating a soft, mysterious blue glow, relentlessly pursues a slowly falling cherry blossom petal. Its wings are covered in velvety black scales, delicate veins visible, while the edges are outlined with a brilliant blue from its inner glow. Each graceful, powerful flap leaves a fleeting, dreamlike streak of light in the air. Its target, the delicate pale pink cherry blossom petal, its edges slightly curled from dehydration, swirls gracefully in the golden shaft of sunlight streaming through the dense treetops. Carried by an invisible breeze, it drifts in an unpredictable path towards the dark green moss-covered ground. The butterfly precisely adjusts its flight path, swooping, circling, always maintaining a delicate distance from the petal, as if performing a silent dance. In the background, the ancient tree trunks and damp soil become blurred due to the depth of field, and only the fragmented and ever-changing spots of light are clearly visible. The blue glow of the butterfly itself is intertwined with the golden sunlight, rendering the entire picture with a tranquil, warm and vibrant magical atmosphere.", + "seed": 20967, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_15/index_6605_seed_20967_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.027030961215496065, + "vlm_logprob": -0.027030961215496065, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.027030961215496065, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011726614087820053 + ], + "opposite_logprob_tokens": [ + -4.886726379394531 + ], + "logprob_reward": -0.011726614087820053, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011726614087820053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16543932259082794, + 0.11999998241662979, + 0.0711546391248703, + 0.1064915582537651, + 0.09439343959093094, + 0.12020937353372574, + 0.1114032045006752, + 0.09027466922998428, + 0.051141709089279175, + 0.06949204951524734 + ], + "credit_map": [ + 0.018658315762877464, + 0.025775106623768806, + 0.0245843306183815, + 0.011167893186211586, + 0.027571190148591995, + 0.022043868899345398, + 0.013187786564230919, + 0.022450843825936317, + 0.01356483530253172, + 0.016173848882317543, + 0.0178933497518301, + 0.013408059254288673, + 0.021928926929831505, + 0.01396851520985365, + 0.010649272240698338, + 0.012413176707923412, + 0.008906529285013676, + 0.009414022788405418, + 0.024754462763667107, + 0.006829605903476477, + 0.006826240103691816, + 0.005335231777280569, + 0.00280745723284781, + 0.006281092297285795, + 0.014577824622392654, + 0.01715957373380661, + 0.012701654806733131, + 0.018152307718992233, + 0.017038997262716293, + 0.006916120648384094, + 0.009503528475761414, + 0.010441556572914124, + 0.012208526954054832, + 0.010649336501955986, + 0.0132907060906291, + 0.012249696999788284, + 0.015468557365238667, + 0.009492756798863411, + 0.005194731522351503, + 0.01583912968635559, + 0.023634368553757668, + 0.019398711621761322, + 0.015831103548407555, + 0.007715897634625435, + 0.0163686815649271, + 0.010134795680642128, + 0.007661133073270321, + 0.019464684650301933, + 0.02338520623743534, + 0.009762544184923172, + 0.017871012911200523, + 0.010229279287159443, + 0.01055492740124464, + 0.011619551107287407, + 0.012368782423436642, + 0.01561190839856863, + 0.014318373054265976, + 0.010177340358495712, + 0.013062424026429653, + 0.008289957419037819, + 0.008489012718200684, + 0.015362685546278954, + 0.009050059132277966, + 0.011524820700287819, + 0.007343585602939129, + 0.00918322429060936, + 0.005562409292906523, + 0.007270656060427427, + 0.005508407484740019, + 0.00555502250790596, + 0.006270332727581263, + 0.004448072053492069, + 0.007603456266224384, + 0.017683126032352448, + 0.008026116527616978, + 0.006763235665857792, + 0.0077448333613574505, + 0.007243098691105843, + 0.00860424805432558, + 0.005823930725455284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.011726614087820053 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011091482825577259 + ], + "opposite_logprob_tokens": [ + -4.511091709136963 + ], + "logprob_reward": -0.011091482825577259, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011091482825577259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1014685183763504, + 0.10478433966636658, + 0.384967565536499, + 0.06827875226736069, + 0.06716614961624146, + 0.07039173692464828, + 0.06071145087480545, + 0.050788763910532, + 0.0383664108812809, + 0.05307632312178612 + ], + "credit_map": [ + 0.01152031496167183, + 0.012463416904211044, + 0.013371460139751434, + 0.00918574444949627, + 0.01990210823714733, + 0.00880408100783825, + 0.008656415157020092, + 0.017564978450536728, + 0.029777368530631065, + 0.01338919810950756, + 0.01682945340871811, + 0.008710402995347977, + 0.011785348877310753, + 0.013452037237584591, + 0.005454110447317362, + 0.005386418662965298, + 0.04652286320924759, + 0.02758037857711315, + 0.09056171029806137, + 0.026717156171798706, + 0.06720070540904999, + 0.07086659222841263, + 0.014006499201059341, + 0.04151167348027229, + 0.010789173655211926, + 0.01078698318451643, + 0.008375616744160652, + 0.012443413026630878, + 0.009562033228576183, + 0.005467191804200411, + 0.005407385528087616, + 0.0054469541646540165, + 0.007449198979884386, + 0.012193056754767895, + 0.00869573000818491, + 0.011334813199937344, + 0.008954087272286415, + 0.00484925601631403, + 0.004989439155906439, + 0.008700571954250336, + 0.009260924533009529, + 0.017146233469247818, + 0.006887903902679682, + 0.008100880309939384, + 0.009079428389668465, + 0.005560093559324741, + 0.006021231412887573, + 0.008335048332810402, + 0.008902044966816902, + 0.004747490398585796, + 0.007875382900238037, + 0.004554819781333208, + 0.0055032409727573395, + 0.00978761538863182, + 0.00936937890946865, + 0.009971478022634983, + 0.007962843403220177, + 0.006180854514241219, + 0.007409308105707169, + 0.0052865175530314445, + 0.004940775688737631, + 0.006797911133617163, + 0.005962160881608725, + 0.006248392630368471, + 0.004994348622858524, + 0.007213148288428783, + 0.005054960492998362, + 0.005490696057677269, + 0.0049581388011574745, + 0.0029885631520301104, + 0.0034559841733425856, + 0.004210571758449078, + 0.007973643951117992, + 0.009600801393389702, + 0.005848041269928217, + 0.004688221029937267, + 0.005567876622080803, + 0.007608989719301462, + 0.0069935680367052555, + 0.004795178305357695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.011091482825577259 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001681625843048 + ], + "opposite_logprob_tokens": [ + -2.695016860961914 + ], + "logprob_reward": -0.07001681625843048, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001681625843048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.145997554063797, + 0.10097434371709824, + 0.29954689741134644, + 0.06900925934314728, + 0.04071298614144325, + 0.07759859412908554, + 0.07906121015548706, + 0.09020712226629257, + 0.03246735408902168, + 0.06442461162805557 + ], + "credit_map": [ + 0.016904260963201523, + 0.021013356745243073, + 0.019439250230789185, + 0.013393165543675423, + 0.02573288604617119, + 0.011810612864792347, + 0.012157939374446869, + 0.025546090677380562, + 0.016854459419846535, + 0.018309107050299644, + 0.009876948781311512, + 0.011607489548623562, + 0.012816617265343666, + 0.017924871295690536, + 0.005988405551761389, + 0.007596444804221392, + 0.03807426989078522, + 0.021834559738636017, + 0.07861599326133728, + 0.015870360657572746, + 0.049691639840602875, + 0.05451684072613716, + 0.01073623076081276, + 0.030207017436623573, + 0.010403892025351524, + 0.010091971606016159, + 0.00868736021220684, + 0.011066656559705734, + 0.00889224000275135, + 0.006440409924834967, + 0.006451241206377745, + 0.006975491531193256, + 0.005126503296196461, + 0.004507144447416067, + 0.005987201351672411, + 0.004569621291011572, + 0.0051284548826515675, + 0.003962809219956398, + 0.004564093891531229, + 0.006867157761007547, + 0.01095485594123602, + 0.014917890541255474, + 0.009409396909177303, + 0.007323678582906723, + 0.010989346541464329, + 0.006046930328011513, + 0.00728178396821022, + 0.010674714110791683, + 0.009645998477935791, + 0.007570766378194094, + 0.010675381869077682, + 0.007273208349943161, + 0.008619685657322407, + 0.006929723545908928, + 0.015543103218078613, + 0.012803338468074799, + 0.014788638800382614, + 0.010565097443759441, + 0.012883536517620087, + 0.008754369802772999, + 0.007872244343161583, + 0.0129702128469944, + 0.009656722657382488, + 0.012716300785541534, + 0.0044220611453056335, + 0.005867467727512121, + 0.003510901937261224, + 0.004889539908617735, + 0.0038318182341754436, + 0.0038987002335488796, + 0.00314179714769125, + 0.002905068686231971, + 0.008661306463181973, + 0.012851132079958916, + 0.00781854148954153, + 0.006254217121750116, + 0.00889785960316658, + 0.007949764840304852, + 0.005997287575155497, + 0.00599450059235096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.07001681625843048 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253464724868536 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.01253464724868536, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253464724868536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11623967438936234, + 0.06348884105682373, + 0.2897244691848755, + 0.06008606404066086, + 0.10793721675872803, + 0.050596754997968674, + 0.11851046234369278, + 0.07735510915517807, + 0.06737774610519409, + 0.04868365824222565 + ], + "credit_map": [ + 0.013883483596146107, + 0.016453709453344345, + 0.014752409420907497, + 0.009830046445131302, + 0.014162551611661911, + 0.012349738739430904, + 0.009992415085434914, + 0.02481532096862793, + 0.007974551990628242, + 0.0143128652125597, + 0.008435294032096863, + 0.011407415382564068, + 0.008844299241900444, + 0.004148878622800112, + 0.004392360802739859, + 0.003973177168518305, + 0.03530528396368027, + 0.01949305273592472, + 0.07071223109960556, + 0.0211530439555645, + 0.048190318048000336, + 0.05586308613419533, + 0.010094278492033482, + 0.028913162648677826, + 0.01666666753590107, + 0.00640345923602581, + 0.014463302679359913, + 0.00564022408798337, + 0.005086406134068966, + 0.00430516479536891, + 0.0032740335445851088, + 0.004246803931891918, + 0.008468172512948513, + 0.018892783671617508, + 0.013955257833003998, + 0.018816499039530754, + 0.01395235862582922, + 0.00984788779169321, + 0.01097069401293993, + 0.013033568859100342, + 0.006435171701014042, + 0.008908574469387531, + 0.00700005516409874, + 0.004710547626018524, + 0.007698508445173502, + 0.004055039491504431, + 0.003955945838242769, + 0.007832911796867847, + 0.010952520184218884, + 0.009757941588759422, + 0.014851135201752186, + 0.010155986063182354, + 0.010485564358532429, + 0.02361663617193699, + 0.01457545068114996, + 0.02411523275077343, + 0.00911271758377552, + 0.010270847007632256, + 0.008351650089025497, + 0.008590576238930225, + 0.008446531370282173, + 0.013449511490762234, + 0.005356867332011461, + 0.013776409439742565, + 0.009934179484844208, + 0.007006221916526556, + 0.013264203444123268, + 0.006639485713094473, + 0.004828156437724829, + 0.012545711360871792, + 0.006394200026988983, + 0.006765585858374834, + 0.006579057313501835, + 0.006008098367601633, + 0.005874021910130978, + 0.005411377642303705, + 0.00520595395937562, + 0.007512837182730436, + 0.0070440517738461494, + 0.005048265215009451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.01253464724868536 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029785245656967163 + ], + "opposite_logprob_tokens": [ + -3.52978515625 + ], + "logprob_reward": -0.029785245656967163, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029785245656967163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2022848278284073, + 0.040962349623441696, + 0.3293844163417816, + 0.021543730050325394, + 0.04070133715867996, + 0.05484791100025177, + 0.07142180949449539, + 0.11285652220249176, + 0.05146010220050812, + 0.07453698664903641 + ], + "credit_map": [ + 0.02621668204665184, + 0.020622290670871735, + 0.029866689816117287, + 0.013714331202208996, + 0.03188584744930267, + 0.017241112887859344, + 0.01713789440691471, + 0.045599982142448425, + 0.0057372599840164185, + 0.005763146094977856, + 0.006133122369647026, + 0.0037282544653862715, + 0.007748092990368605, + 0.004521267022937536, + 0.002597535727545619, + 0.004733672831207514, + 0.04081494361162186, + 0.023304475471377373, + 0.08138438314199448, + 0.019189191982150078, + 0.05786653235554695, + 0.06685172766447067, + 0.010918084532022476, + 0.029055094346404076, + 0.003316415473818779, + 0.0032338793389499187, + 0.0030509447678923607, + 0.0022568153217434883, + 0.0019045493099838495, + 0.0016125560505315661, + 0.001872056513093412, + 0.004296515602618456, + 0.00416187196969986, + 0.006305128801614046, + 0.006386297754943371, + 0.006312578916549683, + 0.005943980533629656, + 0.0040597328916192055, + 0.003032074309885502, + 0.004499673843383789, + 0.0073845963925123215, + 0.012246370315551758, + 0.005439919885247946, + 0.0047818748280406, + 0.010423880070447922, + 0.004591192584484816, + 0.003590641776099801, + 0.006389439105987549, + 0.0102987764403224, + 0.003397209569811821, + 0.0089549720287323, + 0.004561806097626686, + 0.003929079510271549, + 0.020766563713550568, + 0.003157207975164056, + 0.01635619066655636, + 0.016090597957372665, + 0.014667069539427757, + 0.01599210500717163, + 0.010893411934375763, + 0.011034800671041012, + 0.01742819882929325, + 0.012019812129437923, + 0.014730527065694332, + 0.006621540989726782, + 0.005918031092733145, + 0.005288373213261366, + 0.0053658317774534225, + 0.004013311583548784, + 0.011808977462351322, + 0.004685590509325266, + 0.007758451160043478, + 0.011099216528236866, + 0.011400203220546246, + 0.01174680795520544, + 0.007192420307546854, + 0.006474820896983147, + 0.009463268332183361, + 0.008660845458507538, + 0.00849941000342369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.029785245656967163 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward angle, the camera pans steadily to the right, tracking a chase unfolding in microscopic detail. Deep in a sun-dappled forest, a butterfly, its wings radiating a soft, mysterious blue glow, relentlessly pursues a slowly falling cherry blossom petal. Its wings are covered in velvety black scales, delicate veins visible, while the edges are outlined with a brilliant blue from its inner glow. Each graceful, powerful flap leaves a fleeting, dreamlike streak of light in the air. Its target, the delicate pale pink cherry blossom petal, its edges slightly curled from dehydration, swirls gracefully in the golden shaft of sunlight streaming through the dense treetops. Carried by an invisible breeze, it drifts in an unpredictable path towards the dark green moss-covered ground. The butterfly precisely adjusts its flight path, swooping, circling, always maintaining a delicate distance from the petal, as if performing a silent dance. In the background, the ancient tree trunks and damp soil become blurred due to the depth of field, and only the fragmented and ever-changing spots of light are clearly visible. The blue glow of the butterfly itself is intertwined with the golden sunlight, rendering the entire picture with a tranquil, warm and vibrant magical atmosphere.", + "seed": 20968, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_15/index_6605_seed_20968_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.013197119906544685, + "vlm_logprob": -0.013197119906544685, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.013197119906544685, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011378396302461624 + ], + "opposite_logprob_tokens": [ + -5.011378288269043 + ], + "logprob_reward": -0.011378396302461624, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.011378396302461624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.015324573032557964, + 0.0075914631597697735, + 0.005026370752602816, + 0.010129415430128574, + 0.004742533899843693, + 0.0035397724714130163, + 0.010549193248152733, + 0.006107565015554428, + 0.008485355414450169, + 0.9285037517547607 + ], + "credit_map": [ + 0.0008761656354181468, + 0.002211852464824915, + 0.0011356299510225654, + 0.0014033438637852669, + 0.0023903287947177887, + 0.0019370076479390264, + 0.00309568434022367, + 0.0022745609749108553, + 0.0006053555989637971, + 0.0012139161117374897, + 0.0005885651335120201, + 0.0005647827056236565, + 0.0012445602333173156, + 0.0012065611081197858, + 0.0010073967278003693, + 0.0011603262973949313, + 0.0004944565589539707, + 0.0005031775799579918, + 0.0004605343274306506, + 0.0005802615778520703, + 0.0005363902891986072, + 0.000571858196053654, + 0.0011556129902601242, + 0.0007240795530378819, + 0.0007660359842702746, + 0.0009280662052333355, + 0.000550493597984314, + 0.0006829319754615426, + 0.001981091918423772, + 0.002007880248129368, + 0.0019050415139645338, + 0.001307875500060618, + 0.0004857923195231706, + 0.0003385014715604484, + 0.0008447333821095526, + 0.00038863869849592447, + 0.0004698000557255, + 0.0009274018811993301, + 0.0005204525659792125, + 0.000767213583458215, + 0.000345915206708014, + 0.00044090344454161823, + 0.0006449361681006849, + 0.00034967801184393466, + 0.00030652116402052343, + 0.0005019818781875074, + 0.00042848510202020407, + 0.0005213517579250038, + 0.001085414900444448, + 0.0026588791515678167, + 0.0011318179313093424, + 0.000770134967751801, + 0.0013523459201678634, + 0.001541820471175015, + 0.000935834483243525, + 0.0010729468194767833, + 0.0008018111693672836, + 0.00048793121823109686, + 0.0012274913024157286, + 0.00043896082206629217, + 0.0004968080902472138, + 0.0008651106036268175, + 0.0008243798511102796, + 0.0009650725987739861, + 0.000476250919746235, + 0.0002822894894052297, + 0.000418856245232746, + 0.0003610418934840709, + 0.00038715655682608485, + 0.0032841546926647425, + 0.0004088391433469951, + 0.002866766881197691, + 0.028333699330687523, + 0.009315957315266132, + 0.01213910523802042, + 0.007876954972743988, + 0.02860931120812893, + 0.41825518012046814, + 0.01974843628704548, + 0.4042251408100128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.011378396302461624 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007647404447197914 + ], + "opposite_logprob_tokens": [ + -4.882647514343262 + ], + "logprob_reward": -0.007647404447197914, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007647404447197914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03420526161789894, + 0.014263512566685677, + 0.009934688918292522, + 0.01571355015039444, + 0.010991448536515236, + 0.03329591080546379, + 0.020477524027228355, + 0.015696262940764427, + 0.01693943701684475, + 0.828482449054718 + ], + "credit_map": [ + 0.004274929407984018, + 0.0035870627034455538, + 0.005102003924548626, + 0.003797852201387286, + 0.005360420327633619, + 0.0033606889192014933, + 0.004205853212624788, + 0.0045164539478719234, + 0.001336862100288272, + 0.0021919666323810816, + 0.0014561295974999666, + 0.0010982955573126674, + 0.0018238859483972192, + 0.002015088452026248, + 0.0022168918512761593, + 0.0021243931259959936, + 0.000984297483228147, + 0.0012877500848844647, + 0.0009399154805578291, + 0.0020233362447470427, + 0.0008512389031238854, + 0.0013548576971516013, + 0.0011588033521547914, + 0.0013344892067834735, + 0.0013921698555350304, + 0.0016764213796705008, + 0.0017144308658316731, + 0.0012755639618262649, + 0.00215313583612442, + 0.002932108473032713, + 0.0020146025344729424, + 0.002555117942392826, + 0.001146348542533815, + 0.001258435077033937, + 0.0017648626817390323, + 0.0008550216443836689, + 0.001280339783988893, + 0.002053257543593645, + 0.0008968992624431849, + 0.0017362841172143817, + 0.0017384991515427828, + 0.0047386400401592255, + 0.002883564680814743, + 0.0025385688059031963, + 0.005757994018495083, + 0.00461895577609539, + 0.002368277870118618, + 0.008651410229504108, + 0.002391691319644451, + 0.004637578036636114, + 0.0025411942042410374, + 0.001520380494184792, + 0.0026783484499901533, + 0.002817280124872923, + 0.0018561298493295908, + 0.002034921897575259, + 0.0016391376266255975, + 0.002408168278634548, + 0.002649822039529681, + 0.0011348415864631534, + 0.002003403613343835, + 0.002025466412305832, + 0.0020823010709136724, + 0.0017531216144561768, + 0.0008167595369741321, + 0.0009157463209703565, + 0.001047885394655168, + 0.0005851045716553926, + 0.001242263475432992, + 0.005735449492931366, + 0.0010192182380706072, + 0.005577010102570057, + 0.025374488905072212, + 0.008293644525110722, + 0.01074910257011652, + 0.0070042419247329235, + 0.02555890567600727, + 0.37362709641456604, + 0.017526328563690186, + 0.360348641872406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.007647404447197914 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540768831968307 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540768831968307, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540768831968307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03983655571937561, + 0.032117411494255066, + 0.01577136479318142, + 0.03090437687933445, + 0.009787512943148613, + 0.01949433796107769, + 0.017157141119241714, + 0.0059030381962656975, + 0.016926359385252, + 0.8121019005775452 + ], + "credit_map": [ + 0.004805598873645067, + 0.004395992495119572, + 0.005556294694542885, + 0.004488413687795401, + 0.006870551034808159, + 0.004156335722655058, + 0.004009610041975975, + 0.0055537596344947815, + 0.0020406381227076054, + 0.005176829639822245, + 0.003338051959872246, + 0.002609041053801775, + 0.004331684205681086, + 0.00486291479319334, + 0.005071035120636225, + 0.004687216132879257, + 0.0012782602570950985, + 0.0016733967931941152, + 0.001172250253148377, + 0.0028781781438738108, + 0.001410942175425589, + 0.0021159625612199306, + 0.002920488826930523, + 0.002321886830031872, + 0.003460149746388197, + 0.003226407105103135, + 0.0024680455680936575, + 0.0023379207123070955, + 0.0033351401798427105, + 0.004635168705135584, + 0.007110452279448509, + 0.004331093281507492, + 0.001218845252878964, + 0.0008794672321528196, + 0.0016149075236171484, + 0.0007984974654391408, + 0.0010002386989071965, + 0.0017857893835753202, + 0.0011391323059797287, + 0.0013506344985216856, + 0.0013279799604788423, + 0.002694731578230858, + 0.0022010402753949165, + 0.0016672281781211495, + 0.0035209625493735075, + 0.002445153659209609, + 0.0013796063140034676, + 0.004257636144757271, + 0.0018566858489066362, + 0.0029031429439783096, + 0.001991279423236847, + 0.0013294256059452891, + 0.0018320046365261078, + 0.0023688662331551313, + 0.0027779710944741964, + 0.002097765216603875, + 0.0005663428455591202, + 0.0009078492876142263, + 0.000832852441817522, + 0.0004558513464871794, + 0.0008548430632799864, + 0.0007960977382026613, + 0.0007311995723284781, + 0.0007580016972497106, + 0.0008901089895516634, + 0.001104083494283259, + 0.001091727172024548, + 0.0006937152938917279, + 0.0012117072474211454, + 0.005883151665329933, + 0.0008334866142831743, + 0.005218378733843565, + 0.024785861372947693, + 0.008148642256855965, + 0.010519426316022873, + 0.006892818957567215, + 0.025073790922760963, + 0.36601361632347107, + 0.017145080491900444, + 0.3535226881504059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.012540768831968307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634710028767586 + ], + "opposite_logprob_tokens": [ + -5.379634857177734 + ], + "logprob_reward": -0.004634710028767586, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634710028767586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01362779550254345, + 0.0068559469655156136, + 0.0044745709747076035, + 0.014118138700723648, + 0.010419739410281181, + 0.0057120053097605705, + 0.049286071211099625, + 0.00587387615814805, + 0.04000655189156532, + 0.8496253490447998 + ], + "credit_map": [ + 0.0015001025749370456, + 0.001476304605603218, + 0.0019115415634587407, + 0.001617714180611074, + 0.0020352057181298733, + 0.0014740433543920517, + 0.002025842433795333, + 0.0015870403731241822, + 0.001505922176875174, + 0.0009700835798867047, + 0.0007622254779562354, + 0.00048602226888760924, + 0.0006219901842996478, + 0.000845343223772943, + 0.0006434364477172494, + 0.001020922907628119, + 0.0005551103968173265, + 0.0005367519333958626, + 0.0003804996667895466, + 0.0007483343943022192, + 0.00036528336931951344, + 0.0004895052989013493, + 0.0007502730004489422, + 0.0006488120416179299, + 0.0011913954513147473, + 0.0015177145833149552, + 0.0010820080060511827, + 0.0009428024059161544, + 0.0022943599615246058, + 0.0027908766642212868, + 0.002456564921885729, + 0.0018424150766804814, + 0.001092912396416068, + 0.0015938590513542295, + 0.0014922341797500849, + 0.0009307583677582443, + 0.0012716819765046239, + 0.0015745034907013178, + 0.0010365743655711412, + 0.0014272137777879834, + 0.000693201320245862, + 0.000795738713350147, + 0.0011434259358793497, + 0.0004942762316204607, + 0.0005718296742998064, + 0.0007538970676250756, + 0.000526256684679538, + 0.0007333788089454174, + 0.0060647777281701565, + 0.010517233051359653, + 0.007028201129287481, + 0.003919621463865042, + 0.006408174522221088, + 0.006492301356047392, + 0.0034707924351096153, + 0.00538496533408761, + 0.0005414763581939042, + 0.0009436944383196533, + 0.0010203119600191712, + 0.000489255937281996, + 0.0006307157455012202, + 0.0008811994921416044, + 0.0006852761725895107, + 0.0006819452974013984, + 0.0014985426096245646, + 0.001668697688728571, + 0.0021448314655572176, + 0.0016167660942301154, + 0.0018741925014182925, + 0.015131203457713127, + 0.00163347611669451, + 0.014438838697969913, + 0.026013929396867752, + 0.008530676364898682, + 0.010991259478032589, + 0.007266902830451727, + 0.0261270422488451, + 0.3829159736633301, + 0.017993435263633728, + 0.3697860836982727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.004634710028767586 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029784319922327995 + ], + "opposite_logprob_tokens": [ + -3.5297842025756836 + ], + "logprob_reward": -0.029784319922327995, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029784319922327995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024086467921733856, + 0.01016957126557827, + 0.010598148219287395, + 0.01044590212404728, + 0.0071734595112502575, + 0.006129460409283638, + 0.008887642063200474, + 0.008801577612757683, + 0.04002147912979126, + 0.8736861944198608 + ], + "credit_map": [ + 0.003354573156684637, + 0.0024011367931962013, + 0.0037075490690767765, + 0.0024974322877824306, + 0.0034538849722594023, + 0.002201494062319398, + 0.00341302459128201, + 0.0030573720578104258, + 0.0022746557369828224, + 0.0010539047652855515, + 0.001394597813487053, + 0.0007865671650506556, + 0.0009508946095593274, + 0.0012095468118786812, + 0.0010213068453595042, + 0.0014780974015593529, + 0.000589677132666111, + 0.0008155149407684803, + 0.0005031282780691981, + 0.0015678006457164884, + 0.0011221946915611625, + 0.0013565251138061285, + 0.0030027017928659916, + 0.0016406053910031915, + 0.0010032868012785912, + 0.0008650381350889802, + 0.0005801226361654699, + 0.0006370997871272266, + 0.002272195415571332, + 0.0020696602296084166, + 0.0012366396840661764, + 0.001781859784387052, + 0.0007462044013664126, + 0.0010344897164031863, + 0.0009913290850818157, + 0.0006070821546018124, + 0.0009557087905704975, + 0.001041074050590396, + 0.000734904664568603, + 0.0010626668808981776, + 0.0005173236131668091, + 0.0007574456394650042, + 0.0010991087183356285, + 0.0005957056418992579, + 0.0007195249781943858, + 0.000746786710806191, + 0.0007571097812615335, + 0.0009364549769088626, + 0.000584665103815496, + 0.0014257464790716767, + 0.0011798887280747294, + 0.0006815727101638913, + 0.0008381669758819044, + 0.0013145840493962169, + 0.0015030117938295007, + 0.001360006514005363, + 0.0012006955221295357, + 0.0015785645227879286, + 0.001080560963600874, + 0.0005504774744622409, + 0.0012652537552639842, + 0.0010369500378146768, + 0.0011077751405537128, + 0.000981300137937069, + 0.0014911958714947104, + 0.0012669529533013701, + 0.001916328677907586, + 0.0012656071921810508, + 0.0013614956988021731, + 0.016296420246362686, + 0.0012704516993835568, + 0.015153027139604092, + 0.02674235589802265, + 0.008762425743043423, + 0.01147992629557848, + 0.007386423647403717, + 0.027074595913290977, + 0.3936586081981659, + 0.01860770583152771, + 0.37997424602508545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.029784319922327995 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward angle, the camera pans steadily to the right, tracking a chase unfolding in microscopic detail. Deep in a sun-dappled forest, a butterfly, its wings radiating a soft, mysterious blue glow, relentlessly pursues a slowly falling cherry blossom petal. Its wings are covered in velvety black scales, delicate veins visible, while the edges are outlined with a brilliant blue from its inner glow. Each graceful, powerful flap leaves a fleeting, dreamlike streak of light in the air. Its target, the delicate pale pink cherry blossom petal, its edges slightly curled from dehydration, swirls gracefully in the golden shaft of sunlight streaming through the dense treetops. Carried by an invisible breeze, it drifts in an unpredictable path towards the dark green moss-covered ground. The butterfly precisely adjusts its flight path, swooping, circling, always maintaining a delicate distance from the petal, as if performing a silent dance. In the background, the ancient tree trunks and damp soil become blurred due to the depth of field, and only the fragmented and ever-changing spots of light are clearly visible. The blue glow of the butterfly itself is intertwined with the golden sunlight, rendering the entire picture with a tranquil, warm and vibrant magical atmosphere.", + "seed": 20969, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_15/index_6605_seed_20969_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012718737497925759, + "vlm_logprob": -0.012718737497925759, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.012718737497925759, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01849079690873623 + ], + "opposite_logprob_tokens": [ + -4.393490791320801 + ], + "logprob_reward": -0.01849079690873623, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01849079690873623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08994816243648529, + 0.05021151900291443, + 0.13813187181949615, + 0.05046684667468071, + 0.08161651343107224, + 0.041137952357530594, + 0.15566326677799225, + 0.2512383460998535, + 0.05461671203374863, + 0.0869688019156456 + ], + "credit_map": [ + 0.008240362629294395, + 0.0064348564483225346, + 0.0073202792555093765, + 0.007146633230149746, + 0.008281400427222252, + 0.02413681335747242, + 0.010498813353478909, + 0.017889000475406647, + 0.0048158071003854275, + 0.005456606857478619, + 0.003017840441316366, + 0.004996733274310827, + 0.009354664944112301, + 0.007661337032914162, + 0.007675112225115299, + 0.007233417592942715, + 0.011487218551337719, + 0.012206422165036201, + 0.009842299856245518, + 0.012081993743777275, + 0.02804418094456196, + 0.022455329075455666, + 0.01826493628323078, + 0.023749500513076782, + 0.005039981100708246, + 0.005366469733417034, + 0.005609630607068539, + 0.004775319714099169, + 0.008380119688808918, + 0.007980970665812492, + 0.004996851552277803, + 0.00831750687211752, + 0.008668963797390461, + 0.00617432314902544, + 0.01577773503959179, + 0.006957825273275375, + 0.01091022789478302, + 0.016882402822375298, + 0.005412544123828411, + 0.010832492262125015, + 0.003884842386469245, + 0.004396334756165743, + 0.007077337708324194, + 0.004182020202279091, + 0.004551155027002096, + 0.006949273403733969, + 0.004046881105750799, + 0.006050108466297388, + 0.025684956461191177, + 0.026956336572766304, + 0.01995532214641571, + 0.022479994222521782, + 0.021010154858231544, + 0.01790836825966835, + 0.009491842240095139, + 0.012176279909908772, + 0.016927340999245644, + 0.014845076017081738, + 0.07485731691122055, + 0.015218797139823437, + 0.01451794896274805, + 0.09026415646076202, + 0.009792134165763855, + 0.014815573580563068, + 0.006977941375225782, + 0.00634556682780385, + 0.008329244330525398, + 0.009361624717712402, + 0.005480015184730291, + 0.006692932918667793, + 0.005392717197537422, + 0.006036668550223112, + 0.012419895268976688, + 0.017983045428991318, + 0.013497409410774708, + 0.016281070187687874, + 0.006239641457796097, + 0.006501808762550354, + 0.006068340502679348, + 0.00797758623957634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.01849079690873623 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005966471508145332 + ], + "opposite_logprob_tokens": [ + -5.130966663360596 + ], + "logprob_reward": -0.005966471508145332, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005966471508145332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06032863259315491, + 0.02937893196940422, + 0.03133237361907959, + 0.016797536984086037, + 0.04846447706222534, + 0.0146248210221529, + 0.26082202792167664, + 0.4759112298488617, + 0.032214920967817307, + 0.030125059187412262 + ], + "credit_map": [ + 0.0039718858897686005, + 0.0026349930558353662, + 0.00338862556964159, + 0.004056333564221859, + 0.006554923951625824, + 0.023548809811472893, + 0.0025980554055422544, + 0.013575006276369095, + 0.004075400531291962, + 0.0024166270159184933, + 0.0032607505563646555, + 0.0026831296272575855, + 0.004079712089151144, + 0.003998312167823315, + 0.003957378212362528, + 0.00490762060508132, + 0.004257441498339176, + 0.0027383274864405394, + 0.0036524233873933554, + 0.00288120424374938, + 0.005312300752848387, + 0.003902806667611003, + 0.004491883330047131, + 0.004095986019819975, + 0.0018284699181094766, + 0.002171199070289731, + 0.002101195277646184, + 0.001851698849350214, + 0.0025857349392026663, + 0.0031126763205975294, + 0.0011848760768771172, + 0.0019616850186139345, + 0.003632783656939864, + 0.0034996673930436373, + 0.013616572134196758, + 0.0045664082281291485, + 0.005176751408725977, + 0.013565893284976482, + 0.0016088620759546757, + 0.0027975416742265224, + 0.0014688041992485523, + 0.0018667139811441302, + 0.0027965998742729425, + 0.001678853644989431, + 0.0017245430499315262, + 0.0023376471363008022, + 0.0013173373881727457, + 0.0014343204675242305, + 0.043099794536828995, + 0.04297802969813347, + 0.03391407057642937, + 0.034639623016119, + 0.03529997169971466, + 0.03385221213102341, + 0.015140343457460403, + 0.021897979080677032, + 0.02650913968682289, + 0.022277334704995155, + 0.1551443636417389, + 0.023337943479418755, + 0.024481430649757385, + 0.19559380412101746, + 0.01300905179232359, + 0.015558171086013317, + 0.006021150387823582, + 0.004688636399805546, + 0.004197758622467518, + 0.003485831432044506, + 0.004049483686685562, + 0.004398702643811703, + 0.003223934443667531, + 0.0021494273096323013, + 0.003518967656418681, + 0.004859468899667263, + 0.004206446465104818, + 0.0042404187843203545, + 0.003537949174642563, + 0.0033872134517878294, + 0.0033243284560739994, + 0.0030502676963806152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.005966471508145332 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099537618458271 + ], + "opposite_logprob_tokens": [ + -5.504099369049072 + ], + "logprob_reward": -0.004099537618458271, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099537618458271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04898963123559952, + 0.32454612851142883, + 0.12140081822872162, + 0.011802601628005505, + 0.026958594098687172, + 0.010803768411278725, + 0.023441141471266747, + 0.3886958956718445, + 0.023518631234765053, + 0.01984279975295067 + ], + "credit_map": [ + 0.003752355230972171, + 0.0034703968558460474, + 0.002906601410359144, + 0.004014245234429836, + 0.006637500133365393, + 0.01410391740500927, + 0.004935428034514189, + 0.009169189259409904, + 0.033538948744535446, + 0.03884851187467575, + 0.018134286627173424, + 0.028759228065609932, + 0.0680827647447586, + 0.03708885237574577, + 0.060373615473508835, + 0.03971990942955017, + 0.014392309822142124, + 0.008694920688867569, + 0.012939320877194405, + 0.011217164807021618, + 0.02034064568579197, + 0.019627565518021584, + 0.022258736193180084, + 0.011930156499147415, + 0.0009938273578882217, + 0.0010752187808975577, + 0.0017639535944908857, + 0.0009924095356836915, + 0.0016538433264940977, + 0.0033665800001472235, + 0.0006350223557092249, + 0.0013217462692409754, + 0.0026992959901690483, + 0.003709537675604224, + 0.0046253991313278675, + 0.0032132111955434084, + 0.004238687455654144, + 0.00415576109662652, + 0.0021870203781872988, + 0.0021296811755746603, + 0.0011526794405654073, + 0.001256630290299654, + 0.0018928818171843886, + 0.0007899519987404346, + 0.0012192233698442578, + 0.0022044966462999582, + 0.000727807404473424, + 0.0015600972110405564, + 0.0022608707658946514, + 0.0038946359418332577, + 0.003640793263912201, + 0.003597429022192955, + 0.0035107163712382317, + 0.002328417031094432, + 0.0021569570526480675, + 0.002051322255283594, + 0.02087072655558586, + 0.0185793936252594, + 0.12299645692110062, + 0.017278626561164856, + 0.020653460174798965, + 0.156046524643898, + 0.012896857224404812, + 0.01937384530901909, + 0.002391716232523322, + 0.0027650848496705294, + 0.0038522258400917053, + 0.0038248237688094378, + 0.002458935836330056, + 0.002999663818627596, + 0.0018519117729738355, + 0.00337426969781518, + 0.0029438233468681574, + 0.00451712217181921, + 0.003008168190717697, + 0.00253253523260355, + 0.0016421190230175853, + 0.0019447862869128585, + 0.0015571630792692304, + 0.0016970816068351269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 2 + ], + "selected_reward": -0.004099537618458271 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250592716038227 + ], + "opposite_logprob_tokens": [ + -5.255250453948975 + ], + "logprob_reward": -0.005250592716038227, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250592716038227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15611228346824646, + 0.0740327462553978, + 0.03192899748682976, + 0.03540722653269768, + 0.23613642156124115, + 0.016585160046815872, + 0.13295316696166992, + 0.1377682089805603, + 0.05230284109711647, + 0.1267729550600052 + ], + "credit_map": [ + 0.009670203551650047, + 0.007689961232244968, + 0.00485655665397644, + 0.010473412461578846, + 0.01777484081685543, + 0.05950748920440674, + 0.010839440859854221, + 0.035300374031066895, + 0.01313810795545578, + 0.005172925069928169, + 0.007426803931593895, + 0.0062704733572900295, + 0.009715421125292778, + 0.009485381655395031, + 0.010454212315380573, + 0.012369420379400253, + 0.004170161671936512, + 0.0036539945285767317, + 0.0032208056654781103, + 0.0032526024151593447, + 0.004356731660664082, + 0.005837110336869955, + 0.004187428392469883, + 0.00325016095302999, + 0.00462371576577425, + 0.006686608772724867, + 0.005315053276717663, + 0.004172130953520536, + 0.004413863178342581, + 0.004232611507177353, + 0.0020958485547453165, + 0.0038673912640661, + 0.016461890190839767, + 0.011162434704601765, + 0.07369016110897064, + 0.015281339175999165, + 0.016855403780937195, + 0.08034763485193253, + 0.007138597778975964, + 0.015198938548564911, + 0.0014908246230334044, + 0.002435125410556793, + 0.002610633382573724, + 0.002427691360935569, + 0.0023252901155501604, + 0.0018734338227659464, + 0.002091382397338748, + 0.0013307774206623435, + 0.023639144375920296, + 0.018069447949528694, + 0.01756618544459343, + 0.01708134636282921, + 0.017850827425718307, + 0.018194647505879402, + 0.0089961476624012, + 0.011555422097444534, + 0.008548418059945107, + 0.00850699096918106, + 0.0384308397769928, + 0.007152154576033354, + 0.007991624064743519, + 0.048045624047517776, + 0.005973147694021463, + 0.013119389303028584, + 0.0044473884627223015, + 0.005076155066490173, + 0.006855457089841366, + 0.006096239201724529, + 0.008482829667627811, + 0.007139784749597311, + 0.006801609881222248, + 0.007403370458632708, + 0.015620347112417221, + 0.023203445598483086, + 0.0172985028475523, + 0.01544059719890356, + 0.015397004783153534, + 0.01477909181267023, + 0.012430892325937748, + 0.012603068724274635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.005250592716038227 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a sun-dappled forest with ancient tree trunks and damp soil?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a butterfly with velvety black scales and wings outlined in a brilliant blue glow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a pale pink cherry blossom petal with slightly curled edges visible as the target?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the butterfly perform actions such as swooping, circling, and flapping its wings while chasing the petal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a close-up from a slightly upward angle that pans steadily to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029786288738250732 + ], + "opposite_logprob_tokens": [ + -3.5297863483428955 + ], + "logprob_reward": -0.029786288738250732, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029786288738250732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055459365248680115, + 0.06304150819778442, + 0.07070034742355347, + 0.07566932588815689, + 0.1185806468129158, + 0.02519141510128975, + 0.09626882523298264, + 0.17596128582954407, + 0.04909452795982361, + 0.2700328230857849 + ], + "credit_map": [ + 0.004014806821942329, + 0.0032057082280516624, + 0.0033938719425350428, + 0.0037921748589724302, + 0.004295105580240488, + 0.020586524158716202, + 0.003521773498505354, + 0.012649400159716606, + 0.005268904380500317, + 0.006791593041270971, + 0.006461961194872856, + 0.005605337675660849, + 0.01035816315561533, + 0.008512886241078377, + 0.009064548648893833, + 0.010978111065924168, + 0.008859159424901009, + 0.0061228456906974316, + 0.00825120136141777, + 0.007311427500098944, + 0.008943922817707062, + 0.00927248690277338, + 0.00799831934273243, + 0.013940987177193165, + 0.005532538518309593, + 0.00867980532348156, + 0.01059996709227562, + 0.005763690918684006, + 0.013700691051781178, + 0.02141689695417881, + 0.005063127726316452, + 0.0049126036465168, + 0.007756620179861784, + 0.006618799176067114, + 0.036745812743902206, + 0.009545776061713696, + 0.009535637684166431, + 0.03863276168704033, + 0.0032855290919542313, + 0.006459712982177734, + 0.0025020691100507975, + 0.0030292223673313856, + 0.004786825738847256, + 0.002204639371484518, + 0.0030484115704894066, + 0.005653354339301586, + 0.0017348098335787654, + 0.0022320847492665052, + 0.012078920379281044, + 0.01872151531279087, + 0.011937297880649567, + 0.015343748033046722, + 0.01444071065634489, + 0.011091475374996662, + 0.006165664177387953, + 0.0064894878305494785, + 0.010526363737881184, + 0.00970170646905899, + 0.05287275090813637, + 0.009673441760241985, + 0.010500275529921055, + 0.06537642329931259, + 0.007684590760618448, + 0.009625747799873352, + 0.004532926250249147, + 0.005888467188924551, + 0.007011154666543007, + 0.008281941525638103, + 0.004668003413826227, + 0.006981841288506985, + 0.005311216227710247, + 0.006418982986360788, + 0.036078080534935, + 0.04031483083963394, + 0.0413757860660553, + 0.028311973437666893, + 0.03491116687655449, + 0.034199055284261703, + 0.030475901439785957, + 0.024366015568375587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.029786288738250732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_2/index_tensor([1370, 1370])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_2/index_tensor([1370, 1370])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..199c38028e2945efe6629c5e7f1b80ce21a99e74 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_2/index_tensor([1370, 1370])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, sweeping aerial shot unfolds, the entire scene brimming with dynamic tension, reminiscent of a Chinese ink painting, as if the ink is still flowing across the rice paper. The river's surface, rendered with unrestrained, splashed brushstrokes, transforms into a surging, black wave. The surging spray, outlined with the decisive \"feibai\" technique, forms piercing white lines and dots against the dark surface. Amidst this chaos, two simple, black-sailed boats engage in a fierce confrontation, relentlessly tossed and hurled by the waves. One boat tilts sharply to the left, its jet-black sail nearly touching the roiling, inky water; the other, lifted high on the crest of the waves, its bow held high, as if in its final struggle. The blurred figures on the boats are rendered with minimal, calligraphic brushstrokes, their tense postures conveying the full force of their struggle against nature. The distant landscape depicts layered mountains, rendered in shades of gray ink, their outlines looming in the pervasive, moist mist. The scene lacks a specific color palette; instead, the shifting tones of ink, from burnt to thick, heavy, light, and clear, combined with the white space left on the rice paper, create a thrilling, tense atmosphere imbued with primal power and the beauty of Eastern philosophy.", + "seed": 7854, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_2/index_1370_seed_7854_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.026966195052955298, + "vlm_logprob": -0.026966195052955298, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.026966195052955298, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08755490183830261 + ], + "opposite_logprob_tokens": [ + -2.587554931640625 + ], + "logprob_reward": -0.08755490183830261, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08755490183830261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12583306431770325, + 0.1159011572599411, + 0.059289753437042236, + 0.18810833990573883, + 0.06752204895019531, + 0.2829726040363312, + 0.030599182471632957, + 0.02728341706097126, + 0.07602348178625107, + 0.02646695263683796 + ], + "credit_map": [ + 0.014005687087774277, + 0.01437951996922493, + 0.015817496925592422, + 0.025524279102683067, + 0.012681701220571995, + 0.01619592122733593, + 0.013714426197111607, + 0.013514026068150997, + 0.015538903884589672, + 0.017519259825348854, + 0.007479912601411343, + 0.011616488918662071, + 0.023050907999277115, + 0.022095849737524986, + 0.010250041261315346, + 0.008349775336682796, + 0.011682200245559216, + 0.008086301386356354, + 0.005275951698422432, + 0.004927658475935459, + 0.009923139587044716, + 0.007523511536419392, + 0.006832949351519346, + 0.005038038827478886, + 0.023646527901291847, + 0.01980751007795334, + 0.03622463718056679, + 0.0271561648696661, + 0.01863991655409336, + 0.017436573281884193, + 0.016444627195596695, + 0.028752388432621956, + 0.011269976384937763, + 0.008075809106230736, + 0.006184933707118034, + 0.007146748248487711, + 0.008889269083738327, + 0.009441172704100609, + 0.008839795365929604, + 0.007674341555684805, + 0.04013899713754654, + 0.030656173825263977, + 0.028176788240671158, + 0.04096117243170738, + 0.030736669898033142, + 0.03801151365041733, + 0.03657839447259903, + 0.037712886929512024, + 0.003704514354467392, + 0.0034429985098540783, + 0.0035957747604697943, + 0.0034426969941705465, + 0.004596698097884655, + 0.005833509843796492, + 0.0033400363754481077, + 0.00264295213855803, + 0.004114977549761534, + 0.003297771094366908, + 0.0023363756481558084, + 0.004149653483182192, + 0.0033709779381752014, + 0.005766194313764572, + 0.0018193104770034552, + 0.0024281563237309456, + 0.008374967612326145, + 0.00804174691438675, + 0.007967798970639706, + 0.009693830274045467, + 0.007688124664127827, + 0.008907965384423733, + 0.01511482335627079, + 0.010234219022095203, + 0.003279406111687422, + 0.0026850011199712753, + 0.0031589313875883818, + 0.0033086903858929873, + 0.00385810574516654, + 0.0038370524998754263, + 0.00346747855655849, + 0.0028722845017910004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.08755490183830261 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001954194623976946 + ], + "opposite_logprob_tokens": [ + -6.251954078674316 + ], + "logprob_reward": -0.001954194623976946, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001954194623976946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11829962581396103, + 0.05660432577133179, + 0.09922411292791367, + 0.3979019224643707, + 0.05914999917149544, + 0.09136404097080231, + 0.04303564876317978, + 0.046950630843639374, + 0.06747037172317505, + 0.019999314099550247 + ], + "credit_map": [ + 0.01832737773656845, + 0.01561487652361393, + 0.010463240556418896, + 0.015440793707966805, + 0.015342336148023605, + 0.02332303300499916, + 0.009810702875256538, + 0.00997726246714592, + 0.008314588107168674, + 0.006997579708695412, + 0.008723575621843338, + 0.00785495713353157, + 0.004924909211695194, + 0.009409384801983833, + 0.0042123934254050255, + 0.006166933104395866, + 0.017625881358981133, + 0.012243243865668774, + 0.010938582010567188, + 0.011718510650098324, + 0.015980558469891548, + 0.015992818400263786, + 0.007889138534665108, + 0.006835367530584335, + 0.045130759477615356, + 0.03903323784470558, + 0.08761230111122131, + 0.07971061766147614, + 0.03509647026658058, + 0.0375455878674984, + 0.03289017081260681, + 0.04088275879621506, + 0.007387778256088495, + 0.007861660793423653, + 0.005376648623496294, + 0.006952764932066202, + 0.009067014791071415, + 0.007562043145298958, + 0.006606407463550568, + 0.008335673250257969, + 0.013498634099960327, + 0.010490263812243938, + 0.009004801511764526, + 0.012225406244397163, + 0.009745466522872448, + 0.012700000777840614, + 0.011557485908269882, + 0.012141979299485683, + 0.006250207312405109, + 0.005464265588670969, + 0.004241028800606728, + 0.0050851027481257915, + 0.006046297028660774, + 0.005284877493977547, + 0.006727940868586302, + 0.003935923334211111, + 0.005328094586730003, + 0.004950347356498241, + 0.00509145250543952, + 0.004845082759857178, + 0.005828093271702528, + 0.01069402601569891, + 0.005075227934867144, + 0.005138301756232977, + 0.006683027371764183, + 0.0063586244359612465, + 0.010193664580583572, + 0.007768738549202681, + 0.00603764783591032, + 0.005534993950277567, + 0.012922538444399834, + 0.011971126310527325, + 0.002523029688745737, + 0.0024139436427503824, + 0.002086628694087267, + 0.0021572380792349577, + 0.0029610986821353436, + 0.003479970619082451, + 0.0020769876427948475, + 0.0023004156537353992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.001954194623976946 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430646575987339 + ], + "opposite_logprob_tokens": [ + -3.168064594268799 + ], + "logprob_reward": -0.0430646575987339, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430646575987339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15369564294815063, + 0.08866018056869507, + 0.1451415866613388, + 0.17544864118099213, + 0.02944904938340187, + 0.16699136793613434, + 0.026542233303189278, + 0.05284133553504944, + 0.1088201031088829, + 0.05240995064377785 + ], + "credit_map": [ + 0.018511781468987465, + 0.01381758227944374, + 0.02085304819047451, + 0.034423451870679855, + 0.01640189439058304, + 0.015774019062519073, + 0.014808820560574532, + 0.019105028361082077, + 0.011372971348464489, + 0.010644787922501564, + 0.010547671467065811, + 0.010626493021845818, + 0.016417859122157097, + 0.012698796577751637, + 0.008027903735637665, + 0.008323701098561287, + 0.02415483444929123, + 0.020233454182744026, + 0.012582201510667801, + 0.00981950294226408, + 0.033706117421388626, + 0.021434348076581955, + 0.014169299043715, + 0.009041820652782917, + 0.019953586161136627, + 0.016318954527378082, + 0.03477117791771889, + 0.02870864048600197, + 0.016815174371004105, + 0.013068558648228645, + 0.02446765825152397, + 0.021344900131225586, + 0.0042184507474303246, + 0.004036236554384232, + 0.0025957943871617317, + 0.002999525982886553, + 0.0039027556777000427, + 0.005700071342289448, + 0.0032146708108484745, + 0.002781542018055916, + 0.026777060702443123, + 0.02168186567723751, + 0.016658060252666473, + 0.013372659683227539, + 0.020058758556842804, + 0.023959912359714508, + 0.02577849291265011, + 0.018704546615481377, + 0.0030518085695803165, + 0.003590585896745324, + 0.002179557690396905, + 0.0035354606807231903, + 0.00356322736479342, + 0.004494722466915846, + 0.003028110135346651, + 0.0030987609643489122, + 0.006215688772499561, + 0.005607536993920803, + 0.005469226744025946, + 0.007099076174199581, + 0.0066515300422906876, + 0.010810167528688908, + 0.006622219458222389, + 0.00436589028686285, + 0.010300925932824612, + 0.009816405363380909, + 0.013267573900520802, + 0.014413101598620415, + 0.01161493081599474, + 0.010958283208310604, + 0.022927429527044296, + 0.015521451830863953, + 0.006125571206212044, + 0.0052828616462647915, + 0.00517274159938097, + 0.005941175390034914, + 0.006815621163696051, + 0.010007159784436226, + 0.006600105669349432, + 0.006464709993451834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.0430646575987339 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013378250878304243 + ], + "opposite_logprob_tokens": [ + -6.626338005065918 + ], + "logprob_reward": -0.0013378250878304243, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013378250878304243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12436513602733612, + 0.07760022580623627, + 0.0795234739780426, + 0.30437707901000977, + 0.07845507562160492, + 0.08754721283912659, + 0.04099643602967262, + 0.06627923250198364, + 0.10161712020635605, + 0.03923904150724411 + ], + "credit_map": [ + 0.01581018976867199, + 0.014635587111115456, + 0.01408515777438879, + 0.013757264241576195, + 0.01638833060860634, + 0.02152683399617672, + 0.014153268188238144, + 0.014008507132530212, + 0.013317662291228771, + 0.010595414787530899, + 0.011258729733526707, + 0.008765552192926407, + 0.007697943598031998, + 0.015364372171461582, + 0.004538991022855043, + 0.006061557214707136, + 0.011860272847115993, + 0.010611806064844131, + 0.007190515287220478, + 0.00878855586051941, + 0.016565518453717232, + 0.011045794002711773, + 0.007642762269824743, + 0.0058182524517178535, + 0.03334387391805649, + 0.030693423002958298, + 0.06086629629135132, + 0.049636565148830414, + 0.02878722921013832, + 0.03044034354388714, + 0.03828420117497444, + 0.03232510760426521, + 0.011416803114116192, + 0.011142101138830185, + 0.005658942740410566, + 0.009873702190816402, + 0.013367600739002228, + 0.010102353058755398, + 0.007552300114184618, + 0.00934127252548933, + 0.01229183655232191, + 0.012276261113584042, + 0.007208609953522682, + 0.007027036976069212, + 0.01431109756231308, + 0.012098541483283043, + 0.012433599680662155, + 0.009900221601128578, + 0.005099011585116386, + 0.004372104536741972, + 0.004750508815050125, + 0.0023426359985023737, + 0.006157665979117155, + 0.007684797514230013, + 0.006053936202079058, + 0.004535774700343609, + 0.008320319466292858, + 0.007641091011464596, + 0.007793175056576729, + 0.0071726772002875805, + 0.008940019644796848, + 0.013474567793309689, + 0.0069048115983605385, + 0.0060325744561851025, + 0.009576940909028053, + 0.010268772952258587, + 0.011421077884733677, + 0.016174975782632828, + 0.011513211764395237, + 0.01054302416741848, + 0.015684999525547028, + 0.016434116289019585, + 0.006427095737308264, + 0.004805960692465305, + 0.0046644555404782295, + 0.004154181573539972, + 0.005313508678227663, + 0.006155025213956833, + 0.005052288994193077, + 0.0026665241457521915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.0013378250878304243 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009193961159326136 + ], + "opposite_logprob_tokens": [ + -7.000919342041016 + ], + "logprob_reward": -0.0009193961159326136, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009193961159326136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2633204162120819, + 0.06408477574586868, + 0.05588052049279213, + 0.2178882360458374, + 0.1277201920747757, + 0.06334803253412247, + 0.037576500326395035, + 0.026994120329618454, + 0.090674489736557, + 0.052512768656015396 + ], + "credit_map": [ + 0.04072555899620056, + 0.027781160548329353, + 0.020245928317308426, + 0.01982687972486019, + 0.030129868537187576, + 0.09096424281597137, + 0.01591625064611435, + 0.01773049682378769, + 0.00922798179090023, + 0.009148530662059784, + 0.007970270700752735, + 0.009071392007172108, + 0.007839974015951157, + 0.006857066880911589, + 0.006083162967115641, + 0.007886388339102268, + 0.008486206643283367, + 0.006931612268090248, + 0.005466925445944071, + 0.004896190948784351, + 0.011468031443655491, + 0.009374884888529778, + 0.004836907610297203, + 0.004419753327965736, + 0.0299435593187809, + 0.027820313349366188, + 0.030118871480226517, + 0.026051053777337074, + 0.04089890792965889, + 0.0232818815857172, + 0.01691575162112713, + 0.022857872769236565, + 0.015650028362870216, + 0.01414668932557106, + 0.012199041433632374, + 0.022876301780343056, + 0.013856861740350723, + 0.016718387603759766, + 0.012290596030652523, + 0.019982270896434784, + 0.008450014516711235, + 0.006852189544588327, + 0.004775690380483866, + 0.009612392634153366, + 0.008000397123396397, + 0.008930221199989319, + 0.008617040701210499, + 0.008110083639621735, + 0.004696776624768972, + 0.005398295819759369, + 0.003180032828822732, + 0.0035032902378588915, + 0.006540478207170963, + 0.006464301608502865, + 0.004739131312817335, + 0.0030541925225406885, + 0.0028905977960675955, + 0.0034499058965593576, + 0.0024262655060738325, + 0.004515914712101221, + 0.004010768141597509, + 0.00388218741863966, + 0.0037282812409102917, + 0.002090196358039975, + 0.011021321639418602, + 0.008510596118867397, + 0.00968471821397543, + 0.011322703212499619, + 0.00809817761182785, + 0.009800787083804607, + 0.018703890964388847, + 0.01353228185325861, + 0.005014571826905012, + 0.005344496574252844, + 0.005931743886321783, + 0.005760518833994865, + 0.0051053306087851524, + 0.005541789345443249, + 0.00916876271367073, + 0.0106455497443676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0009193961159326136 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, sweeping aerial shot unfolds, the entire scene brimming with dynamic tension, reminiscent of a Chinese ink painting, as if the ink is still flowing across the rice paper. The river's surface, rendered with unrestrained, splashed brushstrokes, transforms into a surging, black wave. The surging spray, outlined with the decisive \"feibai\" technique, forms piercing white lines and dots against the dark surface. Amidst this chaos, two simple, black-sailed boats engage in a fierce confrontation, relentlessly tossed and hurled by the waves. One boat tilts sharply to the left, its jet-black sail nearly touching the roiling, inky water; the other, lifted high on the crest of the waves, its bow held high, as if in its final struggle. The blurred figures on the boats are rendered with minimal, calligraphic brushstrokes, their tense postures conveying the full force of their struggle against nature. The distant landscape depicts layered mountains, rendered in shades of gray ink, their outlines looming in the pervasive, moist mist. The scene lacks a specific color palette; instead, the shifting tones of ink, from burnt to thick, heavy, light, and clear, combined with the white space left on the rice paper, create a thrilling, tense atmosphere imbued with primal power and the beauty of Eastern philosophy.", + "seed": 7855, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_2/index_1370_seed_7855_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10984938677866012, + "vlm_logprob": -0.10984938677866012, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.10984938677866012, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.41656532883644104 + ], + "opposite_logprob_tokens": [ + -1.1665652990341187 + ], + "logprob_reward": -0.41656532883644104, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.41656532883644104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09894242137670517, + 0.17093977332115173, + 0.10678058117628098, + 0.057062938809394836, + 0.03640550747513771, + 0.06022501364350319, + 0.039123810827732086, + 0.07400483638048172, + 0.22789861261844635, + 0.1286165565252304 + ], + "credit_map": [ + 0.009041987359523773, + 0.01305721141397953, + 0.007293217815458775, + 0.013670645654201508, + 0.01572662591934204, + 0.01680794358253479, + 0.009312562644481659, + 0.014032229781150818, + 0.03645901381969452, + 0.02713797427713871, + 0.017899785190820694, + 0.017053263261914253, + 0.019107909873127937, + 0.02494482323527336, + 0.013881910592317581, + 0.014455082826316357, + 0.009819687344133854, + 0.008592595346271992, + 0.007978287525475025, + 0.017342884093523026, + 0.010618255473673344, + 0.013573013246059418, + 0.015306830406188965, + 0.023549024015665054, + 0.005672906059771776, + 0.008846177719533443, + 0.005483509507030249, + 0.004575389437377453, + 0.00850268267095089, + 0.014035822823643684, + 0.003930476028472185, + 0.006015975959599018, + 0.0032023745588958263, + 0.005184566602110863, + 0.0030775482300668955, + 0.00230211578309536, + 0.005878351163119078, + 0.010464370250701904, + 0.0034745018929243088, + 0.0028216794598847628, + 0.008785285986959934, + 0.007063191384077072, + 0.006507692858576775, + 0.007342106197029352, + 0.009161719121038914, + 0.011138509027659893, + 0.005177071783691645, + 0.005049435887485743, + 0.004729290492832661, + 0.005860152654349804, + 0.002254139631986618, + 0.0021428712643682957, + 0.009000792168080807, + 0.009601730853319168, + 0.002541555091738701, + 0.002993278671056032, + 0.009570473805069923, + 0.010725797154009342, + 0.004969062749296427, + 0.007365605793893337, + 0.010691755451261997, + 0.009353588335216045, + 0.008236126974225044, + 0.013092420995235443, + 0.027578124776482582, + 0.0355379618704319, + 0.020689506083726883, + 0.024208420887589455, + 0.04163171350955963, + 0.03534751385450363, + 0.017813339829444885, + 0.025092024356126785, + 0.0362408384680748, + 0.004585823509842157, + 0.00521892448887229, + 0.006855635903775692, + 0.012890055775642395, + 0.05209357663989067, + 0.004061043728142977, + 0.0066706459037959576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.41656532883644104 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001966924872249365 + ], + "opposite_logprob_tokens": [ + -6.251966953277588 + ], + "logprob_reward": -0.001966924872249365, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001966924872249365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1069682314991951, + 0.0810837671160698, + 0.04538458585739136, + 0.042108647525310516, + 0.036647576838731766, + 0.12166843563318253, + 0.0344119518995285, + 0.028424125164747238, + 0.06935425847768784, + 0.4339483678340912 + ], + "credit_map": [ + 0.011363894678652287, + 0.014423497021198273, + 0.011233632452785969, + 0.01788386143743992, + 0.016793353483080864, + 0.014416988007724285, + 0.010888779535889626, + 0.009964236989617348, + 0.01929943636059761, + 0.011074946261942387, + 0.009743794798851013, + 0.009846904315054417, + 0.008967429399490356, + 0.007915940135717392, + 0.0073987822979688644, + 0.006836539600044489, + 0.003661278635263443, + 0.00451508117839694, + 0.0027186861261725426, + 0.004236401524394751, + 0.0046441094018518925, + 0.007249453105032444, + 0.00780116394162178, + 0.01055841613560915, + 0.008224383927881718, + 0.00576073257252574, + 0.004761680960655212, + 0.005192302167415619, + 0.005201814696192741, + 0.007245803717523813, + 0.002824855502694845, + 0.0028970802668482065, + 0.003368607722222805, + 0.004943694919347763, + 0.003999023232609034, + 0.001957671018317342, + 0.00612191716209054, + 0.010807978920638561, + 0.0027091926895081997, + 0.0027394937351346016, + 0.01962442509829998, + 0.018006904050707817, + 0.01012006402015686, + 0.009998473338782787, + 0.014447818510234356, + 0.02311933971941471, + 0.012212855741381645, + 0.014138556085526943, + 0.0037066566292196512, + 0.004378185141831636, + 0.0021359515376389027, + 0.002464056247845292, + 0.006680090446025133, + 0.009663372300565243, + 0.0029374014120548964, + 0.0024462395813316107, + 0.003605001140385866, + 0.003968948032706976, + 0.003016397589817643, + 0.0035039070062339306, + 0.004211674444377422, + 0.003894768888130784, + 0.003291349159553647, + 0.002932078205049038, + 0.007234582677483559, + 0.011760016903281212, + 0.007833980955183506, + 0.009405224584043026, + 0.010806018486618996, + 0.010988451540470123, + 0.004549883306026459, + 0.006776101421564817, + 0.12948735058307648, + 0.013856712728738785, + 0.01283861231058836, + 0.008751211687922478, + 0.04634033888578415, + 0.19319845736026764, + 0.010107271373271942, + 0.01936843991279602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.001966924872249365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704302370548248 + ], + "opposite_logprob_tokens": [ + -2.1270430088043213 + ], + "logprob_reward": -0.12704302370548248, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704302370548248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2338550090789795, + 0.08268623799085617, + 0.08080858737230301, + 0.03543083369731903, + 0.1004958525300026, + 0.07844703644514084, + 0.04552428424358368, + 0.04530204087495804, + 0.0942884087562561, + 0.20316170156002045 + ], + "credit_map": [ + 0.047869253903627396, + 0.02399555593729019, + 0.040047917515039444, + 0.029671069234609604, + 0.02744278497993946, + 0.02813747338950634, + 0.021022189408540726, + 0.015668770298361778, + 0.011475887149572372, + 0.009675787761807442, + 0.007711772806942463, + 0.00542403943836689, + 0.01272970624268055, + 0.013961812481284142, + 0.0073312558233737946, + 0.014375976286828518, + 0.015791941434144974, + 0.010493018664419651, + 0.007590965833514929, + 0.009666718542575836, + 0.008888337761163712, + 0.016705047339200974, + 0.004366325214505196, + 0.007306225132197142, + 0.005868830252438784, + 0.004083782434463501, + 0.0029489356093108654, + 0.0030856679659336805, + 0.005767968483269215, + 0.00577405234798789, + 0.0025803912431001663, + 0.005321204196661711, + 0.008260076865553856, + 0.014427527785301208, + 0.010305413976311684, + 0.00386430649086833, + 0.018139800056815147, + 0.027680814266204834, + 0.008976730518043041, + 0.008841182105243206, + 0.014440489932894707, + 0.008950384333729744, + 0.006515479180961847, + 0.010527088306844234, + 0.010370035655796528, + 0.0122454222291708, + 0.007859050296247005, + 0.007539088372141123, + 0.009598938748240471, + 0.007005271501839161, + 0.00329048209823668, + 0.0029804264195263386, + 0.008842948824167252, + 0.007673699874430895, + 0.0027847690507769585, + 0.003347749589011073, + 0.005371396895498037, + 0.005516137927770615, + 0.00469172652810812, + 0.006225843448191881, + 0.0069624935276806355, + 0.005368893966078758, + 0.005187527742236853, + 0.0059780217707157135, + 0.009689165279269218, + 0.021472234278917313, + 0.00849902629852295, + 0.0160610880702734, + 0.011455663479864597, + 0.011028039269149303, + 0.0067418720573186874, + 0.009341323748230934, + 0.06165454909205437, + 0.008404240943491459, + 0.005664023570716381, + 0.004035839810967445, + 0.02201685681939125, + 0.08868605643510818, + 0.004409682471305132, + 0.00829045195132494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.12704302370548248 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024924185127019882 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024924185127019882, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024924185127019882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0900348648428917, + 0.11473952233791351, + 0.16600772738456726, + 0.047279320657253265, + 0.08742094784975052, + 0.23341384530067444, + 0.06320074945688248, + 0.04599630460143089, + 0.05263044685125351, + 0.09927622973918915 + ], + "credit_map": [ + 0.01213703490793705, + 0.011105694808065891, + 0.011001820676028728, + 0.014017872512340546, + 0.015412682667374611, + 0.011826354078948498, + 0.009061206132173538, + 0.005472202785313129, + 0.013646264560520649, + 0.012682498432695866, + 0.010190543718636036, + 0.013024432584643364, + 0.01568497158586979, + 0.014618623070418835, + 0.014168637804687023, + 0.020723549649119377, + 0.014522384852170944, + 0.013065197505056858, + 0.010259374976158142, + 0.016007617115974426, + 0.013910435140132904, + 0.02059214934706688, + 0.03227236866950989, + 0.045378200709819794, + 0.0039831651374697685, + 0.005498846992850304, + 0.004021522589027882, + 0.004294225480407476, + 0.007354014087468386, + 0.010650765150785446, + 0.0041772471740841866, + 0.007299534045159817, + 0.006402195431292057, + 0.012805158272385597, + 0.009382170625030994, + 0.0036741462536156178, + 0.015670713037252426, + 0.024734120815992355, + 0.0072112311609089375, + 0.007541214115917683, + 0.046576812863349915, + 0.03303554654121399, + 0.020076686516404152, + 0.019886333495378494, + 0.04043135046958923, + 0.03924647718667984, + 0.014313671737909317, + 0.01984696090221405, + 0.009311401285231113, + 0.010234936140477657, + 0.004911225754767656, + 0.0047296867705881596, + 0.008481986820697784, + 0.015117490664124489, + 0.005399941932410002, + 0.005014075897634029, + 0.004570727702230215, + 0.005600083153694868, + 0.004964492749422789, + 0.005333793815225363, + 0.0074126278050243855, + 0.006741940975189209, + 0.005735577084124088, + 0.005637063179165125, + 0.00719071738421917, + 0.007005638442933559, + 0.006240237969905138, + 0.0037857312709093094, + 0.011953921988606453, + 0.008756622672080994, + 0.0038518982473760843, + 0.0038456767797470093, + 0.023232104256749153, + 0.0049246447160840034, + 0.004799657966941595, + 0.0058687832206487656, + 0.011249594390392303, + 0.034883271902799606, + 0.006444119848310947, + 0.007874056696891785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.0024924185127019882 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011792379664257169 + ], + "opposite_logprob_tokens": [ + -6.751179218292236 + ], + "logprob_reward": -0.0011792379664257169, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011792379664257169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17308466136455536, + 0.10724228620529175, + 0.09756278991699219, + 0.061497338116168976, + 0.10198237746953964, + 0.08722662925720215, + 0.05182823911309242, + 0.03581937402486801, + 0.07957823574542999, + 0.2041780948638916 + ], + "credit_map": [ + 0.01869550161063671, + 0.02607976645231247, + 0.016301553696393967, + 0.02352103218436241, + 0.0201865341514349, + 0.026334242895245552, + 0.016470134258270264, + 0.025495896115899086, + 0.021983793005347252, + 0.017103055492043495, + 0.011249818839132786, + 0.008727886714041233, + 0.013366028666496277, + 0.015509139746427536, + 0.008679825812578201, + 0.010622730478644371, + 0.021566718816757202, + 0.012993642129004002, + 0.012077094987034798, + 0.0084860660135746, + 0.009970927610993385, + 0.01627662219107151, + 0.007131041027605534, + 0.00906066969037056, + 0.005043582990765572, + 0.008465028367936611, + 0.005501593928784132, + 0.005982489325106144, + 0.010068967938423157, + 0.015619080513715744, + 0.004330198746174574, + 0.006486391182988882, + 0.010259258560836315, + 0.014869422651827335, + 0.011917510069906712, + 0.006591197568923235, + 0.017727354541420937, + 0.02635718509554863, + 0.00697349151596427, + 0.0072869504801929, + 0.016277605667710304, + 0.011883056722581387, + 0.009056377224624157, + 0.007401538547128439, + 0.01236230880022049, + 0.016079088672995567, + 0.007254570722579956, + 0.0069120763801038265, + 0.007282773498445749, + 0.008057939819991589, + 0.003689779667183757, + 0.003029854269698262, + 0.013636766001582146, + 0.007548979949206114, + 0.0046198247000575066, + 0.003962319809943438, + 0.004488334525376558, + 0.004398614168167114, + 0.002409430919215083, + 0.0029664544854313135, + 0.005524429492652416, + 0.005044547840952873, + 0.004609907511621714, + 0.0063776541501283646, + 0.009075884707272053, + 0.01075297873467207, + 0.0078068021684885025, + 0.008820015005767345, + 0.01604910008609295, + 0.012894436717033386, + 0.005763455294072628, + 0.00841555930674076, + 0.0617622546851635, + 0.006613710895180702, + 0.005749075207859278, + 0.004466614220291376, + 0.022534003481268883, + 0.09069900959730148, + 0.004268165212124586, + 0.008085248991847038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.0011792379664257169 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, sweeping aerial shot unfolds, the entire scene brimming with dynamic tension, reminiscent of a Chinese ink painting, as if the ink is still flowing across the rice paper. The river's surface, rendered with unrestrained, splashed brushstrokes, transforms into a surging, black wave. The surging spray, outlined with the decisive \"feibai\" technique, forms piercing white lines and dots against the dark surface. Amidst this chaos, two simple, black-sailed boats engage in a fierce confrontation, relentlessly tossed and hurled by the waves. One boat tilts sharply to the left, its jet-black sail nearly touching the roiling, inky water; the other, lifted high on the crest of the waves, its bow held high, as if in its final struggle. The blurred figures on the boats are rendered with minimal, calligraphic brushstrokes, their tense postures conveying the full force of their struggle against nature. The distant landscape depicts layered mountains, rendered in shades of gray ink, their outlines looming in the pervasive, moist mist. The scene lacks a specific color palette; instead, the shifting tones of ink, from burnt to thick, heavy, light, and clear, combined with the white space left on the rice paper, create a thrilling, tense atmosphere imbued with primal power and the beauty of Eastern philosophy.", + "seed": 7856, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_2/index_1370_seed_7856_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09628208349458874, + "vlm_logprob": -0.09628208349458874, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.09628208349458874, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42041730880737305 + ], + "opposite_logprob_tokens": [ + -1.170417308807373 + ], + "logprob_reward": -0.42041730880737305, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.42041730880737305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30294111371040344, + 0.1228916198015213, + 0.10236745327711105, + 0.07214576005935669, + 0.03926018625497818, + 0.027164854109287262, + 0.17854495346546173, + 0.06873036175966263, + 0.07239747047424316, + 0.013556175865232944 + ], + "credit_map": [ + 0.049936842173337936, + 0.045089561492204666, + 0.02728700451552868, + 0.041529539972543716, + 0.04060298576951027, + 0.03593304753303528, + 0.023926714435219765, + 0.03863544389605522, + 0.025535354390740395, + 0.018633877858519554, + 0.01167960837483406, + 0.008454198017716408, + 0.019149117171764374, + 0.018336519598960876, + 0.009189445525407791, + 0.011913509108126163, + 0.01351554598659277, + 0.010716940276324749, + 0.011354629881680012, + 0.0166922677308321, + 0.011352654546499252, + 0.011486385948956013, + 0.01275336742401123, + 0.014495670795440674, + 0.008785475976765156, + 0.018548529595136642, + 0.006155741401016712, + 0.00550512271001935, + 0.013114174827933311, + 0.008132686838507652, + 0.004592240322381258, + 0.007311794441193342, + 0.006088798400014639, + 0.0046011158265173435, + 0.004003880545496941, + 0.005177403334528208, + 0.004520935472100973, + 0.0029603063594549894, + 0.006303971633315086, + 0.0056037819012999535, + 0.00487923389300704, + 0.0035463275853544474, + 0.0033519945573061705, + 0.0032455374021083117, + 0.003683371702209115, + 0.003560032229870558, + 0.0019040984334424138, + 0.0029942600522190332, + 0.03405545651912689, + 0.02604042924940586, + 0.015223264694213867, + 0.01707759127020836, + 0.02014528028666973, + 0.012977715581655502, + 0.03346996381878853, + 0.019555263221263885, + 0.01474081166088581, + 0.008425592444837093, + 0.007796304766088724, + 0.006746727507561445, + 0.00819738581776619, + 0.007834206335246563, + 0.007165473885834217, + 0.007823867723345757, + 0.009462825022637844, + 0.011875685304403305, + 0.00721227191388607, + 0.011301642283797264, + 0.009011643938720226, + 0.006451559718698263, + 0.008726495318114758, + 0.008355358615517616, + 0.0018992628902196884, + 0.002298266626894474, + 0.0012524101184681058, + 0.0018760706298053265, + 0.002054249169304967, + 0.0015431162901222706, + 0.0009331895271316171, + 0.0016996109625324607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.42041730880737305 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059771365486085415 + ], + "opposite_logprob_tokens": [ + -5.130977153778076 + ], + "logprob_reward": -0.0059771365486085415, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059771365486085415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20709647238254547, + 0.08566784113645554, + 0.09024510532617569, + 0.21354852616786957, + 0.04333089664578438, + 0.01888122595846653, + 0.12013350427150726, + 0.16177180409431458, + 0.025007328018546104, + 0.03431728482246399 + ], + "credit_map": [ + 0.02336890809237957, + 0.028622811660170555, + 0.030793864279985428, + 0.0562080480158329, + 0.023579107597470284, + 0.011522567830979824, + 0.019983487203717232, + 0.013017687946557999, + 0.016931729391217232, + 0.012319662608206272, + 0.010381661355495453, + 0.011430741287767887, + 0.011420364491641521, + 0.00982954166829586, + 0.0047794533893466, + 0.00857468880712986, + 0.010505300015211105, + 0.010869601741433144, + 0.014381574466824532, + 0.01145985722541809, + 0.010303225368261337, + 0.009426102042198181, + 0.007541924715042114, + 0.015757525339722633, + 0.02942132018506527, + 0.040725670754909515, + 0.03680597245693207, + 0.018307490274310112, + 0.02984161116182804, + 0.011569679714739323, + 0.02899816818535328, + 0.01787860505282879, + 0.00618900079280138, + 0.009219259954988956, + 0.0027596724685281515, + 0.005332109518349171, + 0.007416377775371075, + 0.004596562124788761, + 0.0042037065140903, + 0.003614208195358515, + 0.0038147130981087685, + 0.004050885327160358, + 0.0019510065903887153, + 0.0017666654894128442, + 0.0030635923612862825, + 0.0020482162944972515, + 0.0008815326145850122, + 0.0013046144740656018, + 0.01860249973833561, + 0.015364443883299828, + 0.012298517860472202, + 0.015317771583795547, + 0.012480318546295166, + 0.00789771880954504, + 0.023643678054213524, + 0.014528551138937473, + 0.03442930430173874, + 0.016229551285505295, + 0.016985172405838966, + 0.013701396994292736, + 0.01688762940466404, + 0.01669672131538391, + 0.02406073734164238, + 0.022781290113925934, + 0.003079790621995926, + 0.004215499386191368, + 0.002717613009735942, + 0.00458251778036356, + 0.0029173102229833603, + 0.0023018294014036655, + 0.002596076112240553, + 0.002596690086647868, + 0.006501127500087023, + 0.005971655715256929, + 0.0030059937853366137, + 0.006268797442317009, + 0.004688749089837074, + 0.002375070471316576, + 0.00258254655636847, + 0.0029233419336378574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.0059771365486085415 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866647347807884 + ], + "opposite_logprob_tokens": [ + -3.048666477203369 + ], + "logprob_reward": -0.04866647347807884, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866647347807884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28489452600479126, + 0.08266636729240417, + 0.12751556932926178, + 0.08720714598894119, + 0.05344855412840843, + 0.04518233612179756, + 0.1661987453699112, + 0.029365617781877518, + 0.061650048941373825, + 0.06187102198600769 + ], + "credit_map": [ + 0.05028889328241348, + 0.037411145865917206, + 0.0335908941924572, + 0.03535706177353859, + 0.030833639204502106, + 0.029055384919047356, + 0.03136945143342018, + 0.036988068372011185, + 0.013691571541130543, + 0.011657812632620335, + 0.007550904527306557, + 0.010046474635601044, + 0.011692257598042488, + 0.012961892411112785, + 0.00691685127094388, + 0.008148607797920704, + 0.023514050990343094, + 0.02411816641688347, + 0.01824864186346531, + 0.010471449233591557, + 0.018206454813480377, + 0.01247331965714693, + 0.00931409653276205, + 0.011169394478201866, + 0.01313203014433384, + 0.01501829456537962, + 0.011523597873747349, + 0.007694362197071314, + 0.011767384596168995, + 0.009143143892288208, + 0.008584540337324142, + 0.01034379843622446, + 0.006353302858769894, + 0.010639402084052563, + 0.0037476231809705496, + 0.005093391053378582, + 0.009650231339037418, + 0.007792708929628134, + 0.005180785432457924, + 0.004991110414266586, + 0.00502405920997262, + 0.008534367196261883, + 0.0035534121561795473, + 0.0048751519061625, + 0.006263751536607742, + 0.006693969946354628, + 0.005561160854995251, + 0.00467646261677146, + 0.01843634806573391, + 0.021650396287441254, + 0.022759929299354553, + 0.029301458969712257, + 0.01766197569668293, + 0.012125946581363678, + 0.026159673929214478, + 0.01810302399098873, + 0.005493300501257181, + 0.0036473197396844625, + 0.0020856244955211878, + 0.0022457577288150787, + 0.0033094966784119606, + 0.0031461473554372787, + 0.004933104384690523, + 0.00450486782938242, + 0.008653643541038036, + 0.011821609921753407, + 0.005895301699638367, + 0.008630400523543358, + 0.009866718202829361, + 0.00680339802056551, + 0.004740492906421423, + 0.0052384790033102036, + 0.012865596450865269, + 0.011616243049502373, + 0.006333420518785715, + 0.00795433297753334, + 0.009314948692917824, + 0.004440257791429758, + 0.004004128277301788, + 0.005342097487300634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.04866647347807884 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017159751150757074 + ], + "opposite_logprob_tokens": [ + -6.376716136932373 + ], + "logprob_reward": -0.0017159751150757074, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017159751150757074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3733282685279846, + 0.07875659316778183, + 0.11172089725732803, + 0.08351324498653412, + 0.06267960369586945, + 0.028090614825487137, + 0.031643543392419815, + 0.12358206510543823, + 0.0740499198436737, + 0.03263526409864426 + ], + "credit_map": [ + 0.043261442333459854, + 0.06457611918449402, + 0.04357439652085304, + 0.07497959583997726, + 0.04875485599040985, + 0.03166217729449272, + 0.036864954978227615, + 0.029654718935489655, + 0.014373118057847023, + 0.014385506510734558, + 0.0058387089520692825, + 0.007206871639937162, + 0.012596220709383488, + 0.010760383680462837, + 0.00662677176296711, + 0.006969013717025518, + 0.01990721933543682, + 0.013973227702081203, + 0.026325194165110588, + 0.009436065331101418, + 0.012903446331620216, + 0.0076102036982774734, + 0.007666297256946564, + 0.013899242505431175, + 0.012093993835151196, + 0.014840461313724518, + 0.008339270949363708, + 0.007458855863660574, + 0.012926488183438778, + 0.01257443055510521, + 0.006868584081530571, + 0.008411156013607979, + 0.01122124120593071, + 0.01600397378206253, + 0.005263504572212696, + 0.005504984874278307, + 0.011443909257650375, + 0.005141777917742729, + 0.0034084173385053873, + 0.004691793117672205, + 0.004395000636577606, + 0.006168954074382782, + 0.0026087246369570494, + 0.001934850704856217, + 0.004438532516360283, + 0.003234207397326827, + 0.0021975606214255095, + 0.0031127831898629665, + 0.0038746092468500137, + 0.003929221536964178, + 0.003726125694811344, + 0.008454608730971813, + 0.0032436426263302565, + 0.0027497471310198307, + 0.003130408702418208, + 0.002535182051360607, + 0.023713981732726097, + 0.014167731627821922, + 0.010762368328869343, + 0.011896232143044472, + 0.014427753165364265, + 0.01277063973248005, + 0.018246566876769066, + 0.017596794292330742, + 0.0068131545558571815, + 0.010550823993980885, + 0.00899171456694603, + 0.015503555536270142, + 0.0078021325170993805, + 0.005833752453327179, + 0.01229103747755289, + 0.006263742223381996, + 0.0045479233376681805, + 0.004930592607706785, + 0.002698511816561222, + 0.005580650642514229, + 0.004138152580708265, + 0.003906753845512867, + 0.0033152808900922537, + 0.0035174002405256033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.0017159751150757074 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633523523807526 + ], + "opposite_logprob_tokens": [ + -5.37963342666626 + ], + "logprob_reward": -0.004633523523807526, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633523523807526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15931588411331177, + 0.07789911329746246, + 0.08167973160743713, + 0.0427282489836216, + 0.18127939105033875, + 0.23792128264904022, + 0.026014072820544243, + 0.1372212916612625, + 0.035019341856241226, + 0.02092159539461136 + ], + "credit_map": [ + 0.025995729491114616, + 0.02495349571108818, + 0.014965314418077469, + 0.02352255769073963, + 0.02043943852186203, + 0.017847662791609764, + 0.013365153223276138, + 0.018226535990834236, + 0.015936601907014847, + 0.015083319507539272, + 0.00600726343691349, + 0.006574844475835562, + 0.012078028172254562, + 0.009496958926320076, + 0.0055922456085681915, + 0.0071298498660326, + 0.01383525412529707, + 0.01085721142590046, + 0.011944419704377651, + 0.009701605886220932, + 0.009503252804279327, + 0.0068431394174695015, + 0.009095368906855583, + 0.009899474680423737, + 0.009526698850095272, + 0.008385220542550087, + 0.0041672890074551105, + 0.002621280960738659, + 0.006241579074412584, + 0.004537619184702635, + 0.003903718898072839, + 0.0033448433969169855, + 0.04136016219854355, + 0.032714881002902985, + 0.017094647511839867, + 0.016416961327195168, + 0.02783467248082161, + 0.017918933182954788, + 0.010350505821406841, + 0.017588643357157707, + 0.04228492081165314, + 0.042898029088974, + 0.028117710724473, + 0.026278870180249214, + 0.03216680884361267, + 0.02423767000436783, + 0.018682504072785378, + 0.023254772648215294, + 0.004661111626774073, + 0.004100599326193333, + 0.0018754947232082486, + 0.0025440938770771027, + 0.003611229592934251, + 0.002216475084424019, + 0.0036201151087880135, + 0.0033849540632218122, + 0.028852349147200584, + 0.015910066664218903, + 0.014220139943063259, + 0.011890755966305733, + 0.014994940720498562, + 0.015326450578868389, + 0.017594097182154655, + 0.0184324961155653, + 0.002797391265630722, + 0.004542217124253511, + 0.004444651771336794, + 0.007683025673031807, + 0.003488334594294429, + 0.0020147529430687428, + 0.006072798743844032, + 0.003976169973611832, + 0.002332299714908004, + 0.0036777460481971502, + 0.0016773188253864646, + 0.0030422809068113565, + 0.0028995054308325052, + 0.002140217227861285, + 0.002360295969992876, + 0.002791931154206395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.004633523523807526 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, sweeping aerial shot unfolds, the entire scene brimming with dynamic tension, reminiscent of a Chinese ink painting, as if the ink is still flowing across the rice paper. The river's surface, rendered with unrestrained, splashed brushstrokes, transforms into a surging, black wave. The surging spray, outlined with the decisive \"feibai\" technique, forms piercing white lines and dots against the dark surface. Amidst this chaos, two simple, black-sailed boats engage in a fierce confrontation, relentlessly tossed and hurled by the waves. One boat tilts sharply to the left, its jet-black sail nearly touching the roiling, inky water; the other, lifted high on the crest of the waves, its bow held high, as if in its final struggle. The blurred figures on the boats are rendered with minimal, calligraphic brushstrokes, their tense postures conveying the full force of their struggle against nature. The distant landscape depicts layered mountains, rendered in shades of gray ink, their outlines looming in the pervasive, moist mist. The scene lacks a specific color palette; instead, the shifting tones of ink, from burnt to thick, heavy, light, and clear, combined with the white space left on the rice paper, create a thrilling, tense atmosphere imbued with primal power and the beauty of Eastern philosophy.", + "seed": 7857, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_2/index_1370_seed_7857_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06985881454311312, + "vlm_logprob": -0.06985881454311312, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.06985881454311312, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09030456095933914 + ], + "opposite_logprob_tokens": [ + -2.7153046131134033 + ], + "logprob_reward": -0.09030456095933914, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09030456095933914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16485163569450378, + 0.04840656369924545, + 0.055482834577560425, + 0.06821059435606003, + 0.22227288782596588, + 0.09698515385389328, + 0.04487054422497749, + 0.13355028629302979, + 0.0756215825676918, + 0.08974791318178177 + ], + "credit_map": [ + 0.02680651657283306, + 0.017893150448799133, + 0.02170310541987419, + 0.027077646926045418, + 0.02050459198653698, + 0.014419139362871647, + 0.018901387229561806, + 0.01754608377814293, + 0.007277691271156073, + 0.006114199757575989, + 0.003993842285126448, + 0.006736866198480129, + 0.006542841903865337, + 0.005267787724733353, + 0.005298234522342682, + 0.007175101432949305, + 0.00602733762934804, + 0.008898613974452019, + 0.008142286911606789, + 0.005915287882089615, + 0.0070859575644135475, + 0.006495683919638395, + 0.006216562818735838, + 0.00670110248029232, + 0.009823175147175789, + 0.008731842041015625, + 0.007094412576407194, + 0.00806150771677494, + 0.011343649588525295, + 0.009790980257093906, + 0.005698386114090681, + 0.007666642777621746, + 0.02848595194518566, + 0.024852195754647255, + 0.01707662269473076, + 0.03174089267849922, + 0.029350347816944122, + 0.0193894375115633, + 0.045293115079402924, + 0.026084311306476593, + 0.012557771056890488, + 0.019476348534226418, + 0.006616097409278154, + 0.008674168959259987, + 0.013159036636352539, + 0.008676322177052498, + 0.01382224541157484, + 0.014003162272274494, + 0.004999806173145771, + 0.003484957152977586, + 0.0070550451055169106, + 0.010474244132637978, + 0.00546970684081316, + 0.004016335122287273, + 0.006347429007291794, + 0.0030230225529521704, + 0.012054546736180782, + 0.009671170264482498, + 0.016039520502090454, + 0.010063735768198967, + 0.01120563130825758, + 0.018163112923502922, + 0.020841581746935844, + 0.03551097586750984, + 0.01356167159974575, + 0.00828366819769144, + 0.005943208932876587, + 0.0035510926973074675, + 0.019358523190021515, + 0.012123571708798409, + 0.006273349281400442, + 0.006526492070406675, + 0.011156032793223858, + 0.011859063059091568, + 0.012453969568014145, + 0.010435822419822216, + 0.01091379951685667, + 0.01032758317887783, + 0.008557838387787342, + 0.014043805189430714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.09030456095933914 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003213244490325451 + ], + "opposite_logprob_tokens": [ + -5.753213405609131 + ], + "logprob_reward": -0.003213244490325451, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003213244490325451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3095989227294922, + 0.07293897867202759, + 0.03248060494661331, + 0.07098475843667984, + 0.16867545247077942, + 0.052439142018556595, + 0.07201530039310455, + 0.08883187174797058, + 0.07289823889732361, + 0.05913683772087097 + ], + "credit_map": [ + 0.04529481381177902, + 0.04166325181722641, + 0.043483007699251175, + 0.04502420499920845, + 0.03987734019756317, + 0.025835420936346054, + 0.030985113233327866, + 0.037435706704854965, + 0.010076992213726044, + 0.010907863266766071, + 0.009081888012588024, + 0.012114421464502811, + 0.011437833309173584, + 0.008341548033058643, + 0.004843530710786581, + 0.0061348918825387955, + 0.0034107668325304985, + 0.004431155975908041, + 0.004056221339851618, + 0.00266362470574677, + 0.004275872837752104, + 0.0042120469734072685, + 0.005673796404153109, + 0.003757112193852663, + 0.010348428040742874, + 0.010879782028496265, + 0.005256896372884512, + 0.008300790563225746, + 0.0069244555197656155, + 0.014830722473561764, + 0.007738050539046526, + 0.006705624051392078, + 0.03196436166763306, + 0.02300398424267769, + 0.01770240068435669, + 0.013955261558294296, + 0.020075449720025063, + 0.022632554173469543, + 0.023563487455248833, + 0.015777921304106712, + 0.006310456432402134, + 0.007962786592543125, + 0.003733445191755891, + 0.00325033999979496, + 0.005797136574983597, + 0.008893200196325779, + 0.007815890945494175, + 0.008675877936184406, + 0.010195497423410416, + 0.005746280774474144, + 0.00849191565066576, + 0.014601999893784523, + 0.011164593510329723, + 0.009083561599254608, + 0.007464361842721701, + 0.005267076659947634, + 0.01041124202311039, + 0.007393076084554195, + 0.009593307971954346, + 0.005452955607324839, + 0.00911913625895977, + 0.014451362192630768, + 0.01114293746650219, + 0.021267831325531006, + 0.012312212958931923, + 0.008541098795831203, + 0.006716497242450714, + 0.0035413436125963926, + 0.015083379112184048, + 0.011997108347713947, + 0.006892742123454809, + 0.007813842967152596, + 0.007023794576525688, + 0.007193090859800577, + 0.006769806612282991, + 0.004268171731382608, + 0.014871723018586636, + 0.005967032164335251, + 0.005627399776130915, + 0.0074158040806651115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.003213244490325451 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25210508704185486 + ], + "opposite_logprob_tokens": [ + -1.5021051168441772 + ], + "logprob_reward": -0.25210508704185486, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25210508704185486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10810799896717072, + 0.09049435704946518, + 0.08484720438718796, + 0.03963564336299896, + 0.3705379366874695, + 0.08457496017217636, + 0.038855694234371185, + 0.0932273343205452, + 0.03204174339771271, + 0.05767719820141792 + ], + "credit_map": [ + 0.018155010417103767, + 0.02142881229519844, + 0.009039320051670074, + 0.013693385757505894, + 0.016566481441259384, + 0.009045250713825226, + 0.011865122243762016, + 0.008314601145684719, + 0.014061912894248962, + 0.012677661143243313, + 0.012434742413461208, + 0.011367485858500004, + 0.011161747388541698, + 0.007202317006886005, + 0.00940175075083971, + 0.012186731211841106, + 0.008443860337138176, + 0.016145428642630577, + 0.008922376669943333, + 0.006984627805650234, + 0.011243505403399467, + 0.012761982157826424, + 0.010355296544730663, + 0.009990121237933636, + 0.010775634087622166, + 0.004496677778661251, + 0.003720543347299099, + 0.006384187377989292, + 0.0051360344514250755, + 0.00270428997464478, + 0.0038030229043215513, + 0.0026152506470680237, + 0.05354640260338783, + 0.049254149198532104, + 0.03518906235694885, + 0.033727023750543594, + 0.05179697647690773, + 0.035536352545022964, + 0.07356429100036621, + 0.037923663854599, + 0.01178709790110588, + 0.01786760985851288, + 0.0075549050234258175, + 0.008443457074463367, + 0.010091946460306644, + 0.009168109856545925, + 0.008929399773478508, + 0.010732424445450306, + 0.00691117811948061, + 0.00446569686755538, + 0.003968170844018459, + 0.0033183011692017317, + 0.008473733440041542, + 0.005118238739669323, + 0.002657628618180752, + 0.003942741081118584, + 0.015479675494134426, + 0.015311107970774174, + 0.01086871325969696, + 0.011499874293804169, + 0.015491753816604614, + 0.009113168343901634, + 0.007748260162770748, + 0.007714766543358564, + 0.005595846101641655, + 0.0038087100256234407, + 0.0030116296838968992, + 0.0017908104928210378, + 0.006534879095852375, + 0.00641076872125268, + 0.0022815519478172064, + 0.0026075441855937243, + 0.008276497945189476, + 0.00756223127245903, + 0.004359925165772438, + 0.006056106183677912, + 0.005993066355586052, + 0.006809930317103863, + 0.006585150025784969, + 0.012034285813570023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.25210508704185486 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024924185127019882 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024924185127019882, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024924185127019882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15722379088401794, + 0.13372716307640076, + 0.05594661086797714, + 0.1093805730342865, + 0.11641931533813477, + 0.08317703008651733, + 0.05006040260195732, + 0.14496929943561554, + 0.07085844874382019, + 0.0782373771071434 + ], + "credit_map": [ + 0.021689683198928833, + 0.023165198042988777, + 0.02199782431125641, + 0.018381882458925247, + 0.022413847967982292, + 0.015773991122841835, + 0.016180552542209625, + 0.017620807513594627, + 0.021060550585389137, + 0.018705755472183228, + 0.009876429103314877, + 0.016465697437524796, + 0.021231479942798615, + 0.020220236852765083, + 0.013048235327005386, + 0.013118777424097061, + 0.006231918931007385, + 0.010959893465042114, + 0.005549312103539705, + 0.005169766489416361, + 0.008417473174631596, + 0.007110534701496363, + 0.005354168359190226, + 0.0071535417810082436, + 0.021814921870827675, + 0.017460862174630165, + 0.010825668461620808, + 0.010987218469381332, + 0.01279640756547451, + 0.01373685710132122, + 0.009171939454972744, + 0.012586687691509724, + 0.017904506996273994, + 0.016648145392537117, + 0.01214843150228262, + 0.011639712378382683, + 0.016361258924007416, + 0.017899716272950172, + 0.013223161920905113, + 0.010594380088150501, + 0.012800817377865314, + 0.019313573837280273, + 0.00727920001372695, + 0.00605739513412118, + 0.010622387751936913, + 0.008379044011235237, + 0.009155291132628918, + 0.00956932082772255, + 0.006995113100856543, + 0.005066327750682831, + 0.004132600501179695, + 0.008776369504630566, + 0.008615147322416306, + 0.008964074775576591, + 0.00371521501801908, + 0.003795553930103779, + 0.013748363591730595, + 0.014181798323988914, + 0.016526738181710243, + 0.012871810235083103, + 0.012485460378229618, + 0.019699526950716972, + 0.020080573856830597, + 0.03537502512335777, + 0.01114716101437807, + 0.005789181217551231, + 0.0075964308343827724, + 0.00457503879442811, + 0.009610430337488651, + 0.015426186844706535, + 0.006955212913453579, + 0.009758801199495792, + 0.009956050664186478, + 0.010025189258158207, + 0.006808210164308548, + 0.00609240960329771, + 0.015447347424924374, + 0.012871887534856796, + 0.008502308279275894, + 0.008533972315490246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0024924185127019882 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slow, sweeping aerial perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style reminiscent of a Chinese ink painting with flowing ink on rice paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are two boats with black sails visible engaging in a confrontation on the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the figures on the boats rendered with minimal, calligraphic brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of layered gray mountains shrouded in mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011787617113441229 + ], + "opposite_logprob_tokens": [ + -6.751178741455078 + ], + "logprob_reward": -0.0011787617113441229, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011787617113441229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2863505184650421, + 0.18754735589027405, + 0.03173689544200897, + 0.07332507520914078, + 0.14461082220077515, + 0.040382418781518936, + 0.049770113080739975, + 0.08262153714895248, + 0.06121890991926193, + 0.04243631288409233 + ], + "credit_map": [ + 0.043101198971271515, + 0.036367617547512054, + 0.04600878804922104, + 0.037483882158994675, + 0.035940710455179214, + 0.027242328971624374, + 0.024842845275998116, + 0.035363152623176575, + 0.02544528990983963, + 0.02029579132795334, + 0.02216067723929882, + 0.018092747777700424, + 0.017865994945168495, + 0.013677526265382767, + 0.026515407487750053, + 0.043493933975696564, + 0.0028697652742266655, + 0.0067682391963899136, + 0.0023005891125649214, + 0.004335192497819662, + 0.0050430395640432835, + 0.0026070335879921913, + 0.0031186863780021667, + 0.0046943482011556625, + 0.01737404055893421, + 0.010543474927544594, + 0.006654879078269005, + 0.006397922523319721, + 0.007304024882614613, + 0.009736726060509682, + 0.0055974507704377174, + 0.009716561064124107, + 0.02398594468832016, + 0.02144128642976284, + 0.013753365725278854, + 0.009629043750464916, + 0.021395714953541756, + 0.020874397829174995, + 0.02018345333635807, + 0.013347613625228405, + 0.004588641691952944, + 0.0054348609410226345, + 0.0034433389082551003, + 0.0035561879631131887, + 0.005034240428358316, + 0.004717981442809105, + 0.007928882725536823, + 0.005678285378962755, + 0.007394095882773399, + 0.00527894077822566, + 0.003448453964665532, + 0.00539420684799552, + 0.013729134574532509, + 0.0067534432746469975, + 0.003853192552924156, + 0.003918645437806845, + 0.007111948449164629, + 0.008043060079216957, + 0.009544193744659424, + 0.006929611787199974, + 0.007301462348550558, + 0.009458338841795921, + 0.01455366238951683, + 0.01967925950884819, + 0.009951899759471416, + 0.005916171241551638, + 0.004960407502949238, + 0.0038536270149052143, + 0.010008365847170353, + 0.012509421445429325, + 0.006776245776563883, + 0.007242772728204727, + 0.005961353424936533, + 0.00505185266956687, + 0.005654189269989729, + 0.003284959588199854, + 0.009203827008605003, + 0.004169506020843983, + 0.004647559951990843, + 0.00446306774392724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0011787617113441229 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_3/index_tensor([2516, 2516])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_3/index_tensor([2516, 2516])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0f2977224fa8f483cf113e7417f6b6acbd96d7b8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_3/index_tensor([2516, 2516])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic, sweeping long shot, the camera slowly and steadily tilts upward from the horizon, gradually revealing a nightscape as deep as ink, dotted with distant nebulae. Initially, only a few diamond-bright stars begin to fall, but soon, a silent and magnificent cosmic spectacle unfolds. Thousands of blazing points of light, a blend of icy blue and warm gold, fall like tears from the velvety darkness, each leaving a slender, brilliant trail of light in their wake. These fleeting burning trails interweave into a dynamic web of light. Some are as long as silver threads, gracefully streaking across the sky, while others flash like fleeting sparks, their light briefly illuminating one another before fading, bringing a fleeting vitality to the vast darkness. At the bottom of the frame, the black silhouettes of the mountains stand silent witness to this celestial ballet. The entire scene, shrouded in absolute tranquility, exudes an awe-inspiring grandeur and majesty, creating a visual atmosphere both dreamlike and imbued with cosmic power.", + "seed": 15649, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_3/index_2516_seed_15649_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0793352303095162, + "vlm_logprob": -0.0793352303095162, + "vlm_yes_no_margin": 4.625, + "vlm_token_credit": -0.0793352303095162, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3551538288593292 + ], + "opposite_logprob_tokens": [ + -1.2301537990570068 + ], + "logprob_reward": -0.3551538288593292, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3551538288593292, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48299145698547363, + 0.044473037123680115, + 0.22087214887142181, + 0.0270231906324625, + 0.04693261906504631, + 0.024259792640805244, + 0.021734964102506638, + 0.08921787142753601, + 0.020896291360259056, + 0.021598609164357185 + ], + "credit_map": [ + 0.04351545125246048, + 0.05290798097848892, + 0.04474937543272972, + 0.0522180050611496, + 0.08492960035800934, + 0.100211963057518, + 0.05996265262365341, + 0.044496454298496246, + 0.004915595054626465, + 0.006261578761041164, + 0.003279903205111623, + 0.0033077294938266277, + 0.00960594229400158, + 0.007737775333225727, + 0.004877836909145117, + 0.004486676771193743, + 0.025662481784820557, + 0.021060578525066376, + 0.022544734179973602, + 0.013825860805809498, + 0.032708555459976196, + 0.03450788930058479, + 0.03086848184466362, + 0.03969357907772064, + 0.002467553596943617, + 0.0024782330729067326, + 0.004406322259455919, + 0.0038230009377002716, + 0.003947361372411251, + 0.004396763630211353, + 0.002893107943236828, + 0.0026108496822416782, + 0.0030518528074026108, + 0.0024666241370141506, + 0.00281788338907063, + 0.0027919141575694084, + 0.004607374779880047, + 0.013057291507720947, + 0.004795142449438572, + 0.013344538398087025, + 0.001318487455137074, + 0.0024043743032962084, + 0.002192889805883169, + 0.0027242081705480814, + 0.0023356680758297443, + 0.004282980691641569, + 0.00439493078738451, + 0.004606253933161497, + 0.0021198177710175514, + 0.002176285721361637, + 0.002266538795083761, + 0.0016735836397856474, + 0.003986534662544727, + 0.0036107830237597227, + 0.0030781615059822798, + 0.0028232585173100233, + 0.020174365490674973, + 0.0053895460441708565, + 0.03928200155496597, + 0.005493321921676397, + 0.0037536986637860537, + 0.0027380252722650766, + 0.007676300127059221, + 0.0047106146812438965, + 0.0022858208976686, + 0.00201990595087409, + 0.0020637107081711292, + 0.0024299619253724813, + 0.003289290703833103, + 0.0026029814034700394, + 0.0038703889586031437, + 0.0023342298809438944, + 0.002643452724441886, + 0.004646297078579664, + 0.003126887371763587, + 0.003493329044431448, + 0.002000202890485525, + 0.0022748615592718124, + 0.0016919751651585102, + 0.0017216033302247524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.3551538288593292 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005268855020403862 + ], + "opposite_logprob_tokens": [ + -5.2552690505981445 + ], + "logprob_reward": -0.005268855020403862, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005268855020403862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10937177389860153, + 0.08991523832082748, + 0.18805159628391266, + 0.11536432802677155, + 0.08042068034410477, + 0.06996378302574158, + 0.14953042566776276, + 0.07439152151346207, + 0.03066256083548069, + 0.09232819080352783 + ], + "credit_map": [ + 0.012381509877741337, + 0.01744905859231949, + 0.0077524385415017605, + 0.008497492410242558, + 0.021786773577332497, + 0.019750677049160004, + 0.012006202712655067, + 0.009747596457600594, + 0.00857205968350172, + 0.016399385407567024, + 0.0067312587052583694, + 0.009439783170819283, + 0.014601459726691246, + 0.01689542643725872, + 0.009212289936840534, + 0.008063552901148796, + 0.018196456134319305, + 0.030370328575372696, + 0.016993610188364983, + 0.01219143345952034, + 0.024579260498285294, + 0.029775545001029968, + 0.024372395128011703, + 0.03157253935933113, + 0.011714118532836437, + 0.015128049068152905, + 0.009831120260059834, + 0.014708275906741619, + 0.022068819031119347, + 0.016584977507591248, + 0.015310334041714668, + 0.010018611326813698, + 0.004314167890697718, + 0.004373189993202686, + 0.004354256670922041, + 0.0028268550522625446, + 0.008902176283299923, + 0.024738049134612083, + 0.0050272392109036446, + 0.025884734466671944, + 0.004886014387011528, + 0.005346038844436407, + 0.005456398706883192, + 0.008781866170465946, + 0.006735486444085836, + 0.011851259507238865, + 0.012745526619255543, + 0.014161183498799801, + 0.013235490769147873, + 0.014422183856368065, + 0.02135104313492775, + 0.010046940296888351, + 0.027273114770650864, + 0.01858966052532196, + 0.025847259908914566, + 0.018764706328511238, + 0.007949869148433208, + 0.007324873469769955, + 0.020220501348376274, + 0.00728770112618804, + 0.00726254703477025, + 0.0043574064038693905, + 0.014469796791672707, + 0.005518807098269463, + 0.0036239789333194494, + 0.004719267599284649, + 0.0026104869320988655, + 0.0038511711172759533, + 0.007478990592062473, + 0.0031299020629376173, + 0.002848465461283922, + 0.002400291385129094, + 0.013174338266253471, + 0.013532960787415504, + 0.012236839160323143, + 0.010121929459273815, + 0.012819427996873856, + 0.013783682137727737, + 0.0069356076419353485, + 0.009723387658596039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.005268855020403862 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0006360176485031843 + ], + "opposite_logprob_tokens": [ + -7.375636100769043 + ], + "logprob_reward": -0.0006360176485031843, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0006360176485031843, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1551276445388794, + 0.08858607709407806, + 0.0889354944229126, + 0.16090384125709534, + 0.10030422359704971, + 0.0865684300661087, + 0.09673784673213959, + 0.09272883832454681, + 0.05299349129199982, + 0.07711412012577057 + ], + "credit_map": [ + 0.016335777938365936, + 0.024249650537967682, + 0.015064173378050327, + 0.012609506025910378, + 0.030358897522091866, + 0.01978328637778759, + 0.020534684881567955, + 0.01619165949523449, + 0.009413163177669048, + 0.012136726640164852, + 0.008401148021221161, + 0.009250201284885406, + 0.012530545704066753, + 0.01764744333922863, + 0.010789591819047928, + 0.00841724593192339, + 0.010963722132146358, + 0.015133884735405445, + 0.008249809965491295, + 0.00631288904696703, + 0.01696716994047165, + 0.011180802248418331, + 0.01015283353626728, + 0.00997438095510006, + 0.010368479415774345, + 0.016810407862067223, + 0.012360116466879845, + 0.023670824244618416, + 0.032053109258413315, + 0.023975862190127373, + 0.02529858984053135, + 0.01636643148958683, + 0.00558480154722929, + 0.006275448016822338, + 0.006042905151844025, + 0.004488120321184397, + 0.011375371366739273, + 0.029786095023155212, + 0.007113074418157339, + 0.02963840216398239, + 0.004571768455207348, + 0.007023537065833807, + 0.007670936174690723, + 0.009665350429713726, + 0.0069789160043001175, + 0.012463733553886414, + 0.017017927020788193, + 0.021176258102059364, + 0.007549433968961239, + 0.011770385317504406, + 0.008971727453172207, + 0.007484151981770992, + 0.026944637298583984, + 0.010976036079227924, + 0.01042134128510952, + 0.01262013241648674, + 0.019273580983281136, + 0.006599802058190107, + 0.039127547293901443, + 0.00664164824411273, + 0.004211724270135164, + 0.00329360063187778, + 0.008249100297689438, + 0.00533183803781867, + 0.006080013234168291, + 0.006213071756064892, + 0.00597793236374855, + 0.006434664595872164, + 0.007227647118270397, + 0.004492567386478186, + 0.012130485847592354, + 0.004437102936208248, + 0.0111483009532094, + 0.009412881918251514, + 0.01113025564700365, + 0.010047508403658867, + 0.011605076491832733, + 0.010832059197127819, + 0.005357705056667328, + 0.0075803217478096485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.0006360176485031843 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019382515456527472 + ], + "opposite_logprob_tokens": [ + -6.251938343048096 + ], + "logprob_reward": -0.0019382515456527472, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019382515456527472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4700033664703369, + 0.21794500946998596, + 0.21987047791481018, + 0.01910087838768959, + 0.012265929020941257, + 0.011079831048846245, + 0.008189870975911617, + 0.022519661113619804, + 0.009097146801650524, + 0.009927787818014622 + ], + "credit_map": [ + 0.04014947637915611, + 0.065731942653656, + 0.05216938257217407, + 0.04479426145553589, + 0.08969519287347794, + 0.06818783283233643, + 0.060837145894765854, + 0.04843815416097641, + 0.02106497995555401, + 0.03635513409972191, + 0.017647694796323776, + 0.01645611971616745, + 0.04572458937764168, + 0.04107297211885452, + 0.019686222076416016, + 0.019937297329306602, + 0.01843632385134697, + 0.016849495470523834, + 0.022439630702137947, + 0.015636809170246124, + 0.03160184994339943, + 0.03561810404062271, + 0.03448120877146721, + 0.04480704665184021, + 0.0009299839148297906, + 0.0017056606011465192, + 0.0013983676908537745, + 0.002509171376004815, + 0.0035250855144113302, + 0.002843019785359502, + 0.003580498043447733, + 0.002609091578051448, + 0.0008200913434848189, + 0.0008838929934427142, + 0.0005508112371899188, + 0.0011455207131803036, + 0.00180473318323493, + 0.0028258762322366238, + 0.0015548085793852806, + 0.0026801940985023975, + 0.000543188420124352, + 0.0009174993610940874, + 0.0007503039669245481, + 0.0013875144068151712, + 0.0007992804748937488, + 0.0014036286156624556, + 0.0025295005179941654, + 0.002748915459960699, + 0.0008136865217238665, + 0.0011148392222821712, + 0.0008296245941892266, + 0.0007747100898995996, + 0.0011924623977392912, + 0.001230312162078917, + 0.001131347380578518, + 0.0011028886074200273, + 0.004845976829528809, + 0.0014222859172150493, + 0.010037746280431747, + 0.0013271303614601493, + 0.0007758678984828293, + 0.0006302946130745113, + 0.0022886504884809256, + 0.0011917080264538527, + 0.0009984755888581276, + 0.0008686118526384234, + 0.0013120920630171895, + 0.0012888105120509863, + 0.0010042893700301647, + 0.0009826081804931164, + 0.0019727922044694424, + 0.000669466273393482, + 0.0010713035007938743, + 0.0020903386175632477, + 0.0015004398301243782, + 0.0012467631604522467, + 0.0013475355226546526, + 0.0010336929699406028, + 0.0009527896181680262, + 0.0006849248893558979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0019382515456527472 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03367919847369194 + ], + "opposite_logprob_tokens": [ + -3.408679246902466 + ], + "logprob_reward": -0.03367919847369194, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03367919847369194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42485350370407104, + 0.25850552320480347, + 0.08940514177083969, + 0.03537682443857193, + 0.029438583180308342, + 0.024219738319516182, + 0.024841537699103355, + 0.05771593004465103, + 0.030571887269616127, + 0.025071362033486366 + ], + "credit_map": [ + 0.036313947290182114, + 0.06395980715751648, + 0.04622762277722359, + 0.04225347191095352, + 0.0831553265452385, + 0.05657137930393219, + 0.05552980676293373, + 0.04084210470318794, + 0.028976712375879288, + 0.0322231724858284, + 0.024384764954447746, + 0.023727869614958763, + 0.03650477156043053, + 0.06051578000187874, + 0.027720250189304352, + 0.02445218339562416, + 0.007525783963501453, + 0.007081319577991962, + 0.008919722400605679, + 0.006000474561005831, + 0.013055437244474888, + 0.01403355784714222, + 0.014395484700798988, + 0.01839335262775421, + 0.0017916589276865125, + 0.002920504193753004, + 0.0023143207654356956, + 0.006142765749245882, + 0.006773232016712427, + 0.0051994239911437035, + 0.006266015116125345, + 0.003968903329223394, + 0.0015421088319271803, + 0.0014713306445628405, + 0.0010449836263433099, + 0.0013642623089253902, + 0.0033422703854739666, + 0.009530273266136646, + 0.002063038293272257, + 0.009080315940082073, + 0.0014217074494808912, + 0.0018590866820886731, + 0.0017468766309320927, + 0.002805975265800953, + 0.002169446786865592, + 0.004146394319832325, + 0.0046224999241530895, + 0.005447749979794025, + 0.002824139315634966, + 0.0029945019632577896, + 0.002774663735181093, + 0.001968416152521968, + 0.005513249430805445, + 0.0032058171927928925, + 0.0024857791140675545, + 0.0030749705620110035, + 0.01282145082950592, + 0.003538389690220356, + 0.02519325166940689, + 0.003805015468969941, + 0.0023027125280350447, + 0.0020718821324408054, + 0.004775546956807375, + 0.003207683563232422, + 0.0035255635157227516, + 0.002081620506942272, + 0.0022333920933306217, + 0.004020593129098415, + 0.009594143368303776, + 0.0022259014658629894, + 0.004303700756281614, + 0.002586972899734974, + 0.0029936512000858784, + 0.003826363245025277, + 0.0035198188852518797, + 0.003058911766856909, + 0.0034602961968630552, + 0.0036976707633584738, + 0.0020533534698188305, + 0.002461294876411557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03367919847369194 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, sweeping long shot, the camera slowly and steadily tilts upward from the horizon, gradually revealing a nightscape as deep as ink, dotted with distant nebulae. Initially, only a few diamond-bright stars begin to fall, but soon, a silent and magnificent cosmic spectacle unfolds. Thousands of blazing points of light, a blend of icy blue and warm gold, fall like tears from the velvety darkness, each leaving a slender, brilliant trail of light in their wake. These fleeting burning trails interweave into a dynamic web of light. Some are as long as silver threads, gracefully streaking across the sky, while others flash like fleeting sparks, their light briefly illuminating one another before fading, bringing a fleeting vitality to the vast darkness. At the bottom of the frame, the black silhouettes of the mountains stand silent witness to this celestial ballet. The entire scene, shrouded in absolute tranquility, exudes an awe-inspiring grandeur and majesty, creating a visual atmosphere both dreamlike and imbued with cosmic power.", + "seed": 15650, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_3/index_2516_seed_15650_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13698529894463718, + "vlm_logprob": -0.13698529894463718, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.13698529894463718, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.48455697298049927 + ], + "opposite_logprob_tokens": [ + -0.9845569729804993 + ], + "logprob_reward": -0.48455697298049927, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.48455697298049927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19920887053012848, + 0.10165394097566605, + 0.04295401647686958, + 0.04781656712293625, + 0.033778078854084015, + 0.17495256662368774, + 0.042921457439661026, + 0.09429436177015305, + 0.04210151731967926, + 0.22031858563423157 + ], + "credit_map": [ + 0.01506492868065834, + 0.024449657648801804, + 0.01466769352555275, + 0.02071274258196354, + 0.04407444968819618, + 0.033646345138549805, + 0.02740887738764286, + 0.019184183329343796, + 0.006745190359652042, + 0.008381159044802189, + 0.007301000878214836, + 0.009778490290045738, + 0.01362577173858881, + 0.017182311043143272, + 0.017727266997098923, + 0.020912757143378258, + 0.0031686266884207726, + 0.0034133947920054197, + 0.00510344747453928, + 0.0042479862459003925, + 0.0029922828543931246, + 0.0067412168718874454, + 0.006002457346767187, + 0.011284601874649525, + 0.004436861258000135, + 0.003638845169916749, + 0.007819326594471931, + 0.004347515758126974, + 0.007220278959721327, + 0.006585311144590378, + 0.004938590340316296, + 0.008829840458929539, + 0.0035681023728102446, + 0.004776874557137489, + 0.003372805193066597, + 0.004648348782211542, + 0.0041926829144358635, + 0.0037436557468026876, + 0.004893162753432989, + 0.004582446068525314, + 0.01145286113023758, + 0.016903655603528023, + 0.01448671706020832, + 0.013058075681328773, + 0.02852575294673443, + 0.03040684573352337, + 0.026440037414431572, + 0.033678628504276276, + 0.005026086699217558, + 0.0039925603196024895, + 0.0031162116210907698, + 0.0034637583885341883, + 0.006317063234746456, + 0.0068565066903829575, + 0.005980135872960091, + 0.00816913228482008, + 0.0040562036447227, + 0.0050964136607944965, + 0.004099888261407614, + 0.004267571028321981, + 0.00436524860560894, + 0.027949098497629166, + 0.013619668781757355, + 0.030840272083878517, + 0.005024129990488291, + 0.005107291974127293, + 0.004579368978738785, + 0.005448891781270504, + 0.0034466299694031477, + 0.004888488911092281, + 0.006945839151740074, + 0.00666087819263339, + 0.06298603117465973, + 0.07220206409692764, + 0.012441704981029034, + 0.016351239755749702, + 0.018235744908452034, + 0.01743854023516178, + 0.012170486152172089, + 0.00849277712404728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.48455697298049927 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014218217693269253 + ], + "opposite_logprob_tokens": [ + -4.264218330383301 + ], + "logprob_reward": -0.014218217693269253, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014218217693269253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08671148121356964, + 0.0841432437300682, + 0.05210057273507118, + 0.031332094222307205, + 0.057532817125320435, + 0.08812989294528961, + 0.18579547107219696, + 0.09503967314958572, + 0.032722197473049164, + 0.2864925265312195 + ], + "credit_map": [ + 0.007310974411666393, + 0.01126434188336134, + 0.007127813994884491, + 0.010948560200631618, + 0.018279923126101494, + 0.012989229522645473, + 0.010368820279836655, + 0.008421824313700199, + 0.007477851584553719, + 0.008883487433195114, + 0.006388671230524778, + 0.007099008187651634, + 0.01110260933637619, + 0.01345042698085308, + 0.010093813762068748, + 0.019647380337119102, + 0.003312706248834729, + 0.0034749878104776144, + 0.004405859857797623, + 0.00739221042022109, + 0.006489383988082409, + 0.0070456210523843765, + 0.010390742681920528, + 0.009589067660272121, + 0.002923654392361641, + 0.0029716601129621267, + 0.003158056177198887, + 0.0029018977656960487, + 0.006041808053851128, + 0.0039581600576639175, + 0.004609470255672932, + 0.004767389968037605, + 0.0046414947137236595, + 0.005548491142690182, + 0.0038242749869823456, + 0.003914975095540285, + 0.011463911272585392, + 0.008669250644743443, + 0.0069493940100073814, + 0.012521026656031609, + 0.007339627481997013, + 0.00949897151440382, + 0.0071579706855118275, + 0.007290246896445751, + 0.013911953195929527, + 0.012466654181480408, + 0.016933973878622055, + 0.013530497439205647, + 0.010155157186090946, + 0.008617150597274303, + 0.009684577584266663, + 0.00820714607834816, + 0.025069408118724823, + 0.047382935881614685, + 0.01669672317802906, + 0.05998238921165466, + 0.006162148434668779, + 0.004307507537305355, + 0.004098804667592049, + 0.0031770914793014526, + 0.004060748033225536, + 0.03145719692111015, + 0.008235221728682518, + 0.03354096785187721, + 0.0027658140752464533, + 0.003448024857789278, + 0.0039159576408565044, + 0.004261997528374195, + 0.0033627706579864025, + 0.002985773840919137, + 0.007222900632768869, + 0.0047589619643986225, + 0.08132442831993103, + 0.09602180868387222, + 0.012515358626842499, + 0.022464435547590256, + 0.02448767051100731, + 0.022890137508511543, + 0.015572736039757729, + 0.01121597457677126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.014218217693269253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002201397204771638 + ], + "opposite_logprob_tokens": [ + -6.127201557159424 + ], + "logprob_reward": -0.002201397204771638, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002201397204771638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3365962505340576, + 0.0981457382440567, + 0.0350346639752388, + 0.04852169752120972, + 0.0868854746222496, + 0.03950328379869461, + 0.1776965707540512, + 0.029676653444767, + 0.04292551428079605, + 0.1050141453742981 + ], + "credit_map": [ + 0.022229008376598358, + 0.030819345265626907, + 0.02439400553703308, + 0.022824173793196678, + 0.09624555706977844, + 0.06666410714387894, + 0.048928238451480865, + 0.024491827934980392, + 0.008261867798864841, + 0.008935335092246532, + 0.009421847760677338, + 0.00970072764903307, + 0.014064882881939411, + 0.014822476543486118, + 0.020049260929226875, + 0.012889334931969643, + 0.003179804654791951, + 0.0036752924788743258, + 0.003444844391196966, + 0.003177433041855693, + 0.004525301046669483, + 0.005393391940742731, + 0.004969378001987934, + 0.006669220048934221, + 0.00414002500474453, + 0.005343262571841478, + 0.0038527369033545256, + 0.00524884182959795, + 0.00903475470840931, + 0.006017676088958979, + 0.007648343686014414, + 0.007236052770167589, + 0.007679438684135675, + 0.011870614252984524, + 0.00785231962800026, + 0.007127410266548395, + 0.013078580610454082, + 0.012842519208788872, + 0.01243588700890541, + 0.013998709619045258, + 0.002759183058515191, + 0.0033114133402705193, + 0.002220381749793887, + 0.00232198485173285, + 0.004120166413486004, + 0.007461058907210827, + 0.004392110276967287, + 0.012916984036564827, + 0.010219058021903038, + 0.007698629051446915, + 0.012544816359877586, + 0.00989395473152399, + 0.027417495846748352, + 0.041167423129081726, + 0.017268480733036995, + 0.05148670822381973, + 0.003305459162220359, + 0.002883704612031579, + 0.0016888855025172234, + 0.0020406057592481375, + 0.002270651515573263, + 0.006393826100975275, + 0.004158757161349058, + 0.00693476339802146, + 0.004983444698154926, + 0.005265587009489536, + 0.003831184469163418, + 0.005233550909906626, + 0.003086633747443557, + 0.005074759945273399, + 0.007920905947685242, + 0.0075294505804777145, + 0.030093031004071236, + 0.03456653282046318, + 0.004474739544093609, + 0.008620851673185825, + 0.00859429407864809, + 0.008871727623045444, + 0.005718868225812912, + 0.004074094817042351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.002201397204771638 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040924143977463245 + ], + "opposite_logprob_tokens": [ + -5.504092216491699 + ], + "logprob_reward": -0.0040924143977463245, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040924143977463245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2122337371110916, + 0.04077191650867462, + 0.021180318668484688, + 0.013633384369313717, + 0.006852834019809961, + 0.028689395636320114, + 0.22727619111537933, + 0.044375546276569366, + 0.011427880264818668, + 0.3935587704181671 + ], + "credit_map": [ + 0.012542196549475193, + 0.018798595294356346, + 0.016633648425340652, + 0.015079185366630554, + 0.05493497848510742, + 0.0389912948012352, + 0.04158639535307884, + 0.013667440041899681, + 0.004774364642798901, + 0.0034487880766391754, + 0.004135924857109785, + 0.004321685526520014, + 0.0051916115917265415, + 0.006020287983119488, + 0.006578163709491491, + 0.006301088258624077, + 0.001606380334123969, + 0.0016933107981458306, + 0.0024626252707093954, + 0.001807328430004418, + 0.0020383407827466726, + 0.0034434786066412926, + 0.004118041135370731, + 0.004010813310742378, + 0.0011707248631864786, + 0.0010771804954856634, + 0.0015353377675637603, + 0.0012228096602484584, + 0.0021758058574050665, + 0.00179445487447083, + 0.002612148644402623, + 0.0020449229050427675, + 0.0006584635702893138, + 0.0007333646644838154, + 0.0006555855507031083, + 0.0006773641798645258, + 0.0010674595832824707, + 0.0009827091125771403, + 0.0006481968448497355, + 0.0014296900480985641, + 0.0017254246631637216, + 0.0037390422075986862, + 0.0026682831812649965, + 0.0026986179873347282, + 0.004138980992138386, + 0.004218721762299538, + 0.005815884564071894, + 0.0036844401620328426, + 0.012479842640459538, + 0.010116415098309517, + 0.013322665356099606, + 0.011386127211153507, + 0.031337227672338486, + 0.05625539645552635, + 0.020905571058392525, + 0.07147294282913208, + 0.0016894075088202953, + 0.0017811020370572805, + 0.001854033675044775, + 0.0018286585109308362, + 0.0020075375214219093, + 0.013848534785211086, + 0.006332018878310919, + 0.015034254640340805, + 0.001953340834006667, + 0.0010478950571268797, + 0.001559723517857492, + 0.0010817504953593016, + 0.0010896273888647556, + 0.0010823054471984506, + 0.001987034222111106, + 0.0016262034187093377, + 0.11470448225736618, + 0.1316148340702057, + 0.015950879082083702, + 0.030259907245635986, + 0.03288035839796066, + 0.03210841491818428, + 0.02118789777159691, + 0.014851987361907959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.0040924143977463245 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17985749244689941 + ], + "opposite_logprob_tokens": [ + -1.8048574924468994 + ], + "logprob_reward": -0.17985749244689941, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17985749244689941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3925934433937073, + 0.040872685611248016, + 0.02265574410557747, + 0.009883754886686802, + 0.012342550791800022, + 0.020083075389266014, + 0.34622639417648315, + 0.022186176851391792, + 0.031773924827575684, + 0.10138224065303802 + ], + "credit_map": [ + 0.025411849841475487, + 0.033037908375263214, + 0.029829243198037148, + 0.02540992572903633, + 0.10540211945772171, + 0.07414934784173965, + 0.06983782351016998, + 0.029515238478779793, + 0.0056763882748782635, + 0.003361063776537776, + 0.0036373198963701725, + 0.0045285834930837154, + 0.005648552440106869, + 0.006031944882124662, + 0.006585623137652874, + 0.005403208080679178, + 0.0014864178374409676, + 0.0016114021418616176, + 0.0026203750167042017, + 0.0015842561842873693, + 0.0024024264421314, + 0.0038550281897187233, + 0.004589335061609745, + 0.004506504628807306, + 0.0009594092261977494, + 0.0007521332008764148, + 0.001099656685255468, + 0.0008288666140288115, + 0.0016216650838032365, + 0.0014442414976656437, + 0.0015743132680654526, + 0.001603469019755721, + 0.0013466042000800371, + 0.0017006404232233763, + 0.0013221101835370064, + 0.0009685942204669118, + 0.0018737755017355084, + 0.0015623797662556171, + 0.001257829717360437, + 0.002310617594048381, + 0.0015054515097290277, + 0.0020365086384117603, + 0.0021752233151346445, + 0.0019994403701275587, + 0.0034127316903322935, + 0.002956353360787034, + 0.003959865774959326, + 0.0020375007297843695, + 0.018545806407928467, + 0.01535259559750557, + 0.017887821421027184, + 0.01775708608329296, + 0.05062678083777428, + 0.08631423115730286, + 0.03164075314998627, + 0.10810130834579468, + 0.000989024294540286, + 0.001002232194878161, + 0.0008862836984917521, + 0.0015552602708339691, + 0.0010738956043496728, + 0.005241330713033676, + 0.0053482139483094215, + 0.006089936941862106, + 0.0030359099619090557, + 0.0029068896546959877, + 0.003692119847983122, + 0.003338351147249341, + 0.004156207665801048, + 0.0033197514712810516, + 0.005934993736445904, + 0.005389702506363392, + 0.029566174373030663, + 0.03375265747308731, + 0.004317575599998236, + 0.007761766668409109, + 0.008402504026889801, + 0.008144148625433445, + 0.005555685609579086, + 0.003881720593199134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.17985749244689941 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, sweeping long shot, the camera slowly and steadily tilts upward from the horizon, gradually revealing a nightscape as deep as ink, dotted with distant nebulae. Initially, only a few diamond-bright stars begin to fall, but soon, a silent and magnificent cosmic spectacle unfolds. Thousands of blazing points of light, a blend of icy blue and warm gold, fall like tears from the velvety darkness, each leaving a slender, brilliant trail of light in their wake. These fleeting burning trails interweave into a dynamic web of light. Some are as long as silver threads, gracefully streaking across the sky, while others flash like fleeting sparks, their light briefly illuminating one another before fading, bringing a fleeting vitality to the vast darkness. At the bottom of the frame, the black silhouettes of the mountains stand silent witness to this celestial ballet. The entire scene, shrouded in absolute tranquility, exudes an awe-inspiring grandeur and majesty, creating a visual atmosphere both dreamlike and imbued with cosmic power.", + "seed": 15651, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_3/index_2516_seed_15651_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03525313118007034, + "vlm_logprob": -0.03525313118007034, + "vlm_yes_no_margin": 4.524999976158142, + "vlm_token_credit": -0.03525313118007034, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13343536853790283 + ], + "opposite_logprob_tokens": [ + -2.1334352493286133 + ], + "logprob_reward": -0.13343536853790283, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.13343536853790283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26606160402297974, + 0.14961141347885132, + 0.0925997644662857, + 0.08529415726661682, + 0.14470021426677704, + 0.05681914836168289, + 0.030304592102766037, + 0.11607473343610764, + 0.033289629966020584, + 0.02524476684629917 + ], + "credit_map": [ + 0.02065795660018921, + 0.0532318539917469, + 0.025903457775712013, + 0.01984112337231636, + 0.044920843094587326, + 0.03427811339497566, + 0.020857064053416252, + 0.046371154487133026, + 0.015233833342790604, + 0.022142929956316948, + 0.01216948963701725, + 0.015798384323716164, + 0.022392278537154198, + 0.02871527522802353, + 0.017328031361103058, + 0.01583118364214897, + 0.008202336728572845, + 0.014708122238516808, + 0.007923324592411518, + 0.01097960863262415, + 0.01792112924158573, + 0.013758641667664051, + 0.011085855774581432, + 0.008020737208425999, + 0.009453493170440197, + 0.00640813447535038, + 0.009211421944200993, + 0.012141573242843151, + 0.012108304537832737, + 0.012979712337255478, + 0.009683233685791492, + 0.013308274559676647, + 0.00848415493965149, + 0.010754200629889965, + 0.007093188352882862, + 0.022093381732702255, + 0.009934358298778534, + 0.025931699201464653, + 0.021624786779284477, + 0.0387844443321228, + 0.0053468369878828526, + 0.008700838312506676, + 0.0037417535204440355, + 0.004415544215589762, + 0.007789960131049156, + 0.008612350560724735, + 0.010019328445196152, + 0.008192530833184719, + 0.005446219351142645, + 0.004268571734428406, + 0.003493835451081395, + 0.002875428879633546, + 0.0037618400529026985, + 0.00406267773360014, + 0.0022936619352549314, + 0.004102356731891632, + 0.010160955600440502, + 0.006653136573731899, + 0.007574380841106176, + 0.03156975284218788, + 0.011609957553446293, + 0.013788105919957161, + 0.006215644534677267, + 0.028502797707915306, + 0.004404782317578793, + 0.0037644358817487955, + 0.004604194778949022, + 0.003844101447612047, + 0.00353535870090127, + 0.0028135499451309443, + 0.006069491151720285, + 0.004253714811056852, + 0.0038361893966794014, + 0.003836305346339941, + 0.004091102629899979, + 0.004896420519798994, + 0.0013547211419790983, + 0.0026488788425922394, + 0.002417381154373288, + 0.002163765486329794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.13343536853790283 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006754779256880283 + ], + "opposite_logprob_tokens": [ + -5.0067548751831055 + ], + "logprob_reward": -0.006754779256880283, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006754779256880283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3146294057369232, + 0.16042931377887726, + 0.12204787135124207, + 0.0786934494972229, + 0.1114443764090538, + 0.06739708036184311, + 0.03332972526550293, + 0.03751448169350624, + 0.022373981773853302, + 0.05214034020900726 + ], + "credit_map": [ + 0.025110667571425438, + 0.050827670842409134, + 0.03185790404677391, + 0.026047786697745323, + 0.04830063879489899, + 0.0387321338057518, + 0.022936871275305748, + 0.07081576436758041, + 0.013898106291890144, + 0.027815749868750572, + 0.015677176415920258, + 0.015998167917132378, + 0.028312809765338898, + 0.027276773005723953, + 0.014881068840622902, + 0.016569463536143303, + 0.02213430218398571, + 0.012178018689155579, + 0.011852951720356941, + 0.012709712609648705, + 0.017126085236668587, + 0.024604635313153267, + 0.010898503474891186, + 0.010543652810156345, + 0.005802841857075691, + 0.007460087537765503, + 0.00873582810163498, + 0.009823731146752834, + 0.01100100576877594, + 0.011161704547703266, + 0.013834905810654163, + 0.010873345658183098, + 0.00813201256096363, + 0.008205660618841648, + 0.005928095430135727, + 0.017880406230688095, + 0.006491770036518574, + 0.01873350702226162, + 0.018597465008497238, + 0.027475465089082718, + 0.005054813344031572, + 0.0057645076885819435, + 0.005833850707858801, + 0.0039042376447468996, + 0.011553531512618065, + 0.012726531364023685, + 0.007354026660323143, + 0.015205577947199345, + 0.005391762591898441, + 0.0033628700766712427, + 0.0028416698332875967, + 0.0021966099739074707, + 0.005379591137170792, + 0.005236941389739513, + 0.0027468770276755095, + 0.006173400208353996, + 0.0041356743313372135, + 0.0034519059117883444, + 0.006003260612487793, + 0.004224829841405153, + 0.006746291182935238, + 0.005532211624085903, + 0.004265305586159229, + 0.0031550040002912283, + 0.0034193519968539476, + 0.002038045087829232, + 0.0030593497212976217, + 0.002330839866772294, + 0.001576601411215961, + 0.0019451036350801587, + 0.004516503773629665, + 0.0034881855826824903, + 0.008029657416045666, + 0.007378286682069302, + 0.007445324212312698, + 0.008066537790000439, + 0.003398837521672249, + 0.006511261221021414, + 0.0060801119543612, + 0.005230320617556572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.006754779256880283 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010446811793372035 + ], + "opposite_logprob_tokens": [ + -6.876044750213623 + ], + "logprob_reward": -0.0010446811793372035, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010446811793372035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09609023481607437, + 0.1159428134560585, + 0.07031309604644775, + 0.12272949516773224, + 0.14618980884552002, + 0.09245629608631134, + 0.050494033843278885, + 0.23298804461956024, + 0.03240138292312622, + 0.040394823998212814 + ], + "credit_map": [ + 0.008934472687542439, + 0.015053718350827694, + 0.011508251540362835, + 0.010751831345260143, + 0.014389026910066605, + 0.012102221138775349, + 0.007449537981301546, + 0.015901170670986176, + 0.010091698728501797, + 0.013307450339198112, + 0.02619076892733574, + 0.021921690553426743, + 0.010231433436274529, + 0.012640220113098621, + 0.012001702561974525, + 0.009557853452861309, + 0.009275175631046295, + 0.008597411215305328, + 0.008468588814139366, + 0.008404462598264217, + 0.007630783598870039, + 0.00867318082600832, + 0.008259971626102924, + 0.011003520339727402, + 0.012784168124198914, + 0.012288903817534447, + 0.009943314827978611, + 0.01389277447015047, + 0.018187178298830986, + 0.018985582515597343, + 0.016529669985175133, + 0.020117908716201782, + 0.011550063267350197, + 0.01147613488137722, + 0.008052906021475792, + 0.020957471802830696, + 0.011277534998953342, + 0.024316715076565742, + 0.02453484944999218, + 0.034024130553007126, + 0.010675880126655102, + 0.006728082429617643, + 0.006460200995206833, + 0.008327032439410686, + 0.011727586388587952, + 0.015143468976020813, + 0.015409295447170734, + 0.017984751611948013, + 0.006579618901014328, + 0.007702026516199112, + 0.007143309339880943, + 0.005130454432219267, + 0.006327498704195023, + 0.006250410806387663, + 0.004259804263710976, + 0.0071009122766554356, + 0.037574365735054016, + 0.012888411059975624, + 0.01443258486688137, + 0.0793023481965065, + 0.014282571151852608, + 0.015496557578444481, + 0.015724916011095047, + 0.043286290019750595, + 0.0036490331403911114, + 0.0038105188868939877, + 0.005459851585328579, + 0.0036685392260551453, + 0.0038942827377468348, + 0.003782192710787058, + 0.004263855051249266, + 0.003873111680150032, + 0.006809409707784653, + 0.00408920319750905, + 0.006648296024650335, + 0.005492337979376316, + 0.003219833131879568, + 0.005364309996366501, + 0.003974501509219408, + 0.00479693291708827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.0010446811793372035 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052490513771772385 + ], + "opposite_logprob_tokens": [ + -5.2552490234375 + ], + "logprob_reward": -0.0052490513771772385, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052490513771772385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19363689422607422, + 0.21318000555038452, + 0.06616188585758209, + 0.09828568994998932, + 0.20829136669635773, + 0.062425728887319565, + 0.022519880905747414, + 0.07465547323226929, + 0.026100268587470055, + 0.03474285826086998 + ], + "credit_map": [ + 0.019452432170510292, + 0.024841977283358574, + 0.016424044966697693, + 0.018361764028668404, + 0.023438841104507446, + 0.04796408861875534, + 0.016289442777633667, + 0.02686428278684616, + 0.02168109081685543, + 0.021495675668120384, + 0.0175889041274786, + 0.03250334411859512, + 0.027798449620604515, + 0.05198925733566284, + 0.020121769979596138, + 0.02000150829553604, + 0.008935580030083656, + 0.008212469518184662, + 0.00897008553147316, + 0.008471295237541199, + 0.00626835972070694, + 0.008148116990923882, + 0.006684368010610342, + 0.010471606627106667, + 0.004874823614954948, + 0.007192567922174931, + 0.010689505375921726, + 0.01401508692651987, + 0.013724885880947113, + 0.014219282194972038, + 0.019122041761875153, + 0.01444749254733324, + 0.012670349329710007, + 0.014864323660731316, + 0.009551635012030602, + 0.03001430071890354, + 0.01464927475899458, + 0.03813493251800537, + 0.03377080708742142, + 0.05463571846485138, + 0.0043632108718156815, + 0.004799274727702141, + 0.003620634786784649, + 0.006506378762423992, + 0.006425680126994848, + 0.010350861586630344, + 0.015166491270065308, + 0.01119319349527359, + 0.0036747977137565613, + 0.0031387226190418005, + 0.0028552659787237644, + 0.0020471017342060804, + 0.0028831984382122755, + 0.002719396725296974, + 0.0015786256408318877, + 0.003622770309448242, + 0.009346665814518929, + 0.0040154410526156425, + 0.004766434896737337, + 0.0227672066539526, + 0.005711755715310574, + 0.008192095905542374, + 0.005821113474667072, + 0.014034753665328026, + 0.003879163647070527, + 0.0027655307203531265, + 0.004052328411489725, + 0.0031982026994228363, + 0.002691519446671009, + 0.0022826138883829117, + 0.0036024898290634155, + 0.0036284183152019978, + 0.007329214364290237, + 0.004305815324187279, + 0.007845774292945862, + 0.0036074905656278133, + 0.0022421469911932945, + 0.0031921996269375086, + 0.002994852140545845, + 0.003225363790988922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.0052490513771772385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029781775549054146 + ], + "opposite_logprob_tokens": [ + -3.5297818183898926 + ], + "logprob_reward": -0.029781775549054146, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029781775549054146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2417127937078476, + 0.29813453555107117, + 0.13220550119876862, + 0.09130756556987762, + 0.06529446691274643, + 0.04927567392587662, + 0.03194647282361984, + 0.045894984155893326, + 0.017040381208062172, + 0.02718752808868885 + ], + "credit_map": [ + 0.02260403148829937, + 0.04310735687613487, + 0.023972386494278908, + 0.021206045523285866, + 0.03455005958676338, + 0.028227170929312706, + 0.017162518575787544, + 0.05088326707482338, + 0.023135971277952194, + 0.05970609188079834, + 0.02798612229526043, + 0.027175797149538994, + 0.05888398364186287, + 0.03845682367682457, + 0.037615615874528885, + 0.02517418935894966, + 0.01510866079479456, + 0.01754148304462433, + 0.011861167848110199, + 0.013297602534294128, + 0.02222837321460247, + 0.02072005718946457, + 0.0166572704911232, + 0.014790905639529228, + 0.005787730682641268, + 0.008225226774811745, + 0.0060316696763038635, + 0.010435196571052074, + 0.014776987954974174, + 0.013566332869231701, + 0.020325323566794395, + 0.012159109115600586, + 0.0041122762486338615, + 0.005282296333462, + 0.003053118009120226, + 0.0076481555588543415, + 0.004915711935609579, + 0.012531430460512638, + 0.009319615550339222, + 0.018431873992085457, + 0.003958076238632202, + 0.004874648991972208, + 0.0029050286393612623, + 0.0043191672302782536, + 0.007125032600015402, + 0.008570976555347443, + 0.0071749915368855, + 0.01034776121377945, + 0.0063141342252492905, + 0.00461956113576889, + 0.004428556188941002, + 0.0028186896815896034, + 0.0039577241986989975, + 0.004184544552117586, + 0.0021667396649718285, + 0.0034565269015729427, + 0.003063452197238803, + 0.004614564590156078, + 0.004117416217923164, + 0.0038723256438970566, + 0.009315431118011475, + 0.009110555052757263, + 0.004585447255522013, + 0.0072157979011535645, + 0.002643297426402569, + 0.0013384674675762653, + 0.002752464497461915, + 0.0018551918910816312, + 0.0017990188207477331, + 0.0019515877356752753, + 0.0026505275163799524, + 0.0020498286467045546, + 0.0052048214711248875, + 0.003600857686251402, + 0.004783061798661947, + 0.0043556480668485165, + 0.0020667763892561197, + 0.002541665220633149, + 0.0022802110761404037, + 0.0023544924333691597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.029781775549054146 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, sweeping long shot, the camera slowly and steadily tilts upward from the horizon, gradually revealing a nightscape as deep as ink, dotted with distant nebulae. Initially, only a few diamond-bright stars begin to fall, but soon, a silent and magnificent cosmic spectacle unfolds. Thousands of blazing points of light, a blend of icy blue and warm gold, fall like tears from the velvety darkness, each leaving a slender, brilliant trail of light in their wake. These fleeting burning trails interweave into a dynamic web of light. Some are as long as silver threads, gracefully streaking across the sky, while others flash like fleeting sparks, their light briefly illuminating one another before fading, bringing a fleeting vitality to the vast darkness. At the bottom of the frame, the black silhouettes of the mountains stand silent witness to this celestial ballet. The entire scene, shrouded in absolute tranquility, exudes an awe-inspiring grandeur and majesty, creating a visual atmosphere both dreamlike and imbued with cosmic power.", + "seed": 15652, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_3/index_2516_seed_15652_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.060634191939607264, + "vlm_logprob": -0.060634191939607264, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.060634191939607264, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.26349011063575745 + ], + "opposite_logprob_tokens": [ + -1.513490080833435 + ], + "logprob_reward": -0.26349011063575745, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.26349011063575745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14443433284759521, + 0.13247692584991455, + 0.1225627064704895, + 0.12503384053707123, + 0.133555606007576, + 0.15773262083530426, + 0.08729039877653122, + 0.03816971182823181, + 0.018004950135946274, + 0.04073897749185562 + ], + "credit_map": [ + 0.01244238018989563, + 0.027595629915595055, + 0.01683574728667736, + 0.014549441635608673, + 0.019366925582289696, + 0.023274891078472137, + 0.01480028498917818, + 0.015569017268717289, + 0.011703545227646828, + 0.025023404508829117, + 0.014411713927984238, + 0.020590927451848984, + 0.015764642506837845, + 0.015749553218483925, + 0.014898515306413174, + 0.014334614388644695, + 0.012198773212730885, + 0.01787400245666504, + 0.01857621781527996, + 0.017137369140982628, + 0.017600372433662415, + 0.02000471018254757, + 0.009058021008968353, + 0.010113224387168884, + 0.018997397273778915, + 0.013379707001149654, + 0.011773541569709778, + 0.019618356600403786, + 0.021932238712906837, + 0.018708940595388412, + 0.007772811688482761, + 0.012850836850702763, + 0.02142982929944992, + 0.02010318636894226, + 0.014479144476354122, + 0.011334898881614208, + 0.018949516117572784, + 0.018777361139655113, + 0.013145782053470612, + 0.015335876494646072, + 0.01666741818189621, + 0.018936224281787872, + 0.015040058642625809, + 0.02381017431616783, + 0.01742631010711193, + 0.01846792735159397, + 0.024289118126034737, + 0.023095376789569855, + 0.00789640937000513, + 0.010152600705623627, + 0.003989403136074543, + 0.024663416668772697, + 0.006777515634894371, + 0.012357513420283794, + 0.011629986576735973, + 0.009823544882237911, + 0.006537622306495905, + 0.0071543059311807156, + 0.002415822586044669, + 0.004380018450319767, + 0.003638766473159194, + 0.00508233904838562, + 0.0043914648704230785, + 0.004569364711642265, + 0.0031656636856496334, + 0.002492013154551387, + 0.0019820379093289375, + 0.0020839013159275055, + 0.0016123654786497355, + 0.002378180855885148, + 0.002197862369939685, + 0.0020929223392158747, + 0.0068697514943778515, + 0.005828334484249353, + 0.0064333099871873856, + 0.006110073532909155, + 0.003808497218415141, + 0.004188132472336292, + 0.0028849474620074034, + 0.0046159266494214535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.26349011063575745 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006760344374924898 + ], + "opposite_logprob_tokens": [ + -5.006760120391846 + ], + "logprob_reward": -0.006760344374924898, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006760344374924898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07082246243953705, + 0.06835734844207764, + 0.08619391918182373, + 0.10097233206033707, + 0.19349369406700134, + 0.07311579585075378, + 0.07706104964017868, + 0.026833271607756615, + 0.028174204751849174, + 0.2749759256839752 + ], + "credit_map": [ + 0.007183210924267769, + 0.0128953130915761, + 0.00523525383323431, + 0.00622355192899704, + 0.012809431180357933, + 0.01605495810508728, + 0.005206620320677757, + 0.005214125849306583, + 0.005475467536598444, + 0.009103563614189625, + 0.013178764842450619, + 0.008430663496255875, + 0.008836009539663792, + 0.00834759883582592, + 0.004702101927250624, + 0.01028317492455244, + 0.006979326717555523, + 0.01297325175255537, + 0.008322279900312424, + 0.013891536742448807, + 0.012302477844059467, + 0.01749272085726261, + 0.007258332800120115, + 0.006973998621106148, + 0.013225260190665722, + 0.011196906678378582, + 0.01005579438060522, + 0.0120473001152277, + 0.019456833600997925, + 0.017766863107681274, + 0.007566110230982304, + 0.009657264687120914, + 0.03755760192871094, + 0.027909379452466965, + 0.02270158939063549, + 0.0140152582898736, + 0.027091076597571373, + 0.029335366562008858, + 0.015908021479845047, + 0.018975382670760155, + 0.011447800323367119, + 0.012072842568159103, + 0.006560708861798048, + 0.008037688210606575, + 0.010044167749583721, + 0.009014376439154148, + 0.006925125140696764, + 0.00901308748871088, + 0.005653714295476675, + 0.008268066681921482, + 0.004779360722750425, + 0.011388196609914303, + 0.01163160614669323, + 0.012861973606050014, + 0.009524278342723846, + 0.01295385230332613, + 0.004051532596349716, + 0.0045672254636883736, + 0.001909084734506905, + 0.0034357558470219374, + 0.0026406776160001755, + 0.0031402313616126776, + 0.0037290481850504875, + 0.0033597152214497328, + 0.00474184937775135, + 0.004847485572099686, + 0.002048596739768982, + 0.004108937457203865, + 0.0022515770979225636, + 0.003264104714617133, + 0.003035326022654772, + 0.003876329166814685, + 0.03175320476293564, + 0.03540622442960739, + 0.019001435488462448, + 0.07228987663984299, + 0.030506398528814316, + 0.031708795577287674, + 0.02646632306277752, + 0.027843663468956947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.006760344374924898 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013405631761997938 + ], + "opposite_logprob_tokens": [ + -6.626340389251709 + ], + "logprob_reward": -0.0013405631761997938, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013405631761997938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09430433809757233, + 0.06489013135433197, + 0.04558440297842026, + 0.05238828808069229, + 0.16883817315101624, + 0.05170707404613495, + 0.40753668546676636, + 0.02414780482649803, + 0.020959297195076942, + 0.0696437656879425 + ], + "credit_map": [ + 0.0084553062915802, + 0.013768044300377369, + 0.012131349183619022, + 0.008016303181648254, + 0.014970946125686169, + 0.02140631340444088, + 0.006036986596882343, + 0.009519091807305813, + 0.005362608935683966, + 0.009163076058030128, + 0.011420498602092266, + 0.009465415962040424, + 0.008623439818620682, + 0.007507917005568743, + 0.006065560970455408, + 0.007281618658453226, + 0.004357323050498962, + 0.005606128368526697, + 0.007737568113952875, + 0.00580438831821084, + 0.006295142695307732, + 0.007682847324758768, + 0.00407362449914217, + 0.0040273829363286495, + 0.0046762204729020596, + 0.004154362250119448, + 0.00872915517538786, + 0.012104435823857784, + 0.005534553900361061, + 0.007757862564176321, + 0.004717692267149687, + 0.0047140056267380714, + 0.028289705514907837, + 0.02371673844754696, + 0.01748616248369217, + 0.014827847480773926, + 0.025369489565491676, + 0.023891927674412727, + 0.017751038074493408, + 0.017505262047052383, + 0.0066575245000422, + 0.005926880519837141, + 0.006287123542279005, + 0.006529742851853371, + 0.005946523044258356, + 0.006677401717752218, + 0.006489806808531284, + 0.007192070595920086, + 0.024942941963672638, + 0.03365016356110573, + 0.021488191559910774, + 0.04646570235490799, + 0.0507470928132534, + 0.09429369866847992, + 0.04717453569173813, + 0.08877438306808472, + 0.0031454218551516533, + 0.004267200827598572, + 0.0021383946295827627, + 0.0030757111962884665, + 0.003317475551739335, + 0.003076150082051754, + 0.002356570214033127, + 0.002770880237221718, + 0.0030332638416439295, + 0.003313752356916666, + 0.0017383538652211428, + 0.003824135521426797, + 0.0020070220343768597, + 0.0020335367880761623, + 0.002783925738185644, + 0.0022253082133829594, + 0.0077442508190870285, + 0.009265173226594925, + 0.006456603761762381, + 0.01307184062898159, + 0.009546211920678616, + 0.009525278583168983, + 0.0064942119643092155, + 0.007540194783359766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.0013405631761997938 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250118672847748 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005250118672847748, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250118672847748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3016301095485687, + 0.34774041175842285, + 0.03341757878661156, + 0.07522105425596237, + 0.05091806873679161, + 0.029875632375478745, + 0.08996516466140747, + 0.03418824076652527, + 0.013236620463430882, + 0.023807121440768242 + ], + "credit_map": [ + 0.02374756895005703, + 0.046671945601701736, + 0.04003912955522537, + 0.031502291560173035, + 0.048459894955158234, + 0.06388707458972931, + 0.02516140602529049, + 0.02216079831123352, + 0.029094045981764793, + 0.060329098254442215, + 0.033742502331733704, + 0.04292724281549454, + 0.049294814467430115, + 0.06999233365058899, + 0.02460845373570919, + 0.03775192052125931, + 0.0031141554936766624, + 0.004253850318491459, + 0.005178992170840502, + 0.004976820666342974, + 0.004513517953455448, + 0.004973174072802067, + 0.0035970425233244896, + 0.0028100269846618176, + 0.005966024473309517, + 0.006124471314251423, + 0.00997344870120287, + 0.013261894695460796, + 0.010334526188671589, + 0.011753775179386139, + 0.008921071887016296, + 0.008885848335921764, + 0.007752986624836922, + 0.007037116680294275, + 0.006275833118706942, + 0.004735933616757393, + 0.007561706006526947, + 0.007497178856283426, + 0.005601896438747644, + 0.004455419257283211, + 0.002420402830466628, + 0.0030094683170318604, + 0.003881329670548439, + 0.0036438205279409885, + 0.0035277376882731915, + 0.004317114595323801, + 0.00550302118062973, + 0.0035727382637560368, + 0.0038460581563413143, + 0.0032556115183979273, + 0.0020323239732533693, + 0.009721917100250721, + 0.005122073460370302, + 0.029164817184209824, + 0.005592883098870516, + 0.031229477375745773, + 0.0055670589208602905, + 0.006643774453550577, + 0.0019753144588321447, + 0.004484772682189941, + 0.0036303449887782335, + 0.0040659173391759396, + 0.0037315506488084793, + 0.00408950773999095, + 0.002018383936956525, + 0.0019345867913216352, + 0.0009848270565271378, + 0.0019285681191831827, + 0.0013910895213484764, + 0.001665290561504662, + 0.0020084602292627096, + 0.0013054133160039783, + 0.004678105469793081, + 0.00364622101187706, + 0.0028198969084769487, + 0.0037114270962774754, + 0.0023559951223433018, + 0.0017432378372177482, + 0.00213452847674489, + 0.0027177107986062765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.005250118672847748 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera start with a long shot of the horizon and slowly tilt upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thousands of falling lights visible in the sky, blending icy blue and warm gold colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the falling lights leave slender, brilliant trails of light as they streak across the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are black mountain silhouettes visible at the bottom of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a dark nightscape dotted with distant nebulae and stars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026329822838306427 + ], + "opposite_logprob_tokens": [ + -3.651329755783081 + ], + "logprob_reward": -0.026329822838306427, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026329822838306427, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2253054976463318, + 0.2375936657190323, + 0.08557054400444031, + 0.10477302968502045, + 0.0913003534078598, + 0.06813392788171768, + 0.10098268836736679, + 0.028919203206896782, + 0.017423640936613083, + 0.039997417479753494 + ], + "credit_map": [ + 0.019324922934174538, + 0.03377464413642883, + 0.0255456380546093, + 0.02925817482173443, + 0.035323940217494965, + 0.04880322515964508, + 0.019441206008195877, + 0.013833750039339066, + 0.020131228491663933, + 0.023077379912137985, + 0.042896296828985214, + 0.021629251539707184, + 0.03188575431704521, + 0.046436745673418045, + 0.017008990049362183, + 0.03452801704406738, + 0.006196603178977966, + 0.016673725098371506, + 0.00802683923393488, + 0.014398172497749329, + 0.01444930024445057, + 0.013694331049919128, + 0.006802874617278576, + 0.005328693427145481, + 0.011404924094676971, + 0.012587428092956543, + 0.013292334973812103, + 0.023403948172926903, + 0.012071184813976288, + 0.01125621609389782, + 0.010618278756737709, + 0.010138722136616707, + 0.01562506891787052, + 0.011505726724863052, + 0.01021824311465025, + 0.007053219713270664, + 0.012543362565338612, + 0.01363824587315321, + 0.007930404506623745, + 0.012786087580025196, + 0.0057152630761265755, + 0.009912063367664814, + 0.007266745436936617, + 0.009015378542244434, + 0.010332885198295116, + 0.009122032672166824, + 0.008966542780399323, + 0.007803018670529127, + 0.007111345883458853, + 0.010187210515141487, + 0.006099850405007601, + 0.014151670038700104, + 0.012389807030558586, + 0.021337805315852165, + 0.009841294027864933, + 0.019863713532686234, + 0.004092499613761902, + 0.0041290041990578175, + 0.0018851864151656628, + 0.003601985052227974, + 0.0027174476999789476, + 0.004008552059531212, + 0.003584019374102354, + 0.004900509025901556, + 0.0030887608882039785, + 0.0031482726335525513, + 0.0015650218119844794, + 0.001960056833922863, + 0.0016004890203475952, + 0.0021218140609562397, + 0.0016553803579881787, + 0.0022838464938104153, + 0.006141254678368568, + 0.004997477401047945, + 0.004536911845207214, + 0.00940627045929432, + 0.0037322980351746082, + 0.0042854128405451775, + 0.0033960281871259212, + 0.0035017619375139475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.026329822838306427 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_4/index_tensor([5186, 5186])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_4/index_tensor([5186, 5186])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c1140deda930322de2eb4aedfcd3feb3108a8f50 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_4/index_tensor([5186, 5186])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures a surreal and photoreal confrontation within the booth of a classic American diner. On the left, an anthropomorphic burger, its head and torso formed by a smooth, sesame-studded brioche bun, leans forward aggressively. Two black olive slices act as wide, angry eyes, while a curved pickle slice opens its mouth, revealing layers of thick, grilled patty, crisp lettuce, and tomato slices. Its spindly French fry arms gesture wildly; one defiant, stiff French fry finger points directly at its companion, while another is raised angrily as a drop of melted cheddar cheese drips down its side. Opposite, an anthropomorphic pepperoni pizza slice, its pointy crust forming a head, evokes the same passion. Two glistening black olives half form its narrowed, angry eyes, while its body, covered in glistening oil and melted mozzarella cheese, leans back defiantly. It waved one breadstick arm dismissively, while the other remained poised, as if it had just decisively slapped the stained Formica tabletop. They sat on cracked cherry-red vinyl sofas, their chrome trim and glass condiment shakers gleaming in the cool, sterile glow of the restaurant's overhead fluorescent lights. Through the large windows behind them, the city's blurred nighttime neon signs cast soft, colorful highlights, bathing the tense and comically absurd scene in a dreamy, atmospheric glow.", + "seed": 9253, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_4/index_5186_seed_9253_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03185599073767662, + "vlm_logprob": -0.03185599073767662, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.03185599073767662, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0615057647228241 + ], + "opposite_logprob_tokens": [ + -2.9365057945251465 + ], + "logprob_reward": -0.0615057647228241, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0615057647228241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2740807831287384, + 0.13096006214618683, + 0.06455846130847931, + 0.08515741676092148, + 0.03839024528861046, + 0.1570252776145935, + 0.05619467794895172, + 0.11309652030467987, + 0.04175098240375519, + 0.03878551349043846 + ], + "credit_map": [ + 0.022881846874952316, + 0.04602030664682388, + 0.036061715334653854, + 0.03311467543244362, + 0.022207610309123993, + 0.020026035606861115, + 0.029621083289384842, + 0.06414753198623657, + 0.018001126125454903, + 0.021052910014986992, + 0.01620326191186905, + 0.012225484475493431, + 0.025999736040830612, + 0.012101723812520504, + 0.012557539157569408, + 0.012818297371268272, + 0.006372392643243074, + 0.010469527915120125, + 0.00567887257784605, + 0.0052803028374910355, + 0.007161124609410763, + 0.004634787328541279, + 0.01629890874028206, + 0.008662549778819084, + 0.00687171146273613, + 0.008483132347464561, + 0.007778957020491362, + 0.01940624974668026, + 0.007916045375168324, + 0.007761313114315271, + 0.013166924007236958, + 0.01377309113740921, + 0.004060605075210333, + 0.006621524225920439, + 0.003148299176245928, + 0.0068189837038517, + 0.004934297408908606, + 0.0038201112765818834, + 0.00485493429005146, + 0.004131492227315903, + 0.012730051763355732, + 0.02567349188029766, + 0.009767930954694748, + 0.020666692405939102, + 0.013928455300629139, + 0.008528457023203373, + 0.03542244806885719, + 0.03030775673687458, + 0.006021909881383181, + 0.006913183256983757, + 0.004273961763828993, + 0.006760889198631048, + 0.010648517869412899, + 0.006852410733699799, + 0.008794588968157768, + 0.005929221399128437, + 0.008535520173609257, + 0.007846985012292862, + 0.012358473613858223, + 0.008236194960772991, + 0.008466560393571854, + 0.006663124542683363, + 0.03743080422282219, + 0.023558862507343292, + 0.003631799016147852, + 0.006483498960733414, + 0.003253264119848609, + 0.004426655825227499, + 0.004511274863034487, + 0.004364325664937496, + 0.010264805518090725, + 0.004815361462533474, + 0.004657182842493057, + 0.004702893551439047, + 0.00321433343924582, + 0.003163542365655303, + 0.00288272462785244, + 0.0030264602974057198, + 0.007411903236061335, + 0.00972647313028574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.0615057647228241 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041105784475803375 + ], + "opposite_logprob_tokens": [ + -5.504110813140869 + ], + "logprob_reward": -0.0041105784475803375, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041105784475803375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1525798738002777, + 0.06275840103626251, + 0.09935636818408966, + 0.08273304998874664, + 0.1040482446551323, + 0.2548235058784485, + 0.054788727313280106, + 0.0861661285161972, + 0.04250489920377731, + 0.06024080514907837 + ], + "credit_map": [ + 0.018107503652572632, + 0.021494345739483833, + 0.0157216377556324, + 0.022535154595971107, + 0.016545306891202927, + 0.016045793890953064, + 0.016338562592864037, + 0.025791576132178307, + 0.012795861810445786, + 0.007195004262030125, + 0.006873311474919319, + 0.010227743536233902, + 0.005468800198286772, + 0.005488285794854164, + 0.006815780885517597, + 0.007893613539636135, + 0.013908693566918373, + 0.020458392798900604, + 0.009621676988899708, + 0.007827025838196278, + 0.013679437339305878, + 0.011071225628256798, + 0.012940851971507072, + 0.00984906405210495, + 0.007035571616142988, + 0.017714353278279305, + 0.00659175356850028, + 0.01385200023651123, + 0.008602207526564598, + 0.0048521352000534534, + 0.013859860599040985, + 0.010225167497992516, + 0.009166732430458069, + 0.02443583868443966, + 0.008761939592659473, + 0.013771356083452702, + 0.010212126187980175, + 0.0044979979284107685, + 0.014399307779967785, + 0.0188029445707798, + 0.02015223167836666, + 0.024573801085352898, + 0.02465817518532276, + 0.052451059222221375, + 0.022103242576122284, + 0.019066931679844856, + 0.06092424690723419, + 0.03089381940662861, + 0.0049020880833268166, + 0.0171724371612072, + 0.003823639126494527, + 0.005402496550232172, + 0.004000158980488777, + 0.0039980448782444, + 0.011839624494314194, + 0.0036502403672784567, + 0.009118490852415562, + 0.028016969561576843, + 0.005826720967888832, + 0.00741036469116807, + 0.006846542935818434, + 0.006792224943637848, + 0.008994828909635544, + 0.013159988448023796, + 0.0036917226389050484, + 0.005663268733769655, + 0.004015314392745495, + 0.004423517733812332, + 0.0036585971247404814, + 0.0036235966254025698, + 0.012545332312583923, + 0.004883552435785532, + 0.012560244649648666, + 0.01542147807776928, + 0.004861368797719479, + 0.006058480124920607, + 0.004350188188254833, + 0.003465488087385893, + 0.00814448855817318, + 0.0053790658712387085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0041105784475803375 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073091067373753 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.011073091067373753, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073091067373753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11326520144939423, + 0.07228779047727585, + 0.14313490688800812, + 0.12966133654117584, + 0.05604090169072151, + 0.1700310856103897, + 0.08353862166404724, + 0.110542431473732, + 0.04733709990978241, + 0.07416065782308578 + ], + "credit_map": [ + 0.017732953652739525, + 0.024987468495965004, + 0.012843976728618145, + 0.015810195356607437, + 0.008960246108472347, + 0.008520245552062988, + 0.011602635495364666, + 0.012807471677660942, + 0.013446907512843609, + 0.009250237606465816, + 0.009625949896872044, + 0.01190903503447771, + 0.0059679364785552025, + 0.003831338370218873, + 0.010310282930731773, + 0.007946089841425419, + 0.015766222029924393, + 0.02419164776802063, + 0.014303128235042095, + 0.011691508814692497, + 0.011388433165848255, + 0.0074687618762254715, + 0.029784344136714935, + 0.028540847823023796, + 0.009759512729942799, + 0.037222739309072495, + 0.007658755872398615, + 0.013992289081215858, + 0.015876322984695435, + 0.00639942055568099, + 0.026881001889705658, + 0.011871284805238247, + 0.006662237457931042, + 0.005305571015924215, + 0.006280621979385614, + 0.008132928051054478, + 0.004933840595185757, + 0.0029356414452195168, + 0.010261893272399902, + 0.011528164148330688, + 0.015041252598166466, + 0.05393645539879799, + 0.006902438122779131, + 0.00828129518777132, + 0.02105003595352173, + 0.008494915440678596, + 0.04050295054912567, + 0.01582171954214573, + 0.011782143265008926, + 0.02567591890692711, + 0.007301687262952328, + 0.006549083162099123, + 0.006225390825420618, + 0.006188674364238977, + 0.012975518591701984, + 0.006840196903795004, + 0.00911856908351183, + 0.008376686833798885, + 0.011176381260156631, + 0.007155990228056908, + 0.007757924031466246, + 0.007193293888121843, + 0.03457501158118248, + 0.025188561528921127, + 0.004796864464879036, + 0.010622018948197365, + 0.004472543951123953, + 0.004748850595206022, + 0.005260964389890432, + 0.0045191324315965176, + 0.008078327402472496, + 0.004838393069803715, + 0.00591663271188736, + 0.01351077202707529, + 0.006652451120316982, + 0.00797775387763977, + 0.00483855651691556, + 0.003911102190613747, + 0.024657340720295906, + 0.006696037016808987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.011073091067373753 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533470056951046 + ], + "opposite_logprob_tokens": [ + -4.387533664703369 + ], + "logprob_reward": -0.012533470056951046, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533470056951046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09782196581363678, + 0.052013613283634186, + 0.03876960277557373, + 0.036917757242918015, + 0.0335090346634388, + 0.38887935876846313, + 0.09536778181791306, + 0.12597979605197906, + 0.040438953787088394, + 0.09030214697122574 + ], + "credit_map": [ + 0.010435325093567371, + 0.017167866230010986, + 0.011114038527011871, + 0.009906294755637646, + 0.013164049945771694, + 0.016555646434426308, + 0.008535362780094147, + 0.010943382978439331, + 0.009429743513464928, + 0.004382397513836622, + 0.007977589964866638, + 0.005523255094885826, + 0.006200629752129316, + 0.004016445484012365, + 0.005951414350420237, + 0.008532140403985977, + 0.0068356189876794815, + 0.004352059215307236, + 0.004126737359911203, + 0.004308655392378569, + 0.003635035827755928, + 0.003680312540382147, + 0.008354328572750092, + 0.0034768551122397184, + 0.0031265176367014647, + 0.00394289568066597, + 0.0036370528396219015, + 0.006859578657895327, + 0.00400249008089304, + 0.0030847196467220783, + 0.006038852035999298, + 0.0062256488017737865, + 0.002886355621740222, + 0.006305253598839045, + 0.003075846703723073, + 0.004307339899241924, + 0.0035469485446810722, + 0.0021131436806172132, + 0.007046188227832317, + 0.004227957688271999, + 0.029590358957648277, + 0.10912133753299713, + 0.020949505269527435, + 0.03623562306165695, + 0.045374345034360886, + 0.018606625497341156, + 0.0945405513048172, + 0.03446102887392044, + 0.009680547751486301, + 0.03481637313961983, + 0.005454934202134609, + 0.005006576422601938, + 0.008673718199133873, + 0.009791702963411808, + 0.01486075296998024, + 0.007083179894834757, + 0.011793847195804119, + 0.021989675238728523, + 0.012132642790675163, + 0.009984281845390797, + 0.009642313234508038, + 0.007794934790581465, + 0.03149718791246414, + 0.021144920960068703, + 0.003866003593429923, + 0.012234875932335854, + 0.0027377435471862555, + 0.004444943740963936, + 0.0035282380413264036, + 0.00513794319704175, + 0.003623546566814184, + 0.004865660797804594, + 0.00872823316603899, + 0.02340121939778328, + 0.006341943982988596, + 0.0064687179401516914, + 0.00558427395299077, + 0.0050775716081261635, + 0.025516057386994362, + 0.009184123016893864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.012533470056951046 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005704939365387 + ], + "opposite_logprob_tokens": [ + -2.6950571537017822 + ], + "logprob_reward": -0.07005704939365387, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005704939365387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3954336643218994, + 0.061449650675058365, + 0.07298429310321808, + 0.025827359408140182, + 0.06167786568403244, + 0.17728972434997559, + 0.04481936991214752, + 0.12041635811328888, + 0.018958330154418945, + 0.021143479272723198 + ], + "credit_map": [ + 0.05985386669635773, + 0.11922172456979752, + 0.036666616797447205, + 0.031581323593854904, + 0.030872033908963203, + 0.02106165699660778, + 0.02380988374352455, + 0.07236651331186295, + 0.00804020557552576, + 0.00838809460401535, + 0.007872949354350567, + 0.0081171914935112, + 0.010977059602737427, + 0.0058754812926054, + 0.005858578719198704, + 0.006320086773484945, + 0.00545037817209959, + 0.0187227763235569, + 0.003458435647189617, + 0.004893430508673191, + 0.00899746734648943, + 0.00409217132255435, + 0.019219761714339256, + 0.008149867877364159, + 0.0020135161466896534, + 0.0018195227021351457, + 0.0027254021260887384, + 0.005095988977700472, + 0.0031565839890390635, + 0.002591969445347786, + 0.004234361927956343, + 0.004190011415630579, + 0.0071238260716199875, + 0.004717506933957338, + 0.007453682366758585, + 0.009601481258869171, + 0.0050554703921079636, + 0.003252116497606039, + 0.009998125955462456, + 0.014475652016699314, + 0.012459882535040379, + 0.02747258171439171, + 0.011372850276529789, + 0.015840988606214523, + 0.016702529042959213, + 0.010724716819822788, + 0.04239420220255852, + 0.0403219573199749, + 0.005868819076567888, + 0.015288655646145344, + 0.003880609292536974, + 0.0042526936158537865, + 0.003432679921388626, + 0.003942868206650019, + 0.005085412412881851, + 0.0030676277820020914, + 0.013634138740599155, + 0.042938701808452606, + 0.0056180707179009914, + 0.010352995246648788, + 0.0093939658254385, + 0.009073442779481411, + 0.02056596800684929, + 0.008839058689773083, + 0.001981058157980442, + 0.001890215091407299, + 0.0016197648365050554, + 0.002618823666125536, + 0.0037879508454352617, + 0.002496520522981882, + 0.0022701211273670197, + 0.0022938740439713, + 0.0029935569036751986, + 0.007161908783018589, + 0.001108650816604495, + 0.001172536052763462, + 0.0019600698724389076, + 0.0017860373482108116, + 0.0033822483383119106, + 0.0015784681309014559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.07005704939365387 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a surreal and photoreal confrontation within the booth of a classic American diner. On the left, an anthropomorphic burger, its head and torso formed by a smooth, sesame-studded brioche bun, leans forward aggressively. Two black olive slices act as wide, angry eyes, while a curved pickle slice opens its mouth, revealing layers of thick, grilled patty, crisp lettuce, and tomato slices. Its spindly French fry arms gesture wildly; one defiant, stiff French fry finger points directly at its companion, while another is raised angrily as a drop of melted cheddar cheese drips down its side. Opposite, an anthropomorphic pepperoni pizza slice, its pointy crust forming a head, evokes the same passion. Two glistening black olives half form its narrowed, angry eyes, while its body, covered in glistening oil and melted mozzarella cheese, leans back defiantly. It waved one breadstick arm dismissively, while the other remained poised, as if it had just decisively slapped the stained Formica tabletop. They sat on cracked cherry-red vinyl sofas, their chrome trim and glass condiment shakers gleaming in the cool, sterile glow of the restaurant's overhead fluorescent lights. Through the large windows behind them, the city's blurred nighttime neon signs cast soft, colorful highlights, bathing the tense and comically absurd scene in a dreamy, atmospheric glow.", + "seed": 9254, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_4/index_5186_seed_9254_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05143957454711199, + "vlm_logprob": -0.05143957454711199, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.05143957454711199, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02992686629295349 + ], + "opposite_logprob_tokens": [ + -3.6549267768859863 + ], + "logprob_reward": -0.02992686629295349, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02992686629295349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15471908450126648, + 0.1591024547815323, + 0.021634310483932495, + 0.04112512990832329, + 0.022651370614767075, + 0.058793678879737854, + 0.14925512671470642, + 0.32614558935165405, + 0.029666842892766, + 0.036906372755765915 + ], + "credit_map": [ + 0.023206764832139015, + 0.023959867656230927, + 0.018803196027874947, + 0.016400806605815887, + 0.02109459787607193, + 0.018807699903845787, + 0.013729448430240154, + 0.018716713413596153, + 0.0230997446924448, + 0.031026067212224007, + 0.015630008652806282, + 0.016084831207990646, + 0.02220914140343666, + 0.02178395912051201, + 0.015139135532081127, + 0.014129574410617352, + 0.0021618304308503866, + 0.003050990402698517, + 0.0023377547040581703, + 0.003425326431170106, + 0.0020586170721799135, + 0.002416329924017191, + 0.00270976684987545, + 0.003473693737760186, + 0.00668560154736042, + 0.00933221448212862, + 0.004334792960435152, + 0.005817891098558903, + 0.0036219176836311817, + 0.003965076059103012, + 0.0037863648030906916, + 0.0035812698770314455, + 0.003045555204153061, + 0.004306858405470848, + 0.002554291859269142, + 0.0027078562416136265, + 0.002795255510136485, + 0.002561883069574833, + 0.002139313379302621, + 0.002540354384109378, + 0.009330525994300842, + 0.010970169678330421, + 0.006352100986987352, + 0.009289530105888844, + 0.005650421604514122, + 0.0057802158407866955, + 0.006051804404705763, + 0.005368909332901239, + 0.026437079533934593, + 0.024294627830386162, + 0.014402920380234718, + 0.01635476015508175, + 0.01311946939677, + 0.019702136516571045, + 0.017787283286452293, + 0.017156843096017838, + 0.03793446719646454, + 0.02208784781396389, + 0.03795897215604782, + 0.014390932396054268, + 0.021105889230966568, + 0.08827678859233856, + 0.015763167291879654, + 0.08862753957509995, + 0.00358075019903481, + 0.004759518429636955, + 0.0036589254159480333, + 0.003428023075684905, + 0.004711750894784927, + 0.002856553765013814, + 0.0033591475803405046, + 0.0033121719025075436, + 0.004653960932046175, + 0.005002906545996666, + 0.004376857541501522, + 0.0037803654558956623, + 0.0037311187479645014, + 0.007014622911810875, + 0.004047099966555834, + 0.00429944321513176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.02992686629295349 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553364969789982 + ], + "opposite_logprob_tokens": [ + -4.3875532150268555 + ], + "logprob_reward": -0.012553364969789982, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553364969789982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1585071086883545, + 0.14936180412769318, + 0.18601995706558228, + 0.02932930551469326, + 0.06237237900495529, + 0.036685582250356674, + 0.09666197746992111, + 0.2129935324192047, + 0.027613557875156403, + 0.04045478254556656 + ], + "credit_map": [ + 0.015670662745833397, + 0.01657986082136631, + 0.017776964232325554, + 0.012841151095926762, + 0.026094360277056694, + 0.029945941641926765, + 0.01061511691659689, + 0.02898304909467697, + 0.014536813832819462, + 0.020404988899827003, + 0.013868473470211029, + 0.015360570512712002, + 0.016137627884745598, + 0.035195983946323395, + 0.012276574037969112, + 0.021580770611763, + 0.012700080871582031, + 0.01849118061363697, + 0.018003633245825768, + 0.025321537628769875, + 0.024060605093836784, + 0.03944971412420273, + 0.016874361783266068, + 0.031118834391236305, + 0.003529889974743128, + 0.003951427526772022, + 0.0032542452681809664, + 0.0031692383345216513, + 0.0037424229085445404, + 0.004447205923497677, + 0.0032094179186969995, + 0.004025458358228207, + 0.006721237674355507, + 0.006403602659702301, + 0.006685827858746052, + 0.004588440991938114, + 0.005871904082596302, + 0.015564714558422565, + 0.0046233125030994415, + 0.011913340538740158, + 0.006332179065793753, + 0.004554453305900097, + 0.004280806053429842, + 0.004596646409481764, + 0.003167008515447378, + 0.0049508665688335896, + 0.0036907168105244637, + 0.0051129041239619255, + 0.01758502423763275, + 0.013251987285912037, + 0.008794953115284443, + 0.009053181856870651, + 0.01251953188329935, + 0.013331729918718338, + 0.01131437998265028, + 0.010811185464262962, + 0.01405118964612484, + 0.013368409126996994, + 0.02616199105978012, + 0.00844545941799879, + 0.022070659324526787, + 0.058647602796554565, + 0.010270639322698116, + 0.059977561235427856, + 0.003181018866598606, + 0.003803083673119545, + 0.002983505604788661, + 0.002604137407615781, + 0.004198925569653511, + 0.003158551175147295, + 0.004395908210426569, + 0.0032884262036532164, + 0.00462335767224431, + 0.0060667796060442924, + 0.004595975391566753, + 0.00409004045650363, + 0.005119557026773691, + 0.006396085489541292, + 0.004275535698980093, + 0.005287448409944773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.012553364969789982 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195652678608894 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195652678608894, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195652678608894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18602140247821808, + 0.260537713766098, + 0.05991898477077484, + 0.0402037687599659, + 0.03545866161584854, + 0.08969966322183609, + 0.14383505284786224, + 0.12187889218330383, + 0.030140051618218422, + 0.03230582922697067 + ], + "credit_map": [ + 0.02320370078086853, + 0.022969160228967667, + 0.017754431813955307, + 0.016960032284259796, + 0.024672675877809525, + 0.035502202808856964, + 0.013496633619070053, + 0.031462546437978745, + 0.0319426991045475, + 0.03933359682559967, + 0.026778709143400192, + 0.0384632907807827, + 0.04082894325256348, + 0.02920018695294857, + 0.030298864468932152, + 0.02369142323732376, + 0.006860976107418537, + 0.009892803616821766, + 0.006485048681497574, + 0.006996870040893555, + 0.008049369789659977, + 0.006218353286385536, + 0.006491784937679768, + 0.008923772722482681, + 0.0061931367963552475, + 0.005568547174334526, + 0.003855000250041485, + 0.004027682356536388, + 0.0036677878815680742, + 0.007153955288231373, + 0.00442812405526638, + 0.005309530533850193, + 0.0060547697357833385, + 0.0045686569064855576, + 0.003803040599450469, + 0.0036845924332737923, + 0.004240636248141527, + 0.004170667380094528, + 0.00298164295963943, + 0.005954653024673462, + 0.013802158646285534, + 0.012264011427760124, + 0.009745670482516289, + 0.012340079993009567, + 0.00978910643607378, + 0.011502762325108051, + 0.009279194287955761, + 0.010976669378578663, + 0.025386551395058632, + 0.019251231104135513, + 0.013165237382054329, + 0.013752352446317673, + 0.01786957122385502, + 0.02136871963739395, + 0.01642715372145176, + 0.01661422848701477, + 0.007713747210800648, + 0.00722920149564743, + 0.015268472023308277, + 0.005972197744995356, + 0.011895531788468361, + 0.03296694532036781, + 0.006066497415304184, + 0.03476628661155701, + 0.003558811265975237, + 0.003859257325530052, + 0.00340645550750196, + 0.00352515559643507, + 0.0055307489819824696, + 0.002770321909338236, + 0.004118388518691063, + 0.0033709106501191854, + 0.005669257137924433, + 0.004492305219173431, + 0.0035505276173353195, + 0.0032211891375482082, + 0.003679020795971155, + 0.004676829557865858, + 0.003756594378501177, + 0.0032601025886833668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.014195652678608894 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020571095868945122 + ], + "opposite_logprob_tokens": [ + -3.895570993423462 + ], + "logprob_reward": -0.020571095868945122, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020571095868945122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11699124425649643, + 0.22840151190757751, + 0.2802205979824066, + 0.041410624980926514, + 0.0388481542468071, + 0.0716249942779541, + 0.03725755587220192, + 0.12483067810535431, + 0.03335990756750107, + 0.02705477364361286 + ], + "credit_map": [ + 0.012253720313310623, + 0.01798909343779087, + 0.011820720508694649, + 0.016877422109246254, + 0.020397789776325226, + 0.014938432723283768, + 0.011139943264424801, + 0.01157413236796856, + 0.026791514828801155, + 0.03265239670872688, + 0.020064139738678932, + 0.031194427981972694, + 0.03269360587000847, + 0.038539718836545944, + 0.02550864964723587, + 0.020957060158252716, + 0.021007897332310677, + 0.030649088323116302, + 0.02713410183787346, + 0.040078021585941315, + 0.03343703970313072, + 0.05646725371479988, + 0.028485899791121483, + 0.04296128824353218, + 0.00522993179038167, + 0.007514731492847204, + 0.004615993704646826, + 0.0061319745145738125, + 0.004224867559969425, + 0.004906489979475737, + 0.004352610092610121, + 0.0044340211898088455, + 0.005036177113652229, + 0.004768966697156429, + 0.004178146831691265, + 0.004208457190543413, + 0.0055500902235507965, + 0.005658676847815514, + 0.0041473377496004105, + 0.00530029833316803, + 0.008121825754642487, + 0.01178706530481577, + 0.0071305385790765285, + 0.007705847267061472, + 0.009989121928811073, + 0.013872155919671059, + 0.005442076828330755, + 0.007576362229883671, + 0.004463900346308947, + 0.0048922402784228325, + 0.004321045242249966, + 0.0025949941482394934, + 0.00339074176736176, + 0.008180981501936913, + 0.0027337505016475916, + 0.006679902318865061, + 0.016703695058822632, + 0.009134691208600998, + 0.012513229623436928, + 0.010530305095016956, + 0.010108495131134987, + 0.034022361040115356, + 0.01297954935580492, + 0.01883835531771183, + 0.004757128190249205, + 0.0035665936302393675, + 0.0035875358153134584, + 0.003525781910866499, + 0.0057693952694535255, + 0.003931128419935703, + 0.003367602825164795, + 0.00485474057495594, + 0.0048154545947909355, + 0.003742351196706295, + 0.0028578226920217276, + 0.0030000447295606136, + 0.002839852124452591, + 0.0042403824627399445, + 0.0025394707918167114, + 0.0030193955171853304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.020571095868945122 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17995089292526245 + ], + "opposite_logprob_tokens": [ + -1.8049509525299072 + ], + "logprob_reward": -0.17995089292526245, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17995089292526245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31825321912765503, + 0.11990486830472946, + 0.04967784881591797, + 0.08048650622367859, + 0.08061254024505615, + 0.04109205678105354, + 0.08428431302309036, + 0.16185425221920013, + 0.03023037128150463, + 0.0336039736866951 + ], + "credit_map": [ + 0.03432867303490639, + 0.04216817393898964, + 0.029450250789523125, + 0.03204386308789253, + 0.03621804341673851, + 0.06966375559568405, + 0.030471686273813248, + 0.04390880465507507, + 0.011192213743925095, + 0.014700490981340408, + 0.012115586549043655, + 0.007314731366932392, + 0.016348734498023987, + 0.0281528253108263, + 0.009834706783294678, + 0.020245596766471863, + 0.004691321402788162, + 0.00659923255443573, + 0.004632142838090658, + 0.010231101885437965, + 0.00746607780456543, + 0.004492642357945442, + 0.006601495202630758, + 0.004963838495314121, + 0.01625562272965908, + 0.014157830737531185, + 0.008004809729754925, + 0.00952747743576765, + 0.008180362172424793, + 0.009025812149047852, + 0.008340464904904366, + 0.0069941324181854725, + 0.010700550861656666, + 0.00849836878478527, + 0.008506717160344124, + 0.006798177491873503, + 0.007517550140619278, + 0.01848544552922249, + 0.005911624059081078, + 0.01419411413371563, + 0.007184495683759451, + 0.004905673675239086, + 0.004308050032705069, + 0.005507313180714846, + 0.005357360001653433, + 0.0050454228185117245, + 0.004826988559216261, + 0.0039567588828504086, + 0.01336984895169735, + 0.013727985322475433, + 0.007834265939891338, + 0.008247069083154202, + 0.009878216311335564, + 0.012458577752113342, + 0.009232916869223118, + 0.009535442106425762, + 0.030179044231772423, + 0.01792830601334572, + 0.016714023426175117, + 0.011608476750552654, + 0.02159888669848442, + 0.025861667469143867, + 0.014613303355872631, + 0.023350562900304794, + 0.004499166272580624, + 0.0038352261763066053, + 0.003611101070418954, + 0.0031166449189186096, + 0.0044339438900351524, + 0.003842149395495653, + 0.002868662355467677, + 0.004023480229079723, + 0.007265352178364992, + 0.004946852568536997, + 0.0037566216196864843, + 0.003323674201965332, + 0.003349319100379944, + 0.004403903614729643, + 0.00356286927126348, + 0.002995380898937583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.17995089292526245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a surreal and photoreal confrontation within the booth of a classic American diner. On the left, an anthropomorphic burger, its head and torso formed by a smooth, sesame-studded brioche bun, leans forward aggressively. Two black olive slices act as wide, angry eyes, while a curved pickle slice opens its mouth, revealing layers of thick, grilled patty, crisp lettuce, and tomato slices. Its spindly French fry arms gesture wildly; one defiant, stiff French fry finger points directly at its companion, while another is raised angrily as a drop of melted cheddar cheese drips down its side. Opposite, an anthropomorphic pepperoni pizza slice, its pointy crust forming a head, evokes the same passion. Two glistening black olives half form its narrowed, angry eyes, while its body, covered in glistening oil and melted mozzarella cheese, leans back defiantly. It waved one breadstick arm dismissively, while the other remained poised, as if it had just decisively slapped the stained Formica tabletop. They sat on cracked cherry-red vinyl sofas, their chrome trim and glass condiment shakers gleaming in the cool, sterile glow of the restaurant's overhead fluorescent lights. Through the large windows behind them, the city's blurred nighttime neon signs cast soft, colorful highlights, bathing the tense and comically absurd scene in a dreamy, atmospheric glow.", + "seed": 9255, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_4/index_5186_seed_9255_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.031960845831781624, + "vlm_logprob": -0.031960845831781624, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.031960845831781624, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.040334317833185196 + ], + "opposite_logprob_tokens": [ + -3.4153342247009277 + ], + "logprob_reward": -0.040334317833185196, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.040334317833185196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10002955794334412, + 0.05242011323571205, + 0.04604404419660568, + 0.10481438040733337, + 0.10386279970407486, + 0.10598237812519073, + 0.10934540629386902, + 0.09276442974805832, + 0.16383929550647736, + 0.12089759856462479 + ], + "credit_map": [ + 0.019202321767807007, + 0.00848317239433527, + 0.011057166382670403, + 0.011090340092778206, + 0.011983565054833889, + 0.013666555285453796, + 0.01422808226197958, + 0.01031835563480854, + 0.006695322226732969, + 0.004406487103551626, + 0.009048251435160637, + 0.005759803578257561, + 0.008505160920321941, + 0.008695049211382866, + 0.004364370834082365, + 0.004945666063576937, + 0.004462182056158781, + 0.003167135873809457, + 0.0061346301808953285, + 0.0035613607615232468, + 0.007866887375712395, + 0.007917046546936035, + 0.0078067355789244175, + 0.005128067918121815, + 0.008509916253387928, + 0.0032183839939534664, + 0.034113459289073944, + 0.009951774962246418, + 0.004683065693825483, + 0.005168151576071978, + 0.03162086755037308, + 0.007548760622739792, + 0.00804167240858078, + 0.005528994835913181, + 0.023528387770056725, + 0.007738510146737099, + 0.004845092073082924, + 0.011596223339438438, + 0.030150284990668297, + 0.01243363693356514, + 0.0065534599125385284, + 0.0038225678727030754, + 0.03110276162624359, + 0.00727464584633708, + 0.004475418012589216, + 0.008161148987710476, + 0.03712834417819977, + 0.007464033085852861, + 0.007068909704685211, + 0.0040855747647583485, + 0.02257566526532173, + 0.009049479849636555, + 0.006518310867249966, + 0.013840069994330406, + 0.038843248039484024, + 0.007364145014435053, + 0.007105643395334482, + 0.0030413889326155186, + 0.022465473040938377, + 0.0066577051766216755, + 0.0040135737508535385, + 0.0071803811006248, + 0.03539203852415085, + 0.0069082253612577915, + 0.009452675469219685, + 0.0035028259735554457, + 0.05818881094455719, + 0.012775957584381104, + 0.006667267065495253, + 0.006110163405537605, + 0.05804835259914398, + 0.00909324549138546, + 0.009860008955001831, + 0.006190797314047813, + 0.013938273303210735, + 0.0124258603900671, + 0.005681314039975405, + 0.020638015121221542, + 0.04339798912405968, + 0.008765331469476223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.040334317833185196 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764421047642827 + ], + "opposite_logprob_tokens": [ + -4.882644176483154 + ], + "logprob_reward": -0.00764421047642827, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764421047642827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10269923508167267, + 0.03687221184372902, + 0.08245056122541428, + 0.0856701210141182, + 0.15721824765205383, + 0.06458523869514465, + 0.03560486435890198, + 0.01750672049820423, + 0.2688002586364746, + 0.1485925167798996 + ], + "credit_map": [ + 0.015621980652213097, + 0.009923618286848068, + 0.022286390885710716, + 0.010489284060895443, + 0.009969090111553669, + 0.00664554676041007, + 0.018956487998366356, + 0.008806845173239708, + 0.004740105476230383, + 0.004867915529757738, + 0.007687384728342295, + 0.004792722407728434, + 0.0036808568984270096, + 0.0028398376889526844, + 0.004838623106479645, + 0.003424766706302762, + 0.008348934352397919, + 0.006294197402894497, + 0.013439456932246685, + 0.006937750615179539, + 0.011903260834515095, + 0.01248493604362011, + 0.014987616799771786, + 0.008054414764046669, + 0.005630986299365759, + 0.002517266431823373, + 0.026259679347276688, + 0.00721391849219799, + 0.00497118616476655, + 0.004088821355253458, + 0.02846422605216503, + 0.006524045020341873, + 0.013116247951984406, + 0.007721990346908569, + 0.05198369920253754, + 0.015817251056432724, + 0.00799107737839222, + 0.00913046021014452, + 0.04235837608575821, + 0.009099159389734268, + 0.005906842648983002, + 0.004552187863737345, + 0.01018527615815401, + 0.0057224491611123085, + 0.003748525632545352, + 0.009391485713422298, + 0.019159981980919838, + 0.005918498151004314, + 0.0025409443769603968, + 0.0011906017316505313, + 0.01030903123319149, + 0.002620756858959794, + 0.0022116340696811676, + 0.0028749906923621893, + 0.010860693641006947, + 0.002996215596795082, + 0.0026449160650372505, + 0.0011079419637098908, + 0.0018699807114899158, + 0.002162620425224304, + 0.0010379771701991558, + 0.002139892429113388, + 0.004627915099263191, + 0.001915477798320353, + 0.014628543518483639, + 0.0060186246410012245, + 0.09042438119649887, + 0.018378347158432007, + 0.012316703796386719, + 0.013645512983202934, + 0.09733008593320847, + 0.01605805568397045, + 0.017609113827347755, + 0.00953334104269743, + 0.018654724583029747, + 0.01359556708484888, + 0.008273548446595669, + 0.019483385607600212, + 0.05174718052148819, + 0.009695664048194885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.00764421047642827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066022217273712 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066022217273712, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066022217273712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1011447086930275, + 0.05065205693244934, + 0.07736538350582123, + 0.16240212321281433, + 0.23650281131267548, + 0.051899369806051254, + 0.08496855944395065, + 0.05300694704055786, + 0.1478598415851593, + 0.0341980941593647 + ], + "credit_map": [ + 0.012714500539004803, + 0.007637875620275736, + 0.023954298347234726, + 0.013421819545328617, + 0.008881902322173119, + 0.009011644870042801, + 0.015269476920366287, + 0.010253197513520718, + 0.0086180055513978, + 0.0047513325698673725, + 0.009964093565940857, + 0.005944802891463041, + 0.005144921131432056, + 0.005126898176968098, + 0.006972721312195063, + 0.004129286389797926, + 0.00886465422809124, + 0.006793160457164049, + 0.012633354403078556, + 0.006119630765169859, + 0.011379607953131199, + 0.010440702550113201, + 0.014687946997582912, + 0.006446331739425659, + 0.010144411586225033, + 0.0031098388135433197, + 0.05584600195288658, + 0.01483363937586546, + 0.006318424828350544, + 0.0067015523090958595, + 0.054435256868600845, + 0.011013020761311054, + 0.014769676141440868, + 0.005460874643176794, + 0.08310818672180176, + 0.016931531950831413, + 0.007278556004166603, + 0.009374065324664116, + 0.08606194704771042, + 0.013518008403480053, + 0.004928712733089924, + 0.004000416491180658, + 0.0075279297307133675, + 0.004672139883041382, + 0.0034988424740731716, + 0.008293463848531246, + 0.012904932722449303, + 0.006072936113923788, + 0.004477237816900015, + 0.0026686890050768852, + 0.020835643634200096, + 0.006066972855478525, + 0.004630290903151035, + 0.009806240908801556, + 0.03076680190861225, + 0.005716685671359301, + 0.003248107386752963, + 0.0016368625219911337, + 0.01671493798494339, + 0.004104187712073326, + 0.0022655650973320007, + 0.003304444719105959, + 0.017937498167157173, + 0.0037953464780002832, + 0.008054327219724655, + 0.003314423142001033, + 0.04887352138757706, + 0.009940479882061481, + 0.006402007769793272, + 0.008515841327607632, + 0.054128021001815796, + 0.00863124243915081, + 0.004210309591144323, + 0.004170791245996952, + 0.003768120426684618, + 0.0034279569517821074, + 0.0029506245627999306, + 0.005792200565338135, + 0.00615706667304039, + 0.003721026936545968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.016066022217273712 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033684730529785156 + ], + "opposite_logprob_tokens": [ + -3.408684730529785 + ], + "logprob_reward": -0.033684730529785156, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033684730529785156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1383640170097351, + 0.098378486931324, + 0.06744806468486786, + 0.055170781910419464, + 0.0989624485373497, + 0.07902755588293076, + 0.19313335418701172, + 0.07558812946081161, + 0.06763695925474167, + 0.1262901872396469 + ], + "credit_map": [ + 0.019214710220694542, + 0.010850935243070126, + 0.02337675727903843, + 0.01672467775642872, + 0.014875572174787521, + 0.01989009417593479, + 0.022677049040794373, + 0.01075421366840601, + 0.016370678320527077, + 0.008800207637250423, + 0.02027803659439087, + 0.011681837029755116, + 0.009520984254777431, + 0.010191040113568306, + 0.012940729968249798, + 0.008594982326030731, + 0.012940824031829834, + 0.006725185550749302, + 0.011882197111845016, + 0.0074171205051243305, + 0.006754141300916672, + 0.005767961032688618, + 0.01026915293186903, + 0.00569148687645793, + 0.004006919916719198, + 0.0019279936095699668, + 0.019332824274897575, + 0.0055783456191420555, + 0.0020752374548465014, + 0.002646405715495348, + 0.015677673742175102, + 0.003925379365682602, + 0.010011854581534863, + 0.005288283806294203, + 0.02866497077047825, + 0.01119210198521614, + 0.0033826122526079416, + 0.00747844111174345, + 0.021042855456471443, + 0.011901319026947021, + 0.007043009623885155, + 0.004083751700818539, + 0.021767305210232735, + 0.0058151725679636, + 0.00445905514061451, + 0.006222692783921957, + 0.02363276295363903, + 0.006003802642226219, + 0.010843966156244278, + 0.004274030216038227, + 0.06898000836372375, + 0.015987535938620567, + 0.007585585117340088, + 0.00987500324845314, + 0.06370534747838974, + 0.011881880462169647, + 0.00655569089576602, + 0.0027460847049951553, + 0.02521841786801815, + 0.007722555659711361, + 0.0038613721262663603, + 0.004226943012326956, + 0.019450858235359192, + 0.0058062030002474785, + 0.006383840925991535, + 0.002516370266675949, + 0.01884825900197029, + 0.007384086027741432, + 0.004839032888412476, + 0.0040858229622244835, + 0.01769399829208851, + 0.005885545630007982, + 0.010598267428576946, + 0.0051196557469666, + 0.01612282544374466, + 0.011495662853121758, + 0.00977602694183588, + 0.019583916291594505, + 0.04504808038473129, + 0.00854575727134943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.033684730529785156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062074948102235794 + ], + "opposite_logprob_tokens": [ + -2.812074899673462 + ], + "logprob_reward": -0.062074948102235794, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062074948102235794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21855054795742035, + 0.0502760224044323, + 0.03532038629055023, + 0.2267439365386963, + 0.07538687437772751, + 0.04548623785376549, + 0.14135029911994934, + 0.08550847321748734, + 0.05472852289676666, + 0.06664876639842987 + ], + "credit_map": [ + 0.032687801867723465, + 0.022673113271594048, + 0.03556271642446518, + 0.021706027910113335, + 0.022411227226257324, + 0.026194807142019272, + 0.03208017721772194, + 0.025234688073396683, + 0.007994488812983036, + 0.004655682481825352, + 0.006661428604274988, + 0.005387977696955204, + 0.004686085972934961, + 0.006718907970935106, + 0.006695086136460304, + 0.007476366125047207, + 0.004839913919568062, + 0.003460688516497612, + 0.0055218287743628025, + 0.004418858792632818, + 0.0032788703683763742, + 0.004525857977569103, + 0.005387312266975641, + 0.003887058002874255, + 0.0193672813475132, + 0.008027713745832443, + 0.07495955377817154, + 0.018696816638112068, + 0.008620992302894592, + 0.010254237800836563, + 0.07196004688739777, + 0.014857315458357334, + 0.01040782779455185, + 0.005283178761601448, + 0.01925957016646862, + 0.010851488448679447, + 0.004482556134462357, + 0.011713633313775063, + 0.007950709201395512, + 0.005437913816422224, + 0.0067451163195073605, + 0.004704984836280346, + 0.00715645169839263, + 0.006002191919833422, + 0.003587901359423995, + 0.005544090643525124, + 0.006331732030957937, + 0.005413769744336605, + 0.008845211006700993, + 0.003587811253964901, + 0.05031345412135124, + 0.012407014146447182, + 0.005716218613088131, + 0.0062910281121730804, + 0.04510129988193512, + 0.009088266640901566, + 0.006382760591804981, + 0.002181235235184431, + 0.03154892101883888, + 0.008216035552322865, + 0.003483181819319725, + 0.003470765659585595, + 0.025430897250771523, + 0.004794668406248093, + 0.003168378258123994, + 0.0014216918498277664, + 0.01841224916279316, + 0.005204308312386274, + 0.0030166255310177803, + 0.002819409128278494, + 0.017396191135048866, + 0.003289670217782259, + 0.011886084452271461, + 0.007001555059105158, + 0.007410621270537376, + 0.005456177983433008, + 0.00878333579748869, + 0.010171080008149147, + 0.008655977435410023, + 0.007283934857696295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.062074948102235794 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures a surreal and photoreal confrontation within the booth of a classic American diner. On the left, an anthropomorphic burger, its head and torso formed by a smooth, sesame-studded brioche bun, leans forward aggressively. Two black olive slices act as wide, angry eyes, while a curved pickle slice opens its mouth, revealing layers of thick, grilled patty, crisp lettuce, and tomato slices. Its spindly French fry arms gesture wildly; one defiant, stiff French fry finger points directly at its companion, while another is raised angrily as a drop of melted cheddar cheese drips down its side. Opposite, an anthropomorphic pepperoni pizza slice, its pointy crust forming a head, evokes the same passion. Two glistening black olives half form its narrowed, angry eyes, while its body, covered in glistening oil and melted mozzarella cheese, leans back defiantly. It waved one breadstick arm dismissively, while the other remained poised, as if it had just decisively slapped the stained Formica tabletop. They sat on cracked cherry-red vinyl sofas, their chrome trim and glass condiment shakers gleaming in the cool, sterile glow of the restaurant's overhead fluorescent lights. Through the large windows behind them, the city's blurred nighttime neon signs cast soft, colorful highlights, bathing the tense and comically absurd scene in a dreamy, atmospheric glow.", + "seed": 9256, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_4/index_5186_seed_9256_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.029307875130325556, + "vlm_logprob": -0.029307875130325556, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.029307875130325556, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031547244638204575 + ], + "opposite_logprob_tokens": [ + -3.6565473079681396 + ], + "logprob_reward": -0.031547244638204575, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.031547244638204575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2687405049800873, + 0.09887940436601639, + 0.06672955304384232, + 0.12102186679840088, + 0.05075247585773468, + 0.02508383058011532, + 0.18116770684719086, + 0.12399020791053772, + 0.03560005873441696, + 0.028034409508109093 + ], + "credit_map": [ + 0.030354997143149376, + 0.03325445204973221, + 0.023465339094400406, + 0.0263365525752306, + 0.0313061848282814, + 0.04339775815606117, + 0.025241218507289886, + 0.05538397654891014, + 0.010588164441287518, + 0.019363299012184143, + 0.012993182055652142, + 0.010856792330741882, + 0.013469584286212921, + 0.012587642297148705, + 0.0087264534085989, + 0.01029428094625473, + 0.009861228056252003, + 0.00571170449256897, + 0.007609226740896702, + 0.01522876787930727, + 0.007907266728579998, + 0.005915298592299223, + 0.008066399022936821, + 0.006429661065340042, + 0.00967004057019949, + 0.010480747558176517, + 0.014118368737399578, + 0.03420010581612587, + 0.031150007620453835, + 0.00934561900794506, + 0.007730906829237938, + 0.004326064605265856, + 0.00847978051751852, + 0.006041313987225294, + 0.005606195889413357, + 0.005729537922888994, + 0.008235704153776169, + 0.005840121302753687, + 0.006529704201966524, + 0.004290116019546986, + 0.002740636933594942, + 0.0035859595518559217, + 0.0025122258812189102, + 0.003934397362172604, + 0.0034440599847584963, + 0.003549708751961589, + 0.0025491805281490088, + 0.0027676611207425594, + 0.014866229146718979, + 0.01575564220547676, + 0.01819995976984501, + 0.045253489166498184, + 0.05192127451300621, + 0.019223792478442192, + 0.009781972505152225, + 0.006165345665067434, + 0.01124273706227541, + 0.011782360263168812, + 0.012503516860306263, + 0.028857387602329254, + 0.03179823234677315, + 0.012593173421919346, + 0.009387198835611343, + 0.005825598258525133, + 0.004099920857697725, + 0.004191596992313862, + 0.0035398914478719234, + 0.006476995535194874, + 0.00772160803899169, + 0.004025280009955168, + 0.0019462202908471227, + 0.0035985440481454134, + 0.003785205539315939, + 0.004213318228721619, + 0.0030524772591888905, + 0.004203826189041138, + 0.003984611947089434, + 0.002207148354500532, + 0.002733951900154352, + 0.003853871487081051 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.031547244638204575 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005268024746328592 + ], + "opposite_logprob_tokens": [ + -5.255268096923828 + ], + "logprob_reward": -0.005268024746328592, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005268024746328592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19365090131759644, + 0.10077635943889618, + 0.15341816842556, + 0.0720909908413887, + 0.05312427133321762, + 0.053336724638938904, + 0.1339343786239624, + 0.07500576972961426, + 0.14065901935100555, + 0.024003447964787483 + ], + "credit_map": [ + 0.02074328251183033, + 0.02862854301929474, + 0.029424553737044334, + 0.02554555982351303, + 0.023536430671811104, + 0.029707137495279312, + 0.018499797210097313, + 0.017565587535500526, + 0.009904027916491032, + 0.009480086155235767, + 0.035234484821558, + 0.016569258645176888, + 0.006215117406100035, + 0.008141662925481796, + 0.007776767946779728, + 0.007454946171492338, + 0.017313895747065544, + 0.02176014333963394, + 0.014948566444218159, + 0.025318482890725136, + 0.030887117609381676, + 0.01528295036405325, + 0.0154440151527524, + 0.012462984770536423, + 0.005585132632404566, + 0.009029249660670757, + 0.007824835367500782, + 0.017411543056368828, + 0.018967129290103912, + 0.005988452583551407, + 0.003705742536112666, + 0.0035788975656032562, + 0.006670412607491016, + 0.007042570970952511, + 0.0045511843636631966, + 0.007683304604142904, + 0.012568381614983082, + 0.006342171225696802, + 0.004454322159290314, + 0.003811919130384922, + 0.005892671644687653, + 0.010876180604100227, + 0.005162212532013655, + 0.009728411212563515, + 0.008778328076004982, + 0.005383507814258337, + 0.004057140089571476, + 0.003458266844972968, + 0.012339906767010689, + 0.012360673397779465, + 0.012587741948664188, + 0.02857821434736252, + 0.03580387681722641, + 0.014247297309339046, + 0.011343887075781822, + 0.006672779563814402, + 0.007991167716681957, + 0.0070904879830777645, + 0.007654407527297735, + 0.015535527840256691, + 0.018570423126220703, + 0.008043738082051277, + 0.005888867657631636, + 0.004231142345815897, + 0.01817098632454872, + 0.024792062118649483, + 0.013720300048589706, + 0.024972038343548775, + 0.022090652957558632, + 0.013041666708886623, + 0.012107821181416512, + 0.01176347304135561, + 0.004285149741917849, + 0.0025897168088704348, + 0.0026258211582899094, + 0.003511968068778515, + 0.003420647233724594, + 0.0026447847485542297, + 0.0025636283680796623, + 0.002361730672419071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.005268024746328592 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538060545921326 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012538060545921326, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538060545921326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14022225141525269, + 0.046271972358226776, + 0.13026060163974762, + 0.04536673054099083, + 0.09208700805902481, + 0.0289740152657032, + 0.23384948074817657, + 0.0942574143409729, + 0.10837621241807938, + 0.08033429086208344 + ], + "credit_map": [ + 0.01727217808365822, + 0.015062328428030014, + 0.03606081381440163, + 0.017018912360072136, + 0.012029705569148064, + 0.01804850809276104, + 0.009977730922400951, + 0.014752077870070934, + 0.00476498156785965, + 0.009571935050189495, + 0.007425841875374317, + 0.005880598444491625, + 0.004065738059580326, + 0.005321827717125416, + 0.0034482988994568586, + 0.005792756099253893, + 0.015932422131299973, + 0.01495390385389328, + 0.01351622398942709, + 0.018064158037304878, + 0.025153733789920807, + 0.014382760040462017, + 0.0136210136115551, + 0.014636389911174774, + 0.005347782745957375, + 0.008535265922546387, + 0.005349753890186548, + 0.007115071173757315, + 0.006167906802147627, + 0.004899292252957821, + 0.00432941410690546, + 0.003622249001637101, + 0.011288817971944809, + 0.010439731180667877, + 0.008311492390930653, + 0.016318073496222496, + 0.025008555501699448, + 0.010188575834035873, + 0.005591733381152153, + 0.004940037615597248, + 0.0027114131953567266, + 0.0047982013784348965, + 0.00307550304569304, + 0.004790880251675844, + 0.003687149379402399, + 0.0038019055500626564, + 0.002862055553123355, + 0.003246909473091364, + 0.01857280544936657, + 0.020445752888917923, + 0.02312195859849453, + 0.0581074096262455, + 0.066651351749897, + 0.02494775876402855, + 0.014110678806900978, + 0.007891780696809292, + 0.011009516194462776, + 0.00841109175235033, + 0.009774511680006981, + 0.020134449005126953, + 0.021978631615638733, + 0.009660569950938225, + 0.007800267077982426, + 0.005488384515047073, + 0.014360636472702026, + 0.018774772062897682, + 0.009365742094814777, + 0.017903681844472885, + 0.020156828686594963, + 0.010877094231545925, + 0.009081062860786915, + 0.007856396958231926, + 0.010005243122577667, + 0.01045564841479063, + 0.006978777702897787, + 0.011153348721563816, + 0.0177284125238657, + 0.008873729035258293, + 0.006983396131545305, + 0.008155740797519684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.012538060545921326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176885321736336 + ], + "opposite_logprob_tokens": [ + -4.018177032470703 + ], + "logprob_reward": -0.018176885321736336, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018176885321736336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1582096964120865, + 0.07595372200012207, + 0.10851944237947464, + 0.15225832164287567, + 0.04059696942567825, + 0.11466524749994278, + 0.16499386727809906, + 0.07198718935251236, + 0.04938539117574692, + 0.06343014538288116 + ], + "credit_map": [ + 0.020161278545856476, + 0.022232213988900185, + 0.021648475900292397, + 0.01914437860250473, + 0.015947315841913223, + 0.020354190841317177, + 0.020596927031874657, + 0.01812491938471794, + 0.006593722850084305, + 0.007155483588576317, + 0.02169596590101719, + 0.010513798333704472, + 0.005874880123883486, + 0.008345210924744606, + 0.00627845898270607, + 0.00949619710445404, + 0.013433498330414295, + 0.009492522105574608, + 0.012581137008965015, + 0.01684504747390747, + 0.02306273952126503, + 0.011028656736016273, + 0.011190005578100681, + 0.010885831899940968, + 0.011491311714053154, + 0.010595961473882198, + 0.02005784586071968, + 0.042417529970407486, + 0.03677268326282501, + 0.011604221537709236, + 0.011463322676718235, + 0.007855439558625221, + 0.005592902656644583, + 0.006576331797987223, + 0.00453584548085928, + 0.005961851682513952, + 0.00483915489166975, + 0.006235170643776655, + 0.0034382485318928957, + 0.0034174639731645584, + 0.010294502601027489, + 0.010729698464274406, + 0.013845828361809254, + 0.03264114260673523, + 0.025214916095137596, + 0.009266147390007973, + 0.007025094702839851, + 0.00564791914075613, + 0.01363279391080141, + 0.014003419317305088, + 0.017140375450253487, + 0.04181591421365738, + 0.04689672961831093, + 0.01706593856215477, + 0.009078964591026306, + 0.005359726957976818, + 0.006823770701885223, + 0.007493417244404554, + 0.0070976512506604195, + 0.01473438274115324, + 0.019923076033592224, + 0.00722693745046854, + 0.004864833317697048, + 0.0038231241051107645, + 0.004773849621415138, + 0.006035572849214077, + 0.0047588050365448, + 0.009476026520133018, + 0.01243309210985899, + 0.005903984420001507, + 0.002955635776743293, + 0.0030484232120215893, + 0.007572165224701166, + 0.007734229788184166, + 0.005359227769076824, + 0.00863227341324091, + 0.017827749252319336, + 0.00759150180965662, + 0.0036630204413086176, + 0.005049984902143478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.018176885321736336 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside the booth of a classic American diner?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there an anthropomorphic burger on the left with a sesame-studded brioche bun head and torso?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an anthropomorphic pepperoni pizza slice on the right with a pointy crust head and mozzarella cheese body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the characters sitting on cracked cherry-red vinyl sofas with chrome trim?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting cool and sterile from overhead fluorescent lights with blurred nighttime neon signs visible through the windows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900916039943695 + ], + "opposite_logprob_tokens": [ + -2.5790090560913086 + ], + "logprob_reward": -0.07900916039943695, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900916039943695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3026721179485321, + 0.1388102024793625, + 0.13033445179462433, + 0.07145777344703674, + 0.07060850411653519, + 0.08303763717412949, + 0.11122385412454605, + 0.046637389808893204, + 0.020043227821588516, + 0.025174874812364578 + ], + "credit_map": [ + 0.03946692869067192, + 0.041298575699329376, + 0.03905264660716057, + 0.03935028612613678, + 0.030629940330982208, + 0.04079611971974373, + 0.024513311684131622, + 0.047564297914505005, + 0.018085816875100136, + 0.029050691053271294, + 0.03040769323706627, + 0.018217667937278748, + 0.010850880295038223, + 0.014308936893939972, + 0.009003001265227795, + 0.008885504677891731, + 0.015278719365596771, + 0.015433392487466335, + 0.013048720546066761, + 0.018941214308142662, + 0.027488302439451218, + 0.013778896071016788, + 0.014623385854065418, + 0.011741818860173225, + 0.0059903585352003574, + 0.006736088544130325, + 0.00877404399216175, + 0.01915716379880905, + 0.016174834221601486, + 0.006132221780717373, + 0.005253036506474018, + 0.003240015124902129, + 0.014021129347383976, + 0.009116425178945065, + 0.008049503900110722, + 0.007269673980772495, + 0.00907849706709385, + 0.010083462111651897, + 0.007327796891331673, + 0.005662010051310062, + 0.0063779898919165134, + 0.009259999729692936, + 0.008468987420201302, + 0.020371828228235245, + 0.02163430117070675, + 0.007153940852731466, + 0.00524976197630167, + 0.004520824644714594, + 0.008664215914905071, + 0.009345735423266888, + 0.011721886694431305, + 0.02880602516233921, + 0.030021583661437035, + 0.011777741834521294, + 0.00686857383698225, + 0.0040180878713727, + 0.005855091847479343, + 0.003783400636166334, + 0.005149138160049915, + 0.01040208712220192, + 0.009072874672710896, + 0.00498934043571353, + 0.004530190024524927, + 0.0028552671428769827, + 0.003527670865878463, + 0.002319988561794162, + 0.0022434901911765337, + 0.0028826415073126554, + 0.002474459121003747, + 0.0022466229274868965, + 0.0019584172405302525, + 0.0023899369407445192, + 0.0036514957901090384, + 0.0029100419487804174, + 0.002520187757909298, + 0.0035137070808559656, + 0.0043348949402570724, + 0.002678194548934698, + 0.0021759511437267065, + 0.0033903999719768763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07900916039943695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_5/index_tensor([4265, 4265])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_5/index_tensor([4265, 4265])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..46da54996e61d34bdc9a44badb963864375cbd9e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_5/index_tensor([4265, 4265])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "violently jarring low-angle close-up plunges the viewer into a nightmarish biomechanical landscape, a fusion of bone and metal, framed by a structure resembling a vast ribcage interwoven with dark, weeping pipes. At the center of the frame, a large, ribbed orifice, its texture a grotesque amalgamation of porous bone and pitted, corroded steel, pulsing and contracting organically. From this opening, a viscous, gray, cleansing clay erupts, not in a fluid stream but in a thick, undulating, relentless surge that seems to push against the air. As the viscous substance is expelled spasmodically, it churns and bubbles into a grotesque, living foam, its bubbles swelling to unnatural sizes before bursting with dampness, all the while catching the dim, directional light with an oily, iridescent sheen. The entire scene is rendered in the dark, skeletal aesthetic of H.R. Giger, defined by a stark, monochromatic palette of charcoal and cold metal. A single, harsh light source from above carves out shimmering highlights and deep, cavernous shadows, creating an oppressive and claustrophobic atmosphere of pure visceral terror. The camera continues to shake violently, as if shaken by the vibrations of this horrific birth.", + "seed": 16040, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_5/index_4265_seed_16040_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3470586942974478, + "vlm_logprob": -0.3470586942974478, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.3470586942974478, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007463429588824511 + ], + "opposite_logprob_tokens": [ + -5.257463455200195 + ], + "logprob_reward": -0.007463429588824511, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.007463429588824511, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14946626126766205, + 0.08112069219350815, + 0.09464242309331894, + 0.13410817086696625, + 0.0733889788389206, + 0.05952109023928642, + 0.08270801603794098, + 0.19298233091831207, + 0.06714872270822525, + 0.0649133026599884 + ], + "credit_map": [ + 0.014588410034775734, + 0.023603908717632294, + 0.021311454474925995, + 0.022488635033369064, + 0.007589609362185001, + 0.021836264058947563, + 0.018450824543833733, + 0.019597161561250687, + 0.017281219363212585, + 0.011545684188604355, + 0.0100886020809412, + 0.011393657885491848, + 0.0043978504836559296, + 0.009627382270991802, + 0.0075500765815377235, + 0.009236208163201809, + 0.00713321240618825, + 0.005585019942373037, + 0.01064355205744505, + 0.009379931725561619, + 0.004757286049425602, + 0.005003537517040968, + 0.029703471809625626, + 0.022436410188674927, + 0.012887591496109962, + 0.011717052198946476, + 0.014594707638025284, + 0.02608896978199482, + 0.015999609604477882, + 0.010771443136036396, + 0.01565392315387726, + 0.02639487013220787, + 0.013185624964535236, + 0.007185103837400675, + 0.010252430103719234, + 0.007807811722159386, + 0.006095128133893013, + 0.007463612128049135, + 0.006558405701071024, + 0.014840860851109028, + 0.006710918620228767, + 0.0061539700254797935, + 0.01060394011437893, + 0.009916525334119797, + 0.007066552061587572, + 0.0074910735711455345, + 0.005551914218813181, + 0.006026197224855423, + 0.010127123445272446, + 0.009244764223694801, + 0.008451520465314388, + 0.010941850021481514, + 0.012176957912743092, + 0.010125413537025452, + 0.011763381771743298, + 0.009877003729343414, + 0.02024249918758869, + 0.018940074369311333, + 0.020303910598158836, + 0.02708287164568901, + 0.0150469234213233, + 0.022245537489652634, + 0.03991146758198738, + 0.02920905500650406, + 0.006478163879364729, + 0.007585135288536549, + 0.006304156966507435, + 0.008704008534550667, + 0.008836304768919945, + 0.010154460556805134, + 0.010849369689822197, + 0.008237123489379883, + 0.0059280674904584885, + 0.006443528458476067, + 0.008901932276785374, + 0.007047491148114204, + 0.0066968477331101894, + 0.0066914367489516735, + 0.013453834690153599, + 0.009750164113938808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.007463429588824511 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032089666929095984 + ], + "opposite_logprob_tokens": [ + -5.753209114074707 + ], + "logprob_reward": -0.0032089666929095984, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032089666929095984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05631076544523239, + 0.05180913954973221, + 0.14411751925945282, + 0.12915408611297607, + 0.06422563642263412, + 0.06621304154396057, + 0.061464861035346985, + 0.1990414708852768, + 0.18464669585227966, + 0.043016817420721054 + ], + "credit_map": [ + 0.006903140340000391, + 0.007232477888464928, + 0.006395779084414244, + 0.006994901690632105, + 0.0036408586893230677, + 0.008951973170042038, + 0.00933660939335823, + 0.006855021230876446, + 0.008005132898688316, + 0.0070855990052223206, + 0.0066765788942575455, + 0.008295441046357155, + 0.0032983000855892897, + 0.0073755099438130856, + 0.005484289955347776, + 0.005588284227997065, + 0.014718445017933846, + 0.009017536416649818, + 0.05034509673714638, + 0.0170740969479084, + 0.008150231093168259, + 0.009429431520402431, + 0.016802538186311722, + 0.018580134958028793, + 0.012332865968346596, + 0.012182796373963356, + 0.01029160711914301, + 0.0293966606259346, + 0.012712882831692696, + 0.009232006967067719, + 0.01316582877188921, + 0.02983943372964859, + 0.01171462144702673, + 0.006418358068913221, + 0.010593486949801445, + 0.007690642960369587, + 0.005660192109644413, + 0.004466488491743803, + 0.006225492339581251, + 0.01145634613931179, + 0.006683661136776209, + 0.006529123988002539, + 0.007867339998483658, + 0.01067359559237957, + 0.006885799113661051, + 0.005424462258815765, + 0.01548420824110508, + 0.0066648502834141254, + 0.006981959100812674, + 0.00803158339112997, + 0.00649776216596365, + 0.008365337736904621, + 0.008815785869956017, + 0.007305082865059376, + 0.009095185436308384, + 0.0063721598125994205, + 0.022867413237690926, + 0.02390597015619278, + 0.019836656749248505, + 0.02550978772342205, + 0.015886720269918442, + 0.021248016506433487, + 0.03978397324681282, + 0.030002925544977188, + 0.018848007544875145, + 0.023055385798215866, + 0.014265944249927998, + 0.0217914842069149, + 0.03168340027332306, + 0.020791390910744667, + 0.03456760197877884, + 0.01964346133172512, + 0.0041890437714755535, + 0.00449181767180562, + 0.0057287318632006645, + 0.005244323052465916, + 0.007250093389302492, + 0.005210508592426777, + 0.005472605582326651, + 0.005429691169410944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 2 + ], + "selected_reward": -0.0032089666929095984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294086277484894 + ], + "opposite_logprob_tokens": [ + -2.237940788269043 + ], + "logprob_reward": -0.11294086277484894, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294086277484894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08518075942993164, + 0.05657292157411575, + 0.09667225927114487, + 0.0627325028181076, + 0.16728608310222626, + 0.10739082098007202, + 0.03796334192156792, + 0.29870328307151794, + 0.03196243941783905, + 0.055535633116960526 + ], + "credit_map": [ + 0.008359290659427643, + 0.0140959732234478, + 0.011350495740771294, + 0.015282715670764446, + 0.007604396436363459, + 0.0082537941634655, + 0.010029013268649578, + 0.010205075144767761, + 0.007853022776544094, + 0.00927148386836052, + 0.012251023203134537, + 0.006972837261855602, + 0.0030118420254439116, + 0.006303120404481888, + 0.005585125181823969, + 0.005324465688318014, + 0.009402960538864136, + 0.005205859895795584, + 0.032037198543548584, + 0.010639924556016922, + 0.004185388796031475, + 0.005311664193868637, + 0.007326069753617048, + 0.022563185542821884, + 0.007020925171673298, + 0.006284319795668125, + 0.005465127993375063, + 0.011459075845777988, + 0.0070786974392831326, + 0.004871748853474855, + 0.009030980989336967, + 0.01152162067592144, + 0.04932278022170067, + 0.011577002704143524, + 0.01757768914103508, + 0.009102713316679, + 0.006408557295799255, + 0.013534782454371452, + 0.009347331710159779, + 0.05041522532701492, + 0.01326381042599678, + 0.00999106653034687, + 0.0162049587816, + 0.017668642103672028, + 0.01391832996159792, + 0.011224459856748581, + 0.013791943900287151, + 0.011327614076435566, + 0.003677452215924859, + 0.004102046601474285, + 0.0041465433314442635, + 0.00575563358142972, + 0.005168423987925053, + 0.004637261852622032, + 0.006870672572404146, + 0.003605308011174202, + 0.033746957778930664, + 0.036852654069662094, + 0.029210111126303673, + 0.038462840020656586, + 0.02274148352444172, + 0.03353029489517212, + 0.059590116143226624, + 0.04456884413957596, + 0.0029882334638386965, + 0.005034768953919411, + 0.003838588250800967, + 0.003589496249333024, + 0.004747250583022833, + 0.004444357473403215, + 0.0038730097003281116, + 0.0034467345103621483, + 0.003494821721687913, + 0.005280134733766317, + 0.005730216391384602, + 0.006941637489944696, + 0.012170053087174892, + 0.006253933534026146, + 0.0076822093687951565, + 0.007982629351317883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 5 + ], + "selected_reward": -0.11294086277484894 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011097612790763378 + ], + "opposite_logprob_tokens": [ + -4.511097431182861 + ], + "logprob_reward": -0.011097612790763378, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011097612790763378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13573326170444489, + 0.06984420120716095, + 0.13897107541561127, + 0.07479438930749893, + 0.06975947320461273, + 0.06731784343719482, + 0.03406202420592308, + 0.07603447884321213, + 0.09528866410255432, + 0.23819449543952942 + ], + "credit_map": [ + 0.012688311748206615, + 0.0263991579413414, + 0.020031768828630447, + 0.017550507560372353, + 0.009240810759365559, + 0.01720496453344822, + 0.017159849405288696, + 0.015457912348210812, + 0.01913713663816452, + 0.008038378320634365, + 0.01169799268245697, + 0.007609123829752207, + 0.004352398216724396, + 0.0071836900897324085, + 0.005009686574339867, + 0.006815796718001366, + 0.011342709884047508, + 0.009242120198905468, + 0.020148800686001778, + 0.019486647099256516, + 0.0072941770777106285, + 0.011632682755589485, + 0.037337906658649445, + 0.022486049681901932, + 0.007000854704529047, + 0.0068110814318060875, + 0.006169474218040705, + 0.01693842187523842, + 0.007208174094557762, + 0.00608630059286952, + 0.008923348970711231, + 0.015656741335988045, + 0.01662607491016388, + 0.007088398095220327, + 0.008731639012694359, + 0.007254105061292648, + 0.004749909043312073, + 0.007419159635901451, + 0.004820465575903654, + 0.01306973397731781, + 0.005923092365264893, + 0.005436677951365709, + 0.005951925180852413, + 0.014210020191967487, + 0.00551144452765584, + 0.004616072867065668, + 0.01921379566192627, + 0.006454822141677141, + 0.0036458265967667103, + 0.0030370966996997595, + 0.003683207556605339, + 0.006017997860908508, + 0.003842027857899666, + 0.004049085080623627, + 0.00507240230217576, + 0.004714388400316238, + 0.01099823322147131, + 0.008850456215441227, + 0.007138626184314489, + 0.007036646828055382, + 0.008307748474180698, + 0.009242200292646885, + 0.010124264284968376, + 0.014336316846311092, + 0.008858632296323776, + 0.012195006012916565, + 0.009175523184239864, + 0.009889676235616207, + 0.018092365935444832, + 0.009837617166340351, + 0.015978403389453888, + 0.011261454783380032, + 0.01423223689198494, + 0.014863203279674053, + 0.012069066986441612, + 0.02291656844317913, + 0.019917836412787437, + 0.023473534733057022, + 0.06470353156328201, + 0.06601855903863907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.011097612790763378 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6005825996398926 + ], + "opposite_logprob_tokens": [ + -0.22558264434337616 + ], + "logprob_reward": -1.6005825996398926, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005825996398926, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.196528822183609, + 0.13146400451660156, + 0.13566173613071442, + 0.038658831268548965, + 0.13144053518772125, + 0.10171060264110565, + 0.03269727900624275, + 0.0729915201663971, + 0.04079798236489296, + 0.1180487871170044 + ], + "credit_map": [ + 0.019196338951587677, + 0.037837423384189606, + 0.032450079917907715, + 0.02674228698015213, + 0.016604773700237274, + 0.025336194783449173, + 0.022615164518356323, + 0.015746526420116425, + 0.039422620087862015, + 0.020529359579086304, + 0.01728002540767193, + 0.011860040947794914, + 0.005858117248862982, + 0.015123177319765091, + 0.009679455310106277, + 0.01171119138598442, + 0.013981439173221588, + 0.007366652600467205, + 0.029850997030735016, + 0.011880463920533657, + 0.00686605554074049, + 0.008380258455872536, + 0.02153368666768074, + 0.03580217435956001, + 0.004391141701489687, + 0.005730135831981897, + 0.005202827509492636, + 0.004985105246305466, + 0.004749000072479248, + 0.0035739438608288765, + 0.006047679111361504, + 0.00397899467498064, + 0.039484504610300064, + 0.007498564664274454, + 0.014796348288655281, + 0.005977275315672159, + 0.0051436349749565125, + 0.009907110594213009, + 0.007724737748503685, + 0.040908344089984894, + 0.009955665096640587, + 0.008503136225044727, + 0.015429545193910599, + 0.026822511106729507, + 0.010102061554789543, + 0.00870190653949976, + 0.013520617969334126, + 0.008675146847963333, + 0.0039052620995789766, + 0.0026570127811282873, + 0.0042539131827652454, + 0.005771805066615343, + 0.0029001843649894, + 0.004667902830988169, + 0.005084598902612925, + 0.0034565995447337627, + 0.008651944808661938, + 0.008679651655256748, + 0.0071787782944738865, + 0.008868665434420109, + 0.00834475178271532, + 0.007886970415711403, + 0.01305659580975771, + 0.010324155911803246, + 0.004828735254704952, + 0.005441095679998398, + 0.004039331339299679, + 0.004397306591272354, + 0.005479767918586731, + 0.00610372656956315, + 0.0054679252207279205, + 0.0050400905311107635, + 0.0074667660519480705, + 0.006688984576612711, + 0.005416542757302523, + 0.009301574900746346, + 0.00886170007288456, + 0.010735899209976196, + 0.0335526317358017, + 0.03602467477321625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -1.6005825996398926 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "violently jarring low-angle close-up plunges the viewer into a nightmarish biomechanical landscape, a fusion of bone and metal, framed by a structure resembling a vast ribcage interwoven with dark, weeping pipes. At the center of the frame, a large, ribbed orifice, its texture a grotesque amalgamation of porous bone and pitted, corroded steel, pulsing and contracting organically. From this opening, a viscous, gray, cleansing clay erupts, not in a fluid stream but in a thick, undulating, relentless surge that seems to push against the air. As the viscous substance is expelled spasmodically, it churns and bubbles into a grotesque, living foam, its bubbles swelling to unnatural sizes before bursting with dampness, all the while catching the dim, directional light with an oily, iridescent sheen. The entire scene is rendered in the dark, skeletal aesthetic of H.R. Giger, defined by a stark, monochromatic palette of charcoal and cold metal. A single, harsh light source from above carves out shimmering highlights and deep, cavernous shadows, creating an oppressive and claustrophobic atmosphere of pure visceral terror. The camera continues to shake violently, as if shaken by the vibrations of this horrific birth.", + "seed": 16041, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_5/index_4265_seed_16041_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.29871553180273624, + "vlm_logprob": -0.29871553180273624, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.29871553180273624, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008205866441130638 + ], + "opposite_logprob_tokens": [ + -5.508205890655518 + ], + "logprob_reward": -0.008205866441130638, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.008205866441130638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12990757822990417, + 0.09541913121938705, + 0.15279465913772583, + 0.053931742906570435, + 0.041703708469867706, + 0.2013503909111023, + 0.04122765734791756, + 0.0841117799282074, + 0.12984591722488403, + 0.06970737129449844 + ], + "credit_map": [ + 0.01206136867403984, + 0.01219372358173132, + 0.01598036289215088, + 0.01683678664267063, + 0.014583105221390724, + 0.015784716233611107, + 0.02181066945195198, + 0.020656848326325417, + 0.008416017517447472, + 0.009966013953089714, + 0.013653279282152653, + 0.009698348119854927, + 0.014532562345266342, + 0.010787596926093102, + 0.016388623043894768, + 0.0119766965508461, + 0.008031690493226051, + 0.0132587980479002, + 0.011809377931058407, + 0.012342533096671104, + 0.026968248188495636, + 0.02390015684068203, + 0.026542233303189278, + 0.029941635206341743, + 0.004305747803300619, + 0.004894157871603966, + 0.005520825739949942, + 0.006364261265844107, + 0.006473673973232508, + 0.010656066238880157, + 0.007732123136520386, + 0.007984887808561325, + 0.003551905043423176, + 0.004944326356053352, + 0.0050124055705964565, + 0.004765505902469158, + 0.0054434300400316715, + 0.006335247308015823, + 0.006870181765407324, + 0.00478070555254817, + 0.008222706615924835, + 0.009160640649497509, + 0.0035846165847033262, + 0.006193240638822317, + 0.07700840383768082, + 0.011745060794055462, + 0.007480315864086151, + 0.07795542478561401, + 0.002724595833569765, + 0.004911758936941624, + 0.003385750576853752, + 0.0069126226007938385, + 0.004720914643257856, + 0.004867016337811947, + 0.007910476066172123, + 0.005794523749500513, + 0.006298687309026718, + 0.008366255089640617, + 0.005163844209164381, + 0.008921027183532715, + 0.01336508896201849, + 0.015826169401407242, + 0.013655134476721287, + 0.012515566311776638, + 0.012465239502489567, + 0.013672541826963425, + 0.00806533545255661, + 0.01663890853524208, + 0.013680431060492992, + 0.02506757155060768, + 0.015855463221669197, + 0.02440042793750763, + 0.00955838244408369, + 0.008176968432962894, + 0.010645912028849125, + 0.007597767747938633, + 0.009183285757899284, + 0.009168987162411213, + 0.008115392178297043, + 0.0072606704197824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.008205866441130638 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013525871327146888 + ], + "opposite_logprob_tokens": [ + -6.626352787017822 + ], + "logprob_reward": -0.0013525871327146888, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013525871327146888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04501378536224365, + 0.04355486482381821, + 0.06915891915559769, + 0.017986591905355453, + 0.03611274063587189, + 0.4955558180809021, + 0.04210870712995529, + 0.07023189216852188, + 0.15531811118125916, + 0.02495848946273327 + ], + "credit_map": [ + 0.0045149922370910645, + 0.004332405515015125, + 0.004763065837323666, + 0.005515086930245161, + 0.005715093109756708, + 0.006128325127065182, + 0.007386769168078899, + 0.006658048834651709, + 0.003221836406737566, + 0.003385493764653802, + 0.004315262194722891, + 0.00786452554166317, + 0.00544879911467433, + 0.006143039092421532, + 0.007091471925377846, + 0.0060844384133815765, + 0.005091372411698103, + 0.005611954256892204, + 0.005588814616203308, + 0.005585615988820791, + 0.013969967141747475, + 0.010620960965752602, + 0.007812405936419964, + 0.014877833425998688, + 0.0015065338229760528, + 0.0023608605843037367, + 0.0015661715297028422, + 0.002664023544639349, + 0.002050290582701564, + 0.0026103605050593615, + 0.0035597123205661774, + 0.001668638433329761, + 0.0031714013312011957, + 0.0058738854713737965, + 0.0038035069592297077, + 0.0031846053898334503, + 0.0055526564829051495, + 0.004745127633213997, + 0.005966010969132185, + 0.0038155491929501295, + 0.02210296504199505, + 0.02401321567595005, + 0.006801859010010958, + 0.012797784060239792, + 0.19429147243499756, + 0.02902635559439659, + 0.012499980628490448, + 0.19402222335338593, + 0.005359260365366936, + 0.003194976830855012, + 0.003615953493863344, + 0.006307836156338453, + 0.00429521594196558, + 0.006700465921312571, + 0.006890763994306326, + 0.005744236055761576, + 0.004156673792749643, + 0.010264899581670761, + 0.0037535985466092825, + 0.004973918199539185, + 0.012168070301413536, + 0.012756183743476868, + 0.004906355403363705, + 0.01725219376385212, + 0.013888636603951454, + 0.011164836585521698, + 0.010244522243738174, + 0.019638782367110252, + 0.016168849542737007, + 0.032137755304574966, + 0.02309614233672619, + 0.028978584334254265, + 0.001917838817462325, + 0.002611896488815546, + 0.0038447976112365723, + 0.003667278680950403, + 0.003265919629484415, + 0.0036787025164812803, + 0.003687777556478977, + 0.002284277230501175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.0013525871327146888 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003404200077057 + ], + "opposite_logprob_tokens": [ + -2.6950340270996094 + ], + "logprob_reward": -0.07003404200077057, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003404200077057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07776255905628204, + 0.06004592403769493, + 0.03829130530357361, + 0.047489866614341736, + 0.04046491160988808, + 0.26768699288368225, + 0.04390003904700279, + 0.04904462769627571, + 0.340852290391922, + 0.03446139395236969 + ], + "credit_map": [ + 0.008387620560824871, + 0.006378652062267065, + 0.01133812591433525, + 0.00942984502762556, + 0.009834297001361847, + 0.008971418254077435, + 0.013721131719648838, + 0.009701470844447613, + 0.005559483543038368, + 0.004514199681580067, + 0.007896939292550087, + 0.007571302819997072, + 0.008549398742616177, + 0.007953655906021595, + 0.009423241019248962, + 0.008577704429626465, + 0.0038441424258053303, + 0.0032808964606374502, + 0.004309966694563627, + 0.004360734950751066, + 0.005018511787056923, + 0.006281576585024595, + 0.006586209870874882, + 0.004609265364706516, + 0.0037978182081133127, + 0.004796796012669802, + 0.004355462267994881, + 0.006247499957680702, + 0.005097591318190098, + 0.007595002185553312, + 0.009463584050536156, + 0.00613611564040184, + 0.0039946213364601135, + 0.004432907328009605, + 0.0047162240371108055, + 0.004950887057930231, + 0.005251784343272448, + 0.005370171740651131, + 0.007201256230473518, + 0.004547064192593098, + 0.011268476024270058, + 0.011818582192063332, + 0.006501154508441687, + 0.008175577037036419, + 0.098368339240551, + 0.018811365589499474, + 0.012986166402697563, + 0.09975734353065491, + 0.004286625888198614, + 0.003962470218539238, + 0.003641914576292038, + 0.007707749027758837, + 0.005298440344631672, + 0.005707727279514074, + 0.0071594989858567715, + 0.006135610863566399, + 0.0019840782042592764, + 0.003353892592713237, + 0.003347201505675912, + 0.004848882090300322, + 0.010034250095486641, + 0.005454590544104576, + 0.008597871288657188, + 0.011423859745264053, + 0.028780551627278328, + 0.023010998964309692, + 0.026519939303398132, + 0.04585232585668564, + 0.03822942078113556, + 0.05823729932308197, + 0.0701199620962143, + 0.05010181665420532, + 0.0026643057353794575, + 0.0035041552037000656, + 0.004030084237456322, + 0.005553361028432846, + 0.004630775656551123, + 0.004031740594655275, + 0.007227487396448851, + 0.0028194838669151068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.07003404200077057 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007644683588296175 + ], + "opposite_logprob_tokens": [ + -4.8826446533203125 + ], + "logprob_reward": -0.007644683588296175, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007644683588296175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10935816168785095, + 0.09834414720535278, + 0.04133160784840584, + 0.019714312627911568, + 0.026773996651172638, + 0.5152881145477295, + 0.020032783970236778, + 0.031422678381204605, + 0.11682070046663284, + 0.02091347798705101 + ], + "credit_map": [ + 0.010259483009576797, + 0.009878281503915787, + 0.011568035930395126, + 0.016979418694972992, + 0.014323947951197624, + 0.014643091708421707, + 0.01619415543973446, + 0.015511752106249332, + 0.007501023355871439, + 0.007605025544762611, + 0.010669934563338757, + 0.010949141345918179, + 0.016327304765582085, + 0.013705221936106682, + 0.01833798736333847, + 0.013248513452708721, + 0.004282863344997168, + 0.004463698249310255, + 0.003474602708593011, + 0.00408210139721632, + 0.005791214760392904, + 0.006112468428909779, + 0.0073737334460020065, + 0.005750925745815039, + 0.0018328474834561348, + 0.0017827433766797185, + 0.0020256165880709887, + 0.0020932031329721212, + 0.003057375317439437, + 0.0038540458772331476, + 0.002524402691051364, + 0.002544076880440116, + 0.0036947252228856087, + 0.0037403011228889227, + 0.0024570603854954243, + 0.002634134143590927, + 0.0030954477842897177, + 0.0038694264367222786, + 0.004424343816936016, + 0.0028585591353476048, + 0.0215242188423872, + 0.023793336004018784, + 0.009314832277595997, + 0.01420674379914999, + 0.19989384710788727, + 0.029467513784766197, + 0.017311304807662964, + 0.19977635145187378, + 0.0018293347675353289, + 0.0022237307857722044, + 0.00186125875916332, + 0.003137561958283186, + 0.0023864423856139183, + 0.0029639392159879208, + 0.003139704931527376, + 0.00249081221409142, + 0.0019033728167414665, + 0.0028944106306880713, + 0.001932024140842259, + 0.00358484941534698, + 0.006624559406191111, + 0.004555616527795792, + 0.004331623204052448, + 0.0055962223559618, + 0.008753754198551178, + 0.007715875748544931, + 0.007109601050615311, + 0.01436800230294466, + 0.012821577489376068, + 0.022787896916270256, + 0.0225980244576931, + 0.02066596783697605, + 0.0014204352628439665, + 0.003127369098365307, + 0.003079914255067706, + 0.003212316194549203, + 0.0026494120247662067, + 0.0035273663233965635, + 0.0020111436024308205, + 0.0018855200614780188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.007644683588296175 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.406340479850769 + ], + "opposite_logprob_tokens": [ + -0.28134050965309143 + ], + "logprob_reward": -1.406340479850769, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.406340479850769, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07254823297262192, + 0.06083819642663002, + 0.06080397963523865, + 0.022280747070908546, + 0.07008671760559082, + 0.3670318126678467, + 0.02107931114733219, + 0.04785585775971413, + 0.24804364144802094, + 0.02943151816725731 + ], + "credit_map": [ + 0.008194565773010254, + 0.007584915962070227, + 0.006541131064295769, + 0.008876405656337738, + 0.010506480000913143, + 0.008758898824453354, + 0.012458015233278275, + 0.009627828374505043, + 0.00623239204287529, + 0.00531285535544157, + 0.005241685081273317, + 0.010988782159984112, + 0.007706665433943272, + 0.009179262444376945, + 0.007941044867038727, + 0.008235514163970947, + 0.006059506442397833, + 0.004941752180457115, + 0.005476813763380051, + 0.006445086095482111, + 0.008559989742934704, + 0.00892558228224516, + 0.010927468538284302, + 0.009467783384025097, + 0.0024906585458666086, + 0.0017266995273530483, + 0.002512809354811907, + 0.0026986361481249332, + 0.002558043459430337, + 0.003790648654103279, + 0.0039928629994392395, + 0.002510389778763056, + 0.007787290029227734, + 0.011231123469769955, + 0.006986238993704319, + 0.00752684660255909, + 0.007311615627259016, + 0.008870714344084263, + 0.012112578377127647, + 0.008260310627520084, + 0.016777923330664635, + 0.01705753058195114, + 0.013628005981445312, + 0.013519950211048126, + 0.13478632271289825, + 0.021499697118997574, + 0.017121825367212296, + 0.1326405555009842, + 0.0019272046629339457, + 0.002433427609503269, + 0.002387475688010454, + 0.0029049343429505825, + 0.002290592063218355, + 0.003315465757623315, + 0.003073633648455143, + 0.0027465762104839087, + 0.002836516359820962, + 0.006862144451588392, + 0.0030021669808775187, + 0.005248473025858402, + 0.006273283623158932, + 0.006742706522345543, + 0.009576103650033474, + 0.0073144640773534775, + 0.01833287999033928, + 0.01609361730515957, + 0.0180317685008049, + 0.03215915337204933, + 0.02652336284518242, + 0.04790153726935387, + 0.04722519963979721, + 0.04177612066268921, + 0.0025330453645437956, + 0.0038102755788713694, + 0.004972672089934349, + 0.0034467470832169056, + 0.00390837574377656, + 0.00406290078535676, + 0.0036705220118165016, + 0.0030269816052168608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -1.406340479850769 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "violently jarring low-angle close-up plunges the viewer into a nightmarish biomechanical landscape, a fusion of bone and metal, framed by a structure resembling a vast ribcage interwoven with dark, weeping pipes. At the center of the frame, a large, ribbed orifice, its texture a grotesque amalgamation of porous bone and pitted, corroded steel, pulsing and contracting organically. From this opening, a viscous, gray, cleansing clay erupts, not in a fluid stream but in a thick, undulating, relentless surge that seems to push against the air. As the viscous substance is expelled spasmodically, it churns and bubbles into a grotesque, living foam, its bubbles swelling to unnatural sizes before bursting with dampness, all the while catching the dim, directional light with an oily, iridescent sheen. The entire scene is rendered in the dark, skeletal aesthetic of H.R. Giger, defined by a stark, monochromatic palette of charcoal and cold metal. A single, harsh light source from above carves out shimmering highlights and deep, cavernous shadows, creating an oppressive and claustrophobic atmosphere of pure visceral terror. The camera continues to shake violently, as if shaken by the vibrations of this horrific birth.", + "seed": 16042, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_5/index_4265_seed_16042_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2904675167053938, + "vlm_logprob": -0.2904675167053938, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.2904675167053938, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00888324249535799 + ], + "opposite_logprob_tokens": [ + -5.258883476257324 + ], + "logprob_reward": -0.00888324249535799, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00888324249535799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2908630669116974, + 0.12530100345611572, + 0.0861978530883789, + 0.06473758816719055, + 0.07491226494312286, + 0.07821828126907349, + 0.09050358831882477, + 0.054703932255506516, + 0.07428489625453949, + 0.06027749553322792 + ], + "credit_map": [ + 0.026986679062247276, + 0.04066312313079834, + 0.03683694079518318, + 0.031316354870796204, + 0.04718372970819473, + 0.03528976812958717, + 0.03791738301515579, + 0.03466910496354103, + 0.010185960680246353, + 0.009732019156217575, + 0.016430547460913658, + 0.022190380841493607, + 0.018720855936408043, + 0.008976995013654232, + 0.027616947889328003, + 0.011447308585047722, + 0.010572465136647224, + 0.014163252897560596, + 0.00885680876672268, + 0.010692653246223927, + 0.01106683723628521, + 0.010009760037064552, + 0.010093159042298794, + 0.010742925107479095, + 0.006887235678732395, + 0.005066681187599897, + 0.007097951136529446, + 0.007162661291658878, + 0.005544712767004967, + 0.010716409422457218, + 0.006731835659593344, + 0.015530106611549854, + 0.00823686271905899, + 0.010782876051962376, + 0.01153995655477047, + 0.009577642194926739, + 0.009458610787987709, + 0.007142765913158655, + 0.009935915470123291, + 0.008237644098699093, + 0.00799644086509943, + 0.009322812780737877, + 0.010363824665546417, + 0.00822101254016161, + 0.012364006601274014, + 0.006564723793417215, + 0.011480779387056828, + 0.011904686689376831, + 0.006889731157571077, + 0.008205323480069637, + 0.008298149332404137, + 0.006641410756856203, + 0.009994562715291977, + 0.023347904905676842, + 0.006794045679271221, + 0.020332472398877144, + 0.005093182437121868, + 0.002364058280363679, + 0.003526656422764063, + 0.017072824761271477, + 0.002731186104938388, + 0.004000286106020212, + 0.005157400853931904, + 0.014758345670998096, + 0.009193537756800652, + 0.0060063921846449375, + 0.00788364838808775, + 0.011128487065434456, + 0.00891160685569048, + 0.009733070619404316, + 0.008117380551993847, + 0.013310781680047512, + 0.0061368923634290695, + 0.0063649751245975494, + 0.00616061594337225, + 0.014122672379016876, + 0.007261579390615225, + 0.004708748310804367, + 0.00952126644551754, + 0.0060007511638104916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.00888324249535799 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002833995968103409 + ], + "opposite_logprob_tokens": [ + -5.877833843231201 + ], + "logprob_reward": -0.002833995968103409, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002833995968103409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2547034025192261, + 0.07885244488716125, + 0.08481839299201965, + 0.05891549959778786, + 0.08629390597343445, + 0.054157692939043045, + 0.0749807357788086, + 0.12618529796600342, + 0.11810541898012161, + 0.06298714876174927 + ], + "credit_map": [ + 0.03000308759510517, + 0.036390598863363266, + 0.03197096660733223, + 0.030717067420482635, + 0.03949486091732979, + 0.021621057763695717, + 0.03321854770183563, + 0.03128722310066223, + 0.006301488261669874, + 0.006752068642526865, + 0.010360248386859894, + 0.014376127161085606, + 0.010801460593938828, + 0.005218338221311569, + 0.018318817019462585, + 0.006723904982209206, + 0.011300470679998398, + 0.013275706209242344, + 0.009432940743863583, + 0.009858744218945503, + 0.01348411850631237, + 0.007739703636616468, + 0.0104819405823946, + 0.009244776330888271, + 0.00640252185985446, + 0.005059097893536091, + 0.007954969070851803, + 0.009943688288331032, + 0.007511517498642206, + 0.0060434709303081036, + 0.006865743547677994, + 0.009134497493505478, + 0.00804932601749897, + 0.010995199903845787, + 0.009850047528743744, + 0.012945898808538914, + 0.01120891235768795, + 0.00885852426290512, + 0.015173094347119331, + 0.00921290647238493, + 0.004022540990263224, + 0.0067067863419651985, + 0.005426927004009485, + 0.00510922446846962, + 0.013404764235019684, + 0.004400687757879496, + 0.006983366794884205, + 0.008103400468826294, + 0.005986477714031935, + 0.004941345192492008, + 0.009124931879341602, + 0.00797511637210846, + 0.00874545332044363, + 0.0179822389036417, + 0.006328545976430178, + 0.013896631076931953, + 0.008749738335609436, + 0.004262513481080532, + 0.009325076825916767, + 0.03181753307580948, + 0.00588421942666173, + 0.011576884426176548, + 0.01237048301845789, + 0.04219886288046837, + 0.013267096132040024, + 0.009720312431454659, + 0.012762441299855709, + 0.016424713656306267, + 0.016131248325109482, + 0.014794097281992435, + 0.014210998080670834, + 0.020794520154595375, + 0.005210396833717823, + 0.00819308590143919, + 0.005838289856910706, + 0.013371980749070644, + 0.00799762737005949, + 0.005654129199683666, + 0.008742290548980236, + 0.00797935202717781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.002833995968103409 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293979734182358 + ], + "opposite_logprob_tokens": [ + -2.2379398345947266 + ], + "logprob_reward": -0.11293979734182358, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293979734182358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12572221457958221, + 0.06812146306037903, + 0.05274114012718201, + 0.017975691705942154, + 0.037538573145866394, + 0.047173310071229935, + 0.09483103454113007, + 0.3638339638710022, + 0.12465638667345047, + 0.06740614771842957 + ], + "credit_map": [ + 0.015639275312423706, + 0.014326558448374271, + 0.023064807057380676, + 0.014742042869329453, + 0.01580510288476944, + 0.012672418728470802, + 0.015969140455126762, + 0.013502869755029678, + 0.006932452321052551, + 0.005358507856726646, + 0.009282960556447506, + 0.012610284611582756, + 0.008080721832811832, + 0.00384099711664021, + 0.01610456220805645, + 0.005910973530262709, + 0.009983215481042862, + 0.005142544396221638, + 0.0063705965876579285, + 0.006551344878971577, + 0.006075016688555479, + 0.0059273079968988895, + 0.00662502134218812, + 0.006066087167710066, + 0.0030569236259907484, + 0.0018532306421548128, + 0.002158781047910452, + 0.002019763458520174, + 0.001993580488488078, + 0.002257948275655508, + 0.0016491778660565615, + 0.0029862860683351755, + 0.004660734906792641, + 0.0041330172680318356, + 0.00393020361661911, + 0.003916371148079634, + 0.004656018223613501, + 0.006273041013628244, + 0.005049263592809439, + 0.004919923841953278, + 0.004344499204307795, + 0.005748136900365353, + 0.0054489001631736755, + 0.005372319370508194, + 0.007580969482660294, + 0.003626981284469366, + 0.008308335207402706, + 0.006743166595697403, + 0.016262995079159737, + 0.006916002836078405, + 0.009027981199324131, + 0.012575585395097733, + 0.012180701829493046, + 0.013991625979542732, + 0.008643100038170815, + 0.015233044512569904, + 0.02219967730343342, + 0.013138067908585072, + 0.019976560026407242, + 0.10341586917638779, + 0.014695636928081512, + 0.031903352588415146, + 0.034193914383649826, + 0.12431090325117111, + 0.010125228203833103, + 0.009268457069993019, + 0.011248739436268806, + 0.01841970905661583, + 0.018889056518673897, + 0.015565693378448486, + 0.015698159113526344, + 0.025441346690058708, + 0.0065603856928646564, + 0.006357484497129917, + 0.007740785833448172, + 0.015364200808107853, + 0.0067400382831692696, + 0.006295638624578714, + 0.010282453149557114, + 0.00806516408920288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.11293979734182358 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014221508987247944 + ], + "opposite_logprob_tokens": [ + -4.264221668243408 + ], + "logprob_reward": -0.014221508987247944, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014221508987247944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11749543249607086, + 0.07044799625873566, + 0.04228765517473221, + 0.03923055902123451, + 0.028044426813721657, + 0.04569971188902855, + 0.12687139213085175, + 0.4269208014011383, + 0.06368640810251236, + 0.039315611124038696 + ], + "credit_map": [ + 0.012815852649509907, + 0.0087622394785285, + 0.029504947364330292, + 0.011264712549746037, + 0.011935646645724773, + 0.010775832459330559, + 0.018994683399796486, + 0.013441527262330055, + 0.006170405074954033, + 0.005861172918230295, + 0.0084979934617877, + 0.011226959526538849, + 0.014325723983347416, + 0.005687248427420855, + 0.010254454798996449, + 0.008424048312008381, + 0.005591052584350109, + 0.005459936335682869, + 0.006089142058044672, + 0.004714479204267263, + 0.005995170678943396, + 0.004968368913978338, + 0.0053594172932207584, + 0.0041100927628576756, + 0.006684207357466221, + 0.005411923862993717, + 0.006093223579227924, + 0.004134199116379023, + 0.0038531734608113766, + 0.003983819857239723, + 0.0035065996926277876, + 0.005563410464674234, + 0.00333234085701406, + 0.00409567030146718, + 0.0036006823647767305, + 0.0035983817651867867, + 0.004460920579731464, + 0.002298209583386779, + 0.003553152782842517, + 0.003105070674791932, + 0.005402585491538048, + 0.004885144531726837, + 0.007460793945938349, + 0.004453414585441351, + 0.006828051060438156, + 0.003071151440963149, + 0.00701820757240057, + 0.006580364890396595, + 0.012749593704938889, + 0.010774102061986923, + 0.0134773263707757, + 0.018067529425024986, + 0.017035864293575287, + 0.022705137729644775, + 0.011260499246418476, + 0.020801346749067307, + 0.028288373723626137, + 0.016199592500925064, + 0.020945454016327858, + 0.12883125245571136, + 0.017586875706911087, + 0.03423634171485901, + 0.039158228784799576, + 0.1416746973991394, + 0.005663591902703047, + 0.006209515035152435, + 0.007425131741911173, + 0.009217849001288414, + 0.009162087924778461, + 0.006498951464891434, + 0.012485346756875515, + 0.007023932877928019, + 0.004302842542529106, + 0.004462490789592266, + 0.004265223164111376, + 0.007842118851840496, + 0.004174054134637117, + 0.0038585499860346317, + 0.006681982893496752, + 0.0037283487617969513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.014221508987247944 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.313459038734436 + ], + "opposite_logprob_tokens": [ + -0.31345900893211365 + ], + "logprob_reward": -1.313459038734436, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.313459038734436, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21233214437961578, + 0.0700305849313736, + 0.04435503110289574, + 0.03277646750211716, + 0.06057734042406082, + 0.053482525050640106, + 0.10161104798316956, + 0.2683267295360565, + 0.09928296506404877, + 0.057225171476602554 + ], + "credit_map": [ + 0.020094817504286766, + 0.023144837468862534, + 0.025871945545077324, + 0.024880575016140938, + 0.023195020854473114, + 0.03713050112128258, + 0.0268031507730484, + 0.031211305409669876, + 0.00559049379080534, + 0.005830361973494291, + 0.008728048764169216, + 0.009953226894140244, + 0.014178983867168427, + 0.005513375159353018, + 0.013048325665295124, + 0.007187761832028627, + 0.005637832451611757, + 0.005400068126618862, + 0.005067084450274706, + 0.005799406673759222, + 0.006273780018091202, + 0.005089759360998869, + 0.005762764718383551, + 0.005324333440512419, + 0.004811376333236694, + 0.005518649239093065, + 0.004143594764173031, + 0.0033860800322145224, + 0.004193657077848911, + 0.002671506954357028, + 0.004403748083859682, + 0.003647853387519717, + 0.0036795642226934433, + 0.008161491714417934, + 0.0041510253213346004, + 0.009921264834702015, + 0.009283640421926975, + 0.004755438305437565, + 0.01336001418530941, + 0.007264893036335707, + 0.00459824874997139, + 0.00622023269534111, + 0.007779337000101805, + 0.004890614189207554, + 0.010587820783257484, + 0.003899197792634368, + 0.00794103741645813, + 0.00756603479385376, + 0.01882355287671089, + 0.0083104707300663, + 0.01398609858006239, + 0.009569232352077961, + 0.009623031131923199, + 0.01727268099784851, + 0.008344089612364769, + 0.01568189077079296, + 0.017115946859121323, + 0.011043603532016277, + 0.013225611299276352, + 0.08272071927785873, + 0.010678738355636597, + 0.02081935480237007, + 0.025251762941479683, + 0.08747097849845886, + 0.01577850803732872, + 0.009901199489831924, + 0.011279475875198841, + 0.017260778695344925, + 0.010019296780228615, + 0.01219874992966652, + 0.010878666304051876, + 0.011966298334300518, + 0.005430803634226322, + 0.00680825999006629, + 0.006682992447167635, + 0.01147372741252184, + 0.005559323821216822, + 0.004565367475152016, + 0.008391063660383224, + 0.008313636295497417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -1.313459038734436 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "violently jarring low-angle close-up plunges the viewer into a nightmarish biomechanical landscape, a fusion of bone and metal, framed by a structure resembling a vast ribcage interwoven with dark, weeping pipes. At the center of the frame, a large, ribbed orifice, its texture a grotesque amalgamation of porous bone and pitted, corroded steel, pulsing and contracting organically. From this opening, a viscous, gray, cleansing clay erupts, not in a fluid stream but in a thick, undulating, relentless surge that seems to push against the air. As the viscous substance is expelled spasmodically, it churns and bubbles into a grotesque, living foam, its bubbles swelling to unnatural sizes before bursting with dampness, all the while catching the dim, directional light with an oily, iridescent sheen. The entire scene is rendered in the dark, skeletal aesthetic of H.R. Giger, defined by a stark, monochromatic palette of charcoal and cold metal. A single, harsh light source from above carves out shimmering highlights and deep, cavernous shadows, creating an oppressive and claustrophobic atmosphere of pure visceral terror. The camera continues to shake violently, as if shaken by the vibrations of this horrific birth.", + "seed": 16043, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_5/index_4265_seed_16043_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2943267259281129, + "vlm_logprob": -0.2943267259281129, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.2943267259281129, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00826545525342226 + ], + "opposite_logprob_tokens": [ + -5.633265495300293 + ], + "logprob_reward": -0.00826545525342226, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00826545525342226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6755213141441345, + 0.037117552012205124, + 0.022784501314163208, + 0.07216693460941315, + 0.02377619780600071, + 0.015389129519462585, + 0.031676217913627625, + 0.019684292376041412, + 0.02382618561387062, + 0.07805772125720978 + ], + "credit_map": [ + 0.1632716953754425, + 0.02455270104110241, + 0.27486351132392883, + 0.1022970974445343, + 0.019252782687544823, + 0.014423117972910404, + 0.058058835566043854, + 0.018801501020789146, + 0.008811366744339466, + 0.003131208475679159, + 0.007845270447432995, + 0.005089190788567066, + 0.00258941319771111, + 0.002782169496640563, + 0.003086426993831992, + 0.003782504005357623, + 0.0021513523533940315, + 0.0018737736390903592, + 0.003892978187650442, + 0.00324449734762311, + 0.0017592830117791891, + 0.00227574841119349, + 0.003436488099396229, + 0.0041503761895000935, + 0.009652975015342236, + 0.00932843517512083, + 0.009685658849775791, + 0.013577333651483059, + 0.006255717482417822, + 0.007128888741135597, + 0.007973584346473217, + 0.008564339019358158, + 0.0035752439871430397, + 0.002850757446140051, + 0.0027331530582159758, + 0.0024972637183964252, + 0.0034763189032673836, + 0.002977811498567462, + 0.003322211094200611, + 0.0023434371687471867, + 0.0025817963760346174, + 0.0019412896363064647, + 0.0014646146446466446, + 0.00232974486425519, + 0.001686707022599876, + 0.0014658713480457664, + 0.0018902826122939587, + 0.0020288205705583096, + 0.0033335252664983273, + 0.004440819378942251, + 0.0027762295212596655, + 0.004955759737640619, + 0.0030983577016741037, + 0.003632372710853815, + 0.004344317130744457, + 0.005094834603369236, + 0.0026781714987009764, + 0.0018426880706101656, + 0.002190566621720791, + 0.0026079623494297266, + 0.0023641162551939487, + 0.002958101686090231, + 0.0032826270908117294, + 0.0017600575229153037, + 0.0030117135029286146, + 0.0028811655938625336, + 0.0028405513148754835, + 0.006061739753931761, + 0.0017310408875346184, + 0.0015645259991288185, + 0.004058790858834982, + 0.001676655258052051, + 0.01192169263958931, + 0.003773285774514079, + 0.025426799431443214, + 0.01809084415435791, + 0.003086001379415393, + 0.006752328015863895, + 0.00392233906313777, + 0.0050844186916947365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.00826545525342226 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028363734018057585 + ], + "opposite_logprob_tokens": [ + -5.877836227416992 + ], + "logprob_reward": -0.0028363734018057585, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028363734018057585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3140243887901306, + 0.09929167479276657, + 0.04055752232670784, + 0.0491112656891346, + 0.1154574528336525, + 0.06490026414394379, + 0.07651065289974213, + 0.06135043874382973, + 0.0917259082198143, + 0.08707039803266525 + ], + "credit_map": [ + 0.0598168708384037, + 0.012793553061783314, + 0.12166270613670349, + 0.06353741139173508, + 0.009539683349430561, + 0.007740759756416082, + 0.026084819808602333, + 0.012848597019910812, + 0.02895054966211319, + 0.005126176401972771, + 0.029836656525731087, + 0.009363194927573204, + 0.004662647843360901, + 0.00472303805872798, + 0.00817329902201891, + 0.00845610722899437, + 0.003941656555980444, + 0.0037618891801685095, + 0.006000446621328592, + 0.007057568524032831, + 0.0036008211318403482, + 0.004230606369674206, + 0.004888587631285191, + 0.007075944449752569, + 0.005442531779408455, + 0.006380095612257719, + 0.006932783871889114, + 0.01009384822100401, + 0.004696311894804239, + 0.004469608422368765, + 0.0063800811767578125, + 0.004716010298579931, + 0.016562148928642273, + 0.018314272165298462, + 0.011700109578669071, + 0.01571069285273552, + 0.014094779267907143, + 0.010211977176368237, + 0.016527434810996056, + 0.012336031533777714, + 0.007638745009899139, + 0.006296177394688129, + 0.006922000087797642, + 0.011015055701136589, + 0.0073640188202261925, + 0.005587215535342693, + 0.011964892037212849, + 0.008112161420285702, + 0.008382911793887615, + 0.011414539068937302, + 0.007166476920247078, + 0.01169405598193407, + 0.007233564741909504, + 0.007923884317278862, + 0.010100065730512142, + 0.012595157139003277, + 0.008128063753247261, + 0.005242728162556887, + 0.007688277866691351, + 0.007901842705905437, + 0.007984397932887077, + 0.0074602654203772545, + 0.011719283647835255, + 0.005225580185651779, + 0.011514093726873398, + 0.012899521738290787, + 0.011689811944961548, + 0.01511929277330637, + 0.009144852869212627, + 0.007128655910491943, + 0.018380526453256607, + 0.005849155597388744, + 0.010332338511943817, + 0.004075993318110704, + 0.02049812115728855, + 0.02900831773877144, + 0.0035657149273902178, + 0.004179670941084623, + 0.005605477839708328, + 0.009804768487811089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0028363734018057585 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038108546286821365 + ], + "opposite_logprob_tokens": [ + -3.2881085872650146 + ], + "logprob_reward": -0.038108546286821365, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038108546286821365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10217510908842087, + 0.10329505801200867, + 0.053133003413677216, + 0.05323529243469238, + 0.06529925018548965, + 0.07018169015645981, + 0.24687224626541138, + 0.10089928656816483, + 0.1559385359287262, + 0.048970531672239304 + ], + "credit_map": [ + 0.02137170173227787, + 0.008666768670082092, + 0.015866415575146675, + 0.026949457824230194, + 0.0045192912220954895, + 0.0057100290432572365, + 0.0076192524284124374, + 0.011472187004983425, + 0.02432405948638916, + 0.006186699029058218, + 0.03215805068612099, + 0.009553045965731144, + 0.006215328350663185, + 0.004506074357777834, + 0.012158812023699284, + 0.00819298904389143, + 0.007232488598674536, + 0.005023697391152382, + 0.007651984691619873, + 0.009279552847146988, + 0.0045053851790726185, + 0.006253322586417198, + 0.006691379472613335, + 0.006495190318673849, + 0.008509576320648193, + 0.006687456741929054, + 0.006141540594398975, + 0.006792567204684019, + 0.00592648983001709, + 0.007342082913964987, + 0.0058303168043494225, + 0.006005266215652227, + 0.009526720270514488, + 0.006996103562414646, + 0.007129623554646969, + 0.007860436104238033, + 0.011298662051558495, + 0.008132734335958958, + 0.007201283238828182, + 0.007153687998652458, + 0.009054620750248432, + 0.006025993265211582, + 0.00794932246208191, + 0.011344804428517818, + 0.00860033929347992, + 0.006697464734315872, + 0.013581138104200363, + 0.006928010378032923, + 0.02499529719352722, + 0.03515038639307022, + 0.024058211594820023, + 0.03572532534599304, + 0.027878187596797943, + 0.030433496460318565, + 0.03209475055336952, + 0.036536604166030884, + 0.014188669621944427, + 0.010815059766173363, + 0.01368270255625248, + 0.014402396976947784, + 0.010450158268213272, + 0.009174784645438194, + 0.019293176010251045, + 0.008892333135008812, + 0.019371498376131058, + 0.017518995329737663, + 0.020937638357281685, + 0.02836369350552559, + 0.01822536066174507, + 0.010773353278636932, + 0.027708951383829117, + 0.013039042241871357, + 0.00593333737924695, + 0.0029037194326519966, + 0.011646836996078491, + 0.01288679614663124, + 0.002855513943359256, + 0.00446347938850522, + 0.004535763990134001, + 0.0037450811360031366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.038108546286821365 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016091477125883102 + ], + "opposite_logprob_tokens": [ + -4.141091346740723 + ], + "logprob_reward": -0.016091477125883102, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016091477125883102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32644638419151306, + 0.10208892822265625, + 0.06417368352413177, + 0.03990977630019188, + 0.11844148486852646, + 0.05264003574848175, + 0.05662738159298897, + 0.034449342638254166, + 0.04768357425928116, + 0.15753939747810364 + ], + "credit_map": [ + 0.037290796637535095, + 0.017961880192160606, + 0.10260450839996338, + 0.09618441760540009, + 0.016883475705981255, + 0.012236167676746845, + 0.02801566943526268, + 0.015269458293914795, + 0.01647535152733326, + 0.01232904102653265, + 0.014502130448818207, + 0.00964854285120964, + 0.01231189165264368, + 0.007536699995398521, + 0.019319836050271988, + 0.0099654421210289, + 0.007907151244580746, + 0.006191102322191, + 0.01283180620521307, + 0.010879182256758213, + 0.004858708009123802, + 0.006321519613265991, + 0.009474927559494972, + 0.005709285847842693, + 0.006149503402411938, + 0.004045508801937103, + 0.004971414804458618, + 0.0038436108734458685, + 0.003426529001444578, + 0.005360140930861235, + 0.003976157400757074, + 0.008136911317706108, + 0.01913190446794033, + 0.018316499888896942, + 0.012148318812251091, + 0.02133256569504738, + 0.009670695289969444, + 0.008781423792243004, + 0.020145634189248085, + 0.008914437144994736, + 0.011521570384502411, + 0.006167043000459671, + 0.005332418251782656, + 0.00944820512086153, + 0.004237256478518248, + 0.0036096060648560524, + 0.008689066395163536, + 0.0036348721478134394, + 0.006796133704483509, + 0.007029250264167786, + 0.0056472946889698505, + 0.009013966657221317, + 0.007233110722154379, + 0.005543536506593227, + 0.009083333425223827, + 0.006280754692852497, + 0.00383197795599699, + 0.0035995864309370518, + 0.0032934311311692, + 0.004746831022202969, + 0.004732292145490646, + 0.0047116405330598354, + 0.006189150270074606, + 0.003344437573105097, + 0.005478633567690849, + 0.0069564939476549625, + 0.0056051756255328655, + 0.007912115193903446, + 0.003933390136808157, + 0.004440503194928169, + 0.009805131703615189, + 0.003552132286131382, + 0.021070219576358795, + 0.006003257352858782, + 0.05392201617360115, + 0.05209225416183472, + 0.003835571464151144, + 0.004478370770812035, + 0.009320754557847977, + 0.006816943176090717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.016091477125883102 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a biomechanical landscape featuring a fusion of bone and metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a large, ribbed orifice at the center of the frame with a texture of porous bone and corroded steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thick, gray, viscous clay erupt from the orifice in an undulating surge that churns into living foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a monochromatic palette of charcoal and cold metal with an oily, iridescent sheen on the foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shake violently throughout the video?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063317775726318 + ], + "opposite_logprob_tokens": [ + -0.28133177757263184 + ], + "logprob_reward": -1.4063317775726318, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063317775726318, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.777202844619751, + 0.044782672077417374, + 0.018503431230783463, + 0.027930907905101776, + 0.008591216057538986, + 0.014063980430364609, + 0.023007694631814957, + 0.022138167172670364, + 0.03143377974629402, + 0.03234531730413437 + ], + "credit_map": [ + 0.13410380482673645, + 0.035234395414590836, + 0.2985236942768097, + 0.18214638531208038, + 0.02435021474957466, + 0.014838200062513351, + 0.06433815509080887, + 0.02366805076599121, + 0.00961867906153202, + 0.0037904875352978706, + 0.013475233688950539, + 0.003929965663701296, + 0.003518506186082959, + 0.002643361920490861, + 0.0044923401437699795, + 0.0033140985760837793, + 0.0019527151016518474, + 0.0018324891570955515, + 0.0030564882326871157, + 0.0030173740815371275, + 0.0014383866218850017, + 0.002049531089141965, + 0.0023635076358914375, + 0.002792938379570842, + 0.003421539207920432, + 0.0037430948577821255, + 0.0037080496549606323, + 0.004847331438213587, + 0.0026601464487612247, + 0.0032297009602189064, + 0.003168961498886347, + 0.0031520831398665905, + 0.0012759771198034286, + 0.0014469267334789038, + 0.0009701650706119835, + 0.0011928945314139128, + 0.0007010460249148309, + 0.0008562143775634468, + 0.001479443395510316, + 0.0006685489206574857, + 0.0019759107381105423, + 0.0011801921064034104, + 0.0015915562398731709, + 0.0022976736072450876, + 0.001355502987280488, + 0.0015475853579118848, + 0.002241813112050295, + 0.0018737456994131207, + 0.002774664433673024, + 0.002769813174381852, + 0.0021861593704670668, + 0.0031322496943175793, + 0.0035509485751390457, + 0.0030420771799981594, + 0.0034975619055330753, + 0.00205422006547451, + 0.0032375778537243605, + 0.0023749307729303837, + 0.0028941843193024397, + 0.002992655849084258, + 0.00217803125269711, + 0.0019670603796839714, + 0.004352996591478586, + 0.0021407315507531166, + 0.0038468465209007263, + 0.0033505097962915897, + 0.0047830152325332165, + 0.006563133560121059, + 0.0030189489480108023, + 0.002423574449494481, + 0.005090170074254274, + 0.0023575813975185156, + 0.0059776934795081615, + 0.0025254867505282164, + 0.010202514939010143, + 0.005680793896317482, + 0.0013932613655924797, + 0.002396248746663332, + 0.00226741936057806, + 0.0019018969032913446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -1.4063317775726318 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_6/index_tensor([1434, 1434])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_6/index_tensor([1434, 1434])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ec575e456085585a46b554267fbd259042ac6001 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_6/index_tensor([1434, 1434])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme macro lens begins, then slowly and steadily pulls back, capturing the dynamics of countless mosaic tile fragments, shimmering with an inherent warmth. Composed primarily of ceramic with subtle brushstroke textures and translucent stained glass, these fragments display hues of soft cream, warm ochre, sky-like pale blue, and forest-like emerald green, each possessing imperfect, hand-cut edges. Against a deep, blurred background, they seem pulled by an invisible force, gracefully floating and swirling through the air, gently colliding with one another. Each contact releases a fleeting burst of golden dust accompanied by a crisp, melodious tinkling sound. As the camera continues to zoom out, the fragments begin to assemble in an orderly fashion, like pieces of a living puzzle, quickly and precisely finding their place, fitting seamlessly into place, and gradually piecing together a vast, fantastical mural. Eventually, the image stabilized, revealing the mural in its entirety – depicting floating islands, winding rivers, and strange yet friendly forest creatures. The entire work is imbued with the distinctive hand-painted texture of Studio Ghibli, with rich and soft colors, and delicate light and shadow, creating an atmosphere that is dreamy, tranquil, and full of the miracle of life, like a vivid original animation painting from the hands of a master.", + "seed": 11437, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_6/index_1434_seed_11437_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03712507300078869, + "vlm_logprob": -0.03712507300078869, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.03712507300078869, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03502730652689934 + ], + "opposite_logprob_tokens": [ + -3.535027265548706 + ], + "logprob_reward": -0.03502730652689934, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03502730652689934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25458624958992004, + 0.11872701346874237, + 0.05276225134730339, + 0.07597413659095764, + 0.05408075451850891, + 0.1627993881702423, + 0.09073327481746674, + 0.07979288697242737, + 0.0664447471499443, + 0.04409932345151901 + ], + "credit_map": [ + 0.044798821210861206, + 0.03223147988319397, + 0.03124900534749031, + 0.02595311775803566, + 0.034847673028707504, + 0.0472913458943367, + 0.016723092645406723, + 0.02149171568453312, + 0.015217268839478493, + 0.017403706908226013, + 0.019255347549915314, + 0.009615913033485413, + 0.02209055982530117, + 0.01525899302214384, + 0.010058827698230743, + 0.009826400317251682, + 0.008703538216650486, + 0.007736994884908199, + 0.00667519960552454, + 0.005095934495329857, + 0.006512429099529982, + 0.0056811049580574036, + 0.005689986981451511, + 0.00666706170886755, + 0.0080933952704072, + 0.00954778864979744, + 0.007474890444427729, + 0.010803626850247383, + 0.012380541302263737, + 0.008550272323191166, + 0.011809169314801693, + 0.007314458955079317, + 0.00614542979747057, + 0.007066664751619101, + 0.007225195411592722, + 0.008016870357096195, + 0.007829418405890465, + 0.007241189479827881, + 0.005017050076276064, + 0.005538934376090765, + 0.014897271990776062, + 0.017798710614442825, + 0.02921725995838642, + 0.02839835360646248, + 0.021173277869820595, + 0.01587011106312275, + 0.019624043256044388, + 0.015820367261767387, + 0.01307007484138012, + 0.013185356743633747, + 0.012412410229444504, + 0.015286396257579327, + 0.007229391019791365, + 0.007884645834565163, + 0.012219615280628204, + 0.009445391595363617, + 0.011303822509944439, + 0.009706626646220684, + 0.008154865354299545, + 0.008323992602527142, + 0.009777747094631195, + 0.007249588146805763, + 0.01341694314032793, + 0.011859307065606117, + 0.007265636697411537, + 0.007216347381472588, + 0.011122964322566986, + 0.006422623060643673, + 0.007872093468904495, + 0.010247891768813133, + 0.007757953368127346, + 0.008539238013327122, + 0.006756490562111139, + 0.004679273813962936, + 0.007826385088264942, + 0.005265721585601568, + 0.0055855195969343185, + 0.003346031764522195, + 0.005630523432046175, + 0.0050093806348741055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.03502730652689934 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007649179082363844 + ], + "opposite_logprob_tokens": [ + -4.882648944854736 + ], + "logprob_reward": -0.007649179082363844, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007649179082363844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23627527058124542, + 0.06374652683734894, + 0.054633039981126785, + 0.03714565932750702, + 0.061146654188632965, + 0.11453057825565338, + 0.056737177073955536, + 0.19723232090473175, + 0.1275908350944519, + 0.05096188932657242 + ], + "credit_map": [ + 0.029938215389847755, + 0.032270513474941254, + 0.0302272979170084, + 0.03138003870844841, + 0.03666047751903534, + 0.022431958466768265, + 0.03336004912853241, + 0.02000671625137329, + 0.006520550232380629, + 0.01205163449048996, + 0.012035618536174297, + 0.004324572626501322, + 0.013245063833892345, + 0.006312566343694925, + 0.004354726057499647, + 0.004901799838989973, + 0.0095383170992136, + 0.006517727393656969, + 0.010095755569636822, + 0.007010747212916613, + 0.0065032197162508965, + 0.005882069934159517, + 0.00474578607827425, + 0.004339416045695543, + 0.004636992234736681, + 0.005212557967752218, + 0.004831104539334774, + 0.003920875955373049, + 0.0046425554901361465, + 0.004707252606749535, + 0.005567873828113079, + 0.003626446705311537, + 0.005612259265035391, + 0.007272347342222929, + 0.006869670934975147, + 0.011466906405985355, + 0.011054796166718006, + 0.007547812070697546, + 0.006745966617017984, + 0.004576893523335457, + 0.012684314511716366, + 0.014692159369587898, + 0.016841024160385132, + 0.013648663647472858, + 0.017602747306227684, + 0.010994629934430122, + 0.015507671050727367, + 0.012559370137751102, + 0.006660387851297855, + 0.0081937899813056, + 0.006351287942379713, + 0.007325276266783476, + 0.005659549962729216, + 0.005932103842496872, + 0.008974563330411911, + 0.007640216965228319, + 0.02646283619105816, + 0.025170641019940376, + 0.02069859579205513, + 0.01720266044139862, + 0.03411300480365753, + 0.017384285107254982, + 0.02720511145889759, + 0.028995197266340256, + 0.012665962800383568, + 0.014928142540156841, + 0.02924591302871704, + 0.015246269293129444, + 0.01527775265276432, + 0.011172331869602203, + 0.013929378241300583, + 0.015125090256333351, + 0.005170152522623539, + 0.0066085271537303925, + 0.008894916623830795, + 0.007768070790916681, + 0.006707283668220043, + 0.004348880611360073, + 0.0069047450087964535, + 0.004559315741062164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.007649179082363844 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293000727891922 + ], + "opposite_logprob_tokens": [ + -2.2379300594329834 + ], + "logprob_reward": -0.11293000727891922, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293000727891922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14352932572364807, + 0.05681147426366806, + 0.04644668847322464, + 0.028289757668972015, + 0.059649594128131866, + 0.25808247923851013, + 0.09531594067811966, + 0.1939614862203598, + 0.09126165509223938, + 0.026651598513126373 + ], + "credit_map": [ + 0.015621696598827839, + 0.01658463478088379, + 0.014982405118644238, + 0.020084846764802933, + 0.020405761897563934, + 0.02424089051783085, + 0.02133003994822502, + 0.010279040783643723, + 0.009644321165978909, + 0.008229083381593227, + 0.007164097391068935, + 0.007125491742044687, + 0.00987211987376213, + 0.006176616996526718, + 0.004696196876466274, + 0.0039035428781062365, + 0.006054115481674671, + 0.004051635041832924, + 0.0043588499538600445, + 0.004239959642291069, + 0.006653643678873777, + 0.008632754907011986, + 0.00678264582529664, + 0.005673084873706102, + 0.003658778266981244, + 0.003625344717875123, + 0.0029274404514580965, + 0.0027053384110331535, + 0.00406253058463335, + 0.0028628746513277292, + 0.005293218418955803, + 0.0031542317010462284, + 0.004560353700071573, + 0.005684073083102703, + 0.007661710027605295, + 0.009065256454050541, + 0.01258804090321064, + 0.00812073890119791, + 0.0060136872343719006, + 0.005955731030553579, + 0.027796132490038872, + 0.034652229398489, + 0.03511014208197594, + 0.03155185654759407, + 0.04306584596633911, + 0.026074012741446495, + 0.033264659345149994, + 0.026567602530121803, + 0.012673980556428432, + 0.01626855880022049, + 0.01208910159766674, + 0.013290327973663807, + 0.006371853873133659, + 0.010667573660612106, + 0.011498791165649891, + 0.012455761432647705, + 0.027940373867750168, + 0.02477015182375908, + 0.020684247836470604, + 0.017747674137353897, + 0.028594622388482094, + 0.018429890275001526, + 0.02799331396818161, + 0.027801211923360825, + 0.009556706063449383, + 0.01082752738147974, + 0.017873965203762054, + 0.010227291844785213, + 0.010238608345389366, + 0.011284837499260902, + 0.01099170558154583, + 0.01026101317256689, + 0.0037882765755057335, + 0.0027369027957320213, + 0.0038192886859178543, + 0.0030231340788304806, + 0.002860116073861718, + 0.0022884260397404432, + 0.004976172931492329, + 0.0031592808663845062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.11293000727891922 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734176073223352 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734176073223352, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734176073223352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08022822439670563, + 0.06676733493804932, + 0.11351925134658813, + 0.08786621689796448, + 0.042573027312755585, + 0.08226978778839111, + 0.07574881613254547, + 0.28297755122184753, + 0.07156254351139069, + 0.09648729860782623 + ], + "credit_map": [ + 0.011014004237949848, + 0.009786216542124748, + 0.009702672250568867, + 0.010623638518154621, + 0.008744986727833748, + 0.012141555547714233, + 0.011257611215114594, + 0.00695753050968051, + 0.007445120718330145, + 0.01645732671022415, + 0.006758547853678465, + 0.008422869257628918, + 0.007678109686821699, + 0.00797953736037016, + 0.0054439823143184185, + 0.006581836845725775, + 0.013895348645746708, + 0.01661721244454384, + 0.011944353580474854, + 0.011694393120706081, + 0.018721137195825577, + 0.019585588946938515, + 0.010706921108067036, + 0.010354297235608101, + 0.011604946106672287, + 0.013013157062232494, + 0.010992986150085926, + 0.007652310188859701, + 0.013039806857705116, + 0.009441968984901905, + 0.012328561395406723, + 0.009792479686439037, + 0.004461933858692646, + 0.00653854850679636, + 0.005975008010864258, + 0.004410538822412491, + 0.006928948685526848, + 0.006156283896416426, + 0.003512725932523608, + 0.004589034244418144, + 0.009729019366204739, + 0.010188527405261993, + 0.01332743838429451, + 0.011330155655741692, + 0.00698334164917469, + 0.00860802922397852, + 0.011157500557601452, + 0.010945779271423817, + 0.010587591677904129, + 0.009852820076048374, + 0.010989309288561344, + 0.013812542892992496, + 0.00679819006472826, + 0.006152144633233547, + 0.011040398851037025, + 0.006515826564282179, + 0.041716352105140686, + 0.034873731434345245, + 0.02967534214258194, + 0.0271923765540123, + 0.038553155958652496, + 0.026742108166217804, + 0.042134832590818405, + 0.04208965599536896, + 0.007730016950517893, + 0.008785366080701351, + 0.012193378992378712, + 0.008306450210511684, + 0.0074234288185834885, + 0.009039781987667084, + 0.009834332391619682, + 0.008249781094491482, + 0.01380245666950941, + 0.009164021350443363, + 0.015421008691191673, + 0.011625921353697777, + 0.00970206968486309, + 0.009036153554916382, + 0.016080759465694427, + 0.011654902249574661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.006734176073223352 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328469604253769 + ], + "opposite_logprob_tokens": [ + -3.773284673690796 + ], + "logprob_reward": -0.02328469604253769, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328469604253769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05129680782556534, + 0.03345135971903801, + 0.02743755653500557, + 0.05383164808154106, + 0.046832483261823654, + 0.11287885904312134, + 0.08688123524188995, + 0.2352549284696579, + 0.24184218049049377, + 0.11029297113418579 + ], + "credit_map": [ + 0.005817662924528122, + 0.008027863688766956, + 0.006502558011561632, + 0.008632651530206203, + 0.006235216744244099, + 0.005511670373380184, + 0.005669871810823679, + 0.004899312742054462, + 0.0038107295986264944, + 0.004843743517994881, + 0.005575375631451607, + 0.004184393212199211, + 0.005925450008362532, + 0.003232829738408327, + 0.002573442179709673, + 0.00330539233982563, + 0.004162260331213474, + 0.00292986910790205, + 0.003548207227140665, + 0.0033569943625479937, + 0.004728425294160843, + 0.003270618850365281, + 0.0025038444437086582, + 0.0029373373836278915, + 0.006432213354855776, + 0.006909534335136414, + 0.006233257241547108, + 0.007931509055197239, + 0.007273790426552296, + 0.006033203098922968, + 0.008162043988704681, + 0.004856094717979431, + 0.00520376767963171, + 0.004882762674242258, + 0.00707732979208231, + 0.006664926651865244, + 0.006167000159621239, + 0.00867425836622715, + 0.0040458794683218, + 0.0041165598668158054, + 0.01293504610657692, + 0.016002634540200233, + 0.014090794138610363, + 0.012878818437457085, + 0.010817529633641243, + 0.01017878670245409, + 0.019087782129645348, + 0.016887463629245758, + 0.0097054997459054, + 0.010477619245648384, + 0.011078854091465473, + 0.011936863884329796, + 0.008893943391740322, + 0.011416992172598839, + 0.013308330439031124, + 0.01006313692778349, + 0.03702805936336517, + 0.032400768250226974, + 0.024698268622159958, + 0.02083495818078518, + 0.03016495145857334, + 0.02821614407002926, + 0.030505642294883728, + 0.03140613064169884, + 0.02562883123755455, + 0.030491948127746582, + 0.024794071912765503, + 0.023657221347093582, + 0.03227221593260765, + 0.03990631178021431, + 0.036913640797138214, + 0.028177954256534576, + 0.014177057892084122, + 0.012550398707389832, + 0.021174566820263863, + 0.014900906011462212, + 0.015717610716819763, + 0.007499231956899166, + 0.012176882475614548, + 0.012096316553652287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.02328469604253769 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro lens begins, then slowly and steadily pulls back, capturing the dynamics of countless mosaic tile fragments, shimmering with an inherent warmth. Composed primarily of ceramic with subtle brushstroke textures and translucent stained glass, these fragments display hues of soft cream, warm ochre, sky-like pale blue, and forest-like emerald green, each possessing imperfect, hand-cut edges. Against a deep, blurred background, they seem pulled by an invisible force, gracefully floating and swirling through the air, gently colliding with one another. Each contact releases a fleeting burst of golden dust accompanied by a crisp, melodious tinkling sound. As the camera continues to zoom out, the fragments begin to assemble in an orderly fashion, like pieces of a living puzzle, quickly and precisely finding their place, fitting seamlessly into place, and gradually piecing together a vast, fantastical mural. Eventually, the image stabilized, revealing the mural in its entirety – depicting floating islands, winding rivers, and strange yet friendly forest creatures. The entire work is imbued with the distinctive hand-painted texture of Studio Ghibli, with rich and soft colors, and delicate light and shadow, creating an atmosphere that is dreamy, tranquil, and full of the miracle of life, like a vivid original animation painting from the hands of a master.", + "seed": 11438, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_6/index_1434_seed_11438_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.012652142718434334, + "vlm_logprob": -0.012652142718434334, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.012652142718434334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015363091602921486 + ], + "opposite_logprob_tokens": [ + -4.3903632164001465 + ], + "logprob_reward": -0.015363091602921486, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015363091602921486, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3509480357170105, + 0.14223337173461914, + 0.11156357079744339, + 0.14412015676498413, + 0.07873988151550293, + 0.04780333489179611, + 0.03753219544887543, + 0.016497192904353142, + 0.04434102401137352, + 0.026221219450235367 + ], + "credit_map": [ + 0.027896150946617126, + 0.02069777436554432, + 0.1010018140077591, + 0.02260318584740162, + 0.03235812857747078, + 0.020668314769864082, + 0.04552765563130379, + 0.08019500970840454, + 0.011689550243318081, + 0.012783157639205456, + 0.03023572824895382, + 0.04100579023361206, + 0.009649446234107018, + 0.011257839389145374, + 0.014128448441624641, + 0.011483412235975266, + 0.013878013007342815, + 0.010737272910773754, + 0.012368575669825077, + 0.014896516688168049, + 0.018654143437743187, + 0.014983330853283405, + 0.01502650510519743, + 0.01101921871304512, + 0.018084395676851273, + 0.018774675205349922, + 0.02263944037258625, + 0.014587963931262493, + 0.014704140834510326, + 0.022010495886206627, + 0.014467325992882252, + 0.01885172724723816, + 0.011387046426534653, + 0.009798157028853893, + 0.00844938401132822, + 0.008072227239608765, + 0.00996596459299326, + 0.015178863890469074, + 0.00782103929668665, + 0.008067204616963863, + 0.005101783666759729, + 0.005900442134588957, + 0.004616049584001303, + 0.0054307118989527225, + 0.007025513332337141, + 0.007113568484783173, + 0.004646971821784973, + 0.007968292571604252, + 0.005514294840395451, + 0.004874811973422766, + 0.0042064557783305645, + 0.0036627263762056828, + 0.004535526502877474, + 0.004390742164105177, + 0.005452780518680811, + 0.004894858226180077, + 0.0012359647080302238, + 0.0023950038012117147, + 0.0013883276842534542, + 0.0029530958272516727, + 0.0023026021663099527, + 0.002331252209842205, + 0.0025926472153514624, + 0.0012982998741790652, + 0.011334637179970741, + 0.003952250350266695, + 0.011723469011485577, + 0.003694454673677683, + 0.002964467741549015, + 0.002614404773339629, + 0.002917689736932516, + 0.005139648914337158, + 0.0016342690214514732, + 0.001740357605740428, + 0.0065503497608006, + 0.0017364476807415485, + 0.006945350673049688, + 0.00275438092648983, + 0.002687398111447692, + 0.0021726666018366814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.015363091602921486 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008666286244988441 + ], + "opposite_logprob_tokens": [ + -4.758666515350342 + ], + "logprob_reward": -0.008666286244988441, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008666286244988441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12174906581640244, + 0.32890546321868896, + 0.12246546894311905, + 0.045007526874542236, + 0.06284711509943008, + 0.03359045833349228, + 0.048695553094148636, + 0.10200544446706772, + 0.05133309215307236, + 0.08340077847242355 + ], + "credit_map": [ + 0.01083045732229948, + 0.013175123371183872, + 0.02823026292026043, + 0.0140516497194767, + 0.01260011363774538, + 0.009813770651817322, + 0.018371423706412315, + 0.014676264487206936, + 0.01792668178677559, + 0.01802506111562252, + 0.11555804312229156, + 0.1182502806186676, + 0.015276607125997543, + 0.011655616573989391, + 0.01748465560376644, + 0.014728548936545849, + 0.010380648076534271, + 0.010344598442316055, + 0.01432263758033514, + 0.016223778948187828, + 0.011420897208154202, + 0.023358039557933807, + 0.02947736717760563, + 0.006937501020729542, + 0.005432859994471073, + 0.005767380353063345, + 0.004851500038057566, + 0.006285013630986214, + 0.004048204980790615, + 0.0069896201603114605, + 0.005840962752699852, + 0.005791983101516962, + 0.011795462109148502, + 0.010443504899740219, + 0.006679297890514135, + 0.006459231488406658, + 0.00782057736068964, + 0.009248405694961548, + 0.006372017320245504, + 0.004028619732707739, + 0.004395248834043741, + 0.0034328096080571413, + 0.003477085381746292, + 0.004411165602505207, + 0.004305511713027954, + 0.004052447155117989, + 0.0040957205928862095, + 0.005420467350631952, + 0.003973647020757198, + 0.005398289300501347, + 0.004017856437712908, + 0.004744794685393572, + 0.007259931415319443, + 0.006211582571268082, + 0.008578568696975708, + 0.00851087924093008, + 0.006513602565973997, + 0.007600009441375732, + 0.007543864194303751, + 0.020625397562980652, + 0.01281431782990694, + 0.027803171426057816, + 0.010743686929345131, + 0.008361399173736572, + 0.0049842894077301025, + 0.005284091457724571, + 0.012327481992542744, + 0.005295926239341497, + 0.006425576750189066, + 0.0038717249408364296, + 0.005256866104900837, + 0.00788713339716196, + 0.005321541801095009, + 0.006482386961579323, + 0.02328803576529026, + 0.0045709311962127686, + 0.02467678114771843, + 0.006830739788711071, + 0.005674710962921381, + 0.006555651314556599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.008666286244988441 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350608095526695 + ], + "opposite_logprob_tokens": [ + -3.651350498199463 + ], + "logprob_reward": -0.026350608095526695, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350608095526695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14156457781791687, + 0.41976961493492126, + 0.04963209107518196, + 0.10363978892564774, + 0.05262962728738785, + 0.04222235083580017, + 0.04233682528138161, + 0.05789424479007721, + 0.060422077775001526, + 0.02988884039223194 + ], + "credit_map": [ + 0.01538866013288498, + 0.012806861661374569, + 0.02873779460787773, + 0.01408630795776844, + 0.016708143055438995, + 0.013946954160928726, + 0.015331866219639778, + 0.024557974189519882, + 0.024106603115797043, + 0.02476091869175434, + 0.15136368572711945, + 0.15233062207698822, + 0.011704077944159508, + 0.01566510647535324, + 0.021755337715148926, + 0.018083231523633003, + 0.0042900945991277695, + 0.005599571857601404, + 0.004715782590210438, + 0.007094212342053652, + 0.006753424182534218, + 0.004871359560638666, + 0.007302640471607447, + 0.00900499988347292, + 0.015083574689924717, + 0.016102904453873634, + 0.012849945574998856, + 0.015201196074485779, + 0.010647563263773918, + 0.01171155832707882, + 0.013743987306952477, + 0.008299051783978939, + 0.008103909902274609, + 0.006281808950006962, + 0.005899738520383835, + 0.005526805762201548, + 0.007558994460850954, + 0.00790127832442522, + 0.006939958781003952, + 0.004417127463966608, + 0.007252508774399757, + 0.004546594340354204, + 0.0033159288577735424, + 0.004817079287022352, + 0.006092338357120752, + 0.0044565810821950436, + 0.006331904791295528, + 0.0054094139486551285, + 0.005138072185218334, + 0.004579642787575722, + 0.004904737696051598, + 0.005734789650887251, + 0.0073769716545939445, + 0.00537931127473712, + 0.005727982148528099, + 0.003495312761515379, + 0.00436393404379487, + 0.004491155967116356, + 0.0044947341084480286, + 0.014967236667871475, + 0.004373830743134022, + 0.015731852501630783, + 0.004973176866769791, + 0.004498318303376436, + 0.00810698140412569, + 0.0046760281547904015, + 0.017182445153594017, + 0.006095549091696739, + 0.007551646325737238, + 0.004597477149218321, + 0.004260940011590719, + 0.007951007224619389, + 0.0016527913976460695, + 0.0021601554471999407, + 0.008101355284452438, + 0.0015540223103016615, + 0.010051271878182888, + 0.002317525679245591, + 0.0018798348028212786, + 0.002171883126720786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.026350608095526695 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251660011708736 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251660011708736, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251660011708736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11727962642908096, + 0.10960588604211807, + 0.10967572033405304, + 0.12758855521678925, + 0.08195216208696365, + 0.09667730331420898, + 0.06306586414575577, + 0.10104329884052277, + 0.14112983644008636, + 0.05198178067803383 + ], + "credit_map": [ + 0.01052805781364441, + 0.006652513984590769, + 0.03048715367913246, + 0.008860641159117222, + 0.009058316238224506, + 0.009350469335913658, + 0.010150144807994366, + 0.03219231963157654, + 0.007832696661353111, + 0.009292876347899437, + 0.0331970751285553, + 0.03064880147576332, + 0.007702625822275877, + 0.005758839193731546, + 0.00808623805642128, + 0.007086730096489191, + 0.014393377117812634, + 0.008916807360947132, + 0.010539640672504902, + 0.013058725744485855, + 0.01982264220714569, + 0.018938668072223663, + 0.012408938258886337, + 0.011596908792853355, + 0.01663372665643692, + 0.014505577273666859, + 0.01835772395133972, + 0.012672778218984604, + 0.01377055048942566, + 0.020395400002598763, + 0.012810737825930119, + 0.018442057073116302, + 0.017218472436070442, + 0.008442763239145279, + 0.010461852885782719, + 0.008648456074297428, + 0.008277995511889458, + 0.01359286904335022, + 0.008055590093135834, + 0.00725415488705039, + 0.008514948189258575, + 0.010818341746926308, + 0.009804627858102322, + 0.009536833502352238, + 0.018284950405359268, + 0.013437398709356785, + 0.01132657378911972, + 0.014953629113733768, + 0.007820471189916134, + 0.007247173227369785, + 0.005476599093526602, + 0.006483548320829868, + 0.009523369371891022, + 0.007909676991403103, + 0.010636932216584682, + 0.007968087680637836, + 0.0054082549177110195, + 0.0052595180459320545, + 0.006094429641962051, + 0.02534550242125988, + 0.012451407499611378, + 0.03237204998731613, + 0.00784640945494175, + 0.006265724077820778, + 0.039600685238838196, + 0.018671870231628418, + 0.022630002349615097, + 0.011948350816965103, + 0.015316708944737911, + 0.008289569057524204, + 0.012501955032348633, + 0.012170682661235332, + 0.003168164286762476, + 0.005255849566310644, + 0.013651109300553799, + 0.00302005372941494, + 0.016983389854431152, + 0.003183679888024926, + 0.003101056907325983, + 0.0036184745840728283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.005251660011708736 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762906763702631 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.00762906763702631, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762906763702631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12610319256782532, + 0.21482928097248077, + 0.08272383362054825, + 0.05059469863772392, + 0.0787678137421608, + 0.11114498227834702, + 0.08565639704465866, + 0.127566397190094, + 0.06437310576438904, + 0.058240339159965515 + ], + "credit_map": [ + 0.016842802986502647, + 0.012645634822547436, + 0.01992093212902546, + 0.012809587642550468, + 0.016620369628071785, + 0.015190019272267818, + 0.019501222297549248, + 0.012572618201375008, + 0.014079711399972439, + 0.01911926455795765, + 0.059093985706567764, + 0.06485150009393692, + 0.012537840753793716, + 0.014305994845926762, + 0.01616649143397808, + 0.014674479141831398, + 0.007721890229731798, + 0.006649375427514315, + 0.009541736915707588, + 0.012986642308533192, + 0.008070336654782295, + 0.009794006124138832, + 0.023152967914938927, + 0.004806877579540014, + 0.00962821301072836, + 0.004954629577696323, + 0.0070731425657868385, + 0.005933632142841816, + 0.007455445826053619, + 0.00653521902859211, + 0.004431613255292177, + 0.004582798574119806, + 0.015345667488873005, + 0.010732138529419899, + 0.012499797157943249, + 0.00799549836665392, + 0.00800918135792017, + 0.007459576707333326, + 0.00961859617382288, + 0.007107357960194349, + 0.027556082233786583, + 0.012958040460944176, + 0.011011605150997639, + 0.010234097950160503, + 0.012475309893488884, + 0.013538476079702377, + 0.013080714270472527, + 0.010290647856891155, + 0.008093236945569515, + 0.009160025976598263, + 0.007644629571586847, + 0.012313827872276306, + 0.017620546743273735, + 0.01270484272390604, + 0.009960377588868141, + 0.008158898912370205, + 0.008047065697610378, + 0.008915574289858341, + 0.007873715832829475, + 0.03595099225640297, + 0.009371612221002579, + 0.03970164805650711, + 0.009781650267541409, + 0.007924131117761135, + 0.009499171748757362, + 0.009977461770176888, + 0.01012900099158287, + 0.008503190241754055, + 0.00635371170938015, + 0.004895468708127737, + 0.007352021988481283, + 0.007663073483854532, + 0.005791564006358385, + 0.006532825063914061, + 0.012451983988285065, + 0.004331931937485933, + 0.01131409127265215, + 0.007205861154943705, + 0.005735182669013739, + 0.004876894410699606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.00762906763702631 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro lens begins, then slowly and steadily pulls back, capturing the dynamics of countless mosaic tile fragments, shimmering with an inherent warmth. Composed primarily of ceramic with subtle brushstroke textures and translucent stained glass, these fragments display hues of soft cream, warm ochre, sky-like pale blue, and forest-like emerald green, each possessing imperfect, hand-cut edges. Against a deep, blurred background, they seem pulled by an invisible force, gracefully floating and swirling through the air, gently colliding with one another. Each contact releases a fleeting burst of golden dust accompanied by a crisp, melodious tinkling sound. As the camera continues to zoom out, the fragments begin to assemble in an orderly fashion, like pieces of a living puzzle, quickly and precisely finding their place, fitting seamlessly into place, and gradually piecing together a vast, fantastical mural. Eventually, the image stabilized, revealing the mural in its entirety – depicting floating islands, winding rivers, and strange yet friendly forest creatures. The entire work is imbued with the distinctive hand-painted texture of Studio Ghibli, with rich and soft colors, and delicate light and shadow, creating an atmosphere that is dreamy, tranquil, and full of the miracle of life, like a vivid original animation painting from the hands of a master.", + "seed": 11439, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_6/index_1434_seed_11439_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.040925823897123334, + "vlm_logprob": -0.040925823897123334, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.040925823897123334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019271519035100937 + ], + "opposite_logprob_tokens": [ + -4.144271373748779 + ], + "logprob_reward": -0.019271519035100937, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019271519035100937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3279038369655609, + 0.10184311121702194, + 0.11071404069662094, + 0.1053079143166542, + 0.05301843583583832, + 0.1584344059228897, + 0.036581479012966156, + 0.032480329275131226, + 0.022183936089277267, + 0.05153246596455574 + ], + "credit_map": [ + 0.025443963706493378, + 0.015591474249958992, + 0.09711163491010666, + 0.031604450196027756, + 0.101753368973732, + 0.017460402101278305, + 0.02210746705532074, + 0.0168310534209013, + 0.014039904810488224, + 0.01264264527708292, + 0.01256081648170948, + 0.015422556549310684, + 0.010182062163949013, + 0.011112152598798275, + 0.015431871637701988, + 0.01045110821723938, + 0.012039653025567532, + 0.014145754277706146, + 0.012067235074937344, + 0.011525039561092854, + 0.013497665524482727, + 0.01218689326196909, + 0.016968859359622, + 0.018282942473888397, + 0.01212602574378252, + 0.008940141648054123, + 0.014535623602569103, + 0.015219131484627724, + 0.007999220862984657, + 0.014299686998128891, + 0.01501087099313736, + 0.017177211120724678, + 0.00536113977432251, + 0.005936589557677507, + 0.006953046657145023, + 0.007460309658199549, + 0.005600673612207174, + 0.00397545425221324, + 0.010834433138370514, + 0.00689679104834795, + 0.01851671189069748, + 0.015054898336529732, + 0.014311123639345169, + 0.025730255991220474, + 0.017170069739222527, + 0.014562053605914116, + 0.018402479588985443, + 0.034686822444200516, + 0.005048748571425676, + 0.003940594848245382, + 0.0033365266863256693, + 0.003823427949100733, + 0.006989264860749245, + 0.0036529735662043095, + 0.004988224245607853, + 0.004801715724170208, + 0.00510890269652009, + 0.003703951369971037, + 0.0031973766162991524, + 0.003186065936461091, + 0.0054991720244288445, + 0.003896308597177267, + 0.00451003247871995, + 0.0033785186242312193, + 0.0034437126014381647, + 0.0026149298064410686, + 0.0032239030115306377, + 0.002899473998695612, + 0.002750006038695574, + 0.0026785314548760653, + 0.002734704874455929, + 0.0018386724404990673, + 0.007083073724061251, + 0.010233868844807148, + 0.011296705342829227, + 0.007061767857521772, + 0.004079944454133511, + 0.004321792628616095, + 0.003636130364611745, + 0.0038191843777894974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.019271519035100937 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160912424325943 + ], + "opposite_logprob_tokens": [ + -4.141091346740723 + ], + "logprob_reward": -0.0160912424325943, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160912424325943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3776555359363556, + 0.038549043238162994, + 0.10822442173957825, + 0.03633885830640793, + 0.14641320705413818, + 0.11386817693710327, + 0.041098251938819885, + 0.03514491021633148, + 0.058384355157613754, + 0.04432322829961777 + ], + "credit_map": [ + 0.03072139248251915, + 0.015969960018992424, + 0.13000038266181946, + 0.01904291845858097, + 0.13296028971672058, + 0.020937416702508926, + 0.014106852933764458, + 0.013916329480707645, + 0.004650735761970282, + 0.0033794178161770105, + 0.005962892435491085, + 0.006141840945929289, + 0.004674087278544903, + 0.0032128612510859966, + 0.005360690411180258, + 0.0051665171049535275, + 0.010205901227891445, + 0.017732838168740273, + 0.014812770299613476, + 0.033193282783031464, + 0.008006121963262558, + 0.005643970798701048, + 0.00945264007896185, + 0.009176889434456825, + 0.005213751923292875, + 0.003843844635412097, + 0.005813872441649437, + 0.004270768724381924, + 0.0034440739545971155, + 0.005300660151988268, + 0.0035617591347545385, + 0.004890128970146179, + 0.011775427497923374, + 0.017224406823515892, + 0.012046506628394127, + 0.028371529653668404, + 0.018168902024626732, + 0.020586330443620682, + 0.01871735043823719, + 0.019522758200764656, + 0.037854358553886414, + 0.00879662111401558, + 0.006387816276401281, + 0.007973834872245789, + 0.01196074765175581, + 0.006970079615712166, + 0.007101109717041254, + 0.026823611930012703, + 0.005047207232564688, + 0.003874303540214896, + 0.003207405796274543, + 0.005320499185472727, + 0.0056635295040905476, + 0.005698853638023138, + 0.0068586478009819984, + 0.005427805241197348, + 0.004864031448960304, + 0.004918474704027176, + 0.004483325872570276, + 0.004115062300115824, + 0.0043975189328193665, + 0.003433047793805599, + 0.003907805308699608, + 0.005025642458349466, + 0.008585423231124878, + 0.00666840560734272, + 0.007550377864390612, + 0.007986202836036682, + 0.006854397244751453, + 0.007923968136310577, + 0.006250915117561817, + 0.006564664654433727, + 0.006648237816989422, + 0.0058956746943295, + 0.008677329868078232, + 0.006766427308320999, + 0.0043704272247850895, + 0.004954149015247822, + 0.0033573638647794724, + 0.0036536147817969322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0160912424325943 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620497427880764 + ], + "opposite_logprob_tokens": [ + -2.812049627304077 + ], + "logprob_reward": -0.0620497427880764, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620497427880764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16468696296215057, + 0.1390439122915268, + 0.11027054488658905, + 0.13378320634365082, + 0.08808475732803345, + 0.1916736215353012, + 0.054062966257333755, + 0.030276628211140633, + 0.03273472562432289, + 0.055382732301950455 + ], + "credit_map": [ + 0.01682482287287712, + 0.01873433217406273, + 0.023741589859128, + 0.025335388258099556, + 0.021120144054293633, + 0.02417081966996193, + 0.0178452767431736, + 0.01691458187997341, + 0.014600696973502636, + 0.014055758714675903, + 0.015866370871663094, + 0.020642712712287903, + 0.016887851059436798, + 0.01681256853044033, + 0.021187489852309227, + 0.018990470096468925, + 0.009942984208464622, + 0.011618872173130512, + 0.015129292383790016, + 0.015937980264425278, + 0.013461805880069733, + 0.01400911808013916, + 0.017472948879003525, + 0.012697549536824226, + 0.017525142058730125, + 0.011934293434023857, + 0.016355330124497414, + 0.024151761084794998, + 0.01117396354675293, + 0.015492988750338554, + 0.016483206301927567, + 0.02066650800406933, + 0.009526452980935574, + 0.013935490511357784, + 0.006104551255702972, + 0.012544019147753716, + 0.014239453710615635, + 0.00811429601162672, + 0.01273764856159687, + 0.010882842354476452, + 0.029065972194075584, + 0.022596323862671852, + 0.016607273370027542, + 0.027043715119361877, + 0.026018710806965828, + 0.02090190164744854, + 0.021617161110043526, + 0.027822567149996758, + 0.007167640142142773, + 0.004721909761428833, + 0.005544615909457207, + 0.004086973611265421, + 0.0074242097325623035, + 0.007569716311991215, + 0.009043963626027107, + 0.008503938093781471, + 0.004061846062541008, + 0.004238966852426529, + 0.0031098576728254557, + 0.003160306951031089, + 0.004506823141127825, + 0.0037535002920776606, + 0.0036441052798181772, + 0.0038012214936316013, + 0.0049569918774068356, + 0.0049631018191576, + 0.005547501612454653, + 0.003949996083974838, + 0.0033855035435408354, + 0.0038452311418950558, + 0.0033497430849820375, + 0.002736659487709403, + 0.008594210259616375, + 0.010581258684396744, + 0.012201383709907532, + 0.007608441170305014, + 0.00413704477250576, + 0.00460422458127141, + 0.003837978933006525, + 0.0038181901909410954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.0620497427880764 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817922666668892 + ], + "opposite_logprob_tokens": [ + -4.018179416656494 + ], + "logprob_reward": -0.01817922666668892, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817922666668892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4378836452960968, + 0.0750202089548111, + 0.07313822209835052, + 0.06725044548511505, + 0.05643375962972641, + 0.08976759761571884, + 0.04407887160778046, + 0.07375948131084442, + 0.03981073200702667, + 0.04285705089569092 + ], + "credit_map": [ + 0.035065796226263046, + 0.015700040385127068, + 0.1367114782333374, + 0.03384649008512497, + 0.14200076460838318, + 0.02490745671093464, + 0.029673468321561813, + 0.01997818797826767, + 0.0073900590650737286, + 0.006580850575119257, + 0.00817288737744093, + 0.012456977739930153, + 0.00854527484625578, + 0.007114964071661234, + 0.012536520138382912, + 0.01222267933189869, + 0.006879676133394241, + 0.008877032436430454, + 0.007102032192051411, + 0.013891352340579033, + 0.006168412510305643, + 0.004519912414252758, + 0.013999130576848984, + 0.011700676754117012, + 0.007958642207086086, + 0.006846046540886164, + 0.009205369278788567, + 0.009611863642930984, + 0.007110318168997765, + 0.0072057778015732765, + 0.00940573588013649, + 0.009906702674925327, + 0.00628978107124567, + 0.005814997013658285, + 0.004922545049339533, + 0.009781328961253166, + 0.011930533684790134, + 0.005040364805608988, + 0.00630609504878521, + 0.006348117254674435, + 0.022407997399568558, + 0.006528577301651239, + 0.005629310384392738, + 0.0074020572938025, + 0.007150058634579182, + 0.0068909176625311375, + 0.011324568651616573, + 0.022434113547205925, + 0.006750950589776039, + 0.005301772151142359, + 0.004556325729936361, + 0.005738856270909309, + 0.004876971710473299, + 0.004931730683892965, + 0.005413264501839876, + 0.006509003229439259, + 0.009959569200873375, + 0.010198401287198067, + 0.007626872975379229, + 0.008036828599870205, + 0.01052401028573513, + 0.011978189460933208, + 0.007931854575872421, + 0.0075037661008536816, + 0.004442904144525528, + 0.009299738332629204, + 0.007130708545446396, + 0.004461641889065504, + 0.0034561953507363796, + 0.004379212390631437, + 0.0034251827746629715, + 0.0032151511404663324, + 0.005749756004661322, + 0.007250430062413216, + 0.007020330522209406, + 0.005787491332739592, + 0.004365822300314903, + 0.004320309963077307, + 0.00425786804407835, + 0.004105048719793558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01817922666668892 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903738856315613 + ], + "opposite_logprob_tokens": [ + -2.4640374183654785 + ], + "logprob_reward": -0.08903738856315613, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903738856315613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1310224086046219, + 0.06780170649290085, + 0.09608664363622665, + 0.04831792786717415, + 0.14378516376018524, + 0.23440931737422943, + 0.05131642520427704, + 0.04718963801860809, + 0.07557862251996994, + 0.10449209809303284 + ], + "credit_map": [ + 0.011083927936851978, + 0.008453256450593472, + 0.038102634251117706, + 0.01101929135620594, + 0.0358831025660038, + 0.009069925174117088, + 0.009369530715048313, + 0.008040743879973888, + 0.007388612255454063, + 0.006360246799886227, + 0.009834440425038338, + 0.00953750591725111, + 0.006387150380760431, + 0.00795076321810484, + 0.010542572475969791, + 0.009800414554774761, + 0.009887106716632843, + 0.012287330813705921, + 0.01277188677340746, + 0.008825589902698994, + 0.011755888350307941, + 0.01429942436516285, + 0.0129399374127388, + 0.013319478370249271, + 0.0050875041633844376, + 0.003926340024918318, + 0.006518634967505932, + 0.007039905060082674, + 0.0035209935158491135, + 0.007288243621587753, + 0.007052312139421701, + 0.007883995771408081, + 0.016774676740169525, + 0.01745174452662468, + 0.013519593514502048, + 0.02544333040714264, + 0.019964570179581642, + 0.01696464978158474, + 0.016580943018198013, + 0.01708565466105938, + 0.06037292256951332, + 0.015444759279489517, + 0.010980810970067978, + 0.011651769280433655, + 0.015475199557840824, + 0.014780435711145401, + 0.020544983446598053, + 0.08515844494104385, + 0.007892483845353127, + 0.004917346406728029, + 0.005162754561752081, + 0.006507847458124161, + 0.007002988364547491, + 0.0049426318146288395, + 0.008792351931333542, + 0.00609801709651947, + 0.00648767314851284, + 0.007770756259560585, + 0.005182513501495123, + 0.004546795506030321, + 0.006601210217922926, + 0.005330745596438646, + 0.005505151115357876, + 0.005764791741967201, + 0.009983005002140999, + 0.017450373619794846, + 0.017726201564073563, + 0.007922613061964512, + 0.0065363566391170025, + 0.005806796718388796, + 0.0055039552971720695, + 0.004649318754673004, + 0.015968194231390953, + 0.015907645225524902, + 0.02248171530663967, + 0.012189999222755432, + 0.011669146828353405, + 0.00774855213239789, + 0.010538588277995586, + 0.00798825267702341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.08903738856315613 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro lens begins, then slowly and steadily pulls back, capturing the dynamics of countless mosaic tile fragments, shimmering with an inherent warmth. Composed primarily of ceramic with subtle brushstroke textures and translucent stained glass, these fragments display hues of soft cream, warm ochre, sky-like pale blue, and forest-like emerald green, each possessing imperfect, hand-cut edges. Against a deep, blurred background, they seem pulled by an invisible force, gracefully floating and swirling through the air, gently colliding with one another. Each contact releases a fleeting burst of golden dust accompanied by a crisp, melodious tinkling sound. As the camera continues to zoom out, the fragments begin to assemble in an orderly fashion, like pieces of a living puzzle, quickly and precisely finding their place, fitting seamlessly into place, and gradually piecing together a vast, fantastical mural. Eventually, the image stabilized, revealing the mural in its entirety – depicting floating islands, winding rivers, and strange yet friendly forest creatures. The entire work is imbued with the distinctive hand-painted texture of Studio Ghibli, with rich and soft colors, and delicate light and shadow, creating an atmosphere that is dreamy, tranquil, and full of the miracle of life, like a vivid original animation painting from the hands of a master.", + "seed": 11440, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_6/index_1434_seed_11440_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02023233501240611, + "vlm_logprob": -0.02023233501240611, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.02023233501240611, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01701647974550724 + ], + "opposite_logprob_tokens": [ + -4.267016410827637 + ], + "logprob_reward": -0.01701647974550724, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01701647974550724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2085430771112442, + 0.23179244995117188, + 0.10508494824171066, + 0.06740209460258484, + 0.01797897182404995, + 0.0495416484773159, + 0.09386223554611206, + 0.11954119056463242, + 0.06567640602588654, + 0.0405769981443882 + ], + "credit_map": [ + 0.02297721430659294, + 0.020837822929024696, + 0.03342575579881668, + 0.039587654173374176, + 0.023472202941775322, + 0.021772488951683044, + 0.027753939852118492, + 0.018716001883149147, + 0.021709246560931206, + 0.02287277579307556, + 0.03691296651959419, + 0.03024946339428425, + 0.028804410248994827, + 0.028310609981417656, + 0.025427356362342834, + 0.037505608052015305, + 0.010480465367436409, + 0.0118017066270113, + 0.012517228722572327, + 0.012416098266839981, + 0.01680706813931465, + 0.008311199024319649, + 0.015253493562340736, + 0.01749768853187561, + 0.010048896074295044, + 0.005200206767767668, + 0.011140705086290836, + 0.009947349317371845, + 0.004209603648632765, + 0.004651852883398533, + 0.009849464520812035, + 0.012354015372693539, + 0.002256629057228565, + 0.002849254058673978, + 0.002135993679985404, + 0.0019333348609507084, + 0.001955765299499035, + 0.0022869184613227844, + 0.0020387584809213877, + 0.0025223190896213055, + 0.0058649140410125256, + 0.006156574003398418, + 0.0056761810556054115, + 0.007510128431022167, + 0.007840781472623348, + 0.005664102267473936, + 0.007254834286868572, + 0.0035741296596825123, + 0.008678792975842953, + 0.00880784448236227, + 0.021108431741595268, + 0.0068996623158454895, + 0.028849339112639427, + 0.005423923023045063, + 0.008311927318572998, + 0.0057823131792247295, + 0.020206289365887642, + 0.02100016176700592, + 0.014911547303199768, + 0.014652957208454609, + 0.01913272961974144, + 0.010279531590640545, + 0.010355236940085888, + 0.009002739563584328, + 0.011506968177855015, + 0.00588851235806942, + 0.008298209868371487, + 0.0074574146419763565, + 0.008652270771563053, + 0.006329103838652372, + 0.008198701776564121, + 0.0093452213332057, + 0.006064067594707012, + 0.007255961187183857, + 0.005259710829705, + 0.004099405370652676, + 0.004781673196703196, + 0.002372899791225791, + 0.0048174066469073296, + 0.005925874691456556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.01701647974550724 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012557603418827057 + ], + "opposite_logprob_tokens": [ + -4.387557506561279 + ], + "logprob_reward": -0.012557603418827057, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012557603418827057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28905391693115234, + 0.057960763573646545, + 0.13094289600849152, + 0.0843690037727356, + 0.03613775223493576, + 0.11826244741678238, + 0.07841542363166809, + 0.07926635444164276, + 0.06234624236822128, + 0.06324520707130432 + ], + "credit_map": [ + 0.024771960452198982, + 0.02804887481033802, + 0.035076986998319626, + 0.053274523466825485, + 0.035726115107536316, + 0.044198550283908844, + 0.04155614972114563, + 0.026400741189718246, + 0.006111818831413984, + 0.005298491567373276, + 0.009298133663833141, + 0.00998990423977375, + 0.004682712722569704, + 0.004875941667705774, + 0.006011249963194132, + 0.011692512780427933, + 0.01671423949301243, + 0.015949448570609093, + 0.01779314875602722, + 0.010247988626360893, + 0.021551620215177536, + 0.01206017006188631, + 0.015662362799048424, + 0.020963912829756737, + 0.00920735765248537, + 0.00485233124345541, + 0.023226769641041756, + 0.00725160539150238, + 0.0054512303322553635, + 0.0049964310601353645, + 0.00950599554926157, + 0.019877279177308083, + 0.0035755173303186893, + 0.005249870475381613, + 0.00584853021427989, + 0.005421651527285576, + 0.004038447048515081, + 0.004143796395510435, + 0.004724284168332815, + 0.003135654842481017, + 0.012612706981599331, + 0.014631301164627075, + 0.0114821158349514, + 0.014236400835216045, + 0.013688757084310055, + 0.017243847250938416, + 0.026962392032146454, + 0.007404931355267763, + 0.008083892054855824, + 0.011415104381740093, + 0.010013519786298275, + 0.009294281713664532, + 0.011111299507319927, + 0.010430045425891876, + 0.008593698963522911, + 0.009473584592342377, + 0.009118867106735706, + 0.012593994848430157, + 0.007236130069941282, + 0.0099010169506073, + 0.010026737116277218, + 0.01253743376582861, + 0.009195721708238125, + 0.008656457997858524, + 0.013089162297546864, + 0.007325439248234034, + 0.008287778124213219, + 0.0074793011881411076, + 0.0052176425233483315, + 0.006678096950054169, + 0.006667755544185638, + 0.007601063698530197, + 0.008588945493102074, + 0.01063357014209032, + 0.008200624957680702, + 0.0053495122119784355, + 0.008140492253005505, + 0.004378680605441332, + 0.006823536939918995, + 0.011129838414490223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.012557603418827057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043057575821876526 + ], + "opposite_logprob_tokens": [ + -3.168057680130005 + ], + "logprob_reward": -0.043057575821876526, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043057575821876526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20442074537277222, + 0.10960731655359268, + 0.1715092658996582, + 0.1298765242099762, + 0.034225791692733765, + 0.10301979631185532, + 0.049108412116765976, + 0.058595359325408936, + 0.07221338152885437, + 0.06742339581251144 + ], + "credit_map": [ + 0.034087762236595154, + 0.02500126324594021, + 0.02770817279815674, + 0.02599131129682064, + 0.02107095532119274, + 0.016243092715740204, + 0.035410549491643906, + 0.018907641991972923, + 0.01001522596925497, + 0.011277454905211926, + 0.018890274688601494, + 0.010948511771857738, + 0.01713501289486885, + 0.012967347167432308, + 0.011246184818446636, + 0.017127305269241333, + 0.016904838383197784, + 0.01839521713554859, + 0.03309759870171547, + 0.018807819113135338, + 0.019888026639819145, + 0.02187010645866394, + 0.018466565757989883, + 0.0240790992975235, + 0.01741076074540615, + 0.011636901646852493, + 0.01634378544986248, + 0.018847595900297165, + 0.01201217994093895, + 0.01690247468650341, + 0.017909344285726547, + 0.0188134778290987, + 0.0038989076856523752, + 0.0059396554715931416, + 0.0038862864021211863, + 0.004788077902048826, + 0.003649474587291479, + 0.004739284515380859, + 0.004104210529476404, + 0.0032198948320001364, + 0.013789336197078228, + 0.013946481980383396, + 0.01169377751648426, + 0.014627935364842415, + 0.012450221925973892, + 0.010076361708343029, + 0.01752094179391861, + 0.008914737962186337, + 0.004952572286128998, + 0.007473220117390156, + 0.007196743041276932, + 0.006389151327311993, + 0.005675800610333681, + 0.006586533971130848, + 0.006174237001687288, + 0.004660156089812517, + 0.00872848741710186, + 0.009222588501870632, + 0.007154952734708786, + 0.00956857018172741, + 0.007375231944024563, + 0.005569729954004288, + 0.006059566512703896, + 0.00491623068228364, + 0.010968644171953201, + 0.006618281360715628, + 0.010073833167552948, + 0.008716253563761711, + 0.010356675833463669, + 0.007862916216254234, + 0.008298142813146114, + 0.009318642318248749, + 0.009310359135270119, + 0.013452477753162384, + 0.007387461606413126, + 0.006935914047062397, + 0.007898567244410515, + 0.004388625267893076, + 0.00891219824552536, + 0.009137789718806744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.043057575821876526 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248814355581999 + ], + "opposite_logprob_tokens": [ + -5.2552490234375 + ], + "logprob_reward": -0.005248814355581999, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248814355581999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2424594610929489, + 0.17215050756931305, + 0.07232790440320969, + 0.04513926804065704, + 0.031162269413471222, + 0.12640546262264252, + 0.08600058406591415, + 0.05535290762782097, + 0.12262888997793198, + 0.04637281969189644 + ], + "credit_map": [ + 0.04533601179718971, + 0.03687535971403122, + 0.03662953898310661, + 0.027421139180660248, + 0.028301192447543144, + 0.01939711719751358, + 0.025595229119062424, + 0.022903865203261375, + 0.018621645867824554, + 0.017466051504015923, + 0.025892119854688644, + 0.022578785195946693, + 0.024353275075554848, + 0.02379951812326908, + 0.020639382302761078, + 0.018799712881445885, + 0.007495139725506306, + 0.007352523971349001, + 0.01716884970664978, + 0.007401736918836832, + 0.00804339163005352, + 0.00896176416426897, + 0.007094979751855135, + 0.008809522725641727, + 0.005304367747157812, + 0.005298682022839785, + 0.0072446661069989204, + 0.005535800941288471, + 0.0033037280663847923, + 0.006099830847233534, + 0.006024838890880346, + 0.006327352486550808, + 0.004983056802302599, + 0.004759511444717646, + 0.003598646027967334, + 0.004166250582784414, + 0.003207004629075527, + 0.0032406162936240435, + 0.003940971102565527, + 0.0032662113662809134, + 0.01003754697740078, + 0.009886560961604118, + 0.014712421223521233, + 0.013764902018010616, + 0.015065641142427921, + 0.032563094049692154, + 0.021372750401496887, + 0.009002549573779106, + 0.00782543420791626, + 0.010142278857529163, + 0.014810564927756786, + 0.011501693166792393, + 0.019772427156567574, + 0.0062992163002491, + 0.007434326224029064, + 0.008214637637138367, + 0.006544844713062048, + 0.012112961150705814, + 0.006419551558792591, + 0.0070335338823497295, + 0.0083690220490098, + 0.004844623152166605, + 0.004972061607986689, + 0.005056307651102543, + 0.023238401859998703, + 0.013042077422142029, + 0.013843595050275326, + 0.012189166620373726, + 0.012935447506606579, + 0.015073849819600582, + 0.015316310338675976, + 0.01699002832174301, + 0.005904301069676876, + 0.006445009261369705, + 0.007115328684449196, + 0.004602552857249975, + 0.005556484218686819, + 0.0033084803726524115, + 0.00533164432272315, + 0.00810901541262865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.005248814355581999 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme macro lens view of mosaic tile fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the fragments display hues of soft cream, warm ochre, pale blue, and emerald green with hand-cut edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the fragments float and swirl through the air before assembling into an orderly mural?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pull back (zoom out) throughout the sequence?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the final assembled mural depict floating islands, winding rivers, and friendly forest creatures in a Studio Ghibli style?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023281201720237732 + ], + "opposite_logprob_tokens": [ + -3.7732810974121094 + ], + "logprob_reward": -0.023281201720237732, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023281201720237732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10203176736831665, + 0.06632301956415176, + 0.07705142349004745, + 0.04018126800656319, + 0.042613908648490906, + 0.3469519019126892, + 0.07521049678325653, + 0.08264989405870438, + 0.09789156168699265, + 0.0690946951508522 + ], + "credit_map": [ + 0.008726698346436024, + 0.010229918174445629, + 0.015556253492832184, + 0.01707618311047554, + 0.01244660560041666, + 0.014487674459815025, + 0.01313620526343584, + 0.010372243821620941, + 0.008262925781309605, + 0.006199005991220474, + 0.009993545711040497, + 0.00824026856571436, + 0.008704141713678837, + 0.008824760094285011, + 0.007232570089399815, + 0.008865801617503166, + 0.008834854699671268, + 0.020127085968852043, + 0.010762990452349186, + 0.008078639395534992, + 0.006767364218831062, + 0.005074616055935621, + 0.00847832765430212, + 0.008927551098167896, + 0.005056053400039673, + 0.0056709288619458675, + 0.005182561464607716, + 0.007057443261146545, + 0.003732808865606785, + 0.0031996937468647957, + 0.005770320538431406, + 0.004511462524533272, + 0.006433255970478058, + 0.005850961897522211, + 0.0058479756116867065, + 0.005159825086593628, + 0.0037804869934916496, + 0.005084587726742029, + 0.006263416726142168, + 0.004193403758108616, + 0.023530716076493263, + 0.02065538614988327, + 0.023077668622136116, + 0.023062486201524734, + 0.02019345946609974, + 0.12373951822519302, + 0.09623763710260391, + 0.0164550319314003, + 0.007374189794063568, + 0.006355369929224253, + 0.015587322413921356, + 0.00727818813174963, + 0.017263103276491165, + 0.008461049757897854, + 0.006239700131118298, + 0.006651586387306452, + 0.010031734593212605, + 0.017491217702627182, + 0.009176857769489288, + 0.011506527662277222, + 0.010354367084801197, + 0.006832296960055828, + 0.00992416962981224, + 0.007332731503993273, + 0.014155533164739609, + 0.009271116927266121, + 0.015674740076065063, + 0.013379177078604698, + 0.00951355416327715, + 0.009565153159201145, + 0.013163236901164055, + 0.013169054873287678, + 0.00957382470369339, + 0.010791853070259094, + 0.009496305137872696, + 0.006745076272636652, + 0.008541291579604149, + 0.004249869380146265, + 0.010347176343202591, + 0.009349304251372814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.023281201720237732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_7/index_tensor([9149, 9149])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_7/index_tensor([9149, 9149])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a3d4c8e33af4e2f1454053b93fa71fbbd3e1dd35 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_7/index_tensor([9149, 9149])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, low-angle, static shot captures a vibrant red plastic bottle cap, precariously balanced on its ridged rim. It rests on the textured surface of a dark, weathered wooden board. Warm late afternoon sunlight slants across the scene, casting a striking, dramatic light on the deep veins and tiny imperfections. The cap begins to tremble, a series of tiny, high-frequency vibrations that cause its long, twisted shadow to tremble nervously across the wood. Then, it begins a slow, agonizing struggle, teetering back and forth, cautiously trying to straighten itself, its smooth plastic form catching the golden light with each desperate lurch. After one last, prolonged lurch, its struggle against gravity fails; it collapses decisively, landing on its flat top with a soft, sharp click and finally resting in defeat, creating an atmosphere of quiet, focused, miniature, dramatic defeat.", + "seed": 30243, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_7/index_9149_seed_30243_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02764151096343994, + "vlm_logprob": -0.02764151096343994, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.02764151096343994, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016524244099855423 + ], + "opposite_logprob_tokens": [ + -5.016524314880371 + ], + "logprob_reward": -0.016524244099855423, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.016524244099855423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14617405831813812, + 0.07160872220993042, + 0.2776852250099182, + 0.018475400283932686, + 0.26090121269226074, + 0.033513206988573074, + 0.02355385012924671, + 0.13088268041610718, + 0.025706253945827484, + 0.011499309912323952 + ], + "credit_map": [ + 0.02601679414510727, + 0.022493399679660797, + 0.01515153981745243, + 0.014449935406446457, + 0.0204149279743433, + 0.011256762780249119, + 0.021289072930812836, + 0.015101633965969086, + 0.008864667266607285, + 0.012491961009800434, + 0.008844575844705105, + 0.006781895644962788, + 0.007491628639400005, + 0.010116280056536198, + 0.007371208164840937, + 0.009646505117416382, + 0.029234973713755608, + 0.031104518100619316, + 0.03438711166381836, + 0.0555771104991436, + 0.03291267529129982, + 0.04261549562215805, + 0.032037824392318726, + 0.01981549896299839, + 0.0048279231414198875, + 0.0026083539705723524, + 0.0021699825301766396, + 0.0011999048292636871, + 0.0022030319087207317, + 0.0026596621610224247, + 0.0013631569454446435, + 0.0014433846808969975, + 0.03186514228582382, + 0.023410392925143242, + 0.023451825603842735, + 0.0478760227560997, + 0.04542945325374603, + 0.03185267373919487, + 0.03394528850913048, + 0.023070422932505608, + 0.003960778005421162, + 0.003114553401246667, + 0.003751712152734399, + 0.006653417833149433, + 0.00316377030685544, + 0.004478780552744865, + 0.0051248278468847275, + 0.0032653682865202427, + 0.003311797743663192, + 0.002309134230017662, + 0.004368550144135952, + 0.0024195590522140265, + 0.002211779123172164, + 0.003489289665594697, + 0.0026886421255767345, + 0.0027550996746867895, + 0.03707107529044151, + 0.008623762987554073, + 0.017679598182439804, + 0.009606640785932541, + 0.010015327483415604, + 0.02200625091791153, + 0.010348662734031677, + 0.015531367622315884, + 0.006305073853582144, + 0.002258188324049115, + 0.00376557232812047, + 0.002323709661141038, + 0.0020648047793656588, + 0.0042917788960039616, + 0.0017476241337135434, + 0.0029495027847588062, + 0.0019270854536443949, + 0.0013726763427257538, + 0.0007511202711611986, + 0.0017089865868911147, + 0.0018056053668260574, + 0.0015708880964666605, + 0.0013807883951812983, + 0.0009821585845202208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.016524244099855423 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371965512633324 + ], + "opposite_logprob_tokens": [ + -3.408719539642334 + ], + "logprob_reward": -0.03371965512633324, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371965512633324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12171386182308197, + 0.11636342108249664, + 0.3947955369949341, + 0.04640434309840202, + 0.08290459960699081, + 0.0446581207215786, + 0.03651389107108116, + 0.0565037876367569, + 0.03073013573884964, + 0.06941237300634384 + ], + "credit_map": [ + 0.01782315969467163, + 0.019026612862944603, + 0.013835800811648369, + 0.010213462635874748, + 0.020773516967892647, + 0.021509146317839622, + 0.011461409740149975, + 0.007070728112012148, + 0.014935089275240898, + 0.011895187199115753, + 0.012591834180057049, + 0.008877217769622803, + 0.008997696451842785, + 0.025088543072342873, + 0.01322147622704506, + 0.020756365731358528, + 0.04118817672133446, + 0.04288848116993904, + 0.04885837435722351, + 0.08026453852653503, + 0.046079959720373154, + 0.061258960515260696, + 0.0457310825586319, + 0.028525929898023605, + 0.008497873321175575, + 0.004459545481950045, + 0.005910283420234919, + 0.006797065958380699, + 0.004431456793099642, + 0.006540075875818729, + 0.005371273495256901, + 0.004396763164550066, + 0.01182112842798233, + 0.00830022245645523, + 0.0063822162337601185, + 0.00994183961302042, + 0.017291534692049026, + 0.010170361027121544, + 0.010045241564512253, + 0.008952044881880283, + 0.008920968510210514, + 0.005175093654543161, + 0.004467757884413004, + 0.00439232075586915, + 0.005411175545305014, + 0.006657468620687723, + 0.005584006197750568, + 0.004049325827509165, + 0.006491037551313639, + 0.0037360081914812326, + 0.00426701782271266, + 0.004056536592543125, + 0.003596557304263115, + 0.005020453594624996, + 0.004961743485182524, + 0.004384530242532492, + 0.015138349495828152, + 0.003922420088201761, + 0.008181222714483738, + 0.003903223667293787, + 0.0046784235164523125, + 0.00957626011222601, + 0.004152161534875631, + 0.0069517167285084724, + 0.007863719947636127, + 0.0021046644542366266, + 0.00447411322966218, + 0.003053162945434451, + 0.002075018361210823, + 0.005444152746349573, + 0.0023265606723725796, + 0.003388739423826337, + 0.013166715390980244, + 0.007522631902247667, + 0.008142001926898956, + 0.008243774063885212, + 0.006753467954695225, + 0.010418195277452469, + 0.00734939007088542, + 0.007816186174750328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.03371965512633324 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031989850103855133 + ], + "opposite_logprob_tokens": [ + -5.753199100494385 + ], + "logprob_reward": -0.0031989850103855133, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031989850103855133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06658392399549484, + 0.057853542268276215, + 0.5216181874275208, + 0.02117682807147503, + 0.045638199895620346, + 0.019194936379790306, + 0.12383468449115753, + 0.04533623903989792, + 0.05601660534739494, + 0.0427468903362751 + ], + "credit_map": [ + 0.01400621235370636, + 0.006737401708960533, + 0.012053629383444786, + 0.0055183228105306625, + 0.005106184631586075, + 0.007424516137689352, + 0.007646771147847176, + 0.008090890012681484, + 0.00753114465624094, + 0.007677510846406221, + 0.006492186803370714, + 0.006581434048712254, + 0.005891448352485895, + 0.008689654059708118, + 0.006203068420290947, + 0.008787094615399837, + 0.05536330118775368, + 0.05736711621284485, + 0.0653616338968277, + 0.10270912200212479, + 0.061890847980976105, + 0.08079475909471512, + 0.06043766066431999, + 0.03769376873970032, + 0.004518807400017977, + 0.0020818780176341534, + 0.0020943842828273773, + 0.002017183927819133, + 0.003378289984539151, + 0.002535810461267829, + 0.0020918447989970446, + 0.002458629198372364, + 0.0068960306234657764, + 0.004835437051951885, + 0.0032811767887324095, + 0.005125507712364197, + 0.010040178894996643, + 0.0056256623938679695, + 0.005422414280474186, + 0.004411793313920498, + 0.0026626810431480408, + 0.0021922909654676914, + 0.0023782162461429834, + 0.0021641121711581945, + 0.0025039860047399998, + 0.002954758470878005, + 0.001956145977601409, + 0.0023827459663152695, + 0.015758858993649483, + 0.010852165520191193, + 0.022519951686263084, + 0.011735180392861366, + 0.010200222954154015, + 0.018597979098558426, + 0.014560631476342678, + 0.019609691575169563, + 0.011198673397302628, + 0.002808766672387719, + 0.00565035454928875, + 0.004519064910709858, + 0.003729848423972726, + 0.007382106967270374, + 0.0036724426317960024, + 0.006374978926032782, + 0.014679552055895329, + 0.003595904679968953, + 0.008520355448126793, + 0.005965617019683123, + 0.0038392292335629463, + 0.009031186811625957, + 0.004850948695093393, + 0.005533811170607805, + 0.007904428988695145, + 0.004342057276517153, + 0.005265577230602503, + 0.005586371757090092, + 0.0033613606356084347, + 0.005986216943711042, + 0.005266334395855665, + 0.00503454077988863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.0031989850103855133 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549626462161541 + ], + "opposite_logprob_tokens": [ + -2.929962635040283 + ], + "logprob_reward": -0.0549626462161541, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549626462161541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1402307152748108, + 0.26752376556396484, + 0.1695002317428589, + 0.0486329160630703, + 0.050746168941259384, + 0.03370589390397072, + 0.018973100930452347, + 0.16279837489128113, + 0.06624186784029007, + 0.041646987199783325 + ], + "credit_map": [ + 0.03816729038953781, + 0.01156260073184967, + 0.023545609787106514, + 0.006166548002511263, + 0.011230520904064178, + 0.013177934102714062, + 0.01079564355313778, + 0.025584563612937927, + 0.03885151073336601, + 0.019848138093948364, + 0.02071836031973362, + 0.016626156866550446, + 0.01990313082933426, + 0.07196733355522156, + 0.03439783304929733, + 0.0452113039791584, + 0.018299339339137077, + 0.017568783834576607, + 0.01995656080543995, + 0.0334852896630764, + 0.020586559548974037, + 0.026358336210250854, + 0.02050061896443367, + 0.012744737789034843, + 0.009661595337092876, + 0.005107693839818239, + 0.0050700632855296135, + 0.006615542806684971, + 0.005978744011372328, + 0.006328091491013765, + 0.0048838285729289055, + 0.004987354390323162, + 0.010814903303980827, + 0.0040399557910859585, + 0.005583605263382196, + 0.00802382081747055, + 0.006112128496170044, + 0.007201039232313633, + 0.004713156260550022, + 0.004257565829902887, + 0.007277256343513727, + 0.002899657702073455, + 0.00392888905480504, + 0.004618579056113958, + 0.0034394171088933945, + 0.005257660988718271, + 0.0031507874373346567, + 0.0031336455140262842, + 0.0035996660590171814, + 0.001814215094782412, + 0.001896356581710279, + 0.0019483324140310287, + 0.0020888301078230143, + 0.0025937764439731836, + 0.0024478305131196976, + 0.0025840934831649065, + 0.04626387357711792, + 0.010738268494606018, + 0.022202856838703156, + 0.012063981965184212, + 0.012642721645534039, + 0.027562491595745087, + 0.011793959885835648, + 0.019530214369297028, + 0.017414890229701996, + 0.004226630553603172, + 0.009156567975878716, + 0.006447070278227329, + 0.004597580060362816, + 0.011081838980317116, + 0.006190640851855278, + 0.00712664844468236, + 0.007508911192417145, + 0.004196363501250744, + 0.004850332159548998, + 0.006368433590978384, + 0.0035073563922196627, + 0.00566441286355257, + 0.005111194681376219, + 0.004439981188625097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.0549626462161541 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029802024364471436 + ], + "opposite_logprob_tokens": [ + -3.529802083969116 + ], + "logprob_reward": -0.029802024364471436, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029802024364471436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08076566457748413, + 0.3052802085876465, + 0.15161097049713135, + 0.07356647402048111, + 0.04969706013798714, + 0.09116396307945251, + 0.03478100895881653, + 0.09014804661273956, + 0.02954016625881195, + 0.09344643354415894 + ], + "credit_map": [ + 0.015355507843196392, + 0.006941727362573147, + 0.012079264968633652, + 0.01173704769462347, + 0.007752492092549801, + 0.011430240236222744, + 0.007077442482113838, + 0.00839193444699049, + 0.04194020852446556, + 0.028036199510097504, + 0.02397569827735424, + 0.01932637207210064, + 0.024481339380145073, + 0.0774964690208435, + 0.038589611649513245, + 0.051434289664030075, + 0.01722753420472145, + 0.015615263022482395, + 0.017296694219112396, + 0.029572123661637306, + 0.017583969980478287, + 0.02380147948861122, + 0.018761806190013885, + 0.011752096936106682, + 0.020579518750309944, + 0.00848663691431284, + 0.008387970738112926, + 0.004743211902678013, + 0.00940266065299511, + 0.010116910561919212, + 0.005628927145153284, + 0.006220637354999781, + 0.0071354457177221775, + 0.007183745503425598, + 0.005151722114533186, + 0.006180234719067812, + 0.007611650042235851, + 0.005940334405750036, + 0.005653697066009045, + 0.004840231966227293, + 0.01957169733941555, + 0.008061048574745655, + 0.011090818792581558, + 0.01146437507122755, + 0.00897916592657566, + 0.013792618177831173, + 0.009332869201898575, + 0.008871365338563919, + 0.003309696912765503, + 0.003113367361947894, + 0.004803194664418697, + 0.006507952697575092, + 0.004165827762335539, + 0.004844969604164362, + 0.005063012707978487, + 0.002972989110276103, + 0.027157964184880257, + 0.005357116460800171, + 0.01173890195786953, + 0.0061922017484903336, + 0.007422772701829672, + 0.015586109831929207, + 0.005812168121337891, + 0.010880816727876663, + 0.006109260488301516, + 0.003126251045614481, + 0.0035974793136119843, + 0.004038258921355009, + 0.0033110855147242546, + 0.0041578090749681, + 0.0029163642320781946, + 0.002283659065142274, + 0.014445191249251366, + 0.011133608408272266, + 0.011176913976669312, + 0.014116859063506126, + 0.008381905034184456, + 0.0124186547473073, + 0.011454630643129349, + 0.010318663902580738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 9 + ], + "selected_reward": -0.029802024364471436 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, low-angle, static shot captures a vibrant red plastic bottle cap, precariously balanced on its ridged rim. It rests on the textured surface of a dark, weathered wooden board. Warm late afternoon sunlight slants across the scene, casting a striking, dramatic light on the deep veins and tiny imperfections. The cap begins to tremble, a series of tiny, high-frequency vibrations that cause its long, twisted shadow to tremble nervously across the wood. Then, it begins a slow, agonizing struggle, teetering back and forth, cautiously trying to straighten itself, its smooth plastic form catching the golden light with each desperate lurch. After one last, prolonged lurch, its struggle against gravity fails; it collapses decisively, landing on its flat top with a soft, sharp click and finally resting in defeat, creating an atmosphere of quiet, focused, miniature, dramatic defeat.", + "seed": 30244, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_7/index_9149_seed_30244_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.028065038938075305, + "vlm_logprob": -0.028065038938075305, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.028065038938075305, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008885486982762814 + ], + "opposite_logprob_tokens": [ + -5.258885383605957 + ], + "logprob_reward": -0.008885486982762814, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008885486982762814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1850016564130783, + 0.0769471600651741, + 0.03905707970261574, + 0.042346056550741196, + 0.06629462540149689, + 0.04417894035577774, + 0.06613150238990784, + 0.2041328251361847, + 0.23712633550167084, + 0.03878370299935341 + ], + "credit_map": [ + 0.028779685497283936, + 0.020917799323797226, + 0.02078845538198948, + 0.023957043886184692, + 0.02947578951716423, + 0.02003924362361431, + 0.024033594876527786, + 0.01701006293296814, + 0.007106800097972155, + 0.014522683806717396, + 0.007864044047892094, + 0.010253476910293102, + 0.009715616703033447, + 0.008157651871442795, + 0.010378756560385227, + 0.008948132395744324, + 0.005775505676865578, + 0.00522896321490407, + 0.00358618819154799, + 0.00422727782279253, + 0.008501487784087658, + 0.003742305329069495, + 0.0034345597960054874, + 0.004560796078294516, + 0.005438839551061392, + 0.00326944119296968, + 0.005292245652526617, + 0.003431885503232479, + 0.012453312054276466, + 0.006996274460107088, + 0.00264274631626904, + 0.0028213143814355135, + 0.009766754694283009, + 0.004456222057342529, + 0.00747899292036891, + 0.00764614949002862, + 0.0108948964625597, + 0.007457870990037918, + 0.011004619300365448, + 0.007589123211801052, + 0.006582299713045359, + 0.00437166029587388, + 0.010581105947494507, + 0.0050880503840744495, + 0.005811542738229036, + 0.004073785617947578, + 0.0034964787773787975, + 0.004174019210040569, + 0.01672242023050785, + 0.006938469596207142, + 0.005017389543354511, + 0.003973009996116161, + 0.021846573799848557, + 0.004815928637981415, + 0.0035142130218446255, + 0.003303507110103965, + 0.018824052065610886, + 0.023009616881608963, + 0.028655262663960457, + 0.020841846242547035, + 0.025165855884552002, + 0.03820490837097168, + 0.02463945746421814, + 0.02479184977710247, + 0.052494507282972336, + 0.02428002655506134, + 0.013457877561450005, + 0.009919311851263046, + 0.09195183217525482, + 0.027380945160984993, + 0.010096363723278046, + 0.007545503322035074, + 0.005230749491602182, + 0.0038994005881249905, + 0.005155624821782112, + 0.003953626379370689, + 0.007227661088109016, + 0.006085171364247799, + 0.003593061352148652, + 0.00363841000944376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.008885486982762814 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813356161117554 + ], + "opposite_logprob_tokens": [ + -3.2881336212158203 + ], + "logprob_reward": -0.03813356161117554, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813356161117554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12681561708450317, + 0.10237836837768555, + 0.06354274600744247, + 0.07020195573568344, + 0.08927884697914124, + 0.07400774210691452, + 0.10577643662691116, + 0.06229003146290779, + 0.2539425194263458, + 0.05176571011543274 + ], + "credit_map": [ + 0.01384785957634449, + 0.01902398094534874, + 0.01465845387428999, + 0.011413255706429482, + 0.01740153692662716, + 0.013223242945969105, + 0.017850911244750023, + 0.019396372139453888, + 0.014219921082258224, + 0.019217532128095627, + 0.014468987472355366, + 0.01030364353209734, + 0.01301293633878231, + 0.011353367008268833, + 0.010791631415486336, + 0.009010348469018936, + 0.012012912891805172, + 0.006870257202535868, + 0.012925895862281322, + 0.005237045232206583, + 0.009800137020647526, + 0.006583412643522024, + 0.00403386726975441, + 0.006079215090721846, + 0.007400789763778448, + 0.0066286600194871426, + 0.007369722239673138, + 0.005324146710336208, + 0.02065182849764824, + 0.014110633172094822, + 0.004112125840038061, + 0.00460404995828867, + 0.02145957015454769, + 0.00631704879924655, + 0.010048028081655502, + 0.00983871053904295, + 0.01589052565395832, + 0.007668194826692343, + 0.009869213216006756, + 0.008187564089894295, + 0.009497151710093021, + 0.00904909148812294, + 0.011249740608036518, + 0.01168106496334076, + 0.010569498874247074, + 0.006051536183804274, + 0.009210559539496899, + 0.006699095480144024, + 0.025253022089600563, + 0.008647581562399864, + 0.007236984558403492, + 0.006357010453939438, + 0.03361210227012634, + 0.008899248205125332, + 0.00794217362999916, + 0.007828306406736374, + 0.00581213878467679, + 0.006310860626399517, + 0.010911297053098679, + 0.007247214205563068, + 0.008201463147997856, + 0.009402649477124214, + 0.007838970050215721, + 0.006565433461219072, + 0.05715090036392212, + 0.03027673251926899, + 0.015599404461681843, + 0.007765266578644514, + 0.09332003444433212, + 0.029288673773407936, + 0.0092542739585042, + 0.011287232860922813, + 0.009516641497612, + 0.005226064473390579, + 0.005445179529488087, + 0.0046317121013998985, + 0.012999034486711025, + 0.005950893275439739, + 0.003722632769495249, + 0.004273555241525173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.03813356161117554 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004644795786589384 + ], + "opposite_logprob_tokens": [ + -5.379644870758057 + ], + "logprob_reward": -0.004644795786589384, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004644795786589384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09546661376953125, + 0.10089634358882904, + 0.05725419893860817, + 0.11077096313238144, + 0.04879444092512131, + 0.09405752271413803, + 0.12037643790245056, + 0.09412010759115219, + 0.2225474864244461, + 0.0557158961892128 + ], + "credit_map": [ + 0.012812810018658638, + 0.009329508058726788, + 0.01775651052594185, + 0.00807332806289196, + 0.014935850165784359, + 0.013034963980317116, + 0.009769611060619354, + 0.009754037484526634, + 0.012093898840248585, + 0.015249626711010933, + 0.021597202867269516, + 0.008128630928695202, + 0.012636862695217133, + 0.016433732584118843, + 0.007269653957337141, + 0.007486737333238125, + 0.005864615552127361, + 0.007275110110640526, + 0.01191084086894989, + 0.005047593731433153, + 0.008495533838868141, + 0.007285463158041239, + 0.003768010763451457, + 0.007607028353959322, + 0.02152082696557045, + 0.012506445869803429, + 0.01941252313554287, + 0.009746973402798176, + 0.01612912304699421, + 0.014959161169826984, + 0.006434790790081024, + 0.010061122477054596, + 0.006202594377100468, + 0.00415846798568964, + 0.006925787311047316, + 0.005884495563805103, + 0.006870443932712078, + 0.004779359791427851, + 0.008559738285839558, + 0.005413555074483156, + 0.012678124941885471, + 0.010082254186272621, + 0.01175479032099247, + 0.01412961632013321, + 0.013392157852649689, + 0.009230987168848515, + 0.013420877046883106, + 0.0093687130138278, + 0.03065171279013157, + 0.011032539419829845, + 0.008792615495622158, + 0.004930594470351934, + 0.041785359382629395, + 0.009760896675288677, + 0.007025899365544319, + 0.0063968137837946415, + 0.01778443157672882, + 0.010457731783390045, + 0.011393201537430286, + 0.005609131883829832, + 0.026969432830810547, + 0.009338022209703922, + 0.005557768512517214, + 0.007010395638644695, + 0.05179332196712494, + 0.025308746844530106, + 0.01241244189441204, + 0.008228626102209091, + 0.08430655300617218, + 0.024858219549059868, + 0.007861420512199402, + 0.007778158877044916, + 0.011553560383617878, + 0.005094722844660282, + 0.0063430010341107845, + 0.005701097194105387, + 0.011140930466353893, + 0.0051042851991951466, + 0.006528292782604694, + 0.004250003956258297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.004644795786589384 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495598912239075 + ], + "opposite_logprob_tokens": [ + -2.9299559593200684 + ], + "logprob_reward": -0.05495598912239075, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495598912239075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10889609158039093, + 0.11103864759206772, + 0.05653192847967148, + 0.14866948127746582, + 0.05499472841620445, + 0.04904629662632942, + 0.2450352907180786, + 0.06685920804738998, + 0.09991974383592606, + 0.059008583426475525 + ], + "credit_map": [ + 0.01825089007616043, + 0.013197879306972027, + 0.018997376784682274, + 0.012297326698899269, + 0.010599772445857525, + 0.013426798395812511, + 0.010153739713132381, + 0.011972316540777683, + 0.015505120158195496, + 0.016703283414244652, + 0.0198235884308815, + 0.009108994156122208, + 0.01654411479830742, + 0.018572797998785973, + 0.006755867041647434, + 0.008024882525205612, + 0.0065709068439900875, + 0.008286464028060436, + 0.007597108371555805, + 0.007224204484373331, + 0.008220170624554157, + 0.006248925346881151, + 0.004154410678893328, + 0.008229736238718033, + 0.03281190246343613, + 0.011313105933368206, + 0.00885867141187191, + 0.006059737876057625, + 0.059129588305950165, + 0.02015927992761135, + 0.006176902912557125, + 0.0041602980345487595, + 0.008006309159100056, + 0.004216865636408329, + 0.004953645635396242, + 0.00566841522231698, + 0.010242434218525887, + 0.006772411987185478, + 0.008741912432014942, + 0.006392733193933964, + 0.006167879328131676, + 0.00377963250502944, + 0.007379034534096718, + 0.007344974670559168, + 0.008406766690313816, + 0.005526107270270586, + 0.00549935782328248, + 0.0049425396136939526, + 0.05106951668858528, + 0.02479611709713936, + 0.021283823996782303, + 0.02274486795067787, + 0.060818642377853394, + 0.01686359941959381, + 0.028405848890542984, + 0.019052866846323013, + 0.0072295474819839, + 0.008196902461349964, + 0.010381994768977165, + 0.007472652476280928, + 0.006596154533326626, + 0.012579472735524178, + 0.00719447759911418, + 0.007208009250462055, + 0.020867474377155304, + 0.01008015125989914, + 0.00829552486538887, + 0.006726666819304228, + 0.03143735229969025, + 0.010710052214562893, + 0.006002154666930437, + 0.005800373386591673, + 0.009228231385350227, + 0.005732876248657703, + 0.006953510455787182, + 0.008088486269116402, + 0.008416613563895226, + 0.006268166936933994, + 0.00842874962836504, + 0.005891951732337475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.05495598912239075 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370536118745804 + ], + "opposite_logprob_tokens": [ + -3.408705472946167 + ], + "logprob_reward": -0.03370536118745804, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370536118745804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06746342033147812, + 0.15561626851558685, + 0.04891403391957283, + 0.10503897815942764, + 0.05970737710595131, + 0.05556716024875641, + 0.3320733904838562, + 0.03165541961789131, + 0.040022112429142, + 0.10394182801246643 + ], + "credit_map": [ + 0.011613921262323856, + 0.007466820999979973, + 0.013972674496471882, + 0.005999297369271517, + 0.008053487166762352, + 0.007588673382997513, + 0.00551211042329669, + 0.007256430108100176, + 0.028088467195630074, + 0.026542698964476585, + 0.020040851086378098, + 0.013303500600159168, + 0.02472148835659027, + 0.017602603882551193, + 0.014082012698054314, + 0.011234637349843979, + 0.004968265071511269, + 0.007888034917414188, + 0.0059175328351557255, + 0.007008823566138744, + 0.006789622362703085, + 0.005287928972393274, + 0.00416775094345212, + 0.006886070128530264, + 0.014107255265116692, + 0.007964192889630795, + 0.005571491084992886, + 0.004146312363445759, + 0.041959915310144424, + 0.02229190617799759, + 0.0051817395724356174, + 0.0038161694537848234, + 0.008218562230467796, + 0.005792688112705946, + 0.006712580099701881, + 0.007977833040058613, + 0.010688076727092266, + 0.008087517693638802, + 0.006389198359102011, + 0.005840923171490431, + 0.008327610790729523, + 0.005942042451351881, + 0.008169877342879772, + 0.009145533666014671, + 0.008261038921773434, + 0.005242067854851484, + 0.005361285991966724, + 0.005117703229188919, + 0.08124690502882004, + 0.026720071211457253, + 0.026497989892959595, + 0.0277557373046875, + 0.09482837468385696, + 0.02324823848903179, + 0.03224418684840202, + 0.019531890749931335, + 0.003589714877307415, + 0.00396444508805871, + 0.0058485884219408035, + 0.0034003739710897207, + 0.003555640811100602, + 0.004745301324874163, + 0.003415266051888466, + 0.003136090934276581, + 0.008534284308552742, + 0.003582708304747939, + 0.004585494287312031, + 0.0023241201415657997, + 0.013381759636104107, + 0.0037526285741478205, + 0.0017834206810221076, + 0.0020776926539838314, + 0.01864219456911087, + 0.011065544560551643, + 0.013765012845396996, + 0.015771953389048576, + 0.013332993723452091, + 0.010477705858647823, + 0.013115039095282555, + 0.007771385367959738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 3 + ], + "selected_reward": -0.03370536118745804 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, low-angle, static shot captures a vibrant red plastic bottle cap, precariously balanced on its ridged rim. It rests on the textured surface of a dark, weathered wooden board. Warm late afternoon sunlight slants across the scene, casting a striking, dramatic light on the deep veins and tiny imperfections. The cap begins to tremble, a series of tiny, high-frequency vibrations that cause its long, twisted shadow to tremble nervously across the wood. Then, it begins a slow, agonizing struggle, teetering back and forth, cautiously trying to straighten itself, its smooth plastic form catching the golden light with each desperate lurch. After one last, prolonged lurch, its struggle against gravity fails; it collapses decisively, landing on its flat top with a soft, sharp click and finally resting in defeat, creating an atmosphere of quiet, focused, miniature, dramatic defeat.", + "seed": 30245, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_7/index_9149_seed_30245_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03576953895390034, + "vlm_logprob": -0.03576953895390034, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.03576953895390034, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013904373161494732 + ], + "opposite_logprob_tokens": [ + -5.263904571533203 + ], + "logprob_reward": -0.013904373161494732, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.013904373161494732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1685340851545334, + 0.043625086545944214, + 0.04811772331595421, + 0.04650159180164337, + 0.06153944134712219, + 0.24869243800640106, + 0.11447853595018387, + 0.13572128117084503, + 0.060804903507232666, + 0.0719849020242691 + ], + "credit_map": [ + 0.033502668142318726, + 0.012527399696409702, + 0.01951313391327858, + 0.023717276751995087, + 0.019607271999120712, + 0.02196500450372696, + 0.02180149406194687, + 0.01589982956647873, + 0.005348998121917248, + 0.004674591589719057, + 0.005998004227876663, + 0.007254394236952066, + 0.004393237642943859, + 0.007598480209708214, + 0.004498792812228203, + 0.003858588170260191, + 0.010505110025405884, + 0.003740747459232807, + 0.006475756410509348, + 0.006915603764355183, + 0.0038943239487707615, + 0.0071759577840566635, + 0.005447091534733772, + 0.0039631337858736515, + 0.007290597073733807, + 0.0043020774610340595, + 0.005080883856862783, + 0.006434180773794651, + 0.0046221306547522545, + 0.009176270104944706, + 0.00568885775282979, + 0.003906597848981619, + 0.015132713131606579, + 0.0038103824481368065, + 0.0042285798117518425, + 0.0067757852375507355, + 0.004340495448559523, + 0.018394803628325462, + 0.004359268583357334, + 0.00449741305783391, + 0.03387678787112236, + 0.027927258983254433, + 0.02085864543914795, + 0.027585580945014954, + 0.036810554563999176, + 0.036051951348781586, + 0.036457642912864685, + 0.02912403643131256, + 0.017394699156284332, + 0.015739062801003456, + 0.010527892969548702, + 0.011989526450634003, + 0.017935635522007942, + 0.013592058792710304, + 0.012189673259854317, + 0.01510999258607626, + 0.028269914910197258, + 0.010383548215031624, + 0.01259485725313425, + 0.014881263487040997, + 0.014330040663480759, + 0.016123846173286438, + 0.017797283828258514, + 0.021340517327189445, + 0.010678300634026527, + 0.007481934502720833, + 0.008071927353739738, + 0.0067823585122823715, + 0.007449153810739517, + 0.007645437028259039, + 0.006368833594024181, + 0.006326957140117884, + 0.014929159544408321, + 0.005860656965523958, + 0.006941467523574829, + 0.008830646052956581, + 0.005813910625874996, + 0.009199196472764015, + 0.009033352136611938, + 0.011376512236893177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.013904373161494732 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054966822266578674 + ], + "opposite_logprob_tokens": [ + -2.929966926574707 + ], + "logprob_reward": -0.054966822266578674, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054966822266578674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07212430983781815, + 0.0471949502825737, + 0.0352102555334568, + 0.0455879345536232, + 0.07971111685037613, + 0.33587685227394104, + 0.2778913378715515, + 0.028109075501561165, + 0.05444484204053879, + 0.023849286139011383 + ], + "credit_map": [ + 0.008924377150833607, + 0.006128277629613876, + 0.009255403652787209, + 0.010482480749487877, + 0.0060950759798288345, + 0.012325625866651535, + 0.010073495097458363, + 0.008839576505124569, + 0.006974227726459503, + 0.004157081246376038, + 0.005678858608007431, + 0.00940508209168911, + 0.004296724684536457, + 0.007577214390039444, + 0.004679468926042318, + 0.004426294472068548, + 0.00447007967159152, + 0.0030023311264812946, + 0.005545241758227348, + 0.00672972621396184, + 0.0034191019367426634, + 0.004928608890622854, + 0.0037330600898712873, + 0.0033821051474660635, + 0.00908289197832346, + 0.004063607659190893, + 0.005772066302597523, + 0.00624476931989193, + 0.0036660220939666033, + 0.008191047236323357, + 0.0033001869451254606, + 0.005267343018203974, + 0.018351435661315918, + 0.005514139775186777, + 0.0061217909678816795, + 0.009115989319980145, + 0.0069715059362351894, + 0.01925758272409439, + 0.006279736757278442, + 0.008098936639726162, + 0.050614312291145325, + 0.04713299870491028, + 0.02574373036623001, + 0.02847076766192913, + 0.05221548676490784, + 0.07616948336362839, + 0.028636904433369637, + 0.026893174275755882, + 0.033779438585042953, + 0.05350862070918083, + 0.023612704128026962, + 0.033526260405778885, + 0.0450117252767086, + 0.03323690965771675, + 0.027230989187955856, + 0.027984682470560074, + 0.004848741460591555, + 0.0031742441933602095, + 0.003308058949187398, + 0.003536737523972988, + 0.0034223173279315233, + 0.0030863587744534016, + 0.0034082934726029634, + 0.0033243228681385517, + 0.009603519923985004, + 0.008054028265178204, + 0.004431987646967173, + 0.009117386303842068, + 0.004983639344573021, + 0.006510152015835047, + 0.006748606916517019, + 0.004995522554963827, + 0.003161631291732192, + 0.004257579334080219, + 0.002623997163027525, + 0.003705279203131795, + 0.0021920143626630306, + 0.0029537349473685026, + 0.002783834235742688, + 0.0021712167654186487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.054966822266578674 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046440837904810905 + ], + "opposite_logprob_tokens": [ + -5.37964391708374 + ], + "logprob_reward": -0.0046440837904810905, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046440837904810905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05816180631518364, + 0.07186195999383926, + 0.08158528804779053, + 0.049907200038433075, + 0.1488480567932129, + 0.2248304784297943, + 0.16051951050758362, + 0.07705235481262207, + 0.06196359172463417, + 0.06526970118284225 + ], + "credit_map": [ + 0.010160353034734726, + 0.004646215587854385, + 0.007033087313175201, + 0.008601907640695572, + 0.005009646061807871, + 0.010243477299809456, + 0.006269045174121857, + 0.0061980742029845715, + 0.013243534602224827, + 0.005664462223649025, + 0.008039862848818302, + 0.008521823212504387, + 0.005526086315512657, + 0.0107438825070858, + 0.00786367803812027, + 0.012258627451956272, + 0.018951186910271645, + 0.006474394351243973, + 0.009637018665671349, + 0.01006108894944191, + 0.006146078463643789, + 0.011951586231589317, + 0.011393868364393711, + 0.006970065645873547, + 0.012535768561065197, + 0.0028351345099508762, + 0.004872527439147234, + 0.004617671947926283, + 0.002695320639759302, + 0.013250789605081081, + 0.0037133528385311365, + 0.005386632867157459, + 0.03008045069873333, + 0.009515884332358837, + 0.013471624813973904, + 0.008445654064416885, + 0.009847401641309261, + 0.03968062996864319, + 0.012520944699645042, + 0.02528546191751957, + 0.04782218113541603, + 0.019451197236776352, + 0.019164400175213814, + 0.017602695152163506, + 0.02238279953598976, + 0.055910322815179825, + 0.019652169197797775, + 0.022844720631837845, + 0.019568968564271927, + 0.024760058149695396, + 0.01509007252752781, + 0.018703311681747437, + 0.02540583722293377, + 0.019918132573366165, + 0.015261352993547916, + 0.02181178890168667, + 0.014337471686303616, + 0.00725450087338686, + 0.006994687486439943, + 0.009967806749045849, + 0.008684474974870682, + 0.009422369301319122, + 0.01087238546460867, + 0.00951866339892149, + 0.009758862666785717, + 0.007744453847408295, + 0.005562674719840288, + 0.009345709346234798, + 0.005808042362332344, + 0.007507770322263241, + 0.009409035556018353, + 0.006827045697718859, + 0.011672668159008026, + 0.0075375111773610115, + 0.006016270723193884, + 0.009181560948491096, + 0.005782677326351404, + 0.009074644185602665, + 0.008565827272832394, + 0.0074385409243404865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.0046440837904810905 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898680120706558 + ], + "opposite_logprob_tokens": [ + -2.578986883163452 + ], + "logprob_reward": -0.07898680120706558, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898680120706558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07818277180194855, + 0.040935367345809937, + 0.04222094640135765, + 0.033871520310640335, + 0.07370705902576447, + 0.1842275857925415, + 0.3564135730266571, + 0.07341261953115463, + 0.062062814831733704, + 0.05496573448181152 + ], + "credit_map": [ + 0.011834520846605301, + 0.006667248904705048, + 0.00898910965770483, + 0.008097738027572632, + 0.011073160916566849, + 0.012242330238223076, + 0.009940749034285545, + 0.009337925352156162, + 0.006863432936370373, + 0.0031965558882802725, + 0.00498763844370842, + 0.004339946433901787, + 0.003212355077266693, + 0.011262360960245132, + 0.002910129725933075, + 0.00416295463219285, + 0.007587649393826723, + 0.003621326759457588, + 0.005168996285647154, + 0.00471344543620944, + 0.003202026942744851, + 0.011445886455476284, + 0.00301223318092525, + 0.0034693870693445206, + 0.0042099729180336, + 0.003583647310733795, + 0.0036067888140678406, + 0.004170247819274664, + 0.0036562772002071142, + 0.005100157577544451, + 0.005123831797391176, + 0.004420598968863487, + 0.01559383049607277, + 0.005476565565913916, + 0.004536399617791176, + 0.007387254852801561, + 0.0072058215737342834, + 0.020165277644991875, + 0.004425848834216595, + 0.008916068822145462, + 0.03442499414086342, + 0.02838299423456192, + 0.017067745327949524, + 0.02599208801984787, + 0.02313806861639023, + 0.02040284499526024, + 0.017695650458335876, + 0.01712321862578392, + 0.039843376725912094, + 0.07563412934541702, + 0.026424627751111984, + 0.04984940588474274, + 0.057477016001939774, + 0.04512685164809227, + 0.03344302251935005, + 0.028615150600671768, + 0.012628249824047089, + 0.006161709316074848, + 0.007108396850526333, + 0.009790496900677681, + 0.008123049512505531, + 0.00935458391904831, + 0.010703608393669128, + 0.009542531333863735, + 0.008700575679540634, + 0.008579391054809093, + 0.005595446098595858, + 0.010316379368305206, + 0.006167234852910042, + 0.007125301286578178, + 0.009617703035473824, + 0.005960787180811167, + 0.009682603180408478, + 0.006063051056116819, + 0.0057484786957502365, + 0.007744152564555407, + 0.0047517637722194195, + 0.008814184926450253, + 0.006834748666733503, + 0.005326752085238695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.07898680120706558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026345614343881607 + ], + "opposite_logprob_tokens": [ + -3.651345729827881 + ], + "logprob_reward": -0.026345614343881607, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026345614343881607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06883269548416138, + 0.06842999905347824, + 0.059531670063734055, + 0.04475855454802513, + 0.07169153541326523, + 0.18141843378543854, + 0.36165693402290344, + 0.062446001917123795, + 0.05199912562966347, + 0.029235050082206726 + ], + "credit_map": [ + 0.009408827871084213, + 0.006548506673425436, + 0.006365921348333359, + 0.00889323465526104, + 0.01010141521692276, + 0.009804995730519295, + 0.009819556027650833, + 0.007890237495303154, + 0.009357418864965439, + 0.006267895922064781, + 0.008281894959509373, + 0.008354210294783115, + 0.005521554034203291, + 0.020291047170758247, + 0.0053690713830292225, + 0.004986905492842197, + 0.0067576635628938675, + 0.008115541189908981, + 0.006691917311400175, + 0.007036702241748571, + 0.008409276604652405, + 0.010505317710340023, + 0.006097397767007351, + 0.005917850881814957, + 0.007239534053951502, + 0.004642202518880367, + 0.004295007325708866, + 0.006203751545399427, + 0.0046042525209486485, + 0.0069777267053723335, + 0.006647922098636627, + 0.004148156847804785, + 0.014459905214607716, + 0.006604823749512434, + 0.005158211570233107, + 0.00746392784640193, + 0.00793430395424366, + 0.018455900251865387, + 0.00602850504219532, + 0.005585958249866962, + 0.030906299129128456, + 0.02722649835050106, + 0.01594318076968193, + 0.025444841012358665, + 0.026018405333161354, + 0.020321732386946678, + 0.020415373146533966, + 0.015142093412578106, + 0.04746895283460617, + 0.06669135391712189, + 0.0349225290119648, + 0.04398560896515846, + 0.05652225390076637, + 0.045766524970531464, + 0.032654594630002975, + 0.03364512324333191, + 0.01209132932126522, + 0.00541708106175065, + 0.006863108370453119, + 0.0090720159932971, + 0.006850885227322578, + 0.007355288602411747, + 0.007856207899749279, + 0.006940084975212812, + 0.007624757941812277, + 0.007360209245234728, + 0.006168498191982508, + 0.005915969610214233, + 0.007174476515501738, + 0.0060648201033473015, + 0.005548711400479078, + 0.0061416844837367535, + 0.004871199373155832, + 0.004264971241354942, + 0.0044656554237008095, + 0.003843757091090083, + 0.002843402558937669, + 0.0032073070760816336, + 0.0026478117797523737, + 0.0030909439083188772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.026345614343881607 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, low-angle, static shot captures a vibrant red plastic bottle cap, precariously balanced on its ridged rim. It rests on the textured surface of a dark, weathered wooden board. Warm late afternoon sunlight slants across the scene, casting a striking, dramatic light on the deep veins and tiny imperfections. The cap begins to tremble, a series of tiny, high-frequency vibrations that cause its long, twisted shadow to tremble nervously across the wood. Then, it begins a slow, agonizing struggle, teetering back and forth, cautiously trying to straighten itself, its smooth plastic form catching the golden light with each desperate lurch. After one last, prolonged lurch, its struggle against gravity fails; it collapses decisively, landing on its flat top with a soft, sharp click and finally resting in defeat, creating an atmosphere of quiet, focused, miniature, dramatic defeat.", + "seed": 30246, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_7/index_9149_seed_30246_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02988225594162941, + "vlm_logprob": -0.02988225594162941, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.02988225594162941, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01573515497148037 + ], + "opposite_logprob_tokens": [ + -5.140735149383545 + ], + "logprob_reward": -0.01573515497148037, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01573515497148037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1111588180065155, + 0.03266514837741852, + 0.08039195090532303, + 0.06070287153124809, + 0.04371960461139679, + 0.11696496605873108, + 0.042711831629276276, + 0.04849053919315338, + 0.2959866225719452, + 0.16720765829086304 + ], + "credit_map": [ + 0.012507821433246136, + 0.016194958239793777, + 0.014961079694330692, + 0.013017707504332066, + 0.015003957785665989, + 0.010619944892823696, + 0.011871627531945705, + 0.016981735825538635, + 0.003670378588140011, + 0.007459003012627363, + 0.00343419355340302, + 0.0021280325017869473, + 0.004625700414180756, + 0.0045980303548276424, + 0.0041237180121243, + 0.002626097062602639, + 0.012533440254628658, + 0.009958780370652676, + 0.012567081488668919, + 0.008631656877696514, + 0.011079683899879456, + 0.008135485462844372, + 0.010410372167825699, + 0.007075451780110598, + 0.007524389307945967, + 0.009230063296854496, + 0.00821415614336729, + 0.0071394313126802444, + 0.011140203103423119, + 0.0075038885697722435, + 0.004700882826000452, + 0.005249866284430027, + 0.006386978086084127, + 0.003963345196098089, + 0.00765864597633481, + 0.004258486442267895, + 0.0034698625095188618, + 0.005610693246126175, + 0.004595667123794556, + 0.007775930222123861, + 0.011702808551490307, + 0.01392336469143629, + 0.017962181940674782, + 0.013348030857741833, + 0.01554599404335022, + 0.0120597118511796, + 0.016162173822522163, + 0.01626071333885193, + 0.007420942187309265, + 0.004215063061565161, + 0.009265477769076824, + 0.0038996022194623947, + 0.0025394619442522526, + 0.006533890962600708, + 0.0033707304392009974, + 0.0054666632786393166, + 0.014734278433024883, + 0.0030738734640181065, + 0.008172283880412579, + 0.002068515866994858, + 0.00315665639936924, + 0.011908920481801033, + 0.002940933918580413, + 0.0024350823368877172, + 0.06381309032440186, + 0.0464174821972847, + 0.06379060447216034, + 0.029995519667863846, + 0.011879033409059048, + 0.029521262273192406, + 0.021482093259692192, + 0.029087550938129425, + 0.02161942981183529, + 0.025345679372549057, + 0.019175946712493896, + 0.01804175414144993, + 0.028311915695667267, + 0.02199649065732956, + 0.020475907251238823, + 0.012240555137395859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.01573515497148037 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048677828162908554 + ], + "opposite_logprob_tokens": [ + -3.048677921295166 + ], + "logprob_reward": -0.048677828162908554, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048677828162908554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08774245530366898, + 0.024725070223212242, + 0.09444863349199295, + 0.08097387105226517, + 0.08439677208662033, + 0.22539560496807098, + 0.04997715353965759, + 0.018696725368499756, + 0.29453110694885254, + 0.0391126424074173 + ], + "credit_map": [ + 0.009682203643023968, + 0.010405631735920906, + 0.013665629550814629, + 0.009267857298254967, + 0.016140440478920937, + 0.00947360135614872, + 0.00884588249027729, + 0.010261205025017262, + 0.003361666575074196, + 0.0033657134044915438, + 0.003859912045300007, + 0.0020332112908363342, + 0.0027727370616048574, + 0.0029655899852514267, + 0.0035964606795459986, + 0.0027697780169546604, + 0.011624038219451904, + 0.015178939327597618, + 0.010896285064518452, + 0.013629963621497154, + 0.01579035073518753, + 0.00994910765439272, + 0.010593118146061897, + 0.006786819081753492, + 0.016191042959690094, + 0.008827156387269497, + 0.017715178430080414, + 0.008652559481561184, + 0.007356424815952778, + 0.008000008761882782, + 0.00583602674305439, + 0.008395472541451454, + 0.011183686554431915, + 0.009005882777273655, + 0.011919927783310413, + 0.01130432728677988, + 0.009427237324416637, + 0.010085600428283215, + 0.009304453618824482, + 0.012165656313300133, + 0.024190912023186684, + 0.023749228566884995, + 0.04005701467394829, + 0.026701439172029495, + 0.02558545023202896, + 0.02384837344288826, + 0.02272683009505272, + 0.03853633627295494, + 0.006614698562771082, + 0.0049628219567239285, + 0.007713013794273138, + 0.004912644159048796, + 0.0053854212164878845, + 0.006274828687310219, + 0.0060157012194395065, + 0.008098019286990166, + 0.005713249556720257, + 0.0010712267830967903, + 0.0036476608365774155, + 0.0010096130426973104, + 0.0009131596307270229, + 0.0037691385950893164, + 0.0012667065020650625, + 0.0013059695484116673, + 0.07263300567865372, + 0.041529834270477295, + 0.0665227472782135, + 0.02648903802037239, + 0.01083335280418396, + 0.02718561701476574, + 0.023550063371658325, + 0.02578742615878582, + 0.005711226724088192, + 0.006965650711208582, + 0.010483825579285622, + 0.007463426794856787, + 0.0019450410036370158, + 0.00232845195569098, + 0.0022409381344914436, + 0.0019740797579288483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.048677828162908554 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745306774973869 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.006745306774973869, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745306774973869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1222953274846077, + 0.0774097815155983, + 0.02657543122768402, + 0.030501393601298332, + 0.025009194388985634, + 0.042053453624248505, + 0.02683865651488304, + 0.04006578400731087, + 0.5910800099372864, + 0.018170928582549095 + ], + "credit_map": [ + 0.013961095362901688, + 0.018618494272232056, + 0.020342228934168816, + 0.014331130310893059, + 0.010308478027582169, + 0.013425356708467007, + 0.008765374310314655, + 0.022543178871273994, + 0.012623890303075314, + 0.009474584832787514, + 0.01303005125373602, + 0.006956609431654215, + 0.0060945553705096245, + 0.007529823109507561, + 0.009279005229473114, + 0.012421267107129097, + 0.0032766617368906736, + 0.004606991540640593, + 0.004066119436174631, + 0.0035985256545245647, + 0.0037352610379457474, + 0.0025483074132353067, + 0.002131886314600706, + 0.002611681120470166, + 0.004476685542613268, + 0.005063973367214203, + 0.006819851230829954, + 0.0037180413492023945, + 0.0029051967430859804, + 0.003017469309270382, + 0.0017251464305445552, + 0.002775035332888365, + 0.003676575142890215, + 0.0031575937755405903, + 0.004192098043859005, + 0.0025057715829461813, + 0.003181625623255968, + 0.002763075055554509, + 0.002368832239881158, + 0.003163625719025731, + 0.005448478274047375, + 0.004064068663865328, + 0.009146330878138542, + 0.00503692589700222, + 0.003649029415100813, + 0.0036496720276772976, + 0.00471139932051301, + 0.006347554735839367, + 0.004161156248301268, + 0.003193780081346631, + 0.0057884762063622475, + 0.0030358934309333563, + 0.0020543283317238092, + 0.003350653452798724, + 0.0024396779481321573, + 0.002814694307744503, + 0.009605835191905499, + 0.003608440048992634, + 0.005477606318891048, + 0.0014933369820937514, + 0.0019489160040393472, + 0.014284978620707989, + 0.0015901916194707155, + 0.0020564827136695385, + 0.15600237250328064, + 0.08232296258211136, + 0.14183716475963593, + 0.050136443227529526, + 0.020638225600123405, + 0.049419499933719635, + 0.042258381843566895, + 0.04846502095460892, + 0.0030210025142878294, + 0.002400376833975315, + 0.004319207277148962, + 0.0027535883709788322, + 0.0012814257061108947, + 0.0015070757362991571, + 0.0017444528639316559, + 0.0011438021901994944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.006745306774973869 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054960839450359344 + ], + "opposite_logprob_tokens": [ + -2.9299607276916504 + ], + "logprob_reward": -0.054960839450359344, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054960839450359344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1404724270105362, + 0.03841188922524452, + 0.0883675068616867, + 0.06607067584991455, + 0.025783758610486984, + 0.07464722543954849, + 0.02296086773276329, + 0.03454478085041046, + 0.47575876116752625, + 0.03298214450478554 + ], + "credit_map": [ + 0.015308627858757973, + 0.01751220040023327, + 0.02238868735730648, + 0.0162215456366539, + 0.015276940539479256, + 0.013894865289330482, + 0.013762708753347397, + 0.02610684558749199, + 0.004842097871005535, + 0.007069130428135395, + 0.005608608014881611, + 0.003007703460752964, + 0.00526563310995698, + 0.003984654322266579, + 0.004576217848807573, + 0.004057840909808874, + 0.011664756573736668, + 0.01518794521689415, + 0.012259452603757381, + 0.012457826174795628, + 0.013331347145140171, + 0.009456159546971321, + 0.007306546904146671, + 0.006703465711325407, + 0.007516640704125166, + 0.01115384791046381, + 0.007756248582154512, + 0.009574107825756073, + 0.011242014355957508, + 0.00824777688831091, + 0.004749038256704807, + 0.005831001792103052, + 0.0034511364065110683, + 0.0023915988858789206, + 0.003912235610187054, + 0.003939033020287752, + 0.0022193738259375095, + 0.003156704595312476, + 0.003671711077913642, + 0.003041964489966631, + 0.008482566103339195, + 0.0074247452430427074, + 0.012451812624931335, + 0.008555779233574867, + 0.008468231186270714, + 0.008447001688182354, + 0.007232476491481066, + 0.013584613800048828, + 0.004275944549590349, + 0.0026053746696561575, + 0.005572378169745207, + 0.0021060628350824118, + 0.0018001827411353588, + 0.0027128157671540976, + 0.0016675557708367705, + 0.002220552647486329, + 0.00499262660741806, + 0.00297960895113647, + 0.009190725162625313, + 0.0019004509085789323, + 0.0023780534975230694, + 0.00862243864685297, + 0.0015213872538879514, + 0.002959486562758684, + 0.1130111888051033, + 0.06841795146465302, + 0.10832030326128006, + 0.04442843049764633, + 0.016928959637880325, + 0.044812072068452835, + 0.039351485669612885, + 0.040488384664058685, + 0.004828926641494036, + 0.0071634394116699696, + 0.004055479541420937, + 0.005907636135816574, + 0.002388102002441883, + 0.003621191019192338, + 0.0027301497757434845, + 0.002287218812853098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.054960839450359344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a vibrant red plastic bottle cap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a low angle and remains static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cap rest on the ridged rim of a dark, weathered wooden board with warm late afternoon sunlight casting dramatic light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the cap exhibit a sequence of high-frequency trembling vibrations followed by a slow, teetering struggle against gravity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the cap eventually collapse decisively onto its flat top, landing with a sharp click and coming to rest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329215034842491 + ], + "opposite_logprob_tokens": [ + -3.773292064666748 + ], + "logprob_reward": -0.02329215034842491, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329215034842491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22136355936527252, + 0.05439620092511177, + 0.1302422732114792, + 0.14206545054912567, + 0.0861983448266983, + 0.16125206649303436, + 0.03758300095796585, + 0.03331140801310539, + 0.07201750576496124, + 0.0615701861679554 + ], + "credit_map": [ + 0.017637943848967552, + 0.03563835471868515, + 0.024038225412368774, + 0.032349493354558945, + 0.038351282477378845, + 0.02194121852517128, + 0.03070661798119545, + 0.02070043422281742, + 0.0069391801953315735, + 0.008214695379137993, + 0.009523170068860054, + 0.0042209019884467125, + 0.006796832662075758, + 0.0067613194696605206, + 0.006143081933259964, + 0.005797022022306919, + 0.014580943621695042, + 0.022748364135622978, + 0.015414067544043064, + 0.020728562027215958, + 0.02206309884786606, + 0.013526069931685925, + 0.010255036875605583, + 0.010926122777163982, + 0.01867481879889965, + 0.018652744591236115, + 0.020095691084861755, + 0.01877540349960327, + 0.020408084616065025, + 0.0167282372713089, + 0.0094107985496521, + 0.01931968331336975, + 0.011816777288913727, + 0.009303949773311615, + 0.013945895247161388, + 0.00904065277427435, + 0.010561289265751839, + 0.010513539426028728, + 0.008944186381995678, + 0.012072056531906128, + 0.017809415236115456, + 0.01797143928706646, + 0.02211657725274563, + 0.02198578231036663, + 0.016868233680725098, + 0.017368324100971222, + 0.02245277538895607, + 0.02467953786253929, + 0.006617116741836071, + 0.004184432327747345, + 0.009341646917164326, + 0.0036516713444143534, + 0.0027034904342144728, + 0.003331757616251707, + 0.0032375892624258995, + 0.00451529398560524, + 0.006389622576534748, + 0.0040678950026631355, + 0.006188797764480114, + 0.0037645327392965555, + 0.003733863588422537, + 0.0032164095900952816, + 0.0029704838525503874, + 0.00297980266623199, + 0.014864981174468994, + 0.007098437752574682, + 0.018495069816708565, + 0.005404115188866854, + 0.0036544986069202423, + 0.009778185747563839, + 0.006924587767571211, + 0.005797622259706259, + 0.009244204498827457, + 0.01219527330249548, + 0.009871181100606918, + 0.0077613480389118195, + 0.007088277488946915, + 0.0065753101371228695, + 0.00557235162705183, + 0.0032622392755001783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.02329215034842491 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_8/index_tensor([159, 159])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_8/index_tensor([159, 159])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..84fb25d63f8c11af13a5ee25f9e225dd325a30b5 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_8/index_tensor([159, 159])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level close-up captures a pair of hands meticulously arranging a collection of vintage hardcover books. These hands, belonging to a person wearing the sleeves of a cozy forest green knit sweater, hold a thick hardcover bound in worn crimson leather. The book's spine features a faint, faded gilt pattern, and the pages are a warm, decked ivory. With steady, deliberate precision, the hands guide the book onto a dark, polished mahogany bookshelf already crowded with similar antique volumes. The camera follows the person as they gently and carefully slide the crimson book into the perfectly sized, narrow gap between the aquamarine and dark brown volumes. The book slides smoothly into place until its spine aligns perfectly with the neighboring books, a movement that conveys a deep sense of satisfaction and order. The scene is bathed in the warm, soft glow of an unseen lamp, which casts gentle highlights across the richly textured leather spines and the rich grain of the wood, creating deep, soft shadows. The overall atmosphere is one of tranquil focus and quiet reverence, a peaceful moment dedicated to the love of literature.", + "seed": 7232, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_8/index_159_seed_7232_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.023210163647308945, + "vlm_logprob": -0.023210163647308945, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.023210163647308945, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017287522554397583 + ], + "opposite_logprob_tokens": [ + -4.767287731170654 + ], + "logprob_reward": -0.017287522554397583, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.017287522554397583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06183891370892525, + 0.07053913921117783, + 0.04614068567752838, + 0.05733545869588852, + 0.0651397854089737, + 0.0645553395152092, + 0.06545396894216537, + 0.28234755992889404, + 0.1051439493894577, + 0.18150517344474792 + ], + "credit_map": [ + 0.006041410844773054, + 0.007858537137508392, + 0.007932637818157673, + 0.010564101859927177, + 0.005046213511377573, + 0.006947959307581186, + 0.010225142352283001, + 0.007222906220704317, + 0.004961520899087191, + 0.006905415561050177, + 0.005611894652247429, + 0.00813961960375309, + 0.010912992060184479, + 0.007835824973881245, + 0.017189836129546165, + 0.008982038125395775, + 0.009575586766004562, + 0.004853166174143553, + 0.006294913124293089, + 0.006899118423461914, + 0.00462532602250576, + 0.003837216179817915, + 0.0045777251943945885, + 0.005477630998939276, + 0.0061346436850726604, + 0.009780483320355415, + 0.007043732330203056, + 0.006792075932025909, + 0.004490433260798454, + 0.007027504034340382, + 0.007220848463475704, + 0.008845734409987926, + 0.005565381143242121, + 0.00762089341878891, + 0.004015475045889616, + 0.00928515661507845, + 0.011783551424741745, + 0.009622137062251568, + 0.01070569921284914, + 0.006541488692164421, + 0.007769092917442322, + 0.009743924252688885, + 0.00713560963049531, + 0.010458357632160187, + 0.005220816936343908, + 0.009928732179105282, + 0.008005335927009583, + 0.006293463986366987, + 0.005980641581118107, + 0.0074887461960315704, + 0.003876051865518093, + 0.014153916388750076, + 0.00647661741822958, + 0.006138958968222141, + 0.007716253865510225, + 0.01362277939915657, + 0.011762872338294983, + 0.11596363037824631, + 0.008780387230217457, + 0.04497193172574043, + 0.01628989167511463, + 0.009731329046189785, + 0.06272068619728088, + 0.012126839719712734, + 0.006917708553373814, + 0.007467912044376135, + 0.003575169015675783, + 0.03060711733996868, + 0.006976894102990627, + 0.007441443391144276, + 0.008615483529865742, + 0.0335422158241272, + 0.019988978281617165, + 0.01310728583484888, + 0.011360188014805317, + 0.07456626743078232, + 0.020676473155617714, + 0.01732957921922207, + 0.013643302023410797, + 0.01083308830857277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.017287522554397583 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636384405195713 + ], + "opposite_logprob_tokens": [ + -3.6513638496398926 + ], + "logprob_reward": -0.02636384405195713, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636384405195713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05601360276341438, + 0.022425541654229164, + 0.06794144958257675, + 0.28510549664497375, + 0.04468507319688797, + 0.02635950967669487, + 0.025572139769792557, + 0.3673996925354004, + 0.03666703402996063, + 0.06783048808574677 + ], + "credit_map": [ + 0.005732117220759392, + 0.0061078243888914585, + 0.013349208980798721, + 0.006566909607499838, + 0.004629400558769703, + 0.0034678508527576923, + 0.00842222198843956, + 0.007738069165498018, + 0.001992257544770837, + 0.002641767030581832, + 0.0022349038626998663, + 0.002837727079167962, + 0.00445275055244565, + 0.002782681491225958, + 0.0031394497491419315, + 0.0023440043441951275, + 0.01050766184926033, + 0.009124252945184708, + 0.008795308880507946, + 0.011126753874123096, + 0.008523703552782536, + 0.008116220124065876, + 0.004819903988391161, + 0.006927640177309513, + 0.042359836399555206, + 0.027545800432562828, + 0.051065146923065186, + 0.024446694180369377, + 0.02105218917131424, + 0.07254409044981003, + 0.020937738940119743, + 0.025153998285531998, + 0.0037089306861162186, + 0.005036283750087023, + 0.0027606331277638674, + 0.005532656330615282, + 0.006135267671197653, + 0.00827354472130537, + 0.007303931750357151, + 0.0059338221326470375, + 0.003215295961126685, + 0.003359818598255515, + 0.0026287203654646873, + 0.003914596512913704, + 0.0026646163314580917, + 0.004135590977966785, + 0.003250814275816083, + 0.0031900552567094564, + 0.002724830759689212, + 0.0048143938183784485, + 0.001739074825309217, + 0.0034606631379574537, + 0.004161191638559103, + 0.002206535544246435, + 0.0037047327496111393, + 0.002760718110948801, + 0.010992868803441525, + 0.1358516812324524, + 0.009653019718825817, + 0.14250004291534424, + 0.016772646456956863, + 0.009212746284902096, + 0.02983276918530464, + 0.012583902105689049, + 0.0024314355105161667, + 0.00310090696439147, + 0.0014130211202427745, + 0.010679616592824459, + 0.0020226044580340385, + 0.002294601174071431, + 0.0031307912431657314, + 0.011594056151807308, + 0.0071158260107040405, + 0.008899837732315063, + 0.004315381404012442, + 0.027079058811068535, + 0.004790801089257002, + 0.005830845329910517, + 0.005101815797388554, + 0.0046969223767519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.02636384405195713 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205231696367264 + ], + "opposite_logprob_tokens": [ + -2.8120522499084473 + ], + "logprob_reward": -0.06205231696367264, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205231696367264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049032796174287796, + 0.07812072336673737, + 0.0515158548951149, + 0.10721387714147568, + 0.13661037385463715, + 0.059372130781412125, + 0.07215896993875504, + 0.18675348162651062, + 0.0774446651339531, + 0.18177711963653564 + ], + "credit_map": [ + 0.0037585280369967222, + 0.007475031539797783, + 0.005484779365360737, + 0.009592672809958458, + 0.005023117642849684, + 0.0027648652903735638, + 0.007258110214024782, + 0.007675687316805124, + 0.004508857149630785, + 0.006620720960199833, + 0.005420684814453125, + 0.005910636391490698, + 0.012078607454895973, + 0.01782766357064247, + 0.01567864418029785, + 0.01007490511983633, + 0.008954478427767754, + 0.004379732068628073, + 0.008640659973025322, + 0.006021476350724697, + 0.005174637772142887, + 0.005835024639964104, + 0.005026057828217745, + 0.007483788300305605, + 0.0071127950213849545, + 0.006531538907438517, + 0.01935860700905323, + 0.01284761168062687, + 0.006499765906482935, + 0.02041013352572918, + 0.01976211927831173, + 0.014691313728690147, + 0.012820933014154434, + 0.0102347731590271, + 0.009506885893642902, + 0.0151926064863801, + 0.02952742949128151, + 0.020196497440338135, + 0.01619277149438858, + 0.022938476875424385, + 0.009290526621043682, + 0.0051873112097382545, + 0.004831379279494286, + 0.005667627789080143, + 0.005943343974649906, + 0.011831581592559814, + 0.009346774779260159, + 0.007273579947650433, + 0.005787539761513472, + 0.014376461505889893, + 0.004121717531234026, + 0.0070488713681697845, + 0.0094611132517457, + 0.010228942148387432, + 0.01014972198754549, + 0.010984594002366066, + 0.01114879734814167, + 0.016734927892684937, + 0.025070784613490105, + 0.036466989666223526, + 0.020369257777929306, + 0.015303356572985649, + 0.03536096215248108, + 0.026298394426703453, + 0.007447491399943829, + 0.008442573249340057, + 0.005729276221245527, + 0.008709854446351528, + 0.008537343703210354, + 0.007351140026003122, + 0.011233987286686897, + 0.019992999732494354, + 0.017703454941511154, + 0.013571074232459068, + 0.011849114671349525, + 0.07319773733615875, + 0.022219980135560036, + 0.017971470952033997, + 0.014285125769674778, + 0.010979163460433483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.06205231696367264 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634967729449272 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008634967729449272, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634967729449272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041369106620550156, + 0.05095149576663971, + 0.028996946290135384, + 0.026256727054715157, + 0.030740154907107353, + 0.0527535155415535, + 0.03727085515856743, + 0.511360228061676, + 0.0613146536052227, + 0.15898633003234863 + ], + "credit_map": [ + 0.0032916967757046223, + 0.0029878655914217234, + 0.005454221740365028, + 0.008654185570776463, + 0.003882151562720537, + 0.002535766689106822, + 0.007727175485342741, + 0.006836043205112219, + 0.003981036599725485, + 0.0036797409411519766, + 0.005382867995649576, + 0.005152290221303701, + 0.006033392623066902, + 0.005035996437072754, + 0.0162285678088665, + 0.005457604303956032, + 0.003478275379166007, + 0.00414227694272995, + 0.003910427447408438, + 0.0035744442138820887, + 0.003446059999987483, + 0.003391060046851635, + 0.0035509306471794844, + 0.0035034730099141598, + 0.002761168172582984, + 0.002831326564773917, + 0.0055248974822461605, + 0.0029496445786207914, + 0.0015852537471801043, + 0.004108162596821785, + 0.004054874647408724, + 0.002441397635266185, + 0.002871075877919793, + 0.0025803865864872932, + 0.0020849488209933043, + 0.002767107682302594, + 0.007809098344296217, + 0.005386990960687399, + 0.0041366866789758205, + 0.0031038604211062193, + 0.004957693163305521, + 0.004266634117811918, + 0.003810959868133068, + 0.006183074787259102, + 0.006897854618728161, + 0.012099134735763073, + 0.008314057253301144, + 0.006224109325557947, + 0.004148957319557667, + 0.004660646431148052, + 0.0021532997488975525, + 0.004567945376038551, + 0.005579733289778233, + 0.0035557562950998545, + 0.0059969923458993435, + 0.0066075255163013935, + 0.014382577501237392, + 0.16018034517765045, + 0.014269073493778706, + 0.21685095131397247, + 0.021311916410923004, + 0.011748557910323143, + 0.05571649596095085, + 0.016900300979614258, + 0.005629230290651321, + 0.0051950677298009396, + 0.0026072233449667692, + 0.015554890036582947, + 0.0049169971607625484, + 0.0043480577878654, + 0.00554940290749073, + 0.017513783648610115, + 0.016809044405817986, + 0.029842492192983627, + 0.011099092662334442, + 0.05334624648094177, + 0.007371047046035528, + 0.013107451610267162, + 0.011210737749934196, + 0.016200225800275803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.008634967729449272 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017121669370681047 + ], + "opposite_logprob_tokens": [ + -6.376712322235107 + ], + "logprob_reward": -0.0017121669370681047, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017121669370681047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0772104561328888, + 0.039420273154973984, + 0.07265119254589081, + 0.06793571263551712, + 0.17647282779216766, + 0.06041663512587547, + 0.03652364760637283, + 0.28009143471717834, + 0.11382707208395004, + 0.07545076310634613 + ], + "credit_map": [ + 0.007934056222438812, + 0.007769518066197634, + 0.008935351856052876, + 0.014252906665205956, + 0.006683636922389269, + 0.005431702826172113, + 0.011444606818258762, + 0.014758670702576637, + 0.0027675749734044075, + 0.00423491420224309, + 0.0030191175173968077, + 0.004651117604225874, + 0.006707960274070501, + 0.007936898618936539, + 0.00534515967592597, + 0.004757524933665991, + 0.007568964269012213, + 0.009849325753748417, + 0.009102247655391693, + 0.007140120957046747, + 0.011097200214862823, + 0.01108345203101635, + 0.007503311615437269, + 0.009306566789746284, + 0.006876929197460413, + 0.011708058416843414, + 0.014731988310813904, + 0.005485463887453079, + 0.004042359068989754, + 0.00982400868088007, + 0.009174665436148643, + 0.006092233583331108, + 0.01367892324924469, + 0.014534198679029942, + 0.01321262028068304, + 0.017933439463377, + 0.04111788049340248, + 0.02795976586639881, + 0.0274619460105896, + 0.020574040710926056, + 0.005357513204216957, + 0.0047316341660916805, + 0.0035710767842829227, + 0.00643028412014246, + 0.007178837433457375, + 0.015389150008559227, + 0.008840017020702362, + 0.008918114937841892, + 0.003759833285585046, + 0.006789017468690872, + 0.0020416572224348783, + 0.004175744950771332, + 0.004930874332785606, + 0.003734773024916649, + 0.006181633099913597, + 0.004910110030323267, + 0.008844023570418358, + 0.10267020761966705, + 0.006830195896327496, + 0.10880299657583237, + 0.012826853431761265, + 0.007632011082023382, + 0.02256176620721817, + 0.009923354722559452, + 0.010045395232737064, + 0.007539075333625078, + 0.004684036131948233, + 0.02321959286928177, + 0.010006185621023178, + 0.008374257944524288, + 0.011657215654850006, + 0.03830131143331528, + 0.009214012883603573, + 0.016505436971783638, + 0.006691671907901764, + 0.02155262790620327, + 0.003981430549174547, + 0.005471586249768734, + 0.005316802766174078, + 0.006717189680784941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.0017121669370681047 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures a pair of hands meticulously arranging a collection of vintage hardcover books. These hands, belonging to a person wearing the sleeves of a cozy forest green knit sweater, hold a thick hardcover bound in worn crimson leather. The book's spine features a faint, faded gilt pattern, and the pages are a warm, decked ivory. With steady, deliberate precision, the hands guide the book onto a dark, polished mahogany bookshelf already crowded with similar antique volumes. The camera follows the person as they gently and carefully slide the crimson book into the perfectly sized, narrow gap between the aquamarine and dark brown volumes. The book slides smoothly into place until its spine aligns perfectly with the neighboring books, a movement that conveys a deep sense of satisfaction and order. The scene is bathed in the warm, soft glow of an unseen lamp, which casts gentle highlights across the richly textured leather spines and the rich grain of the wood, creating deep, soft shadows. The overall atmosphere is one of tranquil focus and quiet reverence, a peaceful moment dedicated to the love of literature.", + "seed": 7233, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_8/index_159_seed_7233_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01529574478045106, + "vlm_logprob": -0.01529574478045106, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.01529574478045106, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014387793838977814 + ], + "opposite_logprob_tokens": [ + -5.014387607574463 + ], + "logprob_reward": -0.014387793838977814, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.014387793838977814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050506193190813065, + 0.03842521458864212, + 0.04352192208170891, + 0.03482409194111824, + 0.06843096762895584, + 0.4995207190513611, + 0.07449713349342346, + 0.054281085729599, + 0.042164865881204605, + 0.09382779896259308 + ], + "credit_map": [ + 0.006010567769408226, + 0.004439326003193855, + 0.008421082980930805, + 0.006399035453796387, + 0.004058208782225847, + 0.0056273965165019035, + 0.00933775957673788, + 0.006212812382727861, + 0.003665006021037698, + 0.00503187533468008, + 0.006159535609185696, + 0.004553523380309343, + 0.005909048952162266, + 0.004393679089844227, + 0.0057600331492722034, + 0.0029525072313845158, + 0.004113591276109219, + 0.004074873402714729, + 0.004404474049806595, + 0.006693562958389521, + 0.004978438373655081, + 0.006098569370806217, + 0.0076275295577943325, + 0.005530876573175192, + 0.00322438171133399, + 0.0020693966653198004, + 0.004768937360495329, + 0.0037278933450579643, + 0.003567059990018606, + 0.006491136271506548, + 0.004366465378552675, + 0.006608815863728523, + 0.006060086190700531, + 0.006825456395745277, + 0.012891387566924095, + 0.00798339769244194, + 0.0068823606707155704, + 0.009983735159039497, + 0.005782051011919975, + 0.012022488750517368, + 0.02717006579041481, + 0.008413156494498253, + 0.18336954712867737, + 0.022518722340464592, + 0.024621203541755676, + 0.036073774099349976, + 0.02403334155678749, + 0.17332085967063904, + 0.008900071494281292, + 0.006386845838278532, + 0.012741371057927608, + 0.007176652550697327, + 0.00842295866459608, + 0.011340200901031494, + 0.012276985682547092, + 0.007252035196870565, + 0.005306181963533163, + 0.007241211831569672, + 0.006744931917637587, + 0.004574306309223175, + 0.006717488169670105, + 0.0068458751775324345, + 0.010342932306230068, + 0.006508149206638336, + 0.003955170512199402, + 0.0039909309707582, + 0.0060841660015285015, + 0.0036141874734312296, + 0.008056114427745342, + 0.006352967582643032, + 0.005157212261110544, + 0.004954110831022263, + 0.0069507393054664135, + 0.01036712434142828, + 0.010044272057712078, + 0.013663893565535545, + 0.01503472775220871, + 0.011780069209635258, + 0.011594361625611782, + 0.014392605051398277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.014387793838977814 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007642790675163269 + ], + "opposite_logprob_tokens": [ + -4.88264274597168 + ], + "logprob_reward": -0.007642790675163269, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007642790675163269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061477746814489365, + 0.03773323819041252, + 0.036863695830106735, + 0.03271664306521416, + 0.05454543977975845, + 0.5926462411880493, + 0.070429228246212, + 0.04859257489442825, + 0.015525861643254757, + 0.04946930333971977 + ], + "credit_map": [ + 0.006533346604555845, + 0.005943040829151869, + 0.01049056276679039, + 0.008723357692360878, + 0.005542844999581575, + 0.006729860790073872, + 0.011266721412539482, + 0.006248014513403177, + 0.0033242818899452686, + 0.0036869330797344446, + 0.006942462176084518, + 0.004131333436816931, + 0.006002864800393581, + 0.006360100582242012, + 0.004011936020106077, + 0.003273329231888056, + 0.002641912316903472, + 0.00262266444042325, + 0.004336510319262743, + 0.004508703015744686, + 0.004459759686142206, + 0.009015311487019062, + 0.00526256812736392, + 0.0040162671357393265, + 0.0032483339309692383, + 0.002755327383056283, + 0.0046865385957062244, + 0.004157957620918751, + 0.004501942079514265, + 0.005037416238337755, + 0.003834904870018363, + 0.004494224209338427, + 0.004716694820672274, + 0.004532531835138798, + 0.01068006083369255, + 0.006640331819653511, + 0.004769610706716776, + 0.007812872529029846, + 0.005695351865142584, + 0.009697987698018551, + 0.03420034795999527, + 0.009303390048444271, + 0.20642121136188507, + 0.029713628813624382, + 0.030117757618427277, + 0.04131324589252472, + 0.02779100090265274, + 0.21378570795059204, + 0.007458862848579884, + 0.007164883892983198, + 0.012249773368239403, + 0.007587720640003681, + 0.010935797356069088, + 0.009177948348224163, + 0.008866767399013042, + 0.0069874729961156845, + 0.006547464989125729, + 0.006308936979621649, + 0.006402968894690275, + 0.005042754579335451, + 0.006673425901681185, + 0.006295430473983288, + 0.005839610006660223, + 0.005481978412717581, + 0.001241405843757093, + 0.0013722878647968173, + 0.002131375018507242, + 0.0014162070583552122, + 0.0026802951470017433, + 0.0018325041746720672, + 0.002830216661095619, + 0.002021570224314928, + 0.0032894634641706944, + 0.005919818300753832, + 0.005901744589209557, + 0.006835533306002617, + 0.00648022536188364, + 0.009525962173938751, + 0.005032637156546116, + 0.006483918987214565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.007642790675163269 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307664558291435 + ], + "opposite_logprob_tokens": [ + -3.168076753616333 + ], + "logprob_reward": -0.04307664558291435, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307664558291435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0932258889079094, + 0.058002542704343796, + 0.12275949120521545, + 0.03415124863386154, + 0.08783040940761566, + 0.24447910487651825, + 0.06955429166555405, + 0.055107638239860535, + 0.06102851778268814, + 0.17386090755462646 + ], + "credit_map": [ + 0.011772826313972473, + 0.008512758649885654, + 0.016239825636148453, + 0.012976122088730335, + 0.00784218404442072, + 0.011324632912874222, + 0.01617209054529667, + 0.008385444059967995, + 0.004561309237033129, + 0.005521985236555338, + 0.007049014326184988, + 0.008292890153825283, + 0.008385002613067627, + 0.008755136281251907, + 0.010030142962932587, + 0.00540706142783165, + 0.008487495593726635, + 0.0113541753962636, + 0.015537699684500694, + 0.023238230496644974, + 0.013662750832736492, + 0.018953731283545494, + 0.02064882218837738, + 0.010876578278839588, + 0.004720694851130247, + 0.0028219837695360184, + 0.004994767718017101, + 0.004623835906386375, + 0.0027587341610342264, + 0.005574071779847145, + 0.004669172689318657, + 0.003987989388406277, + 0.008231092244386673, + 0.009069141000509262, + 0.013316725380718708, + 0.01055803894996643, + 0.009630469605326653, + 0.01462812814861536, + 0.009961229749023914, + 0.012435579672455788, + 0.013565434142947197, + 0.004473170265555382, + 0.07454821467399597, + 0.01258478220552206, + 0.021640252321958542, + 0.017442280426621437, + 0.011320047080516815, + 0.08890492469072342, + 0.0061502354219555855, + 0.005186270456761122, + 0.012525168247520924, + 0.006285428069531918, + 0.0119111817330122, + 0.014199400320649147, + 0.006641964428126812, + 0.006654644850641489, + 0.0031017723958939314, + 0.004532654769718647, + 0.008055519312620163, + 0.003872025990858674, + 0.010544192045927048, + 0.007212105207145214, + 0.0115133635699749, + 0.006276004482060671, + 0.00410482520237565, + 0.004542710725218058, + 0.008393611758947372, + 0.006085686851292849, + 0.011133487336337566, + 0.0079300906509161, + 0.00814527366310358, + 0.010692830197513103, + 0.010875736363232136, + 0.018738077953457832, + 0.01647775061428547, + 0.02241046540439129, + 0.02618350274860859, + 0.024988053366541862, + 0.02828306145966053, + 0.025904258713126183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.04307664558291435 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735597271472216 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735597271472216, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735597271472216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1925366222858429, + 0.05397142842411995, + 0.031440772116184235, + 0.0344412587583065, + 0.04413264989852905, + 0.37765368819236755, + 0.06689445674419403, + 0.05041590332984924, + 0.03683251515030861, + 0.11168065667152405 + ], + "credit_map": [ + 0.023001467809081078, + 0.01724303886294365, + 0.03431885689496994, + 0.026438895612955093, + 0.018019704148173332, + 0.022692805156111717, + 0.034324176609516144, + 0.016497693955898285, + 0.006489480379968882, + 0.004410669207572937, + 0.009618633426725864, + 0.00621507503092289, + 0.008998775854706764, + 0.006739144679158926, + 0.005918075330555439, + 0.005581574980169535, + 0.0032646094914525747, + 0.0036535400431603193, + 0.004538488574326038, + 0.005625746212899685, + 0.0032848110422492027, + 0.004381694830954075, + 0.0036864145658910275, + 0.0030054671224206686, + 0.0035493485629558563, + 0.0023246919736266136, + 0.0048992992378771305, + 0.004285359289497137, + 0.0036130354274064302, + 0.006339354440569878, + 0.004144368693232536, + 0.005285801365971565, + 0.004786265082657337, + 0.005127515643835068, + 0.006536561995744705, + 0.0058370050974190235, + 0.006786258891224861, + 0.004861269146203995, + 0.005893190391361713, + 0.0043045817874372005, + 0.022367479279637337, + 0.006526873912662268, + 0.13353659212589264, + 0.019372915849089622, + 0.018036678433418274, + 0.027142973616719246, + 0.01762799359858036, + 0.13304215669631958, + 0.005023970268666744, + 0.004946023691445589, + 0.010815548710525036, + 0.005833971779793501, + 0.015745893120765686, + 0.012537674978375435, + 0.0057289074175059795, + 0.00626246677711606, + 0.0034425805788487196, + 0.0035646462347358465, + 0.007071042899042368, + 0.004690730012953281, + 0.007024405989795923, + 0.004392421338707209, + 0.013289334252476692, + 0.006940742023289204, + 0.0021378714591264725, + 0.003349359380081296, + 0.004862523637712002, + 0.003554973052814603, + 0.004906448535621166, + 0.005493806675076485, + 0.006372346542775631, + 0.006155190523713827, + 0.0077015371061861515, + 0.016045013442635536, + 0.013799779117107391, + 0.01783238723874092, + 0.01069734152406454, + 0.01623094640672207, + 0.014183761551976204, + 0.015189891681075096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.006735597271472216 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635896533727646 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004635896533727646, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635896533727646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09516462683677673, + 0.06611811369657516, + 0.17579205334186554, + 0.07105298340320587, + 0.07209497690200806, + 0.2510891556739807, + 0.042982906103134155, + 0.0870724469423294, + 0.01679224520921707, + 0.12184052914381027 + ], + "credit_map": [ + 0.011705346405506134, + 0.008414625190198421, + 0.01658325083553791, + 0.012874280102550983, + 0.008447873406112194, + 0.010654539801180363, + 0.01803990639746189, + 0.00844479538500309, + 0.004832020960748196, + 0.008584395982325077, + 0.0061029610224068165, + 0.005848499480634928, + 0.011175100691616535, + 0.013591174967586994, + 0.008060482330620289, + 0.007923471741378307, + 0.016605420038104057, + 0.018977468833327293, + 0.030247779563069344, + 0.03036852553486824, + 0.01589040271937847, + 0.02990003675222397, + 0.019246039912104607, + 0.014556361362338066, + 0.00801852997392416, + 0.007340626325458288, + 0.007691610138863325, + 0.009617329575121403, + 0.007982909679412842, + 0.009933484718203545, + 0.009577196091413498, + 0.010891293175518513, + 0.00632527656853199, + 0.0038234407547861338, + 0.015491621568799019, + 0.010111100971698761, + 0.008266146294772625, + 0.012491456232964993, + 0.0068052178248763084, + 0.008780710399150848, + 0.015176933258771896, + 0.00538548082113266, + 0.08937841653823853, + 0.012513728812336922, + 0.02121407724916935, + 0.01779750920832157, + 0.0152377188205719, + 0.07438529282808304, + 0.004046326968818903, + 0.003380330977961421, + 0.006476030219346285, + 0.004795446991920471, + 0.007431712467223406, + 0.007101103197783232, + 0.0051791067235171795, + 0.0045728422701358795, + 0.006303322035819292, + 0.00803878903388977, + 0.012744334526360035, + 0.007057989947497845, + 0.011034483090043068, + 0.01050445158034563, + 0.01975124329328537, + 0.011637830175459385, + 0.0015557868173345923, + 0.001135455910116434, + 0.0025829083751887083, + 0.0017245784401893616, + 0.002416053554043174, + 0.0024189576506614685, + 0.0026830011047422886, + 0.0022755020763725042, + 0.008071274496614933, + 0.015425330959260464, + 0.013192621059715748, + 0.015131060965359211, + 0.013167480938136578, + 0.018170375376939774, + 0.01723054051399231, + 0.02145182155072689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.004635896533727646 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures a pair of hands meticulously arranging a collection of vintage hardcover books. These hands, belonging to a person wearing the sleeves of a cozy forest green knit sweater, hold a thick hardcover bound in worn crimson leather. The book's spine features a faint, faded gilt pattern, and the pages are a warm, decked ivory. With steady, deliberate precision, the hands guide the book onto a dark, polished mahogany bookshelf already crowded with similar antique volumes. The camera follows the person as they gently and carefully slide the crimson book into the perfectly sized, narrow gap between the aquamarine and dark brown volumes. The book slides smoothly into place until its spine aligns perfectly with the neighboring books, a movement that conveys a deep sense of satisfaction and order. The scene is bathed in the warm, soft glow of an unseen lamp, which casts gentle highlights across the richly textured leather spines and the rich grain of the wood, creating deep, soft shadows. The overall atmosphere is one of tranquil focus and quiet reverence, a peaceful moment dedicated to the love of literature.", + "seed": 7234, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_8/index_159_seed_7234_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01158447265625, + "vlm_logprob": -0.01158447265625, + "vlm_yes_no_margin": 4.674999952316284, + "vlm_token_credit": -0.01158447265625, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009362954646348953 + ], + "opposite_logprob_tokens": [ + -5.509363174438477 + ], + "logprob_reward": -0.009362954646348953, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.009362954646348953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13576330244541168, + 0.2868197560310364, + 0.029062826186418533, + 0.14233741164207458, + 0.24524718523025513, + 0.07358277589082718, + 0.01848803460597992, + 0.025130556896328926, + 0.027992483228445053, + 0.015575703233480453 + ], + "credit_map": [ + 0.010145182721316814, + 0.05657177418470383, + 0.0037707851734012365, + 0.008961000479757786, + 0.04038634151220322, + 0.005317656788975, + 0.007502024527639151, + 0.0031085398513823748, + 0.014133652672171593, + 0.02043604850769043, + 0.008844226598739624, + 0.10393615812063217, + 0.03228835016489029, + 0.010209228843450546, + 0.02089332789182663, + 0.07607875019311905, + 0.0025626264978200197, + 0.003930926322937012, + 0.002871739910915494, + 0.005000548902899027, + 0.0035461520310491323, + 0.0026927085127681494, + 0.00496985437348485, + 0.0034882701002061367, + 0.006739114411175251, + 0.05691373720765114, + 0.005047786049544811, + 0.037096403539180756, + 0.011554141528904438, + 0.004310137126594782, + 0.007351086940616369, + 0.013324997387826443, + 0.02309579774737358, + 0.015142117626965046, + 0.06568525731563568, + 0.021567240357398987, + 0.013664568774402142, + 0.019568711519241333, + 0.023549510166049004, + 0.0629739835858345, + 0.0033490818459540606, + 0.005937111098319292, + 0.007201069034636021, + 0.005775537807494402, + 0.02481171116232872, + 0.005922443699091673, + 0.006003253161907196, + 0.014582567848265171, + 0.0013138237409293652, + 0.0017737159505486488, + 0.0017853598110377789, + 0.0027435782831162214, + 0.0035473674070090055, + 0.0032757152803242207, + 0.0021188752725720406, + 0.0019295980455353856, + 0.0018077793065458536, + 0.003329204162582755, + 0.0034558982588350773, + 0.003174928482621908, + 0.0037600959185510874, + 0.004627503454685211, + 0.002906518755480647, + 0.0020686304196715355, + 0.003641401417553425, + 0.003167889779433608, + 0.0033594032283872366, + 0.004483611322939396, + 0.0034790157806128263, + 0.0049719782546162605, + 0.0023685551714152098, + 0.0025206278078258038, + 0.001678233384154737, + 0.0012776704970747232, + 0.0015110581880435348, + 0.002542919246479869, + 0.0030614808201789856, + 0.0021851537749171257, + 0.0018083387985825539, + 0.0015108479419723153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.009362954646348953 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012554424814879894 + ], + "opposite_logprob_tokens": [ + -4.38755464553833 + ], + "logprob_reward": -0.012554424814879894, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012554424814879894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22632484138011932, + 0.5241271257400513, + 0.006472129840403795, + 0.023551108315587044, + 0.11707232892513275, + 0.01447836123406887, + 0.02763039618730545, + 0.024745892733335495, + 0.029134249314665794, + 0.006463552825152874 + ], + "credit_map": [ + 0.01525861769914627, + 0.09115152060985565, + 0.0183577723801136, + 0.012327959761023521, + 0.06431833654642105, + 0.00817116815596819, + 0.011600799858570099, + 0.005138668231666088, + 0.02665039896965027, + 0.03656570985913277, + 0.01575855165719986, + 0.19198544323444366, + 0.054569896310567856, + 0.018751556053757668, + 0.03770696744322777, + 0.14213861525058746, + 0.0008469360182061791, + 0.0008463834528811276, + 0.0007553232135251164, + 0.0011866603745147586, + 0.0005917677190154791, + 0.0005424849805422127, + 0.0010318271815776825, + 0.0006707469001412392, + 0.002225663745775819, + 0.006495241541415453, + 0.002138689858838916, + 0.0036367997527122498, + 0.002179034287109971, + 0.0013197027146816254, + 0.0031093202996999025, + 0.0024466568138450384, + 0.010932357981801033, + 0.0060855187475681305, + 0.03319612145423889, + 0.010039113461971283, + 0.006402067840099335, + 0.008520109578967094, + 0.010270562022924423, + 0.03162648156285286, + 0.0010649360483512282, + 0.0014068952295929193, + 0.0014320602640509605, + 0.0016801696037873626, + 0.003355195978656411, + 0.0012958733132109046, + 0.0019503935473039746, + 0.0022928379476070404, + 0.0021977454889565706, + 0.003334041452035308, + 0.004379169084131718, + 0.002802543807774782, + 0.005640061572194099, + 0.0050712586380541325, + 0.002247557742521167, + 0.001958017935976386, + 0.001824952312745154, + 0.0022246965672820807, + 0.0039037344977259636, + 0.003794230753555894, + 0.00356847676448524, + 0.005144254770129919, + 0.002527877688407898, + 0.001757670077495277, + 0.0035735962446779013, + 0.0034349902998656034, + 0.004442376550287008, + 0.005362942349165678, + 0.0033080156426876783, + 0.004572839941829443, + 0.0023057579528540373, + 0.0021337303332984447, + 0.0005652167601510882, + 0.0004680538841057569, + 0.0007412566337734461, + 0.0012364988215267658, + 0.0011516117956489325, + 0.0008290822152048349, + 0.0009370443876832724, + 0.0005347883561626077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.012554424814879894 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819819025695324 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.01819819025695324, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819819025695324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1863717883825302, + 0.36277905106544495, + 0.019246334210038185, + 0.02545541524887085, + 0.35615643858909607, + 0.007620085030794144, + 0.006224026437848806, + 0.007484490983188152, + 0.01993440091609955, + 0.008727933280169964 + ], + "credit_map": [ + 0.01433657482266426, + 0.07808567583560944, + 0.00913122296333313, + 0.011141972616314888, + 0.05313749238848686, + 0.007019622717052698, + 0.009657139889895916, + 0.003862090641632676, + 0.018633700907230377, + 0.02637946605682373, + 0.010348140262067318, + 0.1327454149723053, + 0.03745095059275627, + 0.012730770744383335, + 0.026345735415816307, + 0.09814486652612686, + 0.002312128432095051, + 0.002427899744361639, + 0.0018492870731279254, + 0.002724972553551197, + 0.0016890076221898198, + 0.001426714938133955, + 0.004330643452703953, + 0.002485680626705289, + 0.0012554065324366093, + 0.00978232454508543, + 0.0010207637678831816, + 0.0066907950676977634, + 0.001978421350941062, + 0.0009013704839162529, + 0.0013789307558909059, + 0.002447402337566018, + 0.040360789746046066, + 0.020813828334212303, + 0.08764283359050751, + 0.03183203190565109, + 0.01929117925465107, + 0.0273252222687006, + 0.03427540883421898, + 0.0946151614189148, + 0.0006102871848270297, + 0.0011219432344660163, + 0.0005267668748274446, + 0.0007785217021591961, + 0.001600898802280426, + 0.0005812384188175201, + 0.0009559549507685006, + 0.0014444743283092976, + 0.00041599266114644706, + 0.000457370828371495, + 0.0006977173616178334, + 0.0007312202942557633, + 0.001224392675794661, + 0.0011071652406826615, + 0.0009869375498965383, + 0.0006032296223565936, + 0.00048396343481726944, + 0.0006911875098012388, + 0.0009453617385588586, + 0.0007670068880543113, + 0.0014864198165014386, + 0.0011477088555693626, + 0.001158717554062605, + 0.0008041252731345594, + 0.0024780977983027697, + 0.002186922123655677, + 0.002985744969919324, + 0.003727244446054101, + 0.0023949919268488884, + 0.003001672448590398, + 0.0016410000389441848, + 0.0015187279786914587, + 0.0006208663107827306, + 0.0005752266151830554, + 0.0007331673405133188, + 0.0014529029140248895, + 0.0017272435361519456, + 0.0012855539098381996, + 0.0013549148570746183, + 0.0009780575055629015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.01819819025695324 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073091067373753 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.011073091067373753, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073091067373753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3008652925491333, + 0.319250226020813, + 0.01864876225590706, + 0.07489162683486938, + 0.18119849264621735, + 0.01532411202788353, + 0.011414148844778538, + 0.023548979312181473, + 0.026979926973581314, + 0.027878403663635254 + ], + "credit_map": [ + 0.02343093976378441, + 0.1279991865158081, + 0.01880033314228058, + 0.017577018588781357, + 0.08118247985839844, + 0.011211594566702843, + 0.014708312228322029, + 0.005955429747700691, + 0.0159410759806633, + 0.022224435582756996, + 0.010192211717367172, + 0.11648853123188019, + 0.03459324315190315, + 0.01142917387187481, + 0.02282778173685074, + 0.08555379509925842, + 0.0020010650623589754, + 0.002296098507940769, + 0.0015965027268975973, + 0.0028591782320290804, + 0.0020165545865893364, + 0.0009558587335050106, + 0.004581335466355085, + 0.002342168940231204, + 0.004107792396098375, + 0.028267506510019302, + 0.0036857922095805407, + 0.016872316598892212, + 0.006980655714869499, + 0.0031335053499788046, + 0.005220672115683556, + 0.006623386405408382, + 0.03386520966887474, + 0.0125639783218503, + 0.02928033284842968, + 0.016544418409466743, + 0.009824534878134727, + 0.014959192834794521, + 0.01775575801730156, + 0.04640506953001022, + 0.0010909956181421876, + 0.0018027536571025848, + 0.0014648593496531248, + 0.0017120724078267813, + 0.002694864524528384, + 0.0013723676092922688, + 0.0021695785690099, + 0.0030166208744049072, + 0.000760519877076149, + 0.0009460184373892844, + 0.0010381881147623062, + 0.001152837066911161, + 0.002661449834704399, + 0.0023117430973798037, + 0.001268387888558209, + 0.001275004236958921, + 0.0011541591957211494, + 0.003458317369222641, + 0.0018831989727914333, + 0.0023672888055443764, + 0.005624169483780861, + 0.0041300831362605095, + 0.00303268083371222, + 0.0018990820972248912, + 0.002456504153087735, + 0.002299724845215678, + 0.0026479943189769983, + 0.003371874336153269, + 0.005186688620597124, + 0.005844518542289734, + 0.0023115368094295263, + 0.0028610879089683294, + 0.0024923549499362707, + 0.0019508077530190349, + 0.003500651102513075, + 0.0062435464933514595, + 0.0044665466994047165, + 0.0031610848382115364, + 0.0037299254909157753, + 0.0023334857542067766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.011073091067373753 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067337024956941605 + ], + "opposite_logprob_tokens": [ + -5.0067338943481445 + ], + "logprob_reward": -0.0067337024956941605, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067337024956941605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0787745788693428, + 0.6427951455116272, + 0.014938597567379475, + 0.05883733928203583, + 0.0713934451341629, + 0.032060109078884125, + 0.01897425577044487, + 0.014192425645887852, + 0.054594721645116806, + 0.013439439237117767 + ], + "credit_map": [ + 0.006884351838380098, + 0.029018834233283997, + 0.021178510040044785, + 0.007245774846524, + 0.0052594211883842945, + 0.003244359977543354, + 0.0037995364982634783, + 0.0021437781397253275, + 0.03234531357884407, + 0.045577291399240494, + 0.018649056553840637, + 0.23636873066425323, + 0.06660407781600952, + 0.022617923095822334, + 0.046441949903964996, + 0.17419080436229706, + 0.0019861424807459116, + 0.0019547438714653254, + 0.0013909911504015326, + 0.0025786294136196375, + 0.0018561171600595117, + 0.0009595605661161244, + 0.002434466267004609, + 0.0017779460176825523, + 0.0030545112676918507, + 0.023288393393158913, + 0.002486448036506772, + 0.01460267510265112, + 0.0051666139625012875, + 0.001812837552279234, + 0.0030728736892342567, + 0.0053529758006334305, + 0.016181739047169685, + 0.005347828380763531, + 0.007193166296929121, + 0.007264973595738411, + 0.004707633052021265, + 0.006521564442664385, + 0.00797735620290041, + 0.016199178993701935, + 0.001519512734375894, + 0.0021857202518731356, + 0.002727772342041135, + 0.0027627251110970974, + 0.01010387297719717, + 0.002571657532826066, + 0.0033659404143691063, + 0.006822902709245682, + 0.001252872054465115, + 0.0014267067890614271, + 0.0014990087365731597, + 0.0016447542002424598, + 0.004769907798618078, + 0.004024006426334381, + 0.002394614275544882, + 0.001962382812052965, + 0.0009972992120310664, + 0.0023792290594428778, + 0.001304330537095666, + 0.0013110098661854863, + 0.0023957511875778437, + 0.003675991902127862, + 0.0010937659535557032, + 0.0010350479278713465, + 0.008389540947973728, + 0.006122208200395107, + 0.007589586544781923, + 0.007716144900768995, + 0.006941923871636391, + 0.006735268514603376, + 0.00447157071903348, + 0.00662847189232707, + 0.001176175195723772, + 0.000933646981138736, + 0.0010853043058887124, + 0.0017092875204980373, + 0.003322698874399066, + 0.0027944352477788925, + 0.001209521316923201, + 0.001208368455991149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0067337024956941605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures a pair of hands meticulously arranging a collection of vintage hardcover books. These hands, belonging to a person wearing the sleeves of a cozy forest green knit sweater, hold a thick hardcover bound in worn crimson leather. The book's spine features a faint, faded gilt pattern, and the pages are a warm, decked ivory. With steady, deliberate precision, the hands guide the book onto a dark, polished mahogany bookshelf already crowded with similar antique volumes. The camera follows the person as they gently and carefully slide the crimson book into the perfectly sized, narrow gap between the aquamarine and dark brown volumes. The book slides smoothly into place until its spine aligns perfectly with the neighboring books, a movement that conveys a deep sense of satisfaction and order. The scene is bathed in the warm, soft glow of an unseen lamp, which casts gentle highlights across the richly textured leather spines and the rich grain of the wood, creating deep, soft shadows. The overall atmosphere is one of tranquil focus and quiet reverence, a peaceful moment dedicated to the love of literature.", + "seed": 7235, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_8/index_159_seed_7235_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010575456591323019, + "vlm_logprob": -0.010575456591323019, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.010575456591323019, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011250981129705906 + ], + "opposite_logprob_tokens": [ + -5.636250972747803 + ], + "logprob_reward": -0.011250981129705906, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.011250981129705906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05132848396897316, + 0.09392698854207993, + 0.17142580449581146, + 0.3112447261810303, + 0.02883405052125454, + 0.02782866545021534, + 0.05692882463335991, + 0.20108073949813843, + 0.023955196142196655, + 0.0334465391933918 + ], + "credit_map": [ + 0.004425895866006613, + 0.0054335095919668674, + 0.009861283004283905, + 0.008106694556772709, + 0.0038997826632112265, + 0.00661986181512475, + 0.008066724985837936, + 0.004914728458970785, + 0.02075345814228058, + 0.006622930057346821, + 0.005692491307854652, + 0.01128863263875246, + 0.010009252466261387, + 0.01001992542296648, + 0.021214082837104797, + 0.008326215669512749, + 0.017453424632549286, + 0.04408656805753708, + 0.0038170393090695143, + 0.00441118935123086, + 0.006668916903436184, + 0.06429876387119293, + 0.009877492673695087, + 0.02081240899860859, + 0.10651447623968124, + 0.011569978669285774, + 0.12031053751707077, + 0.02179023064672947, + 0.011373245157301426, + 0.01470883097499609, + 0.0150278490036726, + 0.009949575178325176, + 0.0020675132982432842, + 0.003370327176526189, + 0.0031578626949340105, + 0.005812705494463444, + 0.002839155960828066, + 0.0027863592840731144, + 0.005100532900542021, + 0.0036995934788137674, + 0.0017114939400926232, + 0.0029614786617457867, + 0.0033815186470746994, + 0.003293606685474515, + 0.0025994705501943827, + 0.005218027625232935, + 0.004000042099505663, + 0.004663029219955206, + 0.0051107099279761314, + 0.010024311020970345, + 0.006382126826792955, + 0.007544122636318207, + 0.005988209508359432, + 0.006016930099576712, + 0.0071398536674678326, + 0.008722559548914433, + 0.0076525090262293816, + 0.010689496994018555, + 0.011002237908542156, + 0.02755877561867237, + 0.01906760223209858, + 0.02427653782069683, + 0.06905338913202286, + 0.03178020566701889, + 0.0018556449795141816, + 0.0015456352848559618, + 0.004879267420619726, + 0.00224307575263083, + 0.0013160948874428868, + 0.004427341278642416, + 0.0028544634114950895, + 0.0048336731269955635, + 0.0023227520287036896, + 0.0021648190449923277, + 0.004035064019262791, + 0.008628591895103455, + 0.0023238989524543285, + 0.00351126492023468, + 0.005381649825721979, + 0.005078497808426619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 2 + ], + "selected_reward": -0.011250981129705906 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004106779582798481 + ], + "opposite_logprob_tokens": [ + -5.5041069984436035 + ], + "logprob_reward": -0.004106779582798481, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004106779582798481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1063615009188652, + 0.19301949441432953, + 0.28691360354423523, + 0.10312570631504059, + 0.053135693073272705, + 0.052145037800073624, + 0.053203001618385315, + 0.1016295999288559, + 0.03002973273396492, + 0.02043658308684826 + ], + "credit_map": [ + 0.008574184030294418, + 0.007933879271149635, + 0.016998736187815666, + 0.015366138890385628, + 0.008972479030489922, + 0.015136126428842545, + 0.01589144952595234, + 0.01748851127922535, + 0.050683535635471344, + 0.012148595415055752, + 0.0142848901450634, + 0.02696315385401249, + 0.009906978346407413, + 0.018964070826768875, + 0.04511014372110367, + 0.014958127401769161, + 0.02644697017967701, + 0.09088589996099472, + 0.006190262734889984, + 0.008151909336447716, + 0.012111308053135872, + 0.10067976266145706, + 0.012225357815623283, + 0.030222125351428986, + 0.02044171467423439, + 0.007390078157186508, + 0.028860928490757942, + 0.007034002803266048, + 0.016996148973703384, + 0.008520499803125858, + 0.009284588508307934, + 0.004597746301442385, + 0.004037174396216869, + 0.006291321478784084, + 0.006660834886133671, + 0.010079292580485344, + 0.005290288012474775, + 0.00583958113566041, + 0.007661044597625732, + 0.007276155054569244, + 0.0036538615822792053, + 0.005264077335596085, + 0.005708251614123583, + 0.009162388741970062, + 0.0059731085784733295, + 0.0068904804065823555, + 0.00852882955223322, + 0.006964037194848061, + 0.0037874062545597553, + 0.008186629973351955, + 0.004947028122842312, + 0.006398159544914961, + 0.007492120377719402, + 0.005808915942907333, + 0.008237856440246105, + 0.00834488496184349, + 0.004038275685161352, + 0.005498957354575396, + 0.0053995633497834206, + 0.01347373891621828, + 0.009853231720626354, + 0.013337645679712296, + 0.03380269557237625, + 0.01622549444437027, + 0.002968277782201767, + 0.002544061979278922, + 0.0028806221671402454, + 0.00294097070582211, + 0.0016869272803887725, + 0.007363128010183573, + 0.0028455746360123158, + 0.0068001700565218925, + 0.0023221024312078953, + 0.001588687882758677, + 0.0036560208536684513, + 0.0032530315220355988, + 0.0009119778405874968, + 0.0024749035947024822, + 0.0028115317691117525, + 0.0034183275420218706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.004106779582798481 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029812784865498543 + ], + "opposite_logprob_tokens": [ + -3.529812812805176 + ], + "logprob_reward": -0.029812784865498543, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029812784865498543, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03445933386683464, + 0.013652733527123928, + 0.14510369300842285, + 0.6435387134552002, + 0.026519939303398132, + 0.02428353764116764, + 0.03150099143385887, + 0.059703465551137924, + 0.011019925586879253, + 0.010217598639428616 + ], + "credit_map": [ + 0.0029786929953843355, + 0.0032968255691230297, + 0.005309916101396084, + 0.004851578269153833, + 0.004514158703386784, + 0.0058884089812636375, + 0.004203784745186567, + 0.0034159692004323006, + 0.002028495306149125, + 0.0009949076920747757, + 0.0014496177900582552, + 0.00164248317014426, + 0.0008643519831821322, + 0.0022869573440402746, + 0.0022226341534405947, + 0.002163286553695798, + 0.013423603028059006, + 0.04608030617237091, + 0.002960849553346634, + 0.004137280397117138, + 0.006185556761920452, + 0.0508413165807724, + 0.006248506251722574, + 0.015226282179355621, + 0.21921198070049286, + 0.02356613427400589, + 0.24662888050079346, + 0.04461820796132088, + 0.027010159566998482, + 0.03140321001410484, + 0.03040960431098938, + 0.02069055289030075, + 0.0016669825417920947, + 0.003379925386980176, + 0.0024786449503153563, + 0.0034049854148179293, + 0.004687783308327198, + 0.00276121124625206, + 0.00423377426341176, + 0.003906630910933018, + 0.0014232350513339043, + 0.0022755952086299658, + 0.0023602391593158245, + 0.0029363983776420355, + 0.00286305695772171, + 0.005158489570021629, + 0.0027932743541896343, + 0.004473247565329075, + 0.002663258695974946, + 0.005897727329283953, + 0.0025947997346520424, + 0.003316961694508791, + 0.004018677864223719, + 0.003165755420923233, + 0.005080080591142178, + 0.004763731732964516, + 0.002639500657096505, + 0.0031807066407054663, + 0.003526300424709916, + 0.008200785145163536, + 0.005403890274465084, + 0.007373135536909103, + 0.019951224327087402, + 0.009427919052541256, + 0.0008444159757345915, + 0.0006886416813358665, + 0.0011900487588718534, + 0.0010485603706911206, + 0.0006650311988778412, + 0.002455910434946418, + 0.00202730530872941, + 0.0021000122651457787, + 0.000730124011170119, + 0.0006557252490893006, + 0.0009804356377571821, + 0.0028112982399761677, + 0.0008520169649273157, + 0.0010617190273478627, + 0.0016773928655311465, + 0.0014488855376839638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 7 + ], + "selected_reward": -0.029812784865498543 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036140859592705965 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036140859592705965, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036140859592705965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0772484615445137, + 0.0183259230107069, + 0.2833124101161957, + 0.32069629430770874, + 0.035686518996953964, + 0.04578731954097748, + 0.07606597244739532, + 0.0601297989487648, + 0.06598909944295883, + 0.016758183017373085 + ], + "credit_map": [ + 0.007101586554199457, + 0.008328537456691265, + 0.01214486826211214, + 0.012039977125823498, + 0.006203918252140284, + 0.011244836263358593, + 0.011149856261909008, + 0.009034878574311733, + 0.0019044035580009222, + 0.0026315865106880665, + 0.0017764955991879106, + 0.0027355903293937445, + 0.0010889485711231828, + 0.003263742895796895, + 0.00265607051551342, + 0.0022690834011882544, + 0.025848623365163803, + 0.09545090794563293, + 0.00639477651566267, + 0.008192389272153378, + 0.011909504421055317, + 0.09707750380039215, + 0.01075387466698885, + 0.027684837579727173, + 0.10342539846897125, + 0.013120251707732677, + 0.1266840398311615, + 0.02346920594573021, + 0.012145265005528927, + 0.013893956318497658, + 0.01741470769047737, + 0.010543475858867168, + 0.003572076791897416, + 0.006471090018749237, + 0.003047293284907937, + 0.004095390439033508, + 0.004553539678454399, + 0.004054811783134937, + 0.0045472136698663235, + 0.0053451042622327805, + 0.00438445620238781, + 0.006348531227558851, + 0.005386394448578358, + 0.006575051229447126, + 0.004175084643065929, + 0.006690801586955786, + 0.0062285615131258965, + 0.005998443812131882, + 0.009125991724431515, + 0.011007989756762981, + 0.00942896492779255, + 0.007882820442318916, + 0.010712443850934505, + 0.008717324584722519, + 0.007789305876940489, + 0.011401129886507988, + 0.0034583737142384052, + 0.00685592507943511, + 0.004114506766200066, + 0.006959929130971432, + 0.004729659296572208, + 0.011859063059091568, + 0.012036527507007122, + 0.010115813463926315, + 0.0030685949604958296, + 0.0026450343430042267, + 0.004497244954109192, + 0.008502530865371227, + 0.004517217632383108, + 0.01266599353402853, + 0.02026335708796978, + 0.009829126298427582, + 0.00195585610345006, + 0.001487326924689114, + 0.0017001566011458635, + 0.003451081458479166, + 0.001250746543519199, + 0.002194332191720605, + 0.0023484996054321527, + 0.0023701840545982122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0036140859592705965 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are a pair of hands visible arranging vintage hardcover books?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the person wearing a forest green knit sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the book being handled bound in worn crimson leather with a faded gilt pattern on the spine?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera follow the hands as the book is slid into a narrow gap between aquamarine and dark brown volumes on a mahogany bookshelf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm, soft light source casting gentle highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092651419341564 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092651419341564, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092651419341564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08609984070062637, + 0.019323935732245445, + 0.07066000252962112, + 0.44400307536125183, + 0.03495465964078903, + 0.024190623313188553, + 0.025984155014157295, + 0.06905334442853928, + 0.03353371098637581, + 0.19219662249088287 + ], + "credit_map": [ + 0.005028648767620325, + 0.0056219203397631645, + 0.00857116375118494, + 0.012893824838101864, + 0.008890549652278423, + 0.020210476592183113, + 0.01152641698718071, + 0.013356832787394524, + 0.0019423927878960967, + 0.0013885276857763529, + 0.0025438570883125067, + 0.002805766649544239, + 0.0024518296122550964, + 0.002190682105720043, + 0.0030145386699587107, + 0.0029863431118428707, + 0.006720315665006638, + 0.020479939877986908, + 0.0035444917157292366, + 0.002790135098621249, + 0.003156582824885845, + 0.02309199422597885, + 0.0038148772437125444, + 0.007061665877699852, + 0.15145710110664368, + 0.016163907945156097, + 0.17473116517066956, + 0.03182545304298401, + 0.014067146927118301, + 0.019953418523073196, + 0.021307416260242462, + 0.014497475698590279, + 0.0023725018836557865, + 0.004692354239523411, + 0.003594875568524003, + 0.007268494926393032, + 0.003493409138172865, + 0.002997988136485219, + 0.006119444500654936, + 0.004415589850395918, + 0.001846491009928286, + 0.0023852719459682703, + 0.0029344551730901003, + 0.0033055883832275867, + 0.0023967870511114597, + 0.00446103373542428, + 0.003808688372373581, + 0.003052307525649667, + 0.0028457222506403923, + 0.005075599532574415, + 0.0032677152194082737, + 0.0032496238127350807, + 0.0024444113951176405, + 0.002722355769947171, + 0.002881822641938925, + 0.0034969050902873278, + 0.007846115157008171, + 0.008274313993752003, + 0.007035719696432352, + 0.008664419874548912, + 0.006029604002833366, + 0.009182586334645748, + 0.011071914806962013, + 0.010948678478598595, + 0.0025577954947948456, + 0.0022434901911765337, + 0.0034720709081739187, + 0.005412808153778315, + 0.0035660110879689455, + 0.005764279048889875, + 0.006259080953896046, + 0.004258172120898962, + 0.013787593692541122, + 0.011189336888492107, + 0.028157178312540054, + 0.043384771794080734, + 0.011103074066340923, + 0.0211884006857872, + 0.03217191621661186, + 0.03121434524655342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.004092651419341564 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_9/index_tensor([3895, 3895])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_9/index_tensor([3895, 3895])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3aab8ec953e0850a6c82b4c45d4e97bb2513504f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_9/index_tensor([3895, 3895])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, sideways tracking shot, viewed from a slightly upward angle, follows a maglev train. Its streamlined body, gleaming with a pearly white metallic sheen, silently and smoothly hovers above the monorail track, maintaining a discernible gap of several centimeters from the concrete track below. At dusk, the train zips through a futuristic urban jungle at breakneck speed. The towering skyscrapers and intricate network of air transportation are stretched into horizontal, iridescent streaks and blurred outlines by the immense relative speed. The setting sun paints the sky a warm orange-gold glow, intertwined with the cool neon lights gradually illuminating the city, and the light reflects off the sleek exterior of the train in a fluid, shimmering pattern. Through the train's expansive, dark-toned panoramic windows, interior details are barely visible, reflecting only the fleeting light from outside. The entire scene is imbued with a powerful sense of speed and futuristic aesthetic. The train's silent gliding motion contrasts sharply with the blurred motion of the background, creating an atmosphere of high-tech, efficient, and slightly alienated tranquility.", + "seed": 30299, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_9/index_3895_seed_30299_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03020472815260291, + "vlm_logprob": -0.03020472815260291, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.03020472815260291, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028581012040376663 + ], + "opposite_logprob_tokens": [ + -4.153581142425537 + ], + "logprob_reward": -0.028581012040376663, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.028581012040376663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2158455103635788, + 0.06329800188541412, + 0.19268691539764404, + 0.060926903039216995, + 0.04199676588177681, + 0.02310933731496334, + 0.039352644234895706, + 0.12176112830638885, + 0.036863744258880615, + 0.20415906608104706 + ], + "credit_map": [ + 0.02732815593481064, + 0.0406854972243309, + 0.012146085500717163, + 0.010768743231892586, + 0.04102697968482971, + 0.04177303612232208, + 0.01691179908812046, + 0.025205200538039207, + 0.010161430574953556, + 0.011275564320385456, + 0.00959863793104887, + 0.005472117103636265, + 0.00816297996789217, + 0.006807313300669193, + 0.006650387309491634, + 0.0051695723086595535, + 0.01808185689151287, + 0.024732453748583794, + 0.01185325626283884, + 0.016165975481271744, + 0.041673220694065094, + 0.026681458577513695, + 0.03255917876958847, + 0.020939528942108154, + 0.008046801201999187, + 0.006991636473685503, + 0.0062258606776595116, + 0.004595456179231405, + 0.011679095216095448, + 0.008871622383594513, + 0.007713547442108393, + 0.006802877876907587, + 0.006140344310551882, + 0.006872468627989292, + 0.003885281039401889, + 0.002872821409255266, + 0.007270138710737228, + 0.007155418861657381, + 0.004006938077509403, + 0.0037933553103357553, + 0.0032290220260620117, + 0.0025459693279117346, + 0.002463399199768901, + 0.0019051922718062997, + 0.003430949291214347, + 0.0034096804447472095, + 0.002530318684875965, + 0.0035948066506534815, + 0.0064156814478337765, + 0.006013012491166592, + 0.003738261526450515, + 0.004761683288961649, + 0.00440468592569232, + 0.0034460369497537613, + 0.006285081617534161, + 0.004288201220333576, + 0.011706099845468998, + 0.019479790702462196, + 0.011801491491496563, + 0.01601354032754898, + 0.019479097798466682, + 0.015904197469353676, + 0.015516504645347595, + 0.011860411614179611, + 0.004366894252598286, + 0.005021256394684315, + 0.0046044886112213135, + 0.005431968718767166, + 0.004645707551389933, + 0.006356803234666586, + 0.003071122569963336, + 0.0033655043225735426, + 0.018349597230553627, + 0.02201749198138714, + 0.024075882509350777, + 0.03791613504290581, + 0.03215494751930237, + 0.03190024942159653, + 0.019909340888261795, + 0.017835427075624466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.028581012040376663 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004771381616592 + ], + "opposite_logprob_tokens": [ + -2.695047616958618 + ], + "logprob_reward": -0.07004771381616592, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004771381616592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08277606964111328, + 0.10571389645338058, + 0.055996548384428024, + 0.19493141770362854, + 0.06000357121229172, + 0.03664146363735199, + 0.052043016999959946, + 0.08417189121246338, + 0.12592920660972595, + 0.2017928808927536 + ], + "credit_map": [ + 0.010233771987259388, + 0.025068391114473343, + 0.006256602704524994, + 0.004019765183329582, + 0.012252239510416985, + 0.006430148612707853, + 0.007113785948604345, + 0.011401360854506493, + 0.018466858193278313, + 0.025661995634436607, + 0.010667175054550171, + 0.010307788848876953, + 0.012683792039752007, + 0.010315905325114727, + 0.0094525758177042, + 0.008157803677022457, + 0.00667008338496089, + 0.007651497144252062, + 0.006498402450233698, + 0.00408947141841054, + 0.012258293107151985, + 0.009090125560760498, + 0.005038779228925705, + 0.004699896555393934, + 0.02129410207271576, + 0.020543090999126434, + 0.017952971160411835, + 0.01779935136437416, + 0.03711480274796486, + 0.02823457308113575, + 0.029316242784261703, + 0.02267627976834774, + 0.008182314224541187, + 0.011272059753537178, + 0.005925119388848543, + 0.004367970395833254, + 0.009577138349413872, + 0.009321589022874832, + 0.006336400751024485, + 0.005020979326218367, + 0.004051506984978914, + 0.0037527522072196007, + 0.0025594900362193584, + 0.0034854072146117687, + 0.005611087195575237, + 0.006781824864447117, + 0.0068878913298249245, + 0.0035115049686282873, + 0.008116628043353558, + 0.00662565091624856, + 0.004477682523429394, + 0.007056752219796181, + 0.005861790385097265, + 0.005053643137216568, + 0.008180798031389713, + 0.006670074071735144, + 0.009015722200274467, + 0.014579810202121735, + 0.007040660362690687, + 0.010542737320065498, + 0.012610726989805698, + 0.013394580222666264, + 0.00975500326603651, + 0.007232642266899347, + 0.01343152578920126, + 0.014755059033632278, + 0.0134193180128932, + 0.01767231896519661, + 0.015872659161686897, + 0.02512345090508461, + 0.01597577892243862, + 0.00967908464372158, + 0.01582002267241478, + 0.01412707008421421, + 0.007427836302667856, + 0.00858040526509285, + 0.07247525453567505, + 0.06752541661262512, + 0.00670267641544342, + 0.009134206920862198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.07004771381616592 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369487449526787 + ], + "opposite_logprob_tokens": [ + -3.4086949825286865 + ], + "logprob_reward": -0.03369487449526787, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369487449526787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06344775855541229, + 0.04687316715717316, + 0.629263699054718, + 0.025133347138762474, + 0.060175299644470215, + 0.030950604006648064, + 0.016037482768297195, + 0.04548824951052666, + 0.04991135746240616, + 0.03271900489926338 + ], + "credit_map": [ + 0.009486249648034573, + 0.008967486210167408, + 0.003016361966729164, + 0.003410704666748643, + 0.017774153500795364, + 0.009620406664907932, + 0.004088850691914558, + 0.007083550561219454, + 0.008276958018541336, + 0.012151312083005905, + 0.0028441508766263723, + 0.003208682406693697, + 0.005611742846667767, + 0.00735247228294611, + 0.003330834209918976, + 0.0040970128029584885, + 0.06068090349435806, + 0.08587107807397842, + 0.03947918489575386, + 0.049183331429958344, + 0.1521480828523636, + 0.10570576786994934, + 0.08053386211395264, + 0.05566153675317764, + 0.0037723735440522432, + 0.003327920800074935, + 0.003220344427973032, + 0.0028159741777926683, + 0.0030262458603829145, + 0.003611726453527808, + 0.0025804718025028706, + 0.002778291003778577, + 0.009948219172656536, + 0.013821479864418507, + 0.004491166677325964, + 0.003352636471390724, + 0.009605263359844685, + 0.010482318699359894, + 0.003471122123301029, + 0.005003099795430899, + 0.005199263337999582, + 0.003226739354431629, + 0.0023925527930259705, + 0.0024014534428715706, + 0.004073415417224169, + 0.005867050029337406, + 0.005103387404233217, + 0.00268674292601645, + 0.0019467821111902595, + 0.002209227532148361, + 0.0016359738074243069, + 0.0013630330795422196, + 0.0031396562699228525, + 0.0019577089697122574, + 0.001606313162483275, + 0.002178789349272847, + 0.004132268484681845, + 0.006475639063864946, + 0.0038439270574599504, + 0.006326759234070778, + 0.007278507575392723, + 0.005329255014657974, + 0.007735049352049828, + 0.004366850480437279, + 0.00596572132781148, + 0.00843746680766344, + 0.004937642253935337, + 0.008430241607129574, + 0.006657752674072981, + 0.007266663014888763, + 0.004813775885850191, + 0.003402098547667265, + 0.0036830168683081865, + 0.0034955679439008236, + 0.0022909606341272593, + 0.0024736057966947556, + 0.008471681736409664, + 0.007932932116091251, + 0.001802791841328144, + 0.0025684507563710213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.03369487449526787 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630132604390383 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630132604390383, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630132604390383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11182692646980286, + 0.08781231939792633, + 0.3195064067840576, + 0.09553924947977066, + 0.07150648534297943, + 0.03275126591324806, + 0.03733188658952713, + 0.06158909946680069, + 0.04812159016728401, + 0.13401471078395844 + ], + "credit_map": [ + 0.014663591980934143, + 0.029439063742756844, + 0.007388014812022448, + 0.005743843037635088, + 0.018048075959086418, + 0.014825400896370411, + 0.008474304340779781, + 0.013244640082120895, + 0.011805379763245583, + 0.011143877170979977, + 0.011338546872138977, + 0.0073249489068984985, + 0.012416571378707886, + 0.013263354077935219, + 0.010393109172582626, + 0.010126537643373013, + 0.022777970880270004, + 0.049759771674871445, + 0.014173662289977074, + 0.021681178361177444, + 0.0868532732129097, + 0.08332142233848572, + 0.02326684631407261, + 0.017672302201390266, + 0.01875099167227745, + 0.013166293501853943, + 0.00866030901670456, + 0.008592413738369942, + 0.013795142993330956, + 0.015222890302538872, + 0.011138062924146652, + 0.006213156506419182, + 0.01125232595950365, + 0.016963975504040718, + 0.008644361980259418, + 0.003261280246078968, + 0.011306731030344963, + 0.011333197355270386, + 0.003375422675162554, + 0.005369197577238083, + 0.006714431568980217, + 0.003047035075724125, + 0.0032600578851997852, + 0.002965780207887292, + 0.004806430544704199, + 0.005960442591458559, + 0.003932518418878317, + 0.0020645721815526485, + 0.003735936712473631, + 0.003553863149136305, + 0.0030391125474125147, + 0.004852444864809513, + 0.005108735989779234, + 0.005192645825445652, + 0.004269671626389027, + 0.007579478435218334, + 0.005049035418778658, + 0.010372515767812729, + 0.00649854214861989, + 0.007251656148582697, + 0.012701212428510189, + 0.006643804721534252, + 0.005921439733356237, + 0.007150903344154358, + 0.005012796260416508, + 0.0067389546893537045, + 0.005737684201449156, + 0.009021451696753502, + 0.005842775106430054, + 0.007528630085289478, + 0.004621532279998064, + 0.0036177716683596373, + 0.010714356787502766, + 0.00901872105896473, + 0.007308196742087603, + 0.008716530166566372, + 0.04490739852190018, + 0.04257886856794357, + 0.0052395532839000225, + 0.005531106144189835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.007630132604390383 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011069907806813717 + ], + "opposite_logprob_tokens": [ + -4.5110697746276855 + ], + "logprob_reward": -0.011069907806813717, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011069907806813717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2290235012769699, + 0.1655855029821396, + 0.13745760917663574, + 0.08541833609342575, + 0.06116090342402458, + 0.04341435804963112, + 0.05283001810312271, + 0.06846122443675995, + 0.06362214684486389, + 0.09302643686532974 + ], + "credit_map": [ + 0.028669310733675957, + 0.0400569885969162, + 0.013692791573703289, + 0.011850533075630665, + 0.03987802192568779, + 0.0416993722319603, + 0.021252453327178955, + 0.03192402049899101, + 0.021808920428156853, + 0.019687138497829437, + 0.027823269367218018, + 0.01495303213596344, + 0.02265314944088459, + 0.0210245531052351, + 0.02394452691078186, + 0.01369091123342514, + 0.01660868525505066, + 0.016930751502513885, + 0.01691022701561451, + 0.008773310109972954, + 0.020567238330841064, + 0.024093657732009888, + 0.020398132503032684, + 0.013175602070987225, + 0.013570939190685749, + 0.011282650753855705, + 0.009588336572051048, + 0.009980976581573486, + 0.014626794494688511, + 0.013179734349250793, + 0.005923975724726915, + 0.007264923304319382, + 0.008344369009137154, + 0.011112871579825878, + 0.006924918852746487, + 0.00450522918254137, + 0.010123219341039658, + 0.009918978437781334, + 0.004977170377969742, + 0.00525414664298296, + 0.007902259938418865, + 0.005636971909552813, + 0.003016020404174924, + 0.004473458044230938, + 0.009096652269363403, + 0.0062764412723481655, + 0.004432393237948418, + 0.0025801588781177998, + 0.00638306699693203, + 0.005924853030592203, + 0.005476589780300856, + 0.006219951435923576, + 0.007182552479207516, + 0.006064284592866898, + 0.0050622946582734585, + 0.010516423732042313, + 0.0065996889024972916, + 0.012318295426666737, + 0.006003053393214941, + 0.007847337983548641, + 0.011149288155138493, + 0.00784226879477501, + 0.00807493831962347, + 0.008626350201666355, + 0.013134381733834743, + 0.009488265961408615, + 0.00555088697001338, + 0.006970921531319618, + 0.008664767257869244, + 0.010925905779004097, + 0.004585698712617159, + 0.004301318433135748, + 0.011295575648546219, + 0.015810184180736542, + 0.008576186373829842, + 0.014484421350061893, + 0.009899858385324478, + 0.012047670781612396, + 0.008849462494254112, + 0.012063072063028812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011069907806813717 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, sideways tracking shot, viewed from a slightly upward angle, follows a maglev train. Its streamlined body, gleaming with a pearly white metallic sheen, silently and smoothly hovers above the monorail track, maintaining a discernible gap of several centimeters from the concrete track below. At dusk, the train zips through a futuristic urban jungle at breakneck speed. The towering skyscrapers and intricate network of air transportation are stretched into horizontal, iridescent streaks and blurred outlines by the immense relative speed. The setting sun paints the sky a warm orange-gold glow, intertwined with the cool neon lights gradually illuminating the city, and the light reflects off the sleek exterior of the train in a fluid, shimmering pattern. Through the train's expansive, dark-toned panoramic windows, interior details are barely visible, reflecting only the fleeting light from outside. The entire scene is imbued with a powerful sense of speed and futuristic aesthetic. The train's silent gliding motion contrasts sharply with the blurred motion of the background, creating an atmosphere of high-tech, efficient, and slightly alienated tranquility.", + "seed": 30300, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_9/index_3895_seed_30300_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0555643524043262, + "vlm_logprob": -0.0555643524043262, + "vlm_yes_no_margin": 3.550000023841858, + "vlm_token_credit": -0.0555643524043262, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012901408597826958 + ], + "opposite_logprob_tokens": [ + -4.887901306152344 + ], + "logprob_reward": -0.012901408597826958, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012901408597826958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44719764590263367, + 0.08322986960411072, + 0.0487496554851532, + 0.11154380440711975, + 0.06371352076530457, + 0.05432814359664917, + 0.03595980256795883, + 0.05896611139178276, + 0.04476388543844223, + 0.0515475794672966 + ], + "credit_map": [ + 0.04338778182864189, + 0.11248579621315002, + 0.04497675970196724, + 0.031944725662469864, + 0.023543767631053925, + 0.03381102904677391, + 0.12685059010982513, + 0.03019719198346138, + 0.009969808161258698, + 0.008635634556412697, + 0.0119554428383708, + 0.009879103861749172, + 0.008480273187160492, + 0.012766875326633453, + 0.009876404888927937, + 0.011666325852274895, + 0.00828162208199501, + 0.0129081467166543, + 0.0057338690385222435, + 0.004405038896948099, + 0.0038730448577553034, + 0.0042783585377037525, + 0.004255587235093117, + 0.0050139836966991425, + 0.014204342849552631, + 0.013945487327873707, + 0.01908961497247219, + 0.009623763151466846, + 0.009242254309356213, + 0.013769496232271194, + 0.015136062167584896, + 0.0165327750146389, + 0.01007823459804058, + 0.00888336542993784, + 0.012868653982877731, + 0.005263853818178177, + 0.003842927049845457, + 0.004852065350860357, + 0.011307335458695889, + 0.006617083679884672, + 0.005803156178444624, + 0.007941416464745998, + 0.007312128320336342, + 0.005802860949188471, + 0.005185049492865801, + 0.006548495031893253, + 0.0052956342697143555, + 0.010439407080411911, + 0.0047558085061609745, + 0.0048751370050013065, + 0.005434634629637003, + 0.004463785793632269, + 0.0032140957191586494, + 0.005202775821089745, + 0.0036864031571894884, + 0.0043271612375974655, + 0.006283685564994812, + 0.009588916786015034, + 0.003535178955644369, + 0.0037394368555396795, + 0.0028074639849364758, + 0.0032917761709541082, + 0.01934938319027424, + 0.01037027221173048, + 0.005375340115278959, + 0.0061164419166743755, + 0.004483493510633707, + 0.006359271705150604, + 0.003684275085106492, + 0.004648575559258461, + 0.008627243340015411, + 0.005469241179525852, + 0.006897910498082638, + 0.004410844761878252, + 0.005778591148555279, + 0.005549360066652298, + 0.003379079047590494, + 0.003626926802098751, + 0.008393223397433758, + 0.013511641882359982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.012901408597826958 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282505214214325 + ], + "opposite_logprob_tokens": [ + -2.017825126647949 + ], + "logprob_reward": -0.14282505214214325, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14282505214214325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3292470872402191, + 0.11245762556791306, + 0.10318902879953384, + 0.10689333826303482, + 0.1208423301577568, + 0.03258901834487915, + 0.05264699086546898, + 0.07872173190116882, + 0.013660619035363197, + 0.04975224286317825 + ], + "credit_map": [ + 0.0703633576631546, + 0.09973621368408203, + 0.029777254909276962, + 0.03229473531246185, + 0.023511607199907303, + 0.030355477705597878, + 0.020525222644209862, + 0.02268321067094803, + 0.013380608521401882, + 0.016706591472029686, + 0.015129541978240013, + 0.012146223336458206, + 0.01557300053536892, + 0.013562045991420746, + 0.013797550462186337, + 0.012162054888904095, + 0.013825367204844952, + 0.011998923495411873, + 0.009279259480535984, + 0.00682309502735734, + 0.007686163764446974, + 0.008137851022183895, + 0.03481566533446312, + 0.010622700676321983, + 0.015443404205143452, + 0.01510875578969717, + 0.014716912060976028, + 0.009232536889612675, + 0.010027029551565647, + 0.013636957854032516, + 0.01451736968010664, + 0.014210374094545841, + 0.023818962275981903, + 0.018560156226158142, + 0.015264366753399372, + 0.008044036105275154, + 0.007281571161001921, + 0.006974367890506983, + 0.027220094576478004, + 0.013678777031600475, + 0.0035073969047516584, + 0.004037807229906321, + 0.005854897201061249, + 0.0031108225230127573, + 0.002901907078921795, + 0.004849976394325495, + 0.003262137994170189, + 0.005064073018729687, + 0.010049832053482533, + 0.008225691504776478, + 0.007032965309917927, + 0.004570263437926769, + 0.005194219294935465, + 0.007556041702628136, + 0.004764773417264223, + 0.005253204144537449, + 0.005518224555999041, + 0.008267629891633987, + 0.0069816806353628635, + 0.005918641574680805, + 0.003924851771444082, + 0.003397966269403696, + 0.026622073724865913, + 0.018090667203068733, + 0.0017585863824933767, + 0.002396238036453724, + 0.0011556270765140653, + 0.0011244600173085928, + 0.0015383854042738676, + 0.0014712511328980327, + 0.0019248508615419269, + 0.0022912193089723587, + 0.005014204420149326, + 0.0030875294469296932, + 0.005414443556219339, + 0.002816235413774848, + 0.0036975264083594084, + 0.003264845348894596, + 0.01033927034586668, + 0.016118187457323074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.14282505214214325 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901972323656082 + ], + "opposite_logprob_tokens": [ + -2.464019775390625 + ], + "logprob_reward": -0.08901972323656082, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901972323656082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19207987189292908, + 0.12719714641571045, + 0.13890795409679413, + 0.1675572395324707, + 0.10496202856302261, + 0.030377710238099098, + 0.01803387701511383, + 0.08874006569385529, + 0.026855945587158203, + 0.10528817027807236 + ], + "credit_map": [ + 0.021364975720643997, + 0.01802685111761093, + 0.012267396785318851, + 0.015148717910051346, + 0.010981171391904354, + 0.015099815092980862, + 0.07516531646251678, + 0.02402561530470848, + 0.02075835131108761, + 0.017717823386192322, + 0.01796136423945427, + 0.015062856487929821, + 0.014147365465760231, + 0.013214030303061008, + 0.014891846105456352, + 0.01344351191073656, + 0.027082890272140503, + 0.034874267876148224, + 0.013857027515769005, + 0.012360958382487297, + 0.010810861364006996, + 0.011806701309978962, + 0.01392327155917883, + 0.014191968366503716, + 0.021187590435147285, + 0.025875484570860863, + 0.03808489069342613, + 0.015289170667529106, + 0.011581570841372013, + 0.016725091263651848, + 0.020430512726306915, + 0.01838291808962822, + 0.01829262264072895, + 0.016340823844075203, + 0.02063862606883049, + 0.007765721995383501, + 0.006188852712512016, + 0.008914239704608917, + 0.015422185882925987, + 0.011398948729038239, + 0.003142142901197076, + 0.004799501970410347, + 0.005209945607930422, + 0.002781766466796398, + 0.0023702667094767094, + 0.0034980690106749535, + 0.003264564322307706, + 0.005311452317982912, + 0.002244639443233609, + 0.0021687494590878487, + 0.0027889017947018147, + 0.002396806376054883, + 0.001473631476983428, + 0.0020777101162821054, + 0.0021372081246227026, + 0.0027462297584861517, + 0.010931371711194515, + 0.012878922745585442, + 0.009213141165673733, + 0.008057285100221634, + 0.0064241704531013966, + 0.007691361475735903, + 0.019464030861854553, + 0.01407978218048811, + 0.004881303291767836, + 0.00491436617448926, + 0.003977486398071051, + 0.002194491680711508, + 0.002107122214511037, + 0.0025531426072120667, + 0.002746518701314926, + 0.0034815152175724506, + 0.019613932818174362, + 0.006189119070768356, + 0.013887779787182808, + 0.007027728948742151, + 0.0058011324144899845, + 0.006677521392703056, + 0.013481739908456802, + 0.03260922059416771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.08901972323656082 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067384387366473675 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.0067384387366473675, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067384387366473675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.398446649312973, + 0.07256518304347992, + 0.11353962123394012, + 0.12187633663415909, + 0.1186518520116806, + 0.01354963332414627, + 0.021368125453591347, + 0.05171000212430954, + 0.015968458727002144, + 0.07232412695884705 + ], + "credit_map": [ + 0.03976273909211159, + 0.1124124750494957, + 0.030271198600530624, + 0.02239646390080452, + 0.012355592101812363, + 0.014731664210557938, + 0.1367916464805603, + 0.029724910855293274, + 0.013440174981951714, + 0.009707953780889511, + 0.007560540921986103, + 0.00965568795800209, + 0.0074744196608662605, + 0.010345438495278358, + 0.007162170019000769, + 0.00721879443153739, + 0.020021578297019005, + 0.008920409716665745, + 0.009787265211343765, + 0.005666075740009546, + 0.004839252680540085, + 0.007229788228869438, + 0.0457192026078701, + 0.0113560501486063, + 0.018971817567944527, + 0.020863886922597885, + 0.01241532526910305, + 0.008978702127933502, + 0.007527013309299946, + 0.009077624417841434, + 0.028034254908561707, + 0.016007713973522186, + 0.027321359142661095, + 0.01781095750629902, + 0.013038156554102898, + 0.006941847503185272, + 0.008206116035580635, + 0.0058877915143966675, + 0.02968859113752842, + 0.009757031686604023, + 0.001163219683803618, + 0.002434531692415476, + 0.0027270764112472534, + 0.0012041163863614202, + 0.0010671092895790935, + 0.001292870962060988, + 0.0013449682155624032, + 0.0023157401010394096, + 0.002892611315473914, + 0.0030245801899582148, + 0.0031191729940474033, + 0.0026276151183992624, + 0.001981752924621105, + 0.0029895601328462362, + 0.0022346784826368093, + 0.0024981554597616196, + 0.004293186590075493, + 0.004345993511378765, + 0.002799154492095113, + 0.0032200063578784466, + 0.002889984054490924, + 0.0027679589111357927, + 0.018822230398654938, + 0.012571493163704872, + 0.001834859256632626, + 0.002151406602934003, + 0.0024313898757100105, + 0.0018513589166104794, + 0.001452430384233594, + 0.0020472079049795866, + 0.002474302425980568, + 0.001725503709167242, + 0.009419228881597519, + 0.005687834229320288, + 0.008818898350000381, + 0.004972180817276239, + 0.0038897988852113485, + 0.004197645001113415, + 0.014558873139321804, + 0.020779669284820557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0067384387366473675 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026337139308452606 + ], + "opposite_logprob_tokens": [ + -3.651337146759033 + ], + "logprob_reward": -0.026337139308452606, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026337139308452606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4854246973991394, + 0.08190292865037918, + 0.062362123280763626, + 0.05247592180967331, + 0.12861578166484833, + 0.03905387967824936, + 0.04672998562455177, + 0.03565911948680878, + 0.01552234310656786, + 0.05225319787859917 + ], + "credit_map": [ + 0.04160327836871147, + 0.08730072528123856, + 0.03281690552830696, + 0.028681490570306778, + 0.016567489132285118, + 0.027720490470528603, + 0.2012404054403305, + 0.049493882805109024, + 0.018305759876966476, + 0.014411774463951588, + 0.007945048622786999, + 0.007956189103424549, + 0.00697130523622036, + 0.010945449583232403, + 0.007717621512711048, + 0.007649789564311504, + 0.00926517229527235, + 0.008156311698257923, + 0.006802861113101244, + 0.0038561951369047165, + 0.003178861690685153, + 0.004600990563631058, + 0.021292006596922874, + 0.005209722090512514, + 0.005257127340883017, + 0.004469903185963631, + 0.007510919123888016, + 0.005402380600571632, + 0.0044870758429169655, + 0.007102664560079575, + 0.007710311561822891, + 0.010535539127886295, + 0.03127973899245262, + 0.01672809012234211, + 0.011308498680591583, + 0.006910970434546471, + 0.0077526806853711605, + 0.005386214703321457, + 0.038756128400564194, + 0.010493465699255466, + 0.003440829459577799, + 0.00534061761572957, + 0.007799201179295778, + 0.004599025472998619, + 0.0028733403887599707, + 0.005162658169865608, + 0.003842185717076063, + 0.00599601911380887, + 0.007563690189272165, + 0.00873966421931982, + 0.005211652722209692, + 0.005318996962159872, + 0.004167662002146244, + 0.0072098528034985065, + 0.004186263307929039, + 0.004332202486693859, + 0.0022212823387235403, + 0.002226821845397353, + 0.006252805236726999, + 0.003333992324769497, + 0.0027770157903432846, + 0.004453178960829973, + 0.00722189387306571, + 0.007172128185629845, + 0.0021025307942181826, + 0.0020275041460990906, + 0.0027622159104794264, + 0.0018056626431643963, + 0.0013855244033038616, + 0.0016183446859940886, + 0.0021268492564558983, + 0.0016937104519456625, + 0.010505959391593933, + 0.004436546005308628, + 0.008443962782621384, + 0.004302804358303547, + 0.0031933614518493414, + 0.003628514939919114, + 0.005550887435674667, + 0.012191158719360828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.026337139308452606 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, sideways tracking shot, viewed from a slightly upward angle, follows a maglev train. Its streamlined body, gleaming with a pearly white metallic sheen, silently and smoothly hovers above the monorail track, maintaining a discernible gap of several centimeters from the concrete track below. At dusk, the train zips through a futuristic urban jungle at breakneck speed. The towering skyscrapers and intricate network of air transportation are stretched into horizontal, iridescent streaks and blurred outlines by the immense relative speed. The setting sun paints the sky a warm orange-gold glow, intertwined with the cool neon lights gradually illuminating the city, and the light reflects off the sleek exterior of the train in a fluid, shimmering pattern. Through the train's expansive, dark-toned panoramic windows, interior details are barely visible, reflecting only the fleeting light from outside. The entire scene is imbued with a powerful sense of speed and futuristic aesthetic. The train's silent gliding motion contrasts sharply with the blurred motion of the background, creating an atmosphere of high-tech, efficient, and slightly alienated tranquility.", + "seed": 30301, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_9/index_3895_seed_30301_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02752868300303817, + "vlm_logprob": -0.02752868300303817, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.02752868300303817, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0201713889837265 + ], + "opposite_logprob_tokens": [ + -4.270171165466309 + ], + "logprob_reward": -0.0201713889837265, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.0201713889837265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06018582358956337, + 0.1361600160598755, + 0.047320932149887085, + 0.12091769278049469, + 0.090938501060009, + 0.05156518146395683, + 0.08939140290021896, + 0.06620212644338608, + 0.1270538568496704, + 0.21026450395584106 + ], + "credit_map": [ + 0.007452746853232384, + 0.010468498803675175, + 0.007512158714234829, + 0.0043144491501152515, + 0.008141608908772469, + 0.007587192580103874, + 0.006908987183123827, + 0.007800180930644274, + 0.03968918323516846, + 0.013105612248182297, + 0.013572104275226593, + 0.012473858892917633, + 0.009476431645452976, + 0.007653370965272188, + 0.010283972136676311, + 0.02990548312664032, + 0.006061731372028589, + 0.005333274602890015, + 0.008226814679801464, + 0.006432054564356804, + 0.005169212818145752, + 0.0049140299670398235, + 0.005800392013043165, + 0.005383421666920185, + 0.021730884909629822, + 0.017473438754677773, + 0.011496089398860931, + 0.0124115701764822, + 0.01467838604003191, + 0.022971494123339653, + 0.00936170294880867, + 0.010794124566018581, + 0.013502420857548714, + 0.013927050866186619, + 0.007896467112004757, + 0.015099674463272095, + 0.009565179236233234, + 0.00976518727838993, + 0.012870777398347855, + 0.008311744779348373, + 0.007448792923241854, + 0.007318568881601095, + 0.00423320522531867, + 0.0059041171334683895, + 0.007343640550971031, + 0.005496568512171507, + 0.0068611204624176025, + 0.006959168706089258, + 0.01277380995452404, + 0.013887214474380016, + 0.010399673134088516, + 0.008556872606277466, + 0.012181350961327553, + 0.01247252244502306, + 0.009538349695503712, + 0.009581607766449451, + 0.009728405624628067, + 0.008911224082112312, + 0.006868769880384207, + 0.007201564963907003, + 0.008307495154440403, + 0.0080035999417305, + 0.00782453641295433, + 0.009356529451906681, + 0.03549006208777428, + 0.00822517741471529, + 0.0078398073092103, + 0.008463647216558456, + 0.010369259864091873, + 0.012987609021365643, + 0.008791113272309303, + 0.03488718345761299, + 0.048039354383945465, + 0.011193017475306988, + 0.009626282379031181, + 0.030290212482213974, + 0.0079914266243577, + 0.006045846734195948, + 0.08397123962640762, + 0.01310713030397892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.0201713889837265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901643216609955 + ], + "opposite_logprob_tokens": [ + -2.5790164470672607 + ], + "logprob_reward": -0.07901643216609955, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901643216609955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0713120624423027, + 0.2636983394622803, + 0.06350985169410706, + 0.2389342188835144, + 0.0940394327044487, + 0.056520432233810425, + 0.023769287392497063, + 0.05499714985489845, + 0.07640206068754196, + 0.05681717023253441 + ], + "credit_map": [ + 0.007848900742828846, + 0.013131442479789257, + 0.010518105700612068, + 0.006801525596529245, + 0.009539106860756874, + 0.00888760108500719, + 0.008225942961871624, + 0.0063594370149075985, + 0.04931759089231491, + 0.028044994920492172, + 0.021820373833179474, + 0.033545032143592834, + 0.0236069243401289, + 0.01404484175145626, + 0.017462434247136116, + 0.07585616409778595, + 0.008488981984555721, + 0.00545294675976038, + 0.011683489196002483, + 0.01061015110462904, + 0.006846156902611256, + 0.0068146223202347755, + 0.007024814374744892, + 0.006588691845536232, + 0.05790003016591072, + 0.024197496473789215, + 0.01170598715543747, + 0.00888758897781372, + 0.021263163536787033, + 0.08396563678979874, + 0.014035878702998161, + 0.016978425905108452, + 0.01967858150601387, + 0.015479031018912792, + 0.00865293387323618, + 0.01145891658961773, + 0.009957552887499332, + 0.007903225719928741, + 0.011414491571485996, + 0.009494698606431484, + 0.012523547746241093, + 0.010596312582492828, + 0.004707365296781063, + 0.005476696416735649, + 0.005591964349150658, + 0.004575992003083229, + 0.006706723012030125, + 0.006341828498989344, + 0.0029973010532557964, + 0.0028416302520781755, + 0.002030021511018276, + 0.0020928350277245045, + 0.0022582572419196367, + 0.00548798032104969, + 0.002833861391991377, + 0.003227399429306388, + 0.00871970597654581, + 0.006661381106823683, + 0.007207305636256933, + 0.007115856744349003, + 0.004809741862118244, + 0.0073445881716907024, + 0.005279991775751114, + 0.007858576253056526, + 0.020369824022054672, + 0.005561902653425932, + 0.005923044867813587, + 0.00653063552454114, + 0.005604229401797056, + 0.005913060158491135, + 0.005433841608464718, + 0.021065523847937584, + 0.013479439541697502, + 0.008374695666134357, + 0.006289900280535221, + 0.004655530210584402, + 0.00272703287191689, + 0.002641606843098998, + 0.014347662217915058, + 0.004301304463297129 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.07901643216609955 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075330898165703 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075330898165703, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075330898165703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04728959500789642, + 0.15221965312957764, + 0.021844714879989624, + 0.5689822435379028, + 0.021607577800750732, + 0.013185523450374603, + 0.015357386320829391, + 0.01958576776087284, + 0.08481085300445557, + 0.05511666089296341 + ], + "credit_map": [ + 0.008056961931288242, + 0.007257582154124975, + 0.004273501690477133, + 0.003525674808770418, + 0.008135771378874779, + 0.0059425407089293, + 0.004765922669321299, + 0.005331641528755426, + 0.04526941478252411, + 0.017550311982631683, + 0.014098873361945152, + 0.014954518526792526, + 0.013472222723066807, + 0.008332549594342709, + 0.012057854793965816, + 0.026483912020921707, + 0.002649974310770631, + 0.0021129529923200607, + 0.0032608110923320055, + 0.004569247364997864, + 0.002102023223415017, + 0.002458809409290552, + 0.0022670752368867397, + 0.002423821948468685, + 0.15600259602069855, + 0.04304487258195877, + 0.022807462140917778, + 0.018902814015746117, + 0.04309035837650299, + 0.22258174419403076, + 0.02610781602561474, + 0.036444637924432755, + 0.0045055849477648735, + 0.003222798928618431, + 0.0019220947287976742, + 0.0031055561266839504, + 0.002366930479183793, + 0.0017853082390502095, + 0.0026876796036958694, + 0.002011626260355115, + 0.001519414596259594, + 0.0021537591237574816, + 0.0013914714800193906, + 0.001311478321440518, + 0.001907023717649281, + 0.0014896182110533118, + 0.001281096483580768, + 0.0021316614001989365, + 0.0021879582200199366, + 0.001512251328676939, + 0.0017857021884992719, + 0.0015399703988805413, + 0.001693716156296432, + 0.002970016561448574, + 0.0015645109815523028, + 0.002103261649608612, + 0.0031233648769557476, + 0.0028192366007715464, + 0.002320757834240794, + 0.0024916992988437414, + 0.0020011875312775373, + 0.0023764593061059713, + 0.0021758077200502157, + 0.0022772569209337234, + 0.026136385276913643, + 0.006561691407114267, + 0.006486576050519943, + 0.007880859076976776, + 0.004146662075072527, + 0.004003957845270634, + 0.005501661915332079, + 0.024093065410852432, + 0.01270088367164135, + 0.003399643814191222, + 0.0035811320412904024, + 0.007494894787669182, + 0.0019333148375153542, + 0.0016802855534479022, + 0.020928949117660522, + 0.0033975662663578987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.011075330898165703 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098943900316954 + ], + "opposite_logprob_tokens": [ + -5.504098892211914 + ], + "logprob_reward": -0.004098943900316954, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098943900316954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09279637038707733, + 0.06533700972795486, + 0.03932361304759979, + 0.3544301688671112, + 0.11677635461091995, + 0.025673411786556244, + 0.06366907805204391, + 0.02939165197312832, + 0.07182324677705765, + 0.14077907800674438 + ], + "credit_map": [ + 0.014376718550920486, + 0.012365289032459259, + 0.009929920546710491, + 0.007952766492962837, + 0.010506306774914265, + 0.009908054955303669, + 0.014864820055663586, + 0.012892503291368484, + 0.019697019830346107, + 0.006523631978780031, + 0.004280674271285534, + 0.004825879354029894, + 0.005173852201551199, + 0.003824191400781274, + 0.005385332740843296, + 0.015626436099410057, + 0.005685724783688784, + 0.004793890751898289, + 0.005172389559447765, + 0.00945963617414236, + 0.0037897282745689154, + 0.004500150214880705, + 0.0030459612607955933, + 0.0028761362191289663, + 0.09377015382051468, + 0.027895499020814896, + 0.01608911156654358, + 0.01280296128243208, + 0.02572382241487503, + 0.13815656304359436, + 0.016973525285720825, + 0.023018551990389824, + 0.021995389834046364, + 0.01753525622189045, + 0.010879532434046268, + 0.018763504922389984, + 0.010931921191513538, + 0.0090903639793396, + 0.016522865742444992, + 0.011057526804506779, + 0.0028896373696625233, + 0.003988177049905062, + 0.002215041546151042, + 0.0026624128222465515, + 0.004212167579680681, + 0.002955595264211297, + 0.003349954029545188, + 0.0034004279877990484, + 0.010832034982740879, + 0.006353974808007479, + 0.006144896149635315, + 0.008720846846699715, + 0.005584462080150843, + 0.008254692889750004, + 0.009108168072998524, + 0.008670001290738583, + 0.007041455712169409, + 0.0020553458016365767, + 0.0026229405775666237, + 0.0022434296552091837, + 0.0023443480022251606, + 0.003057103371247649, + 0.0029363518115133047, + 0.007090677507221699, + 0.019629888236522675, + 0.009218134917318821, + 0.007415036670863628, + 0.004072152078151703, + 0.00440666638314724, + 0.003712899051606655, + 0.004274923820048571, + 0.019093548879027367, + 0.031123870983719826, + 0.0060338107869029045, + 0.007242127321660519, + 0.022644683718681335, + 0.005237480625510216, + 0.0031546736136078835, + 0.05698264762759209, + 0.008359790779650211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.004098943900316954 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023281319066882133 + ], + "opposite_logprob_tokens": [ + -3.7732813358306885 + ], + "logprob_reward": -0.023281319066882133, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023281319066882133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1156591847538948, + 0.09205856174230576, + 0.017874356359243393, + 0.1909329891204834, + 0.11201293766498566, + 0.15093548595905304, + 0.0473640151321888, + 0.029721485450863838, + 0.15075542032718658, + 0.0926855057477951 + ], + "credit_map": [ + 0.014693189412355423, + 0.019849685952067375, + 0.01405300572514534, + 0.008840915746986866, + 0.01610848680138588, + 0.015287165530025959, + 0.011414587497711182, + 0.015412143431603909, + 0.0265821386128664, + 0.012050066143274307, + 0.009829992428421974, + 0.01029727142304182, + 0.00816456601023674, + 0.0052920980378985405, + 0.007674307096749544, + 0.01216812152415514, + 0.002358092460781336, + 0.0020642136223614216, + 0.0028769243508577347, + 0.002448635408654809, + 0.0025231533218175173, + 0.0019763256423175335, + 0.0016664343420416117, + 0.0019605779089033604, + 0.045068640261888504, + 0.019886748865246773, + 0.012172712944447994, + 0.013148129917681217, + 0.018393471837043762, + 0.05260733515024185, + 0.01615997962653637, + 0.01349596306681633, + 0.019463378936052322, + 0.01924065686762333, + 0.009636841714382172, + 0.015468715690076351, + 0.012152351438999176, + 0.008973524905741215, + 0.015166433528065681, + 0.01191103644669056, + 0.03316580131649971, + 0.025018924847245216, + 0.01189043652266264, + 0.015919338911771774, + 0.015085121616721153, + 0.014093142002820969, + 0.01832997426390648, + 0.017432739958167076, + 0.007905085571110249, + 0.00443632435053587, + 0.004567560274153948, + 0.005631738808006048, + 0.004486294463276863, + 0.0045150150544941425, + 0.008274679072201252, + 0.007547313347458839, + 0.005565334111452103, + 0.003485392779111862, + 0.0034650349989533424, + 0.0032925736159086227, + 0.002152628730982542, + 0.003642301307991147, + 0.003216122044250369, + 0.00490209786221385, + 0.04630456492304802, + 0.009979220107197762, + 0.008594030514359474, + 0.008903597481548786, + 0.009669323451817036, + 0.009756284765899181, + 0.00951144564896822, + 0.048036955296993256, + 0.02252030558884144, + 0.008005277253687382, + 0.004929901100695133, + 0.009774127043783665, + 0.0038820537738502026, + 0.00490014161914587, + 0.03234454244375229, + 0.006329155061393976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.023281319066882133 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, sideways tracking shot, viewed from a slightly upward angle, follows a maglev train. Its streamlined body, gleaming with a pearly white metallic sheen, silently and smoothly hovers above the monorail track, maintaining a discernible gap of several centimeters from the concrete track below. At dusk, the train zips through a futuristic urban jungle at breakneck speed. The towering skyscrapers and intricate network of air transportation are stretched into horizontal, iridescent streaks and blurred outlines by the immense relative speed. The setting sun paints the sky a warm orange-gold glow, intertwined with the cool neon lights gradually illuminating the city, and the light reflects off the sleek exterior of the train in a fluid, shimmering pattern. Through the train's expansive, dark-toned panoramic windows, interior details are barely visible, reflecting only the fleeting light from outside. The entire scene is imbued with a powerful sense of speed and futuristic aesthetic. The train's silent gliding motion contrasts sharply with the blurred motion of the background, creating an atmosphere of high-tech, efficient, and slightly alienated tranquility.", + "seed": 30302, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000003/group_9/index_3895_seed_30302_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020424076449126007, + "vlm_logprob": -0.020424076449126007, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.020424076449126007, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01113286241889 + ], + "opposite_logprob_tokens": [ + -6.011132717132568 + ], + "logprob_reward": -0.01113286241889, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.01113286241889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1542627066373825, + 0.05184539780020714, + 0.12260853499174118, + 0.07331939041614532, + 0.08256574720144272, + 0.026435062289237976, + 0.05637340247631073, + 0.050501804798841476, + 0.1719362437725067, + 0.21015165746212006 + ], + "credit_map": [ + 0.02928338758647442, + 0.043504539877176285, + 0.01657705008983612, + 0.006421003956347704, + 0.030191944912075996, + 0.014147968031466007, + 0.0061426120810210705, + 0.007994216866791248, + 0.009837480261921883, + 0.010002348572015762, + 0.005114998202770948, + 0.0076209260150790215, + 0.007103443145751953, + 0.004778625443577766, + 0.004114184528589249, + 0.0032734000124037266, + 0.035084862262010574, + 0.01684139482676983, + 0.015595389530062675, + 0.010989700444042683, + 0.012984725646674633, + 0.013283253647387028, + 0.00941659789532423, + 0.008412613533437252, + 0.0159473717212677, + 0.01696605607867241, + 0.006911767181009054, + 0.00678591150790453, + 0.006299184635281563, + 0.0068811881355941296, + 0.0063403588719666, + 0.007187558338046074, + 0.016608016565442085, + 0.006768133025616407, + 0.005956587381660938, + 0.031128033995628357, + 0.0028728076722472906, + 0.007948813959956169, + 0.00715151010081172, + 0.004131852649152279, + 0.004684350453317165, + 0.0032117748633027077, + 0.004463728051632643, + 0.0032594730146229267, + 0.003146818606182933, + 0.0024174919817596674, + 0.002613063668832183, + 0.002638364676386118, + 0.008211325854063034, + 0.0088707460090518, + 0.008366549387574196, + 0.005824812222272158, + 0.004887421149760485, + 0.00496132206171751, + 0.007903946563601494, + 0.007347282953560352, + 0.0062021370977163315, + 0.011152945458889008, + 0.005821739323437214, + 0.003877397160977125, + 0.00806986540555954, + 0.005074377171695232, + 0.005171073600649834, + 0.005132273305207491, + 0.0291157029569149, + 0.02979469671845436, + 0.03408411890268326, + 0.017461145296692848, + 0.018406054005026817, + 0.018513090908527374, + 0.014062580652534962, + 0.010498874820768833, + 0.015607079491019249, + 0.02902454510331154, + 0.009339229203760624, + 0.047173723578453064, + 0.07577317953109741, + 0.017041977494955063, + 0.00959635991603136, + 0.006595582701265812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.01113286241889 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054971449077129364 + ], + "opposite_logprob_tokens": [ + -2.92997145652771 + ], + "logprob_reward": -0.054971449077129364, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054971449077129364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07471910864114761, + 0.09157243371009827, + 0.04945897310972214, + 0.037973351776599884, + 0.22820763289928436, + 0.017883040010929108, + 0.040105435997247696, + 0.04757022485136986, + 0.06696086376905441, + 0.34554895758628845 + ], + "credit_map": [ + 0.01013724971562624, + 0.011466084979474545, + 0.007517311256378889, + 0.006758809555321932, + 0.017674649134278297, + 0.007558021228760481, + 0.006384360603988171, + 0.0072226193733513355, + 0.021699240431189537, + 0.021277273073792458, + 0.006819550879299641, + 0.008158840239048004, + 0.014990563504397869, + 0.008682487532496452, + 0.005103338975459337, + 0.004841134883463383, + 0.009203455410897732, + 0.0072169615887105465, + 0.005516556557267904, + 0.007725536357611418, + 0.005130624398589134, + 0.005290774628520012, + 0.005492617376148701, + 0.003882448421791196, + 0.005361195653676987, + 0.005643250420689583, + 0.0036801358219236135, + 0.007358456961810589, + 0.00456962501630187, + 0.004726325627416372, + 0.004152373410761356, + 0.0024819904938340187, + 0.0231351088732481, + 0.036642637103796005, + 0.04444378986954689, + 0.02447277307510376, + 0.023425355553627014, + 0.028145600110292435, + 0.02596011757850647, + 0.021982258185744286, + 0.0033459626138210297, + 0.0024864666629582644, + 0.0033692808356136084, + 0.0022488178219646215, + 0.0018876626854762435, + 0.0014388865092769265, + 0.0016863243654370308, + 0.0014196375850588083, + 0.00472450302913785, + 0.004885807167738676, + 0.006879920605570078, + 0.005626354366540909, + 0.004416215233504772, + 0.0036330625880509615, + 0.00579648045822978, + 0.004143095575273037, + 0.005839521065354347, + 0.009142332710325718, + 0.004234992899000645, + 0.004448982421308756, + 0.00909081194549799, + 0.004457545466721058, + 0.005610542371869087, + 0.004745492246001959, + 0.010333309881389141, + 0.011380949057638645, + 0.01147914957255125, + 0.0055508678779006, + 0.009273363277316093, + 0.011133111082017422, + 0.004438784439116716, + 0.0033713355660438538, + 0.020353112369775772, + 0.08094706386327744, + 0.010276950895786285, + 0.05214087292551994, + 0.14157059788703918, + 0.02043968252837658, + 0.012113909237086773, + 0.007706749718636274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.054971449077129364 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096213262528181 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004096213262528181, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096213262528181, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17730467021465302, + 0.28520894050598145, + 0.07117287069559097, + 0.18221983313560486, + 0.09222563356161118, + 0.02386094070971012, + 0.03604407608509064, + 0.036952655762434006, + 0.0526895672082901, + 0.042320869863033295 + ], + "credit_map": [ + 0.025025827810168266, + 0.03778081759810448, + 0.026732154190540314, + 0.0100482739508152, + 0.04105314984917641, + 0.016713062301278114, + 0.0094770397990942, + 0.010474344715476036, + 0.09108442068099976, + 0.08166328072547913, + 0.017098458483815193, + 0.014329236000776291, + 0.03501737490296364, + 0.02583816647529602, + 0.01039915345609188, + 0.009778846055269241, + 0.015348543412983418, + 0.009155281819403172, + 0.009407022967934608, + 0.008075323887169361, + 0.007886329665780067, + 0.008184177801012993, + 0.007449749391525984, + 0.005666445475071669, + 0.028643151745200157, + 0.031346846371889114, + 0.02537860907614231, + 0.028535986319184303, + 0.015836676582694054, + 0.016503572463989258, + 0.017729904502630234, + 0.018245084211230278, + 0.012395897880196571, + 0.012641128152608871, + 0.01564677245914936, + 0.01883680187165737, + 0.008183588273823261, + 0.010071339085698128, + 0.00887263286858797, + 0.00557746971026063, + 0.0056778742000460625, + 0.0035558275412768126, + 0.004018723964691162, + 0.0024022667203098536, + 0.0028565709944814444, + 0.002130248351022601, + 0.001964492490515113, + 0.001254935166798532, + 0.005498844664543867, + 0.005226756911724806, + 0.00429556705057621, + 0.004217101261019707, + 0.004612901248037815, + 0.0049587637186050415, + 0.00384048605337739, + 0.0033936575055122375, + 0.004829654935747385, + 0.00787647720426321, + 0.0042769210413098335, + 0.003329432802274823, + 0.006200904492288828, + 0.004544877912849188, + 0.002565633738413453, + 0.0033287545666098595, + 0.00856759399175644, + 0.00707045616582036, + 0.008063310757279396, + 0.004786955192685127, + 0.00808163546025753, + 0.007297009229660034, + 0.005065819714218378, + 0.0037567871622741222, + 0.005071232095360756, + 0.005120374262332916, + 0.0030828306917101145, + 0.007506241090595722, + 0.007877841591835022, + 0.009099160321056843, + 0.002490379847586155, + 0.0020728090312331915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.004096213262528181 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640285581350327 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640285581350327, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640285581350327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19977496564388275, + 0.20015165209770203, + 0.04139626398682594, + 0.16044485569000244, + 0.14216189086437225, + 0.03290924057364464, + 0.0563071146607399, + 0.06330793350934982, + 0.05906277894973755, + 0.04448333755135536 + ], + "credit_map": [ + 0.02760763093829155, + 0.04841500520706177, + 0.018633641302585602, + 0.01841747760772705, + 0.033408429473638535, + 0.016454117372632027, + 0.023146679624915123, + 0.013691971078515053, + 0.06496546417474747, + 0.05622253194451332, + 0.009216582402586937, + 0.01096915453672409, + 0.01656368188560009, + 0.022270072251558304, + 0.009580938145518303, + 0.010363191366195679, + 0.007862884551286697, + 0.006139383185654879, + 0.004321027547121048, + 0.004571457859128714, + 0.0048016333021223545, + 0.004359039943665266, + 0.00499408645555377, + 0.004346748813986778, + 0.04078133776783943, + 0.03932662308216095, + 0.012800061143934727, + 0.011624358594417572, + 0.013925665058195591, + 0.016465308144688606, + 0.011824840679764748, + 0.013696650974452496, + 0.0268306415528059, + 0.013558577746152878, + 0.009268069639801979, + 0.05590785667300224, + 0.006454924587160349, + 0.012760472483932972, + 0.011873363517224789, + 0.005507986526936293, + 0.006403617560863495, + 0.004335231613367796, + 0.004523663781583309, + 0.004337163642048836, + 0.003895257832482457, + 0.0031016110442578793, + 0.0038367235101759434, + 0.002475967165082693, + 0.008655381388962269, + 0.009706678800284863, + 0.006932506803423166, + 0.005381153430789709, + 0.008894489146769047, + 0.006281562615185976, + 0.004432324320077896, + 0.006023013964295387, + 0.010807939805090427, + 0.011666388250887394, + 0.00500878831371665, + 0.005910803563892841, + 0.014260980300605297, + 0.006496095564216375, + 0.004394923336803913, + 0.004762004129588604, + 0.007004381623119116, + 0.009908568114042282, + 0.006344529800117016, + 0.005405534990131855, + 0.009686507284641266, + 0.009894924238324165, + 0.005222352687269449, + 0.0055959755554795265, + 0.005562718957662582, + 0.0065604946576058865, + 0.002804764313623309, + 0.0050798263400793076, + 0.006503034848719835, + 0.012460468336939812, + 0.002799424109980464, + 0.0027126027271151543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.008640285581350327 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a maglev train with a pearly white metallic sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the train hover above a concrete monorail track with a visible gap?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a futuristic urban environment at dusk with orange-gold sky and neon lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the background skyscrapers and air transportation rendered as horizontal iridescent streaks due to high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a sideways tracking shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279571905732155 + ], + "opposite_logprob_tokens": [ + -3.7732796669006348 + ], + "logprob_reward": -0.023279571905732155, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279571905732155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23792482912540436, + 0.13797205686569214, + 0.05509388819336891, + 0.03092314302921295, + 0.1615278720855713, + 0.04969953000545502, + 0.04192601144313812, + 0.08000016957521439, + 0.13257208466529846, + 0.07236044108867645 + ], + "credit_map": [ + 0.03067546710371971, + 0.07359683513641357, + 0.01793072000145912, + 0.008883011527359486, + 0.06967170536518097, + 0.016446786001324654, + 0.010461230762302876, + 0.01025904156267643, + 0.01994294486939907, + 0.020358553156256676, + 0.012691755779087543, + 0.024764882400631905, + 0.02154478058218956, + 0.018940038979053497, + 0.009689947590231895, + 0.010039142332971096, + 0.013182449154555798, + 0.007080322131514549, + 0.006404242943972349, + 0.006255242973566055, + 0.006492848042398691, + 0.006706566084176302, + 0.004736677277833223, + 0.004235534463077784, + 0.0036536273546516895, + 0.0049276575446128845, + 0.003843429032713175, + 0.005110539495944977, + 0.002470475621521473, + 0.002804168500006199, + 0.0041587394662201405, + 0.003954502288252115, + 0.030900852754712105, + 0.014397694729268551, + 0.00824756734073162, + 0.06592322885990143, + 0.006482087541371584, + 0.015552639961242676, + 0.014117863960564137, + 0.00590590899810195, + 0.01236611045897007, + 0.006747009698301554, + 0.007567443884909153, + 0.005235593765974045, + 0.0039391713216900826, + 0.0034811808727681637, + 0.0059372433461248875, + 0.004425772465765476, + 0.006678249221295118, + 0.008365288376808167, + 0.005345848388969898, + 0.0045523252338171005, + 0.005841321777552366, + 0.004540019202977419, + 0.003723950358107686, + 0.0028790051583200693, + 0.00895219761878252, + 0.015317339450120926, + 0.01038648933172226, + 0.006219898816198111, + 0.01387842744588852, + 0.008310266770422459, + 0.009473123587667942, + 0.007462413515895605, + 0.019071659073233604, + 0.028277922421693802, + 0.015272561460733414, + 0.012137707322835922, + 0.023952508345246315, + 0.01962086744606495, + 0.007664129137992859, + 0.006574710831046104, + 0.006267550401389599, + 0.01996229588985443, + 0.003250360954552889, + 0.00488393334671855, + 0.024220429360866547, + 0.008228020742535591, + 0.002623195294290781, + 0.0029246483463793993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.023279571905732155 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_0/index_tensor([9874, 9874])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_0/index_tensor([9874, 9874])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a866bd36915fe4e654ecbcfb64bb521ef6235c4d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_0/index_tensor([9874, 9874])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle shot, the camera panning steadily to the right, follows a massive, asymmetrical, geometric alien spacecraft as it glides with a majestic, silent speed over a rain-soaked futuristic metropolis. The vessel's hull is composed of an unknown, smooth, obsidian material; rain slides down its icy curves without leaving a trace. The hull is covered in intricate, luminous patterns that pulsate with a cold, cyan-blue light in a rhythmic, inhuman pattern, casting an eerie halo on the wet streets below. Below, skyscrapers pierce the sky like a metallic jungle, shrouded by a thick, dark cloud. Huge, translucent holographic billboards flicker and morph between the buildings, displaying dynamic, color-saturated commercials. As the spacecraft gracefully passed in front of a holographic billboard projecting a huge female face, its powerful energy field even caused slight distortion of the holographic image and data interference. The whole scene, with the contrast between the cold background of the rainy night and the warm glow of the billboard, created a strong sense of cyberpunk melancholy and technological sublimity.", + "seed": 23792, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_0/index_9874_seed_23792_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19952502753585577, + "vlm_logprob": -0.19952502753585577, + "vlm_yes_no_margin": 2.324999988079071, + "vlm_token_credit": -0.19952502753585577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5174247026443481 + ], + "opposite_logprob_tokens": [ + -1.0174247026443481 + ], + "logprob_reward": -0.5174247026443481, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.5174247026443481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5382665991783142, + 0.16472303867340088, + 0.01799948327243328, + 0.05245233327150345, + 0.009123356081545353, + 0.02567882090806961, + 0.0207888912409544, + 0.06294745951890945, + 0.05845506861805916, + 0.049564991146326065 + ], + "credit_map": [ + 0.027396008372306824, + 0.20254570245742798, + 0.013903670944273472, + 0.013032190501689911, + 0.03286173194646835, + 0.221758633852005, + 0.015515956096351147, + 0.011252651922404766, + 0.007745071314275265, + 0.06089431792497635, + 0.0051616583950817585, + 0.005862854421138763, + 0.06421138346195221, + 0.011190958321094513, + 0.005314831156283617, + 0.004341949708759785, + 0.0019369492074474692, + 0.0028116009198129177, + 0.0025609349831938744, + 0.0016379038570448756, + 0.002549871103838086, + 0.0023163328878581524, + 0.0016015705186873674, + 0.002584318397566676, + 0.0073921228758990765, + 0.010910702869296074, + 0.009424971416592598, + 0.003995717037469149, + 0.008273910731077194, + 0.004369624424725771, + 0.004701266996562481, + 0.0033840122632682323, + 0.0007700963760726154, + 0.0013178526423871517, + 0.0011686572106555104, + 0.001103831920772791, + 0.0013003213098272681, + 0.0014328998513519764, + 0.0008633823017589748, + 0.0011663137702271342, + 0.002019437961280346, + 0.0033530748914927244, + 0.003537277225404978, + 0.002587271388620138, + 0.004457592032849789, + 0.004508129321038723, + 0.0022902239579707384, + 0.002925812965258956, + 0.0013372161192819476, + 0.0014219378354027867, + 0.002670304151251912, + 0.002775107743218541, + 0.0017661056481301785, + 0.002348895650357008, + 0.004218578804284334, + 0.004250742495059967, + 0.003637815359979868, + 0.01870339922606945, + 0.02050812542438507, + 0.007038095034658909, + 0.00448556337505579, + 0.002622515195980668, + 0.00365032278932631, + 0.002301619155332446, + 0.005126586649566889, + 0.006160073447972536, + 0.007535871583968401, + 0.011610133573412895, + 0.007978931069374084, + 0.006205124780535698, + 0.00582050159573555, + 0.008017835207283497, + 0.004591592121869326, + 0.004325765650719404, + 0.007939236238598824, + 0.007123784627765417, + 0.007465903181582689, + 0.004884361755102873, + 0.006550537422299385, + 0.0066838073544204235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.5174247026443481 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135109543800354 + ], + "opposite_logprob_tokens": [ + -1.3135108947753906 + ], + "logprob_reward": -0.3135109543800354, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3135109543800354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4323137402534485, + 0.0288836732506752, + 0.016361886635422707, + 0.04595467448234558, + 0.01870594546198845, + 0.04027064889669418, + 0.05634332448244095, + 0.23793849349021912, + 0.07733748108148575, + 0.04589010402560234 + ], + "credit_map": [ + 0.023899516090750694, + 0.16521717607975006, + 0.011513547971844673, + 0.012801497243344784, + 0.02301046811044216, + 0.16814658045768738, + 0.014490035362541676, + 0.013234966434538364, + 0.0027684736996889114, + 0.004519811365753412, + 0.001953443046659231, + 0.002619999460875988, + 0.005052194464951754, + 0.0032120661344379187, + 0.0031635931227356195, + 0.005594093818217516, + 0.0020239714067429304, + 0.0019245174480602145, + 0.0021928749047219753, + 0.0015713840257376432, + 0.0027987216599285603, + 0.0017467867583036423, + 0.0016274950467050076, + 0.0024761357344686985, + 0.005777568556368351, + 0.00915141124278307, + 0.007010017056018114, + 0.004255901090800762, + 0.005651948507875204, + 0.005482363514602184, + 0.003929701633751392, + 0.004695769865065813, + 0.0015031754737719893, + 0.0027626592200249434, + 0.0026450701989233494, + 0.001810921123251319, + 0.0020836920011788607, + 0.0024864592123776674, + 0.0023921059910207987, + 0.003021863056346774, + 0.002649621106684208, + 0.006122146733105183, + 0.005512766540050507, + 0.003836692776530981, + 0.007178130559623241, + 0.007853897288441658, + 0.003819315694272518, + 0.0032980788964778185, + 0.004149402491748333, + 0.004388675093650818, + 0.007014209404587746, + 0.010244633071124554, + 0.00564524857327342, + 0.007546049542725086, + 0.008949256502091885, + 0.008405853994190693, + 0.014977102167904377, + 0.09546437114477158, + 0.07085306942462921, + 0.013492802157998085, + 0.015877453610301018, + 0.014481306076049805, + 0.006939064711332321, + 0.005853328388184309, + 0.00543177779763937, + 0.009638525545597076, + 0.013171017169952393, + 0.010264233686029911, + 0.011916332878172398, + 0.006929168477654457, + 0.00906792189925909, + 0.010918510146439075, + 0.003768486436456442, + 0.003961652051657438, + 0.007966847158968449, + 0.007465618662536144, + 0.005691973026841879, + 0.004193906206637621, + 0.0077194455079734325, + 0.0051221828907728195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.3135109543800354 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12710212171077728 + ], + "opposite_logprob_tokens": [ + -2.1271021366119385 + ], + "logprob_reward": -0.12710212171077728, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12710212171077728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28438758850097656, + 0.4089493155479431, + 0.015284949913620949, + 0.0238477885723114, + 0.008551564998924732, + 0.041445422917604446, + 0.044507574290037155, + 0.06478659063577652, + 0.05714346840977669, + 0.05109569802880287 + ], + "credit_map": [ + 0.014162319712340832, + 0.09449104219675064, + 0.008514275774359703, + 0.010088571347296238, + 0.017852094024419785, + 0.1074843630194664, + 0.015723420307040215, + 0.016071520745754242, + 0.01650986261665821, + 0.16101452708244324, + 0.0064979298040270805, + 0.00710403174161911, + 0.17210783064365387, + 0.02385554276406765, + 0.009842723608016968, + 0.012016914784908295, + 0.0013908343389630318, + 0.002192659070715308, + 0.0021931431256234646, + 0.0018625683151185513, + 0.002082805149257183, + 0.0018374859355390072, + 0.0020085503347218037, + 0.0017169044585898519, + 0.0023538332898169756, + 0.005017720628529787, + 0.0028689319733530283, + 0.0021415394730865955, + 0.004273991100490093, + 0.0031299982219934464, + 0.001886766403913498, + 0.0021750093437731266, + 0.0006634924211539328, + 0.0008365329122170806, + 0.0012029263889417052, + 0.0008753745350986719, + 0.0007593370391987264, + 0.0014150225324556231, + 0.001260195393115282, + 0.0015386845916509628, + 0.002661274978891015, + 0.0051013692282140255, + 0.005987100303173065, + 0.004720540717244148, + 0.006790534593164921, + 0.007533798925578594, + 0.0042481133714318275, + 0.004402694292366505, + 0.0031507001258432865, + 0.0031708895694464445, + 0.007024930324405432, + 0.006538582034409046, + 0.004206071142107248, + 0.005834363400936127, + 0.00806775875389576, + 0.006514286622405052, + 0.004776530433446169, + 0.025738369673490524, + 0.017117053270339966, + 0.0037161430809646845, + 0.004445319529622793, + 0.004315666388720274, + 0.0019900663755834103, + 0.0026874509640038013, + 0.004052902106195688, + 0.006079354789108038, + 0.007730685640126467, + 0.009527342393994331, + 0.007505037356168032, + 0.006304689683020115, + 0.009135227650403976, + 0.006808233913034201, + 0.004534515086561441, + 0.00362125295214355, + 0.01114723552018404, + 0.010302980430424213, + 0.004281728994101286, + 0.004543186631053686, + 0.007264294195920229, + 0.005400505382567644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.12710212171077728 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979274045675993 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.00979274045675993, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979274045675993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40551379323005676, + 0.3597996234893799, + 0.01349126547574997, + 0.018476927652955055, + 0.020666323602199554, + 0.053052693605422974, + 0.01911863498389721, + 0.027350790798664093, + 0.03585634008049965, + 0.04667360335588455 + ], + "credit_map": [ + 0.020581647753715515, + 0.14947247505187988, + 0.010441568680107594, + 0.008812161162495613, + 0.026994815096259117, + 0.16907934844493866, + 0.011975848115980625, + 0.008155900985002518, + 0.013971705920994282, + 0.1408420205116272, + 0.005385808181017637, + 0.005385674070566893, + 0.15340660512447357, + 0.021947074681520462, + 0.008664299733936787, + 0.01019640825688839, + 0.0015619284240528941, + 0.0021578604355454445, + 0.002111495938152075, + 0.0014399500796571374, + 0.0020334997680038214, + 0.001228806097060442, + 0.0012530769454315305, + 0.0017046465072780848, + 0.0019963125232607126, + 0.003249292029067874, + 0.003330305917188525, + 0.0019004341447725892, + 0.0022643243428319693, + 0.0024564892519265413, + 0.0013764677569270134, + 0.0019032997079193592, + 0.0018182243220508099, + 0.002177328336983919, + 0.003798404475674033, + 0.0025184806436300278, + 0.0025411886163055897, + 0.004435471259057522, + 0.0016666280571371317, + 0.0017105971928685904, + 0.0037245166022330523, + 0.007073515560477972, + 0.007005617953836918, + 0.005811135284602642, + 0.009889512322843075, + 0.008012128993868828, + 0.0058569773100316525, + 0.0056792874820530415, + 0.0016408769879490137, + 0.00154493551235646, + 0.003145736176520586, + 0.0021937056444585323, + 0.002020103158429265, + 0.0024226855020970106, + 0.0037067506927996874, + 0.0024438402615487576, + 0.0019098194316029549, + 0.005684664472937584, + 0.008720179088413715, + 0.0038609083276242018, + 0.0015144553035497665, + 0.002002993831411004, + 0.0017103762365877628, + 0.001947392476722598, + 0.002690140623599291, + 0.003686752635985613, + 0.005697745829820633, + 0.005160380154848099, + 0.007502114400267601, + 0.0034464483615010977, + 0.003775712102651596, + 0.0038970441091805696, + 0.0032669990323483944, + 0.003721353830769658, + 0.009086942300200462, + 0.006945242173969746, + 0.005935991648584604, + 0.0048155756667256355, + 0.007653183303773403, + 0.005248312372714281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.00979274045675993 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794618487358093 + ], + "opposite_logprob_tokens": [ + -3.529794692993164 + ], + "logprob_reward": -0.029794618487358093, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794618487358093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4707306921482086, + 0.24540725350379944, + 0.026053376495838165, + 0.029761765152215958, + 0.011974905617535114, + 0.019954198971390724, + 0.04060748219490051, + 0.042546194046735764, + 0.07237830013036728, + 0.040585748851299286 + ], + "credit_map": [ + 0.025290755555033684, + 0.1758487969636917, + 0.011098632588982582, + 0.011726884171366692, + 0.027634354308247566, + 0.19219252467155457, + 0.0166795551776886, + 0.010259218513965607, + 0.009454168379306793, + 0.09571963548660278, + 0.004792443476617336, + 0.005824009422212839, + 0.10382969677448273, + 0.015347078442573547, + 0.0064270999282598495, + 0.0040131318382918835, + 0.002242486458271742, + 0.0036090395879000425, + 0.0039412179030478, + 0.0027309181168675423, + 0.00371160707436502, + 0.0026733872946351767, + 0.00285935471765697, + 0.00428536394611001, + 0.00539380731061101, + 0.00407373858615756, + 0.005376589950174093, + 0.0037604065146297216, + 0.0030390587635338306, + 0.0032602085266262293, + 0.0026054764166474342, + 0.0022524779196828604, + 0.0010155761847272515, + 0.0014728652313351631, + 0.0022406659554690123, + 0.001092693186365068, + 0.0014346850803121924, + 0.00198237132281065, + 0.0011032270267605782, + 0.001632821629755199, + 0.0015078906435519457, + 0.001840189564973116, + 0.0023618615232408047, + 0.002609589137136936, + 0.002273651072755456, + 0.0030058848205953836, + 0.003151021897792816, + 0.003204111475497484, + 0.0028573216404765844, + 0.0030743598472326994, + 0.005689279176294804, + 0.005979636684060097, + 0.0038038974162191153, + 0.0040198517963290215, + 0.00868806429207325, + 0.006495071109384298, + 0.0026178942061960697, + 0.010599308647215366, + 0.011080288328230381, + 0.004009630531072617, + 0.003510913113132119, + 0.002787531353533268, + 0.004524450283497572, + 0.0034161771181970835, + 0.006016492377966642, + 0.011289737187325954, + 0.010341344401240349, + 0.008624648675322533, + 0.011446326039731503, + 0.008237997069954872, + 0.007718686014413834, + 0.008703069761395454, + 0.0028047605883330107, + 0.002489701146259904, + 0.008403033949434757, + 0.008991086855530739, + 0.003966975025832653, + 0.004050372168421745, + 0.005632726009935141, + 0.004247093573212624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.029794618487358093 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, the camera panning steadily to the right, follows a massive, asymmetrical, geometric alien spacecraft as it glides with a majestic, silent speed over a rain-soaked futuristic metropolis. The vessel's hull is composed of an unknown, smooth, obsidian material; rain slides down its icy curves without leaving a trace. The hull is covered in intricate, luminous patterns that pulsate with a cold, cyan-blue light in a rhythmic, inhuman pattern, casting an eerie halo on the wet streets below. Below, skyscrapers pierce the sky like a metallic jungle, shrouded by a thick, dark cloud. Huge, translucent holographic billboards flicker and morph between the buildings, displaying dynamic, color-saturated commercials. As the spacecraft gracefully passed in front of a holographic billboard projecting a huge female face, its powerful energy field even caused slight distortion of the holographic image and data interference. The whole scene, with the contrast between the cold background of the rainy night and the warm glow of the billboard, created a strong sense of cyberpunk melancholy and technological sublimity.", + "seed": 23793, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_0/index_9874_seed_23793_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.056176491733640435, + "vlm_logprob": -0.056176491733640435, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.056176491733640435, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11782754212617874 + ], + "opposite_logprob_tokens": [ + -2.4928276538848877 + ], + "logprob_reward": -0.11782754212617874, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.11782754212617874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23607639968395233, + 0.07823312282562256, + 0.05930820479989052, + 0.10962226986885071, + 0.11859368532896042, + 0.12834426760673523, + 0.11585164815187454, + 0.05229531228542328, + 0.05462472140789032, + 0.04705039784312248 + ], + "credit_map": [ + 0.016137339174747467, + 0.02081512100994587, + 0.02681472897529602, + 0.03921841084957123, + 0.02559269778430462, + 0.041129834949970245, + 0.041001420468091965, + 0.02536684088408947, + 0.003946503158658743, + 0.004817469511181116, + 0.008161920122802258, + 0.012626860290765762, + 0.007227607071399689, + 0.015711385756731033, + 0.016013646498322487, + 0.009727728553116322, + 0.008137679658830166, + 0.006471727509051561, + 0.01175461895763874, + 0.008132516406476498, + 0.005420611705631018, + 0.006660278886556625, + 0.006889651529490948, + 0.0058411178179085255, + 0.008539924398064613, + 0.011941041797399521, + 0.013359337113797665, + 0.018393393605947495, + 0.01050030067563057, + 0.013406860642135143, + 0.016032740473747253, + 0.017448678612709045, + 0.011832812801003456, + 0.015393482521176338, + 0.01540024857968092, + 0.015553622506558895, + 0.019070208072662354, + 0.017914913594722748, + 0.013328438624739647, + 0.010099955834448338, + 0.01046810857951641, + 0.015505196526646614, + 0.028010176494717598, + 0.020615410059690475, + 0.01354584377259016, + 0.018795907497406006, + 0.012756984680891037, + 0.008646635338664055, + 0.01009329967200756, + 0.010576481930911541, + 0.015601355582475662, + 0.011795200407505035, + 0.02243105135858059, + 0.014253136701881886, + 0.018992871046066284, + 0.012108243070542812, + 0.004254857078194618, + 0.006517963949590921, + 0.006572662852704525, + 0.008232430554926395, + 0.00700914952903986, + 0.007163709029555321, + 0.007605463266372681, + 0.0049390788190066814, + 0.00696820393204689, + 0.0047750468365848064, + 0.014052781276404858, + 0.00632766168564558, + 0.00496247410774231, + 0.006881217937916517, + 0.0059152753092348576, + 0.004742059390991926, + 0.005848222877830267, + 0.007848899811506271, + 0.005194056313484907, + 0.006248075980693102, + 0.005349372047930956, + 0.006123384460806847, + 0.0062583074904978275, + 0.004180080257356167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.11782754212617874 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812633082270622 + ], + "opposite_logprob_tokens": [ + -3.288126230239868 + ], + "logprob_reward": -0.03812633082270622, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812633082270622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15224508941173553, + 0.04825183004140854, + 0.03124157153069973, + 0.06421086937189102, + 0.09339383989572525, + 0.11444688588380814, + 0.15982608497142792, + 0.16792327165603638, + 0.10044381022453308, + 0.06801673024892807 + ], + "credit_map": [ + 0.012821746058762074, + 0.01594453677535057, + 0.015860889106988907, + 0.025273723527789116, + 0.017188578844070435, + 0.01719030924141407, + 0.031645093113183975, + 0.016320206224918365, + 0.0031696325168013573, + 0.0032103729899972677, + 0.0052693989127874374, + 0.007775778416544199, + 0.003946008160710335, + 0.0053665852174162865, + 0.011359015479683876, + 0.008155038580298424, + 0.004034394398331642, + 0.0035509613808244467, + 0.006313292775303125, + 0.00533162010833621, + 0.0023166690953075886, + 0.003482967847958207, + 0.002894511679187417, + 0.0033171530812978745, + 0.004594212397933006, + 0.004072750918567181, + 0.009403267875313759, + 0.00972713902592659, + 0.009792791679501534, + 0.00994476955384016, + 0.009743398986756802, + 0.006932539399713278, + 0.01092649344354868, + 0.009057399816811085, + 0.011030704714357853, + 0.01451959554105997, + 0.010613761842250824, + 0.010407580994069576, + 0.013001207262277603, + 0.013837086968123913, + 0.009867634624242783, + 0.015204970724880695, + 0.025942374020814896, + 0.016805145889520645, + 0.011666529811918736, + 0.01592162810266018, + 0.010424735955893993, + 0.008613870479166508, + 0.012868938036262989, + 0.01677880994975567, + 0.022323163226246834, + 0.016033971682190895, + 0.03127598017454147, + 0.018562788143754005, + 0.024618763476610184, + 0.017363665625452995, + 0.01576315425336361, + 0.02304118499159813, + 0.020715177059173584, + 0.019034259021282196, + 0.029185457155108452, + 0.02371870167553425, + 0.019857214763760567, + 0.016608120873570442, + 0.009675404988229275, + 0.00840512104332447, + 0.01903865858912468, + 0.012465725652873516, + 0.013251193799078465, + 0.012089033611118793, + 0.015553136356174946, + 0.009965548291802406, + 0.008220823481678963, + 0.007526765111833811, + 0.006636851467192173, + 0.008780423551797867, + 0.008747790940105915, + 0.007907729595899582, + 0.011347128078341484, + 0.008849221281707287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.03812633082270622 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294203251600266 + ], + "opposite_logprob_tokens": [ + -2.2379419803619385 + ], + "logprob_reward": -0.11294203251600266, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294203251600266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1394999474287033, + 0.0818459764122963, + 0.03984689712524414, + 0.04502755030989647, + 0.1301412582397461, + 0.22369560599327087, + 0.03663691505789757, + 0.06385758519172668, + 0.13924986124038696, + 0.10019846260547638 + ], + "credit_map": [ + 0.01421409286558628, + 0.012849153950810432, + 0.01729660853743553, + 0.014106465503573418, + 0.013794265687465668, + 0.017257286235690117, + 0.028362937271595, + 0.021619126200675964, + 0.007244444917887449, + 0.00844422448426485, + 0.009590402245521545, + 0.014987340196967125, + 0.007290167268365622, + 0.01108494307845831, + 0.01232215203344822, + 0.01088228914886713, + 0.0038407284300774336, + 0.0034992299042642117, + 0.004104380030184984, + 0.0057168519124388695, + 0.004494617227464914, + 0.005312059540301561, + 0.008986135013401508, + 0.003892892273142934, + 0.0024976825807243586, + 0.002760437550023198, + 0.006211564410477877, + 0.009070857428014278, + 0.0038666222244501114, + 0.006071446929126978, + 0.0067222542129457, + 0.007826680317521095, + 0.010746922343969345, + 0.012381399050354958, + 0.011565054766833782, + 0.01694936864078045, + 0.012838571332395077, + 0.014728045091032982, + 0.017818216234445572, + 0.033113669604063034, + 0.018121197819709778, + 0.022518834099173546, + 0.05860317498445511, + 0.029960673302412033, + 0.01659674197435379, + 0.03132987767457962, + 0.0225263349711895, + 0.024038774892687798, + 0.003307928564026952, + 0.004613751079887152, + 0.005271228961646557, + 0.005642073694616556, + 0.003839538199827075, + 0.004355358891189098, + 0.004139479715377092, + 0.005467552691698074, + 0.0051295943558216095, + 0.009718320332467556, + 0.006759722251445055, + 0.007843401283025742, + 0.011015619151294231, + 0.008526964113116264, + 0.00917834509164095, + 0.00568561302497983, + 0.012336835265159607, + 0.01501921471208334, + 0.017190394923090935, + 0.017801254987716675, + 0.02259623445570469, + 0.017699887976050377, + 0.018542127683758736, + 0.018063904717564583, + 0.013200519606471062, + 0.010770611464977264, + 0.018301405012607574, + 0.009539345279335976, + 0.009034794755280018, + 0.016215940937399864, + 0.014100810512900352, + 0.009035027585923672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.11294203251600266 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256166215986013 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256166215986013, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256166215986013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11379961669445038, + 0.12295173853635788, + 0.06591015309095383, + 0.06816476583480835, + 0.07430126518011093, + 0.22788387537002563, + 0.06738819926977158, + 0.08215257525444031, + 0.09300107508897781, + 0.0844467356801033 + ], + "credit_map": [ + 0.008143357932567596, + 0.011760124005377293, + 0.01246299035847187, + 0.016022086143493652, + 0.012400767765939236, + 0.015496108680963516, + 0.023513764142990112, + 0.014000418595969677, + 0.008776724338531494, + 0.011102855205535889, + 0.01564917340874672, + 0.022822249680757523, + 0.014267255552113056, + 0.01859939470887184, + 0.019556976854801178, + 0.012177110649645329, + 0.005665806587785482, + 0.0072415051981806755, + 0.01218220591545105, + 0.009301242418587208, + 0.0072728488594293594, + 0.007466929033398628, + 0.008360172621905804, + 0.008419444784522057, + 0.0047941445372998714, + 0.008334483951330185, + 0.009046463295817375, + 0.010664822533726692, + 0.00859206635504961, + 0.009987306781113148, + 0.009678126312792301, + 0.007067350670695305, + 0.00924756657332182, + 0.00978212058544159, + 0.009210528805851936, + 0.010336308740079403, + 0.008485322818160057, + 0.00879710540175438, + 0.011427350342273712, + 0.007014964707195759, + 0.0170590877532959, + 0.029557209461927414, + 0.05777495354413986, + 0.030861755833029747, + 0.024565866217017174, + 0.0311233252286911, + 0.019432511180639267, + 0.017509175464510918, + 0.005646851845085621, + 0.007773115299642086, + 0.0100208455696702, + 0.010803136974573135, + 0.008930139243602753, + 0.008850854821503162, + 0.009016212075948715, + 0.006347040180116892, + 0.00695182615891099, + 0.008843250572681427, + 0.012829422950744629, + 0.01163738314062357, + 0.01084135938435793, + 0.011591794900596142, + 0.011655811220407486, + 0.007801721803843975, + 0.008176085539162159, + 0.008430086076259613, + 0.00766464788466692, + 0.013391968794167042, + 0.013504913076758385, + 0.012397050857543945, + 0.01208204124122858, + 0.017354285344481468, + 0.012075359001755714, + 0.008758856914937496, + 0.01349575910717249, + 0.010889554396271706, + 0.007615522015839815, + 0.013057448901236057, + 0.009970135986804962, + 0.00858409982174635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.005256166215986013 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730386987328529 + ], + "opposite_logprob_tokens": [ + -5.006730556488037 + ], + "logprob_reward": -0.006730386987328529, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730386987328529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07566524296998978, + 0.10998532921075821, + 0.024619953706860542, + 0.17896683514118195, + 0.05538105592131615, + 0.1640128493309021, + 0.09594099968671799, + 0.04347275570034981, + 0.08655548840761185, + 0.16539952158927917 + ], + "credit_map": [ + 0.006811848375946283, + 0.008397079072892666, + 0.008124344050884247, + 0.011482331901788712, + 0.008012698031961918, + 0.007363993674516678, + 0.01554326806217432, + 0.009929673746228218, + 0.00638990942388773, + 0.008720321580767632, + 0.01232253760099411, + 0.02137351594865322, + 0.010815879330039024, + 0.012875375337898731, + 0.021690700203180313, + 0.015797071158885956, + 0.002036611782386899, + 0.0032842238433659077, + 0.0031221495009958744, + 0.003884761594235897, + 0.0030690019484609365, + 0.0024830710608512163, + 0.004099536221474409, + 0.0026405970565974712, + 0.01147359237074852, + 0.014828269369900227, + 0.0236015934497118, + 0.03326716646552086, + 0.01231457944959402, + 0.023996178060770035, + 0.028922386467456818, + 0.03056306019425392, + 0.00512642040848732, + 0.005469559226185083, + 0.005482188425958157, + 0.0074449279345571995, + 0.0068908752873539925, + 0.006081821396946907, + 0.009438328444957733, + 0.009446932934224606, + 0.014255263842642307, + 0.022958844900131226, + 0.03604039549827576, + 0.028084762394428253, + 0.017033468931913376, + 0.020985856652259827, + 0.013667934574186802, + 0.010986311361193657, + 0.00849984958767891, + 0.008879677392542362, + 0.013374638743698597, + 0.012880333699285984, + 0.01397987175732851, + 0.011994832195341587, + 0.015353665687143803, + 0.01097811944782734, + 0.003276749048382044, + 0.004811163526028395, + 0.00530074629932642, + 0.00586735038086772, + 0.006260978057980537, + 0.0057208179496228695, + 0.0063752951100468636, + 0.0058596511371433735, + 0.008934793062508106, + 0.009350717067718506, + 0.018399249762296677, + 0.009978603571653366, + 0.01131146028637886, + 0.01059133093804121, + 0.011672154068946838, + 0.0063171712681651115, + 0.018233487382531166, + 0.01542698498815298, + 0.018004179000854492, + 0.019536402076482773, + 0.016135001555085182, + 0.019407063722610474, + 0.02764369733631611, + 0.03101268596947193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 5 + ], + "selected_reward": -0.006730386987328529 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, the camera panning steadily to the right, follows a massive, asymmetrical, geometric alien spacecraft as it glides with a majestic, silent speed over a rain-soaked futuristic metropolis. The vessel's hull is composed of an unknown, smooth, obsidian material; rain slides down its icy curves without leaving a trace. The hull is covered in intricate, luminous patterns that pulsate with a cold, cyan-blue light in a rhythmic, inhuman pattern, casting an eerie halo on the wet streets below. Below, skyscrapers pierce the sky like a metallic jungle, shrouded by a thick, dark cloud. Huge, translucent holographic billboards flicker and morph between the buildings, displaying dynamic, color-saturated commercials. As the spacecraft gracefully passed in front of a holographic billboard projecting a huge female face, its powerful energy field even caused slight distortion of the holographic image and data interference. The whole scene, with the contrast between the cold background of the rainy night and the warm glow of the billboard, created a strong sense of cyberpunk melancholy and technological sublimity.", + "seed": 23794, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_0/index_9874_seed_23794_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04790179505944252, + "vlm_logprob": -0.04790179505944252, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.04790179505944252, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10282652825117111 + ], + "opposite_logprob_tokens": [ + -2.7278265953063965 + ], + "logprob_reward": -0.10282652825117111, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.10282652825117111, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14582765102386475, + 0.1296016275882721, + 0.16219139099121094, + 0.08980545401573181, + 0.062134016305208206, + 0.11253389716148376, + 0.08183112740516663, + 0.05552937835454941, + 0.09116459637880325, + 0.06938087940216064 + ], + "credit_map": [ + 0.011875681579113007, + 0.012874073348939419, + 0.014838282018899918, + 0.02702694572508335, + 0.016509762033820152, + 0.017265072092413902, + 0.016164259985089302, + 0.02927357330918312, + 0.013613010756671429, + 0.014976163394749165, + 0.019900517538189888, + 0.014212948270142078, + 0.014471158385276794, + 0.011091848835349083, + 0.024153972044587135, + 0.017182011157274246, + 0.022369328886270523, + 0.0123955262824893, + 0.02748933620750904, + 0.014500842429697514, + 0.013030629605054855, + 0.03800192102789879, + 0.016726404428482056, + 0.01767740584909916, + 0.01344976481050253, + 0.007825344800949097, + 0.02419072575867176, + 0.011397255584597588, + 0.006405723746865988, + 0.006356229539960623, + 0.011187117546796799, + 0.008993295021355152, + 0.005221222992986441, + 0.006657715421169996, + 0.007086654659360647, + 0.009523693472146988, + 0.00804395042359829, + 0.008360430598258972, + 0.008044522255659103, + 0.009195826947689056, + 0.026755020022392273, + 0.009461517445743084, + 0.037007078528404236, + 0.012314613908529282, + 0.006267003249377012, + 0.007594095543026924, + 0.007867504842579365, + 0.005267056170850992, + 0.010925129055976868, + 0.009479638189077377, + 0.015207706019282341, + 0.011310883797705173, + 0.0065581281669437885, + 0.007779999636113644, + 0.007011882495135069, + 0.013557765632867813, + 0.007531391456723213, + 0.008134184405207634, + 0.0071802386082708836, + 0.009194282814860344, + 0.00804381724447012, + 0.00421737227588892, + 0.006967832334339619, + 0.004260255489498377, + 0.018482519313693047, + 0.012474660761654377, + 0.012139171361923218, + 0.005717884749174118, + 0.006855296436697245, + 0.006964432075619698, + 0.01965172588825226, + 0.008878914639353752, + 0.005835584364831448, + 0.003958149813115597, + 0.00583603885024786, + 0.00872555561363697, + 0.007919525727629662, + 0.012002237141132355, + 0.010078114457428455, + 0.015025676228106022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.10282652825117111 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309685528278351 + ], + "opposite_logprob_tokens": [ + -3.1680967807769775 + ], + "logprob_reward": -0.04309685528278351, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309685528278351, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09869088977575302, + 0.051143113523721695, + 0.07243641465902328, + 0.13836145401000977, + 0.05857071280479431, + 0.166778564453125, + 0.1841445118188858, + 0.08906441926956177, + 0.07601135969161987, + 0.0647984966635704 + ], + "credit_map": [ + 0.006956650409847498, + 0.008726435713469982, + 0.009922734461724758, + 0.018735988065600395, + 0.009570082649588585, + 0.010077686049044132, + 0.014431960880756378, + 0.020269347354769707, + 0.004660570994019508, + 0.008017071522772312, + 0.007520680781453848, + 0.006264908704906702, + 0.005261409096419811, + 0.00516233267262578, + 0.008348236791789532, + 0.005907903425395489, + 0.010059329681098461, + 0.007655274588614702, + 0.010912671685218811, + 0.008095084689557552, + 0.007254668977111578, + 0.012157279066741467, + 0.007747399155050516, + 0.008554703556001186, + 0.01598590426146984, + 0.015871716663241386, + 0.02946224994957447, + 0.020335599780082703, + 0.01774648018181324, + 0.011389288119971752, + 0.013013619929552078, + 0.01455660443753004, + 0.004501669201999903, + 0.007712929975241423, + 0.006047351285815239, + 0.009486502036452293, + 0.007693259511142969, + 0.009006844833493233, + 0.009160596877336502, + 0.004961554892361164, + 0.020753348246216774, + 0.024292372167110443, + 0.021160801872611046, + 0.023724189028143883, + 0.02155996672809124, + 0.020125359296798706, + 0.022823337465524673, + 0.012339189648628235, + 0.026663953438401222, + 0.026546329259872437, + 0.038441892713308334, + 0.022892141714692116, + 0.020241916179656982, + 0.015194909647107124, + 0.018353700637817383, + 0.0158096756786108, + 0.011303482577204704, + 0.011384766548871994, + 0.012378041632473469, + 0.016243357211351395, + 0.006462123245000839, + 0.008068138733506203, + 0.015015962533652782, + 0.008208549581468105, + 0.011946392245590687, + 0.011438376270234585, + 0.011272943578660488, + 0.010969799011945724, + 0.006308223586529493, + 0.005066276527941227, + 0.011845233850181103, + 0.007164118345826864, + 0.008408124558627605, + 0.0055811465717852116, + 0.00962707307189703, + 0.008369144052267075, + 0.007479674648493528, + 0.008624590002000332, + 0.006796908099204302, + 0.00991184264421463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.04309685528278351 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899792492389679 + ], + "opposite_logprob_tokens": [ + -2.578997850418091 + ], + "logprob_reward": -0.07899792492389679, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899792492389679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10352914035320282, + 0.07745765894651413, + 0.16856910288333893, + 0.11440657079219818, + 0.06032881885766983, + 0.18090324103832245, + 0.09418520331382751, + 0.07404444366693497, + 0.08603142201900482, + 0.040544383227825165 + ], + "credit_map": [ + 0.00903023686259985, + 0.012142522260546684, + 0.010462377220392227, + 0.01987498626112938, + 0.015111374668776989, + 0.011164627969264984, + 0.010285778902471066, + 0.015457237139344215, + 0.008233810774981976, + 0.00919475220143795, + 0.01115379761904478, + 0.011933645233511925, + 0.007776062004268169, + 0.00820833444595337, + 0.012570880353450775, + 0.008386373519897461, + 0.025281883776187897, + 0.015641914680600166, + 0.02368909679353237, + 0.017220627516508102, + 0.014092388562858105, + 0.029744010418653488, + 0.0160810686647892, + 0.026818111538887024, + 0.016466859728097916, + 0.009454832412302494, + 0.023745665326714516, + 0.01579616591334343, + 0.00849235337227583, + 0.01033725868910551, + 0.015600239858031273, + 0.014513198286294937, + 0.004311987664550543, + 0.006020121742039919, + 0.007354736328125, + 0.00855215359479189, + 0.007621173281222582, + 0.008738542906939983, + 0.008061050437390804, + 0.00966905802488327, + 0.03053831309080124, + 0.023702045902609825, + 0.036663684993982315, + 0.019372116774320602, + 0.02096719667315483, + 0.018077632412314415, + 0.018672896549105644, + 0.012909360229969025, + 0.012335443869233131, + 0.011758281849324703, + 0.015594162978231907, + 0.010783962905406952, + 0.010526749305427074, + 0.008646679110825062, + 0.010670403018593788, + 0.013869513757526875, + 0.011290517635643482, + 0.010438921861350536, + 0.010063269175589085, + 0.011545110493898392, + 0.00681572500616312, + 0.0056342012248933315, + 0.012506058439612389, + 0.005750641226768494, + 0.010034854523837566, + 0.010452968068420887, + 0.017027201130986214, + 0.013897553086280823, + 0.009553375653922558, + 0.005776267498731613, + 0.011686647310853004, + 0.007602549158036709, + 0.003671316895633936, + 0.003934217616915703, + 0.004409163258969784, + 0.006137973628938198, + 0.004775409121066332, + 0.006211034953594208, + 0.004963404964655638, + 0.006441863253712654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.07899792492389679 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644777350127697 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.008644777350127697, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644777350127697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10572116822004318, + 0.10839232802391052, + 0.05516991764307022, + 0.12088219076395035, + 0.04555665701627731, + 0.16346274316310883, + 0.16846337914466858, + 0.11459255963563919, + 0.07816370576620102, + 0.03959539905190468 + ], + "credit_map": [ + 0.009379644878208637, + 0.00878690741956234, + 0.011119048111140728, + 0.015542479231953621, + 0.013644532300531864, + 0.011736742220818996, + 0.014228415675461292, + 0.021283403038978577, + 0.01009808387607336, + 0.01060597412288189, + 0.01943483017385006, + 0.013036469928920269, + 0.010751173831522465, + 0.012929671443998814, + 0.017805639654397964, + 0.013730488717556, + 0.007465976290404797, + 0.005483992397785187, + 0.008592059835791588, + 0.006940241903066635, + 0.004359703045338392, + 0.008897882886230946, + 0.005348344333469868, + 0.008081717416644096, + 0.014584398828446865, + 0.013656726107001305, + 0.02449355646967888, + 0.017126932740211487, + 0.013284267857670784, + 0.009336607530713081, + 0.015083437785506248, + 0.013316262513399124, + 0.0037058463785797358, + 0.00531613826751709, + 0.004825083538889885, + 0.007700928021222353, + 0.006557893007993698, + 0.0057770563289523125, + 0.0061999778263270855, + 0.005473731551319361, + 0.036524850875139236, + 0.01349974051117897, + 0.045239437371492386, + 0.01735663414001465, + 0.014789169654250145, + 0.015804581344127655, + 0.01035395823419094, + 0.009894362650811672, + 0.02324810065329075, + 0.02313878759741783, + 0.03701182082295418, + 0.020648328587412834, + 0.01904512196779251, + 0.014879202470183372, + 0.01812232844531536, + 0.012369678355753422, + 0.017063362523913383, + 0.013424795120954514, + 0.018867801874876022, + 0.01644488051533699, + 0.009610643610358238, + 0.010829213075339794, + 0.018960749730467796, + 0.009391107596457005, + 0.011639483273029327, + 0.009798482991755009, + 0.014774586074054241, + 0.01322767324745655, + 0.007359316572546959, + 0.005441185552626848, + 0.010298270732164383, + 0.005624702665954828, + 0.004040925297886133, + 0.0030847119633108377, + 0.0040826438926160336, + 0.00758263049647212, + 0.0029522841796278954, + 0.007203580345958471, + 0.004073914606124163, + 0.006574704777449369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 3 + ], + "selected_reward": -0.008644777350127697 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942889489233494 + ], + "opposite_logprob_tokens": [ + -5.1309428215026855 + ], + "logprob_reward": -0.005942889489233494, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942889489233494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16085173189640045, + 0.05402624234557152, + 0.06343700736761093, + 0.14574424922466278, + 0.04788590967655182, + 0.18478335440158844, + 0.06624142080545425, + 0.08276733011007309, + 0.08464843034744263, + 0.10961434990167618 + ], + "credit_map": [ + 0.012910784222185612, + 0.014595305547118187, + 0.016973134130239487, + 0.029576757922768593, + 0.016574649140238762, + 0.016509639099240303, + 0.018910137936472893, + 0.03480130061507225, + 0.004870850592851639, + 0.0066671897657215595, + 0.007910199463367462, + 0.005397689063102007, + 0.00840988289564848, + 0.006242688279598951, + 0.006935582496225834, + 0.007592155132442713, + 0.008045335300266743, + 0.005013483576476574, + 0.011176303029060364, + 0.006832744926214218, + 0.00518927862867713, + 0.012855864129960537, + 0.007027527783066034, + 0.007296469062566757, + 0.018759388476610184, + 0.01486424170434475, + 0.03585238754749298, + 0.02003704570233822, + 0.01777239888906479, + 0.01166614517569542, + 0.013097762130200863, + 0.01369487214833498, + 0.004203564953058958, + 0.005139554850757122, + 0.00588841550052166, + 0.006964738946408033, + 0.004864219576120377, + 0.005121063441038132, + 0.009605270810425282, + 0.0060990783385932446, + 0.04789544269442558, + 0.011951589956879616, + 0.06436976045370102, + 0.017627498134970665, + 0.00997744407504797, + 0.011952357366681099, + 0.010955458506941795, + 0.010053801350295544, + 0.009372642263770103, + 0.008901618421077728, + 0.009872299619019032, + 0.008337797597050667, + 0.005909832194447517, + 0.00793315190821886, + 0.007670024875551462, + 0.008244051598012447, + 0.010711575858294964, + 0.009242190048098564, + 0.01189428847283125, + 0.00949590839445591, + 0.011562607251107693, + 0.008214996196329594, + 0.013294223695993423, + 0.008351528085768223, + 0.011907932348549366, + 0.007490272633731365, + 0.020125584676861763, + 0.013191728852689266, + 0.005483593791723251, + 0.007138223387300968, + 0.011882169172167778, + 0.007428920362144709, + 0.009868995286524296, + 0.009369169361889362, + 0.010944264940917492, + 0.01900245063006878, + 0.008228125981986523, + 0.017531678080558777, + 0.014860893599689007, + 0.019808758050203323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.005942889489233494 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, the camera panning steadily to the right, follows a massive, asymmetrical, geometric alien spacecraft as it glides with a majestic, silent speed over a rain-soaked futuristic metropolis. The vessel's hull is composed of an unknown, smooth, obsidian material; rain slides down its icy curves without leaving a trace. The hull is covered in intricate, luminous patterns that pulsate with a cold, cyan-blue light in a rhythmic, inhuman pattern, casting an eerie halo on the wet streets below. Below, skyscrapers pierce the sky like a metallic jungle, shrouded by a thick, dark cloud. Huge, translucent holographic billboards flicker and morph between the buildings, displaying dynamic, color-saturated commercials. As the spacecraft gracefully passed in front of a holographic billboard projecting a huge female face, its powerful energy field even caused slight distortion of the holographic image and data interference. The whole scene, with the contrast between the cold background of the rainy night and the warm glow of the billboard, created a strong sense of cyberpunk melancholy and technological sublimity.", + "seed": 23795, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_0/index_9874_seed_23795_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07705800011754035, + "vlm_logprob": -0.07705800011754035, + "vlm_yes_no_margin": 3.3499999761581423, + "vlm_token_credit": -0.07705800011754035, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20009876787662506 + ], + "opposite_logprob_tokens": [ + -1.9500987529754639 + ], + "logprob_reward": -0.20009876787662506, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.20009876787662506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19125394523143768, + 0.1619410365819931, + 0.1319727897644043, + 0.08279456943273544, + 0.04683994501829147, + 0.0726257711648941, + 0.10239958018064499, + 0.054182592779397964, + 0.050933822989463806, + 0.10505597293376923 + ], + "credit_map": [ + 0.016320426017045975, + 0.029309246689081192, + 0.017865415662527084, + 0.034232575446367264, + 0.02633514814078808, + 0.025330819189548492, + 0.025917042046785355, + 0.015943262726068497, + 0.01292425487190485, + 0.02388959564268589, + 0.02032061479985714, + 0.04067189246416092, + 0.020513808354735374, + 0.020454134792089462, + 0.015265085734426975, + 0.007901636883616447, + 0.013797393068671227, + 0.012872149236500263, + 0.030842473730444908, + 0.014055216684937477, + 0.007720208261162043, + 0.0260170828551054, + 0.013076462782919407, + 0.013591804541647434, + 0.006805942859500647, + 0.00645106378942728, + 0.013162219896912575, + 0.010383286513388157, + 0.00989427138119936, + 0.01341591402888298, + 0.015200952999293804, + 0.00748091796413064, + 0.0034163750242441893, + 0.004630441777408123, + 0.006776254624128342, + 0.007208123803138733, + 0.005648475605994463, + 0.007389850914478302, + 0.007644794415682554, + 0.004125628620386124, + 0.00611988827586174, + 0.010285244323313236, + 0.011789930053055286, + 0.008751544170081615, + 0.00854503083974123, + 0.007925884798169136, + 0.012352507561445236, + 0.006855745799839497, + 0.010211167857050896, + 0.01410350389778614, + 0.017403360456228256, + 0.020127300173044205, + 0.0072114975191652775, + 0.009217165410518646, + 0.013285637833178043, + 0.010839950293302536, + 0.004662070423364639, + 0.005240379832684994, + 0.010589552111923695, + 0.008613253943622112, + 0.004831700120121241, + 0.006345489993691444, + 0.006388933397829533, + 0.007511214353144169, + 0.0048226346261799335, + 0.007657000329345465, + 0.006866625044494867, + 0.00715310825034976, + 0.006933505646884441, + 0.005337516311556101, + 0.005178759340196848, + 0.006984676234424114, + 0.014303106814622879, + 0.011118455789983273, + 0.020782288163900375, + 0.01009023655205965, + 0.0135622788220644, + 0.012180058285593987, + 0.013329461216926575, + 0.009690082632005215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.20009876787662506 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309263080358505 + ], + "opposite_logprob_tokens": [ + -3.168092727661133 + ], + "logprob_reward": -0.04309263080358505, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309263080358505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1119067445397377, + 0.12080053240060806, + 0.09126021713018417, + 0.07385420054197311, + 0.04613692685961723, + 0.04417043551802635, + 0.1457904428243637, + 0.07380347698926926, + 0.11457744985818863, + 0.17769961059093475 + ], + "credit_map": [ + 0.009118836373090744, + 0.01158876158297062, + 0.009598088450729847, + 0.01836450956761837, + 0.011332141235470772, + 0.01361443568021059, + 0.019315434619784355, + 0.018974533304572105, + 0.011095535010099411, + 0.01357568521052599, + 0.02208659239113331, + 0.024288494139909744, + 0.014457912184298038, + 0.01541996467858553, + 0.011277006939053535, + 0.008599339984357357, + 0.010655326768755913, + 0.013010850176215172, + 0.015939651057124138, + 0.013910273090004921, + 0.006262278184294701, + 0.014585303142666817, + 0.009543729946017265, + 0.007352794986218214, + 0.005617029033601284, + 0.005154051817953587, + 0.011859379708766937, + 0.00907615665346384, + 0.008446725085377693, + 0.008863380178809166, + 0.01197834126651287, + 0.012859134934842587, + 0.0032137532252818346, + 0.00487194349989295, + 0.006355212070047855, + 0.005300873424857855, + 0.007482998538762331, + 0.0057585472241044044, + 0.008121836930513382, + 0.005031760316342115, + 0.0034815340768545866, + 0.005403559654951096, + 0.006401457823812962, + 0.007419979199767113, + 0.004151530098170042, + 0.004883744288235903, + 0.007311089430004358, + 0.0051175388507544994, + 0.012929140590131283, + 0.019097881391644478, + 0.024553639814257622, + 0.015583996661007404, + 0.02084367908537388, + 0.014104780741035938, + 0.02365545742213726, + 0.015021857805550098, + 0.01061132550239563, + 0.005713636986911297, + 0.008459392003715038, + 0.010236992500722408, + 0.004536864813417196, + 0.009366738609969616, + 0.009382440708577633, + 0.015496080741286278, + 0.012737204320728779, + 0.009980462491512299, + 0.02703174017369747, + 0.014385752379894257, + 0.01210638228803873, + 0.01418506819754839, + 0.013200541958212852, + 0.010950294323265553, + 0.026904603466391563, + 0.030021298676729202, + 0.03865614905953407, + 0.016544947400689125, + 0.01630108617246151, + 0.012956121936440468, + 0.020634803920984268, + 0.015680592507123947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.04309263080358505 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706337869167328 + ], + "opposite_logprob_tokens": [ + -2.127063274383545 + ], + "logprob_reward": -0.12706337869167328, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12706337869167328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23591072857379913, + 0.3429851531982422, + 0.05741870403289795, + 0.10948727279901505, + 0.04470529779791832, + 0.054747652262449265, + 0.046342745423316956, + 0.030770527198910713, + 0.045695096254348755, + 0.031936727464199066 + ], + "credit_map": [ + 0.02124515175819397, + 0.03359934315085411, + 0.028430616483092308, + 0.050918467342853546, + 0.028661681339144707, + 0.029992904514074326, + 0.026845060288906097, + 0.01621752604842186, + 0.027560710906982422, + 0.05470319092273712, + 0.04020100459456444, + 0.05085361748933792, + 0.07081357389688492, + 0.04729359596967697, + 0.02977810427546501, + 0.021781397983431816, + 0.005512187723070383, + 0.007631842512637377, + 0.009410723112523556, + 0.011510174721479416, + 0.004197351168841124, + 0.00621681846678257, + 0.0054888371378183365, + 0.007450779899954796, + 0.012310678139328957, + 0.009903465397655964, + 0.019089054316282272, + 0.013755726628005505, + 0.00911432970315218, + 0.01631077378988266, + 0.019546162337064743, + 0.009457097388803959, + 0.003018395509570837, + 0.005184740759432316, + 0.006968871224671602, + 0.006111836992204189, + 0.005652121268212795, + 0.008589782752096653, + 0.0061912937089800835, + 0.0029882602393627167, + 0.00371009879745543, + 0.0051725199446082115, + 0.008364828303456306, + 0.006880765315145254, + 0.006080750375986099, + 0.006209484301507473, + 0.012674961239099503, + 0.00565424794331193, + 0.004501387942582369, + 0.006935368292033672, + 0.007194888312369585, + 0.004802368115633726, + 0.007195030804723501, + 0.004823389928787947, + 0.005266753491014242, + 0.005623561330139637, + 0.002072393661364913, + 0.0026595727540552616, + 0.0038013875018805265, + 0.006193771958351135, + 0.0020817117765545845, + 0.0038056562189012766, + 0.006123687606304884, + 0.004032347816973925, + 0.004437895026057959, + 0.008355843834578991, + 0.006284313276410103, + 0.006978187244385481, + 0.0036806066054850817, + 0.004838578402996063, + 0.0064785415306687355, + 0.004641131032258272, + 0.0034955607261508703, + 0.003323328448459506, + 0.006237532943487167, + 0.003314599860459566, + 0.00408676965162158, + 0.0037006845232099295, + 0.004749844782054424, + 0.0030284086242318153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.12706337869167328 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052584195509552956 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.0052584195509552956, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052584195509552956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16667166352272034, + 0.2539525330066681, + 0.07188337296247482, + 0.05981360375881195, + 0.04189793020486832, + 0.06366356462240219, + 0.13683435320854187, + 0.06376528739929199, + 0.095004141330719, + 0.04651352018117905 + ], + "credit_map": [ + 0.01452945452183485, + 0.025075891986489296, + 0.01710541732609272, + 0.0344269834458828, + 0.024407105520367622, + 0.02301858551800251, + 0.01713000424206257, + 0.010978215374052525, + 0.02386251650750637, + 0.04151121899485588, + 0.03777168318629265, + 0.04149423539638519, + 0.04014011099934578, + 0.03263981267809868, + 0.023594988510012627, + 0.012937967665493488, + 0.008477636612951756, + 0.009648039937019348, + 0.009869543835520744, + 0.011081244796514511, + 0.008454722352325916, + 0.009125716984272003, + 0.007628770545125008, + 0.007597699761390686, + 0.005838259123265743, + 0.006999992299824953, + 0.012091949582099915, + 0.00784417986869812, + 0.006516051944345236, + 0.006007723975926638, + 0.008834541775286198, + 0.005680907517671585, + 0.0030868782196193933, + 0.003997731953859329, + 0.006154423579573631, + 0.006837339606136084, + 0.00389894749969244, + 0.005397146102041006, + 0.00770793529227376, + 0.004817526321858168, + 0.004375636577606201, + 0.006883397698402405, + 0.009787002578377724, + 0.007046325132250786, + 0.007021769415587187, + 0.00615578843280673, + 0.012551353313028812, + 0.009842287749052048, + 0.01408373937010765, + 0.013871777802705765, + 0.02288065478205681, + 0.01914016157388687, + 0.013431545346975327, + 0.013649584725499153, + 0.02178335189819336, + 0.017993539571762085, + 0.008007492870092392, + 0.005028270184993744, + 0.011941362172365189, + 0.008690064772963524, + 0.005071296356618404, + 0.0064123510383069515, + 0.006954028271138668, + 0.011660420335829258, + 0.008760878816246986, + 0.01203869003802538, + 0.020790357142686844, + 0.013206100091338158, + 0.009468925185501575, + 0.012304509058594704, + 0.010261770337820053, + 0.00817291159182787, + 0.006207604892551899, + 0.006414640694856644, + 0.006419036071747541, + 0.005817099940031767, + 0.004658692050725222, + 0.005729976575821638, + 0.005240356083959341, + 0.006026113871484995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0052584195509552956 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan steadily to the right while following the spacecraft?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the spacecraft massive, asymmetrical, and geometric in shape?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the spacecraft's hull made of a smooth, obsidian material with luminous cyan-blue patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the spacecraft glide over a rain-soaked futuristic metropolis with skyscrapers and flickering holographic billboards?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977680366486311 + ], + "opposite_logprob_tokens": [ + -4.634776592254639 + ], + "logprob_reward": -0.00977680366486311, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977680366486311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2109677493572235, + 0.07886906713247299, + 0.08061733841896057, + 0.07421493530273438, + 0.08664336055517197, + 0.07241956144571304, + 0.10223598033189774, + 0.205592080950737, + 0.05064595863223076, + 0.03779394552111626 + ], + "credit_map": [ + 0.017050663009285927, + 0.025982171297073364, + 0.021430818364024162, + 0.04887774959206581, + 0.02641274407505989, + 0.02917933650314808, + 0.02674897015094757, + 0.01528529915958643, + 0.007487723138183355, + 0.008979620411992073, + 0.010722275823354721, + 0.011777954176068306, + 0.007842503488063812, + 0.010553145781159401, + 0.012142962776124477, + 0.009362881071865559, + 0.006482838187366724, + 0.009811914525926113, + 0.013502631336450577, + 0.011028489097952843, + 0.008036471903324127, + 0.011873112991452217, + 0.009941385127604008, + 0.009940496645867825, + 0.008339636027812958, + 0.008631420321762562, + 0.010700504295527935, + 0.008668226189911366, + 0.006931599695235491, + 0.009454429149627686, + 0.013515137135982513, + 0.007973983883857727, + 0.005131462588906288, + 0.007270472589880228, + 0.012438821606338024, + 0.014070585370063782, + 0.008702719584107399, + 0.010225901380181313, + 0.01547909900546074, + 0.013324301689863205, + 0.0053480239585042, + 0.007370694540441036, + 0.010475380346179008, + 0.013528183102607727, + 0.006793245207518339, + 0.006637008395045996, + 0.012753545306622982, + 0.00951347779482603, + 0.008152992464601994, + 0.010887869633734226, + 0.01612699031829834, + 0.012969817034900188, + 0.013101709075272083, + 0.010493768379092216, + 0.01802513189613819, + 0.012477703392505646, + 0.01712123304605484, + 0.02269202098250389, + 0.032897934317588806, + 0.03879683092236519, + 0.012196182273328304, + 0.02448130026459694, + 0.029918020591139793, + 0.027488557621836662, + 0.0056225610896945, + 0.011899127624928951, + 0.005537867080420256, + 0.005835331045091152, + 0.007957251742482185, + 0.004730448592454195, + 0.004252112004905939, + 0.004811258986592293, + 0.003731370670720935, + 0.005052482709288597, + 0.006705800537019968, + 0.0042543006129562855, + 0.003923898097127676, + 0.0038983658887445927, + 0.0052860211580991745, + 0.004941707476973534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.00977680366486311 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_1/index_tensor([7881, 7881])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_1/index_tensor([7881, 7881])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6e73ca50d1bc2dc89ece0027d533d0bf88b35557 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_1/index_tensor([7881, 7881])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, sweeping, long shot, shot from a slightly upward perspective, captures a tranquil, moonless summer night. Over a secluded meadow, surrounded by the dark silhouette of an ancient forest, the deep indigo sky is dotted with sparse stars. A gentle breeze rustles the tall grass and distant leaves with a barely audible rustle. At the center of the image is a magnificent ribbon of light, emanating a warm, pulsating yellow-green bioluminescence from countless fireflies. This ethereal river of light slowly and gracefully undulates and swirls through the sky, like a living nebula. Individual fireflies dart and pause within the ribbon, collectively weaving this flowing tapestry of light. Their paths, guided by the gentle breeze, alternately become dense, like brilliant star clusters, and sparse, like a wisp of ethereal light. Their light casts a soft, dreamy halo in the absolute darkness, faintly illuminating the tips of blades of grass and motes of dust floating in the air, creating a magical atmosphere of profound tranquility and the wonder of life.", + "seed": 11784, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_1/index_7881_seed_11784_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11474597230553626, + "vlm_logprob": -0.11474597230553626, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.11474597230553626, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19007375836372375 + ], + "opposite_logprob_tokens": [ + -1.8150737285614014 + ], + "logprob_reward": -0.19007375836372375, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19007375836372375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03210604563355446, + 0.012146041728556156, + 0.014923571608960629, + 0.010628381744027138, + 0.020670071244239807, + 0.030952902510762215, + 0.04356531798839569, + 0.013731746934354305, + 0.013604653999209404, + 0.8076712489128113 + ], + "credit_map": [ + 0.0031603025272488594, + 0.0030356552451848984, + 0.00791220273822546, + 0.0035743932239711285, + 0.002736396389082074, + 0.002891332609578967, + 0.003051693318411708, + 0.005744066089391708, + 0.0010459712939336896, + 0.001273164409212768, + 0.002020755084231496, + 0.0022233447525650263, + 0.0014735718723386526, + 0.0013834057608619332, + 0.0011080149561166763, + 0.001617813017219305, + 0.001220153528265655, + 0.003198952181264758, + 0.0019091351423412561, + 0.0012376756640151143, + 0.0029229093343019485, + 0.0010224080178886652, + 0.0023435032926499844, + 0.0010688332840800285, + 0.0008326121023856103, + 0.0020608261693269014, + 0.0009941504104062915, + 0.002024134388193488, + 0.001707622199319303, + 0.0008987426990643144, + 0.0011877210345119238, + 0.0009225722751580179, + 0.0026006426196545362, + 0.003007967723533511, + 0.0012663973029702902, + 0.0018204919761046767, + 0.004243834409862757, + 0.0025879608001559973, + 0.0023917111102491617, + 0.0027510656509548426, + 0.002632628660649061, + 0.0038345518987625837, + 0.0018130376702174544, + 0.005260874051600695, + 0.004469802603125572, + 0.003642639145255089, + 0.005726683884859085, + 0.0035726851783692837, + 0.0015235479222610593, + 0.006101457867771387, + 0.0012628560652956367, + 0.0010768582578748465, + 0.01861828751862049, + 0.012301141396164894, + 0.0010346290655434132, + 0.0016465430380776525, + 0.0004872462886851281, + 0.0013481560163199902, + 0.0005959137342870235, + 0.00048096885439008474, + 0.004943754058331251, + 0.004176581744104624, + 0.000700945733115077, + 0.0009981805924326181, + 0.0010832847328856587, + 0.0015027394983917475, + 0.0015396879753097892, + 0.0011341954814270139, + 0.0022954351734369993, + 0.0028696588706225157, + 0.001491933362558484, + 0.0016877191374078393, + 0.033793721348047256, + 0.1058548092842102, + 0.017043547704815865, + 0.016184978187084198, + 0.33148109912872314, + 0.26066169142723083, + 0.015322684310376644, + 0.02732877805829048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.19007375836372375 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34869468212127686 + ], + "opposite_logprob_tokens": [ + -1.2236946821212769 + ], + "logprob_reward": -0.34869468212127686, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34869468212127686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08631566166877747, + 0.008115855045616627, + 0.022320877760648727, + 0.022547869011759758, + 0.02027103677392006, + 0.013474640436470509, + 0.027960633859038353, + 0.1676955670118332, + 0.1811370849609375, + 0.45016080141067505 + ], + "credit_map": [ + 0.009033496491611004, + 0.023662766441702843, + 0.011835153214633465, + 0.0058930180966854095, + 0.01222634594887495, + 0.007227687630802393, + 0.005839239805936813, + 0.010597959160804749, + 0.0008339454070664942, + 0.000922026636544615, + 0.0008825946133583784, + 0.00139951694291085, + 0.0010252223582938313, + 0.0012632099678739905, + 0.0007432374986819923, + 0.001046102144755423, + 0.0021565717179328203, + 0.0031988276168704033, + 0.0017016264609992504, + 0.0027385347057133913, + 0.0036519684363156557, + 0.0042070248164236546, + 0.0022773840464651585, + 0.002388943452388048, + 0.001318369060754776, + 0.003991825971752405, + 0.0014312039129436016, + 0.004837927408516407, + 0.003692296799272299, + 0.0014904839918017387, + 0.004270897246897221, + 0.0015148656675592065, + 0.0014601254370063543, + 0.0057266815565526485, + 0.0008343664812855422, + 0.0011167757911607623, + 0.00693139061331749, + 0.0015988099621608853, + 0.001576560316607356, + 0.001026329817250371, + 0.001875605434179306, + 0.0018083978211507201, + 0.0010263052536174655, + 0.0018431802745908499, + 0.003007911378517747, + 0.0015679793432354927, + 0.0011816826881840825, + 0.0011635789414867759, + 0.0016910632839426398, + 0.0031134404707700014, + 0.0010264000156894326, + 0.0010502880904823542, + 0.008339518681168556, + 0.010250601917505264, + 0.0008298085303977132, + 0.0016595189226791263, + 0.005739878863096237, + 0.016258379444479942, + 0.001483492087572813, + 0.0014374128077179193, + 0.0765569806098938, + 0.060709066689014435, + 0.0024041831493377686, + 0.0031061882618814707, + 0.00699321273714304, + 0.02128726616501808, + 0.0034992096479982138, + 0.0034988780971616507, + 0.0733024850487709, + 0.06161414831876755, + 0.004568603355437517, + 0.006373296957463026, + 0.01874595880508423, + 0.05913278087973595, + 0.009300819598138332, + 0.007935709320008755, + 0.18141485750675201, + 0.1512385755777359, + 0.008758309297263622, + 0.013633817434310913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.34869468212127686 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059059962630272 + ], + "opposite_logprob_tokens": [ + -3.8955905437469482 + ], + "logprob_reward": -0.02059059962630272, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059059962630272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047749780118465424, + 0.015089267864823341, + 0.015662450343370438, + 0.017316367477178574, + 0.029865633696317673, + 0.022274311631917953, + 0.06019004434347153, + 0.3109677731990814, + 0.05424925684928894, + 0.42663514614105225 + ], + "credit_map": [ + 0.003359281225129962, + 0.0047064730897545815, + 0.014080776832997799, + 0.005714747589081526, + 0.0033135595731437206, + 0.0032307696528732777, + 0.0038521108217537403, + 0.009492060169577599, + 0.0014509080210700631, + 0.0013787811622023582, + 0.0022488930262625217, + 0.0031600501388311386, + 0.0015160019975155592, + 0.002147263614460826, + 0.0012254054890945554, + 0.001961963251233101, + 0.0014641270972788334, + 0.004251271951943636, + 0.0013451651902869344, + 0.00131992029491812, + 0.004045781213790178, + 0.0012177270837128162, + 0.0009561232873238623, + 0.00106233439873904, + 0.0012488214997574687, + 0.004011236131191254, + 0.001098525826819241, + 0.00165913428645581, + 0.003960317932069302, + 0.0015343150589615107, + 0.0024605521466583014, + 0.001343462965451181, + 0.0022273878566920757, + 0.0077719613909721375, + 0.0013278295518830419, + 0.0020836410112679005, + 0.009361053816974163, + 0.002617052523419261, + 0.002402473473921418, + 0.0020742337219417095, + 0.001590824336744845, + 0.003941501025110483, + 0.001625342178158462, + 0.0022731320932507515, + 0.005387245211750269, + 0.002737196860834956, + 0.002595041412860155, + 0.0021240287460386753, + 0.002613025950267911, + 0.0033675371669232845, + 0.0014019059017300606, + 0.0012231037253513932, + 0.02414369024336338, + 0.02393292263150215, + 0.0013492011930793524, + 0.0021586494985967875, + 0.010277830064296722, + 0.0290053877979517, + 0.0026180294808000326, + 0.0018325959099456668, + 0.1428060382604599, + 0.11499826610088348, + 0.003918255213648081, + 0.0055113499984145164, + 0.002442752243950963, + 0.006629962008446455, + 0.0012132824631407857, + 0.00117533712182194, + 0.02090267464518547, + 0.01854277402162552, + 0.0010826238431036472, + 0.0022598463110625744, + 0.017884301021695137, + 0.05304054170846939, + 0.00854532327502966, + 0.008199231699109077, + 0.17689846456050873, + 0.14005863666534424, + 0.007985636591911316, + 0.014022964052855968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.02059059962630272 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740925367921591 + ], + "opposite_logprob_tokens": [ + -5.006741046905518 + ], + "logprob_reward": -0.006740925367921591, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740925367921591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05079265683889389, + 0.012179714627563953, + 0.02595529332756996, + 0.012199865654110909, + 0.03273647651076317, + 0.0166881512850523, + 0.1729636937379837, + 0.2444085329771042, + 0.01045463141053915, + 0.4216209948062897 + ], + "credit_map": [ + 0.004280485678464174, + 0.005097524728626013, + 0.014609183184802532, + 0.006710740737617016, + 0.004299554508179426, + 0.00409622211009264, + 0.004680579993873835, + 0.007018371019512415, + 0.0010740415891632438, + 0.0021868569310754538, + 0.0010466317180544138, + 0.0017627297202125192, + 0.0019362065941095352, + 0.0016271548811346292, + 0.0009853704832494259, + 0.0015607221284881234, + 0.0020251614041626453, + 0.006517262198030949, + 0.002227877965196967, + 0.001997478073462844, + 0.006021471228450537, + 0.002391138346865773, + 0.0029321997426450253, + 0.0018427054164931178, + 0.0009766402654349804, + 0.0032228354830294847, + 0.0004969756701029837, + 0.001490249065682292, + 0.00296084675937891, + 0.0009098810842260718, + 0.0014245155034586787, + 0.0007179214735515416, + 0.00235087051987648, + 0.009804865345358849, + 0.0014499594690278172, + 0.0021256208419799805, + 0.011716374196112156, + 0.0024454330559819937, + 0.0010636333608999848, + 0.0017797208856791258, + 0.0017580356216058135, + 0.0037982685025781393, + 0.001028949860483408, + 0.001290369895286858, + 0.004570167977362871, + 0.00167948252055794, + 0.001146438647992909, + 0.0014164383755996823, + 0.005534240044653416, + 0.016684845089912415, + 0.0020781203638762236, + 0.0013411169638857245, + 0.07863184064626694, + 0.06321300566196442, + 0.0023695623967796564, + 0.0031109570991247892, + 0.00820439774543047, + 0.022896433249115944, + 0.0019956158939749002, + 0.001548711908981204, + 0.11222401261329651, + 0.09020105004310608, + 0.0030570561066269875, + 0.004281267989426851, + 0.0010787436040118337, + 0.0021694074384868145, + 0.0005223405896686018, + 0.0006997619639150798, + 0.0017570570344105363, + 0.002545240567997098, + 0.0006455104448832572, + 0.00103656982537359, + 0.017126955091953278, + 0.052704352885484695, + 0.008870387449860573, + 0.008527965284883976, + 0.17348836362361908, + 0.1389118731021881, + 0.00833570770919323, + 0.013655396178364754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.006740925367921591 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076298960484564304 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.0076298960484564304, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076298960484564304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13163328170776367, + 0.039005719125270844, + 0.025927379727363586, + 0.02245277352631092, + 0.050696201622486115, + 0.0687328651547432, + 0.16813603043556213, + 0.37614214420318604, + 0.022569430992007256, + 0.09470411390066147 + ], + "credit_map": [ + 0.01067553274333477, + 0.028948254883289337, + 0.019143618643283844, + 0.014604226686060429, + 0.01747189275920391, + 0.013878130353987217, + 0.01224004477262497, + 0.014671597629785538, + 0.0024738300126045942, + 0.007265498861670494, + 0.003820263547822833, + 0.009363911114633083, + 0.005214073695242405, + 0.0031598657369613647, + 0.0027742565143853426, + 0.004934024065732956, + 0.0021776172798126936, + 0.0028199097141623497, + 0.003923832904547453, + 0.003188815899193287, + 0.003135528415441513, + 0.004364199936389923, + 0.0026109612081199884, + 0.0037065164651721716, + 0.0014345388626679778, + 0.007096824701875448, + 0.0010348329087719321, + 0.001603060751222074, + 0.007030986249446869, + 0.001160637242719531, + 0.001880153315141797, + 0.0012117415899410844, + 0.0059904237277805805, + 0.008397403173148632, + 0.0032729364465922117, + 0.007107593584805727, + 0.012296763248741627, + 0.004764895420521498, + 0.0040867626667022705, + 0.004779424984008074, + 0.00727678881958127, + 0.013974836096167564, + 0.004398295655846596, + 0.007127574644982815, + 0.012687907554209232, + 0.008484282530844212, + 0.006987255997955799, + 0.007795932237058878, + 0.00577789731323719, + 0.015939021483063698, + 0.004034075420349836, + 0.002566775307059288, + 0.07218849658966064, + 0.05940947309136391, + 0.00400919932872057, + 0.00421110587194562, + 0.012703930959105492, + 0.0375385656952858, + 0.003453536657616496, + 0.00265504140406847, + 0.17310966551303864, + 0.13511055707931519, + 0.005162663757801056, + 0.006408216431736946, + 0.001895280322059989, + 0.0045488025061786175, + 0.0020730006508529186, + 0.0015246463008224964, + 0.0045207180082798, + 0.0033322935923933983, + 0.0018848032923415303, + 0.002789886435493827, + 0.010727223008871078, + 0.024478061124682426, + 0.006738383788615465, + 0.004858724307268858, + 0.020617853850126266, + 0.012731252238154411, + 0.005885590333491564, + 0.008667037822306156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.0076298960484564304 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, sweeping, long shot, shot from a slightly upward perspective, captures a tranquil, moonless summer night. Over a secluded meadow, surrounded by the dark silhouette of an ancient forest, the deep indigo sky is dotted with sparse stars. A gentle breeze rustles the tall grass and distant leaves with a barely audible rustle. At the center of the image is a magnificent ribbon of light, emanating a warm, pulsating yellow-green bioluminescence from countless fireflies. This ethereal river of light slowly and gracefully undulates and swirls through the sky, like a living nebula. Individual fireflies dart and pause within the ribbon, collectively weaving this flowing tapestry of light. Their paths, guided by the gentle breeze, alternately become dense, like brilliant star clusters, and sparse, like a wisp of ethereal light. Their light casts a soft, dreamy halo in the absolute darkness, faintly illuminating the tips of blades of grass and motes of dust floating in the air, creating a magical atmosphere of profound tranquility and the wonder of life.", + "seed": 11785, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_1/index_7881_seed_11785_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0855975354090333, + "vlm_logprob": -0.0855975354090333, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.0855975354090333, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02039041742682457 + ], + "opposite_logprob_tokens": [ + -4.020390510559082 + ], + "logprob_reward": -0.02039041742682457, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02039041742682457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06195183843374252, + 0.021213041618466377, + 0.12239190191030502, + 0.06112276390194893, + 0.02162995934486389, + 0.03408592566847801, + 0.02144712582230568, + 0.06065160036087036, + 0.031116584315896034, + 0.564389169216156 + ], + "credit_map": [ + 0.007150179706513882, + 0.015348897315561771, + 0.0031223823316395283, + 0.005282142665237188, + 0.006892118137329817, + 0.006252956110984087, + 0.006386346183717251, + 0.011516819708049297, + 0.0020911325700581074, + 0.0034728366881608963, + 0.0016410443931818008, + 0.002765907905995846, + 0.00340136862359941, + 0.0025555421598255634, + 0.0027922373265028, + 0.0024929738137871027, + 0.02998216263949871, + 0.01344167161732912, + 0.0023830053396523, + 0.0023185363970696926, + 0.05188887193799019, + 0.006072802469134331, + 0.0029975061770528555, + 0.013307353481650352, + 0.007924447767436504, + 0.015399124473333359, + 0.0031366501934826374, + 0.003392139682546258, + 0.013583634048700333, + 0.009732679463922977, + 0.002583936555311084, + 0.005370153114199638, + 0.0025238939560949802, + 0.005727671552449465, + 0.0011805803515017033, + 0.0014123977161943913, + 0.004322370048612356, + 0.003071737242862582, + 0.0016993986209854484, + 0.0016919117188081145, + 0.004185707774013281, + 0.005627715960144997, + 0.0020338098984211683, + 0.0035773240961134434, + 0.008583137765526772, + 0.00397550268098712, + 0.003725645365193486, + 0.002377084456384182, + 0.0019231304759159684, + 0.0033165772911161184, + 0.0024023817386478186, + 0.002465129131451249, + 0.0026579962577670813, + 0.003248057793825865, + 0.003350347513332963, + 0.002083507599309087, + 0.006195632740855217, + 0.011796670965850353, + 0.003133450634777546, + 0.004788389429450035, + 0.012270280160009861, + 0.012162523344159126, + 0.0053411624394357204, + 0.004963495768606663, + 0.002472618827596307, + 0.009782109409570694, + 0.0009614457376301289, + 0.0010855338769033551, + 0.010619554668664932, + 0.0025722456630319357, + 0.0018478562124073505, + 0.001775221200659871, + 0.014842752367258072, + 0.23565131425857544, + 0.004488327074795961, + 0.006057179532945156, + 0.2585994303226471, + 0.02908727526664734, + 0.008113219402730465, + 0.007549694739282131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.02039041742682457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38710400462150574 + ], + "opposite_logprob_tokens": [ + -1.1371040344238281 + ], + "logprob_reward": -0.38710400462150574, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38710400462150574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1392633318901062, + 0.050524283200502396, + 0.31587088108062744, + 0.02182203345000744, + 0.057533856481313705, + 0.10156048089265823, + 0.015785448253154755, + 0.02624836564064026, + 0.02999170683324337, + 0.24139957129955292 + ], + "credit_map": [ + 0.018556825816631317, + 0.040756259113550186, + 0.0060177454724907875, + 0.006757847964763641, + 0.019859973341226578, + 0.01194760575890541, + 0.008438030257821083, + 0.026929039508104324, + 0.004188604187220335, + 0.008815327659249306, + 0.002967578824609518, + 0.003599934745579958, + 0.010870796628296375, + 0.0049293022602796555, + 0.009979567490518093, + 0.005173173267394304, + 0.06604887545108795, + 0.035983044654130936, + 0.0073420023545622826, + 0.006373279727995396, + 0.12918508052825928, + 0.017705179750919342, + 0.008720253594219685, + 0.04451315104961395, + 0.003031394677236676, + 0.003200770588591695, + 0.00160907709505409, + 0.0020466898567974567, + 0.0050139231607317924, + 0.0037138238549232483, + 0.0010887366952374578, + 0.002117617055773735, + 0.004624542780220509, + 0.011948944069445133, + 0.002589879324659705, + 0.004047914873808622, + 0.009429802186787128, + 0.015222233720123768, + 0.002816050313413143, + 0.006854483857750893, + 0.009235243313014507, + 0.03008974902331829, + 0.005561667028814554, + 0.004545051604509354, + 0.03266951069235802, + 0.008562569506466389, + 0.006246103905141354, + 0.00465058721601963, + 0.0018373022321611643, + 0.004046731162816286, + 0.0010404778877273202, + 0.0015654085436835885, + 0.0019831571262329817, + 0.0022459456231445074, + 0.001659925328567624, + 0.0014065003488212824, + 0.0028576867189258337, + 0.004135028924793005, + 0.0013302074512466788, + 0.0026273312978446484, + 0.004792869556695223, + 0.005765767768025398, + 0.002329018199816346, + 0.0024104563053697348, + 0.004432990215718746, + 0.0060040331445634365, + 0.001552610076032579, + 0.0021737017668783665, + 0.007908041588962078, + 0.0036984984762966633, + 0.0024465324822813272, + 0.0017752982676029205, + 0.008572372607886791, + 0.09557664394378662, + 0.002795181469991803, + 0.0036694000009447336, + 0.10641373693943024, + 0.012934904545545578, + 0.00633661774918437, + 0.005100721959024668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.38710400462150574 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008655413053929806 + ], + "opposite_logprob_tokens": [ + -4.758655548095703 + ], + "logprob_reward": -0.008655413053929806, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008655413053929806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03772178664803505, + 0.012179039418697357, + 0.7047879099845886, + 0.008821011520922184, + 0.01545708253979683, + 0.047610532492399216, + 0.008594177663326263, + 0.010187620297074318, + 0.021150462329387665, + 0.13349033892154694 + ], + "credit_map": [ + 0.0049851080402731895, + 0.010861078277230263, + 0.0016258121468126774, + 0.0019330729264765978, + 0.00467921094968915, + 0.0035153112839907408, + 0.002083222847431898, + 0.00803897064179182, + 0.001132593723013997, + 0.002257181564345956, + 0.0007644663564860821, + 0.000949005363509059, + 0.002687678439542651, + 0.0013282854342833161, + 0.0018971560057252645, + 0.0011626725317910314, + 0.16491159796714783, + 0.07966134697198868, + 0.014642182737588882, + 0.01131599023938179, + 0.29926738142967224, + 0.03526652231812477, + 0.015913276001811028, + 0.08380960673093796, + 0.0010486069368198514, + 0.0013730868231505156, + 0.0007298532873392105, + 0.000863754830788821, + 0.0015674089081585407, + 0.0017736201407387853, + 0.0005554771632887423, + 0.0009092037216760218, + 0.0014414334436878562, + 0.003328623017296195, + 0.0008201990276575089, + 0.001161798951216042, + 0.002071055583655834, + 0.0035559339448809624, + 0.0014153599040582776, + 0.001662678667344153, + 0.004522272385656834, + 0.013487421907484531, + 0.0027918617706745863, + 0.0023546265438199043, + 0.014817798510193825, + 0.00427933968603611, + 0.0030800874810665846, + 0.0022771270014345646, + 0.0010057201143354177, + 0.002073539886623621, + 0.0006349568720906973, + 0.0009363949066027999, + 0.0014001354575157166, + 0.0008371839066967368, + 0.001020388095639646, + 0.0006858580745756626, + 0.0012281073722988367, + 0.003325042547658086, + 0.00042191307875327766, + 0.000706483842805028, + 0.0017222980968654156, + 0.00112341798376292, + 0.0010993117466568947, + 0.0005610453081317246, + 0.0021287323907017708, + 0.0038721465971320868, + 0.0008449103333987296, + 0.0012855780078098178, + 0.006748305168002844, + 0.0024029798805713654, + 0.002367683220654726, + 0.0015001250430941582, + 0.0037766871973872185, + 0.0551743358373642, + 0.0009566692169755697, + 0.001352371065877378, + 0.061082642525434494, + 0.0069828699342906475, + 0.002285589464008808, + 0.0018791801994666457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 5 + ], + "selected_reward": -0.008655413053929806 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031969649717211723 + ], + "opposite_logprob_tokens": [ + -5.753197193145752 + ], + "logprob_reward": -0.0031969649717211723, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031969649717211723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0646180734038353, + 0.017044635489583015, + 0.3998260498046875, + 0.022917425259947777, + 0.01443792600184679, + 0.043024953454732895, + 0.028136899694800377, + 0.04620378091931343, + 0.05745086073875427, + 0.3063393831253052 + ], + "credit_map": [ + 0.0069762603379786015, + 0.014521309174597263, + 0.0035453634336590767, + 0.005074365064501762, + 0.011274135671555996, + 0.005177009385079145, + 0.007714670151472092, + 0.010334963910281658, + 0.0018634378211572766, + 0.0029629890341311693, + 0.0014378190971910954, + 0.0019793296232819557, + 0.0031925637740641832, + 0.0017050454625859857, + 0.0019902922213077545, + 0.001913160551339388, + 0.10154492408037186, + 0.04519796371459961, + 0.00819298718124628, + 0.00764253456145525, + 0.1709810495376587, + 0.01985926367342472, + 0.009045439772307873, + 0.037361886352300644, + 0.002266785129904747, + 0.0024823127314448357, + 0.0020451927557587624, + 0.002222691895440221, + 0.002367235254496336, + 0.006346618290990591, + 0.0020323372446000576, + 0.0031542533542960882, + 0.001649415004067123, + 0.003289420623332262, + 0.000865478185005486, + 0.0013334176037460566, + 0.002164470963180065, + 0.0023876975756138563, + 0.0014964104630053043, + 0.0012516158167272806, + 0.0028287083841860294, + 0.015344471670687199, + 0.0010409234091639519, + 0.002054274547845125, + 0.01610054075717926, + 0.0024767981376498938, + 0.0017628200585022569, + 0.0014164166059345007, + 0.0036179819144308567, + 0.007278131321072578, + 0.0018250447465106845, + 0.002879166742786765, + 0.0032604204025119543, + 0.003699557390064001, + 0.003371297148987651, + 0.0022052975837141275, + 0.005285330582410097, + 0.0086657190695405, + 0.0020940809044986963, + 0.0036618055310100317, + 0.007923483848571777, + 0.009717835113406181, + 0.00509549118578434, + 0.0037600311916321516, + 0.004559212364256382, + 0.016687069088220596, + 0.0018533575348556042, + 0.0027613702695816755, + 0.019730929285287857, + 0.005744757130742073, + 0.003772015916183591, + 0.002342150080949068, + 0.008196787908673286, + 0.1271662712097168, + 0.002316080965101719, + 0.002926595974713564, + 0.14126364886760712, + 0.01594792678952217, + 0.004399721045047045, + 0.004122336860746145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.0031969649717211723 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640876971185207 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640876971185207, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640876971185207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15618042647838593, + 0.012743554078042507, + 0.20399336516857147, + 0.07128339260816574, + 0.01664205826818943, + 0.09328047931194305, + 0.021644556894898415, + 0.0435982383787632, + 0.1268799602985382, + 0.2537539303302765 + ], + "credit_map": [ + 0.023244746029376984, + 0.05123278498649597, + 0.004847133066505194, + 0.004944152198731899, + 0.01789032854139805, + 0.014821778051555157, + 0.004506594035774469, + 0.0346929207444191, + 0.0010487680556252599, + 0.00222756527364254, + 0.0007527890265919268, + 0.0016964901005849242, + 0.0021024979650974274, + 0.001344582298770547, + 0.002190721221268177, + 0.0013801410095766187, + 0.037994448095560074, + 0.026834413409233093, + 0.00471634604036808, + 0.00488791149109602, + 0.08288808166980743, + 0.01018150057643652, + 0.005041603464633226, + 0.031449075788259506, + 0.008555079810321331, + 0.013429193757474422, + 0.003478047903627157, + 0.003637366695329547, + 0.015119139105081558, + 0.015610192902386189, + 0.0035844596568495035, + 0.007869919762015343, + 0.0017753089778125286, + 0.0027693952433764935, + 0.0008270658436231315, + 0.0012193727307021618, + 0.002439306117594242, + 0.004505678545683622, + 0.001075785607099533, + 0.0020301472395658493, + 0.006989586632698774, + 0.03051070310175419, + 0.004459537100046873, + 0.003709580283612013, + 0.031765565276145935, + 0.007089984603226185, + 0.005041956435889006, + 0.0037135735619813204, + 0.003546708496287465, + 0.0036106144543737173, + 0.0014965332811698318, + 0.002319131512194872, + 0.0027001670096069574, + 0.004205228760838509, + 0.001912859734147787, + 0.00185331457760185, + 0.0044699981808662415, + 0.011496162042021751, + 0.0019896847661584616, + 0.0040680766105651855, + 0.007047176361083984, + 0.00851703342050314, + 0.002862791996449232, + 0.0031473201233893633, + 0.010502679273486137, + 0.039369694888591766, + 0.00370616652071476, + 0.005587999243289232, + 0.04288094490766525, + 0.010779855772852898, + 0.009189111180603504, + 0.0048635187558829784, + 0.007450693752616644, + 0.10636656731367111, + 0.0015893828822299838, + 0.002835286781191826, + 0.11648686975240707, + 0.012849217280745506, + 0.003067947691306472, + 0.003107985481619835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.008640876971185207 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, sweeping, long shot, shot from a slightly upward perspective, captures a tranquil, moonless summer night. Over a secluded meadow, surrounded by the dark silhouette of an ancient forest, the deep indigo sky is dotted with sparse stars. A gentle breeze rustles the tall grass and distant leaves with a barely audible rustle. At the center of the image is a magnificent ribbon of light, emanating a warm, pulsating yellow-green bioluminescence from countless fireflies. This ethereal river of light slowly and gracefully undulates and swirls through the sky, like a living nebula. Individual fireflies dart and pause within the ribbon, collectively weaving this flowing tapestry of light. Their paths, guided by the gentle breeze, alternately become dense, like brilliant star clusters, and sparse, like a wisp of ethereal light. Their light casts a soft, dreamy halo in the absolute darkness, faintly illuminating the tips of blades of grass and motes of dust floating in the air, creating a magical atmosphere of profound tranquility and the wonder of life.", + "seed": 11786, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_1/index_7881_seed_11786_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05769181791692972, + "vlm_logprob": -0.05769181791692972, + "vlm_yes_no_margin": 3.6499999523162843, + "vlm_token_credit": -0.05769181791692972, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015368256717920303 + ], + "opposite_logprob_tokens": [ + -4.390368461608887 + ], + "logprob_reward": -0.015368256717920303, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015368256717920303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10612766444683075, + 0.02183731459081173, + 0.018543295562267303, + 0.01965157687664032, + 0.010728910565376282, + 0.013996001332998276, + 0.0459815189242363, + 0.04499262571334839, + 0.01495994534343481, + 0.7031811475753784 + ], + "credit_map": [ + 0.010064723901450634, + 0.008707938715815544, + 0.02372155524790287, + 0.014424260705709457, + 0.007024376653134823, + 0.011857726611196995, + 0.014852691441774368, + 0.015474381856620312, + 0.0018331416649743915, + 0.0018196947639808059, + 0.002836968982592225, + 0.0029364340007305145, + 0.001833633636124432, + 0.002778969705104828, + 0.004622886423021555, + 0.003175583900883794, + 0.0019835438579320908, + 0.0018114004051312804, + 0.004270295612514019, + 0.0020042064134031534, + 0.0009296732605434954, + 0.0026966468431055546, + 0.0014650959055870771, + 0.0033824315760284662, + 0.0016442055348306894, + 0.0010493469890207052, + 0.004742853809148073, + 0.0035582254640758038, + 0.0009568340028636158, + 0.00179476966150105, + 0.001984774135053158, + 0.0039205653592944145, + 0.0010673213982954621, + 0.0011651403037831187, + 0.003062434494495392, + 0.0011471925536170602, + 0.0006132558337412775, + 0.0012100907042622566, + 0.0014201707672327757, + 0.00104330328758806, + 0.002418022369965911, + 0.002436620183289051, + 0.002138272626325488, + 0.0018171386327594519, + 0.001072552870027721, + 0.0014945463044568896, + 0.001389260753057897, + 0.0012295859633013606, + 0.005096524953842163, + 0.0018511267844587564, + 0.017262600362300873, + 0.005305579397827387, + 0.003139494452625513, + 0.005688960663974285, + 0.002724339487031102, + 0.00491288909688592, + 0.005618848837912083, + 0.0029056640341877937, + 0.007316753268241882, + 0.0033949140924960375, + 0.004055658355355263, + 0.01609085313975811, + 0.0020272103138267994, + 0.0035827243700623512, + 0.0017717863665893674, + 0.0012437145924195647, + 0.0033005457371473312, + 0.0025490159168839455, + 0.0009746557334437966, + 0.002587953582406044, + 0.0012504315236583352, + 0.0012818407267332077, + 0.05237972363829613, + 0.013449950143694878, + 0.2815977931022644, + 0.04328155517578125, + 0.017973551526665688, + 0.2715987265110016, + 0.0107545992359519, + 0.012145196087658405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.015368256717920303 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22561927139759064 + ], + "opposite_logprob_tokens": [ + -1.6006193161010742 + ], + "logprob_reward": -0.22561927139759064, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22561927139759064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21366046369075775, + 0.025091486051678658, + 0.05006510019302368, + 0.024455823004245758, + 0.01651257462799549, + 0.023864159360527992, + 0.20227882266044617, + 0.11708663403987885, + 0.0641796663403511, + 0.2628052830696106 + ], + "credit_map": [ + 0.009650151245296001, + 0.0106696467846632, + 0.055617742240428925, + 0.037421245127916336, + 0.008759084157645702, + 0.011772961355745792, + 0.0264463908970356, + 0.053323253989219666, + 0.002595619298517704, + 0.002539420500397682, + 0.002707939362153411, + 0.002615000819787383, + 0.0020591632928699255, + 0.0032050320878624916, + 0.004739001393318176, + 0.004630310460925102, + 0.00367607525549829, + 0.0025410654488950968, + 0.014915003441274166, + 0.008455192670226097, + 0.002412239322438836, + 0.003510896349325776, + 0.006468703038990498, + 0.008085927926003933, + 0.0025574706960469484, + 0.0018037984846159816, + 0.005417868960648775, + 0.0028942732606083155, + 0.0017803184455260634, + 0.002908510621637106, + 0.0026419060304760933, + 0.0044516753405332565, + 0.0018072533421218395, + 0.002139768796041608, + 0.0031157154589891434, + 0.0017907032743096352, + 0.0008800082141533494, + 0.002222198760136962, + 0.002180911600589752, + 0.002376015530899167, + 0.003389097983017564, + 0.0029822790529578924, + 0.00391386728733778, + 0.0033844716381281614, + 0.001891227555461228, + 0.003622961463406682, + 0.002368774963542819, + 0.002311478368937969, + 0.0231525506824255, + 0.007915317080914974, + 0.07998943328857422, + 0.018247535452246666, + 0.015843460336327553, + 0.028071623295545578, + 0.010854203253984451, + 0.018204696476459503, + 0.01750226691365242, + 0.007381567731499672, + 0.014217516407370567, + 0.016611985862255096, + 0.007064910605549812, + 0.02300667017698288, + 0.010294176638126373, + 0.021007545292377472, + 0.00884926039725542, + 0.0054396395571529865, + 0.011525042355060577, + 0.008488110266625881, + 0.005221112631261349, + 0.012308798730373383, + 0.005908611696213484, + 0.006439085118472576, + 0.018980693072080612, + 0.005770237185060978, + 0.10035064071416855, + 0.01865639165043831, + 0.006753501482307911, + 0.0963020846247673, + 0.007527039851993322, + 0.008464698679745197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.22561927139759064 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201468512415886 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201468512415886, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018201468512415886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13517314195632935, + 0.060266539454460144, + 0.1460595726966858, + 0.019223101437091827, + 0.05425950139760971, + 0.061869896948337555, + 0.10338624566793442, + 0.1350325345993042, + 0.07144701480865479, + 0.213282510638237 + ], + "credit_map": [ + 0.008601839654147625, + 0.0061735608614981174, + 0.03208811953663826, + 0.01861407980322838, + 0.005534645635634661, + 0.008146656677126884, + 0.028034182265400887, + 0.027980053797364235, + 0.005988428369164467, + 0.004203834570944309, + 0.006657678633928299, + 0.00788850151002407, + 0.0026248206850141287, + 0.005516634322702885, + 0.01660536788403988, + 0.010781269520521164, + 0.010022015310823917, + 0.005230289418250322, + 0.05600976571440697, + 0.013514544814825058, + 0.00512453680858016, + 0.010467483662068844, + 0.008936216123402119, + 0.03675471618771553, + 0.001861810334958136, + 0.001758344005793333, + 0.005031417589634657, + 0.0023624184541404247, + 0.00102382677141577, + 0.001745104673318565, + 0.0017289997776970267, + 0.0037111793644726276, + 0.005803057923913002, + 0.007564842235296965, + 0.01019582711160183, + 0.009170395322144032, + 0.0027721270453184843, + 0.006229165941476822, + 0.007383577525615692, + 0.005140502937138081, + 0.008008483797311783, + 0.00525517575442791, + 0.01497576106339693, + 0.011815991252660751, + 0.004207737743854523, + 0.005494371056556702, + 0.006296891253441572, + 0.00581548735499382, + 0.010968311689794064, + 0.004144549369812012, + 0.03788280859589577, + 0.012247518636286259, + 0.006384479347616434, + 0.012412279844284058, + 0.006225742865353823, + 0.01312054693698883, + 0.01675144024193287, + 0.0074882498010993, + 0.02246234379708767, + 0.013860514387488365, + 0.012289297766983509, + 0.04108051583170891, + 0.007586042862385511, + 0.013514126650989056, + 0.006473119370639324, + 0.00568053824827075, + 0.013008412905037403, + 0.014961006119847298, + 0.0047064293175935745, + 0.012343358248472214, + 0.008237333968281746, + 0.006036816630512476, + 0.01496700756251812, + 0.004474938381463289, + 0.08177758753299713, + 0.014244399033486843, + 0.0058488924987614155, + 0.0781603679060936, + 0.0064341360703110695, + 0.0073751662857830524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.018201468512415886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108712051063776 + ], + "opposite_logprob_tokens": [ + -4.511086940765381 + ], + "logprob_reward": -0.01108712051063776, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108712051063776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16789522767066956, + 0.06109803542494774, + 0.18490007519721985, + 0.048679325729608536, + 0.055079810321331024, + 0.026350557804107666, + 0.06921130418777466, + 0.11668925732374191, + 0.1394999474287033, + 0.13059647381305695 + ], + "credit_map": [ + 0.01255861297249794, + 0.011901804246008396, + 0.04552028700709343, + 0.019161300733685493, + 0.008766181766986847, + 0.011497875675559044, + 0.017495226114988327, + 0.04099392890930176, + 0.00543629378080368, + 0.004268815275281668, + 0.008718003518879414, + 0.008309655822813511, + 0.004073312971740961, + 0.006272091064602137, + 0.014287358149886131, + 0.009732497856020927, + 0.012296615168452263, + 0.006336412392556667, + 0.07316448539495468, + 0.017916664481163025, + 0.005770156159996986, + 0.011789296753704548, + 0.013894723728299141, + 0.04373171925544739, + 0.0037729863543063402, + 0.002388287801295519, + 0.014731298200786114, + 0.006074984557926655, + 0.0023961407132446766, + 0.0035651305224746466, + 0.004322158172726631, + 0.011428337544202805, + 0.005718978121876717, + 0.005572210997343063, + 0.014287631958723068, + 0.005024624522775412, + 0.0028385664336383343, + 0.00625219801440835, + 0.009778985753655434, + 0.00560661219060421, + 0.0030392084736377, + 0.0016188316512852907, + 0.007339412812143564, + 0.002906140638515353, + 0.0014190446818247437, + 0.0032893891911953688, + 0.0027050147764384747, + 0.0040335156954824924, + 0.006834070198237896, + 0.003186715766787529, + 0.020372530445456505, + 0.009225987829267979, + 0.003200000151991844, + 0.006572342477738857, + 0.005292906891554594, + 0.014526749029755592, + 0.011839444749057293, + 0.009638064540922642, + 0.013160218484699726, + 0.021929631009697914, + 0.007776126731187105, + 0.017698384821414948, + 0.019045570865273476, + 0.015601815655827522, + 0.013662660494446754, + 0.011128300800919533, + 0.02248663455247879, + 0.0348798930644989, + 0.011607772670686245, + 0.01783451810479164, + 0.017156142741441727, + 0.010744011029601097, + 0.009519841521978378, + 0.003373634535819292, + 0.048901841044425964, + 0.00960763543844223, + 0.004193764179944992, + 0.04765761271119118, + 0.004173720721155405, + 0.0031684052664786577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.01108712051063776 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182972446084023 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.018182972446084023, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018182972446084023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18254032731056213, + 0.02695060335099697, + 0.19886338710784912, + 0.029049184173345566, + 0.10531478375196457, + 0.024178199470043182, + 0.15697015821933746, + 0.091114342212677, + 0.06760234385728836, + 0.11741664260625839 + ], + "credit_map": [ + 0.016328666359186172, + 0.015186376869678497, + 0.027061842381954193, + 0.017833877354860306, + 0.010554535314440727, + 0.024791782721877098, + 0.03095763921737671, + 0.03982561454176903, + 0.002488914877176285, + 0.0028159096837043762, + 0.0035051594022661448, + 0.003937351051717997, + 0.00230529997497797, + 0.003062197705730796, + 0.005352981388568878, + 0.0034827892668545246, + 0.012556448578834534, + 0.0035219124983996153, + 0.08595841377973557, + 0.016967007890343666, + 0.00560199748724699, + 0.01072210818529129, + 0.010225711390376091, + 0.053309787064790726, + 0.004067746456712484, + 0.0034130290150642395, + 0.003639441914856434, + 0.005350334104150534, + 0.0016226860461756587, + 0.004081141669303179, + 0.0027362152468413115, + 0.004138588905334473, + 0.013476278632879257, + 0.015149800106883049, + 0.025277230888605118, + 0.010383154265582561, + 0.0064572542905807495, + 0.011762204580008984, + 0.014321526512503624, + 0.0084873391315341, + 0.0032722598407417536, + 0.0031246438156813383, + 0.005053650122135878, + 0.0026854872703552246, + 0.0016838156152516603, + 0.0033286043908447027, + 0.0020105629228055477, + 0.0030191754922270775, + 0.017703384160995483, + 0.007183992303907871, + 0.06048603355884552, + 0.014459061436355114, + 0.010927306488156319, + 0.020477572456002235, + 0.0088869109749794, + 0.016845904290676117, + 0.012755296193063259, + 0.005866196937859058, + 0.01473397295922041, + 0.00972724799066782, + 0.007444360759109259, + 0.028247548267245293, + 0.003961283713579178, + 0.008378435857594013, + 0.010526370257139206, + 0.006023405585438013, + 0.012472766451537609, + 0.009837036952376366, + 0.004257860127836466, + 0.014728925190865993, + 0.004975983407348394, + 0.004779987037181854, + 0.009320446290075779, + 0.004172006156295538, + 0.041768524795770645, + 0.009230825118720531, + 0.003924479708075523, + 0.042708516120910645, + 0.0032028069254010916, + 0.003089035162702203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.018182972446084023 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, sweeping, long shot, shot from a slightly upward perspective, captures a tranquil, moonless summer night. Over a secluded meadow, surrounded by the dark silhouette of an ancient forest, the deep indigo sky is dotted with sparse stars. A gentle breeze rustles the tall grass and distant leaves with a barely audible rustle. At the center of the image is a magnificent ribbon of light, emanating a warm, pulsating yellow-green bioluminescence from countless fireflies. This ethereal river of light slowly and gracefully undulates and swirls through the sky, like a living nebula. Individual fireflies dart and pause within the ribbon, collectively weaving this flowing tapestry of light. Their paths, guided by the gentle breeze, alternately become dense, like brilliant star clusters, and sparse, like a wisp of ethereal light. Their light casts a soft, dreamy halo in the absolute darkness, faintly illuminating the tips of blades of grass and motes of dust floating in the air, creating a magical atmosphere of profound tranquility and the wonder of life.", + "seed": 11787, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_1/index_7881_seed_11787_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12030500434339046, + "vlm_logprob": -0.12030500434339046, + "vlm_yes_no_margin": 2.849999988079071, + "vlm_token_credit": -0.12030500434339046, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09178373962640762 + ], + "opposite_logprob_tokens": [ + -2.4667837619781494 + ], + "logprob_reward": -0.09178373962640762, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09178373962640762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3706579804420471, + 0.056136175990104675, + 0.03130405768752098, + 0.08180707693099976, + 0.11594170331954956, + 0.055134985595941544, + 0.0363367423415184, + 0.03162761777639389, + 0.02844354510307312, + 0.19261017441749573 + ], + "credit_map": [ + 0.03799040615558624, + 0.10592838376760483, + 0.020360250025987625, + 0.013762960210442543, + 0.09838755428791046, + 0.03856422379612923, + 0.028397969901561737, + 0.027266209945082664, + 0.005405297502875328, + 0.020041676238179207, + 0.002679527271538973, + 0.0026238001883029938, + 0.01066933199763298, + 0.0056863208301365376, + 0.003608695464208722, + 0.00542152114212513, + 0.0037313131615519524, + 0.004310936201363802, + 0.0028856340795755386, + 0.004104622174054384, + 0.006810431834310293, + 0.00386632326990366, + 0.002368898829445243, + 0.003225889755412936, + 0.0071041164919734, + 0.013287030160427094, + 0.0033222115598618984, + 0.005280979443341494, + 0.025249604135751724, + 0.018510831519961357, + 0.004498360678553581, + 0.004553929902613163, + 0.008992897346615791, + 0.0234431903809309, + 0.009367979131639004, + 0.022357244044542313, + 0.014655549079179764, + 0.010620905086398125, + 0.016450723633170128, + 0.01005319319665432, + 0.0041288710199296474, + 0.009578448720276356, + 0.004284703638404608, + 0.009106313809752464, + 0.012563097290694714, + 0.006001794245094061, + 0.004638573620468378, + 0.00483317393809557, + 0.0028716465458273888, + 0.012988202273845673, + 0.001628769445233047, + 0.002631042618304491, + 0.007769204676151276, + 0.0035844934172928333, + 0.0018116217106580734, + 0.003051756415516138, + 0.0022048058453947306, + 0.007318743970245123, + 0.001973900478333235, + 0.003244867315515876, + 0.005676401313394308, + 0.004731867462396622, + 0.00402324041351676, + 0.0024537842255085707, + 0.002603961853310466, + 0.004761320073157549, + 0.0028437592554837465, + 0.0034238153602927923, + 0.0057352506555616856, + 0.003640859154984355, + 0.0033883568830788136, + 0.0020462183747440577, + 0.0135429622605443, + 0.032134149223566055, + 0.015644602477550507, + 0.03423987329006195, + 0.03716342896223068, + 0.01660522259771824, + 0.02640051767230034, + 0.016879385337233543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.09178373962640762 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42893511056900024 + ], + "opposite_logprob_tokens": [ + -1.0539350509643555 + ], + "logprob_reward": -0.42893511056900024, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42893511056900024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32573434710502625, + 0.10107769817113876, + 0.04294048622250557, + 0.08234269171953201, + 0.04871950298547745, + 0.02017025090754032, + 0.0412660650908947, + 0.14644037187099457, + 0.019407952204346657, + 0.17190061509609222 + ], + "credit_map": [ + 0.030765119940042496, + 0.11496491730213165, + 0.01584761217236519, + 0.007156652864068747, + 0.09611774235963821, + 0.03365888074040413, + 0.01318519189953804, + 0.01403826754540205, + 0.008979842066764832, + 0.026468532159924507, + 0.005150454118847847, + 0.004823089111596346, + 0.0276291836053133, + 0.014948872849345207, + 0.005751905497163534, + 0.0073258234187960625, + 0.004487545229494572, + 0.011267106048762798, + 0.0023135251831263304, + 0.00407752487808466, + 0.009369464591145515, + 0.007017266005277634, + 0.002173346234485507, + 0.0022347101476043463, + 0.00612926809117198, + 0.015148468315601349, + 0.00255506276153028, + 0.003950407728552818, + 0.028675807639956474, + 0.018105199560523033, + 0.004885697737336159, + 0.0028927898965775967, + 0.004133610986173153, + 0.013845043256878853, + 0.0037183864042162895, + 0.005578959826380014, + 0.009369269013404846, + 0.006126929074525833, + 0.003116161096841097, + 0.0028311447240412235, + 0.0015579652972519398, + 0.0033386838622391224, + 0.0015948695363476872, + 0.0031228617299348116, + 0.00375957228243351, + 0.0025258376263082027, + 0.002301926026120782, + 0.0019685367587953806, + 0.003348568920046091, + 0.014778473414480686, + 0.0019000651082023978, + 0.00321540841832757, + 0.009431974031031132, + 0.005314123351126909, + 0.0015417137183248997, + 0.0017357415053993464, + 0.010087549686431885, + 0.052872009575366974, + 0.0024851199705153704, + 0.0035668471828103065, + 0.050221752375364304, + 0.01946350373327732, + 0.004147252067923546, + 0.0035963428672403097, + 0.0015418959083035588, + 0.0053978837095201015, + 0.0009956259746104479, + 0.0012096790596842766, + 0.004701576661318541, + 0.002921833423897624, + 0.0015661687357351184, + 0.0010732898954302073, + 0.012178122997283936, + 0.02982465922832489, + 0.013277853839099407, + 0.030905725434422493, + 0.033186107873916626, + 0.013715865090489388, + 0.02407749742269516, + 0.014734789729118347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.42893511056900024 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048669084906578064 + ], + "opposite_logprob_tokens": [ + -3.0486690998077393 + ], + "logprob_reward": -0.048669084906578064, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048669084906578064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19341641664505005, + 0.01671101525425911, + 0.1334109902381897, + 0.05999235063791275, + 0.06872904300689697, + 0.018325086683034897, + 0.018997708335518837, + 0.03461451455950737, + 0.02019534260034561, + 0.4356074929237366 + ], + "credit_map": [ + 0.018655281513929367, + 0.0668562650680542, + 0.01061907410621643, + 0.004572131205350161, + 0.057192977517843246, + 0.018331026658415794, + 0.008446775376796722, + 0.00874291267246008, + 0.0016833072295412421, + 0.003997573629021645, + 0.0015381273115053773, + 0.0014285603538155556, + 0.003200826933607459, + 0.0016714896773919463, + 0.0013933982700109482, + 0.0017977335955947638, + 0.009929973632097244, + 0.02798219583928585, + 0.003556956071406603, + 0.005307030398398638, + 0.04236099496483803, + 0.034031469374895096, + 0.004608515650033951, + 0.005633857566863298, + 0.0038238076958805323, + 0.010789616964757442, + 0.0029992202762514353, + 0.005801004357635975, + 0.01586761511862278, + 0.012175685726106167, + 0.003972283564507961, + 0.004563122522085905, + 0.005621263291686773, + 0.019245382398366928, + 0.004296435508877039, + 0.011348409578204155, + 0.011850334703922272, + 0.007355864159762859, + 0.00428538816049695, + 0.0047259703278541565, + 0.0013778875581920147, + 0.006063246168196201, + 0.0006137387244962156, + 0.000822153000626713, + 0.005681133363395929, + 0.002190283965319395, + 0.0008866144344210625, + 0.0006900301668792963, + 0.0017539619002491236, + 0.004941017832607031, + 0.0008280624751932919, + 0.002415846101939678, + 0.0038910130970180035, + 0.0026348105166107416, + 0.0014827605336904526, + 0.001050237799063325, + 0.003155746264383197, + 0.010873135179281235, + 0.0014104611473158002, + 0.0018431727075949311, + 0.007915860041975975, + 0.004566603805869818, + 0.0026224071625620127, + 0.002227134071290493, + 0.0020829536952078342, + 0.005632259417325258, + 0.0014917278895154595, + 0.0017872401513159275, + 0.0033958510030061007, + 0.0024330492597073317, + 0.0018488928908482194, + 0.001523369923233986, + 0.03140217438340187, + 0.07270990312099457, + 0.034065328538417816, + 0.0786743313074112, + 0.08135131001472473, + 0.03616005927324295, + 0.062245115637779236, + 0.038999322801828384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.048669084906578064 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607493683695793 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.01607493683695793, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607493683695793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20169338583946228, + 0.01794333942234516, + 0.06014060601592064, + 0.07048983126878738, + 0.06428322196006775, + 0.03601434454321861, + 0.031754739582538605, + 0.028909608721733093, + 0.04680348560214043, + 0.4419674575328827 + ], + "credit_map": [ + 0.018838360905647278, + 0.06931754946708679, + 0.01333888154476881, + 0.005281580612063408, + 0.057496119290590286, + 0.01991770602762699, + 0.007745929062366486, + 0.009757265448570251, + 0.0014440168160945177, + 0.004614709410816431, + 0.0015029951464384794, + 0.0012753376504406333, + 0.00341931008733809, + 0.002328337635844946, + 0.0014424044638872147, + 0.0019162279786542058, + 0.0048741744831204414, + 0.014256772585213184, + 0.002912903670221567, + 0.006728804670274258, + 0.01136418990790844, + 0.011994844302535057, + 0.0039043715223670006, + 0.004104543011635542, + 0.0053928508423268795, + 0.0159451961517334, + 0.0037228313740342855, + 0.005693881772458553, + 0.017490999773144722, + 0.01188535988330841, + 0.00580986961722374, + 0.0045488374307751656, + 0.004851828794926405, + 0.014173272997140884, + 0.005433965940028429, + 0.01376170665025711, + 0.00938479509204626, + 0.006810198072344065, + 0.005104552023112774, + 0.004762900061905384, + 0.0031270822510123253, + 0.010376671329140663, + 0.0018374229548498988, + 0.002210969105362892, + 0.008663365617394447, + 0.0056191254407167435, + 0.002244151197373867, + 0.0019355580443516374, + 0.00287004467099905, + 0.010173949413001537, + 0.0012408130569383502, + 0.003229887457564473, + 0.005837467033416033, + 0.004910272546112537, + 0.0018326584249734879, + 0.0016596454661339521, + 0.0023923818953335285, + 0.00926586240530014, + 0.0012324235867708921, + 0.001709180069155991, + 0.006209831219166517, + 0.003882939461618662, + 0.0024145906791090965, + 0.0018024016171693802, + 0.0033590642269700766, + 0.013391450047492981, + 0.0027084103785455227, + 0.003168199211359024, + 0.013180412352085114, + 0.004732612986117601, + 0.0037414641119539738, + 0.0025218704249709845, + 0.030009351670742035, + 0.08572772890329361, + 0.03229939192533493, + 0.07351512461900711, + 0.09060069918632507, + 0.03402561694383621, + 0.05862918868660927, + 0.037160348147153854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.01607493683695793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a secluded meadow surrounded by the dark silhouette of an ancient forest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the sky deep indigo, moonless, and dotted with sparse stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a ribbon of light in the sky emitting a warm, pulsating yellow-green bioluminescence from fireflies?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ribbon of light slowly undulate and swirl through the sky like a living nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slow, sweeping long shot taken from a slightly upward perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606214977800846 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.01606214977800846, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606214977800846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24126096069812775, + 0.02921798638999462, + 0.09101071208715439, + 0.05792410671710968, + 0.07657177001237869, + 0.05504170432686806, + 0.03909456357359886, + 0.060544028878211975, + 0.11196145415306091, + 0.23737271130084991 + ], + "credit_map": [ + 0.024254055693745613, + 0.08113951236009598, + 0.014360425062477589, + 0.008018615655601025, + 0.0644846111536026, + 0.0238976888358593, + 0.012259576469659805, + 0.012846509926021099, + 0.002317833947017789, + 0.00743747828528285, + 0.0018091810634359717, + 0.0016475693555548787, + 0.007465293165296316, + 0.0038657248951494694, + 0.002270116936415434, + 0.002404787577688694, + 0.006317404564470053, + 0.01848132535815239, + 0.0026961087714880705, + 0.004265638999640942, + 0.02793639525771141, + 0.02393166348338127, + 0.0034771820064634085, + 0.00390500552020967, + 0.004102629609405994, + 0.016326166689395905, + 0.0028043650090694427, + 0.0046550678089261055, + 0.015288043767213821, + 0.008058947511017323, + 0.004123389255255461, + 0.0025655103381723166, + 0.00618156511336565, + 0.027887582778930664, + 0.0033384296111762524, + 0.004544716328382492, + 0.01602642424404621, + 0.010247566737234592, + 0.004466257989406586, + 0.0038792332634329796, + 0.0029460096266120672, + 0.013083305209875107, + 0.0031994099263101816, + 0.004944256041198969, + 0.016719752922654152, + 0.006855596322566271, + 0.0038573304191231728, + 0.0034360543359071016, + 0.0030413882341235876, + 0.014128810726106167, + 0.0014417081838473678, + 0.0037143470253795385, + 0.007612096145749092, + 0.005384564865380526, + 0.0018778684316202998, + 0.0018937843851745129, + 0.0023232889361679554, + 0.021665219217538834, + 0.0011975934030488133, + 0.002121452009305358, + 0.023693710565567017, + 0.00518632959574461, + 0.0019690629560500383, + 0.0023873772006481886, + 0.007109710481017828, + 0.037336625158786774, + 0.002087221248075366, + 0.002436352428048849, + 0.041403621435165405, + 0.016080815345048904, + 0.002602012362331152, + 0.0029051064047962427, + 0.015557903796434402, + 0.052559465169906616, + 0.016251444816589355, + 0.03608613461256027, + 0.051347896456718445, + 0.01746307499706745, + 0.029184700921177864, + 0.018922097980976105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.01606214977800846 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_10/index_tensor([5895, 5895])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_10/index_tensor([5895, 5895])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e8317bd186cb50105df33dbadc90d917291a22f8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_10/index_tensor([5895, 5895])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "raw, close-up, still shot taken from a level angle, the composition is clean and minimalist. On the right side of the frame, a plump, vibrant lemon rests peacefully on a pristine, slightly reflective surface. Soft studio lighting streams down from the upper left, creating bright highlights on its waxy, dimpled skin and casting a soft, blurred shadow to the right. The lemon's vibrant yellow, topped by a short, emerald green stem, adds a touch of freshness. In the blank space to the left of the lemon, the words \"Lemon Days\" are drawn stroke by stroke in an elegant, calligraphic animation. As if an invisible pen, dipped in deep, forest-green ink, writes in a fluid and rhythmic pattern, the strokes sometimes delicate and silky, sometimes fuller and more powerful, each letter imbued with a dynamic, artistic beauty. The entire scene is painted in fresh and bright colors, with a tranquil and elegant atmosphere. It perfectly blends the simplicity of nature with the art of handwritten fonts, creating a leisurely, refreshing and poetic visual experience.", + "seed": 7609, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_10/index_5895_seed_7609_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.026115784188732505, + "vlm_logprob": -0.026115784188732505, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.026115784188732505, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003306400263682008 + ], + "opposite_logprob_tokens": [ + -7.1283063888549805 + ], + "logprob_reward": -0.003306400263682008, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.003306400263682008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3873557448387146, + 0.021372061222791672, + 0.010743308812379837, + 0.5168603658676147, + 0.013417028822004795, + 0.008925634436309338, + 0.02339841052889824, + 0.006638139486312866, + 0.004477893467992544, + 0.006811396684497595 + ], + "credit_map": [ + 0.013703946024179459, + 0.1693602055311203, + 0.009089205414056778, + 0.16543708741664886, + 0.009838707745075226, + 0.004065250046551228, + 0.011126738972961903, + 0.0047346376813948154, + 0.0039050388149917126, + 0.004435614217072725, + 0.0037304458674043417, + 0.005515263881534338, + 0.0009470883524045348, + 0.0007910762797109783, + 0.0012860597344115376, + 0.0007614736096002162, + 0.0013517951592803001, + 0.0019562027882784605, + 0.0015329689485952258, + 0.0017029870068654418, + 0.0012182806385681033, + 0.0012755723437294364, + 0.0009729656158015132, + 0.0007325366605073214, + 0.04715368151664734, + 0.20095598697662354, + 0.1583334058523178, + 0.04075191915035248, + 0.022182945162057877, + 0.02125047706067562, + 0.013712690211832523, + 0.012519211508333683, + 0.002612454816699028, + 0.002166829304769635, + 0.0017864189576357603, + 0.001336613087914884, + 0.0013018702156841755, + 0.0015804055146872997, + 0.0012576032895594835, + 0.0013748342171311378, + 0.0012018034467473626, + 0.0013490956043824553, + 0.0016088852426037192, + 0.0008793965680524707, + 0.0007320819422602654, + 0.0009730064775794744, + 0.0009763457928784192, + 0.0012050197692587972, + 0.002808791119605303, + 0.003509541042149067, + 0.003954414743930101, + 0.0019380021840333939, + 0.0020817003678530455, + 0.002798225497826934, + 0.002368009416386485, + 0.003939726389944553, + 0.0011921883560717106, + 0.0007769811782054603, + 0.0008789079147391021, + 0.0005203254404477775, + 0.0007651555351912975, + 0.0008834080654196441, + 0.0006829044432379305, + 0.000938268203753978, + 0.0008206493221223354, + 0.0006517525762319565, + 0.0005238765734247863, + 0.0004481312644202262, + 0.0005213756812736392, + 0.0007757804705761373, + 0.00030546728521585464, + 0.0004308602074161172, + 0.0012151774717494845, + 0.001212207949720323, + 0.001336261397227645, + 0.0005636371206492186, + 0.0007460357155650854, + 0.0006492848042398691, + 0.0006040697917342186, + 0.0004847222298849374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.003306400263682008 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101199563592672 + ], + "opposite_logprob_tokens": [ + -5.504101276397705 + ], + "logprob_reward": -0.004101199563592672, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101199563592672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5002278089523315, + 0.032875098288059235, + 0.01733935996890068, + 0.35000333189964294, + 0.019003896042704582, + 0.016019217669963837, + 0.012319525703787804, + 0.019018495455384254, + 0.02022031880915165, + 0.012972895056009293 + ], + "credit_map": [ + 0.015874624252319336, + 0.2120732069015503, + 0.011408450081944466, + 0.20689283311367035, + 0.016170470044016838, + 0.01292921882122755, + 0.016100673004984856, + 0.008778316900134087, + 0.003434177953749895, + 0.005300209391862154, + 0.0025418209843337536, + 0.005665554199367762, + 0.0045783985406160355, + 0.004040656611323357, + 0.004035794176161289, + 0.003278488526120782, + 0.002267706673592329, + 0.0021889538038522005, + 0.00199936144053936, + 0.002774937776848674, + 0.0022317515686154366, + 0.002098074881359935, + 0.0019013509154319763, + 0.0018772236071527004, + 0.03205060213804245, + 0.13640494644641876, + 0.10630404204130173, + 0.027314094826579094, + 0.015043838880956173, + 0.0148621816188097, + 0.009610787034034729, + 0.008412853814661503, + 0.0018332405015826225, + 0.0037398103158921003, + 0.004937965422868729, + 0.0022658496163785458, + 0.002165075158700347, + 0.0012098544975742698, + 0.0015503932954743505, + 0.0013017060700803995, + 0.0020020429510623217, + 0.002426327206194401, + 0.0037465784698724747, + 0.0015726352576166391, + 0.001675488892942667, + 0.0015545464120805264, + 0.0019280941924080253, + 0.0011135039385408163, + 0.0015722103416919708, + 0.0024186153896152973, + 0.0031308289617300034, + 0.0011452460894361138, + 0.0009402951109223068, + 0.0014828313142061234, + 0.0007637881208211184, + 0.0008657106081955135, + 0.001670571742579341, + 0.0017975128721445799, + 0.001711406628601253, + 0.0015414662193506956, + 0.004446936771273613, + 0.0025635624770075083, + 0.0029541687108576298, + 0.0023328717797994614, + 0.002733977511525154, + 0.0028122770600020885, + 0.0039041393902152777, + 0.0015918753342702985, + 0.002756203059107065, + 0.0028686909936368465, + 0.001771128037944436, + 0.0017820269567891955, + 0.001857604132965207, + 0.0016585374251008034, + 0.001856257556937635, + 0.0010803801706060767, + 0.001725215115584433, + 0.002089152578264475, + 0.001275131362490356, + 0.001430617063306272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.004101199563592672 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031646490097046 + ], + "opposite_logprob_tokens": [ + -2.3503165245056152 + ], + "logprob_reward": -0.10031646490097046, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031646490097046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1609078198671341, + 0.0734199732542038, + 0.034161198884248734, + 0.3545593023300171, + 0.036939527839422226, + 0.04757722094655037, + 0.06027114391326904, + 0.05488086864352226, + 0.06188846752047539, + 0.1153944879770279 + ], + "credit_map": [ + 0.017916645854711533, + 0.03851749002933502, + 0.015254836529493332, + 0.038789618760347366, + 0.011193779297173023, + 0.013373392634093761, + 0.013917104341089725, + 0.011944951489567757, + 0.009953074157238007, + 0.009638605639338493, + 0.0123416967689991, + 0.009945906698703766, + 0.00829914677888155, + 0.007324527483433485, + 0.007711738348007202, + 0.008205280639231205, + 0.005263010039925575, + 0.004491892177611589, + 0.004145157989114523, + 0.004932993091642857, + 0.003602566197514534, + 0.004139429423958063, + 0.00400411756709218, + 0.00358203100040555, + 0.02891627512872219, + 0.13812518119812012, + 0.11486387252807617, + 0.02892046794295311, + 0.014297543093562126, + 0.01330072432756424, + 0.008778958581387997, + 0.007356295362114906, + 0.005444555077701807, + 0.005963278468698263, + 0.00881136767566204, + 0.003078028094023466, + 0.0028700323309749365, + 0.003660960588604212, + 0.003395956242457032, + 0.0037153507582843304, + 0.006515488028526306, + 0.006859124172478914, + 0.00920942984521389, + 0.003362849820405245, + 0.005828644149005413, + 0.005566593259572983, + 0.004536888096481562, + 0.005698205437511206, + 0.0060763428919017315, + 0.01084343995898962, + 0.01496827881783247, + 0.00513219041749835, + 0.005991659127175808, + 0.007403330411761999, + 0.004767712205648422, + 0.00508818868547678, + 0.004599662497639656, + 0.007795755751430988, + 0.009605096653103828, + 0.004423082340508699, + 0.00880895834416151, + 0.009833579882979393, + 0.0056830584071576595, + 0.004131674766540527, + 0.008032018318772316, + 0.007414513733237982, + 0.009365696460008621, + 0.004407600499689579, + 0.007132102269679308, + 0.01078878529369831, + 0.006202222779393196, + 0.008545522578060627, + 0.015669189393520355, + 0.010459888726472855, + 0.011071368120610714, + 0.006754960399121046, + 0.01622401736676693, + 0.02346879430115223, + 0.014203652739524841, + 0.017542624846100807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.10031646490097046 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014215397648513317 + ], + "opposite_logprob_tokens": [ + -4.264215469360352 + ], + "logprob_reward": -0.014215397648513317, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014215397648513317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12754690647125244, + 0.03004416823387146, + 0.017961855977773666, + 0.7476537227630615, + 0.04821489006280899, + 0.006784873548895121, + 0.005683865863829851, + 0.004830627236515284, + 0.003987697418779135, + 0.007291380316019058 + ], + "credit_map": [ + 0.008998190052807331, + 0.04945715144276619, + 0.007137121632695198, + 0.049202267080545425, + 0.003983340226113796, + 0.002715922659263015, + 0.004116852767765522, + 0.001936050713993609, + 0.0048780012875795364, + 0.008019070141017437, + 0.004425584338605404, + 0.006621452514082193, + 0.0011816980550065637, + 0.0011884102132171392, + 0.001678304746747017, + 0.002051642630249262, + 0.0034477715380489826, + 0.0034901199396699667, + 0.0033703630324453115, + 0.002879430539906025, + 0.001327665988355875, + 0.001057530171237886, + 0.0011743373470380902, + 0.0012146368389949203, + 0.07110986113548279, + 0.2894705832004547, + 0.2229064553976059, + 0.05812736228108406, + 0.03308743238449097, + 0.030850747600197792, + 0.023404089733958244, + 0.01869715191423893, + 0.010456553660333157, + 0.007844683714210987, + 0.008858365938067436, + 0.008668598718941212, + 0.0033924856688827276, + 0.0027594256680458784, + 0.002897190861403942, + 0.0033375779166817665, + 0.0015019497368484735, + 0.0010989448055624962, + 0.0015609307447448373, + 0.0010936639737337828, + 0.0004127161228097975, + 0.00039627341902814806, + 0.00034721274278126657, + 0.00037318136310204864, + 0.0007152794278226793, + 0.0007782994653098285, + 0.0012088348157703876, + 0.0005131923826411366, + 0.0004824984061997384, + 0.0006291423342190683, + 0.0005438838270492852, + 0.0008127344772219658, + 0.0004906714311800897, + 0.0009038131101988256, + 0.001125690178014338, + 0.0004217948007863015, + 0.000600927509367466, + 0.0004881187342107296, + 0.000327817106153816, + 0.0004717933770734817, + 0.000532975944224745, + 0.00046134021249599755, + 0.0006873746169731021, + 0.0003791961062233895, + 0.0004999611992388964, + 0.0005979925626888871, + 0.0003945821081288159, + 0.00043427449418231845, + 0.0006782394484616816, + 0.0014906820142641664, + 0.0020798437763005495, + 0.0006027089548297226, + 0.0007287644548341632, + 0.0008142494480125606, + 0.00038255451363511384, + 0.0005143371527083218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.014215397648513317 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639458566904068 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.008639458566904068, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639458566904068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25403761863708496, + 0.017009636387228966, + 0.007671655621379614, + 0.6710749268531799, + 0.020509392023086548, + 0.008809693157672882, + 0.00519370986148715, + 0.0034315669909119606, + 0.005956037435680628, + 0.0063057541847229 + ], + "credit_map": [ + 0.006732822395861149, + 0.11085694283246994, + 0.0053587728179991245, + 0.1099056676030159, + 0.007044951431453228, + 0.004080958664417267, + 0.007323879282921553, + 0.0027336322236806154, + 0.002093780552968383, + 0.0036961131263524294, + 0.0015137498266994953, + 0.0037862625904381275, + 0.0013099453644827008, + 0.0018190611153841019, + 0.001376065076328814, + 0.0014146602479740977, + 0.001484267762862146, + 0.0011079494142904878, + 0.0011285804212093353, + 0.0010261888382956386, + 0.0010389082599431276, + 0.0006896285922266543, + 0.0006840002606622875, + 0.0005121326539665461, + 0.060206782072782516, + 0.26076555252075195, + 0.21102453768253326, + 0.05322914570569992, + 0.028133584186434746, + 0.024920200929045677, + 0.017649775370955467, + 0.015145369805395603, + 0.004494174849241972, + 0.003331720130518079, + 0.0033505961764603853, + 0.0036664896178990602, + 0.0010689065093174577, + 0.0013908195542171597, + 0.0013934712624177337, + 0.0018132139230147004, + 0.0008029479067772627, + 0.000984799931757152, + 0.0011154352687299252, + 0.0008577968692407012, + 0.0015087630599737167, + 0.001333639258518815, + 0.0009274997864849865, + 0.0012788110179826617, + 0.0007741177105344832, + 0.0007910359418019652, + 0.0008514082874171436, + 0.0005480542895384133, + 0.0005693720886483788, + 0.0007718009874224663, + 0.0004278236301615834, + 0.0004600974207278341, + 0.0005463745910674334, + 0.00040665254346095026, + 0.00036887943861074746, + 0.00031393932295031846, + 0.0004291667719371617, + 0.00041629772749729455, + 0.00035494816256687045, + 0.0005953085492365062, + 0.0008658714941702783, + 0.0006775904912501574, + 0.0006388460169546306, + 0.00041856817551888525, + 0.0007426749216392636, + 0.0012479833094403148, + 0.0005518096149899065, + 0.0008126935572363436, + 0.0009510750533081591, + 0.0006344718276523054, + 0.0008507969905622303, + 0.0004267051408533007, + 0.0006786161102354527, + 0.000886701513081789, + 0.0007783840410411358, + 0.0010990038281306624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.008639458566904068 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "raw, close-up, still shot taken from a level angle, the composition is clean and minimalist. On the right side of the frame, a plump, vibrant lemon rests peacefully on a pristine, slightly reflective surface. Soft studio lighting streams down from the upper left, creating bright highlights on its waxy, dimpled skin and casting a soft, blurred shadow to the right. The lemon's vibrant yellow, topped by a short, emerald green stem, adds a touch of freshness. In the blank space to the left of the lemon, the words \"Lemon Days\" are drawn stroke by stroke in an elegant, calligraphic animation. As if an invisible pen, dipped in deep, forest-green ink, writes in a fluid and rhythmic pattern, the strokes sometimes delicate and silky, sometimes fuller and more powerful, each letter imbued with a dynamic, artistic beauty. The entire scene is painted in fresh and bright colors, with a tranquil and elegant atmosphere. It perfectly blends the simplicity of nature with the art of handwritten fonts, creating a leisurely, refreshing and poetic visual experience.", + "seed": 7610, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_10/index_5895_seed_7610_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.012264612410217524, + "vlm_logprob": -0.012264612410217524, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.012264612410217524, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003382914699614048 + ], + "opposite_logprob_tokens": [ + -6.753382682800293 + ], + "logprob_reward": -0.003382914699614048, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.003382914699614048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15490764379501343, + 0.07874754071235657, + 0.05783071368932724, + 0.11438485980033875, + 0.050527237355709076, + 0.0478268526494503, + 0.16389504075050354, + 0.02491842582821846, + 0.18850909173488617, + 0.1184525191783905 + ], + "credit_map": [ + 0.02400687150657177, + 0.03345336392521858, + 0.022921735420823097, + 0.018134087324142456, + 0.015643583610653877, + 0.018909785896539688, + 0.011712627485394478, + 0.01012561284005642, + 0.01039100717753172, + 0.014217544347047806, + 0.010194921866059303, + 0.00806519016623497, + 0.012674057856202126, + 0.01079403143376112, + 0.0063944365829229355, + 0.0060163564048707485, + 0.0062161716632544994, + 0.00958555843681097, + 0.007308017462491989, + 0.006814375054091215, + 0.008641504682600498, + 0.006818526424467564, + 0.006251658778637648, + 0.006194906309247017, + 0.012704587541520596, + 0.016368523240089417, + 0.013616975396871567, + 0.014861414209008217, + 0.016934210434556007, + 0.016647418960928917, + 0.01225778553634882, + 0.010993956588208675, + 0.005198480095714331, + 0.008615651167929173, + 0.0067764222621917725, + 0.005881268531084061, + 0.007693796418607235, + 0.006274963263422251, + 0.005349187646061182, + 0.004737474024295807, + 0.005822695791721344, + 0.007792312651872635, + 0.006085140630602837, + 0.006399590987712145, + 0.006315142381936312, + 0.006193266715854406, + 0.00499307457357645, + 0.004225629381835461, + 0.016431421041488647, + 0.02204505167901516, + 0.017833437770605087, + 0.023769859224557877, + 0.02440704219043255, + 0.02390245534479618, + 0.018279433250427246, + 0.01722634583711624, + 0.0027579525485634804, + 0.0043973252177238464, + 0.0034389463253319263, + 0.00343147455714643, + 0.002763568889349699, + 0.003849295200780034, + 0.002021100604906678, + 0.002258764812722802, + 0.029076674953103065, + 0.03183908015489578, + 0.0269363634288311, + 0.031149497255682945, + 0.014118175022304058, + 0.018494468182325363, + 0.019773367792367935, + 0.01712147705256939, + 0.012057501822710037, + 0.01018845196813345, + 0.005948647856712341, + 0.008067856542766094, + 0.014660410583019257, + 0.0254791509360075, + 0.021032877266407013, + 0.021017635241150856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.003382914699614048 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257115233689547 + ], + "opposite_logprob_tokens": [ + -5.2552571296691895 + ], + "logprob_reward": -0.005257115233689547, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257115233689547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3141295611858368, + 0.31021440029144287, + 0.0684979259967804, + 0.06108132377266884, + 0.029859190806746483, + 0.023192264139652252, + 0.057143986225128174, + 0.043845679610967636, + 0.06111367791891098, + 0.03092198260128498 + ], + "credit_map": [ + 0.03612785413861275, + 0.06238408386707306, + 0.039666008204221725, + 0.0405765026807785, + 0.048631489276885986, + 0.03145324066281319, + 0.03136857599020004, + 0.023921824991703033, + 0.03664505109190941, + 0.0559597872197628, + 0.03820657357573509, + 0.04538965970277786, + 0.04116053506731987, + 0.0356055423617363, + 0.03136249631643295, + 0.02588477171957493, + 0.009094678796827793, + 0.010613557882606983, + 0.008431416936218739, + 0.009946544654667377, + 0.008563159964978695, + 0.008487402461469173, + 0.007331723812967539, + 0.006029448937624693, + 0.0070070489309728146, + 0.00909496657550335, + 0.008208260871469975, + 0.00855986587703228, + 0.008295314386487007, + 0.009300315752625465, + 0.00602599186822772, + 0.0045895627699792385, + 0.0036998894065618515, + 0.005150170531123877, + 0.004181631840765476, + 0.004835180938243866, + 0.003321879543364048, + 0.003803396597504616, + 0.0025873209815472364, + 0.0022797221317887306, + 0.0024995070416480303, + 0.0029853542800992727, + 0.0029584604781121016, + 0.004413771443068981, + 0.0027889455668628216, + 0.0033081683795899153, + 0.0023100466933101416, + 0.0019280139822512865, + 0.007513510063290596, + 0.00787664856761694, + 0.007048139348626137, + 0.007729343604296446, + 0.007399531081318855, + 0.009182061068713665, + 0.0058102598413825035, + 0.004584494512528181, + 0.005046030972152948, + 0.007158838678151369, + 0.00587579607963562, + 0.005240346770733595, + 0.0054065166041255, + 0.00622656149789691, + 0.005233732983469963, + 0.0036578625440597534, + 0.008461471647024155, + 0.008835269138216972, + 0.01107418816536665, + 0.007311483845114708, + 0.006794800516217947, + 0.007373256608843803, + 0.004811394028365612, + 0.006451816763728857, + 0.003615681082010269, + 0.0030481538269668818, + 0.002718007657676935, + 0.003630304243415594, + 0.0040708570741117, + 0.006313486956059933, + 0.003484805580228567, + 0.004040687810629606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005257115233689547 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809121996164322 + ], + "opposite_logprob_tokens": [ + -3.2880911827087402 + ], + "logprob_reward": -0.03809121996164322, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809121996164322, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11994569003582001, + 0.14603671431541443, + 0.13066734373569489, + 0.08748994022607803, + 0.16114945709705353, + 0.09131689369678497, + 0.05272949859499931, + 0.09836997091770172, + 0.07617917656898499, + 0.03611532226204872 + ], + "credit_map": [ + 0.019907576963305473, + 0.01858379878103733, + 0.013262555934488773, + 0.014670358039438725, + 0.017362719401717186, + 0.017641426995396614, + 0.009321230463683605, + 0.009196031838655472, + 0.018497031182050705, + 0.031118469312787056, + 0.0205689687281847, + 0.02281787246465683, + 0.012676717713475227, + 0.015144064091145992, + 0.012537130154669285, + 0.012676473706960678, + 0.018108531832695007, + 0.02834879234433174, + 0.018804237246513367, + 0.020425911992788315, + 0.010193359106779099, + 0.013727269135415554, + 0.011533986777067184, + 0.009525258094072342, + 0.01042319554835558, + 0.011470806784927845, + 0.010494583286345005, + 0.01389045175164938, + 0.008606932125985622, + 0.011961735785007477, + 0.011733932420611382, + 0.008908308111131191, + 0.02111494354903698, + 0.025332357734441757, + 0.022106917575001717, + 0.02080395445227623, + 0.01560795959085226, + 0.022206077352166176, + 0.019092874601483345, + 0.014884384348988533, + 0.011954511515796185, + 0.015979046002030373, + 0.012516274116933346, + 0.008841539733111858, + 0.01219757366925478, + 0.014744602143764496, + 0.008054600097239017, + 0.007028752472251654, + 0.006338594015687704, + 0.006935317534953356, + 0.005977954715490341, + 0.00812700018286705, + 0.006661874707788229, + 0.008499250747263432, + 0.004835039842873812, + 0.005354471504688263, + 0.011906302534043789, + 0.017893565818667412, + 0.01297694630920887, + 0.010803068988025188, + 0.01222741138190031, + 0.013510233722627163, + 0.01010962389409542, + 0.00894282478839159, + 0.011687884107232094, + 0.008887543343007565, + 0.008940489031374454, + 0.0089136753231287, + 0.009074348025023937, + 0.0063679032027721405, + 0.013703403994441032, + 0.00860393512994051, + 0.003176451660692692, + 0.003631212981417775, + 0.003143352223560214, + 0.003762250067666173, + 0.003707300638779998, + 0.005878482013940811, + 0.006689116358757019, + 0.006127156317234039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.03809121996164322 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864548608660698 + ], + "opposite_logprob_tokens": [ + -4.758645534515381 + ], + "logprob_reward": -0.00864548608660698, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864548608660698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12877434492111206, + 0.04472498968243599, + 0.03409209102392197, + 0.07178094238042831, + 0.07030202448368073, + 0.04824134334921837, + 0.04257386550307274, + 0.06032705679535866, + 0.3415020704269409, + 0.15768125653266907 + ], + "credit_map": [ + 0.017154837027192116, + 0.02362765371799469, + 0.017611972987651825, + 0.013398611918091774, + 0.014874763786792755, + 0.017250677570700645, + 0.015320714563131332, + 0.009535113349556923, + 0.005770602263510227, + 0.0063275680877268314, + 0.005902220495045185, + 0.005926905665546656, + 0.005004271399229765, + 0.00770447077229619, + 0.004943810403347015, + 0.0031451424583792686, + 0.004394738003611565, + 0.0046845185570418835, + 0.004023898392915726, + 0.003432962577790022, + 0.005019463133066893, + 0.005115095060318708, + 0.004078961443156004, + 0.003342453623190522, + 0.005089105572551489, + 0.010509219020605087, + 0.007468034513294697, + 0.010106240399181843, + 0.010793434455990791, + 0.007947615347802639, + 0.01087699644267559, + 0.008990291506052017, + 0.006156469229608774, + 0.00929208006709814, + 0.006531687453389168, + 0.008872661739587784, + 0.009964371100068092, + 0.012368223629891872, + 0.009392413310706615, + 0.0077241212129592896, + 0.004246043041348457, + 0.007213565055280924, + 0.005824284628033638, + 0.0076723904348909855, + 0.005528397858142853, + 0.005503979045897722, + 0.006926110479980707, + 0.005326572339981794, + 0.004749802406877279, + 0.007998579181730747, + 0.005085326731204987, + 0.005304143764078617, + 0.006836749613285065, + 0.004715739283710718, + 0.004301459528505802, + 0.0035820624325424433, + 0.0063492488116025925, + 0.008248320780694485, + 0.006644221022725105, + 0.006670738570392132, + 0.009107877500355244, + 0.011392059735953808, + 0.006415795534849167, + 0.005498797632753849, + 0.049143288284540176, + 0.056336380541324615, + 0.08849688619375229, + 0.06507095694541931, + 0.01951754279434681, + 0.015669729560613632, + 0.02535749040544033, + 0.02190982550382614, + 0.029660984873771667, + 0.02546676993370056, + 0.019613783806562424, + 0.019748514518141747, + 0.016014987602829933, + 0.015140554867684841, + 0.016806941479444504, + 0.01522871945053339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.00864548608660698 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946326069533825 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005946326069533825, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946326069533825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14925158023834229, + 0.05618732050061226, + 0.058597080409526825, + 0.0867990255355835, + 0.10239268094301224, + 0.11924823373556137, + 0.2213081270456314, + 0.04863341152667999, + 0.12565438449382782, + 0.031928207725286484 + ], + "credit_map": [ + 0.019167134538292885, + 0.01955699920654297, + 0.015798334032297134, + 0.024566050618886948, + 0.01543488260358572, + 0.023960888385772705, + 0.015056535601615906, + 0.015710746869444847, + 0.006320711225271225, + 0.00862037856131792, + 0.006313813850283623, + 0.008956864476203918, + 0.007105376571416855, + 0.008830447681248188, + 0.005042520817369223, + 0.004997204057872295, + 0.00688235042616725, + 0.008879275992512703, + 0.008010925725102425, + 0.009143969975411892, + 0.00687410868704319, + 0.008662313222885132, + 0.005520783364772797, + 0.004623352084308863, + 0.007187612354755402, + 0.013205892406404018, + 0.009007951244711876, + 0.011350134387612343, + 0.01298270933330059, + 0.010229000821709633, + 0.013289202004671097, + 0.009546519257128239, + 0.009276898577809334, + 0.010538392700254917, + 0.011662102304399014, + 0.012338532134890556, + 0.012245632708072662, + 0.019626980647444725, + 0.015560716390609741, + 0.01114342175424099, + 0.012469972483813763, + 0.021445978432893753, + 0.014616374857723713, + 0.016230331733822823, + 0.015227433294057846, + 0.014769215136766434, + 0.013625365681946278, + 0.010863552801311016, + 0.025339223444461823, + 0.030287040397524834, + 0.026764091104269028, + 0.028552651405334473, + 0.03280719369649887, + 0.034823328256607056, + 0.02185741998255253, + 0.020877163857221603, + 0.0045145126059651375, + 0.006914233323186636, + 0.006526163779199123, + 0.006183243822306395, + 0.006501844618469477, + 0.007616774644702673, + 0.0061563667841255665, + 0.004220267292112112, + 0.019426271319389343, + 0.01951192505657673, + 0.01921725645661354, + 0.02667132019996643, + 0.008737562224268913, + 0.012614522129297256, + 0.010242807678878307, + 0.00923271756619215, + 0.003921550698578358, + 0.004380727186799049, + 0.00404589157551527, + 0.0040961639024317265, + 0.00415031285956502, + 0.0045557632111012936, + 0.0035929474979639053, + 0.0031848496291786432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.005946326069533825 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "raw, close-up, still shot taken from a level angle, the composition is clean and minimalist. On the right side of the frame, a plump, vibrant lemon rests peacefully on a pristine, slightly reflective surface. Soft studio lighting streams down from the upper left, creating bright highlights on its waxy, dimpled skin and casting a soft, blurred shadow to the right. The lemon's vibrant yellow, topped by a short, emerald green stem, adds a touch of freshness. In the blank space to the left of the lemon, the words \"Lemon Days\" are drawn stroke by stroke in an elegant, calligraphic animation. As if an invisible pen, dipped in deep, forest-green ink, writes in a fluid and rhythmic pattern, the strokes sometimes delicate and silky, sometimes fuller and more powerful, each letter imbued with a dynamic, artistic beauty. The entire scene is painted in fresh and bright colors, with a tranquil and elegant atmosphere. It perfectly blends the simplicity of nature with the art of handwritten fonts, creating a leisurely, refreshing and poetic visual experience.", + "seed": 7611, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_10/index_5895_seed_7611_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0375481927767396, + "vlm_logprob": -0.0375481927767396, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.0375481927767396, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002236961852759123 + ], + "opposite_logprob_tokens": [ + -7.252236843109131 + ], + "logprob_reward": -0.002236961852759123, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.002236961852759123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06058846414089203, + 0.12173806130886078, + 0.40262553095817566, + 0.06998302042484283, + 0.08533310890197754, + 0.060874588787555695, + 0.13342487812042236, + 0.007475817576050758, + 0.024236612021923065, + 0.03371991217136383 + ], + "credit_map": [ + 0.010523131117224693, + 0.016561776399612427, + 0.01089845597743988, + 0.007776820100843906, + 0.0033410510513931513, + 0.0046248482540249825, + 0.0028321605641394854, + 0.004030220676213503, + 0.01834084466099739, + 0.04233430698513985, + 0.021148286759853363, + 0.012227222323417664, + 0.007404236122965813, + 0.009401561692357063, + 0.0049164872616529465, + 0.005965124350041151, + 0.04642437770962715, + 0.15490013360977173, + 0.10830704867839813, + 0.030824080109596252, + 0.02119896188378334, + 0.018568051978945732, + 0.011022631078958511, + 0.011380231939256191, + 0.013125112280249596, + 0.016673646867275238, + 0.008077101781964302, + 0.009657972492277622, + 0.004896135535091162, + 0.00865054689347744, + 0.0034688252490013838, + 0.005433675367385149, + 0.012232769280672073, + 0.02916615828871727, + 0.019515497609972954, + 0.007727598771452904, + 0.004372662398964167, + 0.005435212980955839, + 0.0026413521263748407, + 0.004241857212036848, + 0.008704667910933495, + 0.022855643182992935, + 0.013190164230763912, + 0.005231174640357494, + 0.003522200509905815, + 0.0035463699605315924, + 0.0018364350544288754, + 0.001987936208024621, + 0.019328122958540916, + 0.04832283407449722, + 0.025829851627349854, + 0.01226169802248478, + 0.007882689125835896, + 0.009277412667870522, + 0.005232378840446472, + 0.005289889872074127, + 0.001387374708428979, + 0.00128541502635926, + 0.0010718975681811571, + 0.0011775706661865115, + 0.0004716691328212619, + 0.0007947007543407381, + 0.0006893311510793865, + 0.000597858801484108, + 0.0044825104996562, + 0.004583053290843964, + 0.004866079427301884, + 0.002320911269634962, + 0.0017860329244285822, + 0.002665245207026601, + 0.0016401377506554127, + 0.001892642118036747, + 0.005040424410253763, + 0.007722220383584499, + 0.006040211301296949, + 0.0028401101008057594, + 0.0024131701793521643, + 0.003916242625564337, + 0.0022461998742073774, + 0.003501334460452199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.002236961852759123 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979179609566927 + ], + "opposite_logprob_tokens": [ + -4.634791851043701 + ], + "logprob_reward": -0.00979179609566927, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979179609566927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05885174870491028, + 0.03629928082227707, + 0.3330780565738678, + 0.1433115452528, + 0.0380709171295166, + 0.2905983626842499, + 0.026631904765963554, + 0.04810330644249916, + 0.008634713478386402, + 0.016420183703303337 + ], + "credit_map": [ + 0.0068609402514994144, + 0.014661955647170544, + 0.010263335891067982, + 0.009049219079315662, + 0.003879653289914131, + 0.005833367817103863, + 0.00404098117724061, + 0.004262298345565796, + 0.007064588833600283, + 0.009332835674285889, + 0.005349259357899427, + 0.004327720496803522, + 0.002477728296071291, + 0.0036107986234128475, + 0.0017870954470708966, + 0.002349248854443431, + 0.04109945893287659, + 0.12720084190368652, + 0.08162533491849899, + 0.02634345181286335, + 0.018008822575211525, + 0.018645010888576508, + 0.00882780086249113, + 0.011327339336276054, + 0.026089726015925407, + 0.03958730772137642, + 0.016827916726469994, + 0.019559795036911964, + 0.010348128154873848, + 0.015140584670007229, + 0.0061263833194971085, + 0.009631709195673466, + 0.004894276615232229, + 0.013982443138957024, + 0.006731318309903145, + 0.00433719577267766, + 0.002397648524492979, + 0.0024561467580497265, + 0.0017207660712301731, + 0.001551118097268045, + 0.05403181537985802, + 0.07968097925186157, + 0.033256083726882935, + 0.03776523843407631, + 0.020077204331755638, + 0.033973436802625656, + 0.011954248882830143, + 0.019859345629811287, + 0.0034229126758873463, + 0.008873965591192245, + 0.0033523973543196917, + 0.0028559695929288864, + 0.0019924931693822145, + 0.002852736273780465, + 0.0013297944096848369, + 0.0019516360480338335, + 0.007261660415679216, + 0.017988184466958046, + 0.00747414818033576, + 0.004453922621905804, + 0.0032844117376953363, + 0.003748650196939707, + 0.0018557060975581408, + 0.002036623191088438, + 0.0009692959720268846, + 0.0020759927574545145, + 0.001220390317030251, + 0.0010828988160938025, + 0.0006706552812829614, + 0.0009590352419763803, + 0.000877564656548202, + 0.0007788808434270322, + 0.0024085917975753546, + 0.005441511049866676, + 0.003009684383869171, + 0.001578641589730978, + 0.0009284468833357096, + 0.0011897023068740964, + 0.0010247095488011837, + 0.0008388952119275928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.00979179609566927 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906073123216629 + ], + "opposite_logprob_tokens": [ + -2.4640607833862305 + ], + "logprob_reward": -0.08906073123216629, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906073123216629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062492676079273224, + 0.07423963397741318, + 0.14399011433124542, + 0.21418868005275726, + 0.049423698335886, + 0.2687981426715851, + 0.053084976971149445, + 0.07282280176877975, + 0.013886946253478527, + 0.04707237705588341 + ], + "credit_map": [ + 0.009140433743596077, + 0.01844140700995922, + 0.009107576683163643, + 0.0077760606072843075, + 0.004071654751896858, + 0.005147499963641167, + 0.004624692723155022, + 0.004183345939964056, + 0.00922140758484602, + 0.023617468774318695, + 0.015144400298595428, + 0.008692525327205658, + 0.004604377318173647, + 0.005949086043983698, + 0.0038051658775657415, + 0.0032051929738372564, + 0.02076597698032856, + 0.05169258266687393, + 0.02782888151705265, + 0.013276604004204273, + 0.008501528762280941, + 0.010554298758506775, + 0.0043524703942239285, + 0.007017763797193766, + 0.03653951734304428, + 0.06847674399614334, + 0.026306841522455215, + 0.02615066058933735, + 0.014376772567629814, + 0.021012987941503525, + 0.007944118231534958, + 0.013381022959947586, + 0.007350188214331865, + 0.014280201867222786, + 0.007557398173958063, + 0.0077370535582304, + 0.003265491919592023, + 0.0028728849720209837, + 0.0034388715866953135, + 0.002921601990237832, + 0.04570886865258217, + 0.08598630130290985, + 0.036584220826625824, + 0.029114242643117905, + 0.016813548281788826, + 0.028653282672166824, + 0.009237117134034634, + 0.01670054718852043, + 0.0074022505432367325, + 0.019483717158436775, + 0.00992211140692234, + 0.004885044414550066, + 0.0030430294573307037, + 0.00400505680590868, + 0.001836836221627891, + 0.0025069306138902903, + 0.011315455660223961, + 0.024989932775497437, + 0.012799311429262161, + 0.00679190456867218, + 0.004322493448853493, + 0.005589582491666079, + 0.0033885196316987276, + 0.003625597106292844, + 0.0016911494312807918, + 0.002595658181235194, + 0.0012761843390762806, + 0.0018869767664000392, + 0.0011784924427047372, + 0.0022830022498965263, + 0.0011663567274808884, + 0.0018091253004968166, + 0.006698595359921455, + 0.011983484961092472, + 0.008693750016391277, + 0.004095676355063915, + 0.003126106457784772, + 0.005114395637065172, + 0.0033021066337823868, + 0.004058258607983589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.08906073123216629 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902039587497711 + ], + "opposite_logprob_tokens": [ + -2.5790205001831055 + ], + "logprob_reward": -0.07902039587497711, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902039587497711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1061682403087616, + 0.08383823931217194, + 0.0822066217660904, + 0.30339595675468445, + 0.07109338790178299, + 0.17293459177017212, + 0.02552671916782856, + 0.1002122089266777, + 0.01923990622162819, + 0.03538410738110542 + ], + "credit_map": [ + 0.0137977609410882, + 0.02061685360968113, + 0.014711848460137844, + 0.024126308038830757, + 0.007068249862641096, + 0.010347453877329826, + 0.008198844268918037, + 0.007300923578441143, + 0.016385234892368317, + 0.012228642590343952, + 0.014718051999807358, + 0.01565917767584324, + 0.005784688051789999, + 0.006959698162972927, + 0.006398861762136221, + 0.005703884642571211, + 0.01264856569468975, + 0.027355222031474113, + 0.012150936760008335, + 0.009374693967401981, + 0.006570054683834314, + 0.0066415914334356785, + 0.0036445155274122953, + 0.003821050049737096, + 0.04312876984477043, + 0.1099104955792427, + 0.060773130506277084, + 0.029149126261472702, + 0.017801303416490555, + 0.021355075761675835, + 0.008588280528783798, + 0.01268977764993906, + 0.00999009981751442, + 0.024486491456627846, + 0.013972603715956211, + 0.007192080374807119, + 0.004180504009127617, + 0.0051641399040818214, + 0.0026497195940464735, + 0.0034577525220811367, + 0.03081304393708706, + 0.050945233553647995, + 0.01930498145520687, + 0.02185934968292713, + 0.011658333241939545, + 0.019397972151637077, + 0.00690926518291235, + 0.012046417221426964, + 0.0032652767840772867, + 0.00914737582206726, + 0.003727044677361846, + 0.0023332671262323856, + 0.0017256647115573287, + 0.0023713696282356977, + 0.0011703716591000557, + 0.001786348526366055, + 0.014352995902299881, + 0.03752297908067703, + 0.018081434071063995, + 0.009112617932260036, + 0.005862310528755188, + 0.007402030285447836, + 0.00335597968660295, + 0.004521864466369152, + 0.0029547205194830894, + 0.004345344379544258, + 0.003531691152602434, + 0.0022014970891177654, + 0.0014420608058571815, + 0.0020812801085412502, + 0.001423549372702837, + 0.0012597639579325914, + 0.006445123348385096, + 0.010427073575556278, + 0.005645865108817816, + 0.0036743299569934607, + 0.0020043577533215284, + 0.0031418565195053816, + 0.0018043987220153213, + 0.0022411050740629435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.07902039587497711 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631078828126192 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631078828126192, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631078828126192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38792523741722107, + 0.058342136442661285, + 0.13931593298912048, + 0.08302389830350876, + 0.05972680076956749, + 0.03449498862028122, + 0.08305226266384125, + 0.02542196400463581, + 0.06371023505926132, + 0.06498660147190094 + ], + "credit_map": [ + 0.054805319756269455, + 0.14390335977077484, + 0.06042679771780968, + 0.03798399865627289, + 0.027112623676657677, + 0.028217192739248276, + 0.01848124898970127, + 0.016994668170809746, + 0.005963158793747425, + 0.011866753920912743, + 0.009194822981953621, + 0.007036994677037001, + 0.00582745810970664, + 0.008718092925846577, + 0.004967990797013044, + 0.004766856785863638, + 0.02114187180995941, + 0.04788929969072342, + 0.023714862763881683, + 0.013418949209153652, + 0.008525851182639599, + 0.009756957180798054, + 0.005180294159799814, + 0.009687840938568115, + 0.014136084355413914, + 0.010506300255656242, + 0.02427714318037033, + 0.01059171836823225, + 0.004754945635795593, + 0.00909833237528801, + 0.0037517675664275885, + 0.00590759702026844, + 0.007784038782119751, + 0.014660023152828217, + 0.009357539936900139, + 0.011936851777136326, + 0.004023512825369835, + 0.0031526461243629456, + 0.005023132544010878, + 0.0037890500389039516, + 0.0033900293055921793, + 0.007680179551243782, + 0.011009441688656807, + 0.003488411195576191, + 0.0018562697805464268, + 0.0035380315966904163, + 0.0017834346508607268, + 0.0017491887556388974, + 0.012229901738464832, + 0.02989276312291622, + 0.015469393692910671, + 0.007762376684695482, + 0.005188727751374245, + 0.00563019746914506, + 0.0034183620009571314, + 0.003460526932030916, + 0.0029043538961559534, + 0.006712145637720823, + 0.004462947137653828, + 0.002382331294938922, + 0.0016433020355179906, + 0.0029979017563164234, + 0.0017647059867158532, + 0.0025542753282934427, + 0.009268960915505886, + 0.022334305569529533, + 0.010973770171403885, + 0.00624976446852088, + 0.00403613643720746, + 0.004586405586451292, + 0.002997039584442973, + 0.0032638420816510916, + 0.009633803740143776, + 0.015554286539554596, + 0.012902988120913506, + 0.00547646451741457, + 0.003970769699662924, + 0.007066546939313412, + 0.004534473642706871, + 0.0058472598902881145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.007631078828126192 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "raw, close-up, still shot taken from a level angle, the composition is clean and minimalist. On the right side of the frame, a plump, vibrant lemon rests peacefully on a pristine, slightly reflective surface. Soft studio lighting streams down from the upper left, creating bright highlights on its waxy, dimpled skin and casting a soft, blurred shadow to the right. The lemon's vibrant yellow, topped by a short, emerald green stem, adds a touch of freshness. In the blank space to the left of the lemon, the words \"Lemon Days\" are drawn stroke by stroke in an elegant, calligraphic animation. As if an invisible pen, dipped in deep, forest-green ink, writes in a fluid and rhythmic pattern, the strokes sometimes delicate and silky, sometimes fuller and more powerful, each letter imbued with a dynamic, artistic beauty. The entire scene is painted in fresh and bright colors, with a tranquil and elegant atmosphere. It perfectly blends the simplicity of nature with the art of handwritten fonts, creating a leisurely, refreshing and poetic visual experience.", + "seed": 7612, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_10/index_5895_seed_7612_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03902703286148608, + "vlm_logprob": -0.03902703286148608, + "vlm_yes_no_margin": 4.300000023841858, + "vlm_token_credit": -0.03902703286148608, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0023311602417379618 + ], + "opposite_logprob_tokens": [ + -7.127331256866455 + ], + "logprob_reward": -0.0023311602417379618, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0023311602417379618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.165481299161911, + 0.07201908528804779, + 0.14643724262714386, + 0.08086567372083664, + 0.13376428186893463, + 0.05972524359822273, + 0.08387519419193268, + 0.07582635432481766, + 0.11340401321649551, + 0.06860155612230301 + ], + "credit_map": [ + 0.01699339598417282, + 0.018415071070194244, + 0.02712523378431797, + 0.0245247520506382, + 0.020541634410619736, + 0.020092269405722618, + 0.019776461645960808, + 0.01801251992583275, + 0.006330262869596481, + 0.008670561946928501, + 0.01152935903519392, + 0.00714499969035387, + 0.011521857231855392, + 0.00646654237061739, + 0.010791689157485962, + 0.009563818573951721, + 0.01746547408401966, + 0.02065906673669815, + 0.02069399505853653, + 0.01603597216308117, + 0.01951078325510025, + 0.020636530593037605, + 0.014516675844788551, + 0.016918767243623734, + 0.007220377679914236, + 0.009839959442615509, + 0.010987491346895695, + 0.007214083336293697, + 0.011600825935602188, + 0.013990393839776516, + 0.010470773093402386, + 0.009541776962578297, + 0.01591687835752964, + 0.020959220826625824, + 0.02088639698922634, + 0.014209995046257973, + 0.02270766906440258, + 0.0114460913464427, + 0.017452534288167953, + 0.01018551830202341, + 0.006054062861949205, + 0.007112522609531879, + 0.008438519202172756, + 0.0066278427839279175, + 0.010378162376582623, + 0.006327241193503141, + 0.009160705842077732, + 0.005626197438687086, + 0.0072065615095198154, + 0.007658127695322037, + 0.009587974287569523, + 0.008235173299908638, + 0.0102468840777874, + 0.015050484798848629, + 0.014128408394753933, + 0.011761587113142014, + 0.008453737013041973, + 0.011769109405577183, + 0.011930258944630623, + 0.012747506611049175, + 0.0070436312817037106, + 0.008248982019722462, + 0.008558114990592003, + 0.007075024303048849, + 0.014652839861810207, + 0.022690581157803535, + 0.019764769822359085, + 0.020972739905118942, + 0.010395982302725315, + 0.009381260722875595, + 0.007863461039960384, + 0.007682389114052057, + 0.010549977421760559, + 0.009964444674551487, + 0.01121668890118599, + 0.01136492844671011, + 0.005881688091903925, + 0.007330357562750578, + 0.005775120109319687, + 0.0065183560363948345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0023311602417379618 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547949329018593 + ], + "opposite_logprob_tokens": [ + -4.387547969818115 + ], + "logprob_reward": -0.012547949329018593, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547949329018593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15975604951381683, + 0.08409878611564636, + 0.08799280226230621, + 0.1493089497089386, + 0.146889328956604, + 0.12313506752252579, + 0.09854882955551147, + 0.04273758828639984, + 0.04719128832221031, + 0.060341350734233856 + ], + "credit_map": [ + 0.02367694117128849, + 0.01580708473920822, + 0.02071061171591282, + 0.020576752722263336, + 0.022053930908441544, + 0.018534822389483452, + 0.02200876548886299, + 0.016387145966291428, + 0.005935233552008867, + 0.006820065900683403, + 0.00944159273058176, + 0.008361879736185074, + 0.011642447672784328, + 0.013842417858541012, + 0.014053679071366787, + 0.014001472853124142, + 0.00871309731155634, + 0.009135955944657326, + 0.012289049103856087, + 0.009715656749904156, + 0.011614479124546051, + 0.01126840803772211, + 0.012663809582591057, + 0.012592351995408535, + 0.011765691451728344, + 0.01675228215754032, + 0.022102512419223785, + 0.018377969041466713, + 0.015242034569382668, + 0.020364584401249886, + 0.023132186383008957, + 0.021571695804595947, + 0.017674412578344345, + 0.022582557052373886, + 0.023221353068947792, + 0.015476740896701813, + 0.025269201025366783, + 0.011654401198029518, + 0.01969277672469616, + 0.011317884549498558, + 0.00905394647270441, + 0.016444941982626915, + 0.0170707069337368, + 0.01414832379668951, + 0.013181977905333042, + 0.016955194994807243, + 0.01866987906396389, + 0.0176100954413414, + 0.008030595257878304, + 0.008209524676203728, + 0.010887378826737404, + 0.00992195587605238, + 0.013458866626024246, + 0.019268246367573738, + 0.015188266523182392, + 0.013583996333181858, + 0.005336647853255272, + 0.003985155373811722, + 0.005821573082357645, + 0.004817183595150709, + 0.004924939014017582, + 0.005649917759001255, + 0.006858624517917633, + 0.005343547556549311, + 0.006131338886916637, + 0.0094927828758955, + 0.007633735425770283, + 0.008792011067271233, + 0.0034489755053073168, + 0.004767354112118483, + 0.00384191214106977, + 0.0030831790063530207, + 0.005524465814232826, + 0.008274344727396965, + 0.008978363126516342, + 0.008747712709009647, + 0.006143864709883928, + 0.010621647350490093, + 0.0058473460376262665, + 0.006203611381351948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.012547949329018593 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281047880649567 + ], + "opposite_logprob_tokens": [ + -2.017810583114624 + ], + "logprob_reward": -0.14281047880649567, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14281047880649567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11748804152011871, + 0.05921414867043495, + 0.2043420821428299, + 0.09832249581813812, + 0.04909588024020195, + 0.061894796788692474, + 0.07010689377784729, + 0.21284933388233185, + 0.053952500224113464, + 0.0727337971329689 + ], + "credit_map": [ + 0.014780119992792606, + 0.014709316194057465, + 0.019760962575674057, + 0.015297728590667248, + 0.021062126383185387, + 0.00893086101859808, + 0.012637241743505001, + 0.01030968688428402, + 0.007282137405127287, + 0.006577551830559969, + 0.008711844682693481, + 0.006027857773005962, + 0.007800127845257521, + 0.005416867323219776, + 0.008564862422645092, + 0.008832903578877449, + 0.020550932735204697, + 0.023844685405492783, + 0.026423387229442596, + 0.01946154423058033, + 0.024584701284766197, + 0.032332491129636765, + 0.028385410085320473, + 0.02875892072916031, + 0.009217721410095692, + 0.012911125086247921, + 0.013668032363057137, + 0.010514375753700733, + 0.012278660200536251, + 0.014363834634423256, + 0.012814924120903015, + 0.01255381852388382, + 0.004076474811881781, + 0.006187536753714085, + 0.00691612483933568, + 0.005432800855487585, + 0.005388947669416666, + 0.005341527983546257, + 0.007390020415186882, + 0.00836244784295559, + 0.005526330787688494, + 0.007000830955803394, + 0.007946607656776905, + 0.007011645473539829, + 0.00864963885396719, + 0.009219012223184109, + 0.008993885479867458, + 0.007546848151832819, + 0.0056257485412061214, + 0.007446685805916786, + 0.00825325958430767, + 0.0074644871056079865, + 0.00840569008141756, + 0.012880287133157253, + 0.010457240976393223, + 0.00957349594682455, + 0.015966670587658882, + 0.030498940497636795, + 0.02792317233979702, + 0.02510143257677555, + 0.021333839744329453, + 0.03565007448196411, + 0.03139536455273628, + 0.024979839101433754, + 0.00794536154717207, + 0.012033148668706417, + 0.00968040432780981, + 0.009900865145027637, + 0.004711213055998087, + 0.003351653227582574, + 0.0036149166990071535, + 0.002714939182624221, + 0.006001712288707495, + 0.008317413739860058, + 0.009692538529634476, + 0.009466432966291904, + 0.008008833043277264, + 0.011843539774417877, + 0.009736067615449429, + 0.00966726429760456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.14281047880649567 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026369184255599976 + ], + "opposite_logprob_tokens": [ + -3.651369094848633 + ], + "logprob_reward": -0.026369184255599976, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026369184255599976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12926842272281647, + 0.054561734199523926, + 0.06741879135370255, + 0.12643347680568695, + 0.06645910441875458, + 0.076279416680336, + 0.06769618391990662, + 0.16677971184253693, + 0.21524351835250854, + 0.0298596303910017 + ], + "credit_map": [ + 0.024881109595298767, + 0.018227919936180115, + 0.022853991016745567, + 0.026150647550821304, + 0.011783735826611519, + 0.007672418374568224, + 0.008502389304339886, + 0.00919621717184782, + 0.007847495377063751, + 0.011386647820472717, + 0.011075622402131557, + 0.008353357203304768, + 0.00472286157310009, + 0.0038363791536539793, + 0.0037389143835753202, + 0.003600458148866892, + 0.008204596117138863, + 0.007994795218110085, + 0.009826070629060268, + 0.008943499997258186, + 0.008205173537135124, + 0.008812855929136276, + 0.007341328077018261, + 0.008090474642813206, + 0.012993772514164448, + 0.017039066180586815, + 0.01903650537133217, + 0.016011003404855728, + 0.013792726211249828, + 0.015708845108747482, + 0.0156386885792017, + 0.016212867572903633, + 0.009408541023731232, + 0.011392335407435894, + 0.012191949412226677, + 0.009789306670427322, + 0.0068379114381968975, + 0.004262024071067572, + 0.007413228042423725, + 0.005163806490600109, + 0.008328799158334732, + 0.014601324684917927, + 0.015578608959913254, + 0.01949187181890011, + 0.005793250631541014, + 0.0038497145287692547, + 0.004577458370476961, + 0.00405838992446661, + 0.007132500875741243, + 0.011113153770565987, + 0.011123993434011936, + 0.010146190412342548, + 0.00918011274188757, + 0.00577610032632947, + 0.007425815332680941, + 0.0057983179576694965, + 0.01889690011739731, + 0.036164212971925735, + 0.03386833518743515, + 0.033380720764398575, + 0.011677161790430546, + 0.008312036283314228, + 0.01343994215130806, + 0.011040400713682175, + 0.031639911234378815, + 0.04707060009241104, + 0.038146454840898514, + 0.038142651319503784, + 0.017501886934041977, + 0.014778046868741512, + 0.015183898620307446, + 0.012780078686773777, + 0.004382795188575983, + 0.006250464823096991, + 0.005941344890743494, + 0.005626182071864605, + 0.0024913111701607704, + 0.0013496787287294865, + 0.0020171189680695534, + 0.0018007357139140368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.026369184255599976 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a plump, vibrant lemon visible on the right side of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lemon resting on a pristine, slightly reflective surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature soft studio lighting streaming from the upper left?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the words 'Lemon Days' animated in a calligraphic style using deep forest-green ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a still, close-up view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107639167457819 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.01107639167457819, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107639167457819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09940292686223984, + 0.04651743173599243, + 0.031073587015271187, + 0.03994334489107132, + 0.10311736911535263, + 0.12055495381355286, + 0.10140302032232285, + 0.29829728603363037, + 0.11676079034805298, + 0.04292929172515869 + ], + "credit_map": [ + 0.015285325236618519, + 0.012663546949625015, + 0.0170791856944561, + 0.01137822400778532, + 0.014676741324365139, + 0.010212261229753494, + 0.010368002578616142, + 0.007739642169326544, + 0.007079093251377344, + 0.007216555066406727, + 0.010560779832303524, + 0.007585910614579916, + 0.004350806120783091, + 0.0034879373852163553, + 0.0034529136028140783, + 0.002783435396850109, + 0.005040538497269154, + 0.006982820574194193, + 0.0057588559575378895, + 0.003972635604441166, + 0.002785386284813285, + 0.0020348182879388332, + 0.00250824773684144, + 0.001990283839404583, + 0.005427757743746042, + 0.008926537819206715, + 0.007391001563519239, + 0.005343831609934568, + 0.005140685010701418, + 0.00214695418253541, + 0.0031021020840853453, + 0.002464473946020007, + 0.013467107899487019, + 0.021213585510849953, + 0.01972830668091774, + 0.01518669817596674, + 0.011390985921025276, + 0.006127649452537298, + 0.009149502962827682, + 0.006853536237031221, + 0.015296273864805698, + 0.020523864775896072, + 0.020791703835129738, + 0.021134153008461, + 0.016103532165288925, + 0.00812319666147232, + 0.01148673240095377, + 0.00709549942985177, + 0.009992425329983234, + 0.01192439254373312, + 0.013019557110965252, + 0.01137085072696209, + 0.01105548720806837, + 0.017946239560842514, + 0.01362486183643341, + 0.012469204142689705, + 0.029579464346170425, + 0.05874905362725258, + 0.05314412713050842, + 0.053114525973796844, + 0.021434178575873375, + 0.030615994706749916, + 0.027082957327365875, + 0.024576986208558083, + 0.016582276672124863, + 0.02585897222161293, + 0.021335095167160034, + 0.021167704835534096, + 0.009089737199246883, + 0.008316058665513992, + 0.007726641371846199, + 0.006684301421046257, + 0.006594514474272728, + 0.007469279691576958, + 0.00756751699373126, + 0.009741974994540215, + 0.002875419333577156, + 0.0025913123972713947, + 0.002860981971025467, + 0.003228290006518364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.01107639167457819 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_11/index_tensor([7092, 7092])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_11/index_tensor([7092, 7092])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..44eb0254eeb8c1764b2e08bf595e9e62429e0ec8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_11/index_tensor([7092, 7092])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up static shot captures the decadent beauty of a perfectly assembled burger, creating a scene of pure culinary desire. Warm, directional light from the side sculpts the burger, highlighting every texture. A thick, vibrant orange river of melted cheddar cheese slowly and deliberately oozes down the side of the rough, char-grilled beef patty, its smooth surface catching the light as it navigates the rugged terrain of meat. The burger is topped with a toasted, golden-brown sesame seed bun, glistening with color. From beneath the patty, the crisp, bright green edge of the lettuce leaf and a slice of juicy, deep-red tomato can be seen. The entire subject is framed by a softly blurred, dark background, ensuring that the focus is entirely on the appetizing spectacle. As the cheese continues its languid, gooey descent, the entire mouth-watering scene, imbued with an air of indulgent satisfaction, gently and smoothly fades to black.", + "seed": 30146, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_11/index_7092_seed_30146_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.16090529272332788, + "vlm_logprob": -0.16090529272332788, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.16090529272332788, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006441425997763872 + ], + "opposite_logprob_tokens": [ + -6.756441593170166 + ], + "logprob_reward": -0.006441425997763872, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.006441425997763872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2013452649116516, + 0.04460011050105095, + 0.03188332915306091, + 0.04105772450566292, + 0.028368910774588585, + 0.08981751650571823, + 0.05901728942990303, + 0.05528485029935837, + 0.41608738899230957, + 0.032537639141082764 + ], + "credit_map": [ + 0.01751980558037758, + 0.025577044114470482, + 0.029357444494962692, + 0.023537181317806244, + 0.02633672207593918, + 0.028986500576138496, + 0.02847076579928398, + 0.02155979536473751, + 0.003231748705729842, + 0.004125775769352913, + 0.006555899977684021, + 0.005566283594816923, + 0.005350124090909958, + 0.004481697920709848, + 0.010619761422276497, + 0.0046688164584338665, + 0.002608670387417078, + 0.004522942937910557, + 0.004326526541262865, + 0.004511507228016853, + 0.004111181013286114, + 0.004859540145844221, + 0.0038480740040540695, + 0.0030948894564062357, + 0.003478614380583167, + 0.004781650844961405, + 0.01024661771953106, + 0.004302361514419317, + 0.004793995525687933, + 0.004569042474031448, + 0.003951786085963249, + 0.004933656193315983, + 0.00246481504291296, + 0.003718582447618246, + 0.004352159798145294, + 0.0033743574749678373, + 0.0038144129794090986, + 0.004562451038509607, + 0.002923461375758052, + 0.003158669685944915, + 0.007777002640068531, + 0.014904149807989597, + 0.010548854246735573, + 0.01154882088303566, + 0.012490201741456985, + 0.012953788042068481, + 0.009520284831523895, + 0.010074418038129807, + 0.004921159241348505, + 0.007236921228468418, + 0.01140069030225277, + 0.007742933928966522, + 0.008781380951404572, + 0.0073708598501980305, + 0.005554914008826017, + 0.006008430849760771, + 0.0047169821336865425, + 0.008910407312214375, + 0.006352211348712444, + 0.006614786572754383, + 0.007179678417742252, + 0.005743642337620258, + 0.008914660662412643, + 0.006852475926280022, + 0.02688634768128395, + 0.04299169406294823, + 0.05609501153230667, + 0.05464041978120804, + 0.07263655215501785, + 0.05314546078443527, + 0.0687895268201828, + 0.040902383625507355, + 0.002779660513624549, + 0.003553295275196433, + 0.005734536796808243, + 0.003825679188594222, + 0.004374474752694368, + 0.0038490556180477142, + 0.005098668392747641, + 0.0033222700003534555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.006441425997763872 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28135815262794495 + ], + "opposite_logprob_tokens": [ + -1.4063581228256226 + ], + "logprob_reward": -0.28135815262794495, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28135815262794495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16890215873718262, + 0.03960736095905304, + 0.03066358156502247, + 0.03316371515393257, + 0.06037447229027748, + 0.12229327857494354, + 0.06690890341997147, + 0.15569259226322174, + 0.28151512145996094, + 0.0408787876367569 + ], + "credit_map": [ + 0.011317537166178226, + 0.010651549324393272, + 0.01331085804849863, + 0.023558534681797028, + 0.01769418641924858, + 0.011637004092335701, + 0.05827302858233452, + 0.02245945669710636, + 0.0030990554951131344, + 0.003578855888918042, + 0.011606910265982151, + 0.005129151977598667, + 0.004398481920361519, + 0.0038002419751137495, + 0.0031042497139424086, + 0.004890411626547575, + 0.0025608206633478403, + 0.003926998004317284, + 0.005708740092813969, + 0.004034784622490406, + 0.003540706355124712, + 0.0035200761631131172, + 0.003630922408774495, + 0.00374053418636322, + 0.0023774970322847366, + 0.004113130271434784, + 0.004278809297829866, + 0.003810056485235691, + 0.004230780061334372, + 0.0045297835022211075, + 0.006081471685320139, + 0.003742187051102519, + 0.004651253577321768, + 0.009815174154937267, + 0.0060991509817540646, + 0.0066167921759188175, + 0.007361395750194788, + 0.010449932888150215, + 0.007489735726267099, + 0.007891038432717323, + 0.008666092529892921, + 0.020019132643938065, + 0.015527423471212387, + 0.01465828251093626, + 0.014766326174139977, + 0.017033159732818604, + 0.016712738201022148, + 0.014910129830241203, + 0.004584432579576969, + 0.008042455650866032, + 0.010992111638188362, + 0.00824038777500391, + 0.00792205985635519, + 0.00842682458460331, + 0.0093565434217453, + 0.009344086050987244, + 0.010249622166156769, + 0.020505787804722786, + 0.015389105305075645, + 0.026799554005265236, + 0.024552172049880028, + 0.018794402480125427, + 0.02249140292406082, + 0.016910545527935028, + 0.01748925819993019, + 0.02974497526884079, + 0.03468110412359238, + 0.03928471356630325, + 0.04754381626844406, + 0.036305561661720276, + 0.04778791591525078, + 0.028677769005298615, + 0.003469395451247692, + 0.004440721590071917, + 0.007566783577203751, + 0.004864163231104612, + 0.004786423873156309, + 0.004860103130340576, + 0.006445840932428837, + 0.004445354919880629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.28135815262794495 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026364192366600037 + ], + "opposite_logprob_tokens": [ + -3.6513640880584717 + ], + "logprob_reward": -0.026364192366600037, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026364192366600037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31808656454086304, + 0.053898341953754425, + 0.022574903443455696, + 0.021996691823005676, + 0.07393071055412292, + 0.15620888769626617, + 0.04090757295489311, + 0.2091374695301056, + 0.04983408376574516, + 0.05342485010623932 + ], + "credit_map": [ + 0.021894128993153572, + 0.03635169938206673, + 0.03131963685154915, + 0.039557792246341705, + 0.0417013019323349, + 0.03434646874666214, + 0.07070668786764145, + 0.04220883175730705, + 0.0037274251226335764, + 0.007217023987323046, + 0.0073270611464977264, + 0.009018740616738796, + 0.005865634884685278, + 0.00648708688095212, + 0.008760095573961735, + 0.005495268385857344, + 0.001669572782702744, + 0.0031262370757758617, + 0.00299168867059052, + 0.003202886087819934, + 0.0025963862426579, + 0.002129546832293272, + 0.00426571536809206, + 0.00259286817163229, + 0.0016622457187622786, + 0.0026891801971942186, + 0.0022798143327236176, + 0.002840995090082288, + 0.00269240140914917, + 0.0028548366390168667, + 0.004091992508620024, + 0.002885224297642708, + 0.006050983909517527, + 0.008616594597697258, + 0.011183274909853935, + 0.009042038582265377, + 0.008027937263250351, + 0.011514964513480663, + 0.009764830581843853, + 0.009730078279972076, + 0.011773450300097466, + 0.018776115030050278, + 0.019978202879428864, + 0.01880967617034912, + 0.01660344749689102, + 0.024794355034828186, + 0.024896493181586266, + 0.02057712711393833, + 0.002947748638689518, + 0.0046968706883490086, + 0.005339493975043297, + 0.005495995748788118, + 0.0052547394298017025, + 0.0058004953898489475, + 0.006133062765002251, + 0.005239163059741259, + 0.012062584981322289, + 0.023473143577575684, + 0.021092118695378304, + 0.033636424690485, + 0.03399992734193802, + 0.02829979918897152, + 0.03241371735930443, + 0.024159736931324005, + 0.00301602971740067, + 0.0051736654713749886, + 0.005854455288499594, + 0.007840707898139954, + 0.008568355813622475, + 0.0058988044038414955, + 0.007996422238647938, + 0.005485638976097107, + 0.004706583917140961, + 0.00666650477796793, + 0.00978079903870821, + 0.006201509851962328, + 0.006655237637460232, + 0.007875246927142143, + 0.005575979594141245, + 0.00596298323944211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.026364192366600037 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070831567049026 + ], + "opposite_logprob_tokens": [ + -4.14107084274292 + ], + "logprob_reward": -0.016070831567049026, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070831567049026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35318177938461304, + 0.14359532296657562, + 0.04143659770488739, + 0.05225938931107521, + 0.07839899510145187, + 0.06986351311206818, + 0.022623440250754356, + 0.07611299306154251, + 0.1260671764612198, + 0.03646082431077957 + ], + "credit_map": [ + 0.029239490628242493, + 0.04882627725601196, + 0.04166758432984352, + 0.0429813377559185, + 0.04600856080651283, + 0.05081069469451904, + 0.047536931931972504, + 0.046110864728689194, + 0.01164424791932106, + 0.016684595495462418, + 0.022904930636286736, + 0.026138417422771454, + 0.01763743720948696, + 0.012827820144593716, + 0.020979078486561775, + 0.014778779819607735, + 0.0040105837397277355, + 0.004633571486920118, + 0.005562241654843092, + 0.007461823523044586, + 0.00525852432474494, + 0.004943890497088432, + 0.004772350657731295, + 0.0047936043702065945, + 0.004692543763667345, + 0.00710328109562397, + 0.005440782755613327, + 0.008075007237493992, + 0.007015395909547806, + 0.007880562916398048, + 0.006154964677989483, + 0.005896845366805792, + 0.0069776857271790504, + 0.010561542585492134, + 0.011527610942721367, + 0.008652787655591965, + 0.00830615684390068, + 0.011248618364334106, + 0.011797378771007061, + 0.009327207691967487, + 0.005957788787782192, + 0.008238482289016247, + 0.01098411250859499, + 0.007686541881412268, + 0.008535211905837059, + 0.011290512979030609, + 0.008197071962058544, + 0.008973787538707256, + 0.0016144179971888661, + 0.003232091199606657, + 0.003801738377660513, + 0.0027351658791303635, + 0.0029945536516606808, + 0.0031783103477209806, + 0.002721691969782114, + 0.0023454679176211357, + 0.004805030766874552, + 0.007928447797894478, + 0.00786334928125143, + 0.012717953883111477, + 0.012133782729506493, + 0.009667209349572659, + 0.011593488045036793, + 0.009403725154697895, + 0.008532886393368244, + 0.013228408060967922, + 0.018065618351101875, + 0.015358400531113148, + 0.022446688264608383, + 0.016728825867176056, + 0.019628209993243217, + 0.012078114785254002, + 0.0029366915114223957, + 0.0038314438425004482, + 0.005637553986161947, + 0.004003466572612524, + 0.0053107598796486855, + 0.004858938045799732, + 0.005385361146181822, + 0.004496603272855282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.016070831567049026 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742918610572815 + ], + "opposite_logprob_tokens": [ + -0.9742918610572815 + ], + "logprob_reward": -0.4742918610572815, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742918610572815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10619813203811646, + 0.13062790036201477, + 0.04033288359642029, + 0.05922495573759079, + 0.045409779995679855, + 0.06627213954925537, + 0.06326577067375183, + 0.2254347950220108, + 0.21889294683933258, + 0.04434065893292427 + ], + "credit_map": [ + 0.006855368614196777, + 0.014972829259932041, + 0.0106625035405159, + 0.013323220424354076, + 0.01345008797943592, + 0.010703891515731812, + 0.023833846673369408, + 0.012396382167935371, + 0.010560213588178158, + 0.01388628501445055, + 0.01580309495329857, + 0.01673923432826996, + 0.013787674717605114, + 0.014843633398413658, + 0.03229918330907822, + 0.012708590365946293, + 0.00402983371168375, + 0.004809791222214699, + 0.005059781018644571, + 0.005507849156856537, + 0.004654917400330305, + 0.004727706778794527, + 0.007730648387223482, + 0.0038123547565191984, + 0.005289983935654163, + 0.008124664425849915, + 0.009559355676174164, + 0.007765002548694611, + 0.006695430725812912, + 0.006020551081746817, + 0.009398129768669605, + 0.006371833384037018, + 0.00368523970246315, + 0.006191341672092676, + 0.00559749873355031, + 0.004689627327024937, + 0.005390837322920561, + 0.00790763646364212, + 0.006953708827495575, + 0.004993889015167952, + 0.004724827595055103, + 0.011049241758883, + 0.007054391782730818, + 0.00798504613339901, + 0.007805467117577791, + 0.008932625874876976, + 0.010701213963329792, + 0.008019328117370605, + 0.003861720208078623, + 0.005189982242882252, + 0.007764820475131273, + 0.007794377859681845, + 0.0072409226559102535, + 0.008282202295958996, + 0.014438800513744354, + 0.00869294535368681, + 0.014584919437766075, + 0.028242381289601326, + 0.02498360350728035, + 0.031006263568997383, + 0.03676075488328934, + 0.019000740721821785, + 0.038344599306583405, + 0.03251153230667114, + 0.013893503695726395, + 0.0242164209485054, + 0.026747746393084526, + 0.029815180227160454, + 0.035845331847667694, + 0.029439019039273262, + 0.036926355212926865, + 0.022009393200278282, + 0.003709694603458047, + 0.0065187751315534115, + 0.004640358034521341, + 0.004363831132650375, + 0.006337175145745277, + 0.007288946304470301, + 0.0064218551851809025, + 0.005060024093836546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.4742918610572815 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures the decadent beauty of a perfectly assembled burger, creating a scene of pure culinary desire. Warm, directional light from the side sculpts the burger, highlighting every texture. A thick, vibrant orange river of melted cheddar cheese slowly and deliberately oozes down the side of the rough, char-grilled beef patty, its smooth surface catching the light as it navigates the rugged terrain of meat. The burger is topped with a toasted, golden-brown sesame seed bun, glistening with color. From beneath the patty, the crisp, bright green edge of the lettuce leaf and a slice of juicy, deep-red tomato can be seen. The entire subject is framed by a softly blurred, dark background, ensuring that the focus is entirely on the appetizing spectacle. As the cheese continues its languid, gooey descent, the entire mouth-watering scene, imbued with an air of indulgent satisfaction, gently and smoothly fades to black.", + "seed": 30147, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_11/index_7092_seed_30147_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.28119451412931085, + "vlm_logprob": -0.28119451412931085, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.28119451412931085, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005288776475936174 + ], + "opposite_logprob_tokens": [ + -6.755288600921631 + ], + "logprob_reward": -0.005288776475936174, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.005288776475936174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17654350399971008, + 0.10816501080989838, + 0.1345224231481552, + 0.11157067865133286, + 0.08124943822622299, + 0.06393720954656601, + 0.10578637570142746, + 0.04702679440379143, + 0.10697703063488007, + 0.06422152370214462 + ], + "credit_map": [ + 0.017451481893658638, + 0.019563918933272362, + 0.018544532358646393, + 0.023595938459038734, + 0.019063977524638176, + 0.021349089220166206, + 0.030147431418299675, + 0.0268271341919899, + 0.010349390096962452, + 0.015320532955229282, + 0.012943413108587265, + 0.014115910977125168, + 0.011896247044205666, + 0.016203265637159348, + 0.015207654796540737, + 0.012128591537475586, + 0.01835610158741474, + 0.013079481199383736, + 0.017008403316140175, + 0.016204027459025383, + 0.015084813348948956, + 0.013939142227172852, + 0.01871892809867859, + 0.022131523117423058, + 0.011813139542937279, + 0.010112090967595577, + 0.016116103157401085, + 0.014342039823532104, + 0.009570675902068615, + 0.01230741385370493, + 0.01681732013821602, + 0.02049189992249012, + 0.011091494001448154, + 0.005862499121576548, + 0.007701127324253321, + 0.008491073735058308, + 0.007120940834283829, + 0.010655675083398819, + 0.010382061824202538, + 0.019944561645388603, + 0.007380389608442783, + 0.008420553058385849, + 0.008972523733973503, + 0.00850966852158308, + 0.006589158438146114, + 0.007106675300747156, + 0.009012414142489433, + 0.007945820689201355, + 0.0123151745647192, + 0.010467685759067535, + 0.009776740334928036, + 0.012843143194913864, + 0.006456834264099598, + 0.02507135644555092, + 0.014833365567028522, + 0.014022080227732658, + 0.007665383163839579, + 0.0056868321262300014, + 0.006217727903276682, + 0.004896761849522591, + 0.0069702668115496635, + 0.005490019917488098, + 0.005415941588580608, + 0.0046838573180139065, + 0.013516491279006004, + 0.015163276344537735, + 0.011479810811579227, + 0.01082460768520832, + 0.015041402541100979, + 0.011122793890535831, + 0.011586700566112995, + 0.018241940066218376, + 0.007058843970298767, + 0.0072038182988762856, + 0.006580682937055826, + 0.008960096165537834, + 0.006707283202558756, + 0.009861871600151062, + 0.008293060585856438, + 0.009555868804454803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.005288776475936174 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900695502758026 + ], + "opposite_logprob_tokens": [ + -2.5790069103240967 + ], + "logprob_reward": -0.07900695502758026, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900695502758026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08574815094470978, + 0.0984136313199997, + 0.09043224900960922, + 0.16915284097194672, + 0.14447182416915894, + 0.050613272935152054, + 0.11410529911518097, + 0.1007414162158966, + 0.06956393271684647, + 0.07675736397504807 + ], + "credit_map": [ + 0.012898593209683895, + 0.010539841838181019, + 0.01203436404466629, + 0.010171083733439445, + 0.008997969329357147, + 0.0091106491163373, + 0.010049933567643166, + 0.011945712380111217, + 0.01705535687506199, + 0.010125487111508846, + 0.011477247811853886, + 0.010603079572319984, + 0.011324592866003513, + 0.017052005976438522, + 0.00907844491302967, + 0.011697405949234962, + 0.010532628744840622, + 0.008800478652119637, + 0.010166040621697903, + 0.008767063729465008, + 0.00766227301210165, + 0.014697027392685413, + 0.013481232337653637, + 0.01632550358772278, + 0.01507011242210865, + 0.015054545365273952, + 0.027202332392334938, + 0.02121795527637005, + 0.014136592857539654, + 0.0216219499707222, + 0.024594340473413467, + 0.03025500662624836, + 0.01876463182270527, + 0.01408351119607687, + 0.017195872962474823, + 0.016252852976322174, + 0.012346342206001282, + 0.019476572051644325, + 0.019387833774089813, + 0.026964200660586357, + 0.0074292453937232494, + 0.005764861591160297, + 0.005677702836692333, + 0.006121905520558357, + 0.004545312840491533, + 0.006917466875165701, + 0.006603431422263384, + 0.007553350646048784, + 0.016047116369009018, + 0.009862582199275494, + 0.012129698880016804, + 0.012370239943265915, + 0.007993318140506744, + 0.01874159835278988, + 0.015362516976892948, + 0.021598223596811295, + 0.012230009771883488, + 0.012339177541434765, + 0.0115882633253932, + 0.015870967879891396, + 0.011127006262540817, + 0.010096674785017967, + 0.016856428235769272, + 0.010632883757352829, + 0.007039180491119623, + 0.010009314864873886, + 0.008132084272801876, + 0.009905804879963398, + 0.01009488757699728, + 0.007206369191408157, + 0.009400789625942707, + 0.007775499951094389, + 0.007457788567990065, + 0.010602730326354504, + 0.008534365333616734, + 0.010556667111814022, + 0.009378653019666672, + 0.007349611725658178, + 0.009380944073200226, + 0.013496607542037964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 6 + ], + "selected_reward": -0.07900695502758026 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307961463928223 + ], + "opposite_logprob_tokens": [ + -3.1680796146392822 + ], + "logprob_reward": -0.04307961463928223, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307961463928223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14398378133773804, + 0.09813860058784485, + 0.16126015782356262, + 0.2313407063484192, + 0.13762715458869934, + 0.08207297325134277, + 0.028588376939296722, + 0.0387289933860302, + 0.02296566404402256, + 0.055293601006269455 + ], + "credit_map": [ + 0.020590277388691902, + 0.017097093164920807, + 0.017810683697462082, + 0.016209596768021584, + 0.013067261315882206, + 0.02429809793829918, + 0.014831299893558025, + 0.020079461857676506, + 0.010652127675712109, + 0.011563633568584919, + 0.010054150596261024, + 0.009724867530167103, + 0.010358031839132309, + 0.015879448503255844, + 0.011105794459581375, + 0.01880054734647274, + 0.017700234428048134, + 0.018255939707159996, + 0.022846221923828125, + 0.02194545418024063, + 0.015330172143876553, + 0.02073048986494541, + 0.022559132426977158, + 0.021892523393034935, + 0.024280346930027008, + 0.020996281877160072, + 0.035120733082294464, + 0.026947584003210068, + 0.021549398079514503, + 0.026340752840042114, + 0.02933279424905777, + 0.04677281528711319, + 0.013557263650000095, + 0.012416082434356213, + 0.017781855538487434, + 0.014934691600501537, + 0.009626513347029686, + 0.017278920859098434, + 0.016076594591140747, + 0.03595523536205292, + 0.00868112314492464, + 0.009220737963914871, + 0.011893657967448235, + 0.010047641582787037, + 0.008272774517536163, + 0.009183966554701328, + 0.011738375760614872, + 0.013034696690738201, + 0.0027662101201713085, + 0.002689099172130227, + 0.0029403914231806993, + 0.003969642333686352, + 0.0021425846498459578, + 0.006636503152549267, + 0.003851764602586627, + 0.0035921805538237095, + 0.004311990458518267, + 0.0030080873984843493, + 0.003923112526535988, + 0.0052152411080896854, + 0.002589872572571039, + 0.011197112500667572, + 0.003812325419858098, + 0.004671250935643911, + 0.002394999610260129, + 0.003619458759203553, + 0.002207794226706028, + 0.002251646714285016, + 0.0030584114138036966, + 0.0031295912340283394, + 0.002482777927070856, + 0.0038209843914955854, + 0.005956676788628101, + 0.0061063701286911964, + 0.0062204645946621895, + 0.0075925616547465324, + 0.0060387905687093735, + 0.006989406421780586, + 0.007885371334850788, + 0.008503957651555538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.04307961463928223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496727302670479 + ], + "opposite_logprob_tokens": [ + -2.929967164993286 + ], + "logprob_reward": -0.05496727302670479, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496727302670479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08690212666988373, + 0.0696055218577385, + 0.1089092493057251, + 0.13378189504146576, + 0.2742098867893219, + 0.04429825022816658, + 0.09370706230401993, + 0.13370470702648163, + 0.009070776402950287, + 0.04581053927540779 + ], + "credit_map": [ + 0.01049759890884161, + 0.008686588145792484, + 0.008993349969387054, + 0.00970353465527296, + 0.008234376087784767, + 0.0177992545068264, + 0.008632820099592209, + 0.014354601502418518, + 0.009478151798248291, + 0.0058598704636096954, + 0.007072994019836187, + 0.005956302396953106, + 0.008104957640171051, + 0.013563097454607487, + 0.007606936153024435, + 0.011963204480707645, + 0.011068944819271564, + 0.011683100834488869, + 0.013705058954656124, + 0.011972295120358467, + 0.010588199831545353, + 0.012675019912421703, + 0.013257823884487152, + 0.023958811536431313, + 0.013598702847957611, + 0.012009333819150925, + 0.018046915531158447, + 0.014286876656115055, + 0.011513053439557552, + 0.01899579167366028, + 0.016936596482992172, + 0.028394609689712524, + 0.03306856006383896, + 0.02792581543326378, + 0.0373447947204113, + 0.03258923441171646, + 0.026086926460266113, + 0.03506225720047951, + 0.03518235683441162, + 0.04694993421435356, + 0.0054471781477332115, + 0.003945035859942436, + 0.00539242010563612, + 0.005245137959718704, + 0.003689409466460347, + 0.008500196970999241, + 0.005720372311770916, + 0.006358493119478226, + 0.009879833087325096, + 0.009014622308313847, + 0.01053087692707777, + 0.013829494826495647, + 0.008291489444673061, + 0.011548731476068497, + 0.017056161537766457, + 0.013555840589106083, + 0.01595325581729412, + 0.015794845297932625, + 0.012764505110681057, + 0.01750623993575573, + 0.012374863028526306, + 0.025042513385415077, + 0.0200311541557312, + 0.014237327501177788, + 0.0012273252941668034, + 0.001170532195828855, + 0.0009325413266196847, + 0.0010893834987655282, + 0.0010478992480784655, + 0.001090967794880271, + 0.00137701362837106, + 0.0011351132998242974, + 0.0038180253468453884, + 0.005225597880780697, + 0.00409134803339839, + 0.005365326069295406, + 0.0055344426073133945, + 0.008939849212765694, + 0.005987042561173439, + 0.006848901975899935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.05496727302670479 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236299514770508 + ], + "opposite_logprob_tokens": [ + -0.34862998127937317 + ], + "logprob_reward": -1.2236299514770508, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236299514770508, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04610111564397812, + 0.08615870028734207, + 0.03743450343608856, + 0.21386323869228363, + 0.06549961864948273, + 0.1045721247792244, + 0.20522578060626984, + 0.15124091506004333, + 0.02698819898068905, + 0.06291580200195312 + ], + "credit_map": [ + 0.006678313948214054, + 0.005701989866793156, + 0.0053833890706300735, + 0.006247240584343672, + 0.004680907353758812, + 0.004840231500566006, + 0.008280023001134396, + 0.0042890203185379505, + 0.010776225477457047, + 0.00827920250594616, + 0.012179729528725147, + 0.006639150436967611, + 0.010706041939556599, + 0.015658333897590637, + 0.007546539884060621, + 0.014373472891747952, + 0.004245954565703869, + 0.004725235514342785, + 0.004991185385733843, + 0.004677832126617432, + 0.004640099126845598, + 0.0044303466565907, + 0.004288432188332081, + 0.00543541694059968, + 0.021809469908475876, + 0.017972618341445923, + 0.03414732217788696, + 0.02433224394917488, + 0.018350088968873024, + 0.02452240325510502, + 0.027764299884438515, + 0.04496478661894798, + 0.010152439586818218, + 0.0054609449580311775, + 0.011391393840312958, + 0.0076503874734044075, + 0.005583173129707575, + 0.008535548113286495, + 0.007491696160286665, + 0.009234030731022358, + 0.011469033546745777, + 0.010969141498208046, + 0.01224092673510313, + 0.01274093147367239, + 0.008929410949349403, + 0.013175006955862045, + 0.015192005783319473, + 0.019855665042996407, + 0.019610967487096786, + 0.01703016832470894, + 0.021143769845366478, + 0.03114885278046131, + 0.016694338992238045, + 0.02386939898133278, + 0.03900124132633209, + 0.03672703355550766, + 0.017991451546549797, + 0.017277805134654045, + 0.013182055205106735, + 0.02292611449956894, + 0.01728525571525097, + 0.016946492716670036, + 0.02882941998541355, + 0.016802318394184113, + 0.003339045448228717, + 0.0018748192815110087, + 0.003785260021686554, + 0.002919940510764718, + 0.001759010716341436, + 0.0033646030351519585, + 0.0038475235924124718, + 0.006097996607422829, + 0.00490714842453599, + 0.005893297493457794, + 0.00593971274793148, + 0.008475962094962597, + 0.0068060001358389854, + 0.008964731357991695, + 0.009329242631793022, + 0.012599712237715721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -1.2236299514770508 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extreme close-up static shot captures the decadent beauty of a perfectly assembled burger, creating a scene of pure culinary desire. Warm, directional light from the side sculpts the burger, highlighting every texture. A thick, vibrant orange river of melted cheddar cheese slowly and deliberately oozes down the side of the rough, char-grilled beef patty, its smooth surface catching the light as it navigates the rugged terrain of meat. The burger is topped with a toasted, golden-brown sesame seed bun, glistening with color. From beneath the patty, the crisp, bright green edge of the lettuce leaf and a slice of juicy, deep-red tomato can be seen. The entire subject is framed by a softly blurred, dark background, ensuring that the focus is entirely on the appetizing spectacle. As the cheese continues its languid, gooey descent, the entire mouth-watering scene, imbued with an air of indulgent satisfaction, gently and smoothly fades to black.", + "seed": 30148, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_11/index_7092_seed_30148_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.16987739177420735, + "vlm_logprob": -0.16987739177420735, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.16987739177420735, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003957892768085003 + ], + "opposite_logprob_tokens": [ + -8.003957748413086 + ], + "logprob_reward": -0.003957892768085003, + "yes_no_margin_reward": 8.0, + "token_credit_reward": -0.003957892768085003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09751719981431961, + 0.04696181043982506, + 0.03810226917266846, + 0.09433373808860779, + 0.07830540835857391, + 0.1136530190706253, + 0.30484986305236816, + 0.10690593719482422, + 0.03831639885902405, + 0.08105429261922836 + ], + "credit_map": [ + 0.009060961194336414, + 0.013379933312535286, + 0.011362089775502682, + 0.01448904536664486, + 0.017766330391168594, + 0.012357578612864017, + 0.009950177744030952, + 0.009151093661785126, + 0.0035248517524451017, + 0.005781342275440693, + 0.006713441107422113, + 0.0051806531846523285, + 0.0047971345484256744, + 0.005437989719212055, + 0.005655480548739433, + 0.00987092312425375, + 0.002825640607625246, + 0.003912966698408127, + 0.006084935739636421, + 0.004781362600624561, + 0.004557241685688496, + 0.003387004602700472, + 0.006565057206898928, + 0.005988064222037792, + 0.006058970466256142, + 0.012690790928900242, + 0.0115797258913517, + 0.007761434186249971, + 0.010688125155866146, + 0.012121625244617462, + 0.013179083354771137, + 0.020253995433449745, + 0.0052783917635679245, + 0.008282404392957687, + 0.008147669024765491, + 0.00745038315653801, + 0.015105684287846088, + 0.0068923006765544415, + 0.011569580994546413, + 0.015578998252749443, + 0.007613671012222767, + 0.01998359151184559, + 0.012616213411092758, + 0.01311535481363535, + 0.019615819677710533, + 0.011914060451090336, + 0.01095909345895052, + 0.017835231497883797, + 0.01927901804447174, + 0.04691825434565544, + 0.03601968660950661, + 0.031036200001835823, + 0.04883648455142975, + 0.029641183093190193, + 0.0409892201423645, + 0.05212981998920441, + 0.00569158885627985, + 0.024130159988999367, + 0.008644584566354752, + 0.008366851136088371, + 0.031363025307655334, + 0.01231721043586731, + 0.008476664312183857, + 0.007915867492556572, + 0.003402023809030652, + 0.003833172144368291, + 0.005775402765721083, + 0.004360716789960861, + 0.0063691348768770695, + 0.0047969999723136425, + 0.005381330847740173, + 0.004397619049996138, + 0.0049963598139584064, + 0.017141330987215042, + 0.009720784611999989, + 0.008059359155595303, + 0.02265959233045578, + 0.006310492753982544, + 0.006892802659422159, + 0.00527358241379261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.003957892768085003 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761617422103882 + ], + "opposite_logprob_tokens": [ + -0.8261617422103882 + ], + "logprob_reward": -0.5761617422103882, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761617422103882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05834395810961723, + 0.06987834721803665, + 0.12043440341949463, + 0.07411803305149078, + 0.07983259111642838, + 0.18876312673091888, + 0.14586076140403748, + 0.1543482393026352, + 0.06462150812149048, + 0.043799057602882385 + ], + "credit_map": [ + 0.005747694987803698, + 0.010740465484559536, + 0.006418050732463598, + 0.006909569259732962, + 0.009129003621637821, + 0.006090036127716303, + 0.0048622917383909225, + 0.00844684150069952, + 0.005049186293035746, + 0.011371688917279243, + 0.009507616050541401, + 0.009940174408257008, + 0.008878431282937527, + 0.009574934840202332, + 0.007260095328092575, + 0.008296218700706959, + 0.009291330352425575, + 0.019910525530576706, + 0.01805059053003788, + 0.013101213611662388, + 0.017528211697936058, + 0.013638587668538094, + 0.015251511707901955, + 0.01366242952644825, + 0.006295756436884403, + 0.015762800350785255, + 0.009215040132403374, + 0.007662631571292877, + 0.011689889244735241, + 0.007656540721654892, + 0.007219419814646244, + 0.008615956641733646, + 0.00666091637685895, + 0.009075045585632324, + 0.012199613265693188, + 0.009944326244294643, + 0.01237719040364027, + 0.009096689522266388, + 0.010943269357085228, + 0.009535537101328373, + 0.012581519782543182, + 0.030604831874370575, + 0.020152214914560318, + 0.016635870561003685, + 0.039294373244047165, + 0.02032221294939518, + 0.020703652873635292, + 0.028468448668718338, + 0.010931159369647503, + 0.016578178852796555, + 0.019604967907071114, + 0.017595427110791206, + 0.02233288250863552, + 0.01298514660447836, + 0.02312808483839035, + 0.022704916074872017, + 0.014560551382601261, + 0.017843320965766907, + 0.017476720735430717, + 0.014377454295754433, + 0.02090323157608509, + 0.02080484665930271, + 0.02235003188252449, + 0.026032080873847008, + 0.004209638107568026, + 0.012515594251453876, + 0.007615359500050545, + 0.005531016271561384, + 0.01664566434919834, + 0.0053563192486763, + 0.007278965320438147, + 0.005468952003866434, + 0.0018854039954021573, + 0.014865338802337646, + 0.0025952402502298355, + 0.0018666933756321669, + 0.015992475673556328, + 0.002208506688475609, + 0.0021249866113066673, + 0.002260414883494377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.5761617422103882 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012554777786135674 + ], + "opposite_logprob_tokens": [ + -4.38755464553833 + ], + "logprob_reward": -0.012554777786135674, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012554777786135674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09446675330400467, + 0.14338605105876923, + 0.2703213095664978, + 0.13146758079528809, + 0.08959226310253143, + 0.03391076624393463, + 0.0794823169708252, + 0.051395490765571594, + 0.04194597154855728, + 0.06403150409460068 + ], + "credit_map": [ + 0.015379304997622967, + 0.010921184904873371, + 0.014589116908609867, + 0.009101836942136288, + 0.012677475810050964, + 0.009002670645713806, + 0.007622975390404463, + 0.015172184444963932, + 0.007950054481625557, + 0.01817586086690426, + 0.014497058466076851, + 0.016084739938378334, + 0.037800390273332596, + 0.02174813114106655, + 0.012232547625899315, + 0.01489727571606636, + 0.021905891597270966, + 0.05423945188522339, + 0.04105150327086449, + 0.02649766206741333, + 0.0344233438372612, + 0.03042432852089405, + 0.032102420926094055, + 0.02967672236263752, + 0.008981751278042793, + 0.012796636670827866, + 0.016714537516236305, + 0.012462331913411617, + 0.01618540845811367, + 0.013265874236822128, + 0.0226159505546093, + 0.02844509854912758, + 0.0067489007487893105, + 0.015041868202388287, + 0.009554585441946983, + 0.008658327162265778, + 0.017223039641976357, + 0.008892960846424103, + 0.009087852202355862, + 0.01438472606241703, + 0.002833297476172447, + 0.004299485124647617, + 0.004861472640186548, + 0.0048123858869075775, + 0.004740928765386343, + 0.003001403994858265, + 0.004513229709118605, + 0.00484856404364109, + 0.005587394814938307, + 0.011247826740145683, + 0.010065971873700619, + 0.009550456888973713, + 0.012013262137770653, + 0.007293940056115389, + 0.009163493290543556, + 0.014559971168637276, + 0.0036694456357508898, + 0.009160061366856098, + 0.005787307862192392, + 0.006087210029363632, + 0.00963712390512228, + 0.004291110672056675, + 0.006131389643996954, + 0.006631840020418167, + 0.0033143942710012197, + 0.0069289361126720905, + 0.0059730117209255695, + 0.005218897946178913, + 0.008906790055334568, + 0.0035260848235338926, + 0.004260391928255558, + 0.0038174667861312628, + 0.003636739682406187, + 0.012104391120374203, + 0.008618042804300785, + 0.006671696435660124, + 0.016239283606410027, + 0.005535646807402372, + 0.006288595963269472, + 0.004937105346471071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.012554777786135674 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046372017823159695 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.0046372017823159695, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046372017823159695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09363090246915817, + 0.1555924117565155, + 0.07804624736309052, + 0.06144769862294197, + 0.05256432667374611, + 0.0701761469244957, + 0.2257080376148224, + 0.07762708514928818, + 0.10028781741857529, + 0.08491932600736618 + ], + "credit_map": [ + 0.01701909489929676, + 0.013660365715622902, + 0.013312434777617455, + 0.0074514118023216724, + 0.0148362061008811, + 0.009775265119969845, + 0.006220115348696709, + 0.011356010101735592, + 0.011408288963139057, + 0.020408278331160545, + 0.013696047477424145, + 0.022283567115664482, + 0.031713034957647324, + 0.020563345402479172, + 0.01665337383747101, + 0.018866488710045815, + 0.005314776208251715, + 0.007422836497426033, + 0.009161852300167084, + 0.01108855102211237, + 0.01601381227374077, + 0.011578688398003578, + 0.008694120682775974, + 0.00877161044627428, + 0.0047792671248316765, + 0.008068995550274849, + 0.008165514096617699, + 0.005512432660907507, + 0.0074614048935472965, + 0.006849201861768961, + 0.008111248724162579, + 0.012499640695750713, + 0.003871957305818796, + 0.009768911637365818, + 0.0051639084704220295, + 0.0050099859945476055, + 0.011014457792043686, + 0.005115764681249857, + 0.006072612013667822, + 0.006546728778630495, + 0.005033298395574093, + 0.014980172738432884, + 0.006573694292455912, + 0.004651824943721294, + 0.017748741433024406, + 0.008017711341381073, + 0.006822836119681597, + 0.006347863934934139, + 0.013766232877969742, + 0.034218572080135345, + 0.026569150388240814, + 0.024165984243154526, + 0.041190147399902344, + 0.020902078598737717, + 0.029034720733761787, + 0.03586117923259735, + 0.00702071376144886, + 0.00893372017890215, + 0.008866592310369015, + 0.007966523058712482, + 0.013749564066529274, + 0.009045148268342018, + 0.009807517752051353, + 0.012237310409545898, + 0.008683506399393082, + 0.015409661456942558, + 0.016744250431656837, + 0.01044976245611906, + 0.018859989941120148, + 0.008889519609510899, + 0.01237550750374794, + 0.00887562707066536, + 0.007073067594319582, + 0.010652842000126839, + 0.011532315984368324, + 0.014262151904404163, + 0.012422355823218822, + 0.009031691588461399, + 0.008848745375871658, + 0.01109615620225668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.0046372017823159695 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25207534432411194 + ], + "opposite_logprob_tokens": [ + -1.5020753145217896 + ], + "logprob_reward": -0.25207534432411194, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25207534432411194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04081130772829056, + 0.12867920100688934, + 0.12045672535896301, + 0.07948552817106247, + 0.0587085522711277, + 0.06441856175661087, + 0.21085745096206665, + 0.16450613737106323, + 0.08135920763015747, + 0.05071737989783287 + ], + "credit_map": [ + 0.0033643043134361506, + 0.007764067035168409, + 0.0038848971016705036, + 0.006203009281307459, + 0.00555288465693593, + 0.0044560860842466354, + 0.004541619680821896, + 0.005044438410550356, + 0.0084938770160079, + 0.01979627087712288, + 0.014985409565269947, + 0.014171003364026546, + 0.028223274275660515, + 0.016911793500185013, + 0.013888740912079811, + 0.012208825908601284, + 0.007987954653799534, + 0.01890324056148529, + 0.01446421816945076, + 0.011130526661872864, + 0.02565700002014637, + 0.0170300230383873, + 0.013029339723289013, + 0.012254421599209309, + 0.007063922472298145, + 0.012004426680505276, + 0.014376217499375343, + 0.010148335248231888, + 0.009789515286684036, + 0.005396724212914705, + 0.010378952138125896, + 0.010327436029911041, + 0.004472742788493633, + 0.010372022166848183, + 0.007439538836479187, + 0.005850463639944792, + 0.012289457023143768, + 0.0068104756064713, + 0.006527848541736603, + 0.004946004133671522, + 0.004790245555341244, + 0.010435225442051888, + 0.007878551259636879, + 0.006018621847033501, + 0.010374999605119228, + 0.007026042323559523, + 0.0065542785450816154, + 0.011340598575770855, + 0.01491138618439436, + 0.02464429661631584, + 0.02879178151488304, + 0.02425624616444111, + 0.03033703938126564, + 0.020229147747159004, + 0.027788039296865463, + 0.03989949822425842, + 0.012242505326867104, + 0.02437794767320156, + 0.019491782411932945, + 0.016676118597388268, + 0.030687859281897545, + 0.018052980303764343, + 0.022844485938549042, + 0.020132461562752724, + 0.005429315846413374, + 0.009217198006808758, + 0.012640560045838356, + 0.009524431079626083, + 0.017098579555749893, + 0.007534692529588938, + 0.011701385490596294, + 0.008213045075535774, + 0.004112691152840853, + 0.011764204129576683, + 0.004696556832641363, + 0.0035984020214527845, + 0.013345187529921532, + 0.003939827438443899, + 0.004402782768011093, + 0.0048577249981462955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.25207534432411194 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up static shot captures the decadent beauty of a perfectly assembled burger, creating a scene of pure culinary desire. Warm, directional light from the side sculpts the burger, highlighting every texture. A thick, vibrant orange river of melted cheddar cheese slowly and deliberately oozes down the side of the rough, char-grilled beef patty, its smooth surface catching the light as it navigates the rugged terrain of meat. The burger is topped with a toasted, golden-brown sesame seed bun, glistening with color. From beneath the patty, the crisp, bright green edge of the lettuce leaf and a slice of juicy, deep-red tomato can be seen. The entire subject is framed by a softly blurred, dark background, ensuring that the focus is entirely on the appetizing spectacle. As the cheese continues its languid, gooey descent, the entire mouth-watering scene, imbued with an air of indulgent satisfaction, gently and smoothly fades to black.", + "seed": 30149, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_11/index_7092_seed_30149_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16804860197007657, + "vlm_logprob": -0.16804860197007657, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.16804860197007657, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007454673293977976 + ], + "opposite_logprob_tokens": [ + -6.132454872131348 + ], + "logprob_reward": -0.007454673293977976, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.007454673293977976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24450723826885223, + 0.09009253978729248, + 0.06057528778910637, + 0.044952988624572754, + 0.10797642171382904, + 0.060822129249572754, + 0.0881095826625824, + 0.037033434957265854, + 0.20293204486370087, + 0.06299828737974167 + ], + "credit_map": [ + 0.021756257861852646, + 0.033908307552337646, + 0.04116850718855858, + 0.02747650071978569, + 0.0267100241035223, + 0.019017983227968216, + 0.035617388784885406, + 0.0388522669672966, + 0.006794822867959738, + 0.010191662237048149, + 0.01708776131272316, + 0.013878367841243744, + 0.016180671751499176, + 0.006173735484480858, + 0.00975395180284977, + 0.010031567886471748, + 0.005212366580963135, + 0.008577970787882805, + 0.0075944834388792515, + 0.005479868967086077, + 0.01267912145704031, + 0.006029047537595034, + 0.007597451098263264, + 0.007404976524412632, + 0.004128787200897932, + 0.004999516066163778, + 0.006176997907459736, + 0.005690847523510456, + 0.005853308830410242, + 0.004546464886516333, + 0.008118255995213985, + 0.005438807886093855, + 0.009822584688663483, + 0.012304600328207016, + 0.016870079562067986, + 0.010178718715906143, + 0.013217656873166561, + 0.007690319325774908, + 0.020493555814027786, + 0.017398905009031296, + 0.0057298969477415085, + 0.007324593141674995, + 0.015092141926288605, + 0.006805143319070339, + 0.006599223241209984, + 0.004075710196048021, + 0.007527564186602831, + 0.007667855825275183, + 0.008951201103627682, + 0.01296615693718195, + 0.021626513451337814, + 0.008930282667279243, + 0.00928470678627491, + 0.007188680116087198, + 0.010880906134843826, + 0.008281141519546509, + 0.0032668679486960173, + 0.005545743275433779, + 0.005866411607712507, + 0.004693490918725729, + 0.004236768931150436, + 0.0032784524373710155, + 0.005935200955718756, + 0.004210498183965683, + 0.020382022485136986, + 0.027231726795434952, + 0.02406024932861328, + 0.03302566707134247, + 0.03662208467721939, + 0.023535743355751038, + 0.019903328269720078, + 0.018171235918998718, + 0.007037648931145668, + 0.011165912263095379, + 0.00792459025979042, + 0.009526574052870274, + 0.006158496253192425, + 0.004425957333296537, + 0.009535383433103561, + 0.0072237225249409676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.007454673293977976 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762055516242981 + ], + "opposite_logprob_tokens": [ + -0.8262055516242981 + ], + "logprob_reward": -0.5762055516242981, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762055516242981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16616128385066986, + 0.12005101144313812, + 0.074942946434021, + 0.24146971106529236, + 0.07622107863426208, + 0.03845023736357689, + 0.10683071613311768, + 0.021912386640906334, + 0.08368540555238724, + 0.0702752023935318 + ], + "credit_map": [ + 0.013870970346033573, + 0.021909834817051888, + 0.023515356704592705, + 0.017463214695453644, + 0.02999318204820156, + 0.012913315556943417, + 0.02094791829586029, + 0.02554747462272644, + 0.010705199092626572, + 0.013793978840112686, + 0.019971074536442757, + 0.014042653143405914, + 0.0167906004935503, + 0.010446620173752308, + 0.015200838446617126, + 0.019100040197372437, + 0.006275502033531666, + 0.011059350334107876, + 0.009595033712685108, + 0.008169082924723625, + 0.012422717176377773, + 0.0064950967207551, + 0.009672466665506363, + 0.011253699660301208, + 0.0205115657299757, + 0.039064209908246994, + 0.032847922295331955, + 0.026181545108556747, + 0.0444306880235672, + 0.018349414691329002, + 0.03219442442059517, + 0.02788993902504444, + 0.00631307065486908, + 0.008796674199402332, + 0.010298735462129116, + 0.0073931291699409485, + 0.008778666146099567, + 0.005257155280560255, + 0.01691511832177639, + 0.012468523345887661, + 0.002810582285746932, + 0.0069960919208824635, + 0.008948546834290028, + 0.00403795437887311, + 0.004218460060656071, + 0.0028728244360536337, + 0.003909536637365818, + 0.0046562375500798225, + 0.011243544518947601, + 0.012543007731437683, + 0.02342556044459343, + 0.01212288998067379, + 0.011241949163377285, + 0.008947622962296009, + 0.012763605453073978, + 0.014542527496814728, + 0.0028341885190457106, + 0.002608734415844083, + 0.003146703587844968, + 0.0036404321435838938, + 0.002162704011425376, + 0.001954687060788274, + 0.003018095623701811, + 0.0025468398816883564, + 0.0077669937163591385, + 0.01069964561611414, + 0.010010342113673687, + 0.01663023792207241, + 0.011536933481693268, + 0.008809364400804043, + 0.010481464676558971, + 0.007750410120934248, + 0.00982647854834795, + 0.009667040780186653, + 0.009368008002638817, + 0.010409153066575527, + 0.007154541090130806, + 0.005331530701369047, + 0.009524340741336346, + 0.00899410992860794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.5762055516242981 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370925828814507 + ], + "opposite_logprob_tokens": [ + -3.6513710021972656 + ], + "logprob_reward": -0.026370925828814507, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370925828814507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21972033381462097, + 0.12966322898864746, + 0.04291090369224548, + 0.07069466263055801, + 0.18560633063316345, + 0.034211378544569016, + 0.0772288590669632, + 0.014275866560637951, + 0.09016180783510208, + 0.1355266124010086 + ], + "credit_map": [ + 0.016268104314804077, + 0.027731269598007202, + 0.03401802107691765, + 0.019774800166487694, + 0.03219785541296005, + 0.019677147269248962, + 0.028643187135457993, + 0.04140995442867279, + 0.009362191893160343, + 0.014453628100454807, + 0.030587630346417427, + 0.014041503891348839, + 0.018935473635792732, + 0.007752244360744953, + 0.016971686854958534, + 0.017558865249156952, + 0.0037420406006276608, + 0.008781295269727707, + 0.004021108616143465, + 0.003663128474727273, + 0.008246145211160183, + 0.003130507655441761, + 0.006114567629992962, + 0.005212110932916403, + 0.006736543029546738, + 0.011116881854832172, + 0.009409886784851551, + 0.006149740889668465, + 0.012091318145394325, + 0.005766026675701141, + 0.00854289997369051, + 0.010881371796131134, + 0.01452756579965353, + 0.02233247645199299, + 0.029339417815208435, + 0.022227326408028603, + 0.023643692955374718, + 0.014690477401018143, + 0.03500731661915779, + 0.02383805252611637, + 0.003647934878244996, + 0.005763588473200798, + 0.005011436995118856, + 0.0033258269540965557, + 0.003809989197179675, + 0.0033821354154497385, + 0.004593315068632364, + 0.004677152261137962, + 0.0071754190139472485, + 0.010756934061646461, + 0.016421228647232056, + 0.009078913368284702, + 0.009368046186864376, + 0.006051084958016872, + 0.009216354228556156, + 0.009160884656012058, + 0.0017998373368754983, + 0.002072598086670041, + 0.0020493371412158012, + 0.0020413792226463556, + 0.001846245606429875, + 0.0010653914650902152, + 0.0018296170746907592, + 0.0015714607434347272, + 0.005568360444158316, + 0.022255336865782738, + 0.007938072085380554, + 0.02343263290822506, + 0.01103970967233181, + 0.006202707067131996, + 0.008160818368196487, + 0.005564169026911259, + 0.012846449390053749, + 0.01525067538022995, + 0.018606526777148247, + 0.024688083678483963, + 0.01112108863890171, + 0.010782688856124878, + 0.02539862133562565, + 0.01683248206973076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.026370925828814507 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046372017823159695 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.0046372017823159695, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046372017823159695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.102573461830616, + 0.11772483587265015, + 0.12811213731765747, + 0.0881868526339531, + 0.1541949361562729, + 0.06000024825334549, + 0.0426655039191246, + 0.059120867401361465, + 0.20196835696697235, + 0.045452818274497986 + ], + "credit_map": [ + 0.008838438428938389, + 0.01722034066915512, + 0.01204238273203373, + 0.011821187101304531, + 0.017059633508324623, + 0.009961515665054321, + 0.011959371156990528, + 0.013670600950717926, + 0.010475541464984417, + 0.020233621820807457, + 0.016272632405161858, + 0.009943725541234016, + 0.02768484130501747, + 0.0087437080219388, + 0.012154321186244488, + 0.012216444127261639, + 0.016195273026823997, + 0.014107182621955872, + 0.02165857143700123, + 0.011762229725718498, + 0.014554406516253948, + 0.01157327089458704, + 0.013467331416904926, + 0.024793867021799088, + 0.008723443374037743, + 0.016128163784742355, + 0.011855360120534897, + 0.009379318915307522, + 0.015542634762823582, + 0.0063109747134149075, + 0.010922681540250778, + 0.009324280545115471, + 0.011672033928334713, + 0.017417574301362038, + 0.02499292977154255, + 0.020475946366786957, + 0.01825990155339241, + 0.009889109060168266, + 0.034464914351701736, + 0.01702253147959709, + 0.006961096078157425, + 0.008618906140327454, + 0.008719570934772491, + 0.006592215970158577, + 0.008553054183721542, + 0.004774321801960468, + 0.007759355939924717, + 0.00802172813564539, + 0.00667409086599946, + 0.005415302701294422, + 0.008413038216531277, + 0.006186461541801691, + 0.0038548510055989027, + 0.003555901814252138, + 0.0038225988391786814, + 0.0047432612627744675, + 0.007791086565703154, + 0.010412205941975117, + 0.007418544963002205, + 0.005518477875739336, + 0.008203130215406418, + 0.0057623847387731075, + 0.007445283699780703, + 0.00656975107267499, + 0.018001064658164978, + 0.05545731261372566, + 0.020739611238241196, + 0.038967739790678024, + 0.024951687082648277, + 0.01316478755325079, + 0.017181891947984695, + 0.01350426021963358, + 0.00554552860558033, + 0.006573474034667015, + 0.0057725184597074986, + 0.008561406284570694, + 0.004233306739479303, + 0.004589024465531111, + 0.004951789043843746, + 0.005225774832069874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.0046372017823159695 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a burger?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a thick, vibrant orange river of melted cheddar cheese oozing down the side of the beef patty?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background softly blurred and dark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene gently and smoothly fade to black at the end?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557465732097626 + ], + "opposite_logprob_tokens": [ + -1.6005746126174927 + ], + "logprob_reward": -0.22557465732097626, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557465732097626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05275615304708481, + 0.040205154567956924, + 0.040223129093647, + 0.05189986899495125, + 0.05342748388648033, + 0.049507953226566315, + 0.10812424123287201, + 0.07400280237197876, + 0.4625662565231323, + 0.067286916077137 + ], + "credit_map": [ + 0.005375440698117018, + 0.008756760507822037, + 0.008313627913594246, + 0.00675159553065896, + 0.006051267497241497, + 0.003988426178693771, + 0.00680500129237771, + 0.0067140329629182816, + 0.003691855352371931, + 0.004405462183058262, + 0.0060502272099256516, + 0.010585443116724491, + 0.004588651470839977, + 0.002318635117262602, + 0.005194053519517183, + 0.0033708286937326193, + 0.0037624763790518045, + 0.005547301843762398, + 0.005965994670987129, + 0.003827177919447422, + 0.00563048804178834, + 0.00351310963742435, + 0.004629261326044798, + 0.00734731974080205, + 0.005031718406826258, + 0.008364873938262463, + 0.006629550829529762, + 0.0063614570535719395, + 0.0075402273796498775, + 0.004686168394982815, + 0.006329136900603771, + 0.006956734694540501, + 0.004131083842366934, + 0.008007185533642769, + 0.007271565496921539, + 0.007246044464409351, + 0.006330712698400021, + 0.0031647044233977795, + 0.010596673004329205, + 0.006679510697722435, + 0.004757391754537821, + 0.006842080038040876, + 0.010722232982516289, + 0.004858103115111589, + 0.006536717060953379, + 0.0035371151752769947, + 0.006299933884292841, + 0.005954378750175238, + 0.015932969748973846, + 0.01448842603713274, + 0.023301927372813225, + 0.014951254241168499, + 0.009011028334498405, + 0.009269945323467255, + 0.010018819011747837, + 0.011149872094392776, + 0.008408641442656517, + 0.012606045231223106, + 0.009594753384590149, + 0.0074279229156672955, + 0.0091156717389822, + 0.007732040714472532, + 0.010649225674569607, + 0.008468505926430225, + 0.04004376381635666, + 0.12347745895385742, + 0.050768546760082245, + 0.09622638672590256, + 0.052045561373233795, + 0.028809966519474983, + 0.04140811040997505, + 0.02978646196424961, + 0.005900668445974588, + 0.00838185753673315, + 0.009350478649139404, + 0.01847207546234131, + 0.006043093744665384, + 0.006347303278744221, + 0.006585794501006603, + 0.006205645855516195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.22557465732097626 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_12/index_tensor([5867, 5867])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_12/index_tensor([5867, 5867])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f29ab3a68f5dc737369bcdf3515a34d98542f045 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_12/index_tensor([5867, 5867])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static wide-angle shot captures a dark, rain-battered city street in the middle of a nighttime downpour. Rain pelted relentlessly against the slick black asphalt, water pouring from gutters and obscuring the dimly lit brick buildings in the distance. The only illumination came from the faint yellow-orange glow of a few streetlights, their light creating long, distorted reflections on the wet ground. Suddenly, high above, a massive, thick, multi-strand power line, strained taut under the fury of the storm, snapped with a violent, insinuating crack. The severed cable swooped downward, whipping erratically through the rain-laden air like a giant black snake. As it fell, a blinding burst of white sparks erupted from its frayed end, showering the street below in a chaotic, hissing explosion, starkly illuminating every raindrop. The cables struck the wet pavement, twisting and continuing to emit violent, hissing sparks that danced and hissed on contact with the water, enveloping the immediate area in a cloud of steam and creating an atmosphere of pure, sudden chaos and danger.", + "seed": 15657, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_12/index_5867_seed_15657_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03519970737397671, + "vlm_logprob": -0.03519970737397671, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.03519970737397671, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12243186682462692 + ], + "opposite_logprob_tokens": [ + -2.872431755065918 + ], + "logprob_reward": -0.12243186682462692, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.12243186682462692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13390392065048218, + 0.07653278112411499, + 0.08805307745933533, + 0.05805310979485512, + 0.09581196308135986, + 0.05440296232700348, + 0.10330162197351456, + 0.0681988075375557, + 0.2336292564868927, + 0.08811242878437042 + ], + "credit_map": [ + 0.012067181058228016, + 0.014845142140984535, + 0.012155688367784023, + 0.01629653014242649, + 0.017879584804177284, + 0.02385544963181019, + 0.015357364900410175, + 0.021446995437145233, + 0.009857967495918274, + 0.007078568916767836, + 0.008920920081436634, + 0.010700604878365993, + 0.012781516648828983, + 0.007393224630504847, + 0.008693428710103035, + 0.011106549762189388, + 0.011248211376369, + 0.008337841369211674, + 0.012338058091700077, + 0.013085877522826195, + 0.01058264821767807, + 0.010362714529037476, + 0.009655960835516453, + 0.012441776692867279, + 0.01153058372437954, + 0.006404528860002756, + 0.006860543508082628, + 0.006791906431317329, + 0.006701359059661627, + 0.006443653255701065, + 0.005469505209475756, + 0.007851031608879566, + 0.009041843935847282, + 0.010864874348044395, + 0.011037816293537617, + 0.01798783242702484, + 0.01897360011935234, + 0.00794217362999916, + 0.008588085882365704, + 0.01137574203312397, + 0.006863734219223261, + 0.0039949263446033, + 0.004509501159191132, + 0.007301126606762409, + 0.006955526769161224, + 0.011528479866683483, + 0.005147520452737808, + 0.00810215063393116, + 0.015657642856240273, + 0.010039118118584156, + 0.012320377863943577, + 0.011968333274126053, + 0.012411605566740036, + 0.010866215452551842, + 0.01671943999826908, + 0.013318891637027264, + 0.006097506266087294, + 0.0051223947666585445, + 0.006554919295012951, + 0.008389769122004509, + 0.00782789196819067, + 0.010437405668199062, + 0.014223016798496246, + 0.009545903652906418, + 0.02224264293909073, + 0.03227902576327324, + 0.018904155120253563, + 0.040963850915431976, + 0.017723262310028076, + 0.026499230414628983, + 0.05969037860631943, + 0.015326738357543945, + 0.010199165903031826, + 0.017284275963902473, + 0.011053254827857018, + 0.009656699374318123, + 0.010714178904891014, + 0.010208609513938427, + 0.008531303144991398, + 0.010464955121278763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.12243186682462692 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060764841735363 + ], + "opposite_logprob_tokens": [ + -3.8956077098846436 + ], + "logprob_reward": -0.02060764841735363, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060764841735363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3399867117404938, + 0.06797635555267334, + 0.02651199698448181, + 0.13883253931999207, + 0.02273087576031685, + 0.05934162810444832, + 0.06145250424742699, + 0.0571080818772316, + 0.057373132556676865, + 0.16868619620800018 + ], + "credit_map": [ + 0.025382766500115395, + 0.05960087478160858, + 0.036147590726614, + 0.04684467241168022, + 0.04019370675086975, + 0.03953440859913826, + 0.04814485087990761, + 0.04413782060146332, + 0.007471626624464989, + 0.00823600497096777, + 0.007758765481412411, + 0.008303279988467693, + 0.008782792836427689, + 0.0069512659683823586, + 0.011499036103487015, + 0.008973582647740841, + 0.00259711523540318, + 0.0027068168856203556, + 0.003066775156185031, + 0.0032757555600255728, + 0.004690541885793209, + 0.0024606387596577406, + 0.0041680349968373775, + 0.0035463208332657814, + 0.018124869093298912, + 0.024869121611118317, + 0.01079900749027729, + 0.016382578760385513, + 0.02446105144917965, + 0.013116489164531231, + 0.013862976804375648, + 0.01721644029021263, + 0.00313985045067966, + 0.002185889519751072, + 0.0027143042534589767, + 0.0048157451674342155, + 0.0029029883444309235, + 0.0019309904891997576, + 0.002387248445302248, + 0.0026538586243987083, + 0.005170449614524841, + 0.007783330976963043, + 0.0065774936228990555, + 0.013746106065809727, + 0.005027008708566427, + 0.006791429128497839, + 0.005790134891867638, + 0.008455673232674599, + 0.011297610588371754, + 0.006138751283288002, + 0.008056310936808586, + 0.008118992671370506, + 0.006327027454972267, + 0.00583758857101202, + 0.0076895784586668015, + 0.007986647076904774, + 0.006230611819773912, + 0.005369612481445074, + 0.00458728289231658, + 0.005060005933046341, + 0.009626750834286213, + 0.009868137538433075, + 0.006998645607382059, + 0.009367034770548344, + 0.004702415782958269, + 0.008387028239667416, + 0.004138562828302383, + 0.010246312245726585, + 0.005049117375165224, + 0.0060759130865335464, + 0.015177512541413307, + 0.0035962704569101334, + 0.01687472127377987, + 0.019577300176024437, + 0.02458258531987667, + 0.026289112865924835, + 0.019005874171853065, + 0.025333693251013756, + 0.02009720355272293, + 0.016925698146224022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.02060764841735363 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018209312111139297 + ], + "opposite_logprob_tokens": [ + -4.018209457397461 + ], + "logprob_reward": -0.018209312111139297, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018209312111139297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1513262391090393, + 0.04236108809709549, + 0.03773181512951851, + 0.04617326706647873, + 0.06823958456516266, + 0.19950930774211884, + 0.07427123934030533, + 0.040988557040691376, + 0.09095089882612228, + 0.24844804406166077 + ], + "credit_map": [ + 0.01612665317952633, + 0.020934773609042168, + 0.012162833474576473, + 0.015508430078625679, + 0.025511320680379868, + 0.020474927499890327, + 0.019249707460403442, + 0.021357577294111252, + 0.00471503846347332, + 0.0034320917911827564, + 0.005595835391432047, + 0.006248019635677338, + 0.004739794880151749, + 0.0057340869680047035, + 0.00634109228849411, + 0.005555122159421444, + 0.003825957654044032, + 0.006443766877055168, + 0.0037637415807694197, + 0.00630229851230979, + 0.005293598398566246, + 0.003412717953324318, + 0.003998722415417433, + 0.004691009409725666, + 0.007168720476329327, + 0.007185996975749731, + 0.004889638163149357, + 0.006172689143568277, + 0.006394918076694012, + 0.004061956889927387, + 0.004434959031641483, + 0.005864385515451431, + 0.007343604229390621, + 0.008051092736423016, + 0.00780640821903944, + 0.015090636909008026, + 0.010012313723564148, + 0.005609074141830206, + 0.006094655022025108, + 0.008231792598962784, + 0.01943490095436573, + 0.0170998964458704, + 0.016170037910342216, + 0.043392475694417953, + 0.01816501095890999, + 0.0389622300863266, + 0.015229793265461922, + 0.03105495497584343, + 0.01759922318160534, + 0.0061467234045267105, + 0.007360895164310932, + 0.006569665390998125, + 0.011640421114861965, + 0.008403957821428776, + 0.0062704430893063545, + 0.010279908776283264, + 0.0036346022970974445, + 0.003963563125580549, + 0.0044820355251431465, + 0.004827410448342562, + 0.005276148207485676, + 0.004904255270957947, + 0.009281175211071968, + 0.004619364161044359, + 0.01122052501887083, + 0.01638796366751194, + 0.010216032154858112, + 0.013417478650808334, + 0.007949340157210827, + 0.006318616680800915, + 0.017544521018862724, + 0.00789642333984375, + 0.024430057033896446, + 0.027844853699207306, + 0.03590494394302368, + 0.03325873240828514, + 0.0282213743776083, + 0.04490659758448601, + 0.02872752957046032, + 0.025153949856758118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.018209312111139297 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012547360733151436 + ], + "opposite_logprob_tokens": [ + -4.387547492980957 + ], + "logprob_reward": -0.012547360733151436, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012547360733151436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20022770762443542, + 0.05730435997247696, + 0.03266172483563423, + 0.04546381160616875, + 0.08070030808448792, + 0.3848682641983032, + 0.07162487506866455, + 0.01781802624464035, + 0.03606026992201805, + 0.07327070087194443 + ], + "credit_map": [ + 0.018620334565639496, + 0.03497457876801491, + 0.01907338947057724, + 0.027120675891637802, + 0.02424461953341961, + 0.024388855323195457, + 0.023940222337841988, + 0.02786504104733467, + 0.007923107594251633, + 0.006411293987184763, + 0.006862863432615995, + 0.0068355826660990715, + 0.006675533019006252, + 0.007479485124349594, + 0.007768395822495222, + 0.007348096929490566, + 0.005142257083207369, + 0.0037870146334171295, + 0.0035990842152386904, + 0.0037454096600413322, + 0.004796376917511225, + 0.0034883501939475536, + 0.004060580860823393, + 0.004042650572955608, + 0.00737457349896431, + 0.0058905743062496185, + 0.004583095200359821, + 0.005299943964928389, + 0.007134831044822931, + 0.0046242098324000835, + 0.004400113131850958, + 0.006156471092253923, + 0.015524069778621197, + 0.011222807690501213, + 0.007143325638025999, + 0.011888043023645878, + 0.008682183921337128, + 0.009323797188699245, + 0.006646097172051668, + 0.010269984602928162, + 0.03774963319301605, + 0.0358571819961071, + 0.02951808087527752, + 0.06819700449705124, + 0.03637772053480148, + 0.08203858882188797, + 0.031180812045931816, + 0.06394924968481064, + 0.015687279403209686, + 0.006068037357181311, + 0.005931415595114231, + 0.00552314193919301, + 0.011694605462253094, + 0.009493621997535229, + 0.007781906519085169, + 0.009444862604141235, + 0.001721917768009007, + 0.0022890111431479454, + 0.001805457635782659, + 0.0024911481887102127, + 0.0018651112914085388, + 0.0017063119448721409, + 0.003969167359173298, + 0.0019699004478752613, + 0.0044979373924434185, + 0.006899177562445402, + 0.004637475125491619, + 0.004147203639149666, + 0.0034186493139714003, + 0.003584971884265542, + 0.005469747819006443, + 0.003405105322599411, + 0.009079555980861187, + 0.010166879743337631, + 0.009927533566951752, + 0.011340446770191193, + 0.007478897925466299, + 0.009945852681994438, + 0.00747939758002758, + 0.007852140814065933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.012547360733151436 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022023487836122513 + ], + "opposite_logprob_tokens": [ + -6.12720251083374 + ], + "logprob_reward": -0.0022023487836122513, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022023487836122513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.252434641122818, + 0.1464453935623169, + 0.060584377497434616, + 0.05028587579727173, + 0.05486035346984863, + 0.09448534995317459, + 0.04251604527235031, + 0.08565984666347504, + 0.14017684757709503, + 0.07255126535892487 + ], + "credit_map": [ + 0.02207256481051445, + 0.04223308712244034, + 0.026646601036190987, + 0.04540371149778366, + 0.030322225764393806, + 0.026887429878115654, + 0.022845959290862083, + 0.03602305427193642, + 0.015240930020809174, + 0.01805769093334675, + 0.015277693048119545, + 0.027030732482671738, + 0.020155109465122223, + 0.012820407748222351, + 0.019147902727127075, + 0.01871492899954319, + 0.005307732615619898, + 0.0055449726060032845, + 0.007924976758658886, + 0.010774030350148678, + 0.0074074831791222095, + 0.006614530924707651, + 0.008696891367435455, + 0.008313761092722416, + 0.0038132413756102324, + 0.010562950745224953, + 0.004690892994403839, + 0.008061914704740047, + 0.0064843869768083096, + 0.005624732002615929, + 0.004836348816752434, + 0.006211409345269203, + 0.006836391054093838, + 0.006484720390290022, + 0.005768701899796724, + 0.007977630943059921, + 0.009666715748608112, + 0.005561964586377144, + 0.0055196816101670265, + 0.007044544909149408, + 0.009374750778079033, + 0.006664969027042389, + 0.00795305147767067, + 0.015313145704567432, + 0.011400054208934307, + 0.02129187248647213, + 0.007628182414919138, + 0.014859321527183056, + 0.01087307557463646, + 0.0034572144504636526, + 0.003557359566912055, + 0.003712168661877513, + 0.005425483454018831, + 0.005256453529000282, + 0.004691855050623417, + 0.005542434751987457, + 0.013986247591674328, + 0.007946780882775784, + 0.006671201437711716, + 0.007796549238264561, + 0.011274515651166439, + 0.017131146043539047, + 0.0072233788669109344, + 0.013630028814077377, + 0.014548344537615776, + 0.017508773133158684, + 0.0143659682944417, + 0.02488415688276291, + 0.012426796369254589, + 0.00975156482309103, + 0.036863647401332855, + 0.009827595204114914, + 0.009766319766640663, + 0.011340057477355003, + 0.010582894086837769, + 0.008247753605246544, + 0.007624171674251556, + 0.010660318657755852, + 0.006080080755054951, + 0.008249671198427677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0022023487836122513 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide-angle shot captures a dark, rain-battered city street in the middle of a nighttime downpour. Rain pelted relentlessly against the slick black asphalt, water pouring from gutters and obscuring the dimly lit brick buildings in the distance. The only illumination came from the faint yellow-orange glow of a few streetlights, their light creating long, distorted reflections on the wet ground. Suddenly, high above, a massive, thick, multi-strand power line, strained taut under the fury of the storm, snapped with a violent, insinuating crack. The severed cable swooped downward, whipping erratically through the rain-laden air like a giant black snake. As it fell, a blinding burst of white sparks erupted from its frayed end, showering the street below in a chaotic, hissing explosion, starkly illuminating every raindrop. The cables struck the wet pavement, twisting and continuing to emit violent, hissing sparks that danced and hissed on contact with the water, enveloping the immediate area in a cloud of steam and creating an atmosphere of pure, sudden chaos and danger.", + "seed": 15658, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_12/index_5867_seed_15658_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01258439882658422, + "vlm_logprob": -0.01258439882658422, + "vlm_yes_no_margin": 5.85, + "vlm_token_credit": -0.01258439882658422, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.050627999007701874 + ], + "opposite_logprob_tokens": [ + -4.925628185272217 + ], + "logprob_reward": -0.050627999007701874, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.050627999007701874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19689129292964935, + 0.07072633504867554, + 0.06364130228757858, + 0.09334474056959152, + 0.07611096650362015, + 0.04670519009232521, + 0.07617571949958801, + 0.10597201436758041, + 0.19445665180683136, + 0.07597579807043076 + ], + "credit_map": [ + 0.02438735030591488, + 0.01705317385494709, + 0.023899275809526443, + 0.03361101448535919, + 0.030131205916404724, + 0.026873189955949783, + 0.01773606613278389, + 0.02320002391934395, + 0.007588730659335852, + 0.006717845797538757, + 0.010934553109109402, + 0.014838919043540955, + 0.005117154214531183, + 0.005191876087337732, + 0.01326234545558691, + 0.007074910216033459, + 0.009117449633777142, + 0.007034897338598967, + 0.00815976969897747, + 0.014019917696714401, + 0.005272232461720705, + 0.00544324517250061, + 0.0061941202729940414, + 0.008399667218327522, + 0.008545839227735996, + 0.00909098144620657, + 0.008652515709400177, + 0.03312642499804497, + 0.012431085109710693, + 0.004400081001222134, + 0.004556862637400627, + 0.012540950439870358, + 0.0059736380353569984, + 0.00904675293713808, + 0.011135401204228401, + 0.02057092823088169, + 0.007997892796993256, + 0.004806687589734793, + 0.009083088487386703, + 0.0074965739622712135, + 0.0031917684245854616, + 0.005840712692588568, + 0.005676167085766792, + 0.007834802381694317, + 0.010639497078955173, + 0.0038684902247041464, + 0.005730802193284035, + 0.003922951407730579, + 0.0066802180372178555, + 0.01127864420413971, + 0.010930651798844337, + 0.01821008324623108, + 0.009373634122312069, + 0.0050294785760343075, + 0.008799409493803978, + 0.005873596295714378, + 0.0121378805488348, + 0.011158647947013378, + 0.018649498000741005, + 0.011947686783969402, + 0.012874310836195946, + 0.008309617638587952, + 0.015275883488357067, + 0.015618483535945415, + 0.012164285406470299, + 0.023322317749261856, + 0.017671892419457436, + 0.03399576619267464, + 0.026741014793515205, + 0.014966224320232868, + 0.04188433662056923, + 0.02371080406010151, + 0.005527914501726627, + 0.008111141622066498, + 0.008618628606200218, + 0.01435063872486353, + 0.013635281473398209, + 0.006514323875308037, + 0.013534813188016415, + 0.0056830584071576595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.050627999007701874 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007653674576431513 + ], + "opposite_logprob_tokens": [ + -4.882653713226318 + ], + "logprob_reward": -0.007653674576431513, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007653674576431513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20240479707717896, + 0.04769713431596756, + 0.2745349407196045, + 0.16918644309043884, + 0.06720975786447525, + 0.02889309637248516, + 0.031470734626054764, + 0.06470229476690292, + 0.0953368991613388, + 0.01856384053826332 + ], + "credit_map": [ + 0.015396052040159702, + 0.022771215066313744, + 0.02550366520881653, + 0.05081760883331299, + 0.021402433514595032, + 0.013518995605409145, + 0.03298923745751381, + 0.0200056042522192, + 0.004340217448771, + 0.005726171657443047, + 0.0063249277882277966, + 0.010317877866327763, + 0.0035734735429286957, + 0.004135099705308676, + 0.007916394621133804, + 0.005362975876778364, + 0.03330790996551514, + 0.022249950096011162, + 0.03093189001083374, + 0.04745899513363838, + 0.03632383048534393, + 0.021958842873573303, + 0.03828227147459984, + 0.044021282345056534, + 0.017789462581276894, + 0.012182260863482952, + 0.0183368232101202, + 0.03706932067871094, + 0.013101906515657902, + 0.008049839176237583, + 0.011402146890759468, + 0.05125468969345093, + 0.004013506695628166, + 0.006505989003926516, + 0.008985280990600586, + 0.020843341946601868, + 0.006783612072467804, + 0.0035439589992165565, + 0.006706557236611843, + 0.00982751976698637, + 0.0021305561531335115, + 0.0038710562512278557, + 0.0036584013141691685, + 0.005618716590106487, + 0.003477580612525344, + 0.002646881155669689, + 0.004445514176040888, + 0.0030443938449025154, + 0.002987443469464779, + 0.0033244166988879442, + 0.0037035192362964153, + 0.005371663253754377, + 0.005792426411062479, + 0.002856208011507988, + 0.004189709667116404, + 0.0032453557942062616, + 0.0059364791959524155, + 0.007609834428876638, + 0.008690170012414455, + 0.008761899545788765, + 0.007063941098749638, + 0.0056358883157372475, + 0.011136808432638645, + 0.009867277927696705, + 0.007408158853650093, + 0.011070537380874157, + 0.012171855196356773, + 0.016779808327555656, + 0.010228355415165424, + 0.007024628110229969, + 0.01637442409992218, + 0.014279141090810299, + 0.0018081977032124996, + 0.00168305030092597, + 0.0025966675020754337, + 0.0027033421210944653, + 0.0026429281570017338, + 0.0016487014945596457, + 0.002318786922842264, + 0.0031621691305190325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.007653674576431513 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019550274591892958 + ], + "opposite_logprob_tokens": [ + -6.251955032348633 + ], + "logprob_reward": -0.0019550274591892958, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019550274591892958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18434657156467438, + 0.06398608535528183, + 0.10426338762044907, + 0.05791082978248596, + 0.22155436873435974, + 0.045756611973047256, + 0.06217145919799805, + 0.08910728245973587, + 0.13139083981513977, + 0.039512574672698975 + ], + "credit_map": [ + 0.016457591205835342, + 0.02271581068634987, + 0.024970032274723053, + 0.03971359506249428, + 0.022399678826332092, + 0.010038354434072971, + 0.03160896897315979, + 0.016442537307739258, + 0.004797200672328472, + 0.00761779909953475, + 0.007620351389050484, + 0.017294995486736298, + 0.004715091083198786, + 0.003579440526664257, + 0.012095562182366848, + 0.006265648175030947, + 0.01187603734433651, + 0.00916843581944704, + 0.015131457708775997, + 0.019013693556189537, + 0.01219659112393856, + 0.007848459295928478, + 0.012420632876455784, + 0.01660807989537716, + 0.006873210426419973, + 0.010902777314186096, + 0.007826537825167179, + 0.010873458348214626, + 0.008287950418889523, + 0.004460754804313183, + 0.004696282558143139, + 0.0039898608811199665, + 0.015416035428643227, + 0.03698939085006714, + 0.03157500922679901, + 0.05166923999786377, + 0.03013337403535843, + 0.012747032567858696, + 0.019962849095463753, + 0.023061441257596016, + 0.0027086446061730385, + 0.004588959738612175, + 0.008811254985630512, + 0.015244248323142529, + 0.0046048532240092754, + 0.0022982116788625717, + 0.0038975472562015057, + 0.0036028895992785692, + 0.004483737051486969, + 0.006807299330830574, + 0.008742310106754303, + 0.01378125511109829, + 0.010819865390658379, + 0.004558471962809563, + 0.006631815806031227, + 0.006346703507006168, + 0.010031166486442089, + 0.009476840496063232, + 0.01605362631380558, + 0.01942851021885872, + 0.0087653249502182, + 0.006398121360689402, + 0.012752008624374866, + 0.0062016816809773445, + 0.009022116661071777, + 0.015498006716370583, + 0.01511368341743946, + 0.02493162825703621, + 0.015098518691956997, + 0.010136723518371582, + 0.02271457016468048, + 0.018875597044825554, + 0.0041982270777225494, + 0.0037358563859015703, + 0.005949108395725489, + 0.004677719436585903, + 0.0056840102188289165, + 0.005404534749686718, + 0.005669836886227131, + 0.0041932822205126286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.0019550274591892958 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001341753639280796 + ], + "opposite_logprob_tokens": [ + -6.626341819763184 + ], + "logprob_reward": -0.001341753639280796, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001341753639280796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062108300626277924, + 0.026922084391117096, + 0.09910410642623901, + 0.06511489301919937, + 0.055683258920907974, + 0.050611406564712524, + 0.07244192063808441, + 0.06634525209665298, + 0.4310477077960968, + 0.07062108814716339 + ], + "credit_map": [ + 0.005632075946778059, + 0.007391904015094042, + 0.008210552856326103, + 0.012421725317835808, + 0.0077775088138878345, + 0.004786644130945206, + 0.009891364723443985, + 0.005996527150273323, + 0.0027359987143427134, + 0.002887157490476966, + 0.004685335326939821, + 0.0055931080132722855, + 0.0019589599687606096, + 0.0019651157781481743, + 0.0035486475098878145, + 0.003547760657966137, + 0.013138661161065102, + 0.006810491438955069, + 0.009714175947010517, + 0.02114883065223694, + 0.010012584738433361, + 0.00766709353774786, + 0.007747464347630739, + 0.022864801809191704, + 0.006733752321451902, + 0.009831097908318043, + 0.0057542831636965275, + 0.012398689985275269, + 0.0064058322459459305, + 0.0033312104642391205, + 0.004061485640704632, + 0.016598546877503395, + 0.004185883328318596, + 0.010862383060157299, + 0.007098850794136524, + 0.011236763559281826, + 0.005815275479108095, + 0.0027051372453570366, + 0.00709181884303689, + 0.006687148939818144, + 0.0026696661952883005, + 0.00569236371666193, + 0.01240640226751566, + 0.016770554706454277, + 0.004226852674037218, + 0.0023633225355297327, + 0.003559038508683443, + 0.0029232094530016184, + 0.005058539565652609, + 0.008472559973597527, + 0.010187368839979172, + 0.019428854808211327, + 0.00724285002797842, + 0.004456277936697006, + 0.01046629250049591, + 0.007129175588488579, + 0.004425758961588144, + 0.008198181167244911, + 0.009310489520430565, + 0.016813598573207855, + 0.00762600265443325, + 0.0033441749401390553, + 0.011190488934516907, + 0.00543656200170517, + 0.0260966494679451, + 0.04966481775045395, + 0.045419108122587204, + 0.07096372544765472, + 0.059398673474788666, + 0.030822809785604477, + 0.09199803322553635, + 0.056683871895074844, + 0.00665651960298419, + 0.006975740194320679, + 0.006603787653148174, + 0.011614110320806503, + 0.014389030635356903, + 0.0061674281023442745, + 0.010643226094543934, + 0.007571246940642595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 6 + ], + "selected_reward": -0.001341753639280796 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013435394503176212 + ], + "opposite_logprob_tokens": [ + -6.626343727111816 + ], + "logprob_reward": -0.0013435394503176212, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013435394503176212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12097267806529999, + 0.03339608386158943, + 0.13846547901630402, + 0.09216944873332977, + 0.07943052053451538, + 0.07685039192438126, + 0.05002276226878166, + 0.08070016652345657, + 0.18289302289485931, + 0.1450994461774826 + ], + "credit_map": [ + 0.010826203972101212, + 0.013903777115046978, + 0.021197155117988586, + 0.022312456741929054, + 0.015841098502278328, + 0.008238736540079117, + 0.014538036659359932, + 0.014115207828581333, + 0.0039512705989181995, + 0.0032561386469751596, + 0.004389555659145117, + 0.0059011890552937984, + 0.004102736711502075, + 0.0031365137547254562, + 0.004638256039470434, + 0.004020423628389835, + 0.015941057354211807, + 0.009425906464457512, + 0.015497703105211258, + 0.0246108490973711, + 0.01729184202849865, + 0.009958809241652489, + 0.01630108803510666, + 0.029438218101859093, + 0.0131972786039114, + 0.010329755954444408, + 0.011542159132659435, + 0.019834181293845177, + 0.00752728758379817, + 0.006580555811524391, + 0.004789301194250584, + 0.018368929624557495, + 0.005503361579030752, + 0.010919959284365177, + 0.010149678215384483, + 0.023645637556910515, + 0.005844613071531057, + 0.003413411322981119, + 0.006251074839383364, + 0.013702784664928913, + 0.004410959780216217, + 0.006065527908504009, + 0.024369094520807266, + 0.020555680617690086, + 0.006261562928557396, + 0.003043671138584614, + 0.004952579271048307, + 0.007191314361989498, + 0.005409155040979385, + 0.005058948881924152, + 0.0074776541441679, + 0.009989679791033268, + 0.006289335899055004, + 0.004799088463187218, + 0.0052406685426831245, + 0.005758232902735472, + 0.011069043539464474, + 0.008180953562259674, + 0.013475356623530388, + 0.012203115969896317, + 0.008518650196492672, + 0.009297280572354794, + 0.009464610368013382, + 0.008491151966154575, + 0.015269912779331207, + 0.01903821900486946, + 0.02084171399474144, + 0.034001775085926056, + 0.020462412387132645, + 0.014332924038171768, + 0.021722467616200447, + 0.03722357377409935, + 0.013987116515636444, + 0.014812605455517769, + 0.016098732128739357, + 0.016563929617404938, + 0.029213419184088707, + 0.014992951415479183, + 0.018721848726272583, + 0.0207088403403759 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.0013435394503176212 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide-angle shot captures a dark, rain-battered city street in the middle of a nighttime downpour. Rain pelted relentlessly against the slick black asphalt, water pouring from gutters and obscuring the dimly lit brick buildings in the distance. The only illumination came from the faint yellow-orange glow of a few streetlights, their light creating long, distorted reflections on the wet ground. Suddenly, high above, a massive, thick, multi-strand power line, strained taut under the fury of the storm, snapped with a violent, insinuating crack. The severed cable swooped downward, whipping erratically through the rain-laden air like a giant black snake. As it fell, a blinding burst of white sparks erupted from its frayed end, showering the street below in a chaotic, hissing explosion, starkly illuminating every raindrop. The cables struck the wet pavement, twisting and continuing to emit violent, hissing sparks that danced and hissed on contact with the water, enveloping the immediate area in a cloud of steam and creating an atmosphere of pure, sudden chaos and danger.", + "seed": 15659, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_12/index_5867_seed_15659_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017350632697343826, + "vlm_logprob": -0.017350632697343826, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.017350632697343826, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04479408264160156 + ], + "opposite_logprob_tokens": [ + -5.044794082641602 + ], + "logprob_reward": -0.04479408264160156, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.04479408264160156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12496669590473175, + 0.14701922237873077, + 0.14058853685855865, + 0.059675268828868866, + 0.09324301034212112, + 0.1458817571401596, + 0.05434488132596016, + 0.043339408934116364, + 0.09505779296159744, + 0.09588348865509033 + ], + "credit_map": [ + 0.013372029177844524, + 0.018865525722503662, + 0.022096052765846252, + 0.017087463289499283, + 0.013406140729784966, + 0.012414742261171341, + 0.014062422327697277, + 0.0136623065918684, + 0.0203108973801136, + 0.03634168580174446, + 0.03531708940863609, + 0.012622108682990074, + 0.010215149261057377, + 0.0076299807988107204, + 0.009729393757879734, + 0.014852898195385933, + 0.013647674582898617, + 0.012276073917746544, + 0.04977370426058769, + 0.008317259140312672, + 0.005347778555005789, + 0.009388132952153683, + 0.007180084940046072, + 0.03465782478451729, + 0.006388038396835327, + 0.008301577530801296, + 0.008369985967874527, + 0.015070383436977863, + 0.005735739599913359, + 0.0030993467662483454, + 0.006124850362539291, + 0.006585346534848213, + 0.014151368290185928, + 0.0187491737306118, + 0.012148749083280563, + 0.013052556663751602, + 0.010702742263674736, + 0.007064151577651501, + 0.00801094900816679, + 0.00936331320554018, + 0.015501133166253567, + 0.016308987513184547, + 0.024985909461975098, + 0.01143396832048893, + 0.017984116449952126, + 0.023131616413593292, + 0.013108274899423122, + 0.023427734151482582, + 0.009471605531871319, + 0.009822959080338478, + 0.007813019677996635, + 0.00643530860543251, + 0.00613556569442153, + 0.005733672529459, + 0.004109196364879608, + 0.004823545925319195, + 0.005380677524954081, + 0.005697645712643862, + 0.0037403632886707783, + 0.00429429579526186, + 0.006163451354950666, + 0.004027597606182098, + 0.00546403881162405, + 0.008571334183216095, + 0.009928428567945957, + 0.012392877601087093, + 0.01022148597985506, + 0.014504285529255867, + 0.011566883884370327, + 0.011410422623157501, + 0.013752518221735954, + 0.011280888691544533, + 0.009108519181609154, + 0.00972501840442419, + 0.018416697159409523, + 0.019004257395863533, + 0.011332442052662373, + 0.008179593831300735, + 0.011362737976014614, + 0.008754206821322441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.04479408264160156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014219745993614197 + ], + "opposite_logprob_tokens": [ + -4.264219760894775 + ], + "logprob_reward": -0.014219745993614197, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014219745993614197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2008522003889084, + 0.09377408027648926, + 0.029552070423960686, + 0.17789924144744873, + 0.04489366337656975, + 0.05696640536189079, + 0.0544445738196373, + 0.18286845088005066, + 0.12079360336065292, + 0.037955645471811295 + ], + "credit_map": [ + 0.02922338806092739, + 0.051051534712314606, + 0.039939574897289276, + 0.019160334020853043, + 0.020686689764261246, + 0.011004588566720486, + 0.01459486410021782, + 0.015191218815743923, + 0.015340032055974007, + 0.029714249074459076, + 0.015994900837540627, + 0.00887239258736372, + 0.00795047264546156, + 0.0038074785843491554, + 0.0052557094022631645, + 0.0068388432264328, + 0.0033133309334516525, + 0.003015699330717325, + 0.005682071205228567, + 0.005040972493588924, + 0.0021370321046561003, + 0.0024064190220087767, + 0.0031103105284273624, + 0.004846236202865839, + 0.0217715073376894, + 0.03419885039329529, + 0.022940481081604958, + 0.030314797535538673, + 0.01586984284222126, + 0.014208625070750713, + 0.016191281378269196, + 0.02240385301411152, + 0.00611730944365263, + 0.006855276878923178, + 0.007354018744081259, + 0.006196367088705301, + 0.0049236128106713295, + 0.0036860767286270857, + 0.004516166169196367, + 0.005244835279881954, + 0.0062066237442195415, + 0.006576928310096264, + 0.00960464682430029, + 0.00549893407151103, + 0.006195319350808859, + 0.0072512440383434296, + 0.005889880936592817, + 0.009742824360728264, + 0.010832552798092365, + 0.00892883725464344, + 0.008215757086873055, + 0.004420093260705471, + 0.005090234335511923, + 0.005403644405305386, + 0.004551378078758717, + 0.007002071477472782, + 0.015006497502326965, + 0.01286600437015295, + 0.035017672926187515, + 0.032055869698524475, + 0.015602448023855686, + 0.016109811142086983, + 0.03351215645670891, + 0.022697992622852325, + 0.01519707590341568, + 0.01810261979699135, + 0.017691684886813164, + 0.012930981814861298, + 0.014704006724059582, + 0.009401755407452583, + 0.017462588846683502, + 0.015302903950214386, + 0.004043908789753914, + 0.005901908501982689, + 0.00886260811239481, + 0.006793919485062361, + 0.0031784980092197657, + 0.00284485868178308, + 0.0033027250319719315, + 0.0030272207222878933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.014219745993614197 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005962797906249762 + ], + "opposite_logprob_tokens": [ + -5.13096284866333 + ], + "logprob_reward": -0.005962797906249762, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005962797906249762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03825777396559715, + 0.04078076407313347, + 0.09751209616661072, + 0.4533238410949707, + 0.024827752262353897, + 0.11001738160848618, + 0.021812986582517624, + 0.07065393775701523, + 0.11313164234161377, + 0.029681820422410965 + ], + "credit_map": [ + 0.006131534930318594, + 0.005592717789113522, + 0.005868740379810333, + 0.004607212729752064, + 0.00407020328566432, + 0.0032479632645845413, + 0.004776014480739832, + 0.003963387571275234, + 0.004123321734368801, + 0.007370030973106623, + 0.0121390912681818, + 0.003647663164883852, + 0.0033771542366594076, + 0.0028620895463973284, + 0.002829938428476453, + 0.004431473091244698, + 0.009973992593586445, + 0.009354758076369762, + 0.029925735667347908, + 0.008287045173346996, + 0.006152761168777943, + 0.006436878349632025, + 0.007233370561152697, + 0.020147548988461494, + 0.05436249449849129, + 0.12233998626470566, + 0.1135576143860817, + 0.0555211640894413, + 0.024046216160058975, + 0.02055922895669937, + 0.0305262990295887, + 0.03241082653403282, + 0.0033064703457057476, + 0.004197778180241585, + 0.0034614629112184048, + 0.003751327283680439, + 0.0024267423432320356, + 0.003018402960151434, + 0.0017849020659923553, + 0.0028806666377931833, + 0.00979816447943449, + 0.015589169226586819, + 0.04066402092576027, + 0.005156753584742546, + 0.004772207234054804, + 0.007186070084571838, + 0.004300485830754042, + 0.022550512105226517, + 0.0031703240238130093, + 0.0030130294617265463, + 0.005971432197839022, + 0.001827574917115271, + 0.001688756630755961, + 0.001626469660550356, + 0.0017934920033439994, + 0.0027219068724662066, + 0.009008213877677917, + 0.008661702275276184, + 0.014429439790546894, + 0.008393267169594765, + 0.007941320538520813, + 0.004901727195829153, + 0.008118126541376114, + 0.00920014176517725, + 0.012852958403527737, + 0.016200222074985504, + 0.014351382851600647, + 0.01562424935400486, + 0.013847882859408855, + 0.007945776917040348, + 0.01710309460759163, + 0.015206069685518742, + 0.002919726772233844, + 0.0035619514528661966, + 0.006232451647520065, + 0.005681274924427271, + 0.003419922199100256, + 0.0031640154775232077, + 0.002497775247320533, + 0.002204702701419592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.005962797906249762 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588964223861694 + ], + "opposite_logprob_tokens": [ + -3.8955888748168945 + ], + "logprob_reward": -0.020588964223861694, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588964223861694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0850488543510437, + 0.07586871832609177, + 0.04467630013823509, + 0.1386227011680603, + 0.04030138626694679, + 0.04285480082035065, + 0.018433617427945137, + 0.047227486968040466, + 0.3695371448993683, + 0.13742901384830475 + ], + "credit_map": [ + 0.010229760780930519, + 0.014591766521334648, + 0.016645647585392, + 0.011273656971752644, + 0.009600387886166573, + 0.005007818806916475, + 0.0072168149054050446, + 0.01048299577087164, + 0.010726053267717361, + 0.016419822350144386, + 0.01531327422708273, + 0.005983314942568541, + 0.005903656594455242, + 0.004807785153388977, + 0.008027613162994385, + 0.008687205612659454, + 0.004910693969577551, + 0.0044039348140358925, + 0.01263254415243864, + 0.005808488465845585, + 0.003730345983058214, + 0.0027274806052446365, + 0.003125532763078809, + 0.007337280083447695, + 0.018463410437107086, + 0.034702565521001816, + 0.026188118383288383, + 0.016950229182839394, + 0.00879743229597807, + 0.007109935395419598, + 0.011293679475784302, + 0.015117335133254528, + 0.005741878412663937, + 0.009418591856956482, + 0.006430921144783497, + 0.006071349140256643, + 0.003370730672031641, + 0.0022152974270284176, + 0.0031504614744335413, + 0.0039021531119942665, + 0.005506483837962151, + 0.008363429456949234, + 0.007273564580827951, + 0.005585360340774059, + 0.00474383682012558, + 0.003623237367719412, + 0.0034585357643663883, + 0.00430035125464201, + 0.0026243976317346096, + 0.002647336572408676, + 0.0038847068790346384, + 0.0019397379364818335, + 0.0013309033820405602, + 0.0025115974713116884, + 0.001447930815629661, + 0.0020470053423196077, + 0.005561823025345802, + 0.006495763547718525, + 0.007145448587834835, + 0.006172968074679375, + 0.005214175209403038, + 0.0034401435405015945, + 0.006739775650203228, + 0.006457393057644367, + 0.04247504472732544, + 0.05140424147248268, + 0.05544817075133324, + 0.05303514003753662, + 0.04080592840909958, + 0.022067513316869736, + 0.056868553161621094, + 0.0474325530230999, + 0.014288944192230701, + 0.015443178825080395, + 0.025663800537586212, + 0.02429966628551483, + 0.015828752890229225, + 0.015139556489884853, + 0.014092040248215199, + 0.012673069722950459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.020588964223861694 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011875727213919163 + ], + "opposite_logprob_tokens": [ + -6.751187801361084 + ], + "logprob_reward": -0.0011875727213919163, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011875727213919163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10279199481010437, + 0.05833125859498978, + 0.07418512552976608, + 0.28994742035865784, + 0.09436630457639694, + 0.13543054461479187, + 0.021557211875915527, + 0.05545666813850403, + 0.1148390993475914, + 0.05309438705444336 + ], + "credit_map": [ + 0.010032907128334045, + 0.014594734646379948, + 0.020406220108270645, + 0.01390150934457779, + 0.01199948601424694, + 0.007964564487338066, + 0.015114790759980679, + 0.008777783252298832, + 0.006632318254560232, + 0.012978694401681423, + 0.010703178122639656, + 0.0066751413978636265, + 0.006663944106549025, + 0.0041463035158813, + 0.00619393028318882, + 0.004337747115641832, + 0.00841827318072319, + 0.00658140005543828, + 0.023071568459272385, + 0.006755302660167217, + 0.004228739067912102, + 0.00497045973315835, + 0.005453826393932104, + 0.014705559238791466, + 0.03531729802489281, + 0.07515028864145279, + 0.06645320355892181, + 0.03396332636475563, + 0.018994104117155075, + 0.016092289239168167, + 0.019562454894185066, + 0.02441444806754589, + 0.01143934577703476, + 0.016834083944559097, + 0.018702242523431778, + 0.011138824746012688, + 0.007483166642487049, + 0.012491879984736443, + 0.006176331080496311, + 0.01010042428970337, + 0.013087330386042595, + 0.017951814457774162, + 0.05013614892959595, + 0.006463476922363043, + 0.005960636306554079, + 0.009446129202842712, + 0.004197393544018269, + 0.028187615796923637, + 0.003078571753576398, + 0.003805907443165779, + 0.0036210170947015285, + 0.0017519640969112515, + 0.0027059235144406557, + 0.002462010830640793, + 0.0023003413807600737, + 0.0018314752960577607, + 0.007651359308511019, + 0.006691133603453636, + 0.0064733135513961315, + 0.008678359910845757, + 0.006906905211508274, + 0.005734892096370459, + 0.006525967735797167, + 0.006794734392315149, + 0.008357075043022633, + 0.01005315687507391, + 0.020849695429205894, + 0.015911588445305824, + 0.010182201862335205, + 0.00914772879332304, + 0.022538185119628906, + 0.01779947057366371, + 0.006477766670286655, + 0.005996609106659889, + 0.007330742198973894, + 0.009302805177867413, + 0.006303131580352783, + 0.0047906311228871346, + 0.007278114557266235, + 0.0056145875714719296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.0011875727213919163 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static wide-angle shot captures a dark, rain-battered city street in the middle of a nighttime downpour. Rain pelted relentlessly against the slick black asphalt, water pouring from gutters and obscuring the dimly lit brick buildings in the distance. The only illumination came from the faint yellow-orange glow of a few streetlights, their light creating long, distorted reflections on the wet ground. Suddenly, high above, a massive, thick, multi-strand power line, strained taut under the fury of the storm, snapped with a violent, insinuating crack. The severed cable swooped downward, whipping erratically through the rain-laden air like a giant black snake. As it fell, a blinding burst of white sparks erupted from its frayed end, showering the street below in a chaotic, hissing explosion, starkly illuminating every raindrop. The cables struck the wet pavement, twisting and continuing to emit violent, hissing sparks that danced and hissed on contact with the water, enveloping the immediate area in a cloud of steam and creating an atmosphere of pure, sudden chaos and danger.", + "seed": 15660, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_12/index_5867_seed_15660_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013653357257135212, + "vlm_logprob": -0.013653357257135212, + "vlm_yes_no_margin": 5.475, + "vlm_token_credit": -0.013653357257135212, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04907935857772827 + ], + "opposite_logprob_tokens": [ + -4.549079418182373 + ], + "logprob_reward": -0.04907935857772827, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.04907935857772827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16326983273029327, + 0.04399403557181358, + 0.03169231116771698, + 0.09119351953268051, + 0.030159560963511467, + 0.023071054369211197, + 0.07249336689710617, + 0.1516670137643814, + 0.19228070974349976, + 0.20017856359481812 + ], + "credit_map": [ + 0.021429454907774925, + 0.016370490193367004, + 0.02301286906003952, + 0.016756389290094376, + 0.028239523991942406, + 0.020858338102698326, + 0.019620735198259354, + 0.01698203943669796, + 0.006264244206249714, + 0.004259467590600252, + 0.005724051035940647, + 0.005369917023926973, + 0.006722378544509411, + 0.005954220425337553, + 0.004667514003813267, + 0.005032241810113192, + 0.006456969305872917, + 0.003344503464177251, + 0.003620687872171402, + 0.0028047694358974695, + 0.004157263785600662, + 0.004424809478223324, + 0.00347284902818501, + 0.003410461125895381, + 0.007250368595123291, + 0.007754203863441944, + 0.008372757583856583, + 0.006180073134601116, + 0.029057934880256653, + 0.015196125023066998, + 0.009161652997136116, + 0.008220408111810684, + 0.002668331377208233, + 0.0030007215682417154, + 0.0030477310065180063, + 0.0035994581412523985, + 0.006561530288308859, + 0.0039482745341956615, + 0.004343454260379076, + 0.0029900611843913794, + 0.0024460172280669212, + 0.0025011920370161533, + 0.0021792312618345022, + 0.0021299649961292744, + 0.005359561648219824, + 0.0035618841648101807, + 0.0026447896379977465, + 0.0022484154906123877, + 0.007047306280583143, + 0.010406863875687122, + 0.007476691156625748, + 0.006021592300385237, + 0.013675362803041935, + 0.010250268504023552, + 0.008169146254658699, + 0.009446133859455585, + 0.011199606582522392, + 0.01188522670418024, + 0.015184732154011726, + 0.02355624921619892, + 0.023592226207256317, + 0.03281582146883011, + 0.02153763361275196, + 0.011895513162016869, + 0.014326754957437515, + 0.027104340493679047, + 0.018732700496912003, + 0.0694705918431282, + 0.008432784117758274, + 0.008386267349123955, + 0.021541133522987366, + 0.024286141619086266, + 0.015288242138922215, + 0.06771751493215561, + 0.011938490904867649, + 0.016827654093503952, + 0.030356958508491516, + 0.013182158581912518, + 0.011221137829124928, + 0.03364642336964607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.04907935857772827 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00866522267460823 + ], + "opposite_logprob_tokens": [ + -4.758665084838867 + ], + "logprob_reward": -0.00866522267460823, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00866522267460823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19632598757743835, + 0.023604074493050575, + 0.07922012358903885, + 0.13667480647563934, + 0.019769877195358276, + 0.04096297174692154, + 0.040750086307525635, + 0.26522666215896606, + 0.12748447060585022, + 0.06998095661401749 + ], + "credit_map": [ + 0.019567709416151047, + 0.022129220888018608, + 0.022979354485869408, + 0.025035910308361053, + 0.033995673060417175, + 0.025640027597546577, + 0.02788618579506874, + 0.01909192092716694, + 0.0037841512821614742, + 0.0036330290604382753, + 0.002864586189389229, + 0.0028926783706992865, + 0.003389952704310417, + 0.0027530412189662457, + 0.0023267443757504225, + 0.001959893386811018, + 0.008439308032393456, + 0.007261344231665134, + 0.009247386828064919, + 0.008876060135662556, + 0.014812604524195194, + 0.012036456726491451, + 0.009409138932824135, + 0.009137830697000027, + 0.008958997204899788, + 0.010768608190119267, + 0.012057850137352943, + 0.011930362321436405, + 0.046389780938625336, + 0.02275264263153076, + 0.011141657829284668, + 0.012674933299422264, + 0.0021899889688938856, + 0.002016108948737383, + 0.0018483188468962908, + 0.001804800471290946, + 0.003238627454265952, + 0.003058818867430091, + 0.0032563426066190004, + 0.0023568731267005205, + 0.004835788626223803, + 0.0045814006589353085, + 0.004779089707881212, + 0.0036777148488909006, + 0.009545923210680485, + 0.0063350629061460495, + 0.0038156546652317047, + 0.0033923417795449495, + 0.004476638976484537, + 0.0037995227612555027, + 0.0035364488139748573, + 0.003184233559295535, + 0.007770922034978867, + 0.0059614283964037895, + 0.006286158226430416, + 0.005734738428145647, + 0.01074235513806343, + 0.019144553691148758, + 0.007766354829072952, + 0.010991290211677551, + 0.10943373292684555, + 0.08393970876932144, + 0.012709343805909157, + 0.010499320924282074, + 0.0051064640283584595, + 0.006982315331697464, + 0.010731088928878307, + 0.04505440965294838, + 0.00622306764125824, + 0.005220136139541864, + 0.040993910282850266, + 0.0071730827912688255, + 0.005956709384918213, + 0.020859623327851295, + 0.007140247616916895, + 0.007388935890048742, + 0.01016576774418354, + 0.005906134378165007, + 0.004703403450548649, + 0.00786014273762703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.00866522267460823 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076457480899989605 + ], + "opposite_logprob_tokens": [ + -4.882645606994629 + ], + "logprob_reward": -0.0076457480899989605, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076457480899989605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08852135390043259, + 0.03569966182112694, + 0.024670200422406197, + 0.24128441512584686, + 0.03289347514510155, + 0.04351166635751724, + 0.19201786816120148, + 0.13998226821422577, + 0.15614022314548492, + 0.04527885094285011 + ], + "credit_map": [ + 0.011740831658244133, + 0.008711672388017178, + 0.010642929002642632, + 0.010531767271459103, + 0.012660758569836617, + 0.015510411001741886, + 0.00876085739582777, + 0.00996212288737297, + 0.005311299115419388, + 0.0037221878301352262, + 0.004141486249864101, + 0.004628890193998814, + 0.005227630026638508, + 0.004567119292914867, + 0.004041358362883329, + 0.004059690982103348, + 0.002184186829254031, + 0.0019764506723731756, + 0.002866541501134634, + 0.002607254311442375, + 0.005402275361120701, + 0.0032082239631563425, + 0.0038243799936026335, + 0.0026008873246610165, + 0.0179270189255476, + 0.019502408802509308, + 0.023156465962529182, + 0.01808379963040352, + 0.07705120742321014, + 0.04112907126545906, + 0.02216274105012417, + 0.022271718829870224, + 0.002427843399345875, + 0.0031293295323848724, + 0.0037154920864850283, + 0.004536802414804697, + 0.003980670589953661, + 0.004198275040835142, + 0.006875926163047552, + 0.0040291366167366505, + 0.004870782606303692, + 0.006425393745303154, + 0.004574882797896862, + 0.0033952370285987854, + 0.009098636917769909, + 0.006588518153876066, + 0.004006000701338053, + 0.0045522162690758705, + 0.012107836082577705, + 0.02203403227031231, + 0.021083109080791473, + 0.017870264127850533, + 0.03390248864889145, + 0.026283061131834984, + 0.023264016956090927, + 0.0354730449616909, + 0.009201784618198872, + 0.013211722485721111, + 0.007187797222286463, + 0.012549866922199726, + 0.04960634559392929, + 0.026010725647211075, + 0.011269253678619862, + 0.01094476692378521, + 0.010291091166436672, + 0.021693669259548187, + 0.01791955530643463, + 0.036733921617269516, + 0.011841251514852047, + 0.010099993087351322, + 0.033201418817043304, + 0.014359336346387863, + 0.004181073047220707, + 0.00961459893733263, + 0.0045894538052380085, + 0.003715564962476492, + 0.006753937341272831, + 0.0036125120241194963, + 0.0043990337289869785, + 0.008412675000727177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.0076457480899989605 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019464610377326608 + ], + "opposite_logprob_tokens": [ + -6.251946449279785 + ], + "logprob_reward": -0.0019464610377326608, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019464610377326608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05387461185455322, + 0.03744469955563545, + 0.03707630932331085, + 0.057742223143577576, + 0.013761164620518684, + 0.026273753494024277, + 0.11377006024122238, + 0.40541765093803406, + 0.149267315864563, + 0.10537219047546387 + ], + "credit_map": [ + 0.005484197288751602, + 0.004679761361330748, + 0.005289064720273018, + 0.006187976337969303, + 0.011038973927497864, + 0.008117745630443096, + 0.007495708763599396, + 0.00558118149638176, + 0.00419632438570261, + 0.0030516907572746277, + 0.0042551299557089806, + 0.004464688245207071, + 0.008104837499558926, + 0.0052163624204695225, + 0.005128144286572933, + 0.0030275217723101377, + 0.0037354612722992897, + 0.0036254776641726494, + 0.005897022783756256, + 0.004569600336253643, + 0.0055048163048923016, + 0.0045647332444787025, + 0.005099697038531303, + 0.004079500678926706, + 0.004282581154257059, + 0.004907554946839809, + 0.005526429507881403, + 0.004520615562796593, + 0.01851249486207962, + 0.009565992280840874, + 0.00551254628226161, + 0.004914009012281895, + 0.0009239109931513667, + 0.0012956751743331552, + 0.0018741153180599213, + 0.001998097402974963, + 0.0016844685887917876, + 0.001621748204343021, + 0.0028130279388278723, + 0.001550121232867241, + 0.003229867434129119, + 0.0032829726114869118, + 0.0032426610123366117, + 0.003148398594930768, + 0.0036697678733617067, + 0.0037179719656705856, + 0.003016718430444598, + 0.0029653955716639757, + 0.006702028680592775, + 0.009074579924345016, + 0.011299786157906055, + 0.006744903512299061, + 0.015746373683214188, + 0.01666777767241001, + 0.014241890050470829, + 0.033292725682258606, + 0.015855981037020683, + 0.035706523805856705, + 0.008734813891351223, + 0.012512474320828915, + 0.18072617053985596, + 0.12205487489700317, + 0.01892317458987236, + 0.010903610847890377, + 0.015445731580257416, + 0.03253816068172455, + 0.019045637920498848, + 0.015252870507538319, + 0.016279060393571854, + 0.012659617699682713, + 0.016016779467463493, + 0.02202945575118065, + 0.007904375903308392, + 0.03858686611056328, + 0.006647748406976461, + 0.011485068127512932, + 0.01774689555168152, + 0.0065764677710831165, + 0.006270848214626312, + 0.010153922252357006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.0019464610377326608 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot a static wide-angle view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a dark city street at night during a heavy rainstorm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a massive multi-strand power line snap and fall from above like a black snake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the fallen cables emit violent sparks and hiss upon striking the wet pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by the faint yellow-orange glow of streetlights and the white sparks from the cable?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009299959056079388 + ], + "opposite_logprob_tokens": [ + -7.000929832458496 + ], + "logprob_reward": -0.0009299959056079388, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009299959056079388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06592298299074173, + 0.01960194855928421, + 0.023072153329849243, + 0.08677925169467926, + 0.02770945243537426, + 0.04430991783738136, + 0.052612509578466415, + 0.5010052919387817, + 0.0782313272356987, + 0.10075516253709793 + ], + "credit_map": [ + 0.00709545100107789, + 0.006770813371986151, + 0.007489392533898354, + 0.008510095067322254, + 0.009906487539410591, + 0.008446972817182541, + 0.011645863763988018, + 0.006057901307940483, + 0.002987453481182456, + 0.0016758994897827506, + 0.002296332037076354, + 0.0018830784829333425, + 0.003287872299551964, + 0.0028300033882260323, + 0.0025890516117215157, + 0.00205225870013237, + 0.0029088431037962437, + 0.0024934757966548204, + 0.002289686817675829, + 0.0021417534444481134, + 0.004637240897864103, + 0.0032675424590706825, + 0.0026260213926434517, + 0.002707588719204068, + 0.005997091066092253, + 0.0072080097161233425, + 0.008140519261360168, + 0.00719171529635787, + 0.02770496904850006, + 0.014178005047142506, + 0.007788644637912512, + 0.008570301346480846, + 0.0034810116048902273, + 0.0025712328497320414, + 0.0024808617308735847, + 0.002431550296023488, + 0.005585847422480583, + 0.004366546869277954, + 0.003481128718703985, + 0.0033112724777311087, + 0.0029362402856349945, + 0.004483929369598627, + 0.0034680445678532124, + 0.0035207229666411877, + 0.01303386315703392, + 0.007482183165848255, + 0.0054343524388968945, + 0.003950580954551697, + 0.0039970772340893745, + 0.006056064274162054, + 0.0058715688064694405, + 0.005293811671435833, + 0.008422433398663998, + 0.007529432885348797, + 0.007020730059593916, + 0.008421392180025578, + 0.02086608111858368, + 0.041472192853689194, + 0.012196088209748268, + 0.014580199494957924, + 0.21542426943778992, + 0.16213358938694, + 0.018608156591653824, + 0.015724677592515945, + 0.005212744232267141, + 0.01566796749830246, + 0.010385533794760704, + 0.014279519207775593, + 0.008173823356628418, + 0.005473260767757893, + 0.010488438419997692, + 0.008550044149160385, + 0.008016680367290974, + 0.03764894977211952, + 0.005523009225726128, + 0.01097161415964365, + 0.01643536612391472, + 0.0057251546531915665, + 0.004856536164879799, + 0.011577850207686424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 3 + ], + "selected_reward": -0.0009299959056079388 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_13/index_tensor([6, 6])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_13/index_tensor([6, 6])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..45b1e7efaf2e24610047822d28888da95a6626d7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_13/index_tensor([6, 6])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle, wide-angle lens slowly and gracefully sweeps across a classic Chinese landscape, evoking a garden of folly and ignorance, bathed in the warm, nostalgic light of the afternoon. The water of a meticulously designed pond flows so gently and slowly, almost imperceptibly, its surface like polished glass, perfectly reflecting the golden glow of time and the dark, elegant silhouette of an ornate wooden pavilion with its weathered, tilted eaves. A fallen crimson maple leaf floats lazily on the water's surface, its delicate veins visible and its edges slightly curled. As the camera follows its leisurely journey, it passes a rickety stone bridge, its gray slate smoothed by centuries of footsteps. Beneath the tranquil surface, several enormous golden and white mallows swim lazily, their shimmering scales catching the submerged light as they glide in a sinuous, unhurried pattern. The entire scene, suffused with deep greens, warm golds, and earthy browns, is imbued with a sense of peaceful history, creating a visual of profoundly satisfying and tranquil meditation.", + "seed": 25485, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_13/index_6_seed_25485_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.26825504675507544, + "vlm_logprob": -0.26825504675507544, + "vlm_yes_no_margin": 2.175, + "vlm_token_credit": -0.26825504675507544, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.047775644809007645 + ], + "opposite_logprob_tokens": [ + -3.5477757453918457 + ], + "logprob_reward": -0.047775644809007645, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.047775644809007645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08610863238573074, + 0.11887290328741074, + 0.17634300887584686, + 0.16034896671772003, + 0.17540034651756287, + 0.12727421522140503, + 0.02769375406205654, + 0.020609840750694275, + 0.0744178518652916, + 0.03293047472834587 + ], + "credit_map": [ + 0.0057005067355930805, + 0.014444668777287006, + 0.00859738327562809, + 0.011602822691202164, + 0.0127096651121974, + 0.011757375672459602, + 0.01334466878324747, + 0.007951535284519196, + 0.01440407894551754, + 0.007885928265750408, + 0.030747566372156143, + 0.012456397525966167, + 0.012083890847861767, + 0.01842055842280388, + 0.013214179314672947, + 0.00966029055416584, + 0.012260965071618557, + 0.06026434525847435, + 0.01013237051665783, + 0.008621560409665108, + 0.03005600906908512, + 0.013120374642312527, + 0.007485848385840654, + 0.03440151363611221, + 0.017103541642427444, + 0.023972563445568085, + 0.019784053787589073, + 0.02025122195482254, + 0.01822805590927601, + 0.021467486396431923, + 0.01924292929470539, + 0.020299097523093224, + 0.0142943374812603, + 0.02530083991587162, + 0.012970587238669395, + 0.01940762810409069, + 0.02304651401937008, + 0.0258322861045599, + 0.03147047385573387, + 0.023077677935361862, + 0.00569764943793416, + 0.004804180935025215, + 0.00936267152428627, + 0.012547810561954975, + 0.031192272901535034, + 0.03623560070991516, + 0.014999400824308395, + 0.012434614822268486, + 0.0021406011655926704, + 0.0020640441216528416, + 0.0030478828120976686, + 0.0022463102359324694, + 0.0034406755585223436, + 0.004216646309942007, + 0.00591846602037549, + 0.004619124811142683, + 0.0019080713391304016, + 0.0026530728209763765, + 0.002538570435717702, + 0.0037498478777706623, + 0.002105663064867258, + 0.0016463376814499497, + 0.0032472028397023678, + 0.0027610743418335915, + 0.006625721696764231, + 0.006036283913999796, + 0.01756390929222107, + 0.008769357576966286, + 0.00743102328851819, + 0.00584449851885438, + 0.008413969539105892, + 0.01373308151960373, + 0.006169912405312061, + 0.002119584707543254, + 0.0064979600720107555, + 0.004422761499881744, + 0.0022796019911766052, + 0.0031630932353436947, + 0.003621603362262249, + 0.004655955359339714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.047775644809007645 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032239556312561 + ], + "opposite_logprob_tokens": [ + -2.3503224849700928 + ], + "logprob_reward": -0.10032239556312561, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032239556312561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15100371837615967, + 0.14005495607852936, + 0.08780372142791748, + 0.062157079577445984, + 0.08288848400115967, + 0.11614024639129639, + 0.021338967606425285, + 0.03149384260177612, + 0.25653064250946045, + 0.05058836564421654 + ], + "credit_map": [ + 0.01072668842971325, + 0.019072342664003372, + 0.020340830087661743, + 0.020503992214798927, + 0.023270992562174797, + 0.017951413989067078, + 0.020541587844491005, + 0.01859586499631405, + 0.01813015155494213, + 0.009963907301425934, + 0.03533918410539627, + 0.015988722443580627, + 0.016129186376929283, + 0.020084014162421227, + 0.013915077783167362, + 0.010504691861569881, + 0.005921134725213051, + 0.032181136310100555, + 0.005207202397286892, + 0.003757674479857087, + 0.01854240894317627, + 0.006370647344738245, + 0.004290089942514896, + 0.011533406563103199, + 0.006047426722943783, + 0.0074522835202515125, + 0.00973470974713564, + 0.00796844344586134, + 0.0052861846052110195, + 0.0081072598695755, + 0.007507314905524254, + 0.010053444653749466, + 0.0077546159736812115, + 0.01271013729274273, + 0.006397671066224575, + 0.010554985143244267, + 0.012887107208371162, + 0.015368625521659851, + 0.008569546043872833, + 0.008645784109830856, + 0.008065826259553432, + 0.0068623460829257965, + 0.005826723761856556, + 0.017253002151846886, + 0.026878999546170235, + 0.034330274909734726, + 0.010091297328472137, + 0.006831768900156021, + 0.001862863078713417, + 0.00194524135440588, + 0.0018058621790260077, + 0.002878309227526188, + 0.002493827138096094, + 0.002641314174979925, + 0.003398649860173464, + 0.004312899429351091, + 0.0031768635381013155, + 0.0034167002886533737, + 0.004053965676575899, + 0.006592646241188049, + 0.0028721976559609175, + 0.0030481002759188414, + 0.0028299528639763594, + 0.0055034104734659195, + 0.014286298304796219, + 0.010011078789830208, + 0.09390930086374283, + 0.01826060563325882, + 0.015107366256415844, + 0.011328283697366714, + 0.02492544800043106, + 0.06870223581790924, + 0.006128314416855574, + 0.004938594065606594, + 0.008970068767666817, + 0.008370785973966122, + 0.004320601932704449, + 0.006758734118193388, + 0.004418910015374422, + 0.006682348903268576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.10032239556312561 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026350490748882294 + ], + "opposite_logprob_tokens": [ + -3.651350498199463 + ], + "logprob_reward": -0.026350490748882294, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026350490748882294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09853162616491318, + 0.2465011179447174, + 0.32768514752388, + 0.060400597751140594, + 0.042838819324970245, + 0.06573377549648285, + 0.01765742525458336, + 0.039415135979652405, + 0.06114453822374344, + 0.040091872215270996 + ], + "credit_map": [ + 0.008720286190509796, + 0.015860669314861298, + 0.009770710952579975, + 0.010411561466753483, + 0.01798085682094097, + 0.013187809847295284, + 0.013031428679823875, + 0.009568299166858196, + 0.014375820755958557, + 0.009665340185165405, + 0.08951582759618759, + 0.01748008280992508, + 0.017456721514463425, + 0.06835845112800598, + 0.017145643010735512, + 0.012503214180469513, + 0.029909025877714157, + 0.07533100992441177, + 0.022750249132514, + 0.027114329859614372, + 0.04794193059206009, + 0.04780639335513115, + 0.01900792494416237, + 0.057824283838272095, + 0.008501636795699596, + 0.008372428826987743, + 0.007888037711381912, + 0.006947975140064955, + 0.007390434388071299, + 0.007860122248530388, + 0.004998684395104647, + 0.00844127219170332, + 0.003260915633291006, + 0.007828982546925545, + 0.0028455848805606365, + 0.005453092977404594, + 0.006375044584274292, + 0.006316768936812878, + 0.0070120044983923435, + 0.0037464213091880083, + 0.004782241303473711, + 0.005819179117679596, + 0.005859129596501589, + 0.0099294139072299, + 0.012110600247979164, + 0.013421936891973019, + 0.008230188861489296, + 0.005581074394285679, + 0.0024208305403590202, + 0.0017390550347045064, + 0.001688277698121965, + 0.001989487325772643, + 0.002041254425421357, + 0.002301426138728857, + 0.0024270806461572647, + 0.0030500106513500214, + 0.0038079023361206055, + 0.004170392174273729, + 0.005286351311951876, + 0.005155986174941063, + 0.003395841922610998, + 0.004762614611536264, + 0.005842175800353289, + 0.006993870250880718, + 0.003964952193200588, + 0.0028277982491999865, + 0.020972199738025665, + 0.004654115065932274, + 0.003438902087509632, + 0.0032297493889927864, + 0.005756437312811613, + 0.016300378367304802, + 0.005894585512578487, + 0.004612661432474852, + 0.0059151966124773026, + 0.004218692891299725, + 0.004679140169173479, + 0.005522309802472591, + 0.004556195344775915, + 0.004693089053034782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.026350490748882294 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293969303369522 + ], + "opposite_logprob_tokens": [ + -2.2379395961761475 + ], + "logprob_reward": -0.11293969303369522, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293969303369522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13364277780056, + 0.1683729737997055, + 0.19028566777706146, + 0.05077393725514412, + 0.2902424931526184, + 0.04109548404812813, + 0.01651117391884327, + 0.037864070385694504, + 0.04702842980623245, + 0.02418294921517372 + ], + "credit_map": [ + 0.009119203314185143, + 0.02089044451713562, + 0.01597709208726883, + 0.02401212975382805, + 0.014255949296057224, + 0.017486954107880592, + 0.020446937531232834, + 0.011454085819423199, + 0.011893072165548801, + 0.007178173866122961, + 0.057178203016519547, + 0.01582922227680683, + 0.01325458474457264, + 0.03790291026234627, + 0.013647641986608505, + 0.01148919202387333, + 0.014975523576140404, + 0.0704871118068695, + 0.008383319713175297, + 0.008409527130424976, + 0.03726755082607269, + 0.010990812443196774, + 0.006583800073713064, + 0.033188026398420334, + 0.005222443491220474, + 0.007808920461684465, + 0.005955376196652651, + 0.006308771204203367, + 0.00688187638297677, + 0.007327392231673002, + 0.005504783242940903, + 0.005764377769082785, + 0.026736220344901085, + 0.0462215319275856, + 0.02101297490298748, + 0.02917032316327095, + 0.03989006206393242, + 0.04641352593898773, + 0.051865193992853165, + 0.02893269807100296, + 0.003583999350667, + 0.0037066556978970766, + 0.003865642938762903, + 0.0048083290457725525, + 0.00805240124464035, + 0.007246345281600952, + 0.005414856132119894, + 0.00441725691780448, + 0.0013570276787504554, + 0.0012794316280633211, + 0.001445611473172903, + 0.0025193621404469013, + 0.0017186030745506287, + 0.001995574217289686, + 0.0030561364255845547, + 0.0031394297257065773, + 0.004833512008190155, + 0.004549114033579826, + 0.004236072767525911, + 0.0087467972189188, + 0.0024679440539330244, + 0.003460373030975461, + 0.0036132384557276964, + 0.005957020912319422, + 0.002883161185309291, + 0.0023162015713751316, + 0.01489780843257904, + 0.004090496338903904, + 0.0033668752294033766, + 0.0028240575920790434, + 0.004849680699408054, + 0.011800152249634266, + 0.0034525073133409023, + 0.0027995891869068146, + 0.0030353583861142397, + 0.002804205287247896, + 0.0034990867134183645, + 0.002944112056866288, + 0.0028694025240838528, + 0.0027786889113485813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.11293969303369522 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0538870096206665 + ], + "opposite_logprob_tokens": [ + -0.4288869798183441 + ], + "logprob_reward": -1.0538870096206665, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0538870096206665, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11304085701704025, + 0.41774043440818787, + 0.08213937282562256, + 0.09711631387472153, + 0.15225628018379211, + 0.03556692972779274, + 0.020563192665576935, + 0.0359293594956398, + 0.01657838188111782, + 0.029068846255540848 + ], + "credit_map": [ + 0.006597474217414856, + 0.012449592351913452, + 0.016724437475204468, + 0.01901586726307869, + 0.010190595872700214, + 0.00921990443021059, + 0.023940499871969223, + 0.01490249577909708, + 0.016349805518984795, + 0.013403684832155704, + 0.159129336476326, + 0.033372800797224045, + 0.02530999854207039, + 0.11131633818149567, + 0.030741170048713684, + 0.02811730094254017, + 0.0058913156390190125, + 0.022418800741434097, + 0.007090261671692133, + 0.008879585191607475, + 0.012953822501003742, + 0.008992279879748821, + 0.0035126213915646076, + 0.012400691397488117, + 0.007073297630995512, + 0.014724603854119778, + 0.010042975656688213, + 0.010286589153110981, + 0.01580638810992241, + 0.014056461863219738, + 0.01294504664838314, + 0.012180950492620468, + 0.009920576587319374, + 0.02077687904238701, + 0.008787833154201508, + 0.019748372957110405, + 0.024267084896564484, + 0.026863090693950653, + 0.025406869128346443, + 0.016485584899783134, + 0.0036503467708826065, + 0.0038736232090741396, + 0.002843277296051383, + 0.0039024504367262125, + 0.007447064388543367, + 0.006112242583185434, + 0.004082458559423685, + 0.0036554683465510607, + 0.0020672425162047148, + 0.0016596377827227116, + 0.0018105069175362587, + 0.0019952787552028894, + 0.002282430650666356, + 0.002863755216822028, + 0.002933263313025236, + 0.004951076582074165, + 0.00465876329690218, + 0.00426769582554698, + 0.003773109056055546, + 0.006121864076703787, + 0.002497888868674636, + 0.003483567154034972, + 0.00448355358093977, + 0.00664291949942708, + 0.0021079513244330883, + 0.0019982156809419394, + 0.0021488291677087545, + 0.002549027791246772, + 0.0016041095368564129, + 0.001751511124894023, + 0.0026430850848555565, + 0.0017756520537659526, + 0.0032926162239164114, + 0.00389488716609776, + 0.004343576263636351, + 0.003655832028016448, + 0.004195391666144133, + 0.0034761768765747547, + 0.003057134570553899, + 0.0031532314606010914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -1.0538870096206665 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "high-angle, wide-angle lens slowly and gracefully sweeps across a classic Chinese landscape, evoking a garden of folly and ignorance, bathed in the warm, nostalgic light of the afternoon. The water of a meticulously designed pond flows so gently and slowly, almost imperceptibly, its surface like polished glass, perfectly reflecting the golden glow of time and the dark, elegant silhouette of an ornate wooden pavilion with its weathered, tilted eaves. A fallen crimson maple leaf floats lazily on the water's surface, its delicate veins visible and its edges slightly curled. As the camera follows its leisurely journey, it passes a rickety stone bridge, its gray slate smoothed by centuries of footsteps. Beneath the tranquil surface, several enormous golden and white mallows swim lazily, their shimmering scales catching the submerged light as they glide in a sinuous, unhurried pattern. The entire scene, suffused with deep greens, warm golds, and earthy browns, is imbued with a sense of peaceful history, creating a visual of profoundly satisfying and tranquil meditation.", + "seed": 25486, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_13/index_6_seed_25486_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2525939878076315, + "vlm_logprob": -0.2525939878076315, + "vlm_yes_no_margin": 2.6499999523162843, + "vlm_token_credit": -0.2525939878076315, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03415845334529877 + ], + "opposite_logprob_tokens": [ + -4.034158229827881 + ], + "logprob_reward": -0.03415845334529877, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.03415845334529877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12245751172304153, + 0.09343092143535614, + 0.1352389007806778, + 0.39115551114082336, + 0.08234623074531555, + 0.031203320249915123, + 0.04819786921143532, + 0.013830370269715786, + 0.013071405701339245, + 0.06906798481941223 + ], + "credit_map": [ + 0.0066833170130848885, + 0.015435661189258099, + 0.018849404528737068, + 0.025820646435022354, + 0.01339386124163866, + 0.016459202393889427, + 0.017501583322882652, + 0.008313828147947788, + 0.00651384424418211, + 0.007733053062111139, + 0.018847372382879257, + 0.01862066425383091, + 0.012050097808241844, + 0.011641575954854488, + 0.010500753298401833, + 0.007523553445935249, + 0.00967361405491829, + 0.017642568796873093, + 0.015853026881814003, + 0.01875755935907364, + 0.02374894730746746, + 0.0172286257147789, + 0.022114722058176994, + 0.010219822637736797, + 0.012217443436384201, + 0.00844249315559864, + 0.028274372220039368, + 0.024618864059448242, + 0.01453616376966238, + 0.1184576079249382, + 0.0365026481449604, + 0.1481059044599533, + 0.012993969954550266, + 0.010279691778123379, + 0.01173811312764883, + 0.011360524222254753, + 0.00579087994992733, + 0.010246118530631065, + 0.008201893419027328, + 0.011735030449926853, + 0.0021857963874936104, + 0.0033483568113297224, + 0.005299113225191832, + 0.007202846463769674, + 0.0025782675947993994, + 0.002081603277474642, + 0.0044382670894265175, + 0.004069064743816853, + 0.0027051775250583887, + 0.0020752863492816687, + 0.00874309241771698, + 0.004215714521706104, + 0.002808284480124712, + 0.0026113237254321575, + 0.008081763982772827, + 0.016957219690084457, + 0.0010998029028996825, + 0.0012067462084814906, + 0.0016471780836582184, + 0.0011387515114620328, + 0.0011733418796211481, + 0.00284181023016572, + 0.0019748907070606947, + 0.0027478470001369715, + 0.001205061562359333, + 0.0011258762096986175, + 0.002026807051151991, + 0.002413724083453417, + 0.0010607991134747863, + 0.0011701632756739855, + 0.002258995547890663, + 0.0018099781591445208, + 0.005670598242431879, + 0.004361498169600964, + 0.009337585419416428, + 0.010539498180150986, + 0.00519784027710557, + 0.006554082501679659, + 0.01874338462948799, + 0.00866349320858717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.03415845334529877 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810384124517441 + ], + "opposite_logprob_tokens": [ + -3.2881038188934326 + ], + "logprob_reward": -0.03810384124517441, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810384124517441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16021274030208588, + 0.12170468270778656, + 0.13612093031406403, + 0.21530352532863617, + 0.0339178629219532, + 0.08134522289037704, + 0.05092032253742218, + 0.10998979955911636, + 0.05112338811159134, + 0.039361532777547836 + ], + "credit_map": [ + 0.0056099966168403625, + 0.01514322031289339, + 0.01721549779176712, + 0.045594438910484314, + 0.016675347462296486, + 0.02034834213554859, + 0.024876907467842102, + 0.014748972840607166, + 0.011333344504237175, + 0.011138005182147026, + 0.02802352048456669, + 0.013086413033306599, + 0.019686171784996986, + 0.0172114335000515, + 0.012248998507857323, + 0.008976787328720093, + 0.012160672806203365, + 0.020376907661557198, + 0.013834395445883274, + 0.017101941630244255, + 0.02514129877090454, + 0.016855165362358093, + 0.020494455471634865, + 0.010156077332794666, + 0.00936635211110115, + 0.009493958204984665, + 0.017008116468787193, + 0.01635928452014923, + 0.012492859736084938, + 0.05853550136089325, + 0.02057351917028427, + 0.07147391885519028, + 0.004809896927326918, + 0.0040517691522836685, + 0.004083986394107342, + 0.005245749372988939, + 0.0025881247129291296, + 0.004249458201229572, + 0.004270222503691912, + 0.004618652630597353, + 0.005971606820821762, + 0.011581351049244404, + 0.01074720174074173, + 0.019410381093621254, + 0.007682727184146643, + 0.007100355811417103, + 0.011542352847754955, + 0.007309239823371172, + 0.0038640580605715513, + 0.004849622957408428, + 0.006868063937872648, + 0.008316103368997574, + 0.0055161574855446815, + 0.004816613160073757, + 0.0091996556147933, + 0.007490043994039297, + 0.012870362028479576, + 0.01162467896938324, + 0.015077893622219563, + 0.016004810109734535, + 0.011429673060774803, + 0.00841983687132597, + 0.01979142054915428, + 0.014771112240850925, + 0.004616789519786835, + 0.0051568653434515, + 0.006490628235042095, + 0.009480425156652927, + 0.005393046885728836, + 0.004892892204225063, + 0.008260036818683147, + 0.0068327016197144985, + 0.0038662846200168133, + 0.0032246969640254974, + 0.004768725950270891, + 0.004652947187423706, + 0.004732816014438868, + 0.006340499501675367, + 0.006864834576845169, + 0.004910724237561226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.03810384124517441 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784122928977013 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784122928977013, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784122928977013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04202999919652939, + 0.05704600363969803, + 0.11617060750722885, + 0.6640987396240234, + 0.027228156104683876, + 0.021327247843146324, + 0.026279859244823456, + 0.012829024344682693, + 0.011517029255628586, + 0.021473288536071777 + ], + "credit_map": [ + 0.0018549765227362514, + 0.0043998886831104755, + 0.004634227138012648, + 0.009667760692536831, + 0.005244133062660694, + 0.005591989029198885, + 0.00683965440839529, + 0.003797373501583934, + 0.0058640711940824986, + 0.004640874918550253, + 0.01288335770368576, + 0.006869932170957327, + 0.008698425255715847, + 0.007432978600263596, + 0.006239062175154686, + 0.004417305812239647, + 0.00889512337744236, + 0.015598149038851261, + 0.013636338524520397, + 0.015582650899887085, + 0.020400021225214005, + 0.014831184409558773, + 0.01827441342175007, + 0.008952731266617775, + 0.019160404801368713, + 0.015572888776659966, + 0.04761800169944763, + 0.03995941951870918, + 0.02587045729160309, + 0.20246100425720215, + 0.06255508959293365, + 0.2509015202522278, + 0.002059041988104582, + 0.002649229019880295, + 0.0033491593785583973, + 0.0050156912766397, + 0.00252874125726521, + 0.0038313022814691067, + 0.004398545250296593, + 0.00339644867926836, + 0.0014933367492631078, + 0.0023885767441242933, + 0.0028035102877765894, + 0.003784285159781575, + 0.0021530436351895332, + 0.002376777585595846, + 0.003824538318440318, + 0.0025031811092048883, + 0.0016652120975777507, + 0.0017468947917222977, + 0.005290163680911064, + 0.002320394152775407, + 0.0018125615315511823, + 0.0018131498945876956, + 0.003329226514324546, + 0.008302261121571064, + 0.0012027979828417301, + 0.0010718487901613116, + 0.0024606529623270035, + 0.0012645975220948458, + 0.0011691321851685643, + 0.0014064772985875607, + 0.0017853088211268187, + 0.0024682101793587208, + 0.000887798669282347, + 0.001114799757488072, + 0.0017095240764319897, + 0.002214578213170171, + 0.001041117124259472, + 0.000843847868964076, + 0.0018922918243333697, + 0.001813072245568037, + 0.0014951983466744423, + 0.002131895162165165, + 0.002710360335186124, + 0.002209952101111412, + 0.0038858423940837383, + 0.0027923015877604485, + 0.003539928002282977, + 0.002707812236621976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.009784122928977013 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270487904548645 + ], + "opposite_logprob_tokens": [ + -2.1270487308502197 + ], + "logprob_reward": -0.1270487904548645, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270487904548645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14624091982841492, + 0.057606618851423264, + 0.476654976606369, + 0.08903910219669342, + 0.035572536289691925, + 0.021371034905314445, + 0.03378228098154068, + 0.02741587534546852, + 0.06736016273498535, + 0.0449565090239048 + ], + "credit_map": [ + 0.005552385933697224, + 0.012535425834357738, + 0.01461983472108841, + 0.04312421381473541, + 0.01809808798134327, + 0.018019378185272217, + 0.02229345589876175, + 0.011998136527836323, + 0.0055450210347771645, + 0.004423974081873894, + 0.010379714891314507, + 0.00808007176965475, + 0.008281704969704151, + 0.0078412014991045, + 0.006276177242398262, + 0.006778754759579897, + 0.03959216922521591, + 0.06813224405050278, + 0.05325791984796524, + 0.061455048620700836, + 0.08292747288942337, + 0.060741275548934937, + 0.0746227353811264, + 0.035926125943660736, + 0.004095492418855429, + 0.0027212202548980713, + 0.00661473860964179, + 0.007232923526316881, + 0.005710054654628038, + 0.02325025014579296, + 0.011429162696003914, + 0.027985263615846634, + 0.006070280913263559, + 0.004793159198015928, + 0.005109218880534172, + 0.005253831390291452, + 0.002129679312929511, + 0.004366080742329359, + 0.003424969734624028, + 0.004425317980349064, + 0.0015991171821951866, + 0.002205223310738802, + 0.0036098856944590807, + 0.004066308494657278, + 0.001522317063063383, + 0.00266292504966259, + 0.0034357612021267414, + 0.0022694976069033146, + 0.0016351043013855815, + 0.0016703751171007752, + 0.007924381643533707, + 0.0030704496894031763, + 0.0015413588844239712, + 0.0017992969369515777, + 0.004233135376125574, + 0.01190817728638649, + 0.002184024080634117, + 0.0024460090789943933, + 0.003881538985297084, + 0.0027653565630316734, + 0.002489142119884491, + 0.002731506945565343, + 0.005590419284999371, + 0.005327878054231405, + 0.006257131230086088, + 0.006867963820695877, + 0.008853566832840443, + 0.013070115819573402, + 0.007009223569184542, + 0.005478333681821823, + 0.01096793171018362, + 0.008855895139276981, + 0.00533251790329814, + 0.0031166966073215008, + 0.005860308650881052, + 0.005504031199961901, + 0.0061857192777097225, + 0.00596257159486413, + 0.006406831555068493, + 0.00658783596009016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.1270487904548645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0538747310638428 + ], + "opposite_logprob_tokens": [ + -0.4288747310638428 + ], + "logprob_reward": -1.0538747310638428, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0538747310638428, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14382468163967133, + 0.07706635445356369, + 0.2861543893814087, + 0.05455769598484039, + 0.019762476906180382, + 0.07315593957901001, + 0.059131920337677, + 0.023119235411286354, + 0.14085721969604492, + 0.12237003445625305 + ], + "credit_map": [ + 0.008188292384147644, + 0.015575532801449299, + 0.01887076534330845, + 0.03188163787126541, + 0.020447751507163048, + 0.019405681639909744, + 0.019544996321201324, + 0.009910018183290958, + 0.0061696055345237255, + 0.0056761144660413265, + 0.018347209319472313, + 0.013254974037408829, + 0.009799251332879066, + 0.009742789901793003, + 0.008044780232012272, + 0.006031621713191271, + 0.022781826555728912, + 0.039476145058870316, + 0.03323347121477127, + 0.03719601407647133, + 0.05091126263141632, + 0.03650088235735893, + 0.044603969901800156, + 0.021450841799378395, + 0.0049342745915055275, + 0.002099397126585245, + 0.006336150225251913, + 0.0056605711579322815, + 0.004310365300625563, + 0.011879880912601948, + 0.007393679115921259, + 0.011943379417061806, + 0.002126480685546994, + 0.0030185915529727936, + 0.0028522179927676916, + 0.002682392718270421, + 0.0014005257980898023, + 0.0019927294924855232, + 0.002135818125680089, + 0.0035537199582904577, + 0.004048202186822891, + 0.007757557090371847, + 0.01088615134358406, + 0.019292371347546577, + 0.004818134941160679, + 0.005297134630382061, + 0.012064075097441673, + 0.008992310613393784, + 0.0022950712591409683, + 0.0016432789852842689, + 0.013954147696495056, + 0.004713151138275862, + 0.0018660685745999217, + 0.003205831628292799, + 0.007989665493369102, + 0.023464705795049667, + 0.0027255183085799217, + 0.0019982026424258947, + 0.0033657692838460207, + 0.0033166471403092146, + 0.002528340322896838, + 0.00210688728839159, + 0.0035959926899522543, + 0.003481877502053976, + 0.016666419804096222, + 0.016330556944012642, + 0.018092641606926918, + 0.023778880015015602, + 0.015769196674227715, + 0.013224655762314796, + 0.016722500324249268, + 0.02027237042784691, + 0.011943528428673744, + 0.011285086162388325, + 0.0174239594489336, + 0.012874430976808071, + 0.019707081839442253, + 0.013883586972951889, + 0.019782278686761856, + 0.015470082871615887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -1.0538747310638428 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "high-angle, wide-angle lens slowly and gracefully sweeps across a classic Chinese landscape, evoking a garden of folly and ignorance, bathed in the warm, nostalgic light of the afternoon. The water of a meticulously designed pond flows so gently and slowly, almost imperceptibly, its surface like polished glass, perfectly reflecting the golden glow of time and the dark, elegant silhouette of an ornate wooden pavilion with its weathered, tilted eaves. A fallen crimson maple leaf floats lazily on the water's surface, its delicate veins visible and its edges slightly curled. As the camera follows its leisurely journey, it passes a rickety stone bridge, its gray slate smoothed by centuries of footsteps. Beneath the tranquil surface, several enormous golden and white mallows swim lazily, their shimmering scales catching the submerged light as they glide in a sinuous, unhurried pattern. The entire scene, suffused with deep greens, warm golds, and earthy browns, is imbued with a sense of peaceful history, creating a visual of profoundly satisfying and tranquil meditation.", + "seed": 25487, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_13/index_6_seed_25487_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3766306769102812, + "vlm_logprob": -0.3766306769102812, + "vlm_yes_no_margin": 1.825000023841858, + "vlm_token_credit": -0.3766306769102812, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12861868739128113 + ], + "opposite_logprob_tokens": [ + -2.3786187171936035 + ], + "logprob_reward": -0.12861868739128113, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.12861868739128113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11522901803255081, + 0.034084439277648926, + 0.14366820454597473, + 0.18331041932106018, + 0.08550398796796799, + 0.06216265261173248, + 0.047362320125103, + 0.05800141394138336, + 0.22226202487945557, + 0.048415541648864746 + ], + "credit_map": [ + 0.015185892581939697, + 0.01747988909482956, + 0.015427177771925926, + 0.014633066952228546, + 0.00940421037375927, + 0.011101756244897842, + 0.01465947087854147, + 0.017337551340460777, + 0.004730581771582365, + 0.0033854276407510042, + 0.003072743071243167, + 0.002603919478133321, + 0.003082940122112632, + 0.004308971110731363, + 0.007855289615690708, + 0.005044568330049515, + 0.01924367994070053, + 0.013903354294598103, + 0.02089119143784046, + 0.008905401453375816, + 0.011443952098488808, + 0.020104415714740753, + 0.027447398751974106, + 0.02172880992293358, + 0.027563229203224182, + 0.00840748567134142, + 0.04321824386715889, + 0.01184376236051321, + 0.013752306811511517, + 0.02497781254351139, + 0.030878985300660133, + 0.02266860194504261, + 0.00825040228664875, + 0.010010568425059319, + 0.009457004256546497, + 0.00879814475774765, + 0.008627046830952168, + 0.012306206859648228, + 0.011874154210090637, + 0.016180463135242462, + 0.005039750598371029, + 0.0028341684956103563, + 0.004213307518512011, + 0.010353458113968372, + 0.006145010236650705, + 0.011583380401134491, + 0.010281484574079514, + 0.011712092906236649, + 0.004539866466075182, + 0.0034138064365833998, + 0.005606410559266806, + 0.006437105592340231, + 0.0062255156226456165, + 0.007123718969523907, + 0.006861927919089794, + 0.007153963204473257, + 0.008292215876281261, + 0.007558821700513363, + 0.005628686863929033, + 0.006605977658182383, + 0.005789648741483688, + 0.005971273873001337, + 0.010097715072333813, + 0.00805707648396492, + 0.014232752844691277, + 0.024254173040390015, + 0.023411205038428307, + 0.041498225182294846, + 0.021190723404288292, + 0.029998043552041054, + 0.030531786382198334, + 0.03714512661099434, + 0.0023937630467116833, + 0.004608488641679287, + 0.004210531711578369, + 0.006123626139014959, + 0.004023244604468346, + 0.004378370940685272, + 0.012089339084923267, + 0.010588176548480988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.12861868739128113 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905123919248581 + ], + "opposite_logprob_tokens": [ + -2.4640512466430664 + ], + "logprob_reward": -0.08905123919248581, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905123919248581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09053180366754532, + 0.056574393063783646, + 0.034638047218322754, + 0.10933152586221695, + 0.07175090163946152, + 0.07360302656888962, + 0.06659398972988129, + 0.15935884416103363, + 0.293369323015213, + 0.04424813762307167 + ], + "credit_map": [ + 0.009317723102867603, + 0.011038663797080517, + 0.012486343272030354, + 0.011289617978036404, + 0.008928627707064152, + 0.00962604396045208, + 0.012002132833003998, + 0.015842657536268234, + 0.007728511933237314, + 0.008843781426548958, + 0.006280825939029455, + 0.005165422335267067, + 0.004157953429967165, + 0.0055880011059343815, + 0.010556661523878574, + 0.00825323723256588, + 0.004479422699660063, + 0.00354974577203393, + 0.005682807881385088, + 0.0033531750086694956, + 0.0032536282669752836, + 0.0045382920652627945, + 0.005557443480938673, + 0.004223530180752277, + 0.012386469170451164, + 0.0074138399213552475, + 0.01679285801947117, + 0.009177072905004025, + 0.010663572698831558, + 0.017936406657099724, + 0.016199996694922447, + 0.01876130700111389, + 0.009036681614816189, + 0.007815483957529068, + 0.014254453592002392, + 0.00578370550647378, + 0.005176443140953779, + 0.009424763731658459, + 0.011337483301758766, + 0.008921893313527107, + 0.007093305233865976, + 0.004848622716963291, + 0.0068598343059420586, + 0.014190309680998325, + 0.008180910721421242, + 0.011932163499295712, + 0.012723241932690144, + 0.007774636149406433, + 0.00866500660777092, + 0.004579269327223301, + 0.008419970981776714, + 0.01051386259496212, + 0.006858038250356913, + 0.00939403660595417, + 0.010136378929018974, + 0.008027426898479462, + 0.023639770224690437, + 0.022715993225574493, + 0.01480354554951191, + 0.017373714596033096, + 0.017644617706537247, + 0.019049523398280144, + 0.025858698412775993, + 0.018272994086146355, + 0.02022649720311165, + 0.027489664033055305, + 0.020367125049233437, + 0.034147970378398895, + 0.032721832394599915, + 0.034960731863975525, + 0.06486095488071442, + 0.05859456583857536, + 0.0033877750393003225, + 0.004000565968453884, + 0.004015196114778519, + 0.0042760674841701984, + 0.0051344577223062515, + 0.0029818692710250616, + 0.008588514290750027, + 0.011863692663609982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.08905123919248581 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590482279658318 + ], + "opposite_logprob_tokens": [ + -3.8955905437469482 + ], + "logprob_reward": -0.020590482279658318, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590482279658318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09408067911863327, + 0.04585690051317215, + 0.05685954913496971, + 0.062266867607831955, + 0.038827262818813324, + 0.03872290998697281, + 0.046922724694013596, + 0.20703166723251343, + 0.0761137455701828, + 0.33331769704818726 + ], + "credit_map": [ + 0.008818457834422588, + 0.019702788442373276, + 0.010404182597994804, + 0.013284792192280293, + 0.009468885138630867, + 0.008933216333389282, + 0.010434244759380817, + 0.013034111820161343, + 0.007352424319833517, + 0.007674351800233126, + 0.00466315308585763, + 0.003366770688444376, + 0.0033969476353377104, + 0.004759334959089756, + 0.00771981431171298, + 0.006924104411154985, + 0.010064673610031605, + 0.005057324189692736, + 0.007821346633136272, + 0.0036373392213135958, + 0.0036029296461492777, + 0.007729093544185162, + 0.01092943362891674, + 0.0080174021422863, + 0.006715001072734594, + 0.007830251939594746, + 0.008771002292633057, + 0.005380895920097828, + 0.008095827884972095, + 0.009933439083397388, + 0.008681424893438816, + 0.006859019864350557, + 0.004961722530424595, + 0.004188238177448511, + 0.0048211305402219296, + 0.004152422770857811, + 0.003938945475965738, + 0.004951312206685543, + 0.0065084113739430904, + 0.005305080208927393, + 0.003669587429612875, + 0.001961208414286375, + 0.0034775948151946068, + 0.005017547868192196, + 0.005632421933114529, + 0.006443867459893227, + 0.006083397194743156, + 0.006437285337597132, + 0.005133946891874075, + 0.003021401818841696, + 0.0046166772954165936, + 0.006312002893537283, + 0.005073882173746824, + 0.006262821611016989, + 0.008510632440447807, + 0.007991357706487179, + 0.014103403314948082, + 0.015251080505549908, + 0.01953934133052826, + 0.04243927076458931, + 0.016707535833120346, + 0.019046666100621223, + 0.042112238705158234, + 0.03783213347196579, + 0.004146441351622343, + 0.008262326009571552, + 0.006497898604720831, + 0.012423999607563019, + 0.007696711923927069, + 0.009263068437576294, + 0.011322198435664177, + 0.01650109700858593, + 0.010538904927670956, + 0.01659092865884304, + 0.0201776884496212, + 0.025354666635394096, + 0.02444634959101677, + 0.01884320005774498, + 0.10969936102628708, + 0.10766661167144775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.020590482279658318 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280645549297333 + ], + "opposite_logprob_tokens": [ + -2.0178065299987793 + ], + "logprob_reward": -0.14280645549297333, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14280645549297333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08395321667194366, + 0.035037703812122345, + 0.061058904975652695, + 0.06907153129577637, + 0.1584911048412323, + 0.16586491465568542, + 0.08418095111846924, + 0.0671461895108223, + 0.17952753603458405, + 0.09566797316074371 + ], + "credit_map": [ + 0.009483973495662212, + 0.010661021806299686, + 0.012975871562957764, + 0.01038006879389286, + 0.00846755038946867, + 0.00855720229446888, + 0.01090217661112547, + 0.012525352649390697, + 0.004130727145820856, + 0.004280590917915106, + 0.004129176028072834, + 0.0036369997542351484, + 0.003985283896327019, + 0.004681592341512442, + 0.00599780073389411, + 0.0041955322958528996, + 0.010907544754445553, + 0.004640304017812014, + 0.010960130020976067, + 0.00495600001886487, + 0.0042900945991277695, + 0.006088520400226116, + 0.00980998482555151, + 0.009406330063939095, + 0.00784537848085165, + 0.003966395277529955, + 0.0060326652601361275, + 0.007246191147714853, + 0.006179801654070616, + 0.015596595592796803, + 0.010594074614346027, + 0.011610429733991623, + 0.016371455043554306, + 0.009144052863121033, + 0.018757032230496407, + 0.014166004955768585, + 0.01436555478721857, + 0.02505739964544773, + 0.02614179626107216, + 0.03448781371116638, + 0.020050961524248123, + 0.01112117525190115, + 0.015362340956926346, + 0.02071441151201725, + 0.01941961608827114, + 0.030134066939353943, + 0.02510775253176689, + 0.02395458333194256, + 0.00890630204230547, + 0.005883257836103439, + 0.011380059644579887, + 0.0118169616907835, + 0.00844841543585062, + 0.010404775850474834, + 0.013206481002271175, + 0.014134694822132587, + 0.004916341044008732, + 0.0061985221691429615, + 0.007637143135070801, + 0.010758897289633751, + 0.007712816819548607, + 0.008448312059044838, + 0.011928174644708633, + 0.009545981884002686, + 0.008343227207660675, + 0.016514977440238, + 0.013373346999287605, + 0.022562207654118538, + 0.01976701058447361, + 0.01910129375755787, + 0.037557873874902725, + 0.04230758547782898, + 0.004028506111353636, + 0.00602581724524498, + 0.006266497075557709, + 0.008171269670128822, + 0.01004438754171133, + 0.006459270138293505, + 0.0264878012239933, + 0.028184428811073303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.14280645549297333 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5020865201950073 + ], + "opposite_logprob_tokens": [ + -0.2520865499973297 + ], + "logprob_reward": -1.5020865201950073, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5020865201950073, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09430747479200363, + 0.08101125806570053, + 0.04224614053964615, + 0.07113951444625854, + 0.08636114746332169, + 0.10213621705770493, + 0.10479746758937836, + 0.23743975162506104, + 0.11503230035305023, + 0.0655287504196167 + ], + "credit_map": [ + 0.00898313894867897, + 0.012818356044590473, + 0.009885321371257305, + 0.011857282370328903, + 0.01014050841331482, + 0.013001353479921818, + 0.014353367500007153, + 0.013268143869936466, + 0.014748642221093178, + 0.007845275104045868, + 0.011015642434358597, + 0.006483431439846754, + 0.007570497691631317, + 0.008379059843719006, + 0.013536395505070686, + 0.01143230963498354, + 0.008872822858393192, + 0.0025516392197459936, + 0.007648845203220844, + 0.002851595636457205, + 0.0024918843992054462, + 0.006193959154188633, + 0.00581330293789506, + 0.005822088103741407, + 0.005681619979441166, + 0.005972036626189947, + 0.008821231313049793, + 0.00448429211974144, + 0.0062293740920722485, + 0.008773214183747768, + 0.016402246430516243, + 0.014775493182241917, + 0.008396259509027004, + 0.011574741452932358, + 0.009068568237125874, + 0.008173122070729733, + 0.008152049966156483, + 0.013533986173570156, + 0.011797294020652771, + 0.015665119513869286, + 0.010759438388049603, + 0.006755542941391468, + 0.00774780660867691, + 0.018304264172911644, + 0.0129852918908, + 0.018065670505166054, + 0.01464601419866085, + 0.012872178107500076, + 0.010741987265646458, + 0.00623945239931345, + 0.010624188929796219, + 0.013061399571597576, + 0.0155160091817379, + 0.014175760559737682, + 0.01799703575670719, + 0.016441624611616135, + 0.026967614889144897, + 0.026151081547141075, + 0.018140463158488274, + 0.04454370588064194, + 0.023802615702152252, + 0.025610782206058502, + 0.038053758442401886, + 0.03416971117258072, + 0.007068888749927282, + 0.01327553205192089, + 0.011018932797014713, + 0.020824164152145386, + 0.012760655954480171, + 0.015831593424081802, + 0.014462047256529331, + 0.01979047805070877, + 0.004396827891469002, + 0.005954274442046881, + 0.008270171470940113, + 0.014044159092009068, + 0.00629375409334898, + 0.005816595628857613, + 0.010306521318852901, + 0.010446440428495407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -1.5020865201950073 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "high-angle, wide-angle lens slowly and gracefully sweeps across a classic Chinese landscape, evoking a garden of folly and ignorance, bathed in the warm, nostalgic light of the afternoon. The water of a meticulously designed pond flows so gently and slowly, almost imperceptibly, its surface like polished glass, perfectly reflecting the golden glow of time and the dark, elegant silhouette of an ornate wooden pavilion with its weathered, tilted eaves. A fallen crimson maple leaf floats lazily on the water's surface, its delicate veins visible and its edges slightly curled. As the camera follows its leisurely journey, it passes a rickety stone bridge, its gray slate smoothed by centuries of footsteps. Beneath the tranquil surface, several enormous golden and white mallows swim lazily, their shimmering scales catching the submerged light as they glide in a sinuous, unhurried pattern. The entire scene, suffused with deep greens, warm golds, and earthy browns, is imbued with a sense of peaceful history, creating a visual of profoundly satisfying and tranquil meditation.", + "seed": 25488, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_13/index_6_seed_25488_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3617654923349619, + "vlm_logprob": -0.3617654923349619, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.3617654923349619, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2571493685245514 + ], + "opposite_logprob_tokens": [ + -1.757149338722229 + ], + "logprob_reward": -0.2571493685245514, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2571493685245514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08711328357458115, + 0.5697260499000549, + 0.13222825527191162, + 0.03472813218832016, + 0.020031390711665154, + 0.06263052672147751, + 0.026088416576385498, + 0.03322036936879158, + 0.020903754979372025, + 0.013329812325537205 + ], + "credit_map": [ + 0.008780264295637608, + 0.008797462098300457, + 0.012383607216179371, + 0.01412252802401781, + 0.010522090829908848, + 0.00992145948112011, + 0.011080527678132057, + 0.01150534301996231, + 0.025183770805597305, + 0.057312190532684326, + 0.03829905763268471, + 0.08400484174489975, + 0.09939875453710556, + 0.13776162266731262, + 0.055063121020793915, + 0.07270274311304092, + 0.004823447670787573, + 0.00440027192234993, + 0.00622973358258605, + 0.006013140082359314, + 0.010782015509903431, + 0.026686077937483788, + 0.031485628336668015, + 0.04180794209241867, + 0.003862022887915373, + 0.0033406177535653114, + 0.003061051946133375, + 0.002688967389985919, + 0.005544344894587994, + 0.0040407925844192505, + 0.007370093837380409, + 0.004820238333195448, + 0.002276092302054167, + 0.0019722513388842344, + 0.001659771311096847, + 0.002421424724161625, + 0.0028169977013021708, + 0.002395814750343561, + 0.003678216366097331, + 0.0028108227998018265, + 0.007909282110631466, + 0.006617882288992405, + 0.005539199337363243, + 0.0038380222395062447, + 0.00866558775305748, + 0.011209119111299515, + 0.011063244193792343, + 0.007788192015141249, + 0.0030585932545363903, + 0.003625504206866026, + 0.0033342239912599325, + 0.0026888172142207623, + 0.003166099078953266, + 0.002732573077082634, + 0.0053537022322416306, + 0.002128906548023224, + 0.0020569555927067995, + 0.0023229478392750025, + 0.0033358184155076742, + 0.006506992504000664, + 0.0031771305948495865, + 0.005749008152633905, + 0.005975362844765186, + 0.004096154123544693, + 0.00222270586527884, + 0.00161478400696069, + 0.0031529858242720366, + 0.003514318261295557, + 0.0014872215688228607, + 0.0024647971149533987, + 0.003927671816200018, + 0.0025192720349878073, + 0.0013064045924693346, + 0.001336066285148263, + 0.0020400304347276688, + 0.0017516324296593666, + 0.00164792372379452, + 0.0012857220135629177, + 0.0020695794373750687, + 0.0018924541072919965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.2571493685245514 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043089546263217926 + ], + "opposite_logprob_tokens": [ + -3.1680896282196045 + ], + "logprob_reward": -0.043089546263217926, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043089546263217926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07206249982118607, + 0.2204662561416626, + 0.11822865903377533, + 0.036000680178403854, + 0.0330161452293396, + 0.11984860897064209, + 0.08881007134914398, + 0.11561048030853271, + 0.11618034541606903, + 0.07977620512247086 + ], + "credit_map": [ + 0.01112604234367609, + 0.00770749943330884, + 0.015186809003353119, + 0.006944168824702501, + 0.006016516126692295, + 0.00848088227212429, + 0.00845490675419569, + 0.008145684376358986, + 0.014918997883796692, + 0.021850593388080597, + 0.02017979323863983, + 0.026210343465209007, + 0.03750041872262955, + 0.05271022394299507, + 0.01997482217848301, + 0.027121083810925484, + 0.005792940966784954, + 0.0066559151746332645, + 0.006641515996307135, + 0.0060782223008573055, + 0.009884382598102093, + 0.034058865159749985, + 0.017334647476673126, + 0.031782180070877075, + 0.003615352790802717, + 0.004770928528159857, + 0.002874167403206229, + 0.00454970495775342, + 0.005320950411260128, + 0.0028188147116452456, + 0.00756268622353673, + 0.004488077014684677, + 0.004070593509823084, + 0.0044500562362372875, + 0.003044998040422797, + 0.00428383331745863, + 0.0034316405653953552, + 0.003934676293283701, + 0.005671016871929169, + 0.00412933761253953, + 0.011394055560231209, + 0.010292382910847664, + 0.010401809588074684, + 0.009639643132686615, + 0.023197008296847343, + 0.026187781244516373, + 0.013704945333302021, + 0.015030999667942524, + 0.01065509021282196, + 0.009315747767686844, + 0.010879464447498322, + 0.009631091728806496, + 0.011385677382349968, + 0.01205983106046915, + 0.017631052061915398, + 0.007252129726111889, + 0.005853712093085051, + 0.004120085388422012, + 0.01716034859418869, + 0.009919100441038609, + 0.008638446219265461, + 0.03767808899283409, + 0.009049929678440094, + 0.02319077029824257, + 0.01265527680516243, + 0.007943155243992805, + 0.02196688763797283, + 0.01921415701508522, + 0.01099836453795433, + 0.011834890581667423, + 0.019422048702836037, + 0.01214558444917202, + 0.008586419746279716, + 0.013292980380356312, + 0.010397128760814667, + 0.013417181558907032, + 0.011489929631352425, + 0.007221605628728867, + 0.00848331768065691, + 0.006887646857649088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.043089546263217926 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023299487307667732 + ], + "opposite_logprob_tokens": [ + -3.7732994556427 + ], + "logprob_reward": -0.023299487307667732, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023299487307667732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07456005364656448, + 0.2964221239089966, + 0.06751702725887299, + 0.02043023332953453, + 0.03779725730419159, + 0.026222825050354004, + 0.09694907814264297, + 0.1868649274110794, + 0.08598974347114563, + 0.10724679380655289 + ], + "credit_map": [ + 0.009360055439174175, + 0.006865767762064934, + 0.007822123356163502, + 0.011466129682958126, + 0.007302831392735243, + 0.012823579832911491, + 0.00875935796648264, + 0.010160200297832489, + 0.01261627022176981, + 0.030673641711473465, + 0.02325971983373165, + 0.04243358597159386, + 0.051023997366428375, + 0.07177367806434631, + 0.027552055194973946, + 0.03708915784955025, + 0.007451261859387159, + 0.00523521238937974, + 0.009148676879703999, + 0.006724965292960405, + 0.0060228463262319565, + 0.008558639325201511, + 0.010121354833245277, + 0.014254064299166203, + 0.0027999836020171642, + 0.002189791761338711, + 0.0017666014609858394, + 0.0019772881641983986, + 0.0025907824747264385, + 0.002857742365449667, + 0.003319640876725316, + 0.0029284001793712378, + 0.004212674684822559, + 0.003659371519461274, + 0.005090472288429737, + 0.004816174041479826, + 0.0057552289217710495, + 0.004042120184749365, + 0.0059784515760838985, + 0.004242758732289076, + 0.002721997443586588, + 0.0028897507581859827, + 0.002683533588424325, + 0.003189266659319401, + 0.0038565134163945913, + 0.0045660934410989285, + 0.0033866791054606438, + 0.0029289908707141876, + 0.011092266999185085, + 0.013527404516935349, + 0.01369201298803091, + 0.008664981462061405, + 0.014899460598826408, + 0.013884855434298515, + 0.013643069192767143, + 0.007545020896941423, + 0.008470699191093445, + 0.004937354940921068, + 0.0258672833442688, + 0.011453459970653057, + 0.019176265224814415, + 0.06389900296926498, + 0.014830458909273148, + 0.038230396807193756, + 0.0054541053250432014, + 0.0046729049645364285, + 0.01605389453470707, + 0.01646379567682743, + 0.008587157353758812, + 0.009644826874136925, + 0.015439893119037151, + 0.009673161432147026, + 0.01128215529024601, + 0.019265899434685707, + 0.014824007637798786, + 0.01824369840323925, + 0.013990018516778946, + 0.008669503964483738, + 0.010870386846363544, + 0.010101117193698883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.023299487307667732 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898911088705063 + ], + "opposite_logprob_tokens": [ + -2.578989028930664 + ], + "logprob_reward": -0.07898911088705063, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898911088705063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17405439913272858, + 0.20741884410381317, + 0.1892022341489792, + 0.046564217656850815, + 0.04021834954619408, + 0.13829977810382843, + 0.04081391170620918, + 0.10572326183319092, + 0.040580011904239655, + 0.01712499000132084 + ], + "credit_map": [ + 0.025993581861257553, + 0.0207154992967844, + 0.018501348793506622, + 0.02181044965982437, + 0.021039990708231926, + 0.02986227534711361, + 0.01572507619857788, + 0.020406169816851616, + 0.013316101394593716, + 0.02159084938466549, + 0.01779738813638687, + 0.03156711161136627, + 0.03526851534843445, + 0.0448693223297596, + 0.019807567819952965, + 0.023201996460556984, + 0.008440135978162289, + 0.008930020034313202, + 0.010426279157400131, + 0.008000634610652924, + 0.012795787304639816, + 0.0365227647125721, + 0.04387889802455902, + 0.060207728296518326, + 0.006414317525923252, + 0.004643326625227928, + 0.0047957198694348335, + 0.004458009265363216, + 0.006722778547555208, + 0.004674649331718683, + 0.008839039131999016, + 0.006016377359628677, + 0.0031890443060547113, + 0.0039040229748934507, + 0.003633391112089157, + 0.0034412327222526073, + 0.00649098539724946, + 0.004116774071007967, + 0.009821365587413311, + 0.005621534772217274, + 0.012018713168799877, + 0.009872539900243282, + 0.011263103224337101, + 0.012128586880862713, + 0.026927366852760315, + 0.026962870731949806, + 0.0228272695094347, + 0.016299327835440636, + 0.005824741907417774, + 0.004950523376464844, + 0.004652171395719051, + 0.004534500185400248, + 0.004883885849267244, + 0.004879593383520842, + 0.007886609062552452, + 0.003201887011528015, + 0.005443975795060396, + 0.0036880900152027607, + 0.015955934301018715, + 0.006915985606610775, + 0.010817066766321659, + 0.03343908116221428, + 0.009518839418888092, + 0.019944289699196815, + 0.002792369807139039, + 0.0025323296431452036, + 0.008238310925662518, + 0.006331087555736303, + 0.004208608064800501, + 0.004853848833590746, + 0.007208232767879963, + 0.004415225237607956, + 0.001438409206457436, + 0.0015738728689029813, + 0.0020690225064754486, + 0.002577126258984208, + 0.0023945297580212355, + 0.0021717974450439215, + 0.0026235042605549097, + 0.0022767267655581236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.07898911088705063 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and gracefully sweep across the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an ornate wooden pavilion with weathered, tilted eaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a fallen crimson maple leaf floating on the water's surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are enormous golden and white mallows visible swimming beneath the water's surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4062999486923218 + ], + "opposite_logprob_tokens": [ + -0.2812999486923218 + ], + "logprob_reward": -1.4062999486923218, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4062999486923218, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04056667536497116, + 0.3883146047592163, + 0.0759846642613411, + 0.020688936114311218, + 0.025141799822449684, + 0.10074011981487274, + 0.03759466111660004, + 0.1513642519712448, + 0.11225161701440811, + 0.047352589666843414 + ], + "credit_map": [ + 0.005979729816317558, + 0.003251199144870043, + 0.008029586635529995, + 0.0046661593951284885, + 0.003928260877728462, + 0.005388658028095961, + 0.004272229969501495, + 0.0050508566200733185, + 0.016410212963819504, + 0.0397762656211853, + 0.02422942966222763, + 0.062161531299352646, + 0.06767994165420532, + 0.09284672886133194, + 0.03791140764951706, + 0.04729912430047989, + 0.00515484856441617, + 0.003996363840997219, + 0.0038887152913957834, + 0.0040017347782850266, + 0.004942585714161396, + 0.012068281881511211, + 0.024764498695731163, + 0.017167650163173676, + 0.0034107835963368416, + 0.0027385554276406765, + 0.0024330690503120422, + 0.002141161821782589, + 0.0026200469583272934, + 0.001854399568401277, + 0.0029867319390177727, + 0.0025041894987225533, + 0.0029272637329995632, + 0.004880570340901613, + 0.0024985193740576506, + 0.0025857044383883476, + 0.002925977110862732, + 0.002725188387557864, + 0.003761994419619441, + 0.0028365859761834145, + 0.009880763478577137, + 0.009178944863379002, + 0.009262483566999435, + 0.007251692935824394, + 0.016300367191433907, + 0.01957830972969532, + 0.01720269024372101, + 0.01208487618714571, + 0.003447900293394923, + 0.003642325522378087, + 0.004617743194103241, + 0.004485991317778826, + 0.00467090355232358, + 0.004979145247489214, + 0.008380144834518433, + 0.003370514139533043, + 0.004805229604244232, + 0.003683007787913084, + 0.019540995359420776, + 0.005775693338364363, + 0.012369528412818909, + 0.06064893305301666, + 0.007362334989011288, + 0.03717855364084244, + 0.008199798874557018, + 0.006589414551854134, + 0.016275329515337944, + 0.025552894920110703, + 0.009380007162690163, + 0.010624872520565987, + 0.023278959095478058, + 0.012350349687039852, + 0.006177883595228195, + 0.00693455571308732, + 0.006954798940569162, + 0.0073694572784006596, + 0.006398611702024937, + 0.003933633677661419, + 0.004526418633759022, + 0.005057234782725573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 8 + ], + "selected_reward": -1.4062999486923218 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_14/index_tensor([1694, 1694])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_14/index_tensor([1694, 1694])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..abb28c7c8091d0b524714734f35324e59126d340 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_14/index_tensor([1694, 1694])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, static shot, captured from a horizontal angle, captures the mesmerizing moment of a perfectly ripe, vibrant red strawberry slowly sinking into a glass of crystal-clear liquid. Bright, soft sidelight streams from the side, penetrating the glass and precisely outlining the plump contours and moist sheen of the strawberry, dotted with tiny golden seeds. The moment the strawberry is fully submerged, its edges miraculously dissolve, releasing wisps of red pigment that resemble smoke and mist. These delicate threads gracefully unfurl, swirl, and slowly diffuse within the still liquid, as if living tentacles exploring the transparent space around them, creating a dynamic, abstract watercolor painting. The pure tones of the image, dominated by the strawberry's ruby red, contrast sharply with the transparent liquid and the minimalist background, create an atmosphere of tranquility, concentration, and a surreal, poetic beauty.", + "seed": 2599, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_14/index_1694_seed_2599_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12386815343052149, + "vlm_logprob": -0.12386815343052149, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.12386815343052149, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007290544454008341 + ], + "opposite_logprob_tokens": [ + -5.757290363311768 + ], + "logprob_reward": -0.007290544454008341, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.007290544454008341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1251051127910614, + 0.13097131252288818, + 0.4207417666912079, + 0.07497970759868622, + 0.03373124077916145, + 0.027753854170441628, + 0.05280930921435356, + 0.06839053332805634, + 0.028594907373189926, + 0.03692225366830826 + ], + "credit_map": [ + 0.007182830013334751, + 0.013067269697785378, + 0.005202263593673706, + 0.022199394181370735, + 0.009563671424984932, + 0.024323584511876106, + 0.014971642754971981, + 0.02859446033835411, + 0.010207046754658222, + 0.017676955088973045, + 0.010043165646493435, + 0.017957618460059166, + 0.019209593534469604, + 0.014979979954659939, + 0.01787872426211834, + 0.023018237203359604, + 0.024724161252379417, + 0.026700817048549652, + 0.08384805917739868, + 0.14096517860889435, + 0.013836662285029888, + 0.033022768795490265, + 0.08257705718278885, + 0.015067066065967083, + 0.005051387939602137, + 0.00680503249168396, + 0.006501840427517891, + 0.013707159087061882, + 0.0076241763308644295, + 0.007185669615864754, + 0.016133898869156837, + 0.01197054609656334, + 0.003914758563041687, + 0.00404888391494751, + 0.004008558113127947, + 0.004701288416981697, + 0.004363778978586197, + 0.0037712149787694216, + 0.0038752725813537836, + 0.005047486629337072, + 0.0034884423948824406, + 0.0036237624008208513, + 0.0034278614912182093, + 0.003141900058835745, + 0.004274534061551094, + 0.0032345831859856844, + 0.0023265844210982323, + 0.004236188251525164, + 0.008376507088541985, + 0.008428027853369713, + 0.006642250809818506, + 0.006089520175009966, + 0.00810586754232645, + 0.0040290034376084805, + 0.004158751107752323, + 0.006979379802942276, + 0.011605905368924141, + 0.009002775885164738, + 0.008063218556344509, + 0.010485967621207237, + 0.008092275820672512, + 0.005121993366628885, + 0.006321531254798174, + 0.009696868248283863, + 0.003760439110919833, + 0.003579992102459073, + 0.004805941600352526, + 0.003996815532445908, + 0.00343296374194324, + 0.0037067376542836428, + 0.002514103427529335, + 0.002797916531562805, + 0.0031418001744896173, + 0.005384251940995455, + 0.007293493952602148, + 0.0033215568400919437, + 0.004116182681173086, + 0.005290291737765074, + 0.005333574954420328, + 0.0030411004554480314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.007290544454008341 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.576153039932251 + ], + "opposite_logprob_tokens": [ + -0.826153039932251 + ], + "logprob_reward": -0.576153039932251, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.576153039932251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03358105570077896, + 0.0366390235722065, + 0.7997859716415405, + 0.026213783770799637, + 0.017339596524834633, + 0.0076840040273964405, + 0.015473290346562862, + 0.01532378327101469, + 0.023658882826566696, + 0.02430068328976631 + ], + "credit_map": [ + 0.0025922588538378477, + 0.004088118206709623, + 0.002206819597631693, + 0.006725285202264786, + 0.002999418880790472, + 0.00462261913344264, + 0.004919907543808222, + 0.005426629912108183, + 0.002569956937804818, + 0.0050332266837358475, + 0.0024625675287097692, + 0.007205519359558821, + 0.0035197173710912466, + 0.004593173507601023, + 0.004783380776643753, + 0.006471480708569288, + 0.047889724373817444, + 0.050918664783239365, + 0.1615477204322815, + 0.2687312662601471, + 0.02373833768069744, + 0.06306526809930801, + 0.15518911182880402, + 0.028705859556794167, + 0.0017469029407948256, + 0.0021406637970358133, + 0.0027690420392900705, + 0.00376046821475029, + 0.0028644877020269632, + 0.002699572592973709, + 0.004871735814958811, + 0.005360910203307867, + 0.0017241028835996985, + 0.002036083023995161, + 0.0017149992054328322, + 0.001979435095563531, + 0.0024390658363699913, + 0.0022639816161245108, + 0.0022483246866613626, + 0.0029336039442569017, + 0.0006217231857590377, + 0.0009346322040073574, + 0.000780046044383198, + 0.0009769236203283072, + 0.0008212546817958355, + 0.0009280373924411833, + 0.0007732428493909538, + 0.0018481436418369412, + 0.0012777565279975533, + 0.0031664222478866577, + 0.001968798227608204, + 0.0019258209504187107, + 0.0023344147484749556, + 0.001458450686186552, + 0.0013660630211234093, + 0.0019755640532821417, + 0.0026387504767626524, + 0.0016395836137235165, + 0.0016489098779857159, + 0.002052134834229946, + 0.0025472701527178288, + 0.0014779485063627362, + 0.0018296998459845781, + 0.0014894854975864291, + 0.0023527899757027626, + 0.0020032688044011593, + 0.0025864243507385254, + 0.005495211109519005, + 0.002465864410623908, + 0.0030717954505234957, + 0.00353589141741395, + 0.002147634979337454, + 0.003562182653695345, + 0.0031452642288058996, + 0.0056497761979699135, + 0.0025679192040115595, + 0.0026525864377617836, + 0.002173776738345623, + 0.0022141842637211084, + 0.00233499426394701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.576153039932251 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739741191267967 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739741191267967, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739741191267967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09235469996929169, + 0.14472630620002747, + 0.30163058638572693, + 0.10257737338542938, + 0.07963347434997559, + 0.05270574986934662, + 0.04492678493261337, + 0.06317026168107986, + 0.07695852965116501, + 0.04131622985005379 + ], + "credit_map": [ + 0.01003179233521223, + 0.014589238911867142, + 0.004175015725195408, + 0.01929319277405739, + 0.007061721291393042, + 0.011511974968016148, + 0.012547900900244713, + 0.013143859803676605, + 0.008220759220421314, + 0.02008117362856865, + 0.012440821155905724, + 0.020993486046791077, + 0.01765986531972885, + 0.020091695711016655, + 0.01794072426855564, + 0.027297789230942726, + 0.017695456743240356, + 0.01999911665916443, + 0.0584380179643631, + 0.09682624042034149, + 0.01348638255149126, + 0.023018933832645416, + 0.05852152407169342, + 0.013644936494529247, + 0.00883739534765482, + 0.018979908898472786, + 0.010582583956420422, + 0.015973400324583054, + 0.006715065333992243, + 0.009000664576888084, + 0.018068917095661163, + 0.01441943272948265, + 0.0061140055768191814, + 0.008771314285695553, + 0.0072786686941981316, + 0.013231582008302212, + 0.011661509983241558, + 0.008200064301490784, + 0.011448432691395283, + 0.012927904725074768, + 0.0066990009509027, + 0.006588888354599476, + 0.006165201775729656, + 0.007397537585347891, + 0.005989489611238241, + 0.005159861873835325, + 0.005967415403574705, + 0.008738355711102486, + 0.004399981815367937, + 0.008752072229981422, + 0.005509506445378065, + 0.005158423911780119, + 0.006939177867025137, + 0.00375672266818583, + 0.003889746032655239, + 0.006521151401102543, + 0.01189498696476221, + 0.0084080770611763, + 0.007022491190582514, + 0.009003469720482826, + 0.007113793399184942, + 0.004954383242875338, + 0.006518350914120674, + 0.008254708722233772, + 0.009901757352054119, + 0.0065840138122439384, + 0.009516852907836437, + 0.01421278715133667, + 0.008901202119886875, + 0.01294501218944788, + 0.007381505332887173, + 0.007515395525842905, + 0.0030786877032369375, + 0.005014258902519941, + 0.006160889286547899, + 0.0050563993863761425, + 0.007907402701675892, + 0.0051009636372327805, + 0.005793750286102295, + 0.0032038770150393248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.006739741191267967 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819374669343233 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.002819374669343233, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819374669343233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10975269228219986, + 0.051380883902311325, + 0.474116712808609, + 0.10085754841566086, + 0.026827996596693993, + 0.018430152907967567, + 0.03930080682039261, + 0.10011845827102661, + 0.0338619090616703, + 0.04535277932882309 + ], + "credit_map": [ + 0.012866506353020668, + 0.010922796092927456, + 0.00621760031208396, + 0.023714831098914146, + 0.008168434724211693, + 0.017180627211928368, + 0.014321275055408478, + 0.016360633075237274, + 0.003004505066201091, + 0.0073744384571909904, + 0.005398819223046303, + 0.006292157806456089, + 0.008287663571536541, + 0.006299767643213272, + 0.0063419220969080925, + 0.008381618186831474, + 0.02818177454173565, + 0.030193787068128586, + 0.0964450016617775, + 0.15556345880031586, + 0.016305945813655853, + 0.03784767538309097, + 0.09162480384111404, + 0.017954301089048386, + 0.007222194690257311, + 0.009958154521882534, + 0.00901916902512312, + 0.017362816259264946, + 0.009472653269767761, + 0.00919723603874445, + 0.021282954141497612, + 0.017342373728752136, + 0.0030689877457916737, + 0.0036784678231924772, + 0.002460313495248556, + 0.003936120308935642, + 0.0034454292617738247, + 0.0027492158114910126, + 0.003216556506231427, + 0.004272908437997103, + 0.002160874893888831, + 0.002531040459871292, + 0.0021299913059920073, + 0.0026778760366141796, + 0.001955207670107484, + 0.0020975335501134396, + 0.002040325431153178, + 0.002837306587025523, + 0.002664590021595359, + 0.0065047224052250385, + 0.004605419468134642, + 0.0050575388595461845, + 0.0069735487923026085, + 0.003427068470045924, + 0.004083624575287104, + 0.005984299350529909, + 0.008131361566483974, + 0.011717651039361954, + 0.009171255864202976, + 0.017932966351509094, + 0.018011972308158875, + 0.007731788791716099, + 0.016834227368235588, + 0.010587242431938648, + 0.0037555512972176075, + 0.003663848852738738, + 0.00535724638029933, + 0.005320954602211714, + 0.003678989829495549, + 0.004794382490217686, + 0.0042719608172774315, + 0.0030189761891961098, + 0.003571971319615841, + 0.006680068094283342, + 0.005053563974797726, + 0.008557682856917381, + 0.006130059715360403, + 0.005446390714496374, + 0.006335947662591934, + 0.0035771005786955357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.002819374669343233 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026338066905736923 + ], + "opposite_logprob_tokens": [ + -3.6513381004333496 + ], + "logprob_reward": -0.026338066905736923, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026338066905736923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11642751842737198, + 0.08493195474147797, + 0.3812978267669678, + 0.13051292300224304, + 0.04038005694746971, + 0.03227711096405983, + 0.05523139238357544, + 0.03489771485328674, + 0.059994347393512726, + 0.06404918432235718 + ], + "credit_map": [ + 0.011709872633218765, + 0.008991955779492855, + 0.008080782368779182, + 0.00928846001625061, + 0.011151392944157124, + 0.03140570968389511, + 0.009199955500662327, + 0.026599397882819176, + 0.007337562274187803, + 0.012190218083560467, + 0.006335118319839239, + 0.012718313373625278, + 0.011669800616800785, + 0.010320640169084072, + 0.009366629645228386, + 0.014993663877248764, + 0.021666651591658592, + 0.02576998993754387, + 0.07480919361114502, + 0.12235768884420395, + 0.01785285212099552, + 0.030078453943133354, + 0.07324942201375961, + 0.015513560734689236, + 0.011696213856339455, + 0.010520445182919502, + 0.011233608238399029, + 0.024821944534778595, + 0.010821707546710968, + 0.011520246975123882, + 0.026686741039156914, + 0.0232120081782341, + 0.005671621300280094, + 0.004629557020962238, + 0.004945117048919201, + 0.0047029899433255196, + 0.005186029709875584, + 0.005404850468039513, + 0.004700699355453253, + 0.005139191634953022, + 0.005221433471888304, + 0.0035991875920444727, + 0.004381856881082058, + 0.004842718131840229, + 0.0032896604388952255, + 0.0035622662398964167, + 0.00292185484431684, + 0.004458135459572077, + 0.015260064974427223, + 0.006191429682075977, + 0.008971426635980606, + 0.006423399783670902, + 0.005253245588392019, + 0.004329144489020109, + 0.0037387744523584843, + 0.005063906311988831, + 0.006594902370125055, + 0.005336386617273092, + 0.003844551043584943, + 0.006712594069540501, + 0.0033504124730825424, + 0.002138652838766575, + 0.003488610964268446, + 0.003431606339290738, + 0.012959007173776627, + 0.007981979288160801, + 0.012478940188884735, + 0.00625645462423563, + 0.005041946191340685, + 0.0048361653462052345, + 0.0036131637170910835, + 0.006826692260801792, + 0.007823869585990906, + 0.00838130060583353, + 0.006429059896618128, + 0.009556859731674194, + 0.011458813212811947, + 0.005350392311811447, + 0.009655090048909187, + 0.005393797066062689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.026338066905736923 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot, captured from a horizontal angle, captures the mesmerizing moment of a perfectly ripe, vibrant red strawberry slowly sinking into a glass of crystal-clear liquid. Bright, soft sidelight streams from the side, penetrating the glass and precisely outlining the plump contours and moist sheen of the strawberry, dotted with tiny golden seeds. The moment the strawberry is fully submerged, its edges miraculously dissolve, releasing wisps of red pigment that resemble smoke and mist. These delicate threads gracefully unfurl, swirl, and slowly diffuse within the still liquid, as if living tentacles exploring the transparent space around them, creating a dynamic, abstract watercolor painting. The pure tones of the image, dominated by the strawberry's ruby red, contrast sharply with the transparent liquid and the minimalist background, create an atmosphere of tranquility, concentration, and a surreal, poetic beauty.", + "seed": 2600, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_14/index_1694_seed_2600_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07727840119041503, + "vlm_logprob": -0.07727840119041503, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.07727840119041503, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008077218197286129 + ], + "opposite_logprob_tokens": [ + -6.633077144622803 + ], + "logprob_reward": -0.008077218197286129, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.008077218197286129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039235375821590424, + 0.03420306369662285, + 0.12617355585098267, + 0.03908997029066086, + 0.09322208166122437, + 0.07125682383775711, + 0.1107645034790039, + 0.21385611593723297, + 0.10960360616445541, + 0.16259486973285675 + ], + "credit_map": [ + 0.005897077731788158, + 0.0041851140558719635, + 0.006552180275321007, + 0.006197725888341665, + 0.0026534839998930693, + 0.0049698976799845695, + 0.004315070807933807, + 0.004464826546609402, + 0.0043065245263278484, + 0.0028500454500317574, + 0.007304992992430925, + 0.00748197827488184, + 0.00205996073782444, + 0.0021071331575512886, + 0.0022091935388743877, + 0.005883237347006798, + 0.009130974300205708, + 0.006947510410100222, + 0.044924959540367126, + 0.011268030852079391, + 0.004853143822401762, + 0.005034599453210831, + 0.0044332570396363735, + 0.03958108648657799, + 0.005810274742543697, + 0.003680821508169174, + 0.006440558470785618, + 0.008890880271792412, + 0.004108468070626259, + 0.0027262705843895674, + 0.0027255224995315075, + 0.004707178566604853, + 0.013041162863373756, + 0.009221027605235577, + 0.020269690081477165, + 0.016559574753046036, + 0.00930224359035492, + 0.006019327789545059, + 0.006828106939792633, + 0.011980962008237839, + 0.009073643945157528, + 0.013610845431685448, + 0.010905284434556961, + 0.012033591978251934, + 0.006989913061261177, + 0.005897510796785355, + 0.006558489985764027, + 0.006187546066939831, + 0.006980265025049448, + 0.004578604828566313, + 0.043717995285987854, + 0.042315807193517685, + 0.0032078265212476254, + 0.0023568004835397005, + 0.0048147630877792835, + 0.0027924536261707544, + 0.02063627354800701, + 0.011209764517843723, + 0.05875305458903313, + 0.029817357659339905, + 0.01713510975241661, + 0.006896327715367079, + 0.058306753635406494, + 0.011101487092673779, + 0.012365138158202171, + 0.023915909230709076, + 0.012995281256735325, + 0.01815645769238472, + 0.011400211602449417, + 0.011133449152112007, + 0.01155509240925312, + 0.008082076907157898, + 0.022715182974934578, + 0.02363407425582409, + 0.020016884431242943, + 0.021928127855062485, + 0.01449778862297535, + 0.02468230202794075, + 0.019927529618144035, + 0.015192978084087372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 2 + ], + "selected_reward": -0.008077218197286129 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.313448041677475 + ], + "opposite_logprob_tokens": [ + -1.3134480714797974 + ], + "logprob_reward": -0.313448041677475, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.313448041677475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11274588108062744, + 0.03687567263841629, + 0.046652067452669144, + 0.05224550515413284, + 0.10833172500133514, + 0.08028845489025116, + 0.16401289403438568, + 0.21397724747657776, + 0.10382125526666641, + 0.08104932308197021 + ], + "credit_map": [ + 0.01964101754128933, + 0.009184712544083595, + 0.029104109853506088, + 0.008475969545543194, + 0.006263785529881716, + 0.016509855166077614, + 0.0057962131686508656, + 0.01777021214365959, + 0.005018280353397131, + 0.0030869636684656143, + 0.008629750460386276, + 0.004803421441465616, + 0.0023372408468276262, + 0.0028646965511143208, + 0.003999311476945877, + 0.006136002019047737, + 0.007463248912245035, + 0.004092822317034006, + 0.0054494659416377544, + 0.008591673336923122, + 0.00656892592087388, + 0.004529454279690981, + 0.003559844335541129, + 0.006396633107215166, + 0.007364322431385517, + 0.0049698688089847565, + 0.006440089084208012, + 0.01462287362664938, + 0.004435061942785978, + 0.0031199315562844276, + 0.0038773827254772186, + 0.007415976375341415, + 0.014018607325851917, + 0.015478204004466534, + 0.016758184880018234, + 0.01367079559713602, + 0.018894419074058533, + 0.009728743694722652, + 0.010936792008578777, + 0.008845963515341282, + 0.010345159098505974, + 0.01598505489528179, + 0.013769236393272877, + 0.015945827588438988, + 0.005721050314605236, + 0.005450775846838951, + 0.007178146857768297, + 0.005893199238926172, + 0.012521625496447086, + 0.007048956584185362, + 0.06512456387281418, + 0.05739089846611023, + 0.0050030979327857494, + 0.00477991858497262, + 0.0064215874299407005, + 0.00572224473580718, + 0.023480037227272987, + 0.010514874011278152, + 0.05802160128951073, + 0.03070678748190403, + 0.017514795064926147, + 0.0067878891713917255, + 0.05726485699415207, + 0.009686392731964588, + 0.009605009108781815, + 0.020406628027558327, + 0.010147676803171635, + 0.017866430804133415, + 0.012069158256053925, + 0.01039000041782856, + 0.012806766666471958, + 0.010529575869441032, + 0.017556417733430862, + 0.009727121330797672, + 0.0104067362844944, + 0.009213356301188469, + 0.006419266574084759, + 0.012467642314732075, + 0.007148400414735079, + 0.00811038538813591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.313448041677475 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819128356873989 + ], + "opposite_logprob_tokens": [ + -4.018191337585449 + ], + "logprob_reward": -0.01819128356873989, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819128356873989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016107803210616112, + 0.020633574575185776, + 0.04650260508060455, + 0.01967061124742031, + 0.20577391982078552, + 0.3190510869026184, + 0.24094714224338531, + 0.09110378473997116, + 0.024861745536327362, + 0.015347698703408241 + ], + "credit_map": [ + 0.0024732109159231186, + 0.0014374314341694117, + 0.0035282878670841455, + 0.0012751745525747538, + 0.0015719239600002766, + 0.0021277270279824734, + 0.001052050618454814, + 0.002641997067257762, + 0.0031610880978405476, + 0.0015993791166692972, + 0.00445207767188549, + 0.0034858526196330786, + 0.0015165576478466392, + 0.0017291017575189471, + 0.001056565553881228, + 0.0036329522263258696, + 0.003394806757569313, + 0.0025342311710119247, + 0.016791274771094322, + 0.004196622874587774, + 0.0016230582259595394, + 0.0019536742474883795, + 0.0016619206871837378, + 0.014347017742693424, + 0.0026283443439751863, + 0.0017084480496123433, + 0.002501329407095909, + 0.004346395377069712, + 0.0022642051335424185, + 0.0019551280420273542, + 0.0018906984478235245, + 0.002376062097027898, + 0.031007546931505203, + 0.01894937828183174, + 0.048516351729631424, + 0.040976569056510925, + 0.017475944012403488, + 0.012157992459833622, + 0.016958167776465416, + 0.01973198726773262, + 0.038236647844314575, + 0.08011209964752197, + 0.034161508083343506, + 0.05351689085364342, + 0.02524549327790737, + 0.02879558503627777, + 0.03563528135418892, + 0.023347580805420876, + 0.01598532311618328, + 0.009949373081326485, + 0.09995429217815399, + 0.09321517497301102, + 0.006134687922894955, + 0.0030580121092498302, + 0.007612852845340967, + 0.005037439987063408, + 0.008876618929207325, + 0.0050539555959403515, + 0.02564408630132675, + 0.01294094230979681, + 0.007025007624179125, + 0.0031412853859364986, + 0.023899545893073082, + 0.0045223417691886425, + 0.0035650760401040316, + 0.004607857204973698, + 0.0027203725185245275, + 0.004427147563546896, + 0.0022850653622299433, + 0.002557514701038599, + 0.0025913130957633257, + 0.002107398584485054, + 0.0021735820919275284, + 0.0023814912419766188, + 0.0018436913378536701, + 0.0024579684250056744, + 0.0012073916150256991, + 0.0022365355398505926, + 0.0016246826853603125, + 0.001422356115654111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.01819128356873989 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036178866866976023 + ], + "opposite_logprob_tokens": [ + -5.628617763519287 + ], + "logprob_reward": -0.0036178866866976023, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036178866866976023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031343113631010056, + 0.0299820676445961, + 0.01900809071958065, + 0.11397731304168701, + 0.16196277737617493, + 0.20842505991458893, + 0.16821922361850739, + 0.12531669437885284, + 0.0751102864742279, + 0.06665533781051636 + ], + "credit_map": [ + 0.0054804072715342045, + 0.0021468978375196457, + 0.00712371664121747, + 0.0021132228430360556, + 0.0024862773716449738, + 0.0048993658274412155, + 0.0018870142521336675, + 0.005206211470067501, + 0.004413241054862738, + 0.002986447187140584, + 0.00731880497187376, + 0.004035998601466417, + 0.0016969605349004269, + 0.002401593141257763, + 0.0020232058595865965, + 0.005105816293507814, + 0.0030685693491250277, + 0.0023684792686253786, + 0.0027038520202040672, + 0.002695111557841301, + 0.0023477463982999325, + 0.001847913721576333, + 0.0013209680328145623, + 0.0026554500218480825, + 0.016536012291908264, + 0.010596519336104393, + 0.0157458558678627, + 0.032787926495075226, + 0.00910585280507803, + 0.005951411090791225, + 0.00737122492864728, + 0.015882503241300583, + 0.02223566733300686, + 0.01668199710547924, + 0.032750144600868225, + 0.030150722712278366, + 0.018147367984056473, + 0.011808352544903755, + 0.014224149286746979, + 0.01596437208354473, + 0.025893911719322205, + 0.04770529270172119, + 0.026421133428812027, + 0.034914977848529816, + 0.01638147607445717, + 0.01899743638932705, + 0.022672098129987717, + 0.01543873269110918, + 0.011823105625808239, + 0.006450178567320108, + 0.07029387354850769, + 0.06454285979270935, + 0.004112914204597473, + 0.002219711896032095, + 0.005456823855638504, + 0.003319761948660016, + 0.012616751715540886, + 0.0065088942646980286, + 0.036266494542360306, + 0.01937367208302021, + 0.009560395032167435, + 0.004991261754184961, + 0.030268436297774315, + 0.005730785895138979, + 0.00725556630641222, + 0.010266559198498726, + 0.008789111860096455, + 0.011415617540478706, + 0.007120628375560045, + 0.009303913451731205, + 0.012165188789367676, + 0.00879369955509901, + 0.009512382559478283, + 0.007100610062479973, + 0.01168124284595251, + 0.007389816455543041, + 0.006357251666486263, + 0.009760056622326374, + 0.0075740148313343525, + 0.0072799548506736755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.0036178866866976023 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043057575821876526 + ], + "opposite_logprob_tokens": [ + -3.168057680130005 + ], + "logprob_reward": -0.043057575821876526, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043057575821876526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040758486837148666, + 0.057565297931432724, + 0.035246238112449646, + 0.03501666337251663, + 0.13898567855358124, + 0.15936177968978882, + 0.1096724346280098, + 0.08408623933792114, + 0.13906918466091156, + 0.20023798942565918 + ], + "credit_map": [ + 0.004843982867896557, + 0.004303393419831991, + 0.008102834224700928, + 0.004952238406985998, + 0.003783601801842451, + 0.004725937731564045, + 0.0038847308605909348, + 0.00616177124902606, + 0.0069815488532185555, + 0.004226469900459051, + 0.013737509958446026, + 0.009280895814299583, + 0.00329450867138803, + 0.0043232436291873455, + 0.005274828989058733, + 0.010446297936141491, + 0.003332719672471285, + 0.002777396934106946, + 0.009001976810395718, + 0.0038361274637281895, + 0.0024917693808674812, + 0.0032720842864364386, + 0.001972497208043933, + 0.008561665192246437, + 0.003255269257351756, + 0.00438347551971674, + 0.004082290921360254, + 0.004701477941125631, + 0.005981130991131067, + 0.0049684783443808556, + 0.0035852876026183367, + 0.0040592532604932785, + 0.01945427991449833, + 0.015385463833808899, + 0.03249628096818924, + 0.026960978284478188, + 0.01254917774349451, + 0.00836382806301117, + 0.010176864452660084, + 0.013598812744021416, + 0.02150014415383339, + 0.02337014675140381, + 0.03210930898785591, + 0.029944537207484245, + 0.012994546443223953, + 0.0133991539478302, + 0.013503588736057281, + 0.012540371157228947, + 0.008905507624149323, + 0.0052528539672493935, + 0.043318457901477814, + 0.03908226639032364, + 0.003363516880199313, + 0.0029028647113591433, + 0.004070044960826635, + 0.0027769282460212708, + 0.011338953860104084, + 0.0064460523426532745, + 0.018244219943881035, + 0.011525403708219528, + 0.005900131072849035, + 0.004544512834399939, + 0.020118407905101776, + 0.005968557205051184, + 0.018950114026665688, + 0.015043158084154129, + 0.02056882157921791, + 0.016731642186641693, + 0.01522956881672144, + 0.008653038181364536, + 0.03369549289345741, + 0.01019736286252737, + 0.03203672915697098, + 0.02681831829249859, + 0.027548721060156822, + 0.02285562828183174, + 0.01838570088148117, + 0.03404499217867851, + 0.01910734549164772, + 0.019440552219748497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.043057575821876526 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot, captured from a horizontal angle, captures the mesmerizing moment of a perfectly ripe, vibrant red strawberry slowly sinking into a glass of crystal-clear liquid. Bright, soft sidelight streams from the side, penetrating the glass and precisely outlining the plump contours and moist sheen of the strawberry, dotted with tiny golden seeds. The moment the strawberry is fully submerged, its edges miraculously dissolve, releasing wisps of red pigment that resemble smoke and mist. These delicate threads gracefully unfurl, swirl, and slowly diffuse within the still liquid, as if living tentacles exploring the transparent space around them, creating a dynamic, abstract watercolor painting. The pure tones of the image, dominated by the strawberry's ruby red, contrast sharply with the transparent liquid and the minimalist background, create an atmosphere of tranquility, concentration, and a surreal, poetic beauty.", + "seed": 2601, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_14/index_1694_seed_2601_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3330136184114963, + "vlm_logprob": -0.3330136184114963, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.3330136184114963, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0037376086693257093 + ], + "opposite_logprob_tokens": [ + -7.003737449645996 + ], + "logprob_reward": -0.0037376086693257093, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0037376086693257093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11621573567390442, + 0.15873120725154877, + 0.13010987639427185, + 0.044429853558540344, + 0.09067011624574661, + 0.14665666222572327, + 0.06905896216630936, + 0.1145908385515213, + 0.0871824324131012, + 0.04235430061817169 + ], + "credit_map": [ + 0.01468822080641985, + 0.010592065751552582, + 0.018214091658592224, + 0.011221901513636112, + 0.015077626332640648, + 0.010976043529808521, + 0.017243802547454834, + 0.01820198819041252, + 0.016155345365405083, + 0.009400545619428158, + 0.04003969579935074, + 0.007858313620090485, + 0.01850852183997631, + 0.04447179287672043, + 0.01083362102508545, + 0.01146337017416954, + 0.02006896585226059, + 0.006305353250354528, + 0.020016133785247803, + 0.007872621528804302, + 0.01946975849568844, + 0.033234745264053345, + 0.009459054097533226, + 0.01368323527276516, + 0.009788401424884796, + 0.0034154346212744713, + 0.007231720723211765, + 0.0033412377815693617, + 0.005797530524432659, + 0.00377048272639513, + 0.007256639655679464, + 0.0038284084293991327, + 0.011816726066172123, + 0.007349621504545212, + 0.01218637079000473, + 0.010512256994843483, + 0.015542090870440006, + 0.00894223339855671, + 0.014217319898307323, + 0.010103501379489899, + 0.024303270503878593, + 0.01586250215768814, + 0.021308472380042076, + 0.014062989503145218, + 0.018196944147348404, + 0.012413624674081802, + 0.02378472313284874, + 0.01672414131462574, + 0.009891844354569912, + 0.006852220743894577, + 0.009031081572175026, + 0.008584829978644848, + 0.011627410538494587, + 0.006199034862220287, + 0.008226195350289345, + 0.008646348491311073, + 0.01839386485517025, + 0.011310871690511703, + 0.016181811690330505, + 0.011667010374367237, + 0.012922069057822227, + 0.010654651559889317, + 0.012866917066276073, + 0.020593641325831413, + 0.018655406311154366, + 0.017725186422467232, + 0.012477145530283451, + 0.008805361576378345, + 0.008779932744801044, + 0.005180469714105129, + 0.00658835144713521, + 0.008970577269792557, + 0.00446501187980175, + 0.004158561583608389, + 0.004750521387904882, + 0.006378322374075651, + 0.006411159411072731, + 0.005360645242035389, + 0.005233746021986008, + 0.00559632945805788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.0037376086693257093 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.60053288936615 + ], + "opposite_logprob_tokens": [ + -0.2255328893661499 + ], + "logprob_reward": -1.60053288936615, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.60053288936615, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05424118787050247, + 0.16771063208580017, + 0.043023377656936646, + 0.05507605895400047, + 0.032029662281274796, + 0.3224135935306549, + 0.18405884504318237, + 0.039998237043619156, + 0.08773896098136902, + 0.013709424994885921 + ], + "credit_map": [ + 0.007278836332261562, + 0.004076747223734856, + 0.010075811296701431, + 0.005426882300525904, + 0.006752949673682451, + 0.007135763764381409, + 0.006272987462580204, + 0.00722120888531208, + 0.01841270737349987, + 0.022278858348727226, + 0.021544720977544785, + 0.01963598094880581, + 0.028321823105216026, + 0.0221727155148983, + 0.020611194893717766, + 0.014732620678842068, + 0.008299227803945541, + 0.0027317481581121683, + 0.005556182004511356, + 0.0036000602412968874, + 0.007059841882437468, + 0.005597120150923729, + 0.006033203098922968, + 0.004145992454141378, + 0.011075153015553951, + 0.003537131706252694, + 0.007482315879315138, + 0.004579589236527681, + 0.007539082318544388, + 0.005843074060976505, + 0.007691397797316313, + 0.007328319363296032, + 0.0057631307281553745, + 0.003164928173646331, + 0.004800289869308472, + 0.0033325408585369587, + 0.005159178748726845, + 0.0031758868135511875, + 0.0038349165115505457, + 0.002798790344968438, + 0.05660839006304741, + 0.03254230320453644, + 0.047994669526815414, + 0.03145519644021988, + 0.0402991846203804, + 0.025860561057925224, + 0.051798172295093536, + 0.03585512563586235, + 0.029502876102924347, + 0.02061486430466175, + 0.021363135427236557, + 0.023053325712680817, + 0.030510639771819115, + 0.01399601623415947, + 0.022256627678871155, + 0.022761372849345207, + 0.008853249251842499, + 0.004085063934326172, + 0.007064462639391422, + 0.0035677601117640734, + 0.004060434643179178, + 0.0028280606493353844, + 0.0038398781325668097, + 0.005699327681213617, + 0.01672336459159851, + 0.01666266657412052, + 0.014280292205512524, + 0.009651180356740952, + 0.008638745173811913, + 0.006232014857232571, + 0.007105693686753511, + 0.008444997482001781, + 0.0017038198420777917, + 0.00154775264672935, + 0.0018115530256181955, + 0.0021209181286394596, + 0.0021846690215170383, + 0.0011298281606286764, + 0.0014867575373500586, + 0.0017241262830793858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -1.60053288936615 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026360824704170227 + ], + "opposite_logprob_tokens": [ + -3.6513607501983643 + ], + "logprob_reward": -0.026360824704170227, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026360824704170227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10346516221761703, + 0.10177236795425415, + 0.13090603053569794, + 0.11665848642587662, + 0.0863606259226799, + 0.2463083565235138, + 0.049705930054187775, + 0.11017873883247375, + 0.03444221243262291, + 0.02020215056836605 + ], + "credit_map": [ + 0.010541521944105625, + 0.0064714448526501656, + 0.02162225916981697, + 0.008918141014873981, + 0.010203098878264427, + 0.01956997811794281, + 0.01105569675564766, + 0.015083017759025097, + 0.009792972356081009, + 0.006192559842020273, + 0.026551898568868637, + 0.006039195694029331, + 0.009672826156020164, + 0.028928281739354134, + 0.006498338654637337, + 0.008096292614936829, + 0.018684791401028633, + 0.005840354599058628, + 0.017995689064264297, + 0.007001474034041166, + 0.018001027405261993, + 0.033022601157426834, + 0.01557881198823452, + 0.01478127483278513, + 0.01398774329572916, + 0.007121636066585779, + 0.015430979430675507, + 0.006857955828309059, + 0.02958533726632595, + 0.013147001154720783, + 0.020774655044078827, + 0.00975317507982254, + 0.013818937353789806, + 0.010621432214975357, + 0.012338102795183659, + 0.007429768331348896, + 0.013669413514435291, + 0.008519943803548813, + 0.010358745232224464, + 0.009604277089238167, + 0.042289238423109055, + 0.026249045506119728, + 0.03589367866516113, + 0.02302386239171028, + 0.030633389949798584, + 0.020609276369214058, + 0.03989102691411972, + 0.027718855068087578, + 0.008011694066226482, + 0.004987258464097977, + 0.0054249526001513, + 0.005923072341829538, + 0.008613713085651398, + 0.004027737770229578, + 0.006287304684519768, + 0.006430193316191435, + 0.015783559530973434, + 0.014840826392173767, + 0.013790231198072433, + 0.012755471281707287, + 0.020564887672662735, + 0.009156537242233753, + 0.012888900935649872, + 0.010398316197097301, + 0.0057031456381082535, + 0.006136983633041382, + 0.004459228832274675, + 0.00433671660721302, + 0.004035079851746559, + 0.002768336096778512, + 0.003107617609202862, + 0.0038951050955802202, + 0.002135737333446741, + 0.0021268308628350496, + 0.0021293144673109055, + 0.0024984641931951046, + 0.003363808384165168, + 0.0027094478718936443, + 0.002578970044851303, + 0.0026595774106681347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.026360824704170227 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464076129719615 + ], + "opposite_logprob_tokens": [ + -5.379640579223633 + ], + "logprob_reward": -0.00464076129719615, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464076129719615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09086412191390991, + 0.16018962860107422, + 0.044766806066036224, + 0.06113925203680992, + 0.18018724024295807, + 0.11701587587594986, + 0.09480517357587814, + 0.12463770061731339, + 0.0629623681306839, + 0.06343185901641846 + ], + "credit_map": [ + 0.014020560309290886, + 0.006559463683515787, + 0.020447801798582077, + 0.00678306445479393, + 0.009085729718208313, + 0.01243294682353735, + 0.009065905585885048, + 0.012468644417822361, + 0.01677335985004902, + 0.019037369638681412, + 0.0313289649784565, + 0.01425551250576973, + 0.018879732117056847, + 0.033037248998880386, + 0.01286744512617588, + 0.014009999111294746, + 0.0067076245322823524, + 0.004493750166147947, + 0.00546259107068181, + 0.0033401043619960546, + 0.008184457197785378, + 0.00734309246763587, + 0.005103714764118195, + 0.004131467547267675, + 0.00789924617856741, + 0.003876562463119626, + 0.007111005485057831, + 0.003945144824683666, + 0.014273473061621189, + 0.006302595604211092, + 0.011321492493152618, + 0.006409728899598122, + 0.023298067972064018, + 0.016529683023691177, + 0.02581246756017208, + 0.018337134271860123, + 0.027993218973279, + 0.021552177146077156, + 0.02415076084434986, + 0.02251371368765831, + 0.021086394786834717, + 0.012207088991999626, + 0.017789818346500397, + 0.011767384596168995, + 0.014272338710725307, + 0.008234016597270966, + 0.01907886378467083, + 0.012579966336488724, + 0.013831637799739838, + 0.01100007351487875, + 0.010634301230311394, + 0.011564175598323345, + 0.016080614179372787, + 0.007761531043797731, + 0.011496364139020443, + 0.012436473742127419, + 0.017521100118756294, + 0.015657005831599236, + 0.01477397046983242, + 0.016247348859906197, + 0.020011749118566513, + 0.012175624258816242, + 0.015320665203034878, + 0.012930228374898434, + 0.010968820191919804, + 0.012007373385131359, + 0.010546951554715633, + 0.007873931899666786, + 0.0066404761746525764, + 0.004776765592396259, + 0.004746344406157732, + 0.0054016984067857265, + 0.00578744150698185, + 0.00664593605324626, + 0.006808249279856682, + 0.00897444412112236, + 0.010095684789121151, + 0.007409512996673584, + 0.008876727893948555, + 0.008833860047161579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.00464076129719615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979600802063942 + ], + "opposite_logprob_tokens": [ + -3.5297961235046387 + ], + "logprob_reward": -0.02979600802063942, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979600802063942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07199983298778534, + 0.1752086877822876, + 0.0660829097032547, + 0.052075762301683426, + 0.18585962057113647, + 0.06236446276307106, + 0.09836017340421677, + 0.15696664154529572, + 0.051766715943813324, + 0.0793151929974556 + ], + "credit_map": [ + 0.008710034191608429, + 0.0036149360239505768, + 0.015179762616753578, + 0.0061978683806955814, + 0.007570281159132719, + 0.010743563063442707, + 0.0071299029514193535, + 0.012853487394750118, + 0.01653124764561653, + 0.018003717064857483, + 0.04271288588643074, + 0.013049132190644741, + 0.01625482551753521, + 0.04454721137881279, + 0.011906200088560581, + 0.01220347173511982, + 0.007831213064491749, + 0.005890711676329374, + 0.007859688252210617, + 0.007078232709318399, + 0.008105467073619366, + 0.012253775261342525, + 0.008632184937596321, + 0.008431638590991497, + 0.007503914646804333, + 0.004240868147462606, + 0.005955643020570278, + 0.0055053504183888435, + 0.0075458260253071785, + 0.005965001415461302, + 0.008118343539536, + 0.007240813225507736, + 0.03090999275445938, + 0.021722611039876938, + 0.028277253732085228, + 0.014545011334121227, + 0.02670411206781864, + 0.017297381535172462, + 0.024070004001259804, + 0.02233324758708477, + 0.009700872004032135, + 0.005100818816572428, + 0.009126474149525166, + 0.0060219475999474525, + 0.011407398618757725, + 0.0065040551126003265, + 0.0077948314137756824, + 0.006708061788231134, + 0.013339858502149582, + 0.011468207463622093, + 0.01099986769258976, + 0.011167480610311031, + 0.016793930903077126, + 0.008647458627820015, + 0.012611485086381435, + 0.013331888243556023, + 0.02310970239341259, + 0.021446818485856056, + 0.018779432401061058, + 0.02126401662826538, + 0.022259211167693138, + 0.014761398546397686, + 0.01823737658560276, + 0.01710868813097477, + 0.009325182065367699, + 0.008309783414006233, + 0.009105093777179718, + 0.005959572270512581, + 0.00596665358170867, + 0.004478299990296364, + 0.003978175576776266, + 0.004643958993256092, + 0.008863980881869793, + 0.011257536709308624, + 0.008982531726360321, + 0.011014429852366447, + 0.012646653689444065, + 0.007587549276649952, + 0.008726238273084164, + 0.010236275382339954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.02979600802063942 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot, captured from a horizontal angle, captures the mesmerizing moment of a perfectly ripe, vibrant red strawberry slowly sinking into a glass of crystal-clear liquid. Bright, soft sidelight streams from the side, penetrating the glass and precisely outlining the plump contours and moist sheen of the strawberry, dotted with tiny golden seeds. The moment the strawberry is fully submerged, its edges miraculously dissolve, releasing wisps of red pigment that resemble smoke and mist. These delicate threads gracefully unfurl, swirl, and slowly diffuse within the still liquid, as if living tentacles exploring the transparent space around them, creating a dynamic, abstract watercolor painting. The pure tones of the image, dominated by the strawberry's ruby red, contrast sharply with the transparent liquid and the minimalist background, create an atmosphere of tranquility, concentration, and a surreal, poetic beauty.", + "seed": 2602, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_14/index_1694_seed_2602_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.26791307469829917, + "vlm_logprob": -0.26791307469829917, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.26791307469829917, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005021579097956419 + ], + "opposite_logprob_tokens": [ + -6.130021572113037 + ], + "logprob_reward": -0.005021579097956419, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.005021579097956419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14771972596645355, + 0.03311208635568619, + 0.09933485835790634, + 0.053456973284482956, + 0.06755474209785461, + 0.09207332879304886, + 0.11803846806287766, + 0.2535255551338196, + 0.09194212406873703, + 0.043242163956165314 + ], + "credit_map": [ + 0.01751074753701687, + 0.01730787754058838, + 0.015097849071025848, + 0.0201070886105299, + 0.017850875854492188, + 0.019902003929018974, + 0.02152852527797222, + 0.018414754420518875, + 0.0033881328999996185, + 0.0033839289098978043, + 0.00501730153337121, + 0.00462219538167119, + 0.0033329129219055176, + 0.004128619097173214, + 0.004622264765202999, + 0.004616730380803347, + 0.009295754134654999, + 0.011484222486615181, + 0.010691933333873749, + 0.01754038780927658, + 0.013839131221175194, + 0.013206050731241703, + 0.013645213097333908, + 0.009632156230509281, + 0.004837634973227978, + 0.004671440459787846, + 0.008278745226562023, + 0.005947996396571398, + 0.0065274350345134735, + 0.006761969067156315, + 0.008181427605450153, + 0.008250328712165356, + 0.01083141379058361, + 0.005282426252961159, + 0.012699798680841923, + 0.005837045144289732, + 0.005288129206746817, + 0.005706241354346275, + 0.008369685150682926, + 0.01354000624269247, + 0.007090490311384201, + 0.012006370350718498, + 0.00943027064204216, + 0.013804065063595772, + 0.013800513930618763, + 0.009051267988979816, + 0.01631588116288185, + 0.010574473068118095, + 0.01573295146226883, + 0.012955276295542717, + 0.01881069503724575, + 0.014839954674243927, + 0.014403088949620724, + 0.010609744116663933, + 0.012470923364162445, + 0.018215827643871307, + 0.026087773963809013, + 0.04295943304896355, + 0.03142320737242699, + 0.022524012252688408, + 0.02320084162056446, + 0.02198239602148533, + 0.02751108631491661, + 0.057836804538965225, + 0.00882736500352621, + 0.01244302187114954, + 0.01584702357649803, + 0.010773253627121449, + 0.004936222452670336, + 0.004744780249893665, + 0.014469819143414497, + 0.01990063488483429, + 0.004217189271003008, + 0.007386502344161272, + 0.0071784271858632565, + 0.004481012467294931, + 0.005876665934920311, + 0.004204882308840752, + 0.0048243338242173195, + 0.005073146894574165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.005021579097956419 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236173152923584 + ], + "opposite_logprob_tokens": [ + -0.348617285490036 + ], + "logprob_reward": -1.2236173152923584, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2236173152923584, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08047407120466232, + 0.048622015863657, + 0.09145259112119675, + 0.07192100584506989, + 0.0778178796172142, + 0.15630851686000824, + 0.22035746276378632, + 0.1180654987692833, + 0.12285835295915604, + 0.012122644111514091 + ], + "credit_map": [ + 0.0109957754611969, + 0.012146271765232086, + 0.006229233928024769, + 0.012375684455037117, + 0.008081243373453617, + 0.009490384720265865, + 0.013029612600803375, + 0.008125867694616318, + 0.002913612639531493, + 0.0038446199614554644, + 0.005108675919473171, + 0.009923934936523438, + 0.005641140043735504, + 0.004498512018471956, + 0.01157417707145214, + 0.005117340944707394, + 0.009650020860135555, + 0.008150284178555012, + 0.012272948399186134, + 0.011951315216720104, + 0.007733759470283985, + 0.019353792071342468, + 0.012062035501003265, + 0.010278433561325073, + 0.007011679466813803, + 0.005677948240190744, + 0.012194869108498096, + 0.007761096116155386, + 0.00850654672831297, + 0.009083072654902935, + 0.00872582197189331, + 0.01295996643602848, + 0.008790857158601284, + 0.007973159663379192, + 0.010370699688792229, + 0.007778694853186607, + 0.007132027298212051, + 0.006036375183612108, + 0.01514496561139822, + 0.014591099694371223, + 0.011732262559235096, + 0.02470305934548378, + 0.016232425346970558, + 0.025827620178461075, + 0.020689668133854866, + 0.010332978330552578, + 0.024482404813170433, + 0.022308088839054108, + 0.02794540487229824, + 0.02077522687613964, + 0.027985109016299248, + 0.027568211778998375, + 0.020131664350628853, + 0.022546561434864998, + 0.02928597293794155, + 0.04411930963397026, + 0.01177991647273302, + 0.017759760841727257, + 0.015132377855479717, + 0.014932245947420597, + 0.01065068133175373, + 0.013222793117165565, + 0.014160502701997757, + 0.020427213981747627, + 0.009002390317618847, + 0.010987918823957443, + 0.027330048382282257, + 0.011319145560264587, + 0.005323559511452913, + 0.0053770276717841625, + 0.015714090317487717, + 0.03780417889356613, + 0.00105680909473449, + 0.002314419951289892, + 0.0017318857135251164, + 0.0012604338116943836, + 0.001628437195904553, + 0.0008747398387640715, + 0.0019155589397996664, + 0.0013403596822172403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -1.2236173152923584 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900112122297287 + ], + "opposite_logprob_tokens": [ + -2.5790011882781982 + ], + "logprob_reward": -0.07900112122297287, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900112122297287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1214510127902031, + 0.03847215324640274, + 0.18353019654750824, + 0.10600478947162628, + 0.25448715686798096, + 0.07107796519994736, + 0.10235969722270966, + 0.07161881029605865, + 0.026088090613484383, + 0.02491012029349804 + ], + "credit_map": [ + 0.014349518343806267, + 0.014847470447421074, + 0.00916431937366724, + 0.01878235675394535, + 0.0162037443369627, + 0.015112672932446003, + 0.02082412876188755, + 0.012166804634034634, + 0.0023127750027924776, + 0.0029956998769193888, + 0.004741554148495197, + 0.005106116645038128, + 0.004273283760994673, + 0.0076045081950724125, + 0.003908833488821983, + 0.007529381662607193, + 0.013272793032228947, + 0.01906483992934227, + 0.0203256718814373, + 0.02833838202059269, + 0.028214771300554276, + 0.02627221681177616, + 0.03308757767081261, + 0.01495394203811884, + 0.01521674171090126, + 0.010310627520084381, + 0.013678321614861488, + 0.012205682694911957, + 0.012588568963110447, + 0.006857418455183506, + 0.01850312016904354, + 0.016644304618239403, + 0.02705511450767517, + 0.040587421506643295, + 0.03689761087298393, + 0.033591385930776596, + 0.02360828034579754, + 0.021526997908949852, + 0.03454946354031563, + 0.03667087107896805, + 0.005884052254259586, + 0.009255961515009403, + 0.008017020300030708, + 0.007660066708922386, + 0.011014645919203758, + 0.006374412216246128, + 0.01312548853456974, + 0.009746316820383072, + 0.011392900720238686, + 0.012403309345245361, + 0.014298521913588047, + 0.006656619254499674, + 0.016326947137713432, + 0.015168366022408009, + 0.011531815864145756, + 0.014581217430531979, + 0.007207311224192381, + 0.011504494585096836, + 0.007046651095151901, + 0.008818607777357101, + 0.007816639728844166, + 0.008273749612271786, + 0.01155395619571209, + 0.009397393092513084, + 0.0016077248146757483, + 0.002340433420613408, + 0.0062107727862894535, + 0.0021838583052158356, + 0.0013191178441047668, + 0.0012996156001463532, + 0.0025952099822461605, + 0.0085313580930233, + 0.002220823895186186, + 0.00453867157921195, + 0.004453026223927736, + 0.0024805604480206966, + 0.003351704915985465, + 0.0018196922028437257, + 0.0029035930056124926, + 0.0031420490704476833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.07900112122297287 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642058819532394 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642058819532394, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642058819532394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13586853444576263, + 0.051527295261621475, + 0.13906453549861908, + 0.09717612713575363, + 0.17689856886863708, + 0.12477204948663712, + 0.10096945613622665, + 0.08578835427761078, + 0.058696992695331573, + 0.02923808991909027 + ], + "credit_map": [ + 0.014388221316039562, + 0.017892124131321907, + 0.009506220929324627, + 0.023196659982204437, + 0.012677623890340328, + 0.018033882603049278, + 0.02523794025182724, + 0.014935851097106934, + 0.004904902074486017, + 0.006152440793812275, + 0.004170268774032593, + 0.009457618929445744, + 0.0062383501790463924, + 0.005267319735139608, + 0.009168769232928753, + 0.0061676232144236565, + 0.009399829432368279, + 0.015136636793613434, + 0.01576017215847969, + 0.018333662301301956, + 0.022776702418923378, + 0.02190352790057659, + 0.023599043488502502, + 0.012154965661466122, + 0.011713676154613495, + 0.00982748344540596, + 0.013643374666571617, + 0.01167174894362688, + 0.01100707147270441, + 0.007201585453003645, + 0.017060235142707825, + 0.015050950460135937, + 0.021040676161646843, + 0.019824279472231865, + 0.03275880962610245, + 0.019324595108628273, + 0.015252619981765747, + 0.011941101402044296, + 0.020902877673506737, + 0.03585360571742058, + 0.009260126389563084, + 0.02454029954969883, + 0.008510681800544262, + 0.022274168208241463, + 0.019897550344467163, + 0.008200664073228836, + 0.01585291139781475, + 0.01623563840985298, + 0.012806594371795654, + 0.01230231486260891, + 0.014077425934374332, + 0.009429224766790867, + 0.013329599983990192, + 0.01614227704703808, + 0.01046172995120287, + 0.012420286424458027, + 0.008223463781177998, + 0.01222330704331398, + 0.010492187924683094, + 0.007888558320701122, + 0.008748350664973259, + 0.008054244332015514, + 0.008841174654662609, + 0.02131706289947033, + 0.005669751670211554, + 0.006489788647741079, + 0.010259607806801796, + 0.005953243002295494, + 0.004125115927308798, + 0.003244247520342469, + 0.009242625907063484, + 0.013712611049413681, + 0.0031265115831047297, + 0.006971112452447414, + 0.0033940246794372797, + 0.002603900386020541, + 0.00323867192491889, + 0.002609424525871873, + 0.0036247451789677143, + 0.0036696987226605415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.008642058819532394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a vibrant red strawberry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static and captured from a horizontal angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the strawberry slowly sink into a glass of crystal-clear liquid?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do wisps of red pigment resembling smoke and mist dissolve from the strawberry upon submersion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background minimalist and dominated by the strawberry's ruby red tones?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283299058675766 + ], + "opposite_logprob_tokens": [ + -3.7732832431793213 + ], + "logprob_reward": -0.023283299058675766, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283299058675766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11868105828762054, + 0.08929643779993057, + 0.06281644850969315, + 0.13100287318229675, + 0.12242297828197479, + 0.12543708086013794, + 0.08115039765834808, + 0.040500838309526443, + 0.09732718020677567, + 0.13136465847492218 + ], + "credit_map": [ + 0.011063867248594761, + 0.013666183687746525, + 0.012959050014615059, + 0.01682393252849579, + 0.012615988962352276, + 0.015188547782599926, + 0.02406400255858898, + 0.012299488298594952, + 0.004840037319809198, + 0.006426356732845306, + 0.013975732028484344, + 0.010491200722754002, + 0.009945802390575409, + 0.0213057529181242, + 0.010238372720777988, + 0.01207318902015686, + 0.005871118046343327, + 0.007641349453479052, + 0.008282857947051525, + 0.006781759671866894, + 0.006253930740058422, + 0.013904757797718048, + 0.00761184049770236, + 0.006468841340392828, + 0.010264618322253227, + 0.01054010447114706, + 0.03838466480374336, + 0.00918523408472538, + 0.009637173265218735, + 0.008510480634868145, + 0.011649923399090767, + 0.03283068165183067, + 0.012101100757718086, + 0.0182527843862772, + 0.01729041337966919, + 0.017075154930353165, + 0.012638816609978676, + 0.009305328130722046, + 0.019370321184396744, + 0.016389062628149986, + 0.009082893840968609, + 0.023714162409305573, + 0.01303086057305336, + 0.012721164152026176, + 0.018007947131991386, + 0.011117551475763321, + 0.016375934705138206, + 0.02138656936585903, + 0.01233998965471983, + 0.010211245156824589, + 0.010789782740175724, + 0.010819528251886368, + 0.010047907941043377, + 0.009148365817964077, + 0.007906140759587288, + 0.00988744292408228, + 0.0035282073076814413, + 0.0072539434768259525, + 0.004176149144768715, + 0.0050434242002666, + 0.00457428302615881, + 0.005496281664818525, + 0.005181456450372934, + 0.005247097462415695, + 0.008051790297031403, + 0.00794152356684208, + 0.024579627439379692, + 0.0066765714436769485, + 0.004898503422737122, + 0.004078458063304424, + 0.007213588338345289, + 0.033887118101119995, + 0.015657706186175346, + 0.03382789343595505, + 0.01733383908867836, + 0.013158470392227173, + 0.013417176902294159, + 0.007602507248520851, + 0.016520937904715538, + 0.01384612638503313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 5 + ], + "selected_reward": -0.023283299058675766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_15/index_tensor([6049, 6049])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_15/index_tensor([6049, 6049])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..849af7320d4bd0e8cb29c02bba0b97081852f011 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_15/index_tensor([6049, 6049])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, low-angle, wide shot places us at a busy urban intersection at dusk. A police officer in his forties stands steady in the center of the frame, his face stern and his focused gaze beneath neatly trimmed black hair, revealing a time-tested composure. He wears a crisp, dark blue uniform, topped by a fluorescent yellow reflective vest that stands out against the darkening sky. Its silver reflective stripes flicker in the light of car lights. Around him, the city moves at breakneck speed: towering glass skyscrapers reflect the last orange-red glow of the setting sun. On the ground, a stream of yellow taxis, red buses, and private cars of various colors alternates between signals and his direction, creating a motion blur of colored streaks. The sidewalks are thronged with hurried pedestrians. His movements are precise and powerful. One white-gloved hand rises decisively, signaling the traffic in front of him to stop. Then, with a slight turn of his body, his other hand swings sideways in a smooth, crisp motion, directing the vehicles passing sideways. The soft golden afterglow outlines his tall silhouette, and the car lights and the neon lights cast colorful spots of light on the buildings behind him. The whole scene, with the cold gray tones of the city and the warm afterglow of the sunset, creates a tense, orderly and grand atmosphere full of urban vitality.", + "seed": 13908, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_15/index_6049_seed_13908_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.022198439948260783, + "vlm_logprob": -0.022198439948260783, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.022198439948260783, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043700918555259705 + ], + "opposite_logprob_tokens": [ + -3.793700933456421 + ], + "logprob_reward": -0.043700918555259705, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.043700918555259705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14049609005451202, + 0.036845747381448746, + 0.04760992154479027, + 0.07352471351623535, + 0.04434085264801979, + 0.023227233439683914, + 0.053853362798690796, + 0.026236023753881454, + 0.0333758182823658, + 0.5204902291297913 + ], + "credit_map": [ + 0.030765017494559288, + 0.028282757848501205, + 0.020962834358215332, + 0.017457667738199234, + 0.011320249177515507, + 0.011964963749051094, + 0.010246848687529564, + 0.009495755657553673, + 0.0071481987833976746, + 0.005173266399651766, + 0.004826671443879604, + 0.006094899494200945, + 0.0029418705962598324, + 0.004550878424197435, + 0.002269668271765113, + 0.003840296296402812, + 0.009502765722572803, + 0.006896758917719126, + 0.008697250857949257, + 0.006190169136971235, + 0.0034746930468827486, + 0.004970022011548281, + 0.0022142394445836544, + 0.005664023570716381, + 0.01954479143023491, + 0.004804869648069143, + 0.006273431237787008, + 0.02389911748468876, + 0.002514554187655449, + 0.00716813188046217, + 0.0022057995665818453, + 0.00711401691660285, + 0.007956808432936668, + 0.0030278139747679234, + 0.008391465060412884, + 0.005268116015940905, + 0.002435509115457535, + 0.005553824827075005, + 0.0023455035407096148, + 0.0093618119135499, + 0.0036300343926995993, + 0.0022592684254050255, + 0.0024473911616951227, + 0.0024746404960751534, + 0.0029505884740501642, + 0.002972828457131982, + 0.0018346403958275914, + 0.00465784128755331, + 0.006288099568337202, + 0.006793387234210968, + 0.005506195593625307, + 0.01300081331282854, + 0.0044942633248865604, + 0.006218684371560812, + 0.006180249620229006, + 0.005371673498302698, + 0.0036807116121053696, + 0.0027293278835713863, + 0.0031698294915258884, + 0.004118598066270351, + 0.002463194075971842, + 0.0033265589736402035, + 0.0025637915823608637, + 0.004184011835604906, + 0.009747335687279701, + 0.0062264930456876755, + 0.0032312741968780756, + 0.0028979210183024406, + 0.0019559264183044434, + 0.004349976312369108, + 0.0021194317378103733, + 0.0028474589344114065, + 0.1826302856206894, + 0.021505078300833702, + 0.19162733852863312, + 0.03403576463460922, + 0.017176428809762, + 0.04657716676592827, + 0.01031110156327486, + 0.01662706770002842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.043700918555259705 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009815051220357418 + ], + "opposite_logprob_tokens": [ + -4.634815216064453 + ], + "logprob_reward": -0.009815051220357418, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009815051220357418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045132506638765335, + 0.019669020548462868, + 0.016944751143455505, + 0.09620244055986404, + 0.034753184765577316, + 0.05272657424211502, + 0.034493450075387955, + 0.00963529758155346, + 0.025517717003822327, + 0.6649250984191895 + ], + "credit_map": [ + 0.0070780362002551556, + 0.00576719269156456, + 0.008042288944125175, + 0.006844138260930777, + 0.0036808550357818604, + 0.003797413781285286, + 0.0053147380240261555, + 0.004607842303812504, + 0.002516126958653331, + 0.0017961235716938972, + 0.0025251663755625486, + 0.003185340203344822, + 0.0029004556126892567, + 0.002153144683688879, + 0.002017345977947116, + 0.0025753166992217302, + 0.002047738991677761, + 0.002395695075392723, + 0.002327949507161975, + 0.002025092486292124, + 0.0025899154134094715, + 0.002015366218984127, + 0.0013544987887144089, + 0.0021884930320084095, + 0.025727640837430954, + 0.005096430890262127, + 0.008110778406262398, + 0.034721288830041885, + 0.003382159397006035, + 0.005936348810791969, + 0.0034069207031279802, + 0.009820867329835892, + 0.006732565816491842, + 0.0026363253127783537, + 0.005126641131937504, + 0.003496660152450204, + 0.0027342208195477724, + 0.004556278698146343, + 0.0016800076700747013, + 0.0077904826030135155, + 0.005745973438024521, + 0.005035131238400936, + 0.003985797520726919, + 0.006451214198023081, + 0.007479249034076929, + 0.0056600128300487995, + 0.011525457724928856, + 0.006843734998255968, + 0.005806298460811377, + 0.003478993894532323, + 0.003254358656704426, + 0.005256298463791609, + 0.005056893453001976, + 0.004467551130801439, + 0.0031107955146580935, + 0.004062257707118988, + 0.002346033463254571, + 0.001676803338341415, + 0.0011744730873033404, + 0.001270196633413434, + 0.0005763064255006611, + 0.000967090018093586, + 0.0007329168147407472, + 0.0008914771024137735, + 0.0069757308810949326, + 0.004712997004389763, + 0.003001803532242775, + 0.0017159519484266639, + 0.0018681064248085022, + 0.003331258660182357, + 0.001981498207896948, + 0.001930368016473949, + 0.2404579371213913, + 0.026718351989984512, + 0.24201524257659912, + 0.04068644717335701, + 0.02152932435274124, + 0.0594768226146698, + 0.012077863328158855, + 0.021963069215416908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 5 + ], + "selected_reward": -0.009815051220357418 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305963218212128 + ], + "opposite_logprob_tokens": [ + -3.1680595874786377 + ], + "logprob_reward": -0.04305963218212128, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305963218212128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040961239486932755, + 0.03307754173874855, + 0.010570494458079338, + 0.02516724169254303, + 0.057271674275398254, + 0.0273822620511055, + 0.019334517419338226, + 0.023963432759046555, + 0.015670187771320343, + 0.7466014623641968 + ], + "credit_map": [ + 0.007143757306039333, + 0.00898783653974533, + 0.005415155552327633, + 0.005439569242298603, + 0.003288262989372015, + 0.003962796181440353, + 0.003113536862656474, + 0.003610320622101426, + 0.005026847589761019, + 0.0030279161874204874, + 0.004577538464218378, + 0.0074264355935156345, + 0.0019005623180419207, + 0.00542752118781209, + 0.0019871194381266832, + 0.0037035970017313957, + 0.0015550687676295638, + 0.0018156591104343534, + 0.0012385528534650803, + 0.0017647026106715202, + 0.001060193870216608, + 0.0010241026757284999, + 0.0009649255080148578, + 0.0011472883634269238, + 0.00672919349744916, + 0.0017104680882766843, + 0.002303095068782568, + 0.007149109151214361, + 0.001524072140455246, + 0.0020522482227534056, + 0.0009531463147141039, + 0.0027459049597382545, + 0.009843365289270878, + 0.005868173204362392, + 0.005510686431080103, + 0.008171619847416878, + 0.005136770661920309, + 0.0061347708106040955, + 0.00401300098747015, + 0.012593282386660576, + 0.004740519914776087, + 0.0031605742406100035, + 0.0028608872089535, + 0.002441300777718425, + 0.0032887542620301247, + 0.0030566055793315172, + 0.0030389446765184402, + 0.004794672597199678, + 0.002431332366541028, + 0.00325982179492712, + 0.002459288341924548, + 0.002059563994407654, + 0.002754248445853591, + 0.0027129503432661295, + 0.0018572992412373424, + 0.0018000102136284113, + 0.0041575441136956215, + 0.0033395157661288977, + 0.0019446563674136996, + 0.005124666262418032, + 0.0018165201181545854, + 0.003405314404517412, + 0.0019749044440686703, + 0.0022003096528351307, + 0.002381237456575036, + 0.001754579832777381, + 0.002406657673418522, + 0.002243979834020138, + 0.0015064621111378074, + 0.0017663793405517936, + 0.001988417934626341, + 0.001622471958398819, + 0.2677755653858185, + 0.029814058914780617, + 0.2737348973751068, + 0.04681273177266121, + 0.024107705801725388, + 0.06683363765478134, + 0.0136733902618289, + 0.02384941279888153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.04305963218212128 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780345484614372 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780345484614372, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780345484614372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.011645580641925335, + 0.009270881302654743, + 0.005363134201616049, + 0.010125499218702316, + 0.01463504321873188, + 0.011426230892539024, + 0.014451623894274235, + 0.010680481791496277, + 0.038648225367069244, + 0.8737533688545227 + ], + "credit_map": [ + 0.0019031917909160256, + 0.0019177669892087579, + 0.0015477035194635391, + 0.002256997860968113, + 0.0012512593530118465, + 0.0008655567653477192, + 0.0009393958025611937, + 0.0009637079783715308, + 0.001633346895687282, + 0.0011194896651431918, + 0.0011302385246381164, + 0.0018920137081295252, + 0.0008202468161471188, + 0.0009115457651205361, + 0.0005584608297795057, + 0.0012055381666868925, + 0.0008990358910523355, + 0.0006199615891091526, + 0.001051826635375619, + 0.0007433191640302539, + 0.0004242942959535867, + 0.0007361497264355421, + 0.00030509228236041963, + 0.000583454268053174, + 0.0016616318607702851, + 0.000948733591940254, + 0.001102956710383296, + 0.00243956595659256, + 0.0007376260473392904, + 0.0014395714970305562, + 0.0004662755236495286, + 0.0013291372451931238, + 0.0021978148724883795, + 0.0017949582543224096, + 0.0025486862286925316, + 0.0021681420039385557, + 0.0009322430123575032, + 0.0018808846361935139, + 0.0010689041810110211, + 0.0020434088073670864, + 0.0013283583102747798, + 0.0010571409948170185, + 0.0009809861658141017, + 0.0012287640711292624, + 0.0013864979846403003, + 0.0011105636367574334, + 0.0029034209437668324, + 0.0014304976211860776, + 0.0023365854285657406, + 0.0018389204051345587, + 0.0018871633801609278, + 0.0021026786416769028, + 0.0014668647199869156, + 0.0020488505251705647, + 0.0015413520159199834, + 0.001229206915013492, + 0.002320781582966447, + 0.0022641608957201242, + 0.0015706609701737761, + 0.0016853440320119262, + 0.0006031068041920662, + 0.0010917511535808444, + 0.0005476935184560716, + 0.0005969806807115674, + 0.006395426578819752, + 0.00873600970953703, + 0.006469707936048508, + 0.004319500178098679, + 0.0027030380442738533, + 0.004928042180836201, + 0.00311086792498827, + 0.001985625596717, + 0.31399422883987427, + 0.035040803253650665, + 0.3201664686203003, + 0.05500402674078941, + 0.02832203172147274, + 0.07856146991252899, + 0.01575785130262375, + 0.0269064512103796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.009780345484614372 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636252298951149 + ], + "opposite_logprob_tokens": [ + -5.379636287689209 + ], + "logprob_reward": -0.004636252298951149, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636252298951149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04654551297426224, + 0.011649908497929573, + 0.013020687736570835, + 0.028678251430392265, + 0.007822834886610508, + 0.00903706531971693, + 0.018417207524180412, + 0.016360364854335785, + 0.02233206294476986, + 0.8261361122131348 + ], + "credit_map": [ + 0.008519014343619347, + 0.010028315708041191, + 0.007570974063128233, + 0.007320096250623465, + 0.0023317078594118357, + 0.004013862926512957, + 0.0038847229443490505, + 0.002876825165003538, + 0.0020584561862051487, + 0.0012637850595638156, + 0.0014563048025593162, + 0.0023550516925752163, + 0.0010142326354980469, + 0.0014264447381719947, + 0.0006719647790305316, + 0.0014036699431017041, + 0.0029482883401215076, + 0.0014373083831742406, + 0.002553578233346343, + 0.001847358769737184, + 0.0007828439702279866, + 0.001503785140812397, + 0.0005984373856335878, + 0.001349088503047824, + 0.005959609989076853, + 0.0020315770525485277, + 0.0025752061046659946, + 0.007703930605202913, + 0.0015077905263751745, + 0.004173805005848408, + 0.0011226610513404012, + 0.0036036737728863955, + 0.0008632844546809793, + 0.0008720809128135443, + 0.0011239111190661788, + 0.001294981804676354, + 0.00066656299168244, + 0.0013080888893455267, + 0.0006385677843354642, + 0.001055357395671308, + 0.0015286736888810992, + 0.0009777615778148174, + 0.0010221650591120124, + 0.0008943374268710613, + 0.0010080091888085008, + 0.0009257173514924943, + 0.001585301011800766, + 0.001095099956728518, + 0.003196181496605277, + 0.0020335763692855835, + 0.0027874393854290247, + 0.0016924701631069183, + 0.002355538308620453, + 0.0023105198051780462, + 0.002021933440119028, + 0.0020195513498038054, + 0.0022342668380588293, + 0.0029519726522266865, + 0.001411955337971449, + 0.0030694834422320127, + 0.001300686039030552, + 0.002478115726262331, + 0.0014319964684545994, + 0.001481891144067049, + 0.003988184500485659, + 0.005129819270223379, + 0.0028163271490484476, + 0.002557395026087761, + 0.002036829013377428, + 0.002797262277454138, + 0.001448363997042179, + 0.0015578828752040863, + 0.29728618264198303, + 0.03397369384765625, + 0.30198946595191956, + 0.052286434918642044, + 0.02647276222705841, + 0.07428675144910812, + 0.014845345169305801, + 0.024995537474751472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.004636252298951149 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle, wide shot places us at a busy urban intersection at dusk. A police officer in his forties stands steady in the center of the frame, his face stern and his focused gaze beneath neatly trimmed black hair, revealing a time-tested composure. He wears a crisp, dark blue uniform, topped by a fluorescent yellow reflective vest that stands out against the darkening sky. Its silver reflective stripes flicker in the light of car lights. Around him, the city moves at breakneck speed: towering glass skyscrapers reflect the last orange-red glow of the setting sun. On the ground, a stream of yellow taxis, red buses, and private cars of various colors alternates between signals and his direction, creating a motion blur of colored streaks. The sidewalks are thronged with hurried pedestrians. His movements are precise and powerful. One white-gloved hand rises decisively, signaling the traffic in front of him to stop. Then, with a slight turn of his body, his other hand swings sideways in a smooth, crisp motion, directing the vehicles passing sideways. The soft golden afterglow outlines his tall silhouette, and the car lights and the neon lights cast colorful spots of light on the buildings behind him. The whole scene, with the cold gray tones of the city and the warm afterglow of the sunset, creates a tense, orderly and grand atmosphere full of urban vitality.", + "seed": 13909, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_15/index_6049_seed_13909_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.021605563629418612, + "vlm_logprob": -0.021605563629418612, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.021605563629418612, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0372939370572567 + ], + "opposite_logprob_tokens": [ + -5.662293910980225 + ], + "logprob_reward": -0.0372939370572567, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0372939370572567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13285110890865326, + 0.04076552391052246, + 0.03022068738937378, + 0.045718152076005936, + 0.043846502900123596, + 0.0527348667383194, + 0.07078810781240463, + 0.44957947731018066, + 0.07397829741239548, + 0.05951730161905289 + ], + "credit_map": [ + 0.01345575600862503, + 0.01997211202979088, + 0.01691599003970623, + 0.011499688029289246, + 0.017622584477066994, + 0.016382182016968727, + 0.01169378962367773, + 0.025309015065431595, + 0.004189732484519482, + 0.005848588887602091, + 0.0072019100189208984, + 0.0038509115111082792, + 0.006812731269747019, + 0.005158172454684973, + 0.0034217338543385267, + 0.004281741566956043, + 0.0025762221775949, + 0.0030845687724649906, + 0.004579741042107344, + 0.003168016206473112, + 0.005369571968913078, + 0.004558324348181486, + 0.0032273156102746725, + 0.0036569295916706324, + 0.004966416396200657, + 0.006282724440097809, + 0.005528241395950317, + 0.004545947536826134, + 0.00564051466062665, + 0.009782417677342892, + 0.002681681653484702, + 0.006290210876613855, + 0.003923444543033838, + 0.0022015066351741552, + 0.007368794176727533, + 0.003653577994555235, + 0.006039559841156006, + 0.0064735496416687965, + 0.007595799397677183, + 0.006590269505977631, + 0.0042286221869289875, + 0.005399195943027735, + 0.004709500353783369, + 0.008817320689558983, + 0.00824699830263853, + 0.005636736284941435, + 0.009092139080166817, + 0.006604354828596115, + 0.005323399789631367, + 0.006429606582969427, + 0.005396017339080572, + 0.009709169156849384, + 0.012167518027126789, + 0.009982195682823658, + 0.012788143940269947, + 0.008992058224976063, + 0.07763988524675369, + 0.017049573361873627, + 0.024577394127845764, + 0.03445088118314743, + 0.05164838209748268, + 0.1732385754585266, + 0.013594305142760277, + 0.05738045647740364, + 0.006656000390648842, + 0.016266852617263794, + 0.00742728728801012, + 0.007089036516845226, + 0.012890328653156757, + 0.011134197004139423, + 0.005120187532156706, + 0.007394406478852034, + 0.004449551925063133, + 0.005425073206424713, + 0.004766365978866816, + 0.008937261067330837, + 0.010441756807267666, + 0.009371234104037285, + 0.008993292227387428, + 0.007132768630981445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.0372939370572567 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008678458631038666 + ], + "opposite_logprob_tokens": [ + -4.758678436279297 + ], + "logprob_reward": -0.008678458631038666, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008678458631038666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12282148003578186, + 0.051741596311330795, + 0.04511208087205887, + 0.05586089938879013, + 0.10680428147315979, + 0.03887074068188667, + 0.054005566984415054, + 0.35327860713005066, + 0.10130186378955841, + 0.07020293176174164 + ], + "credit_map": [ + 0.011354156769812107, + 0.021934928372502327, + 0.011790801770985126, + 0.013781608082354069, + 0.021185357123613358, + 0.019970593973994255, + 0.009929015301167965, + 0.012875000014901161, + 0.003927099518477917, + 0.007367288693785667, + 0.006943230517208576, + 0.003629680722951889, + 0.011288795620203018, + 0.007089223712682724, + 0.005844210274517536, + 0.005652063060551882, + 0.004127822816371918, + 0.0035100693348795176, + 0.004703475162386894, + 0.002820307621732354, + 0.011698809452354908, + 0.008558925241231918, + 0.0053208963945508, + 0.004371772054582834, + 0.005231933202594519, + 0.006887142080813646, + 0.004748471546918154, + 0.00961209088563919, + 0.00869415607303381, + 0.007348590064793825, + 0.00765484431758523, + 0.005683663301169872, + 0.006743657868355513, + 0.004381986800581217, + 0.0038649095222353935, + 0.005471454467624426, + 0.011356107890605927, + 0.03226685896515846, + 0.012256600894033909, + 0.03046269528567791, + 0.004712865687906742, + 0.004054879769682884, + 0.004243610892444849, + 0.0062581393867731094, + 0.004381120670586824, + 0.005897899623960257, + 0.004651480354368687, + 0.004670741967856884, + 0.004580933600664139, + 0.00265790824778378, + 0.003077596193179488, + 0.005557911936193705, + 0.008916138671338558, + 0.006445383187383413, + 0.008836019784212112, + 0.013933668844401836, + 0.04800926148891449, + 0.01236503291875124, + 0.013923927210271358, + 0.023137608543038368, + 0.0813223272562027, + 0.13427145779132843, + 0.008752998895943165, + 0.03149596229195595, + 0.008750855922698975, + 0.013026940636336803, + 0.006796142086386681, + 0.00640137679874897, + 0.025754272937774658, + 0.020872440189123154, + 0.007208296097815037, + 0.012491527944803238, + 0.008061998523771763, + 0.007063189055770636, + 0.007454770151525736, + 0.009114004671573639, + 0.009753938764333725, + 0.010766889899969101, + 0.008267286233603954, + 0.009720846079289913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.008678458631038666 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865909367799759 + ], + "opposite_logprob_tokens": [ + -3.048659086227417 + ], + "logprob_reward": -0.04865909367799759, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865909367799759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03656851127743721, + 0.016181010752916336, + 0.020147088915109634, + 0.019936997443437576, + 0.023459704592823982, + 0.01857314258813858, + 0.03905598446726799, + 0.6850427389144897, + 0.11950574815273285, + 0.021529005840420723 + ], + "credit_map": [ + 0.0026320721954107285, + 0.004184553865343332, + 0.004182774573564529, + 0.004005326423794031, + 0.004246043041348457, + 0.005405132193118334, + 0.0036987068597227335, + 0.008213900960981846, + 0.001303342985920608, + 0.0023822938092052937, + 0.0024591642431914806, + 0.001893827342428267, + 0.0023894880432635546, + 0.002106262370944023, + 0.0010167668806388974, + 0.002629866125062108, + 0.0033301671501249075, + 0.0015602377243340015, + 0.002015352714806795, + 0.0015238283667713404, + 0.0023946003057062626, + 0.006230509839951992, + 0.0011262120679020882, + 0.0019661798141896725, + 0.0018341096583753824, + 0.0033844676800072193, + 0.002475780202075839, + 0.002369282767176628, + 0.003098617075011134, + 0.0026498124934732914, + 0.0018686521798372269, + 0.002256276085972786, + 0.0012112456606701016, + 0.0030308878049254417, + 0.001658282009884715, + 0.0016648141900077462, + 0.0027494882233440876, + 0.005816280376166105, + 0.0011303755454719067, + 0.006198328919708729, + 0.0018832996720448136, + 0.0019886631052941084, + 0.0025517819449305534, + 0.0018622205825522542, + 0.003222319297492504, + 0.002118524629622698, + 0.002472120337188244, + 0.0024742144159972668, + 0.0033185044303536415, + 0.0017806736286729574, + 0.0029840825591236353, + 0.004006492905318737, + 0.007741972338408232, + 0.005077739246189594, + 0.0031037761364132166, + 0.011042742989957333, + 0.09443394094705582, + 0.022323979064822197, + 0.025397028774023056, + 0.04102178290486336, + 0.11118530482053757, + 0.2705610990524292, + 0.013222947716712952, + 0.10689663141965866, + 0.008406187407672405, + 0.013432134874165058, + 0.008803434669971466, + 0.013962620869278908, + 0.016143565997481346, + 0.016599930822849274, + 0.013691209256649017, + 0.028466667979955673, + 0.0012756726937368512, + 0.00200584321282804, + 0.0017753354040905833, + 0.004553396254777908, + 0.0043009258806705475, + 0.00268930965103209, + 0.0028139748610556126, + 0.0021145474165678024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.04865909367799759 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780817665159702 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780817665159702, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780817665159702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04906345158815384, + 0.013350147753953934, + 0.03392518311738968, + 0.025750406086444855, + 0.027454223483800888, + 0.03292622044682503, + 0.021519863978028297, + 0.5419432520866394, + 0.12525208294391632, + 0.1288152039051056 + ], + "credit_map": [ + 0.0041664643213152885, + 0.0071004279889166355, + 0.0076822564005851746, + 0.00695391371846199, + 0.0067849173210561275, + 0.006403507199138403, + 0.004343449138104916, + 0.005628511309623718, + 0.0010991727467626333, + 0.0019056056626141071, + 0.0018076809355989099, + 0.0014276998117566109, + 0.0019356135744601488, + 0.002096849959343672, + 0.0011352802393957973, + 0.0019422437762841582, + 0.003101003123447299, + 0.003848801599815488, + 0.0047395904548466206, + 0.004471483640372753, + 0.0043538580648601055, + 0.006187343504279852, + 0.0028178354259580374, + 0.004405267536640167, + 0.0032133038621395826, + 0.0030378580559045076, + 0.003145154332742095, + 0.0028672676999121904, + 0.003083287039771676, + 0.0056943874806165695, + 0.0016067319083958864, + 0.003102413844317198, + 0.0020634597167372704, + 0.0018216721946373582, + 0.0016668821917846799, + 0.001797746866941452, + 0.002857052721083164, + 0.007785869762301445, + 0.0023502777330577374, + 0.0071112592704594135, + 0.0026770581025630236, + 0.0035930017475038767, + 0.00857382919639349, + 0.003760030958801508, + 0.003299735253676772, + 0.004546692129224539, + 0.002722611417993903, + 0.0037532576825469732, + 0.0019640715327113867, + 0.0013037604512646794, + 0.0014153303345665336, + 0.0022071890998631716, + 0.0037862646859139204, + 0.0025383911561220884, + 0.004113720264285803, + 0.00419113552197814, + 0.05682593211531639, + 0.014987275935709476, + 0.012112991884350777, + 0.025097226724028587, + 0.111710324883461, + 0.20922768115997314, + 0.00912060122936964, + 0.102861188352108, + 0.01110642496496439, + 0.012629808858036995, + 0.013039707206189632, + 0.010284739546477795, + 0.0229530967772007, + 0.023314127698540688, + 0.007682786323130131, + 0.024241376668214798, + 0.011491883546113968, + 0.016234315931797028, + 0.014740506187081337, + 0.013265632092952728, + 0.02137652225792408, + 0.01630363240838051, + 0.018283596262335777, + 0.01711910590529442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.009780817665159702 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003615511115640402 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.003615511115640402, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003615511115640402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04747394844889641, + 0.03256253898143768, + 0.021920302882790565, + 0.020022692158818245, + 0.040350962430238724, + 0.033308979123830795, + 0.0844651386141777, + 0.5722748041152954, + 0.11560423672199249, + 0.03201635181903839 + ], + "credit_map": [ + 0.004717055708169937, + 0.009423252195119858, + 0.006331704556941986, + 0.005243878811597824, + 0.006815989967435598, + 0.006856330204755068, + 0.0034521075431257486, + 0.0046336292289197445, + 0.0031952322460711002, + 0.005579603835940361, + 0.006139010190963745, + 0.0030479298438876867, + 0.005425876472145319, + 0.004134508315473795, + 0.0020608387421816587, + 0.0029795372392982244, + 0.001618879265151918, + 0.00286954827606678, + 0.00390625698491931, + 0.0020922143012285233, + 0.004110636189579964, + 0.0032227810006588697, + 0.0019761850126087666, + 0.002123801736161113, + 0.002307864371687174, + 0.002404311904683709, + 0.0029473418835550547, + 0.001976264640688896, + 0.0023956827353686094, + 0.004298814572393894, + 0.0013953312300145626, + 0.002297082217410207, + 0.0024846901651471853, + 0.002332355361431837, + 0.003429470816627145, + 0.0019195774802938104, + 0.0022572872694581747, + 0.013761010952293873, + 0.0014596442924812436, + 0.01270692516118288, + 0.003058937843888998, + 0.0035289928782731295, + 0.008236858993768692, + 0.004484592005610466, + 0.003151851939037442, + 0.004061996005475521, + 0.003230817150324583, + 0.0035549309104681015, + 0.007259064354002476, + 0.012128488160669804, + 0.006977676879614592, + 0.010828658938407898, + 0.01116208266466856, + 0.012106391601264477, + 0.01367567665874958, + 0.010327096097171307, + 0.05981269106268883, + 0.01577906683087349, + 0.012052247300744057, + 0.02601667307317257, + 0.11333341151475906, + 0.22072236239910126, + 0.009751877747476101, + 0.11480650305747986, + 0.008670205250382423, + 0.013237705454230309, + 0.006388060282915831, + 0.0069990577176213264, + 0.03858313336968422, + 0.026680422946810722, + 0.005524772219359875, + 0.009520882740616798, + 0.0028751608915627003, + 0.0033451002091169357, + 0.0030421088449656963, + 0.0042131636291742325, + 0.004059928469359875, + 0.006101761478930712, + 0.0048996396362781525, + 0.0034794907551258802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.003615511115640402 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle, wide shot places us at a busy urban intersection at dusk. A police officer in his forties stands steady in the center of the frame, his face stern and his focused gaze beneath neatly trimmed black hair, revealing a time-tested composure. He wears a crisp, dark blue uniform, topped by a fluorescent yellow reflective vest that stands out against the darkening sky. Its silver reflective stripes flicker in the light of car lights. Around him, the city moves at breakneck speed: towering glass skyscrapers reflect the last orange-red glow of the setting sun. On the ground, a stream of yellow taxis, red buses, and private cars of various colors alternates between signals and his direction, creating a motion blur of colored streaks. The sidewalks are thronged with hurried pedestrians. His movements are precise and powerful. One white-gloved hand rises decisively, signaling the traffic in front of him to stop. Then, with a slight turn of his body, his other hand swings sideways in a smooth, crisp motion, directing the vehicles passing sideways. The soft golden afterglow outlines his tall silhouette, and the car lights and the neon lights cast colorful spots of light on the buildings behind him. The whole scene, with the cold gray tones of the city and the warm afterglow of the sunset, creates a tense, orderly and grand atmosphere full of urban vitality.", + "seed": 13910, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_15/index_6049_seed_13910_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.021608806727454067, + "vlm_logprob": -0.021608806727454067, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.021608806727454067, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03731713443994522 + ], + "opposite_logprob_tokens": [ + -4.537317276000977 + ], + "logprob_reward": -0.03731713443994522, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.03731713443994522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2800314426422119, + 0.09765812754631042, + 0.03299538046121597, + 0.08014541864395142, + 0.04427168145775795, + 0.06813488155603409, + 0.09660112112760544, + 0.10294615477323532, + 0.10103396326303482, + 0.09618180245161057 + ], + "credit_map": [ + 0.023331789299845695, + 0.025030670687556267, + 0.033006392419338226, + 0.033375222235918045, + 0.03553692251443863, + 0.03826878219842911, + 0.030248720198869705, + 0.06123296543955803, + 0.012953351251780987, + 0.008518991991877556, + 0.014881541952490807, + 0.010041855275630951, + 0.013470142148435116, + 0.017885873094201088, + 0.011667720973491669, + 0.008238663896918297, + 0.003529289038851857, + 0.0031522128265351057, + 0.005104586482048035, + 0.0035077689681202173, + 0.003213390475139022, + 0.004248860292136669, + 0.005017857067286968, + 0.005221414379775524, + 0.00758197670802474, + 0.007691591512411833, + 0.009994592517614365, + 0.00960370060056448, + 0.010310462675988674, + 0.012402861379086971, + 0.009035148657858372, + 0.013525094836950302, + 0.003933437168598175, + 0.0041728331707417965, + 0.0045915269292891026, + 0.006212239153683186, + 0.005792474374175072, + 0.004812323953956366, + 0.009393814019858837, + 0.005363035015761852, + 0.006132189650088549, + 0.006060766987502575, + 0.008633933030068874, + 0.00918935053050518, + 0.012031885795295238, + 0.008544296957552433, + 0.009552521631121635, + 0.007989950478076935, + 0.008964881300926208, + 0.006868923082947731, + 0.007978859357535839, + 0.012242987751960754, + 0.008313500322401524, + 0.03311260789632797, + 0.00818459689617157, + 0.010934773832559586, + 0.008107357658445835, + 0.007153027690947056, + 0.010776069015264511, + 0.010815945453941822, + 0.01135337259620428, + 0.02363976463675499, + 0.007558096200227737, + 0.023542532697319984, + 0.011211619712412357, + 0.0068642739206552505, + 0.009894439950585365, + 0.012796656228601933, + 0.013857385143637657, + 0.020693913102149963, + 0.013269017450511456, + 0.012446658685803413, + 0.01134051289409399, + 0.021874889731407166, + 0.010299308225512505, + 0.013054869137704372, + 0.00890212133526802, + 0.009115162305533886, + 0.012061805464327335, + 0.009533138014376163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.03731713443994522 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009815877303481102 + ], + "opposite_logprob_tokens": [ + -4.634815692901611 + ], + "logprob_reward": -0.009815877303481102, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009815877303481102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21736112236976624, + 0.10629238933324814, + 0.04655624181032181, + 0.12397521734237671, + 0.08439557254314423, + 0.08650814741849899, + 0.09360093623399734, + 0.06855983287096024, + 0.09895186871290207, + 0.07379870861768723 + ], + "credit_map": [ + 0.01636895351111889, + 0.017154330387711525, + 0.022709853947162628, + 0.027184905484318733, + 0.02542850188910961, + 0.036774441599845886, + 0.04448859393596649, + 0.02725151926279068, + 0.015253025107085705, + 0.009456532076001167, + 0.01559098158031702, + 0.009455952793359756, + 0.013026213273406029, + 0.011217189952731133, + 0.023524155840277672, + 0.00876831915229559, + 0.004075503442436457, + 0.004948064684867859, + 0.004714850336313248, + 0.005613591521978378, + 0.009753184393048286, + 0.007247561123222113, + 0.00470096105709672, + 0.005502515938133001, + 0.010462949983775616, + 0.020365362986922264, + 0.010945731773972511, + 0.014943619258701801, + 0.022226613014936447, + 0.020043348893523216, + 0.012576782144606113, + 0.012410799972712994, + 0.007004454731941223, + 0.006052013486623764, + 0.007919883355498314, + 0.010815056972205639, + 0.012473292648792267, + 0.018431086093187332, + 0.01085660606622696, + 0.010843176394701004, + 0.007016762159764767, + 0.008228452876210213, + 0.01124429889023304, + 0.01031019352376461, + 0.014863328076899052, + 0.009812334552407265, + 0.01444787997752428, + 0.010584887117147446, + 0.00594767602160573, + 0.0076444195583462715, + 0.010361331515014172, + 0.01455327682197094, + 0.01559843122959137, + 0.015025017783045769, + 0.012036314234137535, + 0.012434454634785652, + 0.005717536900192499, + 0.004661526530981064, + 0.008275995962321758, + 0.009013758040964603, + 0.013485600240528584, + 0.007561783771961927, + 0.010252273641526699, + 0.00959134940057993, + 0.007849857211112976, + 0.008517403155565262, + 0.012700052000582218, + 0.011657617054879665, + 0.011011781170964241, + 0.018319081515073776, + 0.010793328285217285, + 0.018102744594216347, + 0.006548176985234022, + 0.012501034885644913, + 0.009002838283777237, + 0.00906007457524538, + 0.008266865275800228, + 0.007730534765869379, + 0.012702970765531063, + 0.007986207492649555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.009815877303481102 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048658184707164764 + ], + "opposite_logprob_tokens": [ + -3.0486581325531006 + ], + "logprob_reward": -0.048658184707164764, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048658184707164764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1534980833530426, + 0.1113186702132225, + 0.028405461460351944, + 0.04793117195367813, + 0.10894447565078735, + 0.07910443097352982, + 0.04976509511470795, + 0.3143922984600067, + 0.05503661930561066, + 0.05160365626215935 + ], + "credit_map": [ + 0.012516012415289879, + 0.010136008262634277, + 0.013930180110037327, + 0.015503945760428905, + 0.02770516462624073, + 0.020008640363812447, + 0.02256322093307972, + 0.03113490343093872, + 0.01143462024629116, + 0.0100827282294631, + 0.014223421923816204, + 0.011198279447853565, + 0.01407836377620697, + 0.021825214847922325, + 0.018836993724107742, + 0.009639041498303413, + 0.0019240387482568622, + 0.0026722182519733906, + 0.0030757670756429434, + 0.0025535698514431715, + 0.004459744319319725, + 0.0050759161822497845, + 0.0048080431297421455, + 0.003836165415123105, + 0.00392633443698287, + 0.005656990688294172, + 0.004202347714453936, + 0.004349302500486374, + 0.00859137810766697, + 0.010459977202117443, + 0.004936147015541792, + 0.005808695685118437, + 0.008086977526545525, + 0.008297992870211601, + 0.009420034475624561, + 0.015622034668922424, + 0.02033628523349762, + 0.021274438127875328, + 0.018103500828146935, + 0.00780321191996336, + 0.007026472128927708, + 0.010875922627747059, + 0.009595579467713833, + 0.00870917085558176, + 0.014222019352018833, + 0.008675651624798775, + 0.0123075470328331, + 0.007692066021263599, + 0.004382990300655365, + 0.0034897061996161938, + 0.004309298470616341, + 0.0057398416101932526, + 0.007314213551580906, + 0.012009498663246632, + 0.007035288959741592, + 0.005484257359057665, + 0.02611834928393364, + 0.015442143194377422, + 0.01987610012292862, + 0.02766093984246254, + 0.029990751296281815, + 0.09604214131832123, + 0.018675535917282104, + 0.08058634400367737, + 0.004457560833543539, + 0.005480112507939339, + 0.005905339494347572, + 0.005519896745681763, + 0.011310587637126446, + 0.0067486949265003204, + 0.007586050778627396, + 0.008028375916182995, + 0.003137177089229226, + 0.004751501139253378, + 0.005945106036961079, + 0.006037767976522446, + 0.006159295327961445, + 0.008980256505310535, + 0.010462718084454536, + 0.006129828747361898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.048658184707164764 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638394996523857 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638394996523857, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638394996523857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21336494386196136, + 0.09342563152313232, + 0.0426170714199543, + 0.06985580176115036, + 0.05395405739545822, + 0.03496680408716202, + 0.19124673306941986, + 0.15471868216991425, + 0.03517559915781021, + 0.11067469418048859 + ], + "credit_map": [ + 0.03190236911177635, + 0.017706725746393204, + 0.03408706933259964, + 0.022672707214951515, + 0.02843729592859745, + 0.023442508652806282, + 0.03093063272535801, + 0.02418564073741436, + 0.00691863801330328, + 0.008484258316457272, + 0.01314062625169754, + 0.009178032167255878, + 0.011529277078807354, + 0.016381187364459038, + 0.00983613170683384, + 0.017957478761672974, + 0.005335833411663771, + 0.004614480771124363, + 0.005693539511412382, + 0.003642236813902855, + 0.005078065674751997, + 0.004760053474456072, + 0.006559428758919239, + 0.006933433469384909, + 0.006703095510601997, + 0.008318634703755379, + 0.006120945326983929, + 0.006881983019411564, + 0.008982613682746887, + 0.015370642766356468, + 0.005792852025479078, + 0.011685040779411793, + 0.005199503619223833, + 0.005536126904189587, + 0.005292579531669617, + 0.006464350502938032, + 0.00913055520504713, + 0.00975295901298523, + 0.006143656559288502, + 0.006434326060116291, + 0.0028709047473967075, + 0.005641797557473183, + 0.005142188165336847, + 0.004863927606493235, + 0.003776401514187455, + 0.0039120265282690525, + 0.004883128218352795, + 0.003876426722854376, + 0.019782835617661476, + 0.02359396032989025, + 0.022460097447037697, + 0.02883058413863182, + 0.020885512232780457, + 0.03328079730272293, + 0.01765434443950653, + 0.02475859969854355, + 0.012725859880447388, + 0.0089850639924407, + 0.011046601459383965, + 0.015404058620333672, + 0.013501927256584167, + 0.04734960198402405, + 0.007496539503335953, + 0.03820904344320297, + 0.003558498341590166, + 0.0038299700245261192, + 0.005077065899968147, + 0.005127214826643467, + 0.003403970506042242, + 0.005225928965955973, + 0.0036463956348598003, + 0.00530655262991786, + 0.01691894792020321, + 0.021993441507220268, + 0.016186222434043884, + 0.011617696844041348, + 0.010811475105583668, + 0.014419748447835445, + 0.010038969106972218, + 0.008688187226653099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.008638394996523857 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003614442190155387 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.003614442190155387, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003614442190155387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28251442313194275, + 0.2006779909133911, + 0.055189099162817, + 0.09503968060016632, + 0.06111797317862511, + 0.07408123463392258, + 0.0603046640753746, + 0.08160459250211716, + 0.04495010897517204, + 0.044520244002342224 + ], + "credit_map": [ + 0.04064703360199928, + 0.024919776245951653, + 0.031261518597602844, + 0.03319777920842171, + 0.0364348329603672, + 0.04448466747999191, + 0.028086857870221138, + 0.043481968343257904, + 0.022477194666862488, + 0.029859362170100212, + 0.02720995806157589, + 0.027558421716094017, + 0.022928085178136826, + 0.022720372304320335, + 0.025481335818767548, + 0.022443251684308052, + 0.006673583760857582, + 0.00694072712212801, + 0.007725033443421125, + 0.0065624164417386055, + 0.00916487816721201, + 0.006152058951556683, + 0.004719825927168131, + 0.007250579539686441, + 0.009777639992535114, + 0.012145113199949265, + 0.007589292712509632, + 0.011195178143680096, + 0.012489608488976955, + 0.022045468911528587, + 0.008975335396826267, + 0.010822041891515255, + 0.0066458601504564285, + 0.004514145199209452, + 0.006345185451209545, + 0.007928569801151752, + 0.009199644438922405, + 0.01309528574347496, + 0.006666297558695078, + 0.006722986698150635, + 0.006709321402013302, + 0.004513729363679886, + 0.007961052469909191, + 0.008984698913991451, + 0.009987875819206238, + 0.013175279833376408, + 0.011625869199633598, + 0.011123411357402802, + 0.006362599786370993, + 0.003982484806329012, + 0.005477214697748423, + 0.008019739761948586, + 0.006627390626817942, + 0.018970344215631485, + 0.004242755006998777, + 0.00662213983014226, + 0.006770436652004719, + 0.004923130385577679, + 0.005427504889667034, + 0.007716367952525616, + 0.007299207150936127, + 0.021555092185735703, + 0.005093182437121868, + 0.022819671779870987, + 0.003714685793966055, + 0.004888614173978567, + 0.004750984255224466, + 0.00433890288695693, + 0.007249435875564814, + 0.007574989460408688, + 0.006386294029653072, + 0.006046202965080738, + 0.0043143946677446365, + 0.007818792946636677, + 0.006550785154104233, + 0.005283594597131014, + 0.004781822673976421, + 0.005764380097389221, + 0.005025337915867567, + 0.004981134552508593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.003614442190155387 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, low-angle, wide shot places us at a busy urban intersection at dusk. A police officer in his forties stands steady in the center of the frame, his face stern and his focused gaze beneath neatly trimmed black hair, revealing a time-tested composure. He wears a crisp, dark blue uniform, topped by a fluorescent yellow reflective vest that stands out against the darkening sky. Its silver reflective stripes flicker in the light of car lights. Around him, the city moves at breakneck speed: towering glass skyscrapers reflect the last orange-red glow of the setting sun. On the ground, a stream of yellow taxis, red buses, and private cars of various colors alternates between signals and his direction, creating a motion blur of colored streaks. The sidewalks are thronged with hurried pedestrians. His movements are precise and powerful. One white-gloved hand rises decisively, signaling the traffic in front of him to stop. Then, with a slight turn of his body, his other hand swings sideways in a smooth, crisp motion, directing the vehicles passing sideways. The soft golden afterglow outlines his tall silhouette, and the car lights and the neon lights cast colorful spots of light on the buildings behind him. The whole scene, with the cold gray tones of the city and the warm afterglow of the sunset, creates a tense, orderly and grand atmosphere full of urban vitality.", + "seed": 13911, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_15/index_6049_seed_13911_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02921989932656288, + "vlm_logprob": -0.02921989932656288, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.02921989932656288, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04268788918852806 + ], + "opposite_logprob_tokens": [ + -5.417687892913818 + ], + "logprob_reward": -0.04268788918852806, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.04268788918852806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14497607946395874, + 0.05756159871816635, + 0.06860671937465668, + 0.04192424938082695, + 0.07907792180776596, + 0.29080119729042053, + 0.06189998984336853, + 0.06372789293527603, + 0.09777972847223282, + 0.09364455938339233 + ], + "credit_map": [ + 0.012206044048070908, + 0.01166695635765791, + 0.022120337933301926, + 0.018702156841754913, + 0.02105456031858921, + 0.024123676121234894, + 0.018027199432253838, + 0.017075147479772568, + 0.0044453018344938755, + 0.004601529333740473, + 0.010222782380878925, + 0.005927258636802435, + 0.008191710337996483, + 0.012235844507813454, + 0.005367608275264502, + 0.006569566670805216, + 0.012840495444834232, + 0.00581379933282733, + 0.006425199098885059, + 0.006995443720370531, + 0.01213599182665348, + 0.009058461524546146, + 0.008279423229396343, + 0.0070579033344984055, + 0.0044837999157607555, + 0.0032338877208530903, + 0.003757558995857835, + 0.0045674992725253105, + 0.00795789621770382, + 0.00523396534845233, + 0.006988704204559326, + 0.00570093747228384, + 0.008748742751777172, + 0.005679074674844742, + 0.009615937247872353, + 0.006409886293113232, + 0.012861890718340874, + 0.009969864040613174, + 0.009840469807386398, + 0.01595204882323742, + 0.03957144916057587, + 0.013557464815676212, + 0.02536461129784584, + 0.019247164949774742, + 0.05022392421960831, + 0.026324762031435966, + 0.0150284543633461, + 0.10148336738348007, + 0.004621704574674368, + 0.006214880384504795, + 0.006124089937657118, + 0.006349928677082062, + 0.016586149111390114, + 0.009122376330196857, + 0.004738755524158478, + 0.008142104372382164, + 0.005979454610496759, + 0.009954182431101799, + 0.009829317219555378, + 0.005369798745959997, + 0.011085034348070621, + 0.00881120190024376, + 0.005173590034246445, + 0.007525312714278698, + 0.013140341266989708, + 0.015001515857875347, + 0.011555472388863564, + 0.015596873126924038, + 0.012368617579340935, + 0.007918584160506725, + 0.011049790307879448, + 0.011148534715175629, + 0.008391656912863255, + 0.013601514510810375, + 0.008338472805917263, + 0.013945125043392181, + 0.011150971055030823, + 0.011166633106768131, + 0.015399816446006298, + 0.011650372296571732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.04268788918852806 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009830987080931664 + ], + "opposite_logprob_tokens": [ + -4.634830951690674 + ], + "logprob_reward": -0.009830987080931664, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009830987080931664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17831069231033325, + 0.08756783604621887, + 0.04713376238942146, + 0.08744193613529205, + 0.11328946799039841, + 0.0603739395737648, + 0.20815213024616241, + 0.06460384279489517, + 0.07583536952733994, + 0.07729102671146393 + ], + "credit_map": [ + 0.020869044587016106, + 0.01812170073390007, + 0.026241585612297058, + 0.020773252472281456, + 0.02204677276313305, + 0.02843935787677765, + 0.019429583102464676, + 0.022389402613043785, + 0.015444697812199593, + 0.006215481553226709, + 0.012097547762095928, + 0.008426137268543243, + 0.01029178686439991, + 0.007301160134375095, + 0.01114934403449297, + 0.016641685739159584, + 0.007036469876766205, + 0.004713631700724363, + 0.006705339532345533, + 0.005635111127048731, + 0.005589447915554047, + 0.005348594859242439, + 0.007012680172920227, + 0.005092486273497343, + 0.006709908600896597, + 0.0075185662135481834, + 0.005858349148184061, + 0.008935293182730675, + 0.021673647686839104, + 0.012285000644624233, + 0.015302877873182297, + 0.009158295579254627, + 0.018235526978969574, + 0.00676725385710597, + 0.010639771819114685, + 0.008412355557084084, + 0.013081679120659828, + 0.01437248382717371, + 0.010991803370416164, + 0.030788596719503403, + 0.008364120498299599, + 0.003812595969066024, + 0.005699269939213991, + 0.004410726483911276, + 0.009539872407913208, + 0.0064889234490692616, + 0.00624194648116827, + 0.01581648178398609, + 0.016965745016932487, + 0.01581522449851036, + 0.019566267728805542, + 0.020141704007983208, + 0.05774354189634323, + 0.03291473537683487, + 0.01587744429707527, + 0.029127469286322594, + 0.006322361063212156, + 0.006021677050739527, + 0.008551553823053837, + 0.007333799730986357, + 0.01102391816675663, + 0.005903556942939758, + 0.00965910404920578, + 0.009787868708372116, + 0.00791421439498663, + 0.008689831010997295, + 0.009628130123019218, + 0.012167776003479958, + 0.012306083925068378, + 0.006326721515506506, + 0.00891641154885292, + 0.009886198677122593, + 0.007877352647483349, + 0.012220257893204689, + 0.0062089646235108376, + 0.009619071148335934, + 0.012412402778863907, + 0.009438905864953995, + 0.012736447155475616, + 0.006777625530958176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.009830987080931664 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899495214223862 + ], + "opposite_logprob_tokens": [ + -2.5789949893951416 + ], + "logprob_reward": -0.07899495214223862, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899495214223862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13085828721523285, + 0.04631379246711731, + 0.055650316178798676, + 0.023824209347367287, + 0.0403071790933609, + 0.07567916065454483, + 0.10019159317016602, + 0.048827338963747025, + 0.41523680090904236, + 0.0631113052368164 + ], + "credit_map": [ + 0.0182674340903759, + 0.011298833414912224, + 0.01721886545419693, + 0.012690501287579536, + 0.017752939835190773, + 0.025058552622795105, + 0.015728456899523735, + 0.012842698022723198, + 0.003962840884923935, + 0.003751271404325962, + 0.007248787675052881, + 0.005104980897158384, + 0.006573879159986973, + 0.0070216720923781395, + 0.004569577053189278, + 0.008080787025392056, + 0.006687304470688105, + 0.004900272469967604, + 0.007282852195203304, + 0.005937689449638128, + 0.0071112592704594135, + 0.007784365210682154, + 0.007624641992151737, + 0.008321929723024368, + 0.0023887758143246174, + 0.0021109518129378557, + 0.0034114925656467676, + 0.0025211733300238848, + 0.0040373350493609905, + 0.0031272952910512686, + 0.0034663942642509937, + 0.0027607916854321957, + 0.0036780366208404303, + 0.0030180008616298437, + 0.0037947804667055607, + 0.004038597922772169, + 0.009522871114313602, + 0.005666770972311497, + 0.005284893326461315, + 0.00530322827398777, + 0.01042177714407444, + 0.0034362864680588245, + 0.008306281641125679, + 0.005187686067074537, + 0.011205187067389488, + 0.008160118944942951, + 0.004604610148817301, + 0.02435721643269062, + 0.010282174684107304, + 0.006490303203463554, + 0.011305099353194237, + 0.010778415016829967, + 0.031061574816703796, + 0.007761073764413595, + 0.007889905013144016, + 0.014623057097196579, + 0.005779423750936985, + 0.00821949914097786, + 0.005007746163755655, + 0.0036673378199338913, + 0.006784206721931696, + 0.0057221041060984135, + 0.006656522862613201, + 0.006990497466176748, + 0.04864395409822464, + 0.052010852843523026, + 0.028116676956415176, + 0.07668182253837585, + 0.0731736421585083, + 0.025216709822416306, + 0.06135139614343643, + 0.05004177987575531, + 0.006191480904817581, + 0.010276460088789463, + 0.004074519034475088, + 0.005646179895848036, + 0.017854470759630203, + 0.007198977284133434, + 0.007572933100163937, + 0.004296286031603813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.07899495214223862 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638868108391762 + ], + "opposite_logprob_tokens": [ + -4.758638858795166 + ], + "logprob_reward": -0.008638868108391762, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638868108391762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11169569194316864, + 0.040824610739946365, + 0.04918890446424484, + 0.048538688570261, + 0.034849829971790314, + 0.23016788065433502, + 0.02819950319826603, + 0.0354568250477314, + 0.323422372341156, + 0.09765563905239105 + ], + "credit_map": [ + 0.011544720269739628, + 0.013300387188792229, + 0.017662089318037033, + 0.009821639396250248, + 0.01344984769821167, + 0.016801318153738976, + 0.008450483903288841, + 0.020665206015110016, + 0.005763646215200424, + 0.0037679297383874655, + 0.006924408487975597, + 0.003640917129814625, + 0.00506573123857379, + 0.004209826700389385, + 0.002859652740880847, + 0.008592499420046806, + 0.009651867672801018, + 0.003986041527241468, + 0.005492425058037043, + 0.005014488007873297, + 0.005315184593200684, + 0.006373908370733261, + 0.005114366766065359, + 0.008240624330937862, + 0.004288140684366226, + 0.004726166371256113, + 0.006995067000389099, + 0.0069060479290783405, + 0.008325384929776192, + 0.005894101224839687, + 0.005585165694355965, + 0.0058186170645058155, + 0.006345206405967474, + 0.0020255029667168856, + 0.004158741794526577, + 0.002980940043926239, + 0.004197975154966116, + 0.0039063128642737865, + 0.003602385288104415, + 0.007632765918970108, + 0.03166930004954338, + 0.01072249561548233, + 0.02178877778351307, + 0.015526136383414268, + 0.03752315416932106, + 0.02100672759115696, + 0.013457262888550758, + 0.0784740298986435, + 0.002810513833537698, + 0.002916165394708514, + 0.0051500070840120316, + 0.0029109951574355364, + 0.00489599397405982, + 0.0035557837691158056, + 0.0017641981830820441, + 0.0041958484798669815, + 0.004542734008282423, + 0.005011364817619324, + 0.004534715320914984, + 0.0026622870936989784, + 0.005804531276226044, + 0.0046662562526762486, + 0.0027473964728415012, + 0.005487538408488035, + 0.04370987415313721, + 0.04936198145151138, + 0.026893559843301773, + 0.04094677045941353, + 0.05666620656847954, + 0.026479704305529594, + 0.031501222401857376, + 0.047863055020570755, + 0.0105294119566679, + 0.013059435412287712, + 0.0097028324380517, + 0.0129872877150774, + 0.018273305147886276, + 0.014483654871582985, + 0.009095218032598495, + 0.009524494409561157 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.008638868108391762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set at a busy urban intersection at dusk?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a police officer in his forties wearing a dark blue uniform and a fluorescent yellow reflective vest visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the officer raise one white-gloved hand to signal traffic to stop and then swing the other hand sideways to direct vehicles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible streams of yellow taxis, red buses, and private cars creating motion blur on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle wide shot with towering glass skyscrapers reflecting the orange-red glow of the setting sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946800112724304 + ], + "opposite_logprob_tokens": [ + -5.130946636199951 + ], + "logprob_reward": -0.005946800112724304, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946800112724304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10658225417137146, + 0.06286240369081497, + 0.10638014227151871, + 0.05189980939030647, + 0.05547936633229256, + 0.1658928543329239, + 0.06180348992347717, + 0.11461299657821655, + 0.08414319902658463, + 0.1903434693813324 + ], + "credit_map": [ + 0.015058644115924835, + 0.007645050063729286, + 0.011746919713914394, + 0.008412659168243408, + 0.010931180790066719, + 0.02325098030269146, + 0.01411074586212635, + 0.015426070429384708, + 0.010219785384833813, + 0.004255141597241163, + 0.01009327918291092, + 0.005207199137657881, + 0.006527367047965527, + 0.006431773770600557, + 0.0048903957940638065, + 0.015237457118928432, + 0.016902277246117592, + 0.010497729294002056, + 0.013253373093903065, + 0.011596713215112686, + 0.013118095695972443, + 0.01541543286293745, + 0.00910993106663227, + 0.01648659259080887, + 0.006257981061935425, + 0.004894617013633251, + 0.01027300488203764, + 0.006194130517542362, + 0.0055385613813996315, + 0.007661715149879456, + 0.005383912939578295, + 0.005695885047316551, + 0.003903005737811327, + 0.0032648153137415648, + 0.006850880570709705, + 0.0036929056514054537, + 0.016284707933664322, + 0.0071477023884654045, + 0.0054021733812987804, + 0.008933180011808872, + 0.016879303380846977, + 0.015514795668423176, + 0.026508167386054993, + 0.01323719136416912, + 0.0237145833671093, + 0.030077530071139336, + 0.00749573390930891, + 0.032465554773807526, + 0.005035156384110451, + 0.004042635206133127, + 0.007716768886893988, + 0.006904180161654949, + 0.017694517970085144, + 0.0074791088700294495, + 0.004516879562288523, + 0.008414243347942829, + 0.01767689175903797, + 0.01564009301364422, + 0.013666420243680477, + 0.008147221989929676, + 0.016027716919779778, + 0.017987126484513283, + 0.010366982780396938, + 0.015100542455911636, + 0.007821772247552872, + 0.011424725875258446, + 0.007504711393266916, + 0.013868122361600399, + 0.012251263484358788, + 0.009016525000333786, + 0.01013200543820858, + 0.012124067172408104, + 0.019870886579155922, + 0.026865357533097267, + 0.025949491187930107, + 0.023200757801532745, + 0.028307972475886345, + 0.029484473168849945, + 0.017934206873178482, + 0.018730325624346733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 7 + ], + "selected_reward": -0.005946800112724304 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_2/index_tensor([9580, 9580])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_2/index_tensor([9580, 9580])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..437d5c38e94f210f3bb2b458183421843ac50616 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_2/index_tensor([9580, 9580])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, tight close-up, positioned at a slightly elevated angle, focuses on a polished stainless steel lid resting on a large stockpot atop a modern, black ceramic cooktop. Soft, warm afternoon sunlight slants through an out-of-frame window, casting a brilliant glow on the lid's curved surface, contrasting sharply with the even, neutral glow from the kitchen's overhead light. The movement begins delicately as the lid, featuring a simple black Bakelite handle, trembles almost imperceptibly. Then, with a sudden jolt, one rim lifts, releasing a sharp puff of white steam that briefly obscures the air before dissipating, and the lid drops with a distinct metallic clang. The movement begins as a steady, deliberate repetition, then evolves into a more frantic, rhythmic rattling as the pressure within builds, each movement more pronounced than the last, creating an atmosphere of tension and anticipation in the quiet kitchen.", + "seed": 33459, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_2/index_9580_seed_33459_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12139663994312286, + "vlm_logprob": -0.12139663994312286, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.12139663994312286, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06772944331169128 + ], + "opposite_logprob_tokens": [ + -3.3177294731140137 + ], + "logprob_reward": -0.06772944331169128, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06772944331169128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0831330344080925, + 0.40612462162971497, + 0.15944232046604156, + 0.02777913212776184, + 0.0474531352519989, + 0.17155693471431732, + 0.047770656645298004, + 0.019482413306832314, + 0.022015942260622978, + 0.015241812914609909 + ], + "credit_map": [ + 0.01048008818179369, + 0.012518180534243584, + 0.006798180285841227, + 0.009444977156817913, + 0.009058605879545212, + 0.012253818102180958, + 0.013683113269507885, + 0.008896071463823318, + 0.055601567029953, + 0.039769552648067474, + 0.029215741902589798, + 0.0636112317442894, + 0.0508543998003006, + 0.046371787786483765, + 0.06574353575706482, + 0.05495680123567581, + 0.014787005260586739, + 0.016630755737423897, + 0.009752640500664711, + 0.011349091306328773, + 0.05675758048892021, + 0.027752956375479698, + 0.012946020811796188, + 0.009466278366744518, + 0.004721911158412695, + 0.0009410958155058324, + 0.001056784763932228, + 0.0013066843384876847, + 0.006546814925968647, + 0.010593876242637634, + 0.0010947195114567876, + 0.0015172461280599236, + 0.003479606471955776, + 0.002137662610039115, + 0.0042844656854867935, + 0.0035025400575250387, + 0.013756070286035538, + 0.013944003731012344, + 0.004375970456749201, + 0.0019728143233805895, + 0.009182467125356197, + 0.006816788576543331, + 0.008104166015982628, + 0.007491439115256071, + 0.0570746585726738, + 0.06752239912748337, + 0.009692978113889694, + 0.005672036204487085, + 0.010995667427778244, + 0.003949516452848911, + 0.003849402302876115, + 0.004465612582862377, + 0.010965114459395409, + 0.004563929513096809, + 0.006201810669153929, + 0.0027796023059636354, + 0.0033838895615190268, + 0.0014178266283124685, + 0.002430841326713562, + 0.002050988143309951, + 0.0014601805014535785, + 0.0038149103056639433, + 0.0030949742067605257, + 0.0018288013525307178, + 0.002566336886957288, + 0.0017422473756596446, + 0.0030351108871400356, + 0.002434725174680352, + 0.003981226589530706, + 0.0042109242640435696, + 0.00174413388594985, + 0.002301236614584923, + 0.0012450387002900243, + 0.001458185026422143, + 0.0016648139571771026, + 0.002491277875378728, + 0.0013981062220409513, + 0.0017243921756744385, + 0.004212487954646349, + 0.0010475103044882417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.06772944331169128 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11302146315574646 + ], + "opposite_logprob_tokens": [ + -2.2380213737487793 + ], + "logprob_reward": -0.11302146315574646, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11302146315574646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11490193009376526, + 0.5414652824401855, + 0.0692654401063919, + 0.053357526659965515, + 0.047894466668367386, + 0.07013561576604843, + 0.01722291298210621, + 0.05198385939002037, + 0.02467397414147854, + 0.009099006652832031 + ], + "credit_map": [ + 0.012683313339948654, + 0.014834760688245296, + 0.00914191734045744, + 0.012996000237762928, + 0.011534003540873528, + 0.022138983011245728, + 0.020308414474129677, + 0.011264534667134285, + 0.07752303034067154, + 0.05149741470813751, + 0.03806150704622269, + 0.08131542056798935, + 0.06803809851408005, + 0.06332839280366898, + 0.0843268632888794, + 0.07737459987401962, + 0.00768217584118247, + 0.006819834932684898, + 0.0038312252145260572, + 0.005910350941121578, + 0.020965881645679474, + 0.011151792481541634, + 0.006280968431383371, + 0.006623212713748217, + 0.011156892403960228, + 0.0015391033375635743, + 0.002271579345688224, + 0.0029547372832894325, + 0.010416504926979542, + 0.019552212208509445, + 0.0027864743024110794, + 0.002680023666471243, + 0.004582359455525875, + 0.002849260112270713, + 0.0059372615069150925, + 0.0041005900129675865, + 0.010521350428462029, + 0.012577509507536888, + 0.004169701598584652, + 0.003156432183459401, + 0.0035170724149793386, + 0.0018729186849668622, + 0.0024327498394995928, + 0.0022002938203513622, + 0.02527967467904091, + 0.029521547257900238, + 0.0032344176433980465, + 0.00207693874835968, + 0.003222828032448888, + 0.0018625300144776702, + 0.0015131003456190228, + 0.0017415708862245083, + 0.003701030509546399, + 0.001954668667167425, + 0.002445654710754752, + 0.000781530630774796, + 0.005905201658606529, + 0.002529505640268326, + 0.002891024574637413, + 0.0033881610725075006, + 0.013334511779248714, + 0.018690895289182663, + 0.002562015550211072, + 0.002682540100067854, + 0.003376040607690811, + 0.0021130526438355446, + 0.0022579324431717396, + 0.0021306464914232492, + 0.004155382048338652, + 0.005182911641895771, + 0.002605283400043845, + 0.002852724166586995, + 0.0010130595182999969, + 0.000795789179392159, + 0.001011051470413804, + 0.0011378662893548608, + 0.0010747453197836876, + 0.0014547856990247965, + 0.002061315579339862, + 0.0005503935972228646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.11302146315574646 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31348028779029846 + ], + "opposite_logprob_tokens": [ + -1.313480257987976 + ], + "logprob_reward": -0.31348028779029846, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31348028779029846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05258764326572418, + 0.11555752903223038, + 0.4873490333557129, + 0.043565694242715836, + 0.12047144770622253, + 0.04618717357516289, + 0.04529980942606926, + 0.06751374900341034, + 0.013049989938735962, + 0.008417913690209389 + ], + "credit_map": [ + 0.005447738338261843, + 0.00634088646620512, + 0.006296399515122175, + 0.005738175008445978, + 0.00493848929181695, + 0.006813645828515291, + 0.01015025470405817, + 0.0068620555102825165, + 0.02025950327515602, + 0.01100094523280859, + 0.009005354717373848, + 0.01731022447347641, + 0.01448965072631836, + 0.013022735714912415, + 0.0172518752515316, + 0.013217242434620857, + 0.051237739622592926, + 0.05467832460999489, + 0.029257027432322502, + 0.04113799333572388, + 0.1512753963470459, + 0.0796530693769455, + 0.04330076649785042, + 0.03680873662233353, + 0.0049633304588496685, + 0.003048427402973175, + 0.003953226841986179, + 0.005335985217243433, + 0.008218903094530106, + 0.008511098101735115, + 0.005092579871416092, + 0.004442144185304642, + 0.017133064568042755, + 0.0117217767983675, + 0.023209718987345695, + 0.016658278182148933, + 0.012548672966659069, + 0.01721966825425625, + 0.014403321780264378, + 0.007576954551041126, + 0.0026041402015835047, + 0.00164816714823246, + 0.0022402170579880476, + 0.00195824122056365, + 0.015223960392177105, + 0.018251484259963036, + 0.0026958701200783253, + 0.0015650958521291614, + 0.007399936206638813, + 0.0038530747406184673, + 0.0031461885664612055, + 0.003453560872003436, + 0.01353700552135706, + 0.008329586125910282, + 0.0035009840503335, + 0.002079472178593278, + 0.005058892536908388, + 0.0033584891352802515, + 0.004951260983943939, + 0.00455258134752512, + 0.01973431557416916, + 0.023957187309861183, + 0.0036813365295529366, + 0.0022196841891855, + 0.0015673479065299034, + 0.0012217918410897255, + 0.0015172030543908477, + 0.0012941696913912892, + 0.0022042321506887674, + 0.0027114471886307, + 0.0014022820396348834, + 0.0011315162992104888, + 0.0010038161417469382, + 0.001251167501322925, + 0.0010270074708387256, + 0.0010722321458160877, + 0.000881427142303437, + 0.0007845895015634596, + 0.0016838563606142998, + 0.000713817251380533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.31348028779029846 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373083472251892 + ], + "opposite_logprob_tokens": [ + -3.4087307453155518 + ], + "logprob_reward": -0.03373083472251892, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373083472251892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04218205064535141, + 0.2578948736190796, + 0.0902135893702507, + 0.05157031863927841, + 0.042829591780900955, + 0.13277554512023926, + 0.09185881912708282, + 0.07444126904010773, + 0.028005827218294144, + 0.18822810053825378 + ], + "credit_map": [ + 0.00560365105047822, + 0.004986852407455444, + 0.004089162684977055, + 0.005490683950483799, + 0.004373990464955568, + 0.006470527965575457, + 0.006039387546479702, + 0.0051277936436235905, + 0.038856472820043564, + 0.024860622361302376, + 0.02018948644399643, + 0.04265427589416504, + 0.0308382585644722, + 0.02781861647963524, + 0.04070919752120972, + 0.03196793422102928, + 0.007873909547924995, + 0.009712066501379013, + 0.005008508451282978, + 0.006264964584261179, + 0.03305383399128914, + 0.016922112554311752, + 0.007272582966834307, + 0.004105603788048029, + 0.011591988615691662, + 0.002048375317826867, + 0.002585026202723384, + 0.0031794102396816015, + 0.009177747182548046, + 0.017731240019202232, + 0.0030820935498923063, + 0.0021744377445429564, + 0.005198238417506218, + 0.0019568833522498608, + 0.0025299438275396824, + 0.0032664660830050707, + 0.010731366463005543, + 0.014171489514410496, + 0.0026420485228300095, + 0.002333155833184719, + 0.007286847103387117, + 0.0031696802470833063, + 0.003322775475680828, + 0.0038118893280625343, + 0.048568807542324066, + 0.05903792753815651, + 0.002954215742647648, + 0.0046234168112277985, + 0.018724234774708748, + 0.005224361550062895, + 0.006729651242494583, + 0.007302684709429741, + 0.02530701644718647, + 0.012195794843137264, + 0.011055915616452694, + 0.00531916506588459, + 0.008736241608858109, + 0.004866356961429119, + 0.007099320646375418, + 0.00817632395774126, + 0.014498271979391575, + 0.021634416654706, + 0.006024012342095375, + 0.0034063183702528477, + 0.0027166579384356737, + 0.0022637262009084225, + 0.003395352978259325, + 0.0030281965155154467, + 0.004785364493727684, + 0.004908054601401091, + 0.003553202608600259, + 0.003355272114276886, + 0.01804344356060028, + 0.025881992653012276, + 0.02341565117239952, + 0.02913505770266056, + 0.017160532996058464, + 0.021274929866194725, + 0.041154395788908005, + 0.01216209214180708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 5 + ], + "selected_reward": -0.03373083472251892 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902117073535919 + ], + "opposite_logprob_tokens": [ + -2.5790212154388428 + ], + "logprob_reward": -0.07902117073535919, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902117073535919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1972142606973648, + 0.0876796543598175, + 0.20443256199359894, + 0.07058770209550858, + 0.030525153502821922, + 0.02791559137403965, + 0.06469792872667313, + 0.04261995851993561, + 0.03259192034602165, + 0.24173520505428314 + ], + "credit_map": [ + 0.030676720663905144, + 0.022309040650725365, + 0.015891576185822487, + 0.02295783720910549, + 0.024158306419849396, + 0.03401406109333038, + 0.02365380898118019, + 0.02355293557047844, + 0.013919955119490623, + 0.008789162151515484, + 0.007694773841649294, + 0.01393142994493246, + 0.007385581731796265, + 0.009986542165279388, + 0.009440485388040543, + 0.01653173565864563, + 0.021830925717949867, + 0.0214622151106596, + 0.012359404005110264, + 0.018679307773709297, + 0.06361591815948486, + 0.03313315659761429, + 0.01869955100119114, + 0.014652105048298836, + 0.017082972452044487, + 0.0030000447295606136, + 0.003456092206761241, + 0.004618972074240446, + 0.011507141403853893, + 0.02358265593647957, + 0.003908487968146801, + 0.0034313485957682133, + 0.004495589528232813, + 0.002946826396510005, + 0.004140364471822977, + 0.004660345613956451, + 0.003440992208197713, + 0.004652991425246, + 0.0041078440845012665, + 0.0020802062936127186, + 0.0026486495044082403, + 0.001864040968939662, + 0.0018985975766554475, + 0.003946720156818628, + 0.005510228220373392, + 0.007650232408195734, + 0.002175605157390237, + 0.0022215216886252165, + 0.011535185389220715, + 0.0037396366242319345, + 0.004709035623818636, + 0.005872612353414297, + 0.017397645860910416, + 0.009438851848244667, + 0.008108023554086685, + 0.0038969451561570168, + 0.009785333648324013, + 0.003514362033456564, + 0.004052911419421434, + 0.0039900061674416065, + 0.0029390333220362663, + 0.007850171998143196, + 0.007246248424053192, + 0.0032418938353657722, + 0.0031578338239341974, + 0.0023356350138783455, + 0.002927966881543398, + 0.002430521184578538, + 0.00814606063067913, + 0.006953179836273193, + 0.003744577057659626, + 0.0028961480129510164, + 0.023497413843870163, + 0.01995835267007351, + 0.025102028623223305, + 0.037856705486774445, + 0.019754882901906967, + 0.02125757560133934, + 0.07580181956291199, + 0.018506459891796112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.07902117073535919 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight close-up, positioned at a slightly elevated angle, focuses on a polished stainless steel lid resting on a large stockpot atop a modern, black ceramic cooktop. Soft, warm afternoon sunlight slants through an out-of-frame window, casting a brilliant glow on the lid's curved surface, contrasting sharply with the even, neutral glow from the kitchen's overhead light. The movement begins delicately as the lid, featuring a simple black Bakelite handle, trembles almost imperceptibly. Then, with a sudden jolt, one rim lifts, releasing a sharp puff of white steam that briefly obscures the air before dissipating, and the lid drops with a distinct metallic clang. The movement begins as a steady, deliberate repetition, then evolves into a more frantic, rhythmic rattling as the pressure within builds, each movement more pronounced than the last, creating an atmosphere of tension and anticipation in the quiet kitchen.", + "seed": 33460, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_2/index_9580_seed_33460_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07087747268378734, + "vlm_logprob": -0.07087747268378734, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.07087747268378734, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06003103405237198 + ], + "opposite_logprob_tokens": [ + -3.435030937194824 + ], + "logprob_reward": -0.06003103405237198, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.06003103405237198, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07413472980260849, + 0.032388996332883835, + 0.11146843433380127, + 0.09850762039422989, + 0.0805882140994072, + 0.20781394839286804, + 0.18800026178359985, + 0.1366848349571228, + 0.026015570387244225, + 0.04439733922481537 + ], + "credit_map": [ + 0.017743531614542007, + 0.004611171782016754, + 0.005209938623011112, + 0.005162402987480164, + 0.006775293964892626, + 0.024327488616108894, + 0.0056002140045166016, + 0.004704691004008055, + 0.0061449105851352215, + 0.0023410713765770197, + 0.002675160299986601, + 0.004553551319986582, + 0.003053868655115366, + 0.007040795870125294, + 0.0030127542559057474, + 0.003566883970052004, + 0.024055685847997665, + 0.007146631367504597, + 0.01499965600669384, + 0.011487912386655807, + 0.007878380827605724, + 0.03100208193063736, + 0.007325134705752134, + 0.007572956848889589, + 0.016617942601442337, + 0.007294675800949335, + 0.003576494986191392, + 0.009980684146285057, + 0.010690667666494846, + 0.0291090477257967, + 0.01611744426190853, + 0.005120668560266495, + 0.008435633964836597, + 0.009489528834819794, + 0.008818031288683414, + 0.013267450965940952, + 0.009203089401125908, + 0.016783425584435463, + 0.008182420395314693, + 0.006408636923879385, + 0.016767777502536774, + 0.030636180192232132, + 0.024798819795250893, + 0.03211066126823425, + 0.03185123950242996, + 0.029297642409801483, + 0.021611444652080536, + 0.020740188658237457, + 0.017864365130662918, + 0.022535573691129684, + 0.016041280701756477, + 0.033974554389715195, + 0.022898303344845772, + 0.023760885000228882, + 0.03660732880234718, + 0.01431796420365572, + 0.011438514105975628, + 0.01685314066708088, + 0.006991348695009947, + 0.028129208832979202, + 0.01657673716545105, + 0.013236362487077713, + 0.036564119160175323, + 0.0068953996524214745, + 0.00192138715647161, + 0.0029797435272485018, + 0.0014870625454932451, + 0.003872770117595792, + 0.003849779488518834, + 0.003668049117550254, + 0.005975786130875349, + 0.002260990208014846, + 0.00425121933221817, + 0.00796960573643446, + 0.003137191291898489, + 0.00794457271695137, + 0.004576300270855427, + 0.00676517840474844, + 0.00653129443526268, + 0.0032219800632447004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.06003103405237198 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07905101776123047 + ], + "opposite_logprob_tokens": [ + -2.5790510177612305 + ], + "logprob_reward": -0.07905101776123047, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07905101776123047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07651793956756592, + 0.01628214679658413, + 0.019041549414396286, + 0.06239645928144455, + 0.24613766372203827, + 0.21736668050289154, + 0.09871049225330353, + 0.12387488782405853, + 0.06908924877643585, + 0.07058290392160416 + ], + "credit_map": [ + 0.017622370272874832, + 0.004790066741406918, + 0.005518930498510599, + 0.0071203685365617275, + 0.006889461074024439, + 0.025296758860349655, + 0.004626878071576357, + 0.0046531143598258495, + 0.002166029764339328, + 0.001317051355727017, + 0.001631168182939291, + 0.0029476704075932503, + 0.0015510638477280736, + 0.0029361434280872345, + 0.0021743380930274725, + 0.0015586826484650373, + 0.0025545931421220303, + 0.0014950711047276855, + 0.0019837887957692146, + 0.003226235043257475, + 0.002631834242492914, + 0.002998997690156102, + 0.002598455874249339, + 0.0015525753842666745, + 0.00855170376598835, + 0.0072243232280015945, + 0.005118881352245808, + 0.007555112708359957, + 0.01151938084512949, + 0.008268887177109718, + 0.009372870437800884, + 0.004785297438502312, + 0.028774093836545944, + 0.028325026854872704, + 0.022127795964479446, + 0.027956722304224968, + 0.040839605033397675, + 0.057460397481918335, + 0.02534961700439453, + 0.01530440617352724, + 0.018997754901647568, + 0.03630495071411133, + 0.0249444842338562, + 0.030623365193605423, + 0.0346807986497879, + 0.02752610854804516, + 0.02117769978940487, + 0.02311152219772339, + 0.008547082543373108, + 0.013530592434108257, + 0.012385144829750061, + 0.01912231557071209, + 0.010879752226173878, + 0.01392064057290554, + 0.012732022441923618, + 0.007592947222292423, + 0.011774354614317417, + 0.022174570709466934, + 0.006630400661379099, + 0.0167859997600317, + 0.014769232831895351, + 0.018067244440317154, + 0.020862119272351265, + 0.012810954824090004, + 0.005711119621992111, + 0.010661355219781399, + 0.008899369277060032, + 0.010876422747969627, + 0.008841214701533318, + 0.0074822804890573025, + 0.010535943321883678, + 0.006081536412239075, + 0.007709944155067205, + 0.012872179970145226, + 0.006798477377742529, + 0.010721408762037754, + 0.006917910650372505, + 0.011385424062609673, + 0.007789078168570995, + 0.006388482637703419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.07905101776123047 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297110468149185 + ], + "opposite_logprob_tokens": [ + -2.237971067428589 + ], + "logprob_reward": -0.11297110468149185, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297110468149185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13175718486309052, + 0.17704899609088898, + 0.05690652132034302, + 0.15570202469825745, + 0.12096583098173141, + 0.04286381974816322, + 0.07764804363250732, + 0.1764947474002838, + 0.04001801460981369, + 0.020594853907823563 + ], + "credit_map": [ + 0.027245692908763885, + 0.013168614357709885, + 0.006253817118704319, + 0.014823160134255886, + 0.021699266508221626, + 0.03057791292667389, + 0.011091627180576324, + 0.00689708860591054, + 0.025031182914972305, + 0.017885832116007805, + 0.013546647503972054, + 0.025893764570355415, + 0.026643289253115654, + 0.030376380309462547, + 0.01939968764781952, + 0.01827220618724823, + 0.005859979894012213, + 0.0052278474904596806, + 0.0037414247635751963, + 0.011378874070942402, + 0.007578379474580288, + 0.008612311445176601, + 0.008505952544510365, + 0.006001748610287905, + 0.02589578367769718, + 0.009420687332749367, + 0.016776613891124725, + 0.01408825721591711, + 0.016273651272058487, + 0.0410834364593029, + 0.01942436210811138, + 0.012739228084683418, + 0.007916971109807491, + 0.008691811934113503, + 0.006880161352455616, + 0.03500546142458916, + 0.03428248316049576, + 0.011182444170117378, + 0.00981674250215292, + 0.007189757190644741, + 0.0030506800394505262, + 0.0068747177720069885, + 0.004384297411888838, + 0.006576019339263439, + 0.0070988815277814865, + 0.006892594043165445, + 0.004203406628221273, + 0.003783223684877157, + 0.006955198012292385, + 0.010126330889761448, + 0.007215951103717089, + 0.016154279932379723, + 0.009289180859923363, + 0.011175235733389854, + 0.010885337367653847, + 0.005846528336405754, + 0.016448214650154114, + 0.03634553402662277, + 0.009545727632939816, + 0.0332828052341938, + 0.02491307631134987, + 0.027720486745238304, + 0.020389730110764503, + 0.007849195040762424, + 0.0027203739155083895, + 0.005851661320775747, + 0.0030095709953457117, + 0.006139765027910471, + 0.006449170410633087, + 0.005659674294292927, + 0.007558051962405443, + 0.002629748312756419, + 0.0021053687669336796, + 0.0026953062042593956, + 0.001610203762538731, + 0.0029438610654324293, + 0.002367790788412094, + 0.0029198969714343548, + 0.0027040692511945963, + 0.003248357679694891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.11297110468149185 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233097355812788 + ], + "opposite_logprob_tokens": [ + -3.7733097076416016 + ], + "logprob_reward": -0.0233097355812788, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233097355812788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060892388224601746, + 0.0312771238386631, + 0.01961096189916134, + 0.18217387795448303, + 0.11957788467407227, + 0.1519411951303482, + 0.15805856883525848, + 0.18111610412597656, + 0.07486046105623245, + 0.02049151435494423 + ], + "credit_map": [ + 0.00787923764437437, + 0.007856674492359161, + 0.004789787344634533, + 0.006886755581945181, + 0.013113409280776978, + 0.00877146515995264, + 0.00687671871855855, + 0.004718333948403597, + 0.004153253044933081, + 0.0032435059547424316, + 0.002348979702219367, + 0.004932428244501352, + 0.005359811708331108, + 0.004305160138756037, + 0.0037162266671657562, + 0.0032177551183849573, + 0.003544118022546172, + 0.0019281309796497226, + 0.0013223282294347882, + 0.00310347112827003, + 0.002888003597036004, + 0.0034783307928591967, + 0.0021603915374726057, + 0.001186187146231532, + 0.038892608135938644, + 0.010167639702558517, + 0.00917586125433445, + 0.018275780603289604, + 0.013569897040724754, + 0.06454047560691833, + 0.018902422860264778, + 0.008649174124002457, + 0.007848715409636497, + 0.008077244274318218, + 0.007606295868754387, + 0.03606387972831726, + 0.033323023468256, + 0.011968052014708519, + 0.010560652241110802, + 0.004130010027438402, + 0.015833387151360512, + 0.026032743975520134, + 0.017886731773614883, + 0.015314456075429916, + 0.02314729616045952, + 0.01735496334731579, + 0.016089610755443573, + 0.02028198353946209, + 0.01260035764425993, + 0.01981528475880623, + 0.012550176121294498, + 0.04281660169363022, + 0.017539052292704582, + 0.018563777208328247, + 0.022391727194190025, + 0.011781571432948112, + 0.017330925911664963, + 0.03673713281750679, + 0.009475037455558777, + 0.030333803966641426, + 0.0241258442401886, + 0.02771640196442604, + 0.02556436136364937, + 0.009832574985921383, + 0.0056455195881426334, + 0.011417488567531109, + 0.006907905917614698, + 0.01169461291283369, + 0.009596925228834152, + 0.01102769747376442, + 0.013133713975548744, + 0.0054365843534469604, + 0.0016973193269222975, + 0.0034914310090243816, + 0.0013165951240807772, + 0.0034826602786779404, + 0.0022914218716323376, + 0.0033155411947518587, + 0.0032865279354155064, + 0.0016100147040560842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.0233097355812788 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07902447134256363 + ], + "opposite_logprob_tokens": [ + -2.57902455329895 + ], + "logprob_reward": -0.07902447134256363, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07902447134256363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030524754896759987, + 0.0333920456469059, + 0.008446414954960346, + 0.04454930126667023, + 0.04133874922990799, + 0.4463624656200409, + 0.13320976495742798, + 0.18672223389148712, + 0.045402128249406815, + 0.03005213662981987 + ], + "credit_map": [ + 0.005521838087588549, + 0.0021286157425493, + 0.00265047256834805, + 0.003713445272296667, + 0.0037760476116091013, + 0.0066853673197329044, + 0.003059003036469221, + 0.0029899626970291138, + 0.004104130901396275, + 0.0034188837744295597, + 0.004302253481000662, + 0.004681732505559921, + 0.005485983099788427, + 0.0037213119212538004, + 0.004335617180913687, + 0.003342132782563567, + 0.001265929196961224, + 0.0005858298391103745, + 0.0011553193908184767, + 0.0014940346591174603, + 0.0006886537303216755, + 0.001601124764420092, + 0.000829791184514761, + 0.0008257321314886212, + 0.008408121764659882, + 0.003297666786238551, + 0.002102285623550415, + 0.004128838889300823, + 0.004368586000055075, + 0.01429652702063322, + 0.005236519500613213, + 0.0027107575442641973, + 0.00536100659519434, + 0.004434044007211924, + 0.0037510173860937357, + 0.006033558864146471, + 0.004810217302292585, + 0.00903540849685669, + 0.004816939122974873, + 0.003096560947597027, + 0.04247374087572098, + 0.0752095878124237, + 0.04306931048631668, + 0.047330714762210846, + 0.0716167539358139, + 0.06352992355823517, + 0.04375242814421654, + 0.05937999486923218, + 0.00994782242923975, + 0.015477131120860577, + 0.011484494432806969, + 0.03462915122509003, + 0.01408994197845459, + 0.011729287914931774, + 0.028060123324394226, + 0.007791819982230663, + 0.017995433881878853, + 0.0385051891207695, + 0.011591242626309395, + 0.031245099380612373, + 0.02589532732963562, + 0.02856779471039772, + 0.02212899550795555, + 0.010793155990540981, + 0.0038786621298640966, + 0.006937211379408836, + 0.004384174477308989, + 0.006324059329926968, + 0.005774296820163727, + 0.006527280434966087, + 0.007823143154382706, + 0.003753300756216049, + 0.0029807882383465767, + 0.004661615006625652, + 0.0020967498421669006, + 0.0054733240976929665, + 0.003193433629348874, + 0.005022658966481686, + 0.003928039222955704, + 0.002695525297895074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.07902447134256363 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight close-up, positioned at a slightly elevated angle, focuses on a polished stainless steel lid resting on a large stockpot atop a modern, black ceramic cooktop. Soft, warm afternoon sunlight slants through an out-of-frame window, casting a brilliant glow on the lid's curved surface, contrasting sharply with the even, neutral glow from the kitchen's overhead light. The movement begins delicately as the lid, featuring a simple black Bakelite handle, trembles almost imperceptibly. Then, with a sudden jolt, one rim lifts, releasing a sharp puff of white steam that briefly obscures the air before dissipating, and the lid drops with a distinct metallic clang. The movement begins as a steady, deliberate repetition, then evolves into a more frantic, rhythmic rattling as the pressure within builds, each movement more pronounced than the last, creating an atmosphere of tension and anticipation in the quiet kitchen.", + "seed": 33461, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_2/index_9580_seed_33461_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.24076521582901478, + "vlm_logprob": -0.24076521582901478, + "vlm_yes_no_margin": 1.825, + "vlm_token_credit": -0.24076521582901478, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16864418983459473 + ], + "opposite_logprob_tokens": [ + -2.1686441898345947 + ], + "logprob_reward": -0.16864418983459473, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.16864418983459473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026729894801974297, + 0.0289141982793808, + 0.03742394223809242, + 0.026622561737895012, + 0.053371816873550415, + 0.43184947967529297, + 0.09645447880029678, + 0.0637538880109787, + 0.16221106052398682, + 0.07266865670681 + ], + "credit_map": [ + 0.005030991043895483, + 0.0033962219022214413, + 0.0022210157476365566, + 0.007600339595228434, + 0.0023189536295831203, + 0.002256706589832902, + 0.0026154485531151295, + 0.001290217973291874, + 0.004534073173999786, + 0.006939105223864317, + 0.003058480331674218, + 0.005405308213084936, + 0.002276025479659438, + 0.002939296420663595, + 0.00234223622828722, + 0.0014196739066392183, + 0.005805517081171274, + 0.006419091019779444, + 0.003272303147241473, + 0.007772014942020178, + 0.0034396618138998747, + 0.003047061851248145, + 0.005597934126853943, + 0.0020703577902168036, + 0.004124751780182123, + 0.00400888966396451, + 0.0025734419468790293, + 0.004109288565814495, + 0.002064335159957409, + 0.006275271065533161, + 0.0021646686363965273, + 0.0013019152684137225, + 0.009602394886314869, + 0.009849738329648972, + 0.00294326338917017, + 0.012860803864896297, + 0.0072600566782057285, + 0.004682101774960756, + 0.003987995442003012, + 0.002185460180044174, + 0.16980832815170288, + 0.15853840112686157, + 0.008829569444060326, + 0.01669052615761757, + 0.02393369749188423, + 0.024372538551688194, + 0.01792232133448124, + 0.011754091829061508, + 0.007700248621404171, + 0.019499678164720535, + 0.0047263894230127335, + 0.00808622594922781, + 0.014667048119008541, + 0.023615745827555656, + 0.011253651231527328, + 0.006905495189130306, + 0.008330031298100948, + 0.00793939083814621, + 0.003064178628847003, + 0.009257684461772442, + 0.006151200737804174, + 0.007731118705123663, + 0.01615672931075096, + 0.005123551934957504, + 0.01572555862367153, + 0.030069321393966675, + 0.009801431559026241, + 0.02772790938615799, + 0.01433740183711052, + 0.027243560180068016, + 0.024015158414840698, + 0.013290712609887123, + 0.0073731704615056515, + 0.01794208586215973, + 0.0056545063853263855, + 0.009618541225790977, + 0.0075506544671952724, + 0.009553194046020508, + 0.008276944048702717, + 0.006699555553495884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.16864418983459473 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6329267024993896 + ], + "opposite_logprob_tokens": [ + -0.7579267024993896 + ], + "logprob_reward": -0.6329267024993896, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6329267024993896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049814578145742416, + 0.020838512107729912, + 0.04629971832036972, + 0.029825778678059578, + 0.02112416923046112, + 0.2105713188648224, + 0.0629788488149643, + 0.0508260540664196, + 0.4851032793521881, + 0.022617822512984276 + ], + "credit_map": [ + 0.00891875196248293, + 0.013213739730417728, + 0.003917638212442398, + 0.007680171635001898, + 0.00474491436034441, + 0.006008395925164223, + 0.0032235586550086737, + 0.0021074065007269382, + 0.003467045957222581, + 0.0029182678554207087, + 0.0020621083676815033, + 0.00443998072296381, + 0.001811136375181377, + 0.0030594603158533573, + 0.0018865838646888733, + 0.0011939278338104486, + 0.010270129889249802, + 0.017043128609657288, + 0.0026716901920735836, + 0.006504622753709555, + 0.0023648191709071398, + 0.003265857230871916, + 0.0029349985998123884, + 0.0012444721069186926, + 0.0045198784209787846, + 0.004414267838001251, + 0.003629181766882539, + 0.004474085755646229, + 0.003789568319916725, + 0.002678354037925601, + 0.0037112669087946415, + 0.0026091772597283125, + 0.0033660412300378084, + 0.002948594279587269, + 0.001330072176642716, + 0.003932921215891838, + 0.0028369491919875145, + 0.003172140335664153, + 0.0025495244190096855, + 0.000987927196547389, + 0.08225664496421814, + 0.07720685005187988, + 0.004273405764251947, + 0.00942082330584526, + 0.011758489534258842, + 0.011611315421760082, + 0.008505054749548435, + 0.005538728088140488, + 0.005933151580393314, + 0.011380805633962154, + 0.003501555183902383, + 0.006562029477208853, + 0.010538358241319656, + 0.012702824547886848, + 0.007857544347643852, + 0.004502580966800451, + 0.004726734012365341, + 0.007462876848876476, + 0.0025952616706490517, + 0.006692021153867245, + 0.006472482345998287, + 0.009955450892448425, + 0.008529240265488625, + 0.004391984082758427, + 0.04840321093797684, + 0.08410682529211044, + 0.02628123015165329, + 0.09984228760004044, + 0.037912797182798386, + 0.06259220838546753, + 0.08813009411096573, + 0.03783461079001427, + 0.002032801741734147, + 0.005549169611185789, + 0.0014443452237173915, + 0.0034015867859125137, + 0.0023963700514286757, + 0.0029283640906214714, + 0.002878558589145541, + 0.0019866256043314934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.6329267024993896 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20165608823299408 + ], + "opposite_logprob_tokens": [ + -1.7016561031341553 + ], + "logprob_reward": -0.20165608823299408, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20165608823299408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05666414648294449, + 0.060139838606119156, + 0.10249224305152893, + 0.024437736719846725, + 0.23987814784049988, + 0.34243738651275635, + 0.05538710206747055, + 0.05872069671750069, + 0.035921741276979446, + 0.023920994251966476 + ], + "credit_map": [ + 0.009141497313976288, + 0.010524616576731205, + 0.0036048688925802708, + 0.012479824014008045, + 0.006436654832214117, + 0.0060374196618795395, + 0.0058919829316437244, + 0.0025472778361290693, + 0.011046238243579865, + 0.0070941452868282795, + 0.007552111987024546, + 0.017437724396586418, + 0.003561175661161542, + 0.004774022847414017, + 0.0057135336101055145, + 0.0029608856420964003, + 0.022315580397844315, + 0.03051774762570858, + 0.006311376579105854, + 0.018682802096009254, + 0.004869015421718359, + 0.006725736428052187, + 0.009860027581453323, + 0.0032099513337016106, + 0.004500360693782568, + 0.004811764694750309, + 0.0020432965829968452, + 0.005527406465262175, + 0.001372308237478137, + 0.0018045251490548253, + 0.0033314754255115986, + 0.0010465977247804403, + 0.03600497171282768, + 0.04262556880712509, + 0.015335563570261002, + 0.03801674395799637, + 0.03860766813158989, + 0.03849053382873535, + 0.019817857071757317, + 0.010979222133755684, + 0.07646527886390686, + 0.08379074186086655, + 0.020408200100064278, + 0.04424956440925598, + 0.03517526015639305, + 0.021334897726774216, + 0.03963189572095871, + 0.021381527185440063, + 0.006138975266367197, + 0.00802452489733696, + 0.003878501011058688, + 0.007700038142502308, + 0.0102618969976902, + 0.007571100722998381, + 0.006878839805722237, + 0.004933218937367201, + 0.006796823814511299, + 0.011108811013400555, + 0.0034255078062415123, + 0.005537870340049267, + 0.007721906527876854, + 0.010770842432975769, + 0.007134555373340845, + 0.006224369630217552, + 0.004416696727275848, + 0.008314927108585835, + 0.0017299031605944037, + 0.007008861284703016, + 0.003390998113900423, + 0.0034541303757578135, + 0.005249709356576204, + 0.0023565124720335007, + 0.0025701771955937147, + 0.006523588206619024, + 0.0019987516570836306, + 0.002505019772797823, + 0.0025883540511131287, + 0.0028056814335286617, + 0.0030677332542836666, + 0.0018616871675476432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.20165608823299408 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02062002755701542 + ], + "opposite_logprob_tokens": [ + -3.895620107650757 + ], + "logprob_reward": -0.02062002755701542, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02062002755701542, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016409991309046745, + 0.0196656733751297, + 0.040272701531648636, + 0.010392873547971249, + 0.04943177103996277, + 0.6136214733123779, + 0.06016954407095909, + 0.03879277780652046, + 0.134868323802948, + 0.016374846920371056 + ], + "credit_map": [ + 0.0023307555820792913, + 0.0029264301992952824, + 0.001710717100650072, + 0.0024859881959855556, + 0.001843793666921556, + 0.002615930512547493, + 0.001740583567880094, + 0.0007557907956652343, + 0.003659708658233285, + 0.00464256526902318, + 0.0012054855469614267, + 0.004473503679037094, + 0.001098067732527852, + 0.0020195855759084225, + 0.0016983592649921775, + 0.0008683964260853827, + 0.006887856405228376, + 0.008879645727574825, + 0.003834352595731616, + 0.00725554907694459, + 0.0032315051648765802, + 0.0034795913379639387, + 0.004690391011536121, + 0.002013805089518428, + 0.0015572827542200685, + 0.0018749526934698224, + 0.0008064256398938596, + 0.002031219657510519, + 0.0009585396037437022, + 0.001089514815248549, + 0.0013803975889459252, + 0.000694539921823889, + 0.008457059040665627, + 0.009249746799468994, + 0.0025463844649493694, + 0.011110799387097359, + 0.00706626009196043, + 0.004964299965649843, + 0.0037433786783367395, + 0.0022938349284231663, + 0.2450217753648758, + 0.2298436462879181, + 0.011302153579890728, + 0.02320815622806549, + 0.03354460746049881, + 0.03399661183357239, + 0.020760657265782356, + 0.01594383269548416, + 0.0064580985344946384, + 0.008361597545444965, + 0.0035622671712189913, + 0.005919846706092358, + 0.012272325344383717, + 0.00867376010864973, + 0.009313623420894146, + 0.005608018487691879, + 0.004888113588094711, + 0.00776927825063467, + 0.002801460912451148, + 0.004774183500558138, + 0.004151580389589071, + 0.005587487481534481, + 0.005829696077853441, + 0.002990973647683859, + 0.013377786614000797, + 0.02612430602312088, + 0.007475437596440315, + 0.026582134887576103, + 0.010604117065668106, + 0.0166794303804636, + 0.023864207789301872, + 0.010160880163311958, + 0.0017992923967540264, + 0.003817520570009947, + 0.0013318239944055676, + 0.0020445468835532665, + 0.0016812441172078252, + 0.0019572440069168806, + 0.0023460853844881058, + 0.0013970868894830346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.02062002755701542 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17997907102108002 + ], + "opposite_logprob_tokens": [ + -1.8049790859222412 + ], + "logprob_reward": -0.17997907102108002, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17997907102108002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.005734069738537073, + 0.008242194540798664, + 0.020269375294446945, + 0.009989168494939804, + 0.05387253686785698, + 0.7997763156890869, + 0.042236264795064926, + 0.028016934171319008, + 0.025004951283335686, + 0.006858183071017265 + ], + "credit_map": [ + 0.0009045126498676836, + 0.0013340885052457452, + 0.0003864777972921729, + 0.0010534647153690457, + 0.0005742310313507915, + 0.0006055709673091769, + 0.0006308744195848703, + 0.00024485011817887425, + 0.0015908004716038704, + 0.0024317537900060415, + 0.0006106715882197022, + 0.001192708732560277, + 0.0007046492537483573, + 0.0007634190260432661, + 0.000642918050289154, + 0.00030527333728969097, + 0.0042679402977228165, + 0.006081626284867525, + 0.0011644673068076372, + 0.0031486244406551123, + 0.001201887265779078, + 0.0014434767654165626, + 0.0023635125253349543, + 0.0005978398839943111, + 0.0016997824423015118, + 0.0019908789545297623, + 0.0006159962504170835, + 0.0016891268314793706, + 0.0006962926709093153, + 0.0008961257408373058, + 0.0017345008673146367, + 0.0006664652610197663, + 0.009248675778508186, + 0.010734283365309238, + 0.0029057702049613, + 0.012110620737075806, + 0.007419304922223091, + 0.004254043567925692, + 0.004679857287555933, + 0.0025199803058058023, + 0.3166736364364624, + 0.2959310710430145, + 0.015039248391985893, + 0.030964313074946404, + 0.04775850847363472, + 0.043819986283779144, + 0.029003432020545006, + 0.020586101338267326, + 0.003043239703401923, + 0.008920914493501186, + 0.0017471332103013992, + 0.003597357077524066, + 0.00816404726356268, + 0.00941486470401287, + 0.00456071924418211, + 0.002787988632917404, + 0.003158268751576543, + 0.005426616873592138, + 0.0016508052358403802, + 0.00254331948235631, + 0.003556460840627551, + 0.004582033958286047, + 0.0045086415484547615, + 0.002590787597000599, + 0.0026934538036584854, + 0.004679864272475243, + 0.0013658468378707767, + 0.005360049661248922, + 0.001828002859838307, + 0.0027338771615177393, + 0.004434801172465086, + 0.0019090553978458047, + 0.0007622050470672548, + 0.00128368497826159, + 0.0005139599088579416, + 0.0011449460871517658, + 0.0007147478172555566, + 0.0008262600167654455, + 0.0009340341784991324, + 0.0006783450953662395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.17997907102108002 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, tight close-up, positioned at a slightly elevated angle, focuses on a polished stainless steel lid resting on a large stockpot atop a modern, black ceramic cooktop. Soft, warm afternoon sunlight slants through an out-of-frame window, casting a brilliant glow on the lid's curved surface, contrasting sharply with the even, neutral glow from the kitchen's overhead light. The movement begins delicately as the lid, featuring a simple black Bakelite handle, trembles almost imperceptibly. Then, with a sudden jolt, one rim lifts, releasing a sharp puff of white steam that briefly obscures the air before dissipating, and the lid drops with a distinct metallic clang. The movement begins as a steady, deliberate repetition, then evolves into a more frantic, rhythmic rattling as the pressure within builds, each movement more pronounced than the last, creating an atmosphere of tension and anticipation in the quiet kitchen.", + "seed": 33462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_2/index_9580_seed_33462_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05815505385398865, + "vlm_logprob": -0.05815505385398865, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.05815505385398865, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06763841956853867 + ], + "opposite_logprob_tokens": [ + -3.317638397216797 + ], + "logprob_reward": -0.06763841956853867, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06763841956853867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1516149789094925, + 0.05353808403015137, + 0.07416300475597382, + 0.03508047014474869, + 0.1838598996400833, + 0.23745203018188477, + 0.07745403796434402, + 0.08063755184412003, + 0.0765596255660057, + 0.02964033931493759 + ], + "credit_map": [ + 0.017295176163315773, + 0.025891350582242012, + 0.011353004723787308, + 0.02213461697101593, + 0.018341965973377228, + 0.022178206592798233, + 0.02155967243015766, + 0.012860974296927452, + 0.006760364398360252, + 0.00511097302660346, + 0.005074265878647566, + 0.009455342777073383, + 0.0056482651270926, + 0.008164468221366405, + 0.007877924479544163, + 0.0054464759305119514, + 0.012727736495435238, + 0.008705676533281803, + 0.009453398175537586, + 0.009625802747905254, + 0.006533736363053322, + 0.0069332849234342575, + 0.012568117119371891, + 0.007615257054567337, + 0.004862263333052397, + 0.0033835559152066708, + 0.004936482757329941, + 0.005968851502984762, + 0.004121996462345123, + 0.0037389928475022316, + 0.0048614805564284325, + 0.003206848166882992, + 0.022862808778882027, + 0.019457167014479637, + 0.014414847828447819, + 0.04327317327260971, + 0.01863788813352585, + 0.021533017978072166, + 0.02708631567656994, + 0.016594668850302696, + 0.03446372225880623, + 0.03938152641057968, + 0.011250886134803295, + 0.02542673982679844, + 0.05830882117152214, + 0.03814372420310974, + 0.01777745969593525, + 0.012699127197265625, + 0.013810799457132816, + 0.00933343917131424, + 0.007412591949105263, + 0.010490508750081062, + 0.01166913565248251, + 0.009328118525445461, + 0.008859618566930294, + 0.006549824960529804, + 0.010015211068093777, + 0.010719849728047848, + 0.005800233222544193, + 0.013614910654723644, + 0.011010694317519665, + 0.013463991694152355, + 0.008460558019578457, + 0.007552104536443949, + 0.00826430693268776, + 0.010093609802424908, + 0.0050428761169314384, + 0.015260335989296436, + 0.008415068499743938, + 0.013181588612496853, + 0.010082597844302654, + 0.006219243165105581, + 0.004415220580995083, + 0.003426040057092905, + 0.002377251163125038, + 0.004643126390874386, + 0.0030174481216818094, + 0.003915335051715374, + 0.0050264629535377026, + 0.0028194542974233627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.06763841956853867 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029858142137527466 + ], + "opposite_logprob_tokens": [ + -3.529858112335205 + ], + "logprob_reward": -0.029858142137527466, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029858142137527466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20286813378334045, + 0.10974647104740143, + 0.05975997447967529, + 0.01846691034734249, + 0.18961404263973236, + 0.14629694819450378, + 0.08750704675912857, + 0.06112237647175789, + 0.08644011616706848, + 0.0381779819726944 + ], + "credit_map": [ + 0.02528229169547558, + 0.023219937458634377, + 0.01576506718993187, + 0.03983316943049431, + 0.020070265978574753, + 0.032347943633794785, + 0.029064662754535675, + 0.01728477142751217, + 0.01581645756959915, + 0.011638147756457329, + 0.013330253772437572, + 0.016702478751540184, + 0.010211576707661152, + 0.0170099139213562, + 0.015207206830382347, + 0.009830431081354618, + 0.009413393214344978, + 0.0061893039382994175, + 0.006459448486566544, + 0.009740609675645828, + 0.005410412792116404, + 0.006089074071496725, + 0.010882155038416386, + 0.005575573071837425, + 0.0029396596364676952, + 0.0017123444704338908, + 0.001937414170242846, + 0.0026524222921580076, + 0.001865388941951096, + 0.002288773190230131, + 0.0035530817694962025, + 0.001517824362963438, + 0.022491060197353363, + 0.022428657859563828, + 0.01290318462997675, + 0.043472107499837875, + 0.019817747175693512, + 0.017844008281826973, + 0.03330563008785248, + 0.01735164038836956, + 0.020364869385957718, + 0.02221800945699215, + 0.00913733709603548, + 0.01862253248691559, + 0.028197769075632095, + 0.018637005239725113, + 0.015167986042797565, + 0.013951432891190052, + 0.01633174903690815, + 0.009889386594295502, + 0.010837974958121777, + 0.011620805598795414, + 0.011403907090425491, + 0.011789659969508648, + 0.007757206447422504, + 0.007876350544393063, + 0.009015152230858803, + 0.007174802012741566, + 0.005487213842570782, + 0.008215043693780899, + 0.006815518252551556, + 0.008076167665421963, + 0.008126967586576939, + 0.008211505599319935, + 0.009282708168029785, + 0.011099656112492085, + 0.0058518885634839535, + 0.016673177480697632, + 0.009622210636734962, + 0.014842902310192585, + 0.0115415770560503, + 0.007525982800871134, + 0.0071267858147621155, + 0.004122220911085606, + 0.003109508892521262, + 0.00498467031866312, + 0.0038836116436868906, + 0.004778662696480751, + 0.0063961585983633995, + 0.0037763623986393213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.029858142137527466 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062088727951049805 + ], + "opposite_logprob_tokens": [ + -2.81208872795105 + ], + "logprob_reward": -0.062088727951049805, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062088727951049805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10551182180643082, + 0.03768903389573097, + 0.13008014857769012, + 0.32595115900039673, + 0.22023780643939972, + 0.039834652096033096, + 0.059335965663194656, + 0.039734307676553726, + 0.019827112555503845, + 0.021798033267259598 + ], + "credit_map": [ + 0.01435108296573162, + 0.010332651436328888, + 0.009881514124572277, + 0.015478403307497501, + 0.009044752456247807, + 0.012904152274131775, + 0.0150827057659626, + 0.018436558544635773, + 0.0047310092486441135, + 0.003541911719366908, + 0.004936557728797197, + 0.005532981362193823, + 0.003863904858008027, + 0.006359905935823917, + 0.005550835281610489, + 0.003171925898641348, + 0.021839730441570282, + 0.018707869574427605, + 0.012319490313529968, + 0.022267824038863182, + 0.015221991576254368, + 0.010809989646077156, + 0.018044721335172653, + 0.010868537239730358, + 0.039406269788742065, + 0.04044797271490097, + 0.04067007824778557, + 0.0498841367661953, + 0.04523390159010887, + 0.041340041905641556, + 0.043714385479688644, + 0.02525438740849495, + 0.033980321139097214, + 0.03335173428058624, + 0.013883540406823158, + 0.03950952738523483, + 0.034332968294620514, + 0.01877572387456894, + 0.03444041684269905, + 0.011963577009737492, + 0.00656492542475462, + 0.005499056074768305, + 0.002520479029044509, + 0.005070216488093138, + 0.007435190491378307, + 0.005740937776863575, + 0.0038512288592755795, + 0.00315261771902442, + 0.007862172089517117, + 0.00821797177195549, + 0.005261103622615337, + 0.010937189683318138, + 0.006745328661054373, + 0.009500520303845406, + 0.006073492579162121, + 0.004738186486065388, + 0.005123901646584272, + 0.004909798502922058, + 0.002743088873103261, + 0.007510125171393156, + 0.005096429493278265, + 0.006074940785765648, + 0.004176934715360403, + 0.004099088255316019, + 0.0025160731747746468, + 0.0027852661442011595, + 0.0016001659678295255, + 0.00372351985424757, + 0.002380550839006901, + 0.002554929116740823, + 0.0022715553641319275, + 0.001995053142309189, + 0.0037884835619479418, + 0.00230867275968194, + 0.0019804821349680424, + 0.0024757490027695894, + 0.002097164513543248, + 0.0034542500507086515, + 0.0028719280380755663, + 0.0028213043697178364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.062088727951049805 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018209077417850494 + ], + "opposite_logprob_tokens": [ + -4.018208980560303 + ], + "logprob_reward": -0.018209077417850494, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018209077417850494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13959459960460663, + 0.054031532257795334, + 0.04525597393512726, + 0.03359610214829445, + 0.3578628897666931, + 0.11604718118906021, + 0.07216887921094894, + 0.07562483847141266, + 0.07271566241979599, + 0.03310234099626541 + ], + "credit_map": [ + 0.016114840283989906, + 0.021258151158690453, + 0.010723737068474293, + 0.024826662614941597, + 0.022178202867507935, + 0.013619683682918549, + 0.02159661054611206, + 0.009276709519326687, + 0.006857651751488447, + 0.007156920153647661, + 0.004567813593894243, + 0.011079260148108006, + 0.005711832083761692, + 0.005326919257640839, + 0.009056820534169674, + 0.004274314269423485, + 0.0073498310521245, + 0.004754418972879648, + 0.0063150241039693356, + 0.005708477925509214, + 0.003895603120326996, + 0.004889478906989098, + 0.007969017140567303, + 0.004374122712761164, + 0.0046918876469135284, + 0.004696671385318041, + 0.0032682884484529495, + 0.005586227867752314, + 0.0052100238390266895, + 0.003856577444821596, + 0.004008037503808737, + 0.002278388012200594, + 0.053995903581380844, + 0.0480535514652729, + 0.021532593294978142, + 0.07876607030630112, + 0.03349396213889122, + 0.03256712481379509, + 0.066072978079319, + 0.023380694910883904, + 0.012683822773396969, + 0.018387669697403908, + 0.006707525346428156, + 0.016319042071700096, + 0.025338442996144295, + 0.01728271134197712, + 0.011457196436822414, + 0.007870765402913094, + 0.011902352795004845, + 0.01051775086671114, + 0.004637746140360832, + 0.009688853286206722, + 0.00969177670776844, + 0.011903396807610989, + 0.00735284062102437, + 0.006474163383245468, + 0.009021452628076077, + 0.008609959855675697, + 0.006786554120481014, + 0.014283860102295876, + 0.0072569092735648155, + 0.011984399519860744, + 0.009492131881415844, + 0.008189569227397442, + 0.0072945193387568, + 0.010356246493756771, + 0.004845073446631432, + 0.017639944329857826, + 0.00736184511333704, + 0.009540393948554993, + 0.009394455701112747, + 0.0062831914983689785, + 0.0051133339293301105, + 0.0036928788758814335, + 0.002713504945859313, + 0.005195661447942257, + 0.004080592188984156, + 0.0036606346257030964, + 0.003760078689083457, + 0.004885658621788025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.018209077417850494 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static, tight close-up of a polished stainless steel lid on a large stockpot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set on a modern black ceramic cooktop with soft warm sunlight and neutral overhead lighting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lid feature a simple black Bakelite handle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the lid initially tremble imperceptibly before suddenly lifting one rim to release a puff of white steam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the lid drop with a metallic clang and then begin a frantic, rhythmic rattling motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129809021949768 + ], + "opposite_logprob_tokens": [ + -2.237980842590332 + ], + "logprob_reward": -0.1129809021949768, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129809021949768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11492690443992615, + 0.04466872662305832, + 0.09250298887491226, + 0.03492538630962372, + 0.14360834658145905, + 0.1543809324502945, + 0.19284071028232574, + 0.09795928001403809, + 0.09258843958377838, + 0.0315982960164547 + ], + "credit_map": [ + 0.013478389009833336, + 0.009516503661870956, + 0.009620619006454945, + 0.02425878494977951, + 0.010017652064561844, + 0.01662004552781582, + 0.018960684537887573, + 0.012454233132302761, + 0.006666647735983133, + 0.004768582060933113, + 0.004758246708661318, + 0.005754877347499132, + 0.004405520390719175, + 0.006747914478182793, + 0.007193449884653091, + 0.004373491741716862, + 0.016788214445114136, + 0.01358240656554699, + 0.009052122011780739, + 0.014482278376817703, + 0.01139958668500185, + 0.009412484243512154, + 0.009826891124248505, + 0.007959009148180485, + 0.00458186212927103, + 0.004340391606092453, + 0.003679972607642412, + 0.005383082665503025, + 0.005168294999748468, + 0.004991688299924135, + 0.004134015645831823, + 0.0026460827793926, + 0.013503668829798698, + 0.014798853546380997, + 0.013874434866011143, + 0.026926835998892784, + 0.01638617366552353, + 0.014433815144002438, + 0.023718055337667465, + 0.01996651664376259, + 0.02011438086628914, + 0.022113773971796036, + 0.012820757925510406, + 0.02491683140397072, + 0.022536935284733772, + 0.02199378050863743, + 0.015758715569972992, + 0.014125755988061428, + 0.029071660712361336, + 0.026346897706389427, + 0.015337638556957245, + 0.03453003242611885, + 0.021400313824415207, + 0.03224887698888779, + 0.019517814740538597, + 0.01438748650252819, + 0.014937366358935833, + 0.01302213966846466, + 0.010039749555289745, + 0.013379727490246296, + 0.012358942069113255, + 0.012910844758152962, + 0.011592642404139042, + 0.009717869572341442, + 0.009755473583936691, + 0.011892532929778099, + 0.005949920509010553, + 0.018408389762043953, + 0.012400606647133827, + 0.016210466623306274, + 0.009340464137494564, + 0.008630583062767982, + 0.003935347311198711, + 0.004215487278997898, + 0.0028182887472212315, + 0.005210034549236298, + 0.004387365188449621, + 0.005111809819936752, + 0.003347944002598524, + 0.0025720230769366026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.1129809021949768 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_3/index_tensor([6154, 6154])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_3/index_tensor([6154, 6154])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..084f7ac9f01571ed4500f8d7afe95d8c708d64fb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_3/index_tensor([6154, 6154])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide shot, slowly and steadily pulling back to an extreme wide angle, captures the catastrophic failure of a massive multi-engine seaplane as it tumbles out of control against a turbulent ochre sky. The entire scene is rendered in the gritty, painterly style of Ashley Wood, with thick, visible brushstrokes and smudged charcoal outlines giving each element a tactile, visceral quality. The seaplane's hulking fuselage, its riveted metal skin scarred, weathered, and stained with soot, rolls violently on its side, one wing hanging precariously low, its engines sputtering and black smoke. The plane's chaotic descent is a desperate, labyrinthine dance against a roiling vortex of sickly yellow and dirty brown clouds. As the viewpoint shifts away, the vast expanse of the storm is revealed, the light an aimless, oppressive glow filtered through a toxic haze. The seaplane, once the dominant subject, gradually shrinks, its violent struggle reduced to a silent, insignificant detail, contrasting with the vast, indifferent, beautiful brutalist sky, emphasizing a profound sense of chaotic doom and isolation.", + "seed": 17998, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_3/index_6154_seed_17998_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2737212519859895, + "vlm_logprob": -0.2737212519859895, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.2737212519859895, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0967535451054573 + ], + "opposite_logprob_tokens": [ + -2.5967535972595215 + ], + "logprob_reward": -0.0967535451054573, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0967535451054573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07244903594255447, + 0.06548886746168137, + 0.3374176323413849, + 0.08057790249586105, + 0.10731962323188782, + 0.0304561797529459, + 0.06197222322225571, + 0.057824909687042236, + 0.14741025865077972, + 0.039083417505025864 + ], + "credit_map": [ + 0.00795766245573759, + 0.00831853412091732, + 0.007067715283483267, + 0.007919362746179104, + 0.01331678032875061, + 0.010929147712886333, + 0.009661203250288963, + 0.007278628181666136, + 0.007359207607805729, + 0.014138826169073582, + 0.006007715594023466, + 0.0068824272602796555, + 0.007920882664620876, + 0.007422166410833597, + 0.009263052605092525, + 0.006494581233710051, + 0.05510350689291954, + 0.0465613529086113, + 0.0701628029346466, + 0.02971397526562214, + 0.03870806097984314, + 0.028695356100797653, + 0.03232964128255844, + 0.036142926663160324, + 0.012571053579449654, + 0.010102939791977406, + 0.014116778038442135, + 0.00870132353156805, + 0.008215432986617088, + 0.009893269278109074, + 0.006921310443431139, + 0.010055790655314922, + 0.009114943444728851, + 0.013602811843156815, + 0.013751644641160965, + 0.011539438739418983, + 0.017809415236115456, + 0.017432790249586105, + 0.00836991798132658, + 0.015698662027716637, + 0.004945370834320784, + 0.002513737417757511, + 0.0030509834177792072, + 0.003080480033531785, + 0.0031659018713980913, + 0.004301880020648241, + 0.0028788188938051462, + 0.00651900889351964, + 0.00631409976631403, + 0.006381509359925985, + 0.007943985052406788, + 0.008754397742450237, + 0.007205537520349026, + 0.012010350823402405, + 0.004183920565992594, + 0.009178422391414642, + 0.004473357927054167, + 0.009483443573117256, + 0.00525639159604907, + 0.005899192299693823, + 0.009902223013341427, + 0.01147968228906393, + 0.005886945873498917, + 0.005443671252578497, + 0.02634672448039055, + 0.016390597447752953, + 0.019159022718667984, + 0.012385251931846142, + 0.013736093416810036, + 0.03204897418618202, + 0.010068955831229687, + 0.0172746479511261, + 0.003922416362911463, + 0.006399974226951599, + 0.0036731120198965073, + 0.005871240049600601, + 0.006382436957210302, + 0.0055025480687618256, + 0.0042500062845647335, + 0.003081684000790119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 4 + ], + "selected_reward": -0.0967535451054573 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033728182315826416 + ], + "opposite_logprob_tokens": [ + -3.4087281227111816 + ], + "logprob_reward": -0.033728182315826416, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033728182315826416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032487448304891586, + 0.019502682611346245, + 0.032086875289678574, + 0.037385500967502594, + 0.042557355016469955, + 0.05472385138273239, + 0.08420497924089432, + 0.04939821735024452, + 0.6010714173316956, + 0.04658162221312523 + ], + "credit_map": [ + 0.003819263307377696, + 0.00282710581086576, + 0.004646893125027418, + 0.004063732456415892, + 0.005936580244451761, + 0.0034820048604160547, + 0.004605764523148537, + 0.0031061051413416862, + 0.0028561041690409184, + 0.0021698044147342443, + 0.002108657732605934, + 0.0021921563893556595, + 0.0025176983326673508, + 0.003112085396423936, + 0.0024837965611368418, + 0.002062381012365222, + 0.004163752775639296, + 0.003913465887308121, + 0.007564218249171972, + 0.004179151728749275, + 0.003824861254543066, + 0.002716806484386325, + 0.001978076994419098, + 0.0037465449422597885, + 0.0056651378981769085, + 0.004653987940400839, + 0.005667451769113541, + 0.003523113438859582, + 0.005024525802582502, + 0.004539964720606804, + 0.00368814286775887, + 0.004623179789632559, + 0.007037379778921604, + 0.0049693095497787, + 0.009549417532980442, + 0.0038528149016201496, + 0.004281282890588045, + 0.004353607073426247, + 0.003271812107414007, + 0.005241733510047197, + 0.005441481247544289, + 0.005446579307317734, + 0.006823481526225805, + 0.006394261494278908, + 0.008146383799612522, + 0.006698233541101217, + 0.005027198698371649, + 0.010746232233941555, + 0.007155923172831535, + 0.009731315076351166, + 0.012997306883335114, + 0.010662897489964962, + 0.010393555276095867, + 0.014907876960933208, + 0.00940354447811842, + 0.008952567353844643, + 0.0042183902114629745, + 0.005249901209026575, + 0.00628640316426754, + 0.0075692967511713505, + 0.00895133800804615, + 0.005741868633776903, + 0.005811803508549929, + 0.005569214001297951, + 0.02757868357002735, + 0.17614679038524628, + 0.023383567109704018, + 0.017388252541422844, + 0.03618893772363663, + 0.2570820748806, + 0.014494137838482857, + 0.04880906268954277, + 0.004397261422127485, + 0.0067363339476287365, + 0.004394064657390118, + 0.004678450524806976, + 0.010716700926423073, + 0.006909899413585663, + 0.004440336022526026, + 0.004308578558266163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.033728182315826416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019468179671093822 + ], + "opposite_logprob_tokens": [ + -6.251946926116943 + ], + "logprob_reward": -0.0019468179671093822, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019468179671093822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05389383062720299, + 0.04058530554175377, + 0.1915668249130249, + 0.04626132920384407, + 0.07721295207738876, + 0.04432215169072151, + 0.056600023061037064, + 0.06505865603685379, + 0.34265372157096863, + 0.08184509724378586 + ], + "credit_map": [ + 0.00936251413077116, + 0.006427494343370199, + 0.0042015002109110355, + 0.006046233233064413, + 0.007393795996904373, + 0.006152329966425896, + 0.008807300589978695, + 0.005502662155777216, + 0.004460800904780626, + 0.005749369505792856, + 0.0036466356832534075, + 0.003950940445065498, + 0.005920445546507835, + 0.005944452248513699, + 0.006722409278154373, + 0.004190250765532255, + 0.03251686692237854, + 0.03088829666376114, + 0.03546055033802986, + 0.01633082516491413, + 0.023742202669382095, + 0.01438214909285307, + 0.01646471954882145, + 0.02178122103214264, + 0.008236674591898918, + 0.005393013823777437, + 0.009230365045368671, + 0.004186025820672512, + 0.004152804613113403, + 0.004619757644832134, + 0.004461575299501419, + 0.005981112830340862, + 0.010827370919287205, + 0.006338594481348991, + 0.016852976754307747, + 0.008192218840122223, + 0.009191612713038921, + 0.011639878153800964, + 0.004167567007243633, + 0.010002733208239079, + 0.006162434350699186, + 0.0046010506339371204, + 0.007106852252036333, + 0.003732200711965561, + 0.005329831503331661, + 0.004691370762884617, + 0.003908890299499035, + 0.008789521642029285, + 0.00559020321816206, + 0.004766963887959719, + 0.00562182767316699, + 0.012124999426305294, + 0.008987460285425186, + 0.008126717060804367, + 0.0041077458299696445, + 0.007274104747921228, + 0.006707218009978533, + 0.006723916158080101, + 0.00811484083533287, + 0.006877815816551447, + 0.008938058279454708, + 0.009685114026069641, + 0.008412855677306652, + 0.009598840028047562, + 0.047174226492643356, + 0.09282762557268143, + 0.02732701785862446, + 0.0121440589427948, + 0.01870724931359291, + 0.07398711144924164, + 0.013802016153931618, + 0.05668444558978081, + 0.010084095411002636, + 0.011795639991760254, + 0.009685637429356575, + 0.009274178184568882, + 0.010259435512125492, + 0.012134457938373089, + 0.00972688477486372, + 0.008884771727025509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 9 + ], + "selected_reward": -0.0019468179671093822 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.223633885383606 + ], + "opposite_logprob_tokens": [ + -0.34863385558128357 + ], + "logprob_reward": -1.223633885383606, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.223633885383606, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.03969263657927513, + 0.02948032133281231, + 0.06896503269672394, + 0.059341125190258026, + 0.12293720245361328, + 0.032180652022361755, + 0.06539258360862732, + 0.07946687936782837, + 0.44512316584587097, + 0.05742044001817703 + ], + "credit_map": [ + 0.004372305702418089, + 0.005045042373239994, + 0.004011207260191441, + 0.004849712830036879, + 0.005255375523120165, + 0.003986329305917025, + 0.008607063442468643, + 0.003565601771697402, + 0.0033353331964462996, + 0.0054720742627978325, + 0.0033115188125520945, + 0.0029732449911534786, + 0.004217910580337048, + 0.0028026236686855555, + 0.004643235355615616, + 0.0027243769727647305, + 0.012174142524600029, + 0.008893168531358242, + 0.012292372062802315, + 0.006420256569981575, + 0.006756778806447983, + 0.009001292288303375, + 0.006354658864438534, + 0.0070723677054047585, + 0.007979242131114006, + 0.006364413537085056, + 0.010984036140143871, + 0.004921716172248125, + 0.009277518838644028, + 0.006065446883440018, + 0.004526895936578512, + 0.00922185555100441, + 0.013533588498830795, + 0.011420369148254395, + 0.02330050989985466, + 0.009819629602134228, + 0.010979436337947845, + 0.02311084233224392, + 0.008451343514025211, + 0.02232147380709648, + 0.005223258398473263, + 0.0031649579759687185, + 0.006027521565556526, + 0.0025666255969554186, + 0.004039136692881584, + 0.00401406129822135, + 0.002166650490835309, + 0.004978439770638943, + 0.006250204984098673, + 0.006299242377281189, + 0.010508265346288681, + 0.011691339313983917, + 0.0074234614148736, + 0.011989329010248184, + 0.0038816339801996946, + 0.0073491111397743225, + 0.006053293123841286, + 0.010896991938352585, + 0.014121126383543015, + 0.007136412896215916, + 0.010430137626826763, + 0.011371237225830555, + 0.009921829216182232, + 0.009535848163068295, + 0.05212101340293884, + 0.12734553217887878, + 0.03668207302689552, + 0.03621422126889229, + 0.02418447472155094, + 0.08911091834306717, + 0.02528955601155758, + 0.05417536199092865, + 0.007430669851601124, + 0.016374999657273293, + 0.003678724868223071, + 0.004356788005679846, + 0.008559721522033215, + 0.006628767587244511, + 0.004687836859375238, + 0.005702932830899954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -1.223633885383606 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543829157948494 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543829157948494, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543829157948494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08961033821105957, + 0.20706556737422943, + 0.08242371678352356, + 0.0726061537861824, + 0.07483875751495361, + 0.0834413543343544, + 0.08211962133646011, + 0.03172893822193146, + 0.21902579069137573, + 0.05713977664709091 + ], + "credit_map": [ + 0.015638582408428192, + 0.008852879516780376, + 0.00624470692127943, + 0.011191545985639095, + 0.01151962298899889, + 0.010118523612618446, + 0.016033364459872246, + 0.01001110952347517, + 0.024881713092327118, + 0.03656037151813507, + 0.018191196024417877, + 0.020284293219447136, + 0.029041126370429993, + 0.03252860903739929, + 0.025084873661398888, + 0.02049337700009346, + 0.013197993859648705, + 0.009369852021336555, + 0.017275307327508926, + 0.008497062139213085, + 0.006576298736035824, + 0.009626769460737705, + 0.008740548975765705, + 0.009139886125922203, + 0.012537197209894657, + 0.007474087178707123, + 0.015342793427407742, + 0.0062865703366696835, + 0.00623348681256175, + 0.008244357071816921, + 0.0067377532832324505, + 0.009749910794198513, + 0.010057740844786167, + 0.007271597161889076, + 0.014306662604212761, + 0.005457879975438118, + 0.00450777355581522, + 0.01481701247394085, + 0.003996883053332567, + 0.014423207379877567, + 0.00925347302109003, + 0.008300254121422768, + 0.011947757564485073, + 0.009505747817456722, + 0.014551034197211266, + 0.009416458196938038, + 0.010685109533369541, + 0.009781522676348686, + 0.006661857012659311, + 0.008212579414248466, + 0.00976741686463356, + 0.0177572313696146, + 0.013042572885751724, + 0.009600700810551643, + 0.006787679623812437, + 0.010289585217833519, + 0.0025378491263836622, + 0.005466998554766178, + 0.0039032476488500834, + 0.003058428643271327, + 0.0036972654052078724, + 0.006229959893971682, + 0.003050315659493208, + 0.0037848735228180885, + 0.018087714910507202, + 0.05371057614684105, + 0.014320162124931812, + 0.005224389955401421, + 0.013343860395252705, + 0.09418582916259766, + 0.00686868280172348, + 0.013284585438668728, + 0.0073840017430484295, + 0.009946390055119991, + 0.006687856279313564, + 0.0065271006897091866, + 0.006955343764275312, + 0.008278574794530869, + 0.006941345985978842, + 0.004419164266437292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.012543829157948494 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "wide shot, slowly and steadily pulling back to an extreme wide angle, captures the catastrophic failure of a massive multi-engine seaplane as it tumbles out of control against a turbulent ochre sky. The entire scene is rendered in the gritty, painterly style of Ashley Wood, with thick, visible brushstrokes and smudged charcoal outlines giving each element a tactile, visceral quality. The seaplane's hulking fuselage, its riveted metal skin scarred, weathered, and stained with soot, rolls violently on its side, one wing hanging precariously low, its engines sputtering and black smoke. The plane's chaotic descent is a desperate, labyrinthine dance against a roiling vortex of sickly yellow and dirty brown clouds. As the viewpoint shifts away, the vast expanse of the storm is revealed, the light an aimless, oppressive glow filtered through a toxic haze. The seaplane, once the dominant subject, gradually shrinks, its violent struggle reduced to a silent, insignificant detail, contrasting with the vast, indifferent, beautiful brutalist sky, emphasizing a profound sense of chaotic doom and isolation.", + "seed": 17999, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_3/index_6154_seed_17999_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2822431106586009, + "vlm_logprob": -0.2822431106586009, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.2822431106586009, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1914852261543274 + ], + "opposite_logprob_tokens": [ + -1.8164851665496826 + ], + "logprob_reward": -0.1914852261543274, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1914852261543274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042499613016843796, + 0.37693339586257935, + 0.18648235499858856, + 0.03715299069881439, + 0.024394644424319267, + 0.07806367427110672, + 0.06028108298778534, + 0.13972659409046173, + 0.028402453288435936, + 0.02606315352022648 + ], + "credit_map": [ + 0.0041741845197975636, + 0.007330013904720545, + 0.003623474622145295, + 0.005363660864531994, + 0.004611011128872633, + 0.005655348766595125, + 0.006808571517467499, + 0.004933347925543785, + 0.048274870961904526, + 0.03055454045534134, + 0.12265452742576599, + 0.013203615322709084, + 0.0157195795327425, + 0.02049103006720543, + 0.01814930886030197, + 0.1078859195113182, + 0.025155765935778618, + 0.025409631431102753, + 0.022226234897971153, + 0.022236865013837814, + 0.02256304770708084, + 0.0196682158857584, + 0.025622613728046417, + 0.023599980399012566, + 0.005108125973492861, + 0.0035100970417261124, + 0.0040078116580843925, + 0.0032793665304780006, + 0.0033072226215153933, + 0.005732994992285967, + 0.0034508679527789354, + 0.008756502531468868, + 0.002248789183795452, + 0.0020723543129861355, + 0.0027220866177231073, + 0.0016753043746575713, + 0.0016814478440210223, + 0.005309822969138622, + 0.0016762528102844954, + 0.007008587475866079, + 0.00783199816942215, + 0.007618701551109552, + 0.007491869851946831, + 0.013805270195007324, + 0.0096288388594985, + 0.013540529645979404, + 0.005974259227514267, + 0.012172208167612553, + 0.004172974731773138, + 0.007420845329761505, + 0.009296533651649952, + 0.0027176239527761936, + 0.0036582332104444504, + 0.014564786106348038, + 0.012898721732199192, + 0.005551361013203859, + 0.011061974801123142, + 0.01669299602508545, + 0.021738190203905106, + 0.01150833535939455, + 0.011056873016059399, + 0.03767460212111473, + 0.01800471358001232, + 0.011988910846412182, + 0.0016666502924636006, + 0.0024524126201868057, + 0.002134207170456648, + 0.00484404806047678, + 0.00404800521209836, + 0.006770650390535593, + 0.0036263142246752977, + 0.002860167296603322, + 0.002503112191334367, + 0.003651312319561839, + 0.003244558582082391, + 0.002152731642127037, + 0.002416973002254963, + 0.006713795010000467, + 0.00366321112960577, + 0.0017174604581668973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 7 + ], + "selected_reward": -0.1914852261543274 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006249576807022 + ], + "opposite_logprob_tokens": [ + -2.6950623989105225 + ], + "logprob_reward": -0.07006249576807022, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006249576807022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06081995368003845, + 0.18769685924053192, + 0.07430579513311386, + 0.08532361686229706, + 0.05662902072072029, + 0.07547187805175781, + 0.2278810441493988, + 0.17515498399734497, + 0.03885863348841667, + 0.017858164384961128 + ], + "credit_map": [ + 0.007271799258887768, + 0.009074809961020947, + 0.006435832008719444, + 0.005941474810242653, + 0.00731799378991127, + 0.008394097909331322, + 0.0069559477269649506, + 0.009427997283637524, + 0.01281385775655508, + 0.011582055129110813, + 0.04816371202468872, + 0.008032815530896187, + 0.022567331790924072, + 0.015339963138103485, + 0.019960440695285797, + 0.04923669248819351, + 0.010326987132430077, + 0.010017279535531998, + 0.011319025419652462, + 0.009225738234817982, + 0.009238878265023232, + 0.006730044726282358, + 0.007891403511166573, + 0.009556438773870468, + 0.011635400354862213, + 0.009976987726986408, + 0.010890871286392212, + 0.011340269818902016, + 0.00909733958542347, + 0.011918331496417522, + 0.008803493343293667, + 0.011660926043987274, + 0.006224434357136488, + 0.008092490956187248, + 0.006949556991457939, + 0.007482900284230709, + 0.0063194516114890575, + 0.008615742437541485, + 0.005990143865346909, + 0.006954297423362732, + 0.008343632332980633, + 0.008874575607478619, + 0.006776147522032261, + 0.015078933909535408, + 0.00991187896579504, + 0.010701953433454037, + 0.006704091560095549, + 0.009080665186047554, + 0.0171859972178936, + 0.027400752529501915, + 0.07348106801509857, + 0.01183771900832653, + 0.012006433680653572, + 0.021526459604501724, + 0.042934052646160126, + 0.021508561447262764, + 0.014166503213346004, + 0.022241555154323578, + 0.02427787147462368, + 0.010366392321884632, + 0.011367025785148144, + 0.05409512296319008, + 0.022388288751244545, + 0.016252225264906883, + 0.003387110773473978, + 0.004873127210885286, + 0.00418016966432333, + 0.007446305826306343, + 0.004805129021406174, + 0.005522591527551413, + 0.005579319782555103, + 0.0030648792162537575, + 0.0010961801744997501, + 0.0017778560286387801, + 0.0012796026421710849, + 0.002611564239487052, + 0.002684260020032525, + 0.003257978009060025, + 0.003422271693125367, + 0.0017284514615312219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.07006249576807022 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028221087995916605 + ], + "opposite_logprob_tokens": [ + -5.877821922302246 + ], + "logprob_reward": -0.0028221087995916605, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028221087995916605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18093211948871613, + 0.1307881772518158, + 0.09508643299341202, + 0.0653192475438118, + 0.05188294127583504, + 0.09777653217315674, + 0.12604668736457825, + 0.1630193591117859, + 0.023217491805553436, + 0.06593099981546402 + ], + "credit_map": [ + 0.018306219950318336, + 0.03740305453538895, + 0.019456375390291214, + 0.022680433467030525, + 0.02046382427215576, + 0.019322827458381653, + 0.024650633335113525, + 0.018648741766810417, + 0.013534246943891048, + 0.00926275085657835, + 0.02821122296154499, + 0.006744156591594219, + 0.01650087907910347, + 0.013024837709963322, + 0.01309148222208023, + 0.030418608337640762, + 0.009445000439882278, + 0.013808456249535084, + 0.012657488696277142, + 0.013337035663425922, + 0.01130201481282711, + 0.008675944991409779, + 0.011677305214107037, + 0.014183185063302517, + 0.008786999620497227, + 0.007013160735368729, + 0.006858916953206062, + 0.009523564949631691, + 0.007390145678073168, + 0.008477465249598026, + 0.007124708965420723, + 0.01014428585767746, + 0.005428011529147625, + 0.004515259061008692, + 0.006973477080464363, + 0.0029818478506058455, + 0.002858920255675912, + 0.013056744821369648, + 0.003160532796755433, + 0.012908151373267174, + 0.008879028260707855, + 0.005740674212574959, + 0.015211807563900948, + 0.011217773891985416, + 0.007928989827632904, + 0.030259855091571808, + 0.005535960663110018, + 0.013002442196011543, + 0.008913788013160229, + 0.012393457815051079, + 0.047837238758802414, + 0.005722451489418745, + 0.00571568263694644, + 0.021501462906599045, + 0.016628386452794075, + 0.007334218360483646, + 0.011262044310569763, + 0.017555417492985725, + 0.02408190444111824, + 0.008925754576921463, + 0.009916230104863644, + 0.05449701100587845, + 0.021581167355179787, + 0.015199815854430199, + 0.0020100681576877832, + 0.0018999113235622644, + 0.002506017917767167, + 0.0032649897038936615, + 0.002625785069540143, + 0.005493753589689732, + 0.003500448539853096, + 0.0019165169214829803, + 0.008468246087431908, + 0.013053139671683311, + 0.01040709763765335, + 0.006297740153968334, + 0.0061704483814537525, + 0.00859278254210949, + 0.008708312176167965, + 0.004233232699334621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0028221087995916605 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370619535446167 + ], + "opposite_logprob_tokens": [ + -0.3870619237422943 + ], + "logprob_reward": -1.1370619535446167, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370619535446167, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1048264354467392, + 0.23641185462474823, + 0.09519477933645248, + 0.06037992238998413, + 0.06371623277664185, + 0.12430011481046677, + 0.1709003895521164, + 0.07449749857187271, + 0.020227797329425812, + 0.049544986337423325 + ], + "credit_map": [ + 0.00897836871445179, + 0.012462131679058075, + 0.007684035692363977, + 0.01176025252789259, + 0.011666134931147099, + 0.012962833046913147, + 0.017132239416241646, + 0.02218043804168701, + 0.024618975818157196, + 0.022723808884620667, + 0.02571551874279976, + 0.019270583987236023, + 0.04990985989570618, + 0.029355622828006744, + 0.03740345314145088, + 0.027414055541157722, + 0.011639091186225414, + 0.012648573145270348, + 0.012308921664953232, + 0.00853827316313982, + 0.013347606174647808, + 0.010845217853784561, + 0.011091754771769047, + 0.014775347895920277, + 0.006396740209311247, + 0.005301682278513908, + 0.00569312646985054, + 0.007471856195479631, + 0.00909437332302332, + 0.007941274903714657, + 0.006323339883238077, + 0.012157532386481762, + 0.009106598794460297, + 0.007459485437721014, + 0.008347910828888416, + 0.0066044870764017105, + 0.005598052404820919, + 0.013988547958433628, + 0.004285342991352081, + 0.008325805887579918, + 0.009190781973302364, + 0.008794429711997509, + 0.020893702283501625, + 0.014576333574950695, + 0.010728678666055202, + 0.03593004122376442, + 0.007986635901033878, + 0.016199510544538498, + 0.019789576530456543, + 0.029180273413658142, + 0.02889896370470524, + 0.00930060539394617, + 0.01075374148786068, + 0.029038794338703156, + 0.029963495209813118, + 0.013974959962069988, + 0.006564647890627384, + 0.00989778432995081, + 0.01027006097137928, + 0.011951732449233532, + 0.011299524456262589, + 0.010698196478188038, + 0.007142399437725544, + 0.006673155818134546, + 0.0015123736811801791, + 0.0025336022954434156, + 0.0017095041694119573, + 0.002454081317409873, + 0.0028623638208955526, + 0.0042242929339408875, + 0.0022938072215765715, + 0.002637772588059306, + 0.00452159671112895, + 0.006728418171405792, + 0.005458048544824123, + 0.006362308748066425, + 0.004912587348371744, + 0.008964098058640957, + 0.00813788641244173, + 0.004460041876882315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 5 + ], + "selected_reward": -1.1370619535446167 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783769026398659 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783769026398659, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783769026398659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22525250911712646, + 0.1131923645734787, + 0.09673275798559189, + 0.07752881944179535, + 0.050370439887046814, + 0.11936107277870178, + 0.13701452314853668, + 0.07930183410644531, + 0.08045452833175659, + 0.02079116553068161 + ], + "credit_map": [ + 0.02162446267902851, + 0.04439593106508255, + 0.021140797063708305, + 0.029205992817878723, + 0.024940790608525276, + 0.02296931855380535, + 0.03398823365569115, + 0.026986990123987198, + 0.008781627751886845, + 0.008742071688175201, + 0.030072586610913277, + 0.008282394148409367, + 0.009086486883461475, + 0.01132558286190033, + 0.012018305249512196, + 0.02488330937922001, + 0.010979276150465012, + 0.008146720938384533, + 0.009151487611234188, + 0.012504352256655693, + 0.01441851258277893, + 0.011314010247588158, + 0.013819647021591663, + 0.016398757696151733, + 0.011202423833310604, + 0.007154382765293121, + 0.008083919063210487, + 0.008080162107944489, + 0.006763024255633354, + 0.01256841141730547, + 0.006420853082090616, + 0.017255643382668495, + 0.0036351431626826525, + 0.00413363566622138, + 0.004378201439976692, + 0.004197217058390379, + 0.003654272761195898, + 0.013138221576809883, + 0.003412938676774502, + 0.013820811174809933, + 0.013888147659599781, + 0.010795961134135723, + 0.013789133168756962, + 0.0165921188890934, + 0.011693569831550121, + 0.028411900624632835, + 0.010146915912628174, + 0.014043315313756466, + 0.010500850155949593, + 0.015996262431144714, + 0.05032128840684891, + 0.007967028766870499, + 0.006350954994559288, + 0.014965682290494442, + 0.02211952768266201, + 0.008792921900749207, + 0.004476800095289946, + 0.007002730388194323, + 0.01183479093015194, + 0.0056679085828363895, + 0.005708496086299419, + 0.027013318613171577, + 0.01079407799988985, + 0.0068037183955311775, + 0.009407782927155495, + 0.01420973613858223, + 0.012184469029307365, + 0.009407746605575085, + 0.00892822165042162, + 0.008407020010054111, + 0.009663593024015427, + 0.008245965465903282, + 0.002427993807941675, + 0.0033051632344722748, + 0.00287053850479424, + 0.0018269247375428677, + 0.0016589529113844037, + 0.004532674327492714, + 0.0026251666713505983, + 0.0015437524998560548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.009783769026398659 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "wide shot, slowly and steadily pulling back to an extreme wide angle, captures the catastrophic failure of a massive multi-engine seaplane as it tumbles out of control against a turbulent ochre sky. The entire scene is rendered in the gritty, painterly style of Ashley Wood, with thick, visible brushstrokes and smudged charcoal outlines giving each element a tactile, visceral quality. The seaplane's hulking fuselage, its riveted metal skin scarred, weathered, and stained with soot, rolls violently on its side, one wing hanging precariously low, its engines sputtering and black smoke. The plane's chaotic descent is a desperate, labyrinthine dance against a roiling vortex of sickly yellow and dirty brown clouds. As the viewpoint shifts away, the vast expanse of the storm is revealed, the light an aimless, oppressive glow filtered through a toxic haze. The seaplane, once the dominant subject, gradually shrinks, its violent struggle reduced to a silent, insignificant detail, contrasting with the vast, indifferent, beautiful brutalist sky, emphasizing a profound sense of chaotic doom and isolation.", + "seed": 18000, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_3/index_6154_seed_18000_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11076758473645895, + "vlm_logprob": -0.11076758473645895, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.11076758473645895, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03883133828639984 + ], + "opposite_logprob_tokens": [ + -3.6638312339782715 + ], + "logprob_reward": -0.03883133828639984, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03883133828639984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08222464472055435, + 0.07759516686201096, + 0.1075611487030983, + 0.13971835374832153, + 0.1045997217297554, + 0.04860552027821541, + 0.06795139610767365, + 0.027088509872555733, + 0.31329667568206787, + 0.03135885298252106 + ], + "credit_map": [ + 0.006393586751073599, + 0.010600212030112743, + 0.01207562442868948, + 0.011057251133024693, + 0.009936832822859287, + 0.012438404373824596, + 0.009447257034480572, + 0.010275480337440968, + 0.005588495172560215, + 0.007998394779860973, + 0.007978459820151329, + 0.016299720853567123, + 0.00710359076038003, + 0.014701279811561108, + 0.009378939867019653, + 0.00854629185050726, + 0.012820565141737461, + 0.018556328490376472, + 0.00890436302870512, + 0.01041010208427906, + 0.015415453352034092, + 0.018098581582307816, + 0.009396368637681007, + 0.013959395699203014, + 0.02884063310921192, + 0.02397751435637474, + 0.015242332592606544, + 0.01398259587585926, + 0.015803929418325424, + 0.014453161507844925, + 0.012952062301337719, + 0.01446611899882555, + 0.011646152473986149, + 0.012417791411280632, + 0.012502924539148808, + 0.013772700913250446, + 0.01421352569013834, + 0.015981400385499, + 0.012221820652484894, + 0.011843408457934856, + 0.005463607143610716, + 0.0033817149233072996, + 0.008741095662117004, + 0.003337997244670987, + 0.005830615293234587, + 0.012294037267565727, + 0.005409543868154287, + 0.004146910272538662, + 0.009478882886469364, + 0.0077491807751357555, + 0.01235805731266737, + 0.005052823573350906, + 0.006046900060027838, + 0.007900922559201717, + 0.004239572677761316, + 0.015125053003430367, + 0.0047346740029752254, + 0.0024499299470335245, + 0.003427100135013461, + 0.0033815104980021715, + 0.001804620260372758, + 0.004753576125949621, + 0.0019887308590114117, + 0.004548369906842709, + 0.017965739592909813, + 0.021557779982686043, + 0.10184647142887115, + 0.029685650020837784, + 0.018960407003760338, + 0.019898878410458565, + 0.016577312722802162, + 0.08680444210767746, + 0.0032547072041779757, + 0.0024497509002685547, + 0.0029628078918904066, + 0.004247425124049187, + 0.0019459688337519765, + 0.0037741195410490036, + 0.008417712524533272, + 0.00430635828524828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.03883133828639984 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560015320777893 + ], + "opposite_logprob_tokens": [ + -1.6006001234054565 + ], + "logprob_reward": -0.22560015320777893, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560015320777893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06632455438375473, + 0.08770432323217392, + 0.09578205645084381, + 0.041246622800827026, + 0.047547198832035065, + 0.029470279812812805, + 0.05040990561246872, + 0.028377853333950043, + 0.5307227373123169, + 0.022414637729525566 + ], + "credit_map": [ + 0.004427451640367508, + 0.011775633320212364, + 0.008533906191587448, + 0.008122618310153484, + 0.005993589758872986, + 0.011766292154788971, + 0.007026661187410355, + 0.008678394369781017, + 0.007593480404466391, + 0.00838081818073988, + 0.008863371796905994, + 0.020532887428998947, + 0.007900874130427837, + 0.013711790554225445, + 0.009463010355830193, + 0.011258076876401901, + 0.007237768266350031, + 0.014083540998399258, + 0.006060291081666946, + 0.009090936742722988, + 0.016113227233290672, + 0.01998080126941204, + 0.01087184902280569, + 0.012343628332018852, + 0.00495894206687808, + 0.004883649758994579, + 0.006336174439638853, + 0.005168836563825607, + 0.004176127724349499, + 0.005528613924980164, + 0.005784073378890753, + 0.004410200752317905, + 0.007462457288056612, + 0.005021997261792421, + 0.005773379001766443, + 0.004163768142461777, + 0.006041893269866705, + 0.005099490750581026, + 0.007932564243674278, + 0.006051640957593918, + 0.0028147168923169374, + 0.0020083540584892035, + 0.005390476901084185, + 0.0017794981831684709, + 0.0046914005652070045, + 0.006831842474639416, + 0.00260098441503942, + 0.003353002481162548, + 0.007299394346773624, + 0.006312042009085417, + 0.008495605550706387, + 0.00546414265409112, + 0.004059390164911747, + 0.005359611939638853, + 0.0032994544599205256, + 0.01012025773525238, + 0.005250917281955481, + 0.0030120457522571087, + 0.003244672203436494, + 0.0028819837607443333, + 0.0022322654258459806, + 0.0044301231391727924, + 0.003058468224480748, + 0.004267373587936163, + 0.032629601657390594, + 0.03452003374695778, + 0.17635416984558105, + 0.04707009717822075, + 0.03431924059987068, + 0.036244265735149384, + 0.027343248948454857, + 0.14224198460578918, + 0.0018201947677880526, + 0.0029067283030599356, + 0.002421974204480648, + 0.003208982991054654, + 0.0029705921187996864, + 0.002125391038134694, + 0.0056360093876719475, + 0.0013247635215520859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.22560015320777893 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00134151556994766 + ], + "opposite_logprob_tokens": [ + -6.626341342926025 + ], + "logprob_reward": -0.00134151556994766, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.00134151556994766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2985844314098358, + 0.1156083419919014, + 0.06113646179437637, + 0.09455163031816483, + 0.060654740780591965, + 0.05510370805859566, + 0.07613980025053024, + 0.1121862456202507, + 0.07889517396688461, + 0.04713943973183632 + ], + "credit_map": [ + 0.017275692895054817, + 0.05521070584654808, + 0.03707348182797432, + 0.03194168955087662, + 0.03402339667081833, + 0.06413231790065765, + 0.021627137437462807, + 0.03730001300573349, + 0.007390593644231558, + 0.014810873195528984, + 0.008871737867593765, + 0.01244788896292448, + 0.013975357636809349, + 0.03230707719922066, + 0.01175645086914301, + 0.014048362150788307, + 0.006288556382060051, + 0.008000091649591923, + 0.005486675072461367, + 0.007582089398056269, + 0.009514478966593742, + 0.009159037843346596, + 0.0072740488685667515, + 0.007831484079360962, + 0.011059020645916462, + 0.01808236353099346, + 0.00933887530118227, + 0.009840183891355991, + 0.011833708733320236, + 0.011134750209748745, + 0.012907618656754494, + 0.01035511214286089, + 0.009915279224514961, + 0.004516758024692535, + 0.010746580548584461, + 0.005059320945292711, + 0.00609131483361125, + 0.006970728747546673, + 0.007718329783529043, + 0.009636434726417065, + 0.006717741023749113, + 0.004105340223759413, + 0.011543083004653454, + 0.004022682551294565, + 0.005539935082197189, + 0.00963190570473671, + 0.00728296535089612, + 0.006260059308260679, + 0.011358075775206089, + 0.007335586939007044, + 0.010950899682939053, + 0.008747673593461514, + 0.006212161388248205, + 0.012692474760115147, + 0.005691616330295801, + 0.013151314109563828, + 0.01693027652800083, + 0.010480515658855438, + 0.011906391941010952, + 0.01573673076927662, + 0.010356221348047256, + 0.019543224945664406, + 0.009506620466709137, + 0.017726276069879532, + 0.006352938245981932, + 0.008025845512747765, + 0.01876203343272209, + 0.008618725463747978, + 0.0063019925728440285, + 0.006760232616215944, + 0.004695831798017025, + 0.019377579912543297, + 0.006421937607228756, + 0.006012680474668741, + 0.006058405619114637, + 0.006968799512833357, + 0.0031406483612954617, + 0.007223541382700205, + 0.006550163496285677, + 0.004763265140354633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.00134151556994766 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28132718801498413 + ], + "opposite_logprob_tokens": [ + -1.406327247619629 + ], + "logprob_reward": -0.28132718801498413, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28132718801498413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14046506583690643, + 0.09866698831319809, + 0.10322520136833191, + 0.09748349338769913, + 0.07063695788383484, + 0.16557352244853973, + 0.20080071687698364, + 0.036292463541030884, + 0.06566115468740463, + 0.021194512024521828 + ], + "credit_map": [ + 0.009173497557640076, + 0.015404382720589638, + 0.018874799832701683, + 0.03405730798840523, + 0.015595820732414722, + 0.01515519805252552, + 0.013356250710785389, + 0.018847811967134476, + 0.006844469346106052, + 0.017160695046186447, + 0.009868171066045761, + 0.01843240112066269, + 0.010126772336661816, + 0.011462927795946598, + 0.010289838537573814, + 0.01448170468211174, + 0.008456827141344547, + 0.010840210132300854, + 0.007564932107925415, + 0.011557997204363346, + 0.017383964732289314, + 0.019406532868742943, + 0.01583734340965748, + 0.012177404016256332, + 0.018460704013705254, + 0.018510164692997932, + 0.00944459717720747, + 0.00959662813693285, + 0.010038478299975395, + 0.010235236026346684, + 0.010303769260644913, + 0.010893920436501503, + 0.011889812536537647, + 0.005833881441503763, + 0.015203203074634075, + 0.004830501042306423, + 0.009158819913864136, + 0.007563699968159199, + 0.006345963571220636, + 0.009811070747673512, + 0.014938047155737877, + 0.013960701413452625, + 0.027992989867925644, + 0.01327204518020153, + 0.025467531755566597, + 0.02455323003232479, + 0.025363421067595482, + 0.020025549456477165, + 0.018823126330971718, + 0.027352776378393173, + 0.04030150920152664, + 0.02169930748641491, + 0.021327562630176544, + 0.019471891224384308, + 0.01590181142091751, + 0.03592272475361824, + 0.004537689499557018, + 0.004048444330692291, + 0.0036642844788730145, + 0.005422607995569706, + 0.0033850655891001225, + 0.006903748493641615, + 0.002680470934137702, + 0.005650147795677185, + 0.0046193297021090984, + 0.005370058584958315, + 0.01877220906317234, + 0.006626354996114969, + 0.006439657416194677, + 0.006198108196258545, + 0.003995705861598253, + 0.013639728538691998, + 0.0018334862543269992, + 0.0023413558956235647, + 0.0013972799060866237, + 0.0024008103646337986, + 0.0019795196130871773, + 0.0021443674340844154, + 0.0057942820712924, + 0.003303410718217492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.28132718801498413 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737728603184223 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.006737728603184223, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737728603184223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1927158534526825, + 0.11500008404254913, + 0.06063487008213997, + 0.09570340067148209, + 0.08264382928609848, + 0.05647962540388107, + 0.08215609192848206, + 0.10292252153158188, + 0.18097783625125885, + 0.03076588548719883 + ], + "credit_map": [ + 0.009958332404494286, + 0.040678106248378754, + 0.02124236337840557, + 0.028695903718471527, + 0.018943557515740395, + 0.0413019023835659, + 0.013585619628429413, + 0.01831008307635784, + 0.009813003242015839, + 0.011688759550452232, + 0.009461910463869572, + 0.01973401941359043, + 0.014800215139985085, + 0.025692235678434372, + 0.012355122715234756, + 0.011454821564257145, + 0.006790038663893938, + 0.00743495486676693, + 0.006719327066093683, + 0.007604371756315231, + 0.009573950432240963, + 0.010772583074867725, + 0.006346384063363075, + 0.005393270403146744, + 0.017172569409012794, + 0.023633677512407303, + 0.005636909045279026, + 0.009658380411565304, + 0.011731818318367004, + 0.008363984525203705, + 0.00893441028892994, + 0.010571654886007309, + 0.01127525046467781, + 0.006264253985136747, + 0.014284064993262291, + 0.008207092992961407, + 0.009264613501727581, + 0.009327624924480915, + 0.011909796856343746, + 0.012111134827136993, + 0.0051988898776471615, + 0.003938005771487951, + 0.012814884074032307, + 0.0038579371757805347, + 0.005634387955069542, + 0.010562069714069366, + 0.007702800445258617, + 0.006770651787519455, + 0.010060487315058708, + 0.009233714081346989, + 0.013824944384396076, + 0.00996349286288023, + 0.009389451704919338, + 0.012552181258797646, + 0.005786860827356577, + 0.011344967409968376, + 0.01612899638712406, + 0.010279346257448196, + 0.011151556856930256, + 0.01949063502252102, + 0.006856165826320648, + 0.016742687672376633, + 0.009722154587507248, + 0.01255098544061184, + 0.009570007212460041, + 0.009189320728182793, + 0.06662849336862564, + 0.014099369756877422, + 0.009035872295498848, + 0.013342575170099735, + 0.0090519730001688, + 0.05006024241447449, + 0.0033124745823442936, + 0.0038976753130555153, + 0.003257878590375185, + 0.0037958871107548475, + 0.0031902771443128586, + 0.005966451484709978, + 0.0040090675465762615, + 0.0033361767418682575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.006737728603184223 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide shot, slowly and steadily pulling back to an extreme wide angle, captures the catastrophic failure of a massive multi-engine seaplane as it tumbles out of control against a turbulent ochre sky. The entire scene is rendered in the gritty, painterly style of Ashley Wood, with thick, visible brushstrokes and smudged charcoal outlines giving each element a tactile, visceral quality. The seaplane's hulking fuselage, its riveted metal skin scarred, weathered, and stained with soot, rolls violently on its side, one wing hanging precariously low, its engines sputtering and black smoke. The plane's chaotic descent is a desperate, labyrinthine dance against a roiling vortex of sickly yellow and dirty brown clouds. As the viewpoint shifts away, the vast expanse of the storm is revealed, the light an aimless, oppressive glow filtered through a toxic haze. The seaplane, once the dominant subject, gradually shrinks, its violent struggle reduced to a silent, insignificant detail, contrasting with the vast, indifferent, beautiful brutalist sky, emphasizing a profound sense of chaotic doom and isolation.", + "seed": 18001, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_3/index_6154_seed_18001_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3099502284661867, + "vlm_logprob": -0.3099502284661867, + "vlm_yes_no_margin": 3.250000023841858, + "vlm_token_credit": -0.3099502284661867, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15330065786838531 + ], + "opposite_logprob_tokens": [ + -2.0283007621765137 + ], + "logprob_reward": -0.15330065786838531, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.15330065786838531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14670608937740326, + 0.08258230984210968, + 0.11737870424985886, + 0.1556737869977951, + 0.04986843839287758, + 0.04926745593547821, + 0.037411198019981384, + 0.12811928987503052, + 0.17005860805511475, + 0.06293407827615738 + ], + "credit_map": [ + 0.009849915280938148, + 0.01842782087624073, + 0.016166578978300095, + 0.019850948825478554, + 0.03928646072745323, + 0.020031029358506203, + 0.012388505972921848, + 0.010704846121370792, + 0.012469632551074028, + 0.00954435858875513, + 0.009237845428287983, + 0.007279327604919672, + 0.010804221034049988, + 0.009857366792857647, + 0.013231102377176285, + 0.010158460587263107, + 0.00651887571439147, + 0.008953835815191269, + 0.005649344064295292, + 0.006360316649079323, + 0.007064701057970524, + 0.03728727623820305, + 0.007331070024520159, + 0.03821329399943352, + 0.01871434412896633, + 0.02108384110033512, + 0.019239142537117004, + 0.017614036798477173, + 0.0148239741101861, + 0.025750204920768738, + 0.01689656265079975, + 0.02155168727040291, + 0.006501895375549793, + 0.005998404696583748, + 0.006226317025721073, + 0.007284261751919985, + 0.005897577852010727, + 0.006688656285405159, + 0.005532305687665939, + 0.005739023443311453, + 0.005608854815363884, + 0.004405294079333544, + 0.009392962791025639, + 0.004450595006346703, + 0.005866785068064928, + 0.010078957304358482, + 0.0037538083270192146, + 0.005710201803594828, + 0.004714113660156727, + 0.0034067260567098856, + 0.00447052950039506, + 0.004741994198411703, + 0.004430191125720739, + 0.005128927528858185, + 0.005603752098977566, + 0.004914966877549887, + 0.005109964404255152, + 0.005268820561468601, + 0.04338879883289337, + 0.011300324462354183, + 0.006724087987095118, + 0.04287762567400932, + 0.007184197660535574, + 0.006265487987548113, + 0.005405902862548828, + 0.007771060336381197, + 0.008371951058506966, + 0.007429580204188824, + 0.05726098641753197, + 0.06071830913424492, + 0.007620423100888729, + 0.01548040471971035, + 0.005894887261092663, + 0.006983565166592598, + 0.010529784485697746, + 0.00798970740288496, + 0.010097768157720566, + 0.007749415468424559, + 0.006882058456540108, + 0.006806892808526754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.15330065786838531 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901136577129364 + ], + "opposite_logprob_tokens": [ + -2.5790114402770996 + ], + "logprob_reward": -0.07901136577129364, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901136577129364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08034265041351318, + 0.053644243627786636, + 0.045502979308366776, + 0.10964982211589813, + 0.04133390635251999, + 0.02783198654651642, + 0.022710878401994705, + 0.19949400424957275, + 0.381662517786026, + 0.0378270149230957 + ], + "credit_map": [ + 0.007057760842144489, + 0.009330127388238907, + 0.00853497814387083, + 0.008519100956618786, + 0.02124963141977787, + 0.009933333843946457, + 0.007771893870085478, + 0.007945836521685123, + 0.008097884245216846, + 0.00671380152925849, + 0.006246825680136681, + 0.00474590715020895, + 0.006750393658876419, + 0.006742116995155811, + 0.008429653942584991, + 0.005917664617300034, + 0.005760149098932743, + 0.006900938693434, + 0.00460796570405364, + 0.00458461232483387, + 0.006461724638938904, + 0.0061331260949373245, + 0.004636907484382391, + 0.006417561788111925, + 0.010503437370061874, + 0.016976138576865196, + 0.01576533541083336, + 0.01490518357604742, + 0.01158932875841856, + 0.014152894727885723, + 0.009937531314790249, + 0.015819977968931198, + 0.004227125085890293, + 0.005016738548874855, + 0.00482795387506485, + 0.0054120407439768314, + 0.0055612316355109215, + 0.006246416829526424, + 0.005264955572783947, + 0.004777449183166027, + 0.002285493304952979, + 0.002001741901040077, + 0.005329258739948273, + 0.0031435592100024223, + 0.0027576328720897436, + 0.007609835360199213, + 0.002060872735455632, + 0.0026435954496264458, + 0.0024406015872955322, + 0.002384054008871317, + 0.0030843389686197042, + 0.002308313036337495, + 0.0027101209852844477, + 0.003918495029211044, + 0.0030983593314886093, + 0.0027665968518704176, + 0.00709221325814724, + 0.005132649559527636, + 0.07408646494150162, + 0.010368152521550655, + 0.008349670097231865, + 0.08091788738965988, + 0.006349173840135336, + 0.007197825703769922, + 0.012862885370850563, + 0.02192457765340805, + 0.022337859496474266, + 0.007846416905522346, + 0.15870293974876404, + 0.125284343957901, + 0.013489029370248318, + 0.019214486703276634, + 0.003250792156904936, + 0.0038644017186015844, + 0.005512033589184284, + 0.007954408414661884, + 0.005223127547651529, + 0.0038839145563542843, + 0.005196775309741497, + 0.0029415623284876347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -0.07901136577129364 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008169650682248175 + ], + "opposite_logprob_tokens": [ + -7.125816822052002 + ], + "logprob_reward": -0.0008169650682248175, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008169650682248175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21010459959506989, + 0.21981793642044067, + 0.08876476436853409, + 0.09592326730489731, + 0.05340925231575966, + 0.027485152706503868, + 0.03674999251961708, + 0.09214048832654953, + 0.05354028940200806, + 0.12206429243087769 + ], + "credit_map": [ + 0.027272623032331467, + 0.02950691618025303, + 0.02789534442126751, + 0.02704799734055996, + 0.02659567818045616, + 0.025692394003272057, + 0.026864632964134216, + 0.019228992983698845, + 0.034992631524801254, + 0.02611791156232357, + 0.0236305333673954, + 0.019610147923231125, + 0.02420460619032383, + 0.028713183477520943, + 0.037310242652893066, + 0.025238672271370888, + 0.01218736544251442, + 0.006599761545658112, + 0.007579104509204626, + 0.006759874522686005, + 0.007763403933495283, + 0.01790371909737587, + 0.009609243832528591, + 0.020362289622426033, + 0.01848793216049671, + 0.009980153292417526, + 0.01279299333691597, + 0.008352801203727722, + 0.009344374760985374, + 0.013162988238036633, + 0.00824520830065012, + 0.01555680762976408, + 0.005038257222622633, + 0.004415292292833328, + 0.008087227120995522, + 0.004853288177400827, + 0.00506358128041029, + 0.007785265799611807, + 0.008510406129062176, + 0.00965593010187149, + 0.0038836731109768152, + 0.0025199956726282835, + 0.004093957133591175, + 0.002640488790348172, + 0.0029063341207802296, + 0.005609435960650444, + 0.0023296738509088755, + 0.00350159197114408, + 0.003046754514798522, + 0.0046648746356368065, + 0.005215928424149752, + 0.00463499128818512, + 0.004819952882826328, + 0.004558992572128773, + 0.005604161880910397, + 0.004204333759844303, + 0.006189141888171434, + 0.004622929263859987, + 0.029062239453196526, + 0.004721452947705984, + 0.00557058397680521, + 0.0293654203414917, + 0.0052649276331067085, + 0.007343784905970097, + 0.006101174745708704, + 0.003992062993347645, + 0.0065456279553473, + 0.005411034449934959, + 0.011628225445747375, + 0.008600576780736446, + 0.006669668946415186, + 0.004591915290802717, + 0.009972617030143738, + 0.01355138048529625, + 0.01967606693506241, + 0.0169661957770586, + 0.019793376326560974, + 0.015389353968203068, + 0.014361534267663956, + 0.01235375925898552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0008169650682248175 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134268522262573 + ], + "opposite_logprob_tokens": [ + -0.3134268522262573 + ], + "logprob_reward": -1.3134268522262573, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134268522262573, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.052846599370241165, + 0.04878968745470047, + 0.06089026853442192, + 0.0637449100613594, + 0.04171036183834076, + 0.0413345992565155, + 0.020377784967422485, + 0.1315351277589798, + 0.48909464478492737, + 0.04967600479722023 + ], + "credit_map": [ + 0.004851937759667635, + 0.0069189295172691345, + 0.008190854452550411, + 0.006319284904748201, + 0.006801462732255459, + 0.006085734814405441, + 0.006148308515548706, + 0.007530085742473602, + 0.004080924671143293, + 0.006614050827920437, + 0.004648053087294102, + 0.00508568249642849, + 0.007050519809126854, + 0.007730427198112011, + 0.0057751829735934734, + 0.007804845925420523, + 0.00683620385825634, + 0.004694069270044565, + 0.005173555109649897, + 0.005602832417935133, + 0.00645276578143239, + 0.0135597949847579, + 0.004180647898465395, + 0.014390398748219013, + 0.008909299038350582, + 0.00786431971937418, + 0.007140690460801125, + 0.007174012716859579, + 0.006793587002903223, + 0.011399967595934868, + 0.00477017043158412, + 0.009692857973277569, + 0.004760253243148327, + 0.004700783174484968, + 0.005133027210831642, + 0.004390382207930088, + 0.0038875218015164137, + 0.006305105052888393, + 0.0052681295201182365, + 0.007265162188559771, + 0.004251070786267519, + 0.002769390819594264, + 0.008768842555582523, + 0.002383057028055191, + 0.0038293569814413786, + 0.012631677091121674, + 0.002499941736459732, + 0.004201261792331934, + 0.002314568031579256, + 0.0021170666441321373, + 0.0026963294949382544, + 0.0030659197364002466, + 0.002428022213280201, + 0.002363108331337571, + 0.0033120340667665005, + 0.002080737380310893, + 0.007061749696731567, + 0.006881247740238905, + 0.037369441241025925, + 0.012753801420331001, + 0.008769125677645206, + 0.040343791246414185, + 0.007525371387600899, + 0.010830595158040524, + 0.024538198485970497, + 0.034244757145643234, + 0.039983026683330536, + 0.01540705282241106, + 0.21999435126781464, + 0.11274833977222443, + 0.03013937920331955, + 0.012039581313729286, + 0.003661892144009471, + 0.004642265383154154, + 0.00609931955114007, + 0.012131906114518642, + 0.007125395815819502, + 0.005244813859462738, + 0.006731170229613781, + 0.004039243329316378 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 3 + ], + "selected_reward": -1.3134268522262573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive multi-engine seaplane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the seaplane shown tumbling out of control with its fuselage rolling violently on its side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a turbulent sky with ochre, sickly yellow, and dirty brown colors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a wide shot and slowly pull back to an extreme wide angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style gritty and painterly with thick visible brushstrokes and smudged charcoal outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031953013967722654 + ], + "opposite_logprob_tokens": [ + -5.753195285797119 + ], + "logprob_reward": -0.0031953013967722654, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031953013967722654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22854912281036377, + 0.22694651782512665, + 0.03792133554816246, + 0.09203426539897919, + 0.03043295256793499, + 0.016886554658412933, + 0.026248805224895477, + 0.14903788268566132, + 0.15052475035190582, + 0.04141787812113762 + ], + "credit_map": [ + 0.029737189412117004, + 0.033616580069065094, + 0.02308826707303524, + 0.034658417105674744, + 0.04514963924884796, + 0.021781351417303085, + 0.02356885001063347, + 0.01694883406162262, + 0.03068395145237446, + 0.027933720499277115, + 0.02399473264813423, + 0.02184796892106533, + 0.023187527433037758, + 0.032997265458106995, + 0.04184507206082344, + 0.024456268176436424, + 0.005027354694902897, + 0.00309938145801425, + 0.0038634813390672207, + 0.0033137546852231026, + 0.005205431487411261, + 0.005755975842475891, + 0.004524060059338808, + 0.007131895516067743, + 0.01034588273614645, + 0.015157779678702354, + 0.009587129577994347, + 0.010432533919811249, + 0.011586167849600315, + 0.01444509718567133, + 0.008317654952406883, + 0.01216201577335596, + 0.0020759485196322203, + 0.002324614906683564, + 0.004367380402982235, + 0.0037810462526977062, + 0.0025455059949308634, + 0.005057777278125286, + 0.003939970396459103, + 0.0063407085835933685, + 0.0019141571829095483, + 0.0013397386064752936, + 0.002790102269500494, + 0.0014321046182885766, + 0.0017574072116985917, + 0.0035324585624039173, + 0.0015179658075794578, + 0.0026026188861578703, + 0.0023033032193779945, + 0.0023288081865757704, + 0.0038464742247015238, + 0.002629086608067155, + 0.0034511073026806116, + 0.004282204434275627, + 0.0044531021267175674, + 0.002954719588160515, + 0.0054902201518416405, + 0.004554344806820154, + 0.05073176696896553, + 0.006809698883444071, + 0.009657175280153751, + 0.05840441584587097, + 0.005915243178606033, + 0.007475006394088268, + 0.010699247010052204, + 0.011878837831318378, + 0.014673552475869656, + 0.009050650522112846, + 0.06303046643733978, + 0.020457671955227852, + 0.013987923040986061, + 0.006746397353708744, + 0.004227573983371258, + 0.0038643984589725733, + 0.005777305457741022, + 0.008843067102134228, + 0.006519000977277756, + 0.00417210441082716, + 0.004552668426185846, + 0.0034617597702890635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0031953013967722654 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_4/index_tensor([8882, 8882])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_4/index_tensor([8882, 8882])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4cca50182a2c6b9bd5bd56cbc9d356355397ea14 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_4/index_tensor([8882, 8882])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, mid-shot shot, taken from a horizontal angle, captures a tranquil moment in documentary style. In a Tang Dynasty courtyard bathed in a cool, silvery glow, the swaying shadows of trees are reflected on the stone pavement. The poet Li Bai wears a wide-sleeved white silk robe, which flutters gently in the night breeze, revealing its delicate texture. His long, dark hair is held together only by a simple hairpin. His face is thin, and his eyes hold a hint of drunkenness and contemplation. He turns slowly and calmly, a movement that gently sways a small, white porcelain wine glass in his hand. Ripples ripple on the surface of the clear wine, momentarily catching and reflecting the bright moonlight. The entire scene is bathed in moonlight, creating a striking contrast between light and dark. The light outlines his graceful figure and the deep shadows of the courtyard scenery, creating a timeless, serene, and poetic classical atmosphere.", + "seed": 4781, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_4/index_8882_seed_4781_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10609041629359126, + "vlm_logprob": -0.10609041629359126, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.10609041629359126, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.39428409934043884 + ], + "opposite_logprob_tokens": [ + -1.1442841291427612 + ], + "logprob_reward": -0.39428409934043884, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.39428409934043884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15931880474090576, + 0.05864948406815529, + 0.035439200699329376, + 0.07407023012638092, + 0.13878503441810608, + 0.1296859234571457, + 0.2572210133075714, + 0.01763262413442135, + 0.06149697303771973, + 0.06770069152116776 + ], + "credit_map": [ + 0.021697625517845154, + 0.025291042402386665, + 0.01579919271171093, + 0.022739771753549576, + 0.024156834930181503, + 0.02407754212617874, + 0.011157317087054253, + 0.014399465173482895, + 0.005878128111362457, + 0.0076913475058972836, + 0.012679065577685833, + 0.0075292885303497314, + 0.008584782481193542, + 0.004922864958643913, + 0.005793248303234577, + 0.0055707585997879505, + 0.0032604297157377005, + 0.0037202804815024137, + 0.004553156904876232, + 0.006636844482272863, + 0.0036859577521681786, + 0.004296977538615465, + 0.004391991533339024, + 0.00489356042817235, + 0.006212966050952673, + 0.007825220935046673, + 0.007160430308431387, + 0.018065541982650757, + 0.007866398431360722, + 0.00689576705917716, + 0.007833574898540974, + 0.012210321612656116, + 0.01572336256504059, + 0.01142363715916872, + 0.02232319675385952, + 0.029471708461642265, + 0.009904826059937477, + 0.012597613036632538, + 0.010504667647182941, + 0.026836011558771133, + 0.031221546232700348, + 0.012563592754304409, + 0.024768011644482613, + 0.015421788208186626, + 0.007670528255403042, + 0.01931646279990673, + 0.008833998814225197, + 0.009889982640743256, + 0.05923797935247421, + 0.03411052003502846, + 0.0350654274225235, + 0.025034701451659203, + 0.02292552962899208, + 0.03484262526035309, + 0.024938782677054405, + 0.021065426990389824, + 0.002481434028595686, + 0.0022290251217782497, + 0.0018154100980609655, + 0.0019618459045886993, + 0.0018418332329019904, + 0.0027966597117483616, + 0.002307395450770855, + 0.002199017908424139, + 0.009698363952338696, + 0.0046755592338740826, + 0.004602588713169098, + 0.0056949518620967865, + 0.00433701416477561, + 0.005884097423404455, + 0.014363140799105167, + 0.012241248972713947, + 0.013463249430060387, + 0.006695079617202282, + 0.006973718758672476, + 0.007025492377579212, + 0.008404869586229324, + 0.009194783866405487, + 0.00736224465072155, + 0.008581247180700302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.39428409934043884 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054951589554548264 + ], + "opposite_logprob_tokens": [ + -2.9299516677856445 + ], + "logprob_reward": -0.054951589554548264, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054951589554548264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26689690351486206, + 0.083729088306427, + 0.05819661170244217, + 0.1929960548877716, + 0.12695926427841187, + 0.10388129204511642, + 0.049428779631853104, + 0.03013867326080799, + 0.05562806874513626, + 0.03214533254504204 + ], + "credit_map": [ + 0.02283538319170475, + 0.05311058089137077, + 0.020262515172362328, + 0.03841705247759819, + 0.06101977452635765, + 0.01834859699010849, + 0.026786422356963158, + 0.026116563007235527, + 0.007572430185973644, + 0.010329815559089184, + 0.0071486495435237885, + 0.01863822713494301, + 0.014727446250617504, + 0.006388536188751459, + 0.010897933505475521, + 0.008026055991649628, + 0.003520991187542677, + 0.006363934837281704, + 0.005718245171010494, + 0.012752421200275421, + 0.007399622350931168, + 0.004344620276242495, + 0.009020783938467503, + 0.009075992740690708, + 0.028915446251630783, + 0.018017418682575226, + 0.01520920917391777, + 0.03441354259848595, + 0.02712227776646614, + 0.018469486385583878, + 0.021804913878440857, + 0.02904375083744526, + 0.009230266325175762, + 0.007998785935342312, + 0.0063499207608401775, + 0.03849225863814354, + 0.016833944246172905, + 0.006986583583056927, + 0.011819343082606792, + 0.029248159378767014, + 0.018219517543911934, + 0.007004859857261181, + 0.007384956814348698, + 0.012552091851830482, + 0.01437555905431509, + 0.010471692308783531, + 0.015057874843478203, + 0.01881473697721958, + 0.0124772097915411, + 0.0066579245030879974, + 0.006203501019626856, + 0.003908196929842234, + 0.005938967224210501, + 0.0049352687783539295, + 0.0048405504785478115, + 0.004467161372303963, + 0.0025535975582897663, + 0.0028719366528093815, + 0.0020881935488432646, + 0.004352818243205547, + 0.005180115345865488, + 0.0036178273148834705, + 0.00445923674851656, + 0.005014945287257433, + 0.0066663650795817375, + 0.006332437042146921, + 0.005978681147098541, + 0.006536076311022043, + 0.006291115190833807, + 0.005761459935456514, + 0.008754067122936249, + 0.009307867847383022, + 0.006077907048165798, + 0.002039940096437931, + 0.0033629999961704016, + 0.0031661393586546183, + 0.006130724214017391, + 0.00379609945230186, + 0.0037003345787525177, + 0.0038711882662028074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.054951589554548264 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253083072602749 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253083072602749, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253083072602749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15228834748268127, + 0.04906245321035385, + 0.045220665633678436, + 0.3951069712638855, + 0.03585095703601837, + 0.06996665894985199, + 0.11800867319107056, + 0.02557026594877243, + 0.05780006945133209, + 0.051124949008226395 + ], + "credit_map": [ + 0.01751551404595375, + 0.02172231115400791, + 0.012994658201932907, + 0.024929402396082878, + 0.02065407671034336, + 0.02214910089969635, + 0.016827603802084923, + 0.015495692379772663, + 0.00833938643336296, + 0.005875877570360899, + 0.005451208911836147, + 0.007155061233788729, + 0.005916777998209, + 0.005933430977165699, + 0.004829721059650183, + 0.005560990888625383, + 0.004314560443162918, + 0.004393529146909714, + 0.003782591549679637, + 0.00592290423810482, + 0.007849420420825481, + 0.006292484700679779, + 0.007854876108467579, + 0.004810301586985588, + 0.05216168239712715, + 0.04193229600787163, + 0.031664732843637466, + 0.10002689808607101, + 0.03725532814860344, + 0.032773662358522415, + 0.044855840504169464, + 0.05443656072020531, + 0.00246410071849823, + 0.004356148652732372, + 0.003093153005465865, + 0.007398558314889669, + 0.0064222910441458225, + 0.002108550863340497, + 0.004805401433259249, + 0.005202753469347954, + 0.010319989174604416, + 0.005738162901252508, + 0.004955984652042389, + 0.0070477938279509544, + 0.010397912003099918, + 0.006787354592233896, + 0.011828583665192127, + 0.012890882790088654, + 0.023066455498337746, + 0.016617609187960625, + 0.015303169377148151, + 0.011719143018126488, + 0.008958959951996803, + 0.014630799181759357, + 0.014558008871972561, + 0.013154527172446251, + 0.0025186368729919195, + 0.0024123084731400013, + 0.002443082630634308, + 0.0036383760161697865, + 0.00425414415076375, + 0.00280113541521132, + 0.004165430553257465, + 0.00333715183660388, + 0.009042218327522278, + 0.005426280200481415, + 0.00394526356831193, + 0.004669134970754385, + 0.005258458666503429, + 0.005263406317681074, + 0.012633627280592918, + 0.011561678722500801, + 0.010360968299210072, + 0.005464594345539808, + 0.0049507468938827515, + 0.0060631451196968555, + 0.005644447170197964, + 0.007015789393335581, + 0.005462535656988621, + 0.0061627221293747425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.005253083072602749 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001248002052307 + ], + "opposite_logprob_tokens": [ + -2.6950125694274902 + ], + "logprob_reward": -0.07001248002052307, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001248002052307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23667630553245544, + 0.0763363316655159, + 0.14477285742759705, + 0.09547770768404007, + 0.04956423118710518, + 0.09702544659376144, + 0.12865862250328064, + 0.054260823875665665, + 0.0557248629629612, + 0.06150278449058533 + ], + "credit_map": [ + 0.04137038066983223, + 0.03703707456588745, + 0.026919737458229065, + 0.026492610573768616, + 0.02997283823788166, + 0.02896314300596714, + 0.020369883626699448, + 0.02555062063038349, + 0.006583577021956444, + 0.009073757566511631, + 0.006378154270350933, + 0.021983694285154343, + 0.008320077322423458, + 0.005600736476480961, + 0.008802395313978195, + 0.009593942202627659, + 0.010063418187201023, + 0.01565168984234333, + 0.011358805000782013, + 0.026844913139939308, + 0.021245403215289116, + 0.012345114722847939, + 0.023218601942062378, + 0.024044904857873917, + 0.014027331955730915, + 0.01003266591578722, + 0.007887279614806175, + 0.020796680822968483, + 0.008251933380961418, + 0.00913635641336441, + 0.011680316179990768, + 0.01366515178233385, + 0.003684796392917633, + 0.004191495478153229, + 0.0048903520219028, + 0.01070257369428873, + 0.007167238276451826, + 0.00287544634193182, + 0.005744526162743568, + 0.010307803750038147, + 0.01912612095475197, + 0.007880554534494877, + 0.014899815432727337, + 0.013471442274749279, + 0.006535365246236324, + 0.016875380650162697, + 0.008548139594495296, + 0.00968862883746624, + 0.011922772042453289, + 0.012623731046915054, + 0.009092026390135288, + 0.021488163620233536, + 0.013904851861298084, + 0.012673423625528812, + 0.027074379846453667, + 0.01987927407026291, + 0.004332328215241432, + 0.004853288177400827, + 0.00478851655498147, + 0.009501265361905098, + 0.006386866327375174, + 0.0039392998442053795, + 0.01212996244430542, + 0.008329298347234726, + 0.007377428002655506, + 0.005519625265151262, + 0.00458902632817626, + 0.006662264466285706, + 0.004590787459164858, + 0.006147261708974838, + 0.010748868808150291, + 0.010089598596096039, + 0.014428999274969101, + 0.005133019294589758, + 0.0066746873781085014, + 0.007163500413298607, + 0.006325560621917248, + 0.007468942552804947, + 0.007309244014322758, + 0.006998829543590546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.07001248002052307 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950829479843378 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950829479843378, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950829479843378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16759274899959564, + 0.1830868422985077, + 0.08951709419488907, + 0.06722549349069595, + 0.035127922892570496, + 0.25732675194740295, + 0.05105471983551979, + 0.03412662819027901, + 0.060111936181783676, + 0.05482984706759453 + ], + "credit_map": [ + 0.014187829568982124, + 0.031526464968919754, + 0.015201520174741745, + 0.02420954778790474, + 0.021186282858252525, + 0.02679814025759697, + 0.017841026186943054, + 0.01664193533360958, + 0.023077450692653656, + 0.028777578845620155, + 0.015549752861261368, + 0.05751075968146324, + 0.011634591966867447, + 0.009506721049547195, + 0.021723808720707893, + 0.01530617568641901, + 0.015204384922981262, + 0.010678204707801342, + 0.01092179398983717, + 0.01308799721300602, + 0.007279197219759226, + 0.009684659540653229, + 0.007762622088193893, + 0.014898237772285938, + 0.005998160224407911, + 0.007471971679478884, + 0.004558900836855173, + 0.017631858587265015, + 0.006403378676623106, + 0.005838513839989901, + 0.009015797637403011, + 0.01030691433697939, + 0.003937505651265383, + 0.0033235896844416857, + 0.003862595884129405, + 0.00589801836758852, + 0.0048057325184345245, + 0.0031352215446531773, + 0.0038322401233017445, + 0.0063330200500786304, + 0.08847716450691223, + 0.022643206641077995, + 0.03710724413394928, + 0.028775034472346306, + 0.01293747778981924, + 0.02755911462008953, + 0.01984509639441967, + 0.019982412457466125, + 0.010515782982110977, + 0.006625684909522533, + 0.0071448530070483685, + 0.005823066923767328, + 0.004422079771757126, + 0.006431691348552704, + 0.0058533623814582825, + 0.00423820223659277, + 0.002595692640170455, + 0.0034192607272416353, + 0.002910817740485072, + 0.0041697365231812, + 0.0026392813306301832, + 0.004508434794843197, + 0.006968740373849869, + 0.006914661265909672, + 0.005553492810577154, + 0.004720258992165327, + 0.00409694155678153, + 0.004177496302872896, + 0.005413527134805918, + 0.004767348524183035, + 0.016092829406261444, + 0.015290038660168648, + 0.0067652752622962, + 0.005742603447288275, + 0.004562173970043659, + 0.006174837239086628, + 0.008537574671208858, + 0.007934153079986572, + 0.007280332036316395, + 0.007832900620996952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.005950829479843378 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-shot shot, taken from a horizontal angle, captures a tranquil moment in documentary style. In a Tang Dynasty courtyard bathed in a cool, silvery glow, the swaying shadows of trees are reflected on the stone pavement. The poet Li Bai wears a wide-sleeved white silk robe, which flutters gently in the night breeze, revealing its delicate texture. His long, dark hair is held together only by a simple hairpin. His face is thin, and his eyes hold a hint of drunkenness and contemplation. He turns slowly and calmly, a movement that gently sways a small, white porcelain wine glass in his hand. Ripples ripple on the surface of the clear wine, momentarily catching and reflecting the bright moonlight. The entire scene is bathed in moonlight, creating a striking contrast between light and dark. The light outlines his graceful figure and the deep shadows of the courtyard scenery, creating a timeless, serene, and poetic classical atmosphere.", + "seed": 4782, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_4/index_8882_seed_4782_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06853353017941118, + "vlm_logprob": -0.06853353017941118, + "vlm_yes_no_margin": 3.449999976158142, + "vlm_token_credit": -0.06853353017941118, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17191581428050995 + ], + "opposite_logprob_tokens": [ + -1.9219157695770264 + ], + "logprob_reward": -0.17191581428050995, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17191581428050995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13512863218784332, + 0.1898975521326065, + 0.040827397257089615, + 0.04200083762407303, + 0.0630570650100708, + 0.0795988216996193, + 0.03136993944644928, + 0.19101236760616302, + 0.16931147873401642, + 0.05779585614800453 + ], + "credit_map": [ + 0.01389998383820057, + 0.017089156433939934, + 0.011273862794041634, + 0.020187299698591232, + 0.01600477658212185, + 0.023710379377007484, + 0.014800884760916233, + 0.018162304535508156, + 0.011968701146543026, + 0.014620148576796055, + 0.006926641333848238, + 0.011299006640911102, + 0.05601127818226814, + 0.017588764429092407, + 0.05021863803267479, + 0.021264391019940376, + 0.003902134718373418, + 0.0051025343127548695, + 0.004390334710478783, + 0.005860134493559599, + 0.0062606194987893105, + 0.007478027604520321, + 0.0032164244912564754, + 0.00461718812584877, + 0.003550548106431961, + 0.004323489964008331, + 0.004159065429121256, + 0.0030254132580012083, + 0.007452764548361301, + 0.0051200916059315205, + 0.004467978607863188, + 0.009901494719088078, + 0.005029409658163786, + 0.012998916208744049, + 0.0062424070201814175, + 0.008726567029953003, + 0.005090464372187853, + 0.005222532898187637, + 0.014516163617372513, + 0.005230612587183714, + 0.00795064214617014, + 0.011963593773543835, + 0.016205038875341415, + 0.00930954422801733, + 0.013147217221558094, + 0.004918681923300028, + 0.006416605319827795, + 0.009687508456408978, + 0.0027543173637241125, + 0.0037569624837487936, + 0.0033805870916694403, + 0.00478255795314908, + 0.0033556404523551464, + 0.006205941084772348, + 0.0027459801640361547, + 0.004387956112623215, + 0.015840893611311913, + 0.018675589933991432, + 0.06369553506374359, + 0.024888481944799423, + 0.008064008317887783, + 0.03970470651984215, + 0.01199953630566597, + 0.008143647573888302, + 0.015632497146725655, + 0.012038931250572205, + 0.038483135402202606, + 0.019058896228671074, + 0.01598244346678257, + 0.021088020876049995, + 0.019085265696048737, + 0.027942297980189323, + 0.005550938192754984, + 0.006892499979585409, + 0.0052665541879832745, + 0.007263293024152517, + 0.008055695332586765, + 0.006608366500586271, + 0.00956733152270317, + 0.00859118439257145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.17191581428050995 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981787547469139 + ], + "opposite_logprob_tokens": [ + -3.529817819595337 + ], + "logprob_reward": -0.02981787547469139, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981787547469139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07129118591547012, + 0.04642900079488754, + 0.022873908281326294, + 0.07814484089612961, + 0.09146096557378769, + 0.02546720951795578, + 0.04857345297932625, + 0.5436710119247437, + 0.025368554517626762, + 0.04671984910964966 + ], + "credit_map": [ + 0.007862674072384834, + 0.008970721624791622, + 0.0059899380430579185, + 0.010962857864797115, + 0.007267916575074196, + 0.007933131419122219, + 0.0095405587926507, + 0.012763393111526966, + 0.007301419973373413, + 0.0073318262584507465, + 0.0035403226502239704, + 0.005157955456525087, + 0.006913581397384405, + 0.004328504204750061, + 0.0056136599741876125, + 0.006241735070943832, + 0.002541022375226021, + 0.002491212682798505, + 0.0021768761798739433, + 0.004881443455815315, + 0.0041094692423939705, + 0.0018142999615520239, + 0.002057544654235244, + 0.002802040660753846, + 0.00582736823707819, + 0.00875443872064352, + 0.006754299160093069, + 0.005644768010824919, + 0.013179649598896503, + 0.009652805514633656, + 0.008771302178502083, + 0.019560210406780243, + 0.006179537624120712, + 0.01795794628560543, + 0.00946843158453703, + 0.012823302298784256, + 0.007640556897968054, + 0.005340994335711002, + 0.02258252538740635, + 0.009467682801187038, + 0.002443657023832202, + 0.0042173839174211025, + 0.0027020424604415894, + 0.002700801705941558, + 0.005183112807571888, + 0.0020405040122568607, + 0.003633300308138132, + 0.0025464105419814587, + 0.003557806834578514, + 0.005901452153921127, + 0.003964848816394806, + 0.004413376096636057, + 0.009527478367090225, + 0.00837028119713068, + 0.008281494490802288, + 0.004556721076369286, + 0.07116279751062393, + 0.05904710292816162, + 0.10197264701128006, + 0.05313967540860176, + 0.03806958720088005, + 0.14839762449264526, + 0.027244988828897476, + 0.044636644423007965, + 0.0033530774526298046, + 0.001718658022582531, + 0.007161223795264959, + 0.0031689386814832687, + 0.0025232434272766113, + 0.0021709559950977564, + 0.003206675173714757, + 0.0020657842978835106, + 0.005219621583819389, + 0.004605117253959179, + 0.006291994825005531, + 0.007143309805542231, + 0.007342963479459286, + 0.006581980735063553, + 0.0072860270738601685, + 0.00224883621558547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.02981787547469139 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254269111901522 + ], + "opposite_logprob_tokens": [ + -5.25525426864624 + ], + "logprob_reward": -0.005254269111901522, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254269111901522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06631579995155334, + 0.12827277183532715, + 0.03091450408101082, + 0.13507938385009766, + 0.10180952399969101, + 0.03527621552348137, + 0.03594722971320152, + 0.3557533025741577, + 0.03806197643280029, + 0.07256927341222763 + ], + "credit_map": [ + 0.0062080551870167255, + 0.010971903800964355, + 0.0046445573680102825, + 0.0119846286252141, + 0.007345096208155155, + 0.006395139265805483, + 0.008247132413089275, + 0.01051928848028183, + 0.009982640855014324, + 0.006416300777345896, + 0.003956087864935398, + 0.008389828726649284, + 0.00935251172631979, + 0.005894308444112539, + 0.03955287113785744, + 0.04472822695970535, + 0.0025790808722376823, + 0.003525807522237301, + 0.0034610917791724205, + 0.005657561123371124, + 0.004291442688554525, + 0.0021222783252596855, + 0.003231419948861003, + 0.006045823451131582, + 0.010554634965956211, + 0.014277083799242973, + 0.011532053351402283, + 0.008506824262440205, + 0.027339063584804535, + 0.016264477744698524, + 0.014241155236959457, + 0.03236409276723862, + 0.007380860857665539, + 0.024395031854510307, + 0.004290415905416012, + 0.008278386667370796, + 0.009617292322218418, + 0.006626506336033344, + 0.02736606076359749, + 0.013854963704943657, + 0.002825644798576832, + 0.004070516210049391, + 0.0033141830936074257, + 0.00474362401291728, + 0.00567922368645668, + 0.0028711820486932993, + 0.005623342469334602, + 0.0061485012993216515, + 0.0029021205846220255, + 0.004105446394532919, + 0.0033935909159481525, + 0.003954523243010044, + 0.00511690229177475, + 0.006045850925147533, + 0.003645758843049407, + 0.0067830379121005535, + 0.037336189299821854, + 0.03837675601243973, + 0.09057653695344925, + 0.03964834660291672, + 0.021518904715776443, + 0.09293745458126068, + 0.01952892728149891, + 0.015830207616090775, + 0.004515024833381176, + 0.0029920628294348717, + 0.004022830631583929, + 0.005032997578382492, + 0.004658624529838562, + 0.0036710998974740505, + 0.004032403230667114, + 0.009136931970715523, + 0.006884580943733454, + 0.005846734624356031, + 0.008496949449181557, + 0.011552435345947742, + 0.010577653534710407, + 0.006318049971014261, + 0.01212235912680626, + 0.010770515538752079 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.005254269111901522 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703976035118103 + ], + "opposite_logprob_tokens": [ + -2.127039670944214 + ], + "logprob_reward": -0.12703976035118103, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12703976035118103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08873332291841507, + 0.2070920616388321, + 0.03769076615571976, + 0.03473581001162529, + 0.07600855827331543, + 0.07119828462600708, + 0.13712801039218903, + 0.10126368701457977, + 0.21374613046646118, + 0.03240345045924187 + ], + "credit_map": [ + 0.009093568660318851, + 0.010544314980506897, + 0.006935233250260353, + 0.01603250578045845, + 0.011581091210246086, + 0.01090327650308609, + 0.009035377763211727, + 0.014607958495616913, + 0.010379059240221977, + 0.011580808088183403, + 0.004741033539175987, + 0.010330375283956528, + 0.038040850311517715, + 0.012399130500853062, + 0.06857717782258987, + 0.051043614745140076, + 0.003539362223818898, + 0.0037880372256040573, + 0.004068337846547365, + 0.005649428814649582, + 0.0043714046478271484, + 0.0032557575032114983, + 0.004497982095927, + 0.008520453236997128, + 0.0031733063515275717, + 0.003191182855516672, + 0.001912561827339232, + 0.0027993719559162855, + 0.004773764871060848, + 0.0037320845294743776, + 0.004997180309146643, + 0.0101563585922122, + 0.006388577166944742, + 0.009131547063589096, + 0.008116868324577808, + 0.01158339437097311, + 0.01061264704912901, + 0.0063576651737093925, + 0.010286207310855389, + 0.013531651347875595, + 0.005577214527875185, + 0.010365132242441177, + 0.007450419012457132, + 0.005465719383209944, + 0.01932385563850403, + 0.007056727074086666, + 0.008283603936433792, + 0.007675614673644304, + 0.011417911387979984, + 0.014417462050914764, + 0.01857822574675083, + 0.013779710978269577, + 0.014695804566144943, + 0.022086450830101967, + 0.01380028948187828, + 0.0283521581441164, + 0.021151218563318253, + 0.013913125731050968, + 0.005076773930341005, + 0.0054918862879276276, + 0.0077004022896289825, + 0.034237224608659744, + 0.005434788763523102, + 0.008258257061243057, + 0.020737025886774063, + 0.01362377218902111, + 0.052159350365400314, + 0.02127170003950596, + 0.01785852015018463, + 0.02498195320367813, + 0.023264802992343903, + 0.03984900936484337, + 0.0037812571972608566, + 0.0036501854192465544, + 0.0042457845993340015, + 0.0031088509131222963, + 0.005682051181793213, + 0.004217150155454874, + 0.004125851672142744, + 0.0035923158284276724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.12703976035118103 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639931678771973 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639931678771973, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639931678771973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1472739279270172, + 0.05050381273031235, + 0.024303792044520378, + 0.16047142446041107, + 0.054196517914533615, + 0.019154664129018784, + 0.028353020548820496, + 0.3507530093193054, + 0.07042884081602097, + 0.09456098079681396 + ], + "credit_map": [ + 0.016873477026820183, + 0.01751681976020336, + 0.011662459000945091, + 0.02341713011264801, + 0.017742525786161423, + 0.01756882295012474, + 0.02083662897348404, + 0.021656058728694916, + 0.0032269994262605906, + 0.0034827350173145533, + 0.0022600293159484863, + 0.0034618708305060863, + 0.008593982085585594, + 0.004068251233547926, + 0.014498380944132805, + 0.010911562480032444, + 0.0038890878204256296, + 0.0029750701505690813, + 0.0019326163455843925, + 0.003697129664942622, + 0.0042128232307732105, + 0.0029223598539829254, + 0.0018240652279928327, + 0.0028506394010037184, + 0.011889682151377201, + 0.016568342223763466, + 0.014997736550867558, + 0.011628147214651108, + 0.02579311840236187, + 0.021089449524879456, + 0.0179771576076746, + 0.04052779823541641, + 0.004487989470362663, + 0.008915768936276436, + 0.007641298230737448, + 0.010051107034087181, + 0.004068306647241116, + 0.0030289930291473866, + 0.010918582789599895, + 0.005084474571049213, + 0.0022847591899335384, + 0.0026168583426624537, + 0.0018151706317439675, + 0.0022743286099284887, + 0.00309063121676445, + 0.00224321149289608, + 0.002720983698964119, + 0.0021087229251861572, + 0.0025653743650764227, + 0.0031288776081055403, + 0.007033410482108593, + 0.0038697998970746994, + 0.0026084966957569122, + 0.0020230028312653303, + 0.0025561456568539143, + 0.004567914642393589, + 0.03803001716732979, + 0.039480771869421005, + 0.08925985544919968, + 0.0379539430141449, + 0.019660314545035362, + 0.095367431640625, + 0.01791244000196457, + 0.013088254258036613, + 0.008330674842000008, + 0.004881780594587326, + 0.01829848252236843, + 0.00896758958697319, + 0.005299542564898729, + 0.007745686452835798, + 0.007142384070903063, + 0.009762702509760857, + 0.010932275094091892, + 0.00976865366101265, + 0.013379834592342377, + 0.008359001018106937, + 0.011682333424687386, + 0.012449108064174652, + 0.008617712184786797, + 0.019372060894966125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.008639931678771973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-shot shot, taken from a horizontal angle, captures a tranquil moment in documentary style. In a Tang Dynasty courtyard bathed in a cool, silvery glow, the swaying shadows of trees are reflected on the stone pavement. The poet Li Bai wears a wide-sleeved white silk robe, which flutters gently in the night breeze, revealing its delicate texture. His long, dark hair is held together only by a simple hairpin. His face is thin, and his eyes hold a hint of drunkenness and contemplation. He turns slowly and calmly, a movement that gently sways a small, white porcelain wine glass in his hand. Ripples ripple on the surface of the clear wine, momentarily catching and reflecting the bright moonlight. The entire scene is bathed in moonlight, creating a striking contrast between light and dark. The light outlines his graceful figure and the deep shadows of the courtyard scenery, creating a timeless, serene, and poetic classical atmosphere.", + "seed": 4783, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_4/index_8882_seed_4783_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08455925863236188, + "vlm_logprob": -0.08455925863236188, + "vlm_yes_no_margin": 3.799999988079071, + "vlm_token_credit": -0.08455925863236188, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.36054831743240356 + ], + "opposite_logprob_tokens": [ + -1.2355482578277588 + ], + "logprob_reward": -0.36054831743240356, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.36054831743240356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2882881760597229, + 0.03872352093458176, + 0.10224185138940811, + 0.10478583723306656, + 0.0954679474234581, + 0.11015953123569489, + 0.06519296020269394, + 0.07928872853517532, + 0.0351824089884758, + 0.08066897839307785 + ], + "credit_map": [ + 0.02055356092751026, + 0.031954485923051834, + 0.02453404851257801, + 0.03123863972723484, + 0.04569720849394798, + 0.023506881669163704, + 0.04178685322403908, + 0.06901651620864868, + 0.0037908281665295362, + 0.005575675051659346, + 0.00377931771799922, + 0.004533016122877598, + 0.006156575866043568, + 0.004162365570664406, + 0.005468752235174179, + 0.0052569895051419735, + 0.006227189674973488, + 0.010186034254729748, + 0.010439349338412285, + 0.008471975103020668, + 0.024186942726373672, + 0.01611090637743473, + 0.011862521059811115, + 0.01475693192332983, + 0.008366971276700497, + 0.013474407605826855, + 0.00820257980376482, + 0.009584144689142704, + 0.014858738519251347, + 0.019131332635879517, + 0.015685508027672768, + 0.015482156537473202, + 0.00962921418249607, + 0.009625024162232876, + 0.008816523477435112, + 0.00933724083006382, + 0.012481438927352428, + 0.018729377537965775, + 0.01259435061365366, + 0.014254775829613209, + 0.010039347223937511, + 0.01243514847010374, + 0.015588643960654736, + 0.01110260933637619, + 0.014500175602734089, + 0.018022561445832253, + 0.01186494529247284, + 0.01660609431564808, + 0.005289182998239994, + 0.011495574377477169, + 0.008281861431896687, + 0.006390892900526524, + 0.008482292294502258, + 0.0070990812964737415, + 0.00943752657622099, + 0.008716550655663013, + 0.007542667910456657, + 0.010725939646363258, + 0.007146724034100771, + 0.008873672224581242, + 0.010603709146380424, + 0.012374057434499264, + 0.010177908465266228, + 0.011844048276543617, + 0.005272968672215939, + 0.003880035365000367, + 0.00441258167847991, + 0.004483758006244898, + 0.0040137856267392635, + 0.004018782638013363, + 0.003853096393868327, + 0.005247399210929871, + 0.00570592051371932, + 0.006615712307393551, + 0.005625362507998943, + 0.026157589629292488, + 0.01765657775104046, + 0.006952625699341297, + 0.006397009361535311, + 0.005558187607675791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.36054831743240356 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023301467299461365 + ], + "opposite_logprob_tokens": [ + -3.773301362991333 + ], + "logprob_reward": -0.023301467299461365, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023301467299461365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1510683000087738, + 0.0787346363067627, + 0.07170499861240387, + 0.09598720073699951, + 0.06949643790721893, + 0.08772438764572144, + 0.06329087167978287, + 0.07108449935913086, + 0.06845241039991379, + 0.2424563318490982 + ], + "credit_map": [ + 0.010855354368686676, + 0.012909231707453728, + 0.017166554927825928, + 0.021810127422213554, + 0.014838863164186478, + 0.014438675716519356, + 0.023463279008865356, + 0.035586193203926086, + 0.00520545756444335, + 0.017619742080569267, + 0.007471967022866011, + 0.01414389070123434, + 0.007379069458693266, + 0.008874312974512577, + 0.009044747799634933, + 0.008995447307825089, + 0.005110531114041805, + 0.008384013548493385, + 0.0070165484212338924, + 0.007912932895123959, + 0.014931181445717812, + 0.008648846298456192, + 0.008655763231217861, + 0.011045178398489952, + 0.011823403649032116, + 0.014065367169678211, + 0.010864273644983768, + 0.012372703291475773, + 0.009495214559137821, + 0.01004054769873619, + 0.013331260532140732, + 0.013994409702718258, + 0.007170192897319794, + 0.008495986461639404, + 0.008083072490990162, + 0.009906665422022343, + 0.008047174662351608, + 0.0088728666305542, + 0.008999516256153584, + 0.00992096308618784, + 0.006570762954652309, + 0.009566049091517925, + 0.011038036085665226, + 0.017050931230187416, + 0.00921971257776022, + 0.0076706684194505215, + 0.01274810079485178, + 0.013860125094652176, + 0.004997187294065952, + 0.008403025567531586, + 0.006940999999642372, + 0.006073335651308298, + 0.007098040077835321, + 0.007002652622759342, + 0.013823917135596275, + 0.008951711468398571, + 0.006921148858964443, + 0.011090708896517754, + 0.008410399779677391, + 0.00781719945371151, + 0.006738537456840277, + 0.011739795096218586, + 0.0076055703684687614, + 0.010761129669845104, + 0.01849934086203575, + 0.005912469699978828, + 0.006727361585944891, + 0.006724800448864698, + 0.0075725531205534935, + 0.009305018000304699, + 0.004933396354317665, + 0.00877746008336544, + 0.014225400984287262, + 0.011207331903278828, + 0.008380027487874031, + 0.046785131096839905, + 0.09450967609882355, + 0.016196012496948242, + 0.013322672806680202, + 0.03783005103468895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.023301467299461365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031949449330568314 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031949449330568314, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031949449330568314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21100685000419617, + 0.06515566259622574, + 0.14255817234516144, + 0.10902267694473267, + 0.06567326188087463, + 0.05172773450613022, + 0.08475563675165176, + 0.1685442477464676, + 0.02062499336898327, + 0.08093079179525375 + ], + "credit_map": [ + 0.013199643231928349, + 0.016433628275990486, + 0.017879802733659744, + 0.02029437944293022, + 0.03914138302206993, + 0.018306246027350426, + 0.02996245212852955, + 0.05578930675983429, + 0.005000376142561436, + 0.009710123762488365, + 0.006000394467264414, + 0.011075431481003761, + 0.008289660327136517, + 0.009109231643378735, + 0.008306330069899559, + 0.007664112839847803, + 0.008349606767296791, + 0.018630515784025192, + 0.011527884751558304, + 0.01109794806689024, + 0.03180638328194618, + 0.020942751318216324, + 0.01995948702096939, + 0.020243583247065544, + 0.006668019574135542, + 0.016170836985111237, + 0.00767558766528964, + 0.010058450512588024, + 0.014855720102787018, + 0.016734028235077858, + 0.01920558325946331, + 0.017654435709118843, + 0.007156055420637131, + 0.00706168357282877, + 0.0071924952790141106, + 0.006213367450982332, + 0.009616623632609844, + 0.011768385767936707, + 0.007060325238853693, + 0.009604324586689472, + 0.004192707594484091, + 0.006007250864058733, + 0.006603187415748835, + 0.006785528734326363, + 0.006079715210944414, + 0.007772146724164486, + 0.006946598645299673, + 0.007340592797845602, + 0.00714336009696126, + 0.011438221670687199, + 0.009335197508335114, + 0.0075155883096158504, + 0.011156592518091202, + 0.009500931948423386, + 0.01663370057940483, + 0.012032031081616879, + 0.016856154426932335, + 0.019717736169695854, + 0.014851624146103859, + 0.016377629712224007, + 0.02692907489836216, + 0.028249498456716537, + 0.018950141966342926, + 0.026612378656864166, + 0.002525629010051489, + 0.002469718223437667, + 0.002213046420365572, + 0.002607960719615221, + 0.0027964101172983646, + 0.0031543890945613384, + 0.0025064311921596527, + 0.0023514071945101023, + 0.007276136428117752, + 0.006549079902470112, + 0.003472372656688094, + 0.011241462081670761, + 0.02671399526298046, + 0.010229702107608318, + 0.006116464734077454, + 0.009331570006906986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.0031949449330568314 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029801445081830025 + ], + "opposite_logprob_tokens": [ + -3.529801368713379 + ], + "logprob_reward": -0.029801445081830025, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801445081830025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11462432891130447, + 0.07609029114246368, + 0.14492926001548767, + 0.13585680723190308, + 0.1275579184293747, + 0.08500904589891434, + 0.07343943417072296, + 0.0859358012676239, + 0.05818932503461838, + 0.09836771339178085 + ], + "credit_map": [ + 0.011606470681726933, + 0.018838683143258095, + 0.013897404074668884, + 0.014782684855163097, + 0.015335654839873314, + 0.016461146995425224, + 0.012537294067442417, + 0.011164993979036808, + 0.005457710474729538, + 0.010801758617162704, + 0.009171586483716965, + 0.015590229071676731, + 0.007471743505448103, + 0.008689277805387974, + 0.00965107697993517, + 0.00925691332668066, + 0.008512568660080433, + 0.017617665231227875, + 0.01767396181821823, + 0.013118376955389977, + 0.021648908033967018, + 0.017904357984662056, + 0.028876937925815582, + 0.019576480612158775, + 0.011911597102880478, + 0.021843256428837776, + 0.009660255163908005, + 0.012239907868206501, + 0.01917012594640255, + 0.019916057586669922, + 0.02102092280983925, + 0.020094703882932663, + 0.008682332001626492, + 0.012613341212272644, + 0.013659143820405006, + 0.016061043366789818, + 0.016763361170887947, + 0.013545548543334007, + 0.02787640690803528, + 0.018356749787926674, + 0.005866931285709143, + 0.010006819851696491, + 0.008034385740756989, + 0.01058702077716589, + 0.010089321993291378, + 0.007599685341119766, + 0.02020806446671486, + 0.012616826221346855, + 0.006352908909320831, + 0.00837424211204052, + 0.007531106472015381, + 0.00821894034743309, + 0.009224318899214268, + 0.011938407085835934, + 0.011723059229552746, + 0.010076453909277916, + 0.0073724049143493176, + 0.013706074096262455, + 0.009008265100419521, + 0.007375644985586405, + 0.015375160612165928, + 0.008651603013277054, + 0.010966594330966473, + 0.013480057939887047, + 0.009703781455755234, + 0.006872529163956642, + 0.0057850126177072525, + 0.00737657118588686, + 0.00812553707510233, + 0.006920914631336927, + 0.006297789514064789, + 0.007107199169695377, + 0.009854050353169441, + 0.008853944018483162, + 0.007338198833167553, + 0.02253788709640503, + 0.014742233790457249, + 0.0075982133857905865, + 0.008675971068441868, + 0.018767226487398148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.029801445081830025 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950118415057659 + ], + "opposite_logprob_tokens": [ + -5.130949974060059 + ], + "logprob_reward": -0.005950118415057659, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950118415057659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2765369415283203, + 0.07562398910522461, + 0.07177167385816574, + 0.077669158577919, + 0.05226120725274086, + 0.09828321635723114, + 0.051256753504276276, + 0.09038211405277252, + 0.030321786180138588, + 0.1758931577205658 + ], + "credit_map": [ + 0.013484353199601173, + 0.01782147027552128, + 0.01757478155195713, + 0.016891619190573692, + 0.06676950305700302, + 0.019328325986862183, + 0.03693059831857681, + 0.08773630112409592, + 0.006993310991674662, + 0.01001181360334158, + 0.007774298544973135, + 0.010069540701806545, + 0.012033511884510517, + 0.008398099802434444, + 0.01189636904746294, + 0.008447041735053062, + 0.003663355018943548, + 0.0074201482348144054, + 0.009138091467320919, + 0.006345769856125116, + 0.01732579991221428, + 0.010376238264143467, + 0.00795016624033451, + 0.009552109986543655, + 0.004684953950345516, + 0.008767800405621529, + 0.0057639917358756065, + 0.006108327768743038, + 0.011666744947433472, + 0.01425107754766941, + 0.013742159120738506, + 0.012684098444879055, + 0.0037130373530089855, + 0.006094557233154774, + 0.005223408807069063, + 0.005900600925087929, + 0.0073388791643083096, + 0.0077475556172430515, + 0.008128318935632706, + 0.008114849217236042, + 0.008914932608604431, + 0.015429606661200523, + 0.012338836677372456, + 0.010360106825828552, + 0.011902278289198875, + 0.013810570351779461, + 0.011457704938948154, + 0.014069177210330963, + 0.005354135762900114, + 0.006874029990285635, + 0.008262569084763527, + 0.0042126900516450405, + 0.006402191706001759, + 0.008307385258376598, + 0.004715021699666977, + 0.0071287271566689014, + 0.0074584875255823135, + 0.007209598086774349, + 0.006975581403821707, + 0.006836616899818182, + 0.018291832879185677, + 0.01639290153980255, + 0.011058194562792778, + 0.016158906742930412, + 0.004325482994318008, + 0.0033057311084121466, + 0.003032095031812787, + 0.004294096026569605, + 0.003588619874790311, + 0.004445169121026993, + 0.002955494448542595, + 0.004375098738819361, + 0.01333802193403244, + 0.008897748775780201, + 0.005397763568907976, + 0.011084258556365967, + 0.06937123090028763, + 0.014247912913560867, + 0.01115118432790041, + 0.04240504652261734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.005950118415057659 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-shot shot, taken from a horizontal angle, captures a tranquil moment in documentary style. In a Tang Dynasty courtyard bathed in a cool, silvery glow, the swaying shadows of trees are reflected on the stone pavement. The poet Li Bai wears a wide-sleeved white silk robe, which flutters gently in the night breeze, revealing its delicate texture. His long, dark hair is held together only by a simple hairpin. His face is thin, and his eyes hold a hint of drunkenness and contemplation. He turns slowly and calmly, a movement that gently sways a small, white porcelain wine glass in his hand. Ripples ripple on the surface of the clear wine, momentarily catching and reflecting the bright moonlight. The entire scene is bathed in moonlight, creating a striking contrast between light and dark. The light outlines his graceful figure and the deep shadows of the courtyard scenery, creating a timeless, serene, and poetic classical atmosphere.", + "seed": 4784, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_4/index_8882_seed_4784_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.22215127777308225, + "vlm_logprob": -0.22215127777308225, + "vlm_yes_no_margin": 2.674999976158142, + "vlm_token_credit": -0.22215127777308225, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9050654768943787 + ], + "opposite_logprob_tokens": [ + -0.5300654768943787 + ], + "logprob_reward": -0.9050654768943787, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9050654768943787, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16452088952064514, + 0.05048789829015732, + 0.35124221444129944, + 0.030265435576438904, + 0.24264761805534363, + 0.023213258013129234, + 0.03449912369251251, + 0.021754920482635498, + 0.018642805516719818, + 0.0627257451415062 + ], + "credit_map": [ + 0.01500645000487566, + 0.018270252272486687, + 0.03478711470961571, + 0.01080665085464716, + 0.00925310980528593, + 0.04712396487593651, + 0.011352135799825191, + 0.017921218648552895, + 0.0082358717918396, + 0.009779773652553558, + 0.005330351181328297, + 0.006349132861942053, + 0.005165203008800745, + 0.007750844117254019, + 0.004390978254377842, + 0.003485746681690216, + 0.01976258121430874, + 0.028734665364027023, + 0.016674797981977463, + 0.08095603436231613, + 0.012944584712386131, + 0.1488179862499237, + 0.02194831147789955, + 0.021403271704912186, + 0.0024984225165098906, + 0.004029311705380678, + 0.0042551239021122456, + 0.005603699013590813, + 0.005007818806916475, + 0.003404084127396345, + 0.0025698705576360226, + 0.0028971051797270775, + 0.026099756360054016, + 0.015349767170846462, + 0.045659761875867844, + 0.1061200499534607, + 0.0061724246479570866, + 0.015916645526885986, + 0.01710353046655655, + 0.010225705802440643, + 0.002368608256801963, + 0.0024751697201281786, + 0.004477555863559246, + 0.007465966045856476, + 0.0019774280954152346, + 0.0018577530281618237, + 0.001362113282084465, + 0.0012286644196137786, + 0.006142789032310247, + 0.0038516297936439514, + 0.004454542882740498, + 0.004056296311318874, + 0.0040053450502455235, + 0.00555308535695076, + 0.0028091378044337034, + 0.003626297926530242, + 0.003949148114770651, + 0.003734288737177849, + 0.002021192340180278, + 0.0012548165395855904, + 0.0016633273335173726, + 0.0064244624227285385, + 0.001524896826595068, + 0.0011827893322333694, + 0.00314259622246027, + 0.003330506617203355, + 0.0017268587835133076, + 0.0010410096729174256, + 0.004233922343701124, + 0.0021164054051041603, + 0.0015619398327544332, + 0.0014895661734044552, + 0.009965852834284306, + 0.013310601934790611, + 0.0046658883802592754, + 0.0040177470073103905, + 0.007436337880790234, + 0.01232241652905941, + 0.004985509905964136, + 0.006021388806402683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.9050654768943787 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029812553897500038 + ], + "opposite_logprob_tokens": [ + -3.5298125743865967 + ], + "logprob_reward": -0.029812553897500038, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029812553897500038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42232343554496765, + 0.09438221901655197, + 0.23287717998027802, + 0.04018668085336685, + 0.08768714219331741, + 0.02387707307934761, + 0.028994236141443253, + 0.01943771168589592, + 0.02491971105337143, + 0.02531461976468563 + ], + "credit_map": [ + 0.025389567017555237, + 0.02373768389225006, + 0.12188884615898132, + 0.02908380888402462, + 0.025486517697572708, + 0.14626650512218475, + 0.025997979566454887, + 0.024472538381814957, + 0.012006891891360283, + 0.012544495984911919, + 0.007004005368798971, + 0.01232828851789236, + 0.020847799256443977, + 0.0122337331995368, + 0.009402885101735592, + 0.00801412295550108, + 0.018145255744457245, + 0.019148072227835655, + 0.031477369368076324, + 0.04670552536845207, + 0.008429476991295815, + 0.05781041085720062, + 0.01731107197701931, + 0.03385001793503761, + 0.0035002827644348145, + 0.0039054204244166613, + 0.004722287878394127, + 0.009069817140698433, + 0.0066522592678666115, + 0.003934170585125685, + 0.003247038461267948, + 0.005155405029654503, + 0.008451925590634346, + 0.007226087152957916, + 0.012302973307669163, + 0.02852095291018486, + 0.007463772315531969, + 0.0086038364097476, + 0.006237976718693972, + 0.008879616856575012, + 0.0032507546711713076, + 0.00271041551604867, + 0.0020334217697381973, + 0.004093131981790066, + 0.0035347219090908766, + 0.0030939343851059675, + 0.0018432034412398934, + 0.0033174895215779543, + 0.0062892138957977295, + 0.0029374670702964067, + 0.0028355533722788095, + 0.002462656470015645, + 0.005736258812248707, + 0.002777555724605918, + 0.0026119521353393793, + 0.0033435800578445196, + 0.0014426849083974957, + 0.0015146172372624278, + 0.001423720852471888, + 0.00176941545214504, + 0.004963958635926247, + 0.003964472562074661, + 0.001946910284459591, + 0.0024119324516505003, + 0.002653267700225115, + 0.0023607362527400255, + 0.0015377141535282135, + 0.00220838887616992, + 0.007240587845444679, + 0.0029609622433781624, + 0.002690161345526576, + 0.0032678935676813126, + 0.0036525968462228775, + 0.0037436799611896276, + 0.001624021795578301, + 0.0018197305034846067, + 0.00460911076515913, + 0.0042740474455058575, + 0.0026091602630913258, + 0.0029822734650224447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.029812553897500038 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780935011804104 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780935011804104, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780935011804104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17309816181659698, + 0.04981265217065811, + 0.5109113454818726, + 0.059629883617162704, + 0.0876348465681076, + 0.017961980774998665, + 0.035232916474342346, + 0.013165983371436596, + 0.018147684633731842, + 0.034404605627059937 + ], + "credit_map": [ + 0.013521878980100155, + 0.014269586652517319, + 0.041666429489851, + 0.0118313143029809, + 0.010281069204211235, + 0.049221742898225784, + 0.012644192203879356, + 0.019661935046315193, + 0.0064422315917909145, + 0.006155210081487894, + 0.005446872673928738, + 0.006545816082507372, + 0.00655811931937933, + 0.00585533631965518, + 0.004385403357446194, + 0.008423661813139915, + 0.02748706005513668, + 0.03838956728577614, + 0.021291954442858696, + 0.11867675185203552, + 0.019186805933713913, + 0.21823690831661224, + 0.03003920055925846, + 0.037603043019771576, + 0.005556132644414902, + 0.006276991683989763, + 0.005149989388883114, + 0.005279280245304108, + 0.007746702991425991, + 0.005698663182556629, + 0.00861191563308239, + 0.015310199931263924, + 0.009213401935994625, + 0.008081807754933834, + 0.01391010545194149, + 0.02960834465920925, + 0.0063593219965696335, + 0.005473604425787926, + 0.006543505471199751, + 0.008444750681519508, + 0.0025875114370137453, + 0.002409163862466812, + 0.0032338195014744997, + 0.0036093490198254585, + 0.0020925053395330906, + 0.0015026748878881335, + 0.0011776488972827792, + 0.0013493066653609276, + 0.007753653917461634, + 0.004564470145851374, + 0.005324417259544134, + 0.003442948218435049, + 0.0029629224445670843, + 0.004245075862854719, + 0.003156653605401516, + 0.0037827733904123306, + 0.001715244841761887, + 0.001545775798149407, + 0.001116003841161728, + 0.0009790041949599981, + 0.0014551755739375949, + 0.0036455958615988493, + 0.0012161890044808388, + 0.0014929932076483965, + 0.002480820519849658, + 0.0022952803410589695, + 0.001547188381664455, + 0.001673795050010085, + 0.0043500736355781555, + 0.0019445543875917792, + 0.0017653227550908923, + 0.00209064781665802, + 0.006888787727802992, + 0.007852018810808659, + 0.0028672260232269764, + 0.0018599667819216847, + 0.0035889591090381145, + 0.006435563787817955, + 0.0026292153634130955, + 0.002282865345478058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.009780935011804104 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280375838279724 + ], + "opposite_logprob_tokens": [ + -2.01780366897583 + ], + "logprob_reward": -0.14280375838279724, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14280375838279724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19268600642681122, + 0.0646691620349884, + 0.46132320165634155, + 0.035514745861291885, + 0.07456931471824646, + 0.03416815772652626, + 0.0539245642721653, + 0.023630212992429733, + 0.017597399652004242, + 0.04191723093390465 + ], + "credit_map": [ + 0.013015602715313435, + 0.011311790905892849, + 0.055828388780355453, + 0.015532375313341618, + 0.009618696756660938, + 0.06497623026371002, + 0.011054066009819508, + 0.011348849162459373, + 0.006937744561582804, + 0.00889511313289404, + 0.004396616946905851, + 0.006948523223400116, + 0.008434456773102283, + 0.009937543421983719, + 0.005298085510730743, + 0.013821080327033997, + 0.02147396095097065, + 0.030731096863746643, + 0.01925288513302803, + 0.15575315058231354, + 0.013679856434464455, + 0.171474426984787, + 0.02754858136177063, + 0.021409226581454277, + 0.0026733586564660072, + 0.0030343299731612206, + 0.003257031785324216, + 0.0044668130576610565, + 0.0059884837828576565, + 0.0031944119837135077, + 0.005325317848473787, + 0.007574999239295721, + 0.008914696983993053, + 0.00985068827867508, + 0.0110270781442523, + 0.012660722248256207, + 0.009179800748825073, + 0.005887523293495178, + 0.0069154477678239346, + 0.010133352130651474, + 0.0027401612605899572, + 0.0024539190344512463, + 0.006108884233981371, + 0.011331312358379364, + 0.0034999747294932604, + 0.00222177361138165, + 0.002701979596167803, + 0.003110154764726758, + 0.008863294497132301, + 0.005229210015386343, + 0.005951358005404472, + 0.007996984757483006, + 0.006894819438457489, + 0.0063688065856695175, + 0.0062057203613221645, + 0.006414365489035845, + 0.004119512625038624, + 0.0031556289177387953, + 0.0021057273261249065, + 0.0022530234418809414, + 0.0025455718860030174, + 0.006304590031504631, + 0.0015021080616861582, + 0.001644051051698625, + 0.002063668565824628, + 0.0019084811210632324, + 0.0011578714475035667, + 0.0015531058888882399, + 0.0044356342405080795, + 0.002739385236054659, + 0.0017021859530359507, + 0.002037067199125886, + 0.006529196165502071, + 0.00996712688356638, + 0.0031633516773581505, + 0.0025653159245848656, + 0.003823536913841963, + 0.007072979584336281, + 0.0038113247137516737, + 0.004984400700777769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.14280375838279724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a Tang Dynasty courtyard with stone pavement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character depicted as a thin man with long dark hair held by a simple hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the character wearing a wide-sleeved white silk robe that flutters in the breeze?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the character holding a small white porcelain wine glass with clear wine showing ripples and moonlight reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool silvery moonlight creating high contrast between light and dark shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023293664678931236 + ], + "opposite_logprob_tokens": [ + -3.7732937335968018 + ], + "logprob_reward": -0.023293664678931236, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023293664678931236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.397688627243042, + 0.14667564630508423, + 0.07713250070810318, + 0.08720406889915466, + 0.12137281149625778, + 0.046086445450782776, + 0.03888873755931854, + 0.025355685502290726, + 0.012870020233094692, + 0.04672542214393616 + ], + "credit_map": [ + 0.02755551226437092, + 0.034733131527900696, + 0.10763686895370483, + 0.029435178264975548, + 0.018008152022957802, + 0.13872721791267395, + 0.0229195524007082, + 0.018673041835427284, + 0.018702145665884018, + 0.024027444422245026, + 0.01601044274866581, + 0.01633547432720661, + 0.014224431477487087, + 0.036531008780002594, + 0.010179424658417702, + 0.010665299370884895, + 0.006076055113226175, + 0.007429522927850485, + 0.007550001610070467, + 0.011334353126585484, + 0.0031962234061211348, + 0.021594176068902016, + 0.004864581860601902, + 0.015087597072124481, + 0.00675172870978713, + 0.013537871651351452, + 0.010769029147922993, + 0.020546922460198402, + 0.008247646503150463, + 0.008194904774427414, + 0.009910482913255692, + 0.009245485067367554, + 0.011605193838477135, + 0.01202971488237381, + 0.017481433227658272, + 0.03730832412838936, + 0.009583493694663048, + 0.01186391245573759, + 0.008568395860493183, + 0.012932355515658855, + 0.009509028866887093, + 0.005643956363201141, + 0.006806373130530119, + 0.006646304856985807, + 0.00548971863463521, + 0.004607116803526878, + 0.0036975820548832417, + 0.0036863707937300205, + 0.005674757529050112, + 0.00623681116849184, + 0.006209222599864006, + 0.0042322054505348206, + 0.0036255549639463425, + 0.00583729287609458, + 0.003393711755052209, + 0.0036791879683732986, + 0.005142014008015394, + 0.005154680460691452, + 0.0022883815690875053, + 0.001655989675782621, + 0.002058898564428091, + 0.005151090212166309, + 0.0017742420313879848, + 0.002130393637344241, + 0.002598897088319063, + 0.001959484303370118, + 0.0008854743791744113, + 0.0015920802252367139, + 0.0009933795081451535, + 0.001562955672852695, + 0.0012833954533562064, + 0.001994355348870158, + 0.00834745541214943, + 0.009969192557036877, + 0.004093879368156195, + 0.00285648787394166, + 0.004254776518791914, + 0.007499951869249344, + 0.004305532667785883, + 0.005398150067776442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.023293664678931236 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_5/index_tensor([7437, 7437])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_5/index_tensor([7437, 7437])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..caba0f4aa2549ec9d0cc57a61cd45ed334b484da --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_5/index_tensor([7437, 7437])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up shot from a level angle, the camera slowly and steadily moves forward, focusing on a vast wheat field bathed in the golden evening sunlight. A young mother, with long, warm brown hair that flutters gently in the breeze, wears a gentle, contented smile, her eyes filled with maternal love. Before her stands a child, perhaps five or six years old, with short, soft golden hair and cheeks slightly flushed by the sun, looking up at her with her tender face. The mother slowly reaches out her right hand and, with the back of her slender fingers, carefully and lovingly strokes the child's smooth cheek. The child, with eyes slightly closed, savours this moment of tenderness with complete trust. The gentle sunset casts a warm glow on their hair and figures. Behind them, heavy ears of wheat ripple gently in the breeze, enveloping the entire scene in a tranquil, warm, and affectionate atmosphere.", + "seed": 33471, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_5/index_7437_seed_33471_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11269193589687347, + "vlm_logprob": -0.11269193589687347, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.11269193589687347, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03653210401535034 + ], + "opposite_logprob_tokens": [ + -5.036531925201416 + ], + "logprob_reward": -0.03653210401535034, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.03653210401535034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1102786734700203, + 0.052852652966976166, + 0.17807212471961975, + 0.05647572502493858, + 0.07392025738954544, + 0.27676162123680115, + 0.06540592014789581, + 0.05992179363965988, + 0.0748036727309227, + 0.05150749906897545 + ], + "credit_map": [ + 0.013395095244050026, + 0.00613813940435648, + 0.014575495384633541, + 0.009456118568778038, + 0.007409598678350449, + 0.017501240596175194, + 0.013673791661858559, + 0.028129195794463158, + 0.009399943053722382, + 0.004885225556790829, + 0.0059636542573571205, + 0.006448186002671719, + 0.0034859750885516405, + 0.012261928990483284, + 0.005070733372122049, + 0.0053370059467852116, + 0.009201684035360813, + 0.008449718356132507, + 0.0121364276856184, + 0.0590374618768692, + 0.00804259441792965, + 0.046016398817300797, + 0.008576100692152977, + 0.02661173604428768, + 0.004814366810023785, + 0.0063325888477265835, + 0.011572305113077164, + 0.010544348508119583, + 0.0056722527369856834, + 0.004098643083125353, + 0.007266397587954998, + 0.006174822337925434, + 0.005419075954705477, + 0.006926761008799076, + 0.015402339398860931, + 0.010068393312394619, + 0.005527260713279247, + 0.007321072742342949, + 0.013918308541178703, + 0.009337044321000576, + 0.016455264762043953, + 0.013340056873857975, + 0.018023112788796425, + 0.03228786960244179, + 0.018343692645430565, + 0.059196677058935165, + 0.09649208188056946, + 0.022622860968112946, + 0.005351023282855749, + 0.006549017503857613, + 0.008116024546325207, + 0.011093148030340672, + 0.0063940915279090405, + 0.005768126808106899, + 0.013920065946877003, + 0.00821442436426878, + 0.004092592746019363, + 0.005804464686661959, + 0.01151940319687128, + 0.006625671871006489, + 0.005184873007237911, + 0.009939765557646751, + 0.010560224764049053, + 0.006194799207150936, + 0.006412420887500048, + 0.010706364177167416, + 0.007424265146255493, + 0.01228680182248354, + 0.006483889650553465, + 0.007054663263261318, + 0.01723693311214447, + 0.0071983328089118, + 0.004082633648067713, + 0.006798948626965284, + 0.006459792610257864, + 0.0069457790814340115, + 0.003889249637722969, + 0.0034987800754606724, + 0.012093573808670044, + 0.0077387443743646145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.03653210401535034 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742891788482666 + ], + "opposite_logprob_tokens": [ + -0.9742891788482666 + ], + "logprob_reward": -0.4742891788482666, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742891788482666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029379423707723618, + 0.034055717289447784, + 0.06358671188354492, + 0.056096844375133514, + 0.04534371942281723, + 0.6591106057167053, + 0.06371407955884933, + 0.01352451741695404, + 0.026359472423791885, + 0.00882893055677414 + ], + "credit_map": [ + 0.005347617901861668, + 0.0025245121214538813, + 0.003880253992974758, + 0.003920333925634623, + 0.001462931977584958, + 0.004444627556949854, + 0.003352764993906021, + 0.004446378443390131, + 0.007986409589648247, + 0.002380814403295517, + 0.003427016083151102, + 0.0025842422619462013, + 0.0020324934739619493, + 0.009379500523209572, + 0.0035214857198297977, + 0.0027437515091151, + 0.003686537966132164, + 0.002988708671182394, + 0.0043040188029408455, + 0.021839190274477005, + 0.0030873622745275497, + 0.01365366019308567, + 0.0035634110681712627, + 0.010463817045092583, + 0.0046564438380301, + 0.004304161295294762, + 0.014650489203631878, + 0.012987278401851654, + 0.004209965001791716, + 0.003496521385386586, + 0.006473967339843512, + 0.005318008828908205, + 0.004280623514205217, + 0.0039558554999530315, + 0.006444031372666359, + 0.005368656944483519, + 0.003909135237336159, + 0.004327908158302307, + 0.009054613299667835, + 0.008002884685993195, + 0.0356748066842556, + 0.018723784014582634, + 0.04111495986580849, + 0.09577387571334839, + 0.02809326723217964, + 0.12372922897338867, + 0.2770450711250305, + 0.03895556926727295, + 0.003907750826328993, + 0.004434145987033844, + 0.01594301499426365, + 0.005380796268582344, + 0.00376061606220901, + 0.014650355093181133, + 0.010206767357885838, + 0.0054306271485984325, + 0.001212658709846437, + 0.0010845657670870423, + 0.0018758868100121617, + 0.0019209665479138494, + 0.001147611066699028, + 0.002052937401458621, + 0.0023939316160976887, + 0.0018359568202868104, + 0.0019455101573839784, + 0.003233382012695074, + 0.003170865122228861, + 0.004337445832788944, + 0.002952879760414362, + 0.0031468544621020555, + 0.00387800601311028, + 0.00369452522136271, + 0.0006704786792397499, + 0.001301366719417274, + 0.0013220033142715693, + 0.0012529530795291066, + 0.0005978117696940899, + 0.0006756259244866669, + 0.001654506428167224, + 0.0013541840016841888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 2 + ], + "selected_reward": -0.4742891788482666 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181988924741745 + ], + "opposite_logprob_tokens": [ + -4.0181989669799805 + ], + "logprob_reward": -0.0181988924741745, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0181988924741745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11653876304626465, + 0.18144990503787994, + 0.205635204911232, + 0.03348411247134209, + 0.1012439876794815, + 0.12889578938484192, + 0.1541275531053543, + 0.030321329832077026, + 0.03406465798616409, + 0.014238651841878891 + ], + "credit_map": [ + 0.019958704710006714, + 0.008920012041926384, + 0.015601899474859238, + 0.008655903860926628, + 0.006093358155339956, + 0.014420750550925732, + 0.009849080815911293, + 0.033039066940546036, + 0.06368987262248993, + 0.006943878252059221, + 0.009557047858834267, + 0.008659986779093742, + 0.007271443959325552, + 0.06653891503810883, + 0.008536525070667267, + 0.010252262465655804, + 0.013621164485812187, + 0.00941675528883934, + 0.0127241937443614, + 0.07227111607789993, + 0.00784002710133791, + 0.04590125009417534, + 0.010265541262924671, + 0.033595163375139236, + 0.004256812389940023, + 0.003794158110395074, + 0.00645500747486949, + 0.0064640287309885025, + 0.003426916664466262, + 0.003088418859988451, + 0.0026094953063875437, + 0.0033892798237502575, + 0.007355174981057644, + 0.011502847075462341, + 0.02638297714293003, + 0.01067942101508379, + 0.007864855229854584, + 0.013132218271493912, + 0.013913437724113464, + 0.010413059964776039, + 0.00584245752543211, + 0.003260845085605979, + 0.007130119483917952, + 0.017646556720137596, + 0.0058518643490970135, + 0.027460981160402298, + 0.05425845831632614, + 0.007444530259817839, + 0.008771243505179882, + 0.006786905694752932, + 0.04311583191156387, + 0.011725082993507385, + 0.008043170906603336, + 0.03947719931602478, + 0.025377914309501648, + 0.010830226354300976, + 0.002199075650423765, + 0.0027419605758041143, + 0.0026150948833674192, + 0.004673638381063938, + 0.0030616396106779575, + 0.004095995798707008, + 0.005194417200982571, + 0.005739510990679264, + 0.002304880414158106, + 0.00414859177544713, + 0.0033396657090634108, + 0.0057732900604605675, + 0.004013272933661938, + 0.0029277061112225056, + 0.005997566506266594, + 0.00555968564003706, + 0.0013392532709985971, + 0.0022306693717837334, + 0.0014212009264156222, + 0.002250916790217161, + 0.0011717935558408499, + 0.0013316968688741326, + 0.002338386606425047, + 0.0021547365467995405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.0181988924741745 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639100283384323 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639100283384323, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639100283384323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05008593574166298, + 0.03273481875658035, + 0.06993599236011505, + 0.019195634871721268, + 0.18690232932567596, + 0.10660945624113083, + 0.4942450523376465, + 0.01448256615549326, + 0.009202101267874241, + 0.016606057062745094 + ], + "credit_map": [ + 0.006679725833237171, + 0.0029617687687277794, + 0.009740691632032394, + 0.00420737499371171, + 0.003130118828266859, + 0.006251510698348284, + 0.006940133403986692, + 0.010174615308642387, + 0.01096495334059, + 0.0013732207007706165, + 0.0019617807120084763, + 0.0015526479110121727, + 0.0012677066260948777, + 0.01219349354505539, + 0.0013127644779160619, + 0.0021082519087940454, + 0.004274048376828432, + 0.003163723275065422, + 0.004436060786247253, + 0.024291055276989937, + 0.00293157109990716, + 0.017401687800884247, + 0.003453077981248498, + 0.009984767064452171, + 0.003008488565683365, + 0.0018196834716945887, + 0.00365676311776042, + 0.0014880943344905972, + 0.0016284673474729061, + 0.0020623693708330393, + 0.0022249252069741488, + 0.0033068445045500994, + 0.013631561771035194, + 0.015251745469868183, + 0.03986283019185066, + 0.018697910010814667, + 0.016643812879920006, + 0.031911395490169525, + 0.027650952339172363, + 0.02325214073061943, + 0.004308012779802084, + 0.002596096368506551, + 0.005769060458987951, + 0.014563399367034435, + 0.005142006557434797, + 0.02303251251578331, + 0.04536040872335434, + 0.005837968084961176, + 0.028962992131710052, + 0.0180987361818552, + 0.1429516226053238, + 0.03092629462480545, + 0.025051716715097427, + 0.13710366189479828, + 0.0782427042722702, + 0.03290736675262451, + 0.0011770398123189807, + 0.0017968087922781706, + 0.0018572293920442462, + 0.002002628752961755, + 0.0011365487007424235, + 0.0017244709888473153, + 0.002382104517892003, + 0.002405735896900296, + 0.0005907011218369007, + 0.0010791022796183825, + 0.0010501841315999627, + 0.0013667112216353416, + 0.0010099265491589904, + 0.0010489050764590502, + 0.0018028351478278637, + 0.0012537359725683928, + 0.0010524190729483962, + 0.001504950923845172, + 0.0026418743655085564, + 0.0019498976180329919, + 0.0012446523178368807, + 0.0016851581167429686, + 0.002045102883130312, + 0.004482003394514322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.004639100283384323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029800403863191605 + ], + "opposite_logprob_tokens": [ + -3.5298004150390625 + ], + "logprob_reward": -0.029800403863191605, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029800403863191605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14979152381420135, + 0.08187813311815262, + 0.08141510933637619, + 0.046831801533699036, + 0.092160664498806, + 0.2081606239080429, + 0.1772100031375885, + 0.0639883503317833, + 0.04341399297118187, + 0.05514973774552345 + ], + "credit_map": [ + 0.02319176308810711, + 0.009562277235090733, + 0.026710256934165955, + 0.012731722556054592, + 0.008656828664243221, + 0.015875713899731636, + 0.016231192275881767, + 0.03683178126811981, + 0.02442481368780136, + 0.005225055385380983, + 0.00711980601772666, + 0.0048003545962274075, + 0.0035507765132933855, + 0.02788618393242359, + 0.0037510113324970007, + 0.00512014189735055, + 0.011542849242687225, + 0.006758029572665691, + 0.012065826915204525, + 0.009799099527299404, + 0.008077786304056644, + 0.01827438920736313, + 0.005184111651033163, + 0.009713034145534039, + 0.004217057954519987, + 0.005183541215956211, + 0.010999372228980064, + 0.010719473473727703, + 0.0035916671622544527, + 0.002936015371233225, + 0.004733286798000336, + 0.0044513908214867115, + 0.00865214318037033, + 0.009696953929960728, + 0.011461683548986912, + 0.013390226289629936, + 0.008620590902864933, + 0.010421490296721458, + 0.012896123342216015, + 0.017021464183926582, + 0.009317589923739433, + 0.007762085180729628, + 0.011714983731508255, + 0.026461772620677948, + 0.011573391035199165, + 0.0472581647336483, + 0.0793035700917244, + 0.014769098721444607, + 0.00949113443493843, + 0.011010820046067238, + 0.044735949486494064, + 0.012960060499608517, + 0.010985711589455605, + 0.04887889698147774, + 0.024845555424690247, + 0.014301900751888752, + 0.003415505401790142, + 0.005551309324800968, + 0.009775468148291111, + 0.0072763841599226, + 0.004155920352786779, + 0.009568773210048676, + 0.00988332461565733, + 0.01436167024075985, + 0.0037481822073459625, + 0.006069838535040617, + 0.004371451213955879, + 0.005831748712807894, + 0.004992923233658075, + 0.005380562972277403, + 0.0069486903958022594, + 0.006070602685213089, + 0.003410613862797618, + 0.004744275938719511, + 0.010821700096130371, + 0.005912876687943935, + 0.0034764432348310947, + 0.004817531444132328, + 0.007502635009586811, + 0.014463664963841438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.029800403863191605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from a level angle, the camera slowly and steadily moves forward, focusing on a vast wheat field bathed in the golden evening sunlight. A young mother, with long, warm brown hair that flutters gently in the breeze, wears a gentle, contented smile, her eyes filled with maternal love. Before her stands a child, perhaps five or six years old, with short, soft golden hair and cheeks slightly flushed by the sun, looking up at her with her tender face. The mother slowly reaches out her right hand and, with the back of her slender fingers, carefully and lovingly strokes the child's smooth cheek. The child, with eyes slightly closed, savours this moment of tenderness with complete trust. The gentle sunset casts a warm glow on their hair and figures. Behind them, heavy ears of wheat ripple gently in the breeze, enveloping the entire scene in a tranquil, warm, and affectionate atmosphere.", + "seed": 33472, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_5/index_7437_seed_33472_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09498895131982862, + "vlm_logprob": -0.09498895131982862, + "vlm_yes_no_margin": 3.8500000119209288, + "vlm_token_credit": -0.09498895131982862, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029976151883602142 + ], + "opposite_logprob_tokens": [ + -5.6549763679504395 + ], + "logprob_reward": -0.029976151883602142, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.029976151883602142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3744979202747345, + 0.08903106302022934, + 0.06745216250419617, + 0.04033922404050827, + 0.04149702563881874, + 0.06106589362025261, + 0.0630117654800415, + 0.0830007791519165, + 0.05886898562312126, + 0.121235191822052 + ], + "credit_map": [ + 0.04350002855062485, + 0.02455105073750019, + 0.09433203935623169, + 0.019377661868929863, + 0.12881778180599213, + 0.024993641301989555, + 0.025614600628614426, + 0.013311125338077545, + 0.020574551075696945, + 0.008897782303392887, + 0.010032609105110168, + 0.0070922598242759705, + 0.02552480436861515, + 0.005240288563072681, + 0.0062958961352705956, + 0.005372866988182068, + 0.009785212576389313, + 0.009286880493164062, + 0.012333608232438564, + 0.009726266376674175, + 0.006891332101076841, + 0.005829170346260071, + 0.00957780983299017, + 0.004021888133138418, + 0.004326794762164354, + 0.006466229446232319, + 0.006728455889970064, + 0.005009026732295752, + 0.005432231351733208, + 0.004598174709826708, + 0.004753678105771542, + 0.003024633275344968, + 0.007137184031307697, + 0.0042970869690179825, + 0.005908802151679993, + 0.0046211290173232555, + 0.004199739079922438, + 0.006325528025627136, + 0.005033890716731548, + 0.003973666112869978, + 0.006900198291987181, + 0.008647656068205833, + 0.006963726133108139, + 0.011573797091841698, + 0.006610759533941746, + 0.0052349381148815155, + 0.011400161311030388, + 0.0037346507888287306, + 0.005477614235132933, + 0.009739753790199757, + 0.007008231710642576, + 0.010627456940710545, + 0.00713117141276598, + 0.006633384618908167, + 0.012095636688172817, + 0.0042985184118151665, + 0.0074431514367461205, + 0.006984212435781956, + 0.005623590666800737, + 0.013298764824867249, + 0.00931757502257824, + 0.024268651381134987, + 0.009323793463408947, + 0.006741035729646683, + 0.004893882665783167, + 0.007644022349268198, + 0.004731033928692341, + 0.011913364753127098, + 0.005852170754224062, + 0.009418990463018417, + 0.010550402104854584, + 0.003865119768306613, + 0.010051506571471691, + 0.012772290036082268, + 0.009079692885279655, + 0.01731160469353199, + 0.024228285998106003, + 0.009622294455766678, + 0.03147555887699127, + 0.0066939666867256165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.029976151883602142 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38707202672958374 + ], + "opposite_logprob_tokens": [ + -1.1370720863342285 + ], + "logprob_reward": -0.38707202672958374, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38707202672958374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18602254986763, + 0.052375853061676025, + 0.041218120604753494, + 0.041141342371702194, + 0.05069331079721451, + 0.09477799385786057, + 0.4215402603149414, + 0.018598193302750587, + 0.059799354523420334, + 0.03383311629295349 + ], + "credit_map": [ + 0.014768379740417004, + 0.008031515404582024, + 0.0680171325802803, + 0.010233549401164055, + 0.05248630791902542, + 0.012199247255921364, + 0.014305670745670795, + 0.005980729125440121, + 0.007754297461360693, + 0.007831824943423271, + 0.006898138672113419, + 0.004927916452288628, + 0.01127297431230545, + 0.003796942299231887, + 0.006049863062798977, + 0.003843891667202115, + 0.006335136946290731, + 0.006890537217259407, + 0.0056741610169410706, + 0.005575005430728197, + 0.00447005033493042, + 0.00387881463393569, + 0.005957543384283781, + 0.002436866983771324, + 0.005870386026799679, + 0.005886401515454054, + 0.008221568539738655, + 0.003991741221398115, + 0.005642100237309933, + 0.004140871111303568, + 0.004984697792679071, + 0.0024035752285271883, + 0.009797574952244759, + 0.005947716534137726, + 0.005788449663668871, + 0.007203402929008007, + 0.006685527041554451, + 0.005789221730083227, + 0.006442552898079157, + 0.0030388582963496447, + 0.009377453476190567, + 0.010554581880569458, + 0.008365470916032791, + 0.02626432478427887, + 0.009660652838647366, + 0.009452849626541138, + 0.014443500898778439, + 0.006659148260951042, + 0.03138596937060356, + 0.04910651966929436, + 0.04582873731851578, + 0.08633305132389069, + 0.044407203793525696, + 0.039211492985486984, + 0.10388827323913574, + 0.021378973498940468, + 0.0014310143887996674, + 0.0018791316542774439, + 0.0016837762668728828, + 0.002936069155111909, + 0.0018415744416415691, + 0.005046659614890814, + 0.0020822894293814898, + 0.001697676139883697, + 0.00920963380485773, + 0.007747863885015249, + 0.006196919362992048, + 0.011463535949587822, + 0.006934843957424164, + 0.007689458318054676, + 0.006721412762999535, + 0.0038356815930455923, + 0.0035382751375436783, + 0.004131636116653681, + 0.0033350579906255007, + 0.004455076530575752, + 0.005861395504325628, + 0.002706334227696061, + 0.008110917173326015, + 0.0016944243106991053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.38707202672958374 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370685875415802 + ], + "opposite_logprob_tokens": [ + -3.4087069034576416 + ], + "logprob_reward": -0.03370685875415802, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370685875415802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5036319494247437, + 0.10351131111383438, + 0.08490391075611115, + 0.050547752529382706, + 0.03299472853541374, + 0.06042595952749252, + 0.05702739208936691, + 0.043779950588941574, + 0.024408286437392235, + 0.0387687161564827 + ], + "credit_map": [ + 0.1084054484963417, + 0.030056366696953773, + 0.20450322329998016, + 0.026719525456428528, + 0.05017058923840523, + 0.023408684879541397, + 0.042437952011823654, + 0.01793021522462368, + 0.012382425367832184, + 0.012631411664187908, + 0.023690715432167053, + 0.01158235315233469, + 0.018938865512609482, + 0.006475735921412706, + 0.009964732453227043, + 0.00784507766366005, + 0.013537627644836903, + 0.013116003014147282, + 0.014328300021588802, + 0.008789677172899246, + 0.009847945533692837, + 0.010053183883428574, + 0.009491033852100372, + 0.00574014475569129, + 0.0036973096430301666, + 0.006246246863156557, + 0.004960023798048496, + 0.007001889869570732, + 0.005166743416339159, + 0.010529802180826664, + 0.0075349146500229836, + 0.005410822574049234, + 0.0034063614439219236, + 0.003979336470365524, + 0.0026598884724080563, + 0.006173952482640743, + 0.0033718689810484648, + 0.005560584366321564, + 0.004809897858649492, + 0.003032838925719261, + 0.008065410889685154, + 0.008712917566299438, + 0.00744609022513032, + 0.009609023109078407, + 0.008129964582622051, + 0.008083179593086243, + 0.007452473044395447, + 0.0029269049409776926, + 0.004024526569992304, + 0.006709549576044083, + 0.005800735205411911, + 0.01083520520478487, + 0.006311573088169098, + 0.004981013480573893, + 0.01486468780785799, + 0.0035001039505004883, + 0.004254410974681377, + 0.002905446570366621, + 0.0028221230022609234, + 0.004530655685812235, + 0.0048526618629693985, + 0.013101678341627121, + 0.006841771770268679, + 0.004471201915293932, + 0.003209585789591074, + 0.0027324038092046976, + 0.0027019730769097805, + 0.0045734066516160965, + 0.002909081056714058, + 0.0035632660146802664, + 0.002823436865583062, + 0.0018951332895085216, + 0.006691778544336557, + 0.007041935808956623, + 0.005319830495864153, + 0.005735944956541061, + 0.0046058460138738155, + 0.00375162810087204, + 0.003525101812556386, + 0.002096648095175624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03370685875415802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613016800954938 + ], + "opposite_logprob_tokens": [ + -5.628612995147705 + ], + "logprob_reward": -0.003613016800954938, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613016800954938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23441626131534576, + 0.21778452396392822, + 0.0551077239215374, + 0.03973860666155815, + 0.043886810541152954, + 0.032898105680942535, + 0.03860284760594368, + 0.0629516988992691, + 0.046352971345186234, + 0.22826048731803894 + ], + "credit_map": [ + 0.058974362909793854, + 0.015225859358906746, + 0.08875627815723419, + 0.011777886189520359, + 0.01892951875925064, + 0.010367115028202534, + 0.021984588354825974, + 0.008400658145546913, + 0.03880630433559418, + 0.03401185944676399, + 0.042461197823286057, + 0.023126985877752304, + 0.024009929969906807, + 0.015530930832028389, + 0.023760048672556877, + 0.016077259555459023, + 0.008202401921153069, + 0.0076763927936553955, + 0.01072207186371088, + 0.008444925770163536, + 0.005275364499539137, + 0.004186310339719057, + 0.007118000648915768, + 0.003482254222035408, + 0.0033988584764301777, + 0.004481079522520304, + 0.005334403365850449, + 0.008901894092559814, + 0.00491804676130414, + 0.003929848782718182, + 0.006376639008522034, + 0.002397835021838546, + 0.007432971149682999, + 0.00615228246897459, + 0.007612542249262333, + 0.0056314473040401936, + 0.004922304768115282, + 0.004864673595875502, + 0.004049892071634531, + 0.0032206980977207422, + 0.0031289912294596434, + 0.0038763554766774178, + 0.0029953434132039547, + 0.007097958121448755, + 0.0036938937846571207, + 0.0042669749818742275, + 0.004909626208245754, + 0.0029289615340530872, + 0.004531682934612036, + 0.004865798633545637, + 0.0027780141681432724, + 0.005468522198498249, + 0.004540777299553156, + 0.005498702637851238, + 0.007834769785404205, + 0.0030845811124891043, + 0.005258133169263601, + 0.006271593272686005, + 0.005989558063447475, + 0.008839979767799377, + 0.0070198941975831985, + 0.014582324773073196, + 0.009807559661567211, + 0.005182653199881315, + 0.004077170975506306, + 0.004888048395514488, + 0.003635207423940301, + 0.007316939067095518, + 0.0059282719157636166, + 0.011650660075247288, + 0.005220553372055292, + 0.0036361224483698606, + 0.019072799012064934, + 0.0326373353600502, + 0.018383868038654327, + 0.04839964210987091, + 0.03905558958649635, + 0.01704220660030842, + 0.04348915070295334, + 0.010179882869124413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.003613016800954938 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020576702430844307 + ], + "opposite_logprob_tokens": [ + -3.8955767154693604 + ], + "logprob_reward": -0.020576702430844307, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020576702430844307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7300159335136414, + 0.1048748791217804, + 0.02375032939016819, + 0.020368922501802444, + 0.016982194036245346, + 0.012529819272458553, + 0.010066316463053226, + 0.017879148945212364, + 0.01871846430003643, + 0.044813964515924454 + ], + "credit_map": [ + 0.1648675948381424, + 0.036502450704574585, + 0.29845383763313293, + 0.03758998215198517, + 0.07593800127506256, + 0.029696442186832428, + 0.06275428831577301, + 0.024213289842009544, + 0.012289359234273434, + 0.012437629513442516, + 0.02353001944720745, + 0.007318774238228798, + 0.028061307966709137, + 0.006195054855197668, + 0.009798035956919193, + 0.005244691856205463, + 0.0030813964549452066, + 0.003984401002526283, + 0.0033325734548270702, + 0.0030658121686428785, + 0.002346181310713291, + 0.002348258625715971, + 0.003739722305908799, + 0.0018519816221669316, + 0.0018097356660291553, + 0.00271452683955431, + 0.003643527626991272, + 0.003502711420878768, + 0.0024118367582559586, + 0.002556808991357684, + 0.002588086761534214, + 0.0011416856432333589, + 0.0024979133158922195, + 0.0024927498307079077, + 0.0026281909085810184, + 0.002457538153976202, + 0.0017837544437497854, + 0.002266665454953909, + 0.001718588056974113, + 0.0011367901461198926, + 0.0008659114828333259, + 0.0014075242215767503, + 0.001186216832138598, + 0.0036117469426244497, + 0.0012174594448879361, + 0.0013154353946447372, + 0.0018482605228200555, + 0.0010772626847028732, + 0.0012807495659217238, + 0.001177514554001391, + 0.0009454935207031667, + 0.0014766575768589973, + 0.001146607450209558, + 0.001242534490302205, + 0.0019041901687160134, + 0.0008925683214329183, + 0.0021662290673702955, + 0.0020093389321118593, + 0.0016946591204032302, + 0.0024687773548066616, + 0.002061270410194993, + 0.0033419898245483637, + 0.002535135019570589, + 0.0016017466550692916, + 0.0018055784748867154, + 0.0021227377001196146, + 0.001999077619984746, + 0.003677141387015581, + 0.002106564352288842, + 0.003285578452050686, + 0.002404139842838049, + 0.0013176434440538287, + 0.004797447007149458, + 0.0057442509569227695, + 0.005729055963456631, + 0.00719107361510396, + 0.006245267111808062, + 0.004814901854842901, + 0.007042813580483198, + 0.0032491483725607395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.020576702430844307 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from a level angle, the camera slowly and steadily moves forward, focusing on a vast wheat field bathed in the golden evening sunlight. A young mother, with long, warm brown hair that flutters gently in the breeze, wears a gentle, contented smile, her eyes filled with maternal love. Before her stands a child, perhaps five or six years old, with short, soft golden hair and cheeks slightly flushed by the sun, looking up at her with her tender face. The mother slowly reaches out her right hand and, with the back of her slender fingers, carefully and lovingly strokes the child's smooth cheek. The child, with eyes slightly closed, savours this moment of tenderness with complete trust. The gentle sunset casts a warm glow on their hair and figures. Behind them, heavy ears of wheat ripple gently in the breeze, enveloping the entire scene in a tranquil, warm, and affectionate atmosphere.", + "seed": 33473, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_5/index_7437_seed_33473_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07911470741964877, + "vlm_logprob": -0.07911470741964877, + "vlm_yes_no_margin": 3.8249999403953554, + "vlm_token_credit": -0.07911470741964877, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03093533031642437 + ], + "opposite_logprob_tokens": [ + -4.905935287475586 + ], + "logprob_reward": -0.03093533031642437, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.03093533031642437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09858186542987823, + 0.15194733440876007, + 0.21899613738059998, + 0.11556646972894669, + 0.039426688104867935, + 0.048645373433828354, + 0.035389531403779984, + 0.04935187101364136, + 0.16497956216335297, + 0.07711523771286011 + ], + "credit_map": [ + 0.008467826060950756, + 0.006730289198458195, + 0.010361277498304844, + 0.02189485915005207, + 0.007457513827830553, + 0.01964275725185871, + 0.011473197489976883, + 0.012554135173559189, + 0.014329658821225166, + 0.018166104331612587, + 0.01678616926074028, + 0.024733778089284897, + 0.014734849333763123, + 0.015307830646634102, + 0.030678488314151764, + 0.01721044071018696, + 0.02280815690755844, + 0.020722121000289917, + 0.024436820298433304, + 0.04834912344813347, + 0.016061924397945404, + 0.016586313024163246, + 0.05214517191052437, + 0.017886485904455185, + 0.019590189680457115, + 0.015971152111887932, + 0.017251096665859222, + 0.016517676413059235, + 0.010837115347385406, + 0.008172738365828991, + 0.017647074535489082, + 0.00957942008972168, + 0.006142094731330872, + 0.005144881550222635, + 0.004036081489175558, + 0.006475860718637705, + 0.004806892946362495, + 0.002630239585414529, + 0.006642383988946676, + 0.0035482514649629593, + 0.006980428472161293, + 0.004561604931950569, + 0.005008046515285969, + 0.006598223466426134, + 0.009297377429902554, + 0.004456172231584787, + 0.006600800435990095, + 0.005142716690897942, + 0.004392426460981369, + 0.0027344622649252415, + 0.005976449232548475, + 0.00594278983771801, + 0.003113426035270095, + 0.0035983130801469088, + 0.006694041192531586, + 0.002937619574368, + 0.007701284717768431, + 0.004186875186860561, + 0.005515066906809807, + 0.007248336914926767, + 0.004410782363265753, + 0.006762275472283363, + 0.0094283577054739, + 0.004098885226994753, + 0.02014344371855259, + 0.01633593440055847, + 0.030825113877654076, + 0.03346332162618637, + 0.011336257681250572, + 0.011583174578845501, + 0.021477235481142998, + 0.019815051928162575, + 0.011603575199842453, + 0.007200509775429964, + 0.012757926248013973, + 0.014313156716525555, + 0.006056143436580896, + 0.006384409964084625, + 0.010486967861652374, + 0.008312540128827095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.03093533031642437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134477734565735 + ], + "opposite_logprob_tokens": [ + -1.3134477138519287 + ], + "logprob_reward": -0.3134477734565735, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3134477734565735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19361001253128052, + 0.04920418560504913, + 0.28314635157585144, + 0.03714923560619354, + 0.07942838966846466, + 0.0439322330057621, + 0.03996727615594864, + 0.06810671836137772, + 0.17211367189884186, + 0.03334196284413338 + ], + "credit_map": [ + 0.020035451278090477, + 0.019251029938459396, + 0.02805548906326294, + 0.046016618609428406, + 0.024270549416542053, + 0.01773618906736374, + 0.02534947171807289, + 0.012895205989480019, + 0.003952059429138899, + 0.004573160782456398, + 0.007760810665786266, + 0.0069939084351062775, + 0.004510184749960899, + 0.004711393266916275, + 0.009489507414400578, + 0.007213156204670668, + 0.027127614244818687, + 0.02466222271323204, + 0.03213976323604584, + 0.06303606927394867, + 0.01994449459016323, + 0.01754489727318287, + 0.07974930107593536, + 0.0189419724047184, + 0.003305807476863265, + 0.004321717191487551, + 0.0038191070780158043, + 0.0048070065677165985, + 0.0070579200983047485, + 0.0035375587176531553, + 0.005670556332916021, + 0.004629560746252537, + 0.01096542738378048, + 0.009774894453585148, + 0.009969593957066536, + 0.018029537051916122, + 0.0078421076759696, + 0.00482350355014205, + 0.010235695168375969, + 0.007787629030644894, + 0.005704680923372507, + 0.005836217664182186, + 0.003932515159249306, + 0.00527947535738349, + 0.008544380776584148, + 0.0036776612978428602, + 0.00545950373634696, + 0.0054977997206151485, + 0.007282261271029711, + 0.0026026919949799776, + 0.004298120271414518, + 0.007151411846280098, + 0.002663872903212905, + 0.003009721403941512, + 0.010287684388458729, + 0.0026715153362601995, + 0.008958414196968079, + 0.005360580515116453, + 0.006078827660530806, + 0.01093331165611744, + 0.010770405642688274, + 0.008751428686082363, + 0.011566858738660812, + 0.005686893593519926, + 0.018927276134490967, + 0.016246439889073372, + 0.02497905120253563, + 0.03135044500231743, + 0.030620528385043144, + 0.012465900741517544, + 0.02442905865609646, + 0.013094968162477016, + 0.005322466604411602, + 0.0023868130519986153, + 0.006072380114346743, + 0.00590331619605422, + 0.0025393320247530937, + 0.0031688541639596224, + 0.0036172373220324516, + 0.004331563133746386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.3134477734565735 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029812438413500786 + ], + "opposite_logprob_tokens": [ + -3.5298123359680176 + ], + "logprob_reward": -0.029812438413500786, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029812438413500786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1619773805141449, + 0.10149508714675903, + 0.16608808934688568, + 0.11017734557390213, + 0.15512211620807648, + 0.05239008739590645, + 0.05328494310379028, + 0.054160527884960175, + 0.10120309144258499, + 0.044101327657699585 + ], + "credit_map": [ + 0.01781204529106617, + 0.024486815556883812, + 0.022279469296336174, + 0.020956309512257576, + 0.028010603040456772, + 0.015154820866882801, + 0.019661350175738335, + 0.013615965843200684, + 0.007563407067209482, + 0.0064955162815749645, + 0.008803632110357285, + 0.031077958643436432, + 0.009149644523859024, + 0.012377734296023846, + 0.014898871071636677, + 0.011128319427371025, + 0.01791951432824135, + 0.016211872920393944, + 0.020407821983098984, + 0.03300768882036209, + 0.013439280912280083, + 0.012791856192052364, + 0.037215229123830795, + 0.01509481854736805, + 0.006943027023226023, + 0.010126571170985699, + 0.009011296555399895, + 0.014813994988799095, + 0.0236499086022377, + 0.01231404859572649, + 0.017872784286737442, + 0.015445704571902752, + 0.017612816765904427, + 0.01887761615216732, + 0.01959553174674511, + 0.03311607241630554, + 0.02135453186929226, + 0.009023336693644524, + 0.01885533519089222, + 0.016686875373125076, + 0.007359354756772518, + 0.00468062236905098, + 0.00859195925295353, + 0.011178246699273586, + 0.00500050513073802, + 0.0051840986125171185, + 0.004961605649441481, + 0.0054336958564817905, + 0.006680676247924566, + 0.0040428717620670795, + 0.007943028584122658, + 0.009508887305855751, + 0.004772209096699953, + 0.006065643858164549, + 0.010231444612145424, + 0.004040186293423176, + 0.006869692821055651, + 0.005357280373573303, + 0.007256434299051762, + 0.007929334416985512, + 0.010640868917107582, + 0.004887375980615616, + 0.0062308297492563725, + 0.004988707602024078, + 0.01344737596809864, + 0.00902500282973051, + 0.014126607216894627, + 0.015329934656620026, + 0.018240975216031075, + 0.0092494310811162, + 0.01221146434545517, + 0.00957231130450964, + 0.0053921230137348175, + 0.003147758776322007, + 0.006008823402225971, + 0.006193655543029308, + 0.005259597208350897, + 0.005806263070553541, + 0.0072941104881465435, + 0.0049989959225058556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.029812438413500786 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031922117341309786 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.0031922117341309786, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031922117341309786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2852189540863037, + 0.10335713624954224, + 0.08582006394863129, + 0.022353805601596832, + 0.11459407955408096, + 0.019760742783546448, + 0.06539313495159149, + 0.041798047721385956, + 0.21756762266159058, + 0.044136423617601395 + ], + "credit_map": [ + 0.03291008621454239, + 0.053565941751003265, + 0.04780348017811775, + 0.04389219358563423, + 0.022808751091361046, + 0.03098350763320923, + 0.030713606625795364, + 0.02254137583076954, + 0.00819018017500639, + 0.008778647519648075, + 0.011974714696407318, + 0.022295759990811348, + 0.010699016973376274, + 0.012988715432584286, + 0.018149878829717636, + 0.01028022263199091, + 0.007798325270414352, + 0.008709695190191269, + 0.012623054906725883, + 0.019554395228624344, + 0.00500357523560524, + 0.006467566359788179, + 0.020013410598039627, + 0.005650043021887541, + 0.003096554894000292, + 0.0030499822460114956, + 0.00212706602178514, + 0.0031939037144184113, + 0.0024152162950485945, + 0.0022523817606270313, + 0.0032301542814821005, + 0.002988546621054411, + 0.01457185298204422, + 0.016305284574627876, + 0.008880367502570152, + 0.01512070745229721, + 0.027486169710755348, + 0.007254529744386673, + 0.012816842645406723, + 0.012158321216702461, + 0.0022110501304268837, + 0.0017645296175032854, + 0.0021987224463373423, + 0.003924204036593437, + 0.0024962234310805798, + 0.0020229939837008715, + 0.0023366182576864958, + 0.0028064006473869085, + 0.012861112132668495, + 0.006154267583042383, + 0.006875316612422466, + 0.008865551091730595, + 0.005820256192237139, + 0.004629277624189854, + 0.016591623425483704, + 0.003595726564526558, + 0.007271864451467991, + 0.005785795394331217, + 0.0046445103362202644, + 0.004784951452165842, + 0.004438209813088179, + 0.007087794132530689, + 0.003973772283643484, + 0.0038111486937850714, + 0.03800426051020622, + 0.021726321429014206, + 0.032812464982271194, + 0.024837706238031387, + 0.035763781517744064, + 0.02459689788520336, + 0.017527226358652115, + 0.02229895256459713, + 0.005265093874186277, + 0.005861467216163874, + 0.0052198139019310474, + 0.006762874778360128, + 0.005495539866387844, + 0.0034064967185258865, + 0.008126921020448208, + 0.003998220432549715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.0031922117341309786 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185783177614212 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.018185783177614212, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018185783177614212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10208330303430557, + 0.04398086667060852, + 0.11592262238264084, + 0.05914492532610893, + 0.17555959522724152, + 0.024772001430392265, + 0.07119126617908478, + 0.07290427386760712, + 0.2687642276287079, + 0.06567689776420593 + ], + "credit_map": [ + 0.013702384196221828, + 0.009633002802729607, + 0.011376474052667618, + 0.01176177803426981, + 0.013127267360687256, + 0.013279564678668976, + 0.013932272791862488, + 0.015270556323230267, + 0.0038364145439118147, + 0.0036937829572707415, + 0.004625329747796059, + 0.010504014790058136, + 0.0026159428525716066, + 0.005998330190777779, + 0.005023197270929813, + 0.007683850359171629, + 0.011142335832118988, + 0.012493066489696503, + 0.016768142580986023, + 0.024192750453948975, + 0.008675056509673595, + 0.009127800352871418, + 0.023372000083327293, + 0.01015146728605032, + 0.011101565323770046, + 0.007609511259943247, + 0.0058726053684949875, + 0.007781293708831072, + 0.005148188676685095, + 0.005383340176194906, + 0.010209153406322002, + 0.006039265543222427, + 0.02250748686492443, + 0.023627465590834618, + 0.02041724883019924, + 0.03542513772845268, + 0.02612803317606449, + 0.00990821048617363, + 0.02082081139087677, + 0.0167252067476511, + 0.005019767675548792, + 0.002360629616305232, + 0.0033154392149299383, + 0.004370054230093956, + 0.0022426971700042486, + 0.0021675419993698597, + 0.0025545659009367228, + 0.002741307718679309, + 0.015405082143843174, + 0.0037314477376639843, + 0.0073913647793233395, + 0.012046712450683117, + 0.005099847912788391, + 0.0049529410898685455, + 0.018690412864089012, + 0.0038734483532607555, + 0.013119609095156193, + 0.010349790565669537, + 0.00879464391618967, + 0.00798130501061678, + 0.0065440027974545956, + 0.014085368253290653, + 0.007177662570029497, + 0.004851894918829203, + 0.026569802314043045, + 0.023933986201882362, + 0.03690340742468834, + 0.06396716088056564, + 0.04478518292307854, + 0.019401799887418747, + 0.037160445004701614, + 0.016042444854974747, + 0.008070622570812702, + 0.005604412406682968, + 0.011257709935307503, + 0.010344413109123707, + 0.00646371953189373, + 0.0059481095522642136, + 0.01130890753120184, + 0.006679003592580557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 2 + ], + "selected_reward": -0.018185783177614212 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot from a level angle, the camera slowly and steadily moves forward, focusing on a vast wheat field bathed in the golden evening sunlight. A young mother, with long, warm brown hair that flutters gently in the breeze, wears a gentle, contented smile, her eyes filled with maternal love. Before her stands a child, perhaps five or six years old, with short, soft golden hair and cheeks slightly flushed by the sun, looking up at her with her tender face. The mother slowly reaches out her right hand and, with the back of her slender fingers, carefully and lovingly strokes the child's smooth cheek. The child, with eyes slightly closed, savours this moment of tenderness with complete trust. The gentle sunset casts a warm glow on their hair and figures. Behind them, heavy ears of wheat ripple gently in the breeze, enveloping the entire scene in a tranquil, warm, and affectionate atmosphere.", + "seed": 33474, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_5/index_7437_seed_33474_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14284093887545168, + "vlm_logprob": -0.14284093887545168, + "vlm_yes_no_margin": 3.8999999523162843, + "vlm_token_credit": -0.14284093887545168, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04220287501811981 + ], + "opposite_logprob_tokens": [ + -5.542202949523926 + ], + "logprob_reward": -0.04220287501811981, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.04220287501811981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16009539365768433, + 0.029356209561228752, + 0.045859694480895996, + 0.0514632910490036, + 0.33825957775115967, + 0.1465660184621811, + 0.02936054766178131, + 0.06482371687889099, + 0.08273905515670776, + 0.05147651210427284 + ], + "credit_map": [ + 0.021990051493048668, + 0.02719028666615486, + 0.025348816066980362, + 0.02634214423596859, + 0.016336454078555107, + 0.010210837237536907, + 0.017411615699529648, + 0.015265175141394138, + 0.004759624600410461, + 0.003344300203025341, + 0.004645942710340023, + 0.005180978681892157, + 0.0032267405185848475, + 0.0023193401284515858, + 0.0036840669345110655, + 0.002195213455706835, + 0.007541764061897993, + 0.0064172339625656605, + 0.005991658661514521, + 0.005564797203987837, + 0.006278121843934059, + 0.004236527718603611, + 0.006489626131951809, + 0.0033399600069969893, + 0.0094014136120677, + 0.009620696306228638, + 0.005837070755660534, + 0.00471798749640584, + 0.010532055981457233, + 0.003961334470659494, + 0.004332026932388544, + 0.0030607043299824, + 0.027927260845899582, + 0.11792195588350296, + 0.019975371658802032, + 0.12316812574863434, + 0.020575862377882004, + 0.010424630716443062, + 0.008600637316703796, + 0.009665724821388721, + 0.03852398693561554, + 0.020638957619667053, + 0.012873833999037743, + 0.015387127175927162, + 0.03165668249130249, + 0.007889599539339542, + 0.01029449887573719, + 0.009301317855715752, + 0.0057136681862175465, + 0.004483136348426342, + 0.005627486389130354, + 0.0035163455177098513, + 0.003160288091748953, + 0.0018356000073254108, + 0.0035084346309304237, + 0.0015155865112319589, + 0.010181033983826637, + 0.008797407150268555, + 0.014686529524624348, + 0.008492615073919296, + 0.005523862782865763, + 0.004625897388905287, + 0.008266827091574669, + 0.00424954341724515, + 0.013203519396483898, + 0.011212754994630814, + 0.01739649288356304, + 0.011697313748300076, + 0.00819954089820385, + 0.007704618852585554, + 0.00929257832467556, + 0.004032233264297247, + 0.005477204918861389, + 0.007256085518747568, + 0.00833725556731224, + 0.005395827814936638, + 0.007357200141996145, + 0.005054018460214138, + 0.00948875118046999, + 0.0031101664062589407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.04220287501811981 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328050494194031 + ], + "opposite_logprob_tokens": [ + -0.7578050494194031 + ], + "logprob_reward": -0.6328050494194031, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328050494194031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10902038961648941, + 0.0203581340610981, + 0.033167168498039246, + 0.04437887668609619, + 0.4286257326602936, + 0.0412420853972435, + 0.06503383815288544, + 0.030892688781023026, + 0.13343775272369385, + 0.09384334832429886 + ], + "credit_map": [ + 0.02629328891634941, + 0.024404017254710197, + 0.021761149168014526, + 0.008347533643245697, + 0.008095216937363148, + 0.004808844532817602, + 0.010451499372720718, + 0.004858838859945536, + 0.002850426360964775, + 0.002345044631510973, + 0.003876573871821165, + 0.0038584948051720858, + 0.0018818317912518978, + 0.0013182749971747398, + 0.0027699852362275124, + 0.0014575027162209153, + 0.004660970065742731, + 0.004278304986655712, + 0.005811460316181183, + 0.0038826651871204376, + 0.00440288707613945, + 0.0033073911909013987, + 0.003868349362164736, + 0.0029551403131335974, + 0.006400668062269688, + 0.007009285036474466, + 0.0029861503280699253, + 0.0046445936895906925, + 0.009450556710362434, + 0.004321235232055187, + 0.005774344317615032, + 0.0037920435424894094, + 0.034030210226774216, + 0.1500784009695053, + 0.023129049688577652, + 0.16128544509410858, + 0.025084886699914932, + 0.012830585241317749, + 0.010115403681993484, + 0.012071754783391953, + 0.008100093342363834, + 0.007839161902666092, + 0.005790565628558397, + 0.007581620942801237, + 0.005074725486338139, + 0.002691564615815878, + 0.0024058804847300053, + 0.001758473925292492, + 0.009192518889904022, + 0.01811218447983265, + 0.009163115173578262, + 0.012232287786900997, + 0.005116972140967846, + 0.003302806755527854, + 0.004203039687126875, + 0.003710914868861437, + 0.003223731881007552, + 0.003975456580519676, + 0.0037571925204247236, + 0.0046832384541630745, + 0.003911666106432676, + 0.002515320433303714, + 0.005872501991689205, + 0.002953579882159829, + 0.015916617587208748, + 0.02727113477885723, + 0.025400929152965546, + 0.017617063596844673, + 0.014549802988767624, + 0.009212605655193329, + 0.01852824166417122, + 0.004941354505717754, + 0.011133776046335697, + 0.01464204490184784, + 0.012794313952326775, + 0.007271047681570053, + 0.013531710021197796, + 0.00880111288279295, + 0.01932985894382, + 0.0063394820317626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.6328050494194031 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193157389760017 + ], + "opposite_logprob_tokens": [ + -4.018193244934082 + ], + "logprob_reward": -0.018193157389760017, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193157389760017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12299656867980957, + 0.0716489925980568, + 0.05214604362845421, + 0.06440278887748718, + 0.18342481553554535, + 0.05672159418463707, + 0.08263763785362244, + 0.17179852724075317, + 0.1278689205646515, + 0.06635410338640213 + ], + "credit_map": [ + 0.017163552343845367, + 0.01435372605919838, + 0.03064209409058094, + 0.01619565300643444, + 0.013018435798585415, + 0.009730163961648941, + 0.013473519124090672, + 0.008419410325586796, + 0.009750847704708576, + 0.010166926309466362, + 0.00905314739793539, + 0.009115338325500488, + 0.007658876944333315, + 0.00883492361754179, + 0.010384110733866692, + 0.006684810854494572, + 0.009797463193535805, + 0.006735599599778652, + 0.007953125983476639, + 0.0062506734393537045, + 0.005691079888492823, + 0.0054755075834691525, + 0.006209866609424353, + 0.004032724071294069, + 0.014352314174175262, + 0.007530987728387117, + 0.006439251359552145, + 0.00740621704608202, + 0.010230652056634426, + 0.004479137249290943, + 0.008544408716261387, + 0.005419818684458733, + 0.015119784511625767, + 0.05233186483383179, + 0.008967947214841843, + 0.06342563033103943, + 0.02079812064766884, + 0.007759314961731434, + 0.007725999224931002, + 0.007296124007552862, + 0.012457550503313541, + 0.006942831911146641, + 0.010528660379350185, + 0.005088370759040117, + 0.00995780248194933, + 0.004321239423006773, + 0.004004794172942638, + 0.003420339897274971, + 0.015417863614857197, + 0.009716127999126911, + 0.013550604693591595, + 0.011935276910662651, + 0.010487287305295467, + 0.0058318935334682465, + 0.009831313043832779, + 0.00586725864559412, + 0.027721520513296127, + 0.02401013672351837, + 0.019083578139543533, + 0.01871596835553646, + 0.019995082169771194, + 0.016804352402687073, + 0.03086141310632229, + 0.01460646279156208, + 0.012816852889955044, + 0.03015078604221344, + 0.015538795851171017, + 0.011201162822544575, + 0.016666414216160774, + 0.014590528793632984, + 0.019630076363682747, + 0.007274298928678036, + 0.006802764721214771, + 0.01108645275235176, + 0.0067711398005485535, + 0.0060303532518446445, + 0.008262619376182556, + 0.005819173064082861, + 0.017790375277400017, + 0.003791216528043151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.018193157389760017 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028178293723613024 + ], + "opposite_logprob_tokens": [ + -5.877817630767822 + ], + "logprob_reward": -0.0028178293723613024, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028178293723613024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12585967779159546, + 0.1375974416732788, + 0.03680640086531639, + 0.03709279000759125, + 0.2243560254573822, + 0.10446576774120331, + 0.0653204619884491, + 0.061836179345846176, + 0.15019218623638153, + 0.05647307261824608 + ], + "credit_map": [ + 0.01782558299601078, + 0.016268927603960037, + 0.030900876969099045, + 0.01436852291226387, + 0.01552778109908104, + 0.00988359097391367, + 0.01130244042724371, + 0.00978195108473301, + 0.02377750165760517, + 0.019029561430215836, + 0.03185352310538292, + 0.014835904352366924, + 0.01333056390285492, + 0.010363009758293629, + 0.011793971061706543, + 0.012613407336175442, + 0.005647588521242142, + 0.0042304168455302715, + 0.004038945306092501, + 0.004027375020086765, + 0.005204575136303902, + 0.0036493095103651285, + 0.006216596812009811, + 0.003791594645008445, + 0.008905135095119476, + 0.004893855657428503, + 0.007113019935786724, + 0.0036940700374543667, + 0.004602173343300819, + 0.0025524103548377752, + 0.0030666792299598455, + 0.002265445189550519, + 0.019666926935315132, + 0.07646451145410538, + 0.013765670359134674, + 0.08002091199159622, + 0.015159822069108486, + 0.006741654127836227, + 0.006088626570999622, + 0.006447901483625174, + 0.022193245589733124, + 0.01586412638425827, + 0.015093631111085415, + 0.01557320449501276, + 0.016794558614492416, + 0.006987499073147774, + 0.006042013876140118, + 0.005917493253946304, + 0.008057435043156147, + 0.014438011683523655, + 0.009247317910194397, + 0.014459042809903622, + 0.006732452195137739, + 0.0034869208466261625, + 0.005582624115049839, + 0.0033166632056236267, + 0.008959072642028332, + 0.008640795014798641, + 0.006701983977109194, + 0.008911347016692162, + 0.006464826408773661, + 0.00496288575232029, + 0.012940628454089165, + 0.004254638683050871, + 0.018870295956730843, + 0.035638805478811264, + 0.027967236936092377, + 0.017474882304668427, + 0.015358657576143742, + 0.010290643200278282, + 0.016333552077412605, + 0.008258115500211716, + 0.007163556292653084, + 0.012439197860658169, + 0.011532510630786419, + 0.005705921910703182, + 0.004536284599453211, + 0.0037195649929344654, + 0.008595007471740246, + 0.0027810297906398773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.0028178293723613024 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward slowly and steadily?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a young mother with long brown hair stroking a child's cheek with her right hand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a vast wheat field bathed in golden evening sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are heavy ears of wheat visible rippling gently in the breeze behind the subjects?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185783177614212 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.018185783177614212, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018185783177614212, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27375420928001404, + 0.025837447494268417, + 0.029077427461743355, + 0.043935079127550125, + 0.47266334295272827, + 0.022246910259127617, + 0.0273516234010458, + 0.031083736568689346, + 0.03971794992685318, + 0.03433225676417351 + ], + "credit_map": [ + 0.039220161736011505, + 0.04983053728938103, + 0.05012117698788643, + 0.048274893313646317, + 0.02202601172029972, + 0.02194148488342762, + 0.024070002138614655, + 0.01826995611190796, + 0.005061686970293522, + 0.0027484449092298746, + 0.004376486875116825, + 0.004309092648327351, + 0.0029409381095319986, + 0.0021787972655147314, + 0.0025183584075421095, + 0.0017036414938047528, + 0.005210267845541239, + 0.003810248803347349, + 0.0064467755146324635, + 0.00309216370806098, + 0.003441320965066552, + 0.0026055886410176754, + 0.0027470423374325037, + 0.001724019180983305, + 0.007891285233199596, + 0.00571346003562212, + 0.006877923384308815, + 0.0067854225635528564, + 0.003971044439822435, + 0.004074269440025091, + 0.005277790129184723, + 0.0033438808750361204, + 0.0394476018846035, + 0.16202643513679504, + 0.01772119477391243, + 0.18009133636951447, + 0.03062182292342186, + 0.01447470672428608, + 0.013851584866642952, + 0.014428675174713135, + 0.0033209691755473614, + 0.0023749954998493195, + 0.0031695461366325617, + 0.003041486721485853, + 0.003920088522136211, + 0.002202519914135337, + 0.0022374321706593037, + 0.001979873050004244, + 0.0034223792608827353, + 0.002909316448494792, + 0.005900230258703232, + 0.004027528688311577, + 0.0035555579233914614, + 0.0022181603126227856, + 0.003179933177307248, + 0.0021385177969932556, + 0.003853695699945092, + 0.0044257985427975655, + 0.0035129166208207607, + 0.003479840001091361, + 0.006052787881344557, + 0.0031734651420265436, + 0.0037182222586125135, + 0.0028670108877122402, + 0.004825173877179623, + 0.005509559065103531, + 0.008611193858087063, + 0.005328497849404812, + 0.005190230440348387, + 0.003480688901618123, + 0.004569561220705509, + 0.0022030447144061327, + 0.004007828421890736, + 0.008084828965365887, + 0.004294614773243666, + 0.0026850844733417034, + 0.0038810691330581903, + 0.0021676619071513414, + 0.007471349090337753, + 0.0017398198833689094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.018185783177614212 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_6/index_tensor([946, 946])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_6/index_tensor([946, 946])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..47f475952b52b92a7d229304e89e729ddc90401d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_6/index_tensor([946, 946])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight, static close-up focuses on the car's interior rearview mirror, its black plastic casing slightly worn and textured. Reflected in the pristine glass is a perfectly framed, rapidly receding view of a long, straight asphalt highway, its white lane markings rapidly converging to disappear on the horizon. The blurred landscape of trees and rolling hills on either side of the road is rendered in a soft motion blur of deep greens and earthy browns, powerfully conveying the vehicle's high speed. The entire reflected scene is dramatically illuminated by the setting sun, which paints the sky in a stunning gradient of fiery orange, deep burgundy, and soft purple. This golden hour light spills out of the reflection, casting a warm, nostalgic glow around the edges of the frame and creating subtle lens flare on the glass's surface, evoking both a poignant atmosphere of departure and a quiet, hopeful feeling of a journey just beginning.", + "seed": 12650, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_6/index_946_seed_12650_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02120963253546506, + "vlm_logprob": -0.02120963253546506, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.02120963253546506, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010576269589364529 + ], + "opposite_logprob_tokens": [ + -6.260576248168945 + ], + "logprob_reward": -0.010576269589364529, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.010576269589364529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26540759205818176, + 0.09227250516414642, + 0.05844387039542198, + 0.0733015313744545, + 0.061037417501211166, + 0.09959587454795837, + 0.1322994977235794, + 0.07188719511032104, + 0.08636585623025894, + 0.059388697147369385 + ], + "credit_map": [ + 0.02544683776795864, + 0.025119129568338394, + 0.028635689988732338, + 0.05177769437432289, + 0.018199509009718895, + 0.021933862939476967, + 0.05572778731584549, + 0.038567058742046356, + 0.008518684655427933, + 0.008217568509280682, + 0.010605364106595516, + 0.01825319603085518, + 0.008331621997058392, + 0.012029462493956089, + 0.014432922936975956, + 0.011883677914738655, + 0.007997923530638218, + 0.0059973252937197685, + 0.00916231144219637, + 0.00901617482304573, + 0.006580342072993517, + 0.005453656427562237, + 0.008042768575251102, + 0.006193361710757017, + 0.012314954772591591, + 0.008485314436256886, + 0.00882387813180685, + 0.011923464946448803, + 0.008141330443322659, + 0.0069143809378147125, + 0.009137097746133804, + 0.00756111228838563, + 0.007380870170891285, + 0.004828601609915495, + 0.007546666078269482, + 0.009460439905524254, + 0.008316400460898876, + 0.005099287256598473, + 0.010171741247177124, + 0.008233404718339443, + 0.01693287491798401, + 0.00858055055141449, + 0.013835825026035309, + 0.014694180339574814, + 0.008925494737923145, + 0.009736590087413788, + 0.01450306922197342, + 0.012387286871671677, + 0.020511023700237274, + 0.014836207032203674, + 0.01669749803841114, + 0.02144348993897438, + 0.011905185878276825, + 0.014208782464265823, + 0.015417422167956829, + 0.017279881983995438, + 0.012940957210958004, + 0.006500776391476393, + 0.010101460851728916, + 0.012249557301402092, + 0.006910232827067375, + 0.0051171742379665375, + 0.010147840715944767, + 0.007919183000922203, + 0.015611345879733562, + 0.007471576798707247, + 0.011744028888642788, + 0.015564478933811188, + 0.010449039749801159, + 0.006334331817924976, + 0.009454737417399883, + 0.009736314415931702, + 0.0060942163690924644, + 0.003586407983675599, + 0.010127917863428593, + 0.014880383387207985, + 0.003763386746868491, + 0.004037703387439251, + 0.006567609030753374, + 0.01033107005059719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.010576269589364529 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020608583465218544 + ], + "opposite_logprob_tokens": [ + -3.89560866355896 + ], + "logprob_reward": -0.020608583465218544, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020608583465218544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11999602615833282, + 0.07596325874328613, + 0.10767115652561188, + 0.08716020733118057, + 0.16128891706466675, + 0.2670549154281616, + 0.057300541549921036, + 0.03593036159873009, + 0.023787235841155052, + 0.0638473704457283 + ], + "credit_map": [ + 0.016442066058516502, + 0.009933620691299438, + 0.014231075532734394, + 0.020773708820343018, + 0.010057421401143074, + 0.008268449455499649, + 0.024107802659273148, + 0.016181878745555878, + 0.010205374099314213, + 0.006244538351893425, + 0.010295933112502098, + 0.012008391320705414, + 0.00831008329987526, + 0.008192148990929127, + 0.011873428709805012, + 0.008833365514874458, + 0.015983683988451958, + 0.010865282267332077, + 0.015349064022302628, + 0.016567213460803032, + 0.006786656565964222, + 0.009891855530440807, + 0.015729203820228577, + 0.016498196870088577, + 0.015069314278662205, + 0.009356261231005192, + 0.010870097205042839, + 0.014927545562386513, + 0.0064767273142933846, + 0.008365994319319725, + 0.011218825355172157, + 0.010875437408685684, + 0.020863894373178482, + 0.01484723761677742, + 0.02154071256518364, + 0.030435513705015182, + 0.014226897619664669, + 0.011147555895149708, + 0.0270638857036829, + 0.021163223311305046, + 0.05051671341061592, + 0.02754567377269268, + 0.04029800370335579, + 0.04298076406121254, + 0.02017391473054886, + 0.024532154202461243, + 0.03572973608970642, + 0.025277938693761826, + 0.005194862838834524, + 0.005588417872786522, + 0.0068821595050394535, + 0.011716468259692192, + 0.008644603192806244, + 0.005723306443542242, + 0.008239135146141052, + 0.0053115831688046455, + 0.0031782551668584347, + 0.0030524672474712133, + 0.004647970665246248, + 0.005961535032838583, + 0.004626797512173653, + 0.0034454830456525087, + 0.0065869418904185295, + 0.004430912435054779, + 0.004446153994649649, + 0.0021805542055517435, + 0.0033780436497181654, + 0.004102781414985657, + 0.0020830691792070866, + 0.002140789059922099, + 0.0030513075180351734, + 0.002404535422101617, + 0.006795680616050959, + 0.00396349374204874, + 0.009310426190495491, + 0.014048014767467976, + 0.0065653882920742035, + 0.006821615621447563, + 0.008286806754767895, + 0.008055946789681911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.020608583465218544 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206564977765083 + ], + "opposite_logprob_tokens": [ + -2.812065601348877 + ], + "logprob_reward": -0.06206564977765083, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206564977765083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2488982230424881, + 0.13929109275341034, + 0.0752214640378952, + 0.0648331269621849, + 0.14748506247997284, + 0.13341908156871796, + 0.08015550673007965, + 0.03131326287984848, + 0.03081403858959675, + 0.04856908693909645 + ], + "credit_map": [ + 0.03051668405532837, + 0.018298033624887466, + 0.03213365375995636, + 0.04027118906378746, + 0.03096761927008629, + 0.019725043326616287, + 0.04693787917494774, + 0.030048131942749023, + 0.014173682779073715, + 0.010027834214270115, + 0.017833415418863297, + 0.021165812388062477, + 0.01818259432911873, + 0.021976353600621223, + 0.018451640382409096, + 0.017479779198765755, + 0.009995547123253345, + 0.006023034453392029, + 0.011487273499369621, + 0.012025322765111923, + 0.009584372863173485, + 0.005935443099588156, + 0.009517267346382141, + 0.010653214529156685, + 0.01234505046159029, + 0.0064160460606217384, + 0.00900617428123951, + 0.009674482978880405, + 0.006592991761863232, + 0.006617891136556864, + 0.007570656482130289, + 0.0066098375245928764, + 0.014947389252483845, + 0.011790935881435871, + 0.01703241840004921, + 0.02648271434009075, + 0.0156536977738142, + 0.014177659526467323, + 0.027224808931350708, + 0.020175443962216377, + 0.01726461574435234, + 0.013790072873234749, + 0.018682410940527916, + 0.01935204677283764, + 0.018956448882818222, + 0.014457250013947487, + 0.019387170672416687, + 0.011529071256518364, + 0.008505167439579964, + 0.007485006470233202, + 0.011098322458565235, + 0.012518857605755329, + 0.012769129127264023, + 0.008524204604327679, + 0.01172651071101427, + 0.007528319954872131, + 0.0026492471806705, + 0.0021486945915967226, + 0.0041299923323094845, + 0.005847874563187361, + 0.005012317094951868, + 0.002950238296762109, + 0.005721536464989185, + 0.002853365847840905, + 0.004676167853176594, + 0.0026514879427850246, + 0.004252320155501366, + 0.005323226563632488, + 0.004025217145681381, + 0.002870503580197692, + 0.0036560017615556717, + 0.0033591159153729677, + 0.004419584758579731, + 0.002861193846911192, + 0.0076013715006411076, + 0.009443879127502441, + 0.004794764798134565, + 0.0036125837359577417, + 0.007670195773243904, + 0.008165518753230572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.06206564977765083 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011082993820309639 + ], + "opposite_logprob_tokens": [ + -4.511083126068115 + ], + "logprob_reward": -0.011082993820309639, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011082993820309639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21396546065807343, + 0.11896016448736191, + 0.04717177897691727, + 0.11611483246088028, + 0.07438147813081741, + 0.17096349596977234, + 0.1336272656917572, + 0.06727699190378189, + 0.03182390332221985, + 0.02571459673345089 + ], + "credit_map": [ + 0.024385463446378708, + 0.016706719994544983, + 0.023870185017585754, + 0.05312507599592209, + 0.018215933814644814, + 0.016121715307235718, + 0.036412954330444336, + 0.025127390399575233, + 0.009371718391776085, + 0.009521552361547947, + 0.014517218805849552, + 0.02479645609855652, + 0.018547069281339645, + 0.012033791281282902, + 0.018141303211450577, + 0.012031054124236107, + 0.006343398243188858, + 0.004327597562223673, + 0.0065456414595246315, + 0.008123612031340599, + 0.004430024418979883, + 0.004600541200488806, + 0.006701806094497442, + 0.006099160294979811, + 0.013551400043070316, + 0.00994656141847372, + 0.014591355808079243, + 0.021375097334384918, + 0.007935180328786373, + 0.015777401626110077, + 0.01602417416870594, + 0.01691366918385029, + 0.008664820343255997, + 0.004338070750236511, + 0.009517577476799488, + 0.010802563279867172, + 0.007949316874146461, + 0.00973350927233696, + 0.01410658098757267, + 0.009269039146602154, + 0.015829768031835556, + 0.01842029206454754, + 0.023472318425774574, + 0.02958499640226364, + 0.015691719949245453, + 0.02097150683403015, + 0.02306649275124073, + 0.023926394060254097, + 0.010304889641702175, + 0.012455988675355911, + 0.01779637672007084, + 0.02312687411904335, + 0.02381967008113861, + 0.016333110630512238, + 0.01907646469771862, + 0.010713889263570309, + 0.006124650593847036, + 0.0054493010975420475, + 0.00932994019240141, + 0.010984313674271107, + 0.00846048817038536, + 0.006199598778039217, + 0.012290120124816895, + 0.00843858066946268, + 0.00432334141805768, + 0.0031278948299586773, + 0.004853179212659597, + 0.005641489755362272, + 0.0021792687475681305, + 0.002726010512560606, + 0.00443294458091259, + 0.004539774265140295, + 0.0033735441975295544, + 0.0018515520496293902, + 0.004789475351572037, + 0.0051262336783111095, + 0.0015923723112791777, + 0.001420237123966217, + 0.004155319184064865, + 0.0034058629535138607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.011082993820309639 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017146660247817636 + ], + "opposite_logprob_tokens": [ + -6.376714706420898 + ], + "logprob_reward": -0.0017146660247817636, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017146660247817636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20429262518882751, + 0.1232551708817482, + 0.11630852520465851, + 0.06233619898557663, + 0.09824252128601074, + 0.09344857186079025, + 0.11524098366498947, + 0.05188523605465889, + 0.060472723096609116, + 0.07451742887496948 + ], + "credit_map": [ + 0.025201786309480667, + 0.013222713023424149, + 0.028650855645537376, + 0.031680990010499954, + 0.026067497208714485, + 0.024493345990777016, + 0.03579727187752724, + 0.019178183749318123, + 0.013209543190896511, + 0.008664466440677643, + 0.014789172448217869, + 0.0197039395570755, + 0.014086565002799034, + 0.018877387046813965, + 0.017734840512275696, + 0.016189252957701683, + 0.012278571724891663, + 0.008035324513912201, + 0.014123760163784027, + 0.019968107342720032, + 0.013394569046795368, + 0.01745540276169777, + 0.018037868663668633, + 0.013014920055866241, + 0.0055669210851192474, + 0.0036607887595891953, + 0.008329909294843674, + 0.01634821482002735, + 0.004932067357003689, + 0.008383277803659439, + 0.00754838390275836, + 0.007566637825220823, + 0.010100767016410828, + 0.008109175600111485, + 0.011696665547788143, + 0.018493831157684326, + 0.009128664620220661, + 0.010228377766907215, + 0.015593115240335464, + 0.014891921542584896, + 0.012687773443758488, + 0.008093674667179585, + 0.01184091530740261, + 0.014656273648142815, + 0.012308874167501926, + 0.010516682639718056, + 0.013158750720322132, + 0.010185622610151768, + 0.01155915018171072, + 0.00994555838406086, + 0.01598457247018814, + 0.019018011167645454, + 0.020185839384794235, + 0.013707175850868225, + 0.015284465625882149, + 0.009556209668517113, + 0.0057001360692083836, + 0.004177723545581102, + 0.006001666653901339, + 0.007319796830415726, + 0.007881006225943565, + 0.008823554031550884, + 0.006655022501945496, + 0.00532632926478982, + 0.00876509491354227, + 0.00403252337127924, + 0.008999797515571117, + 0.012730507180094719, + 0.00541135435923934, + 0.005507734138518572, + 0.008804087527096272, + 0.006221624091267586, + 0.006562861613929272, + 0.00529978983104229, + 0.011511337012052536, + 0.011620688252151012, + 0.008922294713556767, + 0.006938260514289141, + 0.012038104236125946, + 0.011624098755419254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0017146660247817636 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, static close-up focuses on the car's interior rearview mirror, its black plastic casing slightly worn and textured. Reflected in the pristine glass is a perfectly framed, rapidly receding view of a long, straight asphalt highway, its white lane markings rapidly converging to disappear on the horizon. The blurred landscape of trees and rolling hills on either side of the road is rendered in a soft motion blur of deep greens and earthy browns, powerfully conveying the vehicle's high speed. The entire reflected scene is dramatically illuminated by the setting sun, which paints the sky in a stunning gradient of fiery orange, deep burgundy, and soft purple. This golden hour light spills out of the reflection, casting a warm, nostalgic glow around the edges of the frame and creating subtle lens flare on the glass's surface, evoking both a poignant atmosphere of departure and a quiet, hopeful feeling of a journey just beginning.", + "seed": 12651, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_6/index_946_seed_12651_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02600622484460473, + "vlm_logprob": -0.02600622484460473, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.02600622484460473, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021842744201421738 + ], + "opposite_logprob_tokens": [ + -4.271842956542969 + ], + "logprob_reward": -0.021842744201421738, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.021842744201421738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2299298197031021, + 0.0782591700553894, + 0.051099881529808044, + 0.054795943200588226, + 0.05777483806014061, + 0.12199127674102783, + 0.2873779535293579, + 0.05193975567817688, + 0.02247787080705166, + 0.04435353726148605 + ], + "credit_map": [ + 0.022413035854697227, + 0.017133494839072227, + 0.020276367664337158, + 0.02054702863097191, + 0.02148772031068802, + 0.043391916900873184, + 0.03448612987995148, + 0.05019412934780121, + 0.005210236646234989, + 0.004670047666877508, + 0.012256923131644726, + 0.008122642524540424, + 0.008836060762405396, + 0.01315495278686285, + 0.013839361257851124, + 0.012168942019343376, + 0.0044410922564566135, + 0.003589973086491227, + 0.006636559963226318, + 0.004524767864495516, + 0.004869889002293348, + 0.011327925138175488, + 0.007475836668163538, + 0.008233838714659214, + 0.0046662273816764355, + 0.005363922566175461, + 0.006254741456359625, + 0.00648869201540947, + 0.007719486020505428, + 0.007901663891971111, + 0.00804674718528986, + 0.008354462683200836, + 0.0050118910148739815, + 0.0031478179153054953, + 0.007914228364825249, + 0.005524990148842335, + 0.0040992675349116325, + 0.012243112549185753, + 0.009652059525251389, + 0.010181471705436707, + 0.01086850930005312, + 0.00923220720142126, + 0.014976613223552704, + 0.01422994676977396, + 0.01212043222039938, + 0.022314945235848427, + 0.017973776906728745, + 0.020274845883250237, + 0.020748386159539223, + 0.022032493725419044, + 0.04036519676446915, + 0.03049948811531067, + 0.026273522526025772, + 0.045111652463674545, + 0.04878741130232811, + 0.0535597950220108, + 0.003871256485581398, + 0.0035837609320878983, + 0.008398572914302349, + 0.007633568719029427, + 0.004560887813568115, + 0.00875174906104803, + 0.008372848853468895, + 0.006767112296074629, + 0.0018726675771176815, + 0.0011264384957030416, + 0.003840372897684574, + 0.004040743224322796, + 0.0013706039171665907, + 0.003747364738956094, + 0.0026232372038066387, + 0.003856442403048277, + 0.0032065492123365402, + 0.00217931205406785, + 0.006388239562511444, + 0.004469794686883688, + 0.002700811019167304, + 0.00894321221858263, + 0.006457787938416004, + 0.010007829405367374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.021842744201421738 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03815697133541107 + ], + "opposite_logprob_tokens": [ + -3.2881569862365723 + ], + "logprob_reward": -0.03815697133541107, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03815697133541107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3443906903266907, + 0.03283586725592613, + 0.04431697726249695, + 0.08146411925554276, + 0.12291458249092102, + 0.09112467616796494, + 0.07180433720350266, + 0.062332529574632645, + 0.029913829639554024, + 0.11890225112438202 + ], + "credit_map": [ + 0.03433026373386383, + 0.023177651688456535, + 0.0284194927662611, + 0.030643785372376442, + 0.02944149449467659, + 0.06708794832229614, + 0.052087876945734024, + 0.07920225709676743, + 0.0029007112607359886, + 0.001984470756724477, + 0.0036316041368991137, + 0.003068284597247839, + 0.003982992377132177, + 0.0073969862423837185, + 0.0034511766862124205, + 0.006419647950679064, + 0.004137336276471615, + 0.0025526208337396383, + 0.00613422179594636, + 0.004454130306839943, + 0.002776581794023514, + 0.010083065368235111, + 0.006625927984714508, + 0.007553100120276213, + 0.006551898084580898, + 0.005999211221933365, + 0.01149839535355568, + 0.009237733669579029, + 0.007047405932098627, + 0.01622728817164898, + 0.01177926268428564, + 0.013122938573360443, + 0.008501159958541393, + 0.00927077978849411, + 0.018561208620667458, + 0.011819557286798954, + 0.01540343463420868, + 0.016841022297739983, + 0.022688351571559906, + 0.01982908323407173, + 0.009817923419177532, + 0.005933744832873344, + 0.010086501017212868, + 0.009320897050201893, + 0.0064765652641654015, + 0.02052793651819229, + 0.012487102299928665, + 0.01647401601076126, + 0.0070665148086845875, + 0.006165789440274239, + 0.012206644751131535, + 0.007720272988080978, + 0.006977722514420748, + 0.010428747162222862, + 0.010453544557094574, + 0.0107851168140769, + 0.004649735055863857, + 0.0035075191408395767, + 0.005416822154074907, + 0.006732087582349777, + 0.00463896943256259, + 0.010661031119525433, + 0.014926568605005741, + 0.011799810454249382, + 0.0020734125282615423, + 0.0019138487987220287, + 0.0038432993460446596, + 0.004265764728188515, + 0.0017367559485137463, + 0.006079371552914381, + 0.0030245163943618536, + 0.006976868025958538, + 0.008438116870820522, + 0.005574778188019991, + 0.015383526682853699, + 0.010389580391347408, + 0.007695627398788929, + 0.02453998103737831, + 0.02001139149069786, + 0.026869269087910652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.03815697133541107 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498780682682991 + ], + "opposite_logprob_tokens": [ + -2.929987907409668 + ], + "logprob_reward": -0.05498780682682991, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498780682682991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10789096355438232, + 0.13396170735359192, + 0.0888373926281929, + 0.24835632741451263, + 0.14240573346614838, + 0.0665903389453888, + 0.08204042911529541, + 0.03571479395031929, + 0.02969995327293873, + 0.06450232863426208 + ], + "credit_map": [ + 0.009461216628551483, + 0.007170713506639004, + 0.01764455996453762, + 0.010177779011428356, + 0.01103234477341175, + 0.018361283466219902, + 0.012407067231833935, + 0.021636003628373146, + 0.008472693152725697, + 0.01422701682895422, + 0.017769282683730125, + 0.01802934892475605, + 0.01414460875093937, + 0.022575054317712784, + 0.019306538626551628, + 0.01943717896938324, + 0.00689789978787303, + 0.005288579035550356, + 0.01235687080770731, + 0.009963056072592735, + 0.008633232675492764, + 0.016730526462197304, + 0.015590354800224304, + 0.01337688323110342, + 0.01954684406518936, + 0.01783783733844757, + 0.03199612721800804, + 0.027959156781435013, + 0.021296845749020576, + 0.05206764116883278, + 0.03768465667963028, + 0.03996724262833595, + 0.010705183260142803, + 0.009693380445241928, + 0.022485392168164253, + 0.015383089892566204, + 0.01572093553841114, + 0.023041142150759697, + 0.019898217171430588, + 0.025478407740592957, + 0.00765240378677845, + 0.004325543064624071, + 0.00818957295268774, + 0.005990273784846067, + 0.005820092745125294, + 0.011855868622660637, + 0.010628881864249706, + 0.012127711437642574, + 0.006113030482083559, + 0.0064721968956291676, + 0.010557308793067932, + 0.008813278749585152, + 0.007975875400006771, + 0.01363595575094223, + 0.014184265397489071, + 0.014288526959717274, + 0.002494982210919261, + 0.002171241445466876, + 0.004479324445128441, + 0.003861697856336832, + 0.0029591082129627466, + 0.006608054507523775, + 0.0064969793893396854, + 0.006643409840762615, + 0.0017450818559154868, + 0.002053669421002269, + 0.004350415896624327, + 0.005229891277849674, + 0.0020663095638155937, + 0.005129469092935324, + 0.0036389438901096582, + 0.00548617634922266, + 0.0040407804772257805, + 0.0030566600617021322, + 0.006265490315854549, + 0.01130849402397871, + 0.004035288468003273, + 0.012782289646565914, + 0.009836641140282154, + 0.0131766926497221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.05498780682682991 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257826764136553 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.005257826764136553, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257826764136553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17892111837863922, + 0.07156895101070404, + 0.06474626064300537, + 0.17668122053146362, + 0.07702688872814178, + 0.1694021373987198, + 0.08826906979084015, + 0.04196171090006828, + 0.04069395735859871, + 0.09072870761156082 + ], + "credit_map": [ + 0.02393098548054695, + 0.014174917712807655, + 0.021447861567139626, + 0.016938691958785057, + 0.018103819340467453, + 0.029020361602306366, + 0.023011954501271248, + 0.03229251503944397, + 0.0048566535115242004, + 0.005754060577601194, + 0.011516690254211426, + 0.008975100703537464, + 0.006418429780751467, + 0.011916873045265675, + 0.010577154345810413, + 0.011553981341421604, + 0.0054216389544308186, + 0.005546353757381439, + 0.010980179533362389, + 0.005498063284903765, + 0.005134278442710638, + 0.014160563237965107, + 0.007844009436666965, + 0.010161174461245537, + 0.013822279870510101, + 0.014692495577037334, + 0.021065060049295425, + 0.016534732654690742, + 0.017299506813287735, + 0.038788873702287674, + 0.02737423963844776, + 0.027104033157229424, + 0.006171018350869417, + 0.0069948467426002026, + 0.012872427701950073, + 0.007283157203346491, + 0.0078356321901083, + 0.013571502640843391, + 0.009685435332357883, + 0.012612873688340187, + 0.020218364894390106, + 0.011337285861372948, + 0.016937412321567535, + 0.017257878556847572, + 0.010422509163618088, + 0.03413334861397743, + 0.03199301287531853, + 0.027102330699563026, + 0.008692319504916668, + 0.0074169631116092205, + 0.01166804600507021, + 0.0083263348788023, + 0.010461908765137196, + 0.014708474278450012, + 0.016751715913414955, + 0.010243306867778301, + 0.0033585703931748867, + 0.002480670576915145, + 0.006105226930230856, + 0.006345795001834631, + 0.002472942229360342, + 0.008163764141499996, + 0.0048421951942145824, + 0.008192545734345913, + 0.0027547520585358143, + 0.0026988075114786625, + 0.005667299963533878, + 0.004831062629818916, + 0.002750016050413251, + 0.008082636632025242, + 0.0046255928464233875, + 0.009283790364861488, + 0.006645824760198593, + 0.004517371766269207, + 0.014080507680773735, + 0.01704058051109314, + 0.0065000602044165134, + 0.01547397393733263, + 0.010940495878458023, + 0.015529890544712543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.005257826764136553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978577509522438 + ], + "opposite_logprob_tokens": [ + -4.6347856521606445 + ], + "logprob_reward": -0.00978577509522438, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978577509522438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2790259122848511, + 0.12438180297613144, + 0.11532977223396301, + 0.055702220648527145, + 0.11483607441186905, + 0.05885287746787071, + 0.06154526397585869, + 0.0628165677189827, + 0.037383511662483215, + 0.09012602269649506 + ], + "credit_map": [ + 0.025374989956617355, + 0.019317008554935455, + 0.03598252683877945, + 0.027189988642930984, + 0.02831588126718998, + 0.05044160783290863, + 0.03737952932715416, + 0.05502437427639961, + 0.008466781117022038, + 0.008136174641549587, + 0.018519623205065727, + 0.014573819935321808, + 0.01448859553784132, + 0.02111820876598358, + 0.0199736338108778, + 0.019104963168501854, + 0.00765224639326334, + 0.00790395587682724, + 0.015907367691397667, + 0.010711520910263062, + 0.01274495292454958, + 0.02009112946689129, + 0.021481169387698174, + 0.018837429583072662, + 0.0039064413867890835, + 0.003167543327435851, + 0.010292361490428448, + 0.008220956660807133, + 0.004202244803309441, + 0.009353084489703178, + 0.007826426066458225, + 0.008733163587749004, + 0.008579100482165813, + 0.009473700076341629, + 0.016844376921653748, + 0.009996965527534485, + 0.013810444623231888, + 0.017467809841036797, + 0.019834857434034348, + 0.01882881671190262, + 0.004282705951482058, + 0.0050549935549497604, + 0.005292660091072321, + 0.00691172806546092, + 0.003964098170399666, + 0.011315734125673771, + 0.011569692753255367, + 0.010461260564625263, + 0.006219041999429464, + 0.004304294008761644, + 0.007661828771233559, + 0.009009179659187794, + 0.005697284359484911, + 0.008874175138771534, + 0.009033046662807465, + 0.010746413841843605, + 0.00435038423165679, + 0.0036562501918524504, + 0.0075144898146390915, + 0.005952700972557068, + 0.004743070341646671, + 0.011275012046098709, + 0.015148981474339962, + 0.010175679810345173, + 0.002760283648967743, + 0.002307602670043707, + 0.003981047309935093, + 0.004856737330555916, + 0.002692439826205373, + 0.007153613492846489, + 0.005624203011393547, + 0.00800758320838213, + 0.0069082751870155334, + 0.004956538323312998, + 0.013938051648437977, + 0.011580741964280605, + 0.006888785865157843, + 0.016360269859433174, + 0.01395785715430975, + 0.015535497106611729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00978577509522438 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, static close-up focuses on the car's interior rearview mirror, its black plastic casing slightly worn and textured. Reflected in the pristine glass is a perfectly framed, rapidly receding view of a long, straight asphalt highway, its white lane markings rapidly converging to disappear on the horizon. The blurred landscape of trees and rolling hills on either side of the road is rendered in a soft motion blur of deep greens and earthy browns, powerfully conveying the vehicle's high speed. The entire reflected scene is dramatically illuminated by the setting sun, which paints the sky in a stunning gradient of fiery orange, deep burgundy, and soft purple. This golden hour light spills out of the reflection, casting a warm, nostalgic glow around the edges of the frame and creating subtle lens flare on the glass's surface, evoking both a poignant atmosphere of departure and a quiet, hopeful feeling of a journey just beginning.", + "seed": 12652, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_6/index_946_seed_12652_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010605190205387771, + "vlm_logprob": -0.010605190205387771, + "vlm_yes_no_margin": 5.1, + "vlm_token_credit": -0.010605190205387771, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008257769979536533 + ], + "opposite_logprob_tokens": [ + -6.508257865905762 + ], + "logprob_reward": -0.008257769979536533, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.008257769979536533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1265682429075241, + 0.11153232306241989, + 0.17351137101650238, + 0.25831910967826843, + 0.032972823828458786, + 0.08086903393268585, + 0.04186895862221718, + 0.06140194460749626, + 0.043245524168014526, + 0.06971067935228348 + ], + "credit_map": [ + 0.01412586122751236, + 0.00966875534504652, + 0.015796924009919167, + 0.018348202109336853, + 0.010687571950256824, + 0.01665772683918476, + 0.012362473644316196, + 0.028920715674757957, + 0.013238279148936272, + 0.013652564957737923, + 0.01302097924053669, + 0.016485948115587234, + 0.008881671354174614, + 0.015489434823393822, + 0.014345726929605007, + 0.016417725011706352, + 0.017917321994900703, + 0.01357173640280962, + 0.020801501348614693, + 0.021848348900675774, + 0.01385426800698042, + 0.021692240610718727, + 0.028240866959095, + 0.035585083067417145, + 0.023446913808584213, + 0.023765459656715393, + 0.03316735103726387, + 0.040180616080760956, + 0.023848844692111015, + 0.023224906995892525, + 0.051077187061309814, + 0.039607830345630646, + 0.005101148039102554, + 0.003619862487539649, + 0.004974445328116417, + 0.0038255457766354084, + 0.003178395563736558, + 0.004298350773751736, + 0.0032419925555586815, + 0.004733082372695208, + 0.010968632064759731, + 0.00664407666772604, + 0.009902216494083405, + 0.010290618985891342, + 0.0066648186184465885, + 0.008538292720913887, + 0.012661740183830261, + 0.01519863959401846, + 0.006139167584478855, + 0.005708080250769854, + 0.0043791853822767735, + 0.004227584693580866, + 0.003738699248060584, + 0.008540764451026917, + 0.003437178675085306, + 0.005698299966752529, + 0.006984637584537268, + 0.007247368339449167, + 0.008510754443705082, + 0.008326531387865543, + 0.006415289826691151, + 0.007438566070050001, + 0.0077766659669578075, + 0.008702130988240242, + 0.004441691096872091, + 0.004078672267496586, + 0.005812515504658222, + 0.006888830568641424, + 0.003745444817468524, + 0.00654870318248868, + 0.0054802317172288895, + 0.0062494357116520405, + 0.010160868987441063, + 0.008074486628174782, + 0.008482315577566624, + 0.008700387552380562, + 0.005575807765126228, + 0.010898870415985584, + 0.007822306826710701, + 0.009995630010962486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.008257769979536533 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016093354672193527 + ], + "opposite_logprob_tokens": [ + -4.1410932540893555 + ], + "logprob_reward": -0.016093354672193527, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016093354672193527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30064597725868225, + 0.05554046481847763, + 0.1760222464799881, + 0.04822046682238579, + 0.20600783824920654, + 0.05062097683548927, + 0.05788913741707802, + 0.022337159141898155, + 0.059523142874240875, + 0.023192565888166428 + ], + "credit_map": [ + 0.031184496358036995, + 0.026002168655395508, + 0.034294337034225464, + 0.05551527440547943, + 0.029956040903925896, + 0.04121113568544388, + 0.0409831628203392, + 0.041499387472867966, + 0.0070265354588627815, + 0.00732865184545517, + 0.006860268767923117, + 0.006398622877895832, + 0.004642734304070473, + 0.010309130884706974, + 0.006669078022241592, + 0.006305452901870012, + 0.0174059197306633, + 0.012714766897261143, + 0.02084120735526085, + 0.02905573509633541, + 0.01686285063624382, + 0.023262565955519676, + 0.026541002094745636, + 0.029338210821151733, + 0.0049446201883256435, + 0.0049355290830135345, + 0.004224810749292374, + 0.008613010868430138, + 0.004326497670263052, + 0.00639000628143549, + 0.00713001936674118, + 0.007655977737158537, + 0.02449633553624153, + 0.013485822826623917, + 0.02478756569325924, + 0.02955980971455574, + 0.02258511818945408, + 0.024878133088350296, + 0.019078580662608147, + 0.04713648185133934, + 0.0072488971054553986, + 0.004394248593598604, + 0.006245104596018791, + 0.006014422979205847, + 0.004009375348687172, + 0.007571190595626831, + 0.006436583120375872, + 0.008701160550117493, + 0.007745878770947456, + 0.006718247197568417, + 0.0053742919117212296, + 0.006862522568553686, + 0.005157472100108862, + 0.009968012571334839, + 0.007093001622706652, + 0.008969717659056187, + 0.002773764543235302, + 0.001935067935846746, + 0.003056626534089446, + 0.0034304801374673843, + 0.0023156802635639906, + 0.003212948329746723, + 0.003064124146476388, + 0.002548469230532646, + 0.0065254042856395245, + 0.005389621946960688, + 0.006040043663233519, + 0.008178692311048508, + 0.005358237307518721, + 0.006534504238516092, + 0.011002955958247185, + 0.010493692941963673, + 0.002803841605782509, + 0.002145978156477213, + 0.0030585667118430138, + 0.003601408563554287, + 0.0015518894651904702, + 0.0030449843034148216, + 0.0033081406727433205, + 0.0036777593195438385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.016093354672193527 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014196592383086681 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.014196592383086681, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014196592383086681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16837729513645172, + 0.21554909646511078, + 0.0987558588385582, + 0.07852252572774887, + 0.07765787839889526, + 0.027120625600218773, + 0.04907837510108948, + 0.03989160805940628, + 0.22840851545333862, + 0.01663816347718239 + ], + "credit_map": [ + 0.018404189497232437, + 0.013680645264685154, + 0.0216256994754076, + 0.03285030275583267, + 0.011403768323361874, + 0.01625101827085018, + 0.028996672481298447, + 0.025164997205138206, + 0.02846655808389187, + 0.011401084251701832, + 0.027775131165981293, + 0.03305421769618988, + 0.01356518268585205, + 0.03027130290865898, + 0.030806250870227814, + 0.04020937904715538, + 0.010931811295449734, + 0.007660288829356432, + 0.013341628946363926, + 0.013151644729077816, + 0.008863940834999084, + 0.013843036256730556, + 0.013715783134102821, + 0.017247721552848816, + 0.009854018688201904, + 0.007154393941164017, + 0.008636955171823502, + 0.010953997261822224, + 0.0073102363385260105, + 0.011877547949552536, + 0.013239005580544472, + 0.009496370330452919, + 0.010059656575322151, + 0.005287511739879847, + 0.008661708794534206, + 0.008954166434705257, + 0.008033566176891327, + 0.009004088118672371, + 0.009669122286140919, + 0.017988063395023346, + 0.0037101355846971273, + 0.002848479663953185, + 0.003622167743742466, + 0.0026675451081246138, + 0.002590383403003216, + 0.004345695022493601, + 0.0032468275167047977, + 0.004089390393346548, + 0.005886787082999945, + 0.005648545455187559, + 0.005777490325272083, + 0.005344319622963667, + 0.005500985309481621, + 0.009435711428523064, + 0.00446449127048254, + 0.007020042277872562, + 0.0050751264207065105, + 0.0035680150613188744, + 0.004065959248691797, + 0.0050216419622302055, + 0.0034151382278651, + 0.004462162498384714, + 0.005523648113012314, + 0.008759916760027409, + 0.0206330306828022, + 0.01807049848139286, + 0.029573792591691017, + 0.03483743593096733, + 0.017305778339505196, + 0.02715338207781315, + 0.04522440582513809, + 0.03561020269989967, + 0.0020572037901729345, + 0.0021485795732587576, + 0.0017951828194782138, + 0.002197312656790018, + 0.0014712322736158967, + 0.002030479721724987, + 0.0026378717739135027, + 0.0023003006353974342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.014196592383086681 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019418209558352828 + ], + "opposite_logprob_tokens": [ + -6.251941680908203 + ], + "logprob_reward": -0.0019418209558352828, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019418209558352828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13740494847297668, + 0.14346687495708466, + 0.17657232284545898, + 0.09383510798215866, + 0.06699366122484207, + 0.05503106862306595, + 0.08010733872652054, + 0.05098225921392441, + 0.18257667124271393, + 0.013029765337705612 + ], + "credit_map": [ + 0.018371442332863808, + 0.009559640660881996, + 0.02018502540886402, + 0.026051441207528114, + 0.009120765142142773, + 0.012183008715510368, + 0.017433907836675644, + 0.024499719962477684, + 0.018027596175670624, + 0.016154572367668152, + 0.017159583047032356, + 0.01844025030732155, + 0.011577512137591839, + 0.01941237971186638, + 0.017608780413866043, + 0.02508619986474514, + 0.020627904683351517, + 0.013844321481883526, + 0.021594371646642685, + 0.022089501842856407, + 0.013020404614508152, + 0.025319090113043785, + 0.025347400456666946, + 0.03472931683063507, + 0.010667425580322742, + 0.005768982693552971, + 0.010148911736905575, + 0.012448384426534176, + 0.007622180040925741, + 0.01926203817129135, + 0.013057352975010872, + 0.014859835617244244, + 0.005830332171171904, + 0.0037128624971956015, + 0.00984124094247818, + 0.005352707579731941, + 0.006196770817041397, + 0.009587886743247509, + 0.006474032532423735, + 0.01999782770872116, + 0.006767554208636284, + 0.005353710148483515, + 0.005809524562209845, + 0.008670986630022526, + 0.00453934958204627, + 0.011519620195031166, + 0.005857320036739111, + 0.006513004191219807, + 0.009020606055855751, + 0.00801265612244606, + 0.0101711954921484, + 0.009551496244966984, + 0.008380758576095104, + 0.014091485179960728, + 0.008616369217634201, + 0.012262765318155289, + 0.00619898596778512, + 0.006479077506810427, + 0.005961390677839518, + 0.005922718904912472, + 0.005946370307356119, + 0.007191771641373634, + 0.005515749100595713, + 0.007766196504235268, + 0.01614133082330227, + 0.014260184019804, + 0.02309577725827694, + 0.026642348617315292, + 0.014744640327990055, + 0.021781841292977333, + 0.03762996569275856, + 0.028280574828386307, + 0.0015333446208387613, + 0.0015082138124853373, + 0.0014477120712399483, + 0.0019800555892288685, + 0.001001899130642414, + 0.0015830694464966655, + 0.0018138136947527528, + 0.0021616569720208645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 1 + ], + "selected_reward": -0.0019418209558352828 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536413036286831 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536413036286831, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536413036286831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1354552060365677, + 0.03648979961872101, + 0.20295388996601105, + 0.33543121814727783, + 0.020390307530760765, + 0.008139141835272312, + 0.010386479087173939, + 0.04847843945026398, + 0.19228073954582214, + 0.009994721040129662 + ], + "credit_map": [ + 0.016450747847557068, + 0.01147470809519291, + 0.01765424758195877, + 0.0215547326952219, + 0.011109528131783009, + 0.020091943442821503, + 0.017225464805960655, + 0.01989384926855564, + 0.0035965864080935717, + 0.004643813241273165, + 0.0041581979021430016, + 0.00503777340054512, + 0.003364398842677474, + 0.004950232338160276, + 0.005901378113776445, + 0.004837419372051954, + 0.02110312320291996, + 0.01679282821714878, + 0.026680275797843933, + 0.0290983896702528, + 0.016927232965826988, + 0.022411955520510674, + 0.0336301364004612, + 0.03630994260311127, + 0.03701374679803848, + 0.029941901564598083, + 0.04364931955933571, + 0.0501827746629715, + 0.02540002577006817, + 0.03434428945183754, + 0.060024190694093704, + 0.05487499386072159, + 0.00182521459646523, + 0.0011871785391122103, + 0.003065569791942835, + 0.0015892165247350931, + 0.0016479906626045704, + 0.00341565883718431, + 0.0018477820558473468, + 0.005811696406453848, + 0.0010201282566413283, + 0.0009079374140128493, + 0.0010952043812721968, + 0.0010128734866157174, + 0.0006949355592951179, + 0.0013190357713028789, + 0.0010113458847627044, + 0.0010776810813695192, + 0.0016667022136971354, + 0.0011001270031556487, + 0.001130569027736783, + 0.0013318852288648486, + 0.0008136339602060616, + 0.0017998011317104101, + 0.0011378851486369967, + 0.0014058746164664626, + 0.005981882568448782, + 0.006603199057281017, + 0.005293219815939665, + 0.004941868595778942, + 0.006377209443598986, + 0.006004544906318188, + 0.005723573267459869, + 0.007552943658083677, + 0.019794756546616554, + 0.015708155930042267, + 0.02386421337723732, + 0.028085311874747276, + 0.013883600942790508, + 0.01885293237864971, + 0.03384280204772949, + 0.03824895992875099, + 0.0013687782920897007, + 0.0010431630071252584, + 0.001160524901933968, + 0.0012623314978554845, + 0.0008753355359658599, + 0.0017228754004463553, + 0.0009371175547130406, + 0.0016245945589616895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 8 + ], + "selected_reward": -0.012536413036286831 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, static close-up focuses on the car's interior rearview mirror, its black plastic casing slightly worn and textured. Reflected in the pristine glass is a perfectly framed, rapidly receding view of a long, straight asphalt highway, its white lane markings rapidly converging to disappear on the horizon. The blurred landscape of trees and rolling hills on either side of the road is rendered in a soft motion blur of deep greens and earthy browns, powerfully conveying the vehicle's high speed. The entire reflected scene is dramatically illuminated by the setting sun, which paints the sky in a stunning gradient of fiery orange, deep burgundy, and soft purple. This golden hour light spills out of the reflection, casting a warm, nostalgic glow around the edges of the frame and creating subtle lens flare on the glass's surface, evoking both a poignant atmosphere of departure and a quiet, hopeful feeling of a journey just beginning.", + "seed": 12653, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_6/index_946_seed_12653_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012194754974916577, + "vlm_logprob": -0.012194754974916577, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.012194754974916577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013898847624659538 + ], + "opposite_logprob_tokens": [ + -5.888898849487305 + ], + "logprob_reward": -0.013898847624659538, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.013898847624659538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21062569320201874, + 0.07654063403606415, + 0.07874973118305206, + 0.15294791758060455, + 0.07297518849372864, + 0.06402449309825897, + 0.07279643416404724, + 0.10695723444223404, + 0.07500376552343369, + 0.08937890082597733 + ], + "credit_map": [ + 0.01998237520456314, + 0.014478432945907116, + 0.030090928077697754, + 0.022013183683156967, + 0.022674966603517532, + 0.029365161433815956, + 0.04203837364912033, + 0.029982266947627068, + 0.01269669272005558, + 0.004860491491854191, + 0.014436697587370872, + 0.007710192818194628, + 0.005512701813131571, + 0.011344147846102715, + 0.007739599794149399, + 0.012240109033882618, + 0.010127129033207893, + 0.008344889618456364, + 0.009822333231568336, + 0.009778378531336784, + 0.009707617573440075, + 0.011308145709335804, + 0.00942469947040081, + 0.010236538015305996, + 0.01262346375733614, + 0.011912954039871693, + 0.03246781975030899, + 0.016544917598366737, + 0.009112298488616943, + 0.023067321628332138, + 0.025974063202738762, + 0.021245082840323448, + 0.006909489165991545, + 0.009905883111059666, + 0.010592365637421608, + 0.009839756414294243, + 0.00557341193780303, + 0.009184102527797222, + 0.010814267210662365, + 0.010155907832086086, + 0.006919345818459988, + 0.0064099011942744255, + 0.008347186259925365, + 0.008125858381390572, + 0.006735480856150389, + 0.009111344814300537, + 0.010890373960137367, + 0.00748500507324934, + 0.005556801799684763, + 0.011826250702142715, + 0.010353010147809982, + 0.01002670917659998, + 0.006496692541986704, + 0.009459149092435837, + 0.00806699600070715, + 0.011010824702680111, + 0.01023445837199688, + 0.009848644025623798, + 0.016805294901132584, + 0.008355467580258846, + 0.011874872259795666, + 0.02518008090555668, + 0.012328002601861954, + 0.012330415658652782, + 0.00837775506079197, + 0.007140745874494314, + 0.013301141560077667, + 0.005658030975610018, + 0.007848311215639114, + 0.013894964009523392, + 0.006062834523618221, + 0.01271998044103384, + 0.011826889589428902, + 0.009805037640035152, + 0.012860453687608242, + 0.009871873073279858, + 0.012647715397179127, + 0.011768723838031292, + 0.011094638146460056, + 0.009503569453954697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.013898847624659538 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0182165689766407 + ], + "opposite_logprob_tokens": [ + -4.018216609954834 + ], + "logprob_reward": -0.0182165689766407, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0182165689766407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14627660810947418, + 0.18041624128818512, + 0.1745658814907074, + 0.06824560463428497, + 0.05522971972823143, + 0.0698539987206459, + 0.06219422072172165, + 0.08418654650449753, + 0.07936324924230576, + 0.07966792583465576 + ], + "credit_map": [ + 0.012577993795275688, + 0.009293617680668831, + 0.019346127286553383, + 0.017209812998771667, + 0.008413824252784252, + 0.017803970724344254, + 0.03706780821084976, + 0.024563441053032875, + 0.02385350689291954, + 0.012319964356720448, + 0.04761610925197601, + 0.01496970746666193, + 0.010579845868051052, + 0.027403945103287697, + 0.02165060117840767, + 0.022022545337677002, + 0.024998296052217484, + 0.011982179246842861, + 0.02489413321018219, + 0.015252947807312012, + 0.010763896629214287, + 0.04360103979706764, + 0.015383262187242508, + 0.02769012749195099, + 0.009659817442297935, + 0.004629911854863167, + 0.014952127821743488, + 0.005868991836905479, + 0.004976604599505663, + 0.008330615237355232, + 0.01009115856140852, + 0.009736375883221626, + 0.003969856072217226, + 0.003415789920836687, + 0.005984557792544365, + 0.008519407361745834, + 0.0033304535318166018, + 0.004635000601410866, + 0.01612827554345131, + 0.009246375411748886, + 0.01114964485168457, + 0.005164366215467453, + 0.010887210257351398, + 0.004160620737820864, + 0.006953363306820393, + 0.016168180853128433, + 0.005938931833952665, + 0.00943167507648468, + 0.004820569884032011, + 0.010934785939753056, + 0.006975226569920778, + 0.006958065088838339, + 0.008602642454206944, + 0.006665351800620556, + 0.006977418437600136, + 0.010260151699185371, + 0.008801300078630447, + 0.005698584020137787, + 0.010391541756689548, + 0.010101928375661373, + 0.009322045370936394, + 0.01165948249399662, + 0.01665974035859108, + 0.011551917530596256, + 0.007061755750328302, + 0.010043691843748093, + 0.011602777056396008, + 0.008904115296900272, + 0.007443519309163094, + 0.015511683188378811, + 0.008431184105575085, + 0.010364517569541931, + 0.009212854318320751, + 0.009626660495996475, + 0.010224288329482079, + 0.00920131802558899, + 0.013598877936601639, + 0.010943986475467682, + 0.00981968268752098, + 0.007040257100015879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0182165689766407 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070127487182617 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.016070127487182617, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070127487182617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.034817177802324295, + 0.027928614988923073, + 0.743080735206604, + 0.03699827194213867, + 0.011617032811045647, + 0.015521078370511532, + 0.011767682619392872, + 0.07798481732606888, + 0.03270192816853523, + 0.007582671940326691 + ], + "credit_map": [ + 0.00407438725233078, + 0.0018370106117799878, + 0.006146930158138275, + 0.0032159967813640833, + 0.0024014099035412073, + 0.005813803058117628, + 0.005783178377896547, + 0.005544455721974373, + 0.0030367975123226643, + 0.0016909665428102016, + 0.005787596572190523, + 0.002604326931759715, + 0.001477441401220858, + 0.0037136347964406013, + 0.005544856656342745, + 0.004072991665452719, + 0.08800442516803741, + 0.047507572919130325, + 0.11920585483312607, + 0.05927950143814087, + 0.0388590432703495, + 0.17646394670009613, + 0.09390839189291, + 0.11985199898481369, + 0.0031141669023782015, + 0.0015394218498840928, + 0.008040782064199448, + 0.004117801319807768, + 0.0012674753088504076, + 0.0051405890844762325, + 0.00832401867955923, + 0.005454016383737326, + 0.0015175308799371123, + 0.0010282347211614251, + 0.0015962757170200348, + 0.001152387703768909, + 0.0007760750595480204, + 0.002159806899726391, + 0.0015440512215718627, + 0.0018426699098199606, + 0.002480843337252736, + 0.0011604215251281857, + 0.002372627379372716, + 0.0009702434181235731, + 0.0014533842913806438, + 0.003599734976887703, + 0.0013100209180265665, + 0.0021738011855632067, + 0.0011887259315699339, + 0.001413576421327889, + 0.001672936137765646, + 0.0009158303146250546, + 0.0014259547460824251, + 0.0019814549013972282, + 0.0015318416990339756, + 0.0016373618273064494, + 0.00707349693402648, + 0.00403039064258337, + 0.009935404174029827, + 0.007830538786947727, + 0.005515561439096928, + 0.018323257565498352, + 0.013931485824286938, + 0.011344674043357372, + 0.0023152055218815804, + 0.0014351395657286048, + 0.006010123062878847, + 0.0035517041105777025, + 0.002184416865929961, + 0.006630557589232922, + 0.005878825671970844, + 0.004695956129580736, + 0.0011815796606242657, + 0.0006456719129346311, + 0.0014170199865475297, + 0.000833258731290698, + 0.0005548137123696506, + 0.0010634236969053745, + 0.0007774893892928958, + 0.0011094140354543924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.016070127487182617 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017149040941148996 + ], + "opposite_logprob_tokens": [ + -6.376714706420898 + ], + "logprob_reward": -0.0017149040941148996, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017149040941148996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11600279062986374, + 0.031732384115457535, + 0.3874869644641876, + 0.09334798902273178, + 0.10019398480653763, + 0.05325812101364136, + 0.08199534565210342, + 0.03714786469936371, + 0.059493646025657654, + 0.03934093564748764 + ], + "credit_map": [ + 0.008683640509843826, + 0.007438190281391144, + 0.015880411490797997, + 0.014796935953199863, + 0.0064779180102050304, + 0.01405355241149664, + 0.03047075867652893, + 0.018201379105448723, + 0.004295624326914549, + 0.002050782786682248, + 0.0074174790643155575, + 0.0031794675160199404, + 0.002512905513867736, + 0.004215158987790346, + 0.003980299457907677, + 0.004080665297806263, + 0.048606451600790024, + 0.024674082174897194, + 0.06017674505710602, + 0.030953846871852875, + 0.02078116685152054, + 0.09469842165708542, + 0.0454571358859539, + 0.0621391199529171, + 0.006869066506624222, + 0.005444055423140526, + 0.012939385138452053, + 0.009259435348212719, + 0.005279433447867632, + 0.016048651188611984, + 0.024629749357700348, + 0.012878216803073883, + 0.00796584039926529, + 0.009251686744391918, + 0.012116252444684505, + 0.012719121761620045, + 0.007012471091002226, + 0.010838929563760757, + 0.02465653233230114, + 0.015633154660463333, + 0.008776552975177765, + 0.002873810473829508, + 0.007937547750771046, + 0.0030783615075051785, + 0.003526474116370082, + 0.013799311593174934, + 0.005384211428463459, + 0.00788185466080904, + 0.006052007433027029, + 0.01619943045079708, + 0.008311845362186432, + 0.00997628178447485, + 0.011884983628988266, + 0.006754993926733732, + 0.009297039359807968, + 0.013518759049475193, + 0.005281704477965832, + 0.0020506626460701227, + 0.005820613820105791, + 0.002677640412002802, + 0.0025647312868386507, + 0.008530070073902607, + 0.004566596355289221, + 0.005655842833220959, + 0.0043565658852458, + 0.0026302188634872437, + 0.009242364205420017, + 0.007019169628620148, + 0.004904765170067549, + 0.011468845419585705, + 0.011690537445247173, + 0.008181173354387283, + 0.003932937514036894, + 0.005127692129462957, + 0.004628544207662344, + 0.0043230983428657055, + 0.007474228739738464, + 0.005675863940268755, + 0.004566025920212269, + 0.003612543223425746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.0017149040941148996 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, static close-up focused on a car's interior rearview mirror?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rearview mirror's casing black, plastic, slightly worn, and textured?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the reflection show a long, straight asphalt highway with white lane markings converging to the horizon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the reflected landscape of trees and hills rendered in a soft motion blur of deep greens and earthy browns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a setting sun creating a gradient of fiery orange, deep burgundy, and soft purple in the reflection?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107332669198513 + ], + "opposite_logprob_tokens": [ + -4.511073112487793 + ], + "logprob_reward": -0.01107332669198513, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107332669198513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09778483957052231, + 0.0254044272005558, + 0.3977094292640686, + 0.06395740807056427, + 0.14018875360488892, + 0.040482308715581894, + 0.018664466217160225, + 0.09613689035177231, + 0.0334232859313488, + 0.08624818176031113 + ], + "credit_map": [ + 0.017440414056181908, + 0.005003766622394323, + 0.01856464147567749, + 0.007441279944032431, + 0.0049466220661997795, + 0.012768213637173176, + 0.01501901913434267, + 0.01660088077187538, + 0.003112046280875802, + 0.0015202920185402036, + 0.005502145737409592, + 0.002272711368277669, + 0.0025527875404804945, + 0.0036214112769812346, + 0.003570288885384798, + 0.0032527411822229624, + 0.046086523681879044, + 0.026350965723395348, + 0.06247926503419876, + 0.03257015720009804, + 0.020259730517864227, + 0.09724187105894089, + 0.047595277428627014, + 0.06512560695409775, + 0.007196906488388777, + 0.0030151742976158857, + 0.010101333260536194, + 0.0065068635158240795, + 0.0024358953814953566, + 0.01048120204359293, + 0.014095292426645756, + 0.010124732740223408, + 0.0173345897346735, + 0.008876009844243526, + 0.02813323587179184, + 0.013207157142460346, + 0.007143579889088869, + 0.027241887524724007, + 0.01629924401640892, + 0.0219530388712883, + 0.005651741288602352, + 0.00361585128121078, + 0.006030122749507427, + 0.003932129591703415, + 0.0028542783111333847, + 0.006886625196784735, + 0.005686495453119278, + 0.005825062748044729, + 0.0020434982143342495, + 0.0023232679814100266, + 0.0027371798641979694, + 0.0021376933436840773, + 0.0012789276661351323, + 0.0028346192557364702, + 0.0025554157327860594, + 0.002753862179815769, + 0.0075797648169100285, + 0.005357312969863415, + 0.010456188581883907, + 0.008088610135018826, + 0.01134024653583765, + 0.027350038290023804, + 0.0081033268943429, + 0.017861386761069298, + 0.002805703319609165, + 0.0017484023701399565, + 0.0056341649033129215, + 0.00398848345503211, + 0.002450336702167988, + 0.0069871158339083195, + 0.005452150944620371, + 0.004356927704066038, + 0.009299591183662415, + 0.010438786819577217, + 0.009868234395980835, + 0.009762355126440525, + 0.016324425116181374, + 0.01225019060075283, + 0.011071472428739071, + 0.007233117241412401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.01107332669198513 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_7/index_tensor([5156, 5156])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_7/index_tensor([5156, 5156])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a4f022aa3a1de68c4118216aa0e1f5b4ba5c8e3c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_7/index_tensor([5156, 5156])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a slightly low-angle medium shot, the camera tracking steadily alongside a magnificent golden retriever as it parades across a lush, sun-drenched backyard lawn. The dog's thick, honey-gold fur gleams in the warm afternoon sun as it proudly holds a sturdy, white rubber bone firmly but gently in its mouth, the object showing faint tooth marks from previous play. It moves with a brisk, trotting gait, head held high, its dark, intelligent eyes gleaming with happiness. Central to its joyful expression is the rhythmic, sweeping motion of its long, feathered tail, which sways enthusiastically from side to side. In the soft-focus background, a weathered wooden fence surrounds the yard. The low-hanging sunlight casts long, soft shadows across the vibrant green grass, bathing the entire scene in a golden glow and creating a warm atmosphere of pure, simple canine joy.", + "seed": 19054, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_7/index_5156_seed_19054_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03634330281056464, + "vlm_logprob": -0.03634330281056464, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.03634330281056464, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005810987204313278 + ], + "opposite_logprob_tokens": [ + -6.3808112144470215 + ], + "logprob_reward": -0.005810987204313278, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.005810987204313278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08048190176486969, + 0.06577504426240921, + 0.1096082329750061, + 0.038832295686006546, + 0.06917504966259003, + 0.02677154168486595, + 0.04966762289404869, + 0.16546960175037384, + 0.3743487298488617, + 0.019869958981871605 + ], + "credit_map": [ + 0.012361536733806133, + 0.008708389475941658, + 0.011808373965322971, + 0.00988817773759365, + 0.007529253605753183, + 0.01029475312680006, + 0.007070257794111967, + 0.01282116211950779, + 0.006579609587788582, + 0.0066859726794064045, + 0.011010938324034214, + 0.008767412975430489, + 0.009437155909836292, + 0.00489206425845623, + 0.009159126318991184, + 0.009242761880159378, + 0.014510268345475197, + 0.010505188256502151, + 0.010404790751636028, + 0.017906388267874718, + 0.02195858210325241, + 0.009666552767157555, + 0.01266787014901638, + 0.011988588608801365, + 0.007740363944321871, + 0.004944285377860069, + 0.004643716849386692, + 0.00520817656069994, + 0.0052533526904881, + 0.003612208180129528, + 0.0037245687562972307, + 0.0037056240253150463, + 0.007539297454059124, + 0.008617359213531017, + 0.010093066841363907, + 0.009611404500901699, + 0.007239693310111761, + 0.010403226129710674, + 0.008900429122149944, + 0.0067705693654716015, + 0.0032925817649811506, + 0.003134021768346429, + 0.003936283756047487, + 0.004526929464191198, + 0.003819828387349844, + 0.003522485727444291, + 0.002488513709977269, + 0.0020508975721895695, + 0.00910274125635624, + 0.006328057497739792, + 0.006745690479874611, + 0.0059960996732115746, + 0.0057081617414951324, + 0.006227447185665369, + 0.005313222296535969, + 0.004246204160153866, + 0.015540898777544498, + 0.017645208165049553, + 0.01629256270825863, + 0.022374046966433525, + 0.032500047236680984, + 0.019000958651304245, + 0.02631702646613121, + 0.015798868611454964, + 0.055811572819948196, + 0.04601897671818733, + 0.05068311095237732, + 0.05874588340520859, + 0.03782440721988678, + 0.047764651477336884, + 0.04539065435528755, + 0.03210948407649994, + 0.0034037414006888866, + 0.002245161449536681, + 0.002068910049274564, + 0.0028391859959810972, + 0.0019084915984421968, + 0.002358656143769622, + 0.002730234991759062, + 0.0023155780509114265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.005810987204313278 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14284613728523254 + ], + "opposite_logprob_tokens": [ + -2.01784610748291 + ], + "logprob_reward": -0.14284613728523254, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14284613728523254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11372121423482895, + 0.18744413554668427, + 0.0666145458817482, + 0.09647651761770248, + 0.05507466942071915, + 0.03609311953186989, + 0.11285016685724258, + 0.1981540024280548, + 0.08568848669528961, + 0.04788315296173096 + ], + "credit_map": [ + 0.02509172447025776, + 0.012555920518934727, + 0.010177398100495338, + 0.012778096832334995, + 0.015667874366044998, + 0.007415579166263342, + 0.010485603474080563, + 0.01954900100827217, + 0.01639373041689396, + 0.028888685628771782, + 0.02331150509417057, + 0.016414161771535873, + 0.028621286153793335, + 0.013196156360208988, + 0.038959428668022156, + 0.021659167483448982, + 0.005184026435017586, + 0.006803223863244057, + 0.0071112168952822685, + 0.00781985279172659, + 0.01145531702786684, + 0.00992288626730442, + 0.009428394958376884, + 0.008889622986316681, + 0.009443609975278378, + 0.00912997405976057, + 0.016128886491060257, + 0.009095350280404091, + 0.01866118051111698, + 0.011455993168056011, + 0.01119188778102398, + 0.011369626969099045, + 0.005155312828719616, + 0.006626158021390438, + 0.007095754146575928, + 0.006889824755489826, + 0.006987469736486673, + 0.009209827519953251, + 0.006441367790102959, + 0.006668949034065008, + 0.005223116371780634, + 0.0034730418119579554, + 0.005623426288366318, + 0.004605405498296022, + 0.005162370856851339, + 0.004054482560604811, + 0.0038565516006201506, + 0.004094721283763647, + 0.011657142080366611, + 0.012436145916581154, + 0.015484239906072617, + 0.013100218959152699, + 0.011504041962325573, + 0.016072828322649002, + 0.01566910743713379, + 0.01692643016576767, + 0.02145390398800373, + 0.02158764936029911, + 0.018644072115421295, + 0.028602249920368195, + 0.03698692470788956, + 0.021614806726574898, + 0.030748611316084862, + 0.018515771254897118, + 0.01288208644837141, + 0.011739811860024929, + 0.01063921581953764, + 0.012431823648512363, + 0.008196392096579075, + 0.011782617308199406, + 0.010798397473990917, + 0.007218133192509413, + 0.004742853343486786, + 0.004605218302458525, + 0.005746678449213505, + 0.004931464325636625, + 0.007301305420696735, + 0.003583930665627122, + 0.00560612790286541, + 0.01136557012796402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.14284613728523254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418977603316307 + ], + "opposite_logprob_tokens": [ + -4.264189720153809 + ], + "logprob_reward": -0.01418977603316307, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418977603316307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09060729295015335, + 0.15783952176570892, + 0.07055707275867462, + 0.09499325603246689, + 0.3232942819595337, + 0.06744023412466049, + 0.07883614301681519, + 0.02821105532348156, + 0.04580244794487953, + 0.042418673634529114 + ], + "credit_map": [ + 0.010122348554432392, + 0.0116177499294281, + 0.01200302504003048, + 0.017514560371637344, + 0.014502415433526039, + 0.006406256929039955, + 0.01169916894286871, + 0.006741771474480629, + 0.017258064821362495, + 0.025823308154940605, + 0.01454605720937252, + 0.0260793324559927, + 0.02753310836851597, + 0.01703154295682907, + 0.018123462796211243, + 0.01144465059041977, + 0.00553300604224205, + 0.006337963044643402, + 0.00808773748576641, + 0.011244367808103561, + 0.012421605177223682, + 0.009813915938138962, + 0.009724856354296207, + 0.007393626030534506, + 0.009735827334225178, + 0.006470744032412767, + 0.019532524049282074, + 0.009294663555920124, + 0.013282744213938713, + 0.015110069885849953, + 0.009570402093231678, + 0.01199627947062254, + 0.04387339949607849, + 0.03240625560283661, + 0.053842294961214066, + 0.038465410470962524, + 0.0345933772623539, + 0.041999492794275284, + 0.04374895244836807, + 0.03436511382460594, + 0.00789495650678873, + 0.006691161077469587, + 0.00683729350566864, + 0.012173091061413288, + 0.011019380763173103, + 0.005627043545246124, + 0.009215257130563259, + 0.007982051931321621, + 0.010619017295539379, + 0.008806396275758743, + 0.009957768954336643, + 0.011185825802385807, + 0.009017418138682842, + 0.01026470772922039, + 0.009523688815534115, + 0.009461328387260437, + 0.0038042122032493353, + 0.0035961316898465157, + 0.00295618805103004, + 0.004486917518079281, + 0.003160573076456785, + 0.003140093991532922, + 0.0036457269452512264, + 0.003421213012188673, + 0.005066287238150835, + 0.004531429149210453, + 0.0063857585191726685, + 0.006729018874466419, + 0.005879225209355354, + 0.004949100315570831, + 0.007794296834617853, + 0.004467334132641554, + 0.006608457770198584, + 0.006215742323547602, + 0.0039009370375424623, + 0.005904768593609333, + 0.005109300371259451, + 0.004063243046402931, + 0.00516278063878417, + 0.0054534440860152245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.01418977603316307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028145008254796267 + ], + "opposite_logprob_tokens": [ + -5.877814292907715 + ], + "logprob_reward": -0.0028145008254796267, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028145008254796267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14071397483348846, + 0.04208693653345108, + 0.13900554180145264, + 0.059005238115787506, + 0.04714827612042427, + 0.0881381630897522, + 0.04135473817586899, + 0.10608282685279846, + 0.288296103477478, + 0.048168182373046875 + ], + "credit_map": [ + 0.029826801270246506, + 0.01118630450218916, + 0.0165297519415617, + 0.007753867655992508, + 0.007593153510242701, + 0.019074877724051476, + 0.011706046760082245, + 0.03704315796494484, + 0.005097670014947653, + 0.00498682726174593, + 0.004334104247391224, + 0.00729439128190279, + 0.005652825813740492, + 0.004261801950633526, + 0.004370939452201128, + 0.006088375113904476, + 0.02732192538678646, + 0.01586735248565674, + 0.018806157633662224, + 0.009946675971150398, + 0.010289506986737251, + 0.028207221999764442, + 0.010552814230322838, + 0.01801389269530773, + 0.009001780301332474, + 0.006778738461434841, + 0.008732648566365242, + 0.006764627061784267, + 0.008791225031018257, + 0.006401938386261463, + 0.006824715528637171, + 0.0057095689699053764, + 0.008965771645307541, + 0.0038769261445850134, + 0.006397423800081015, + 0.006448326166719198, + 0.00399290444329381, + 0.007272204849869013, + 0.0058405231684446335, + 0.0043541924096643925, + 0.01603526808321476, + 0.006489151157438755, + 0.016915904358029366, + 0.007997551932930946, + 0.006277403794229031, + 0.018615078181028366, + 0.0080174645408988, + 0.007790339179337025, + 0.004497789312154055, + 0.00421780114993453, + 0.007839667610824108, + 0.005215927492827177, + 0.0036728070117533207, + 0.006207760889083147, + 0.005352829582989216, + 0.004350155591964722, + 0.009747901000082493, + 0.01255199033766985, + 0.011590377427637577, + 0.011932800523936749, + 0.02083340846002102, + 0.012813592329621315, + 0.015830080956220627, + 0.01078267116099596, + 0.04306456074118614, + 0.03575582057237625, + 0.039440400898456573, + 0.042485516518354416, + 0.030534563586115837, + 0.036591339856386185, + 0.03450445458292961, + 0.025919463485479355, + 0.011064332909882069, + 0.004595212638378143, + 0.0077499570325016975, + 0.004627184942364693, + 0.003549770684912801, + 0.004458296112716198, + 0.004520848859101534, + 0.00760257663205266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.0028145008254796267 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016055112704634666 + ], + "opposite_logprob_tokens": [ + -4.141055107116699 + ], + "logprob_reward": -0.016055112704634666, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016055112704634666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16949975490570068, + 0.12669824063777924, + 0.09104464203119278, + 0.1542772799730301, + 0.06298422813415527, + 0.060321103781461716, + 0.0777614638209343, + 0.06449110805988312, + 0.1276615411043167, + 0.06526059657335281 + ], + "credit_map": [ + 0.03658919036388397, + 0.01481206901371479, + 0.017108658328652382, + 0.01820686273276806, + 0.016571789979934692, + 0.008223077282309532, + 0.013314513489603996, + 0.04467359930276871, + 0.011163774877786636, + 0.020900441333651543, + 0.010908232070505619, + 0.020174095407128334, + 0.024762550368905067, + 0.010094706900417805, + 0.018240192905068398, + 0.01045424584299326, + 0.00827467255294323, + 0.007341557182371616, + 0.009760339744389057, + 0.013869409449398518, + 0.018767818808555603, + 0.01009333785623312, + 0.013619371689856052, + 0.00931814406067133, + 0.021047299727797508, + 0.01975974254310131, + 0.02417229861021042, + 0.01931956224143505, + 0.021239148452878, + 0.016123883426189423, + 0.014458918944001198, + 0.01815643347799778, + 0.007986845448613167, + 0.006358105223625898, + 0.009497280232608318, + 0.007025864906609058, + 0.007063871715217829, + 0.009491904638707638, + 0.008218442089855671, + 0.007341914344578981, + 0.007945999503135681, + 0.005825311876833439, + 0.007113033905625343, + 0.011556963436305523, + 0.007368609309196472, + 0.007423842791467905, + 0.007330020423978567, + 0.0057573262602090836, + 0.012366734445095062, + 0.010359501466155052, + 0.007936427369713783, + 0.009377826005220413, + 0.009706824086606503, + 0.009218119084835052, + 0.010766723193228245, + 0.008029316551983356, + 0.005550106521695852, + 0.006464404985308647, + 0.007532611954957247, + 0.009405942633748055, + 0.00999855250120163, + 0.008926402777433395, + 0.009679459035396576, + 0.006933632772415876, + 0.018653705716133118, + 0.016272690147161484, + 0.01836833730340004, + 0.017798973247408867, + 0.014635514467954636, + 0.015601178631186485, + 0.014135386794805527, + 0.01219576969742775, + 0.008601300418376923, + 0.011090011335909367, + 0.007971810176968575, + 0.007707332260906696, + 0.007352070417255163, + 0.005941708572208881, + 0.006624286063015461, + 0.009972085244953632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.016055112704634666 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a slightly low-angle medium shot, the camera tracking steadily alongside a magnificent golden retriever as it parades across a lush, sun-drenched backyard lawn. The dog's thick, honey-gold fur gleams in the warm afternoon sun as it proudly holds a sturdy, white rubber bone firmly but gently in its mouth, the object showing faint tooth marks from previous play. It moves with a brisk, trotting gait, head held high, its dark, intelligent eyes gleaming with happiness. Central to its joyful expression is the rhythmic, sweeping motion of its long, feathered tail, which sways enthusiastically from side to side. In the soft-focus background, a weathered wooden fence surrounds the yard. The low-hanging sunlight casts long, soft shadows across the vibrant green grass, bathing the entire scene in a golden glow and creating a warm atmosphere of pure, simple canine joy.", + "seed": 19055, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_7/index_5156_seed_19055_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.021299811406061055, + "vlm_logprob": -0.021299811406061055, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.021299811406061055, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006442847196012735 + ], + "opposite_logprob_tokens": [ + -5.881443023681641 + ], + "logprob_reward": -0.006442847196012735, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.006442847196012735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08413506299257278, + 0.35916799306869507, + 0.05807318538427353, + 0.024245141074061394, + 0.17972922325134277, + 0.07258562743663788, + 0.018449177965521812, + 0.14140155911445618, + 0.02483796887099743, + 0.03737504780292511 + ], + "credit_map": [ + 0.007298659533262253, + 0.011134427040815353, + 0.008943784050643444, + 0.008743912912905216, + 0.00915773306041956, + 0.011598342098295689, + 0.016752498224377632, + 0.010505708865821362, + 0.023539302870631218, + 0.11450561881065369, + 0.009926284663379192, + 0.020674636587500572, + 0.03148559108376503, + 0.10112311691045761, + 0.025859566405415535, + 0.0320538692176342, + 0.006048272829502821, + 0.011539309285581112, + 0.005793814081698656, + 0.00694003701210022, + 0.006205014884471893, + 0.004467868246138096, + 0.008153997361660004, + 0.008924868889153004, + 0.002584448317065835, + 0.002251545898616314, + 0.0018540407763794065, + 0.0045628622174263, + 0.003590645268559456, + 0.0038522561080753803, + 0.0024516014382243156, + 0.003097742097452283, + 0.012564516626298428, + 0.009703192859888077, + 0.013543532229959965, + 0.06739860773086548, + 0.01317835133522749, + 0.009232287295162678, + 0.01869717799127102, + 0.03541155904531479, + 0.005503875203430653, + 0.00829260516911745, + 0.006404015701264143, + 0.008322319947183132, + 0.020505839958786964, + 0.004325376357883215, + 0.006516889203339815, + 0.012714710086584091, + 0.0016404130728915334, + 0.004376537632197142, + 0.00241064908914268, + 0.0026136531960219145, + 0.0017499036621302366, + 0.0019064538646489382, + 0.0016974107129499316, + 0.00205415696837008, + 0.0054558878764510155, + 0.009215262718498707, + 0.005483364686369896, + 0.03749037906527519, + 0.004815701395273209, + 0.06419383734464645, + 0.005065485835075378, + 0.009681650437414646, + 0.003367084078490734, + 0.0028597775381058455, + 0.002739908406510949, + 0.0031731470953673124, + 0.0038171361666172743, + 0.0027176588773727417, + 0.0036941603757441044, + 0.002469096565619111, + 0.003146461443975568, + 0.006815996952354908, + 0.0028589002322405577, + 0.006243204232305288, + 0.005960679147392511, + 0.003379182191565633, + 0.005047944374382496, + 0.003922677133232355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.006442847196012735 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054982054978609085 + ], + "opposite_logprob_tokens": [ + -2.9299819469451904 + ], + "logprob_reward": -0.054982054978609085, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054982054978609085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10782264918088913, + 0.25991204380989075, + 0.06491296738386154, + 0.03227245807647705, + 0.09111540764570236, + 0.11131993681192398, + 0.06817638874053955, + 0.09733869880437851, + 0.07859055697917938, + 0.08853892236948013 + ], + "credit_map": [ + 0.009539956226944923, + 0.023028884083032608, + 0.008180784061551094, + 0.012300913222134113, + 0.0139573123306036, + 0.010275572538375854, + 0.014779015444219112, + 0.015760209411382675, + 0.01551083754748106, + 0.08625232428312302, + 0.007696019951254129, + 0.014693581499159336, + 0.017522012814879417, + 0.01984558068215847, + 0.01623488776385784, + 0.08215679973363876, + 0.006413294468075037, + 0.011494827456772327, + 0.014088943600654602, + 0.009459993802011013, + 0.005057408008724451, + 0.0050014471635222435, + 0.007508224807679653, + 0.005888823885470629, + 0.0032725087366998196, + 0.006848201155662537, + 0.0017453039763495326, + 0.003633070969954133, + 0.005691651254892349, + 0.004225186537951231, + 0.0026867766864597797, + 0.0041697584092617035, + 0.0064382231794297695, + 0.007565339561551809, + 0.006324291229248047, + 0.03296247497200966, + 0.005066565703600645, + 0.0051353671588003635, + 0.0087896678596735, + 0.01883346401154995, + 0.009513461031019688, + 0.04210129380226135, + 0.014580977149307728, + 0.008495060727000237, + 0.008929207921028137, + 0.004106334410607815, + 0.006213064305484295, + 0.017380524426698685, + 0.004623202607035637, + 0.010810318402945995, + 0.00899259839206934, + 0.018733365461230278, + 0.006448710337281227, + 0.005589178763329983, + 0.008540899492800236, + 0.004438112489879131, + 0.007762097753584385, + 0.010173101909458637, + 0.004901237320154905, + 0.03452069312334061, + 0.0051569631323218346, + 0.008178705349564552, + 0.007696827873587608, + 0.01894906908273697, + 0.00793890655040741, + 0.008509024977684021, + 0.009113681502640247, + 0.016797246411442757, + 0.008037201128900051, + 0.012105511501431465, + 0.00878168735653162, + 0.0073072887025773525, + 0.009152216836810112, + 0.012578007765114307, + 0.008189447224140167, + 0.0113949254155159, + 0.012558633461594582, + 0.011505483649671078, + 0.014546790160238743, + 0.008613405749201775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.054982054978609085 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023284463211894035 + ], + "opposite_logprob_tokens": [ + -3.773284435272217 + ], + "logprob_reward": -0.023284463211894035, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023284463211894035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0906592309474945, + 0.45629650354385376, + 0.04777219146490097, + 0.032912492752075195, + 0.04596509784460068, + 0.1277342289686203, + 0.10228226333856583, + 0.029990091919898987, + 0.0519665963947773, + 0.014421321451663971 + ], + "credit_map": [ + 0.00812061782926321, + 0.018744366243481636, + 0.0049184709787368774, + 0.017285728827118874, + 0.008486904203891754, + 0.007873211987316608, + 0.01138788741081953, + 0.01384204626083374, + 0.02842140384018421, + 0.154190331697464, + 0.01081776898354292, + 0.02709009312093258, + 0.038228996098041534, + 0.10644537955522537, + 0.03227599710226059, + 0.05882653594017029, + 0.005250759422779083, + 0.011045414954423904, + 0.004926731344312429, + 0.006617547012865543, + 0.005335405468940735, + 0.0037835349794477224, + 0.007146927528083324, + 0.0036658712197095156, + 0.0028752614744007587, + 0.006330459378659725, + 0.0026580544654279947, + 0.0060518765822052956, + 0.005006317514926195, + 0.004037986043840647, + 0.0032450833823531866, + 0.0027074546087533236, + 0.0034676191862672567, + 0.004408862441778183, + 0.0034873641561716795, + 0.0057564713060855865, + 0.002922496059909463, + 0.008736951276659966, + 0.0046736556105315685, + 0.01251167617738247, + 0.008697409182786942, + 0.01836448349058628, + 0.007644364610314369, + 0.013016629964113235, + 0.042032450437545776, + 0.009526172652840614, + 0.010813234373927116, + 0.017639486119151115, + 0.008429436944425106, + 0.013530018739402294, + 0.011565363965928555, + 0.023097719997167587, + 0.011595414020121098, + 0.011773405596613884, + 0.014778116717934608, + 0.007512778975069523, + 0.0021158072631806135, + 0.003466750727966428, + 0.0015145203797146678, + 0.004944557789713144, + 0.0017293636919930577, + 0.008506027050316334, + 0.0016001942567527294, + 0.006112873088568449, + 0.006473892834037542, + 0.009553126990795135, + 0.004227836150676012, + 0.008812043815851212, + 0.005563539452850819, + 0.0059388186782598495, + 0.0067915236577391624, + 0.004605812951922417, + 0.0015119909076020122, + 0.0022519591730087996, + 0.0011965958401560783, + 0.0018699024803936481, + 0.0021763560362160206, + 0.0017548931064084172, + 0.0018901634030044079, + 0.001769460504874587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.023284463211894035 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036145609337836504 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.0036145609337836504, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036145609337836504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13987138867378235, + 0.16493546962738037, + 0.05070868507027626, + 0.03118310682475567, + 0.06789886951446533, + 0.13562586903572083, + 0.019738443195819855, + 0.3503905236721039, + 0.017990238964557648, + 0.02165742591023445 + ], + "credit_map": [ + 0.013654852285981178, + 0.015677018091082573, + 0.01390549074858427, + 0.01685985177755356, + 0.015800030902028084, + 0.02050413191318512, + 0.02641751989722252, + 0.01705249212682247, + 0.01036829873919487, + 0.012642459943890572, + 0.005796082783490419, + 0.009714918211102486, + 0.009179847314953804, + 0.05683956667780876, + 0.009922271594405174, + 0.05047202482819557, + 0.006139208097010851, + 0.010956719517707825, + 0.006234388332813978, + 0.004814461339265108, + 0.004444092512130737, + 0.006913561373949051, + 0.005146220792084932, + 0.006060032173991203, + 0.0032097597140818834, + 0.005683451425284147, + 0.002503808122128248, + 0.0049177431501448154, + 0.0034822586458176374, + 0.005047948565334082, + 0.0027366310823708773, + 0.0036015049554407597, + 0.012697044759988785, + 0.008750523440539837, + 0.005442014429718256, + 0.01230490393936634, + 0.00481428112834692, + 0.006242366041988134, + 0.005965627729892731, + 0.01168210618197918, + 0.007542487233877182, + 0.01953829452395439, + 0.015049737878143787, + 0.013233420439064503, + 0.03622036799788475, + 0.009572429582476616, + 0.00903865322470665, + 0.025430474430322647, + 0.0018663418013602495, + 0.0048547908663749695, + 0.002038865815848112, + 0.002029313938692212, + 0.002205503173172474, + 0.002018656814470887, + 0.0023314855061471462, + 0.0023934838827699423, + 0.013432078994810581, + 0.017860664054751396, + 0.0122666135430336, + 0.12453000247478485, + 0.010036814026534557, + 0.14493247866630554, + 0.01538847852498293, + 0.011943396180868149, + 0.001578963827341795, + 0.00336456298828125, + 0.0012276193592697382, + 0.0028291847556829453, + 0.0021333603654056787, + 0.0033959178254008293, + 0.0017554797232151031, + 0.0017051496542990208, + 0.001691275043413043, + 0.00326411216519773, + 0.0018621818162500858, + 0.0023456630297005177, + 0.001968488097190857, + 0.004042853135615587, + 0.00353760807774961, + 0.0029452454764395952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.0036145609337836504 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817513071000576 + ], + "opposite_logprob_tokens": [ + -4.01817512512207 + ], + "logprob_reward": -0.01817513071000576, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817513071000576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09266320616006851, + 0.41961443424224854, + 0.05486637353897095, + 0.04203690588474274, + 0.07645201683044434, + 0.12818364799022675, + 0.11297349631786346, + 0.025046464055776596, + 0.031061766669154167, + 0.01710166409611702 + ], + "credit_map": [ + 0.009529897943139076, + 0.021003512665629387, + 0.005133903585374355, + 0.011543264612555504, + 0.010997160337865353, + 0.009639034047722816, + 0.01072861347347498, + 0.014087827876210213, + 0.02645280212163925, + 0.14215071499347687, + 0.010566404089331627, + 0.024156952276825905, + 0.034579724073410034, + 0.1034240797162056, + 0.028611404821276665, + 0.04967240244150162, + 0.0051910351030528545, + 0.006975211203098297, + 0.009334428235888481, + 0.004679040051996708, + 0.0039948089979588985, + 0.014122666791081429, + 0.005504075903445482, + 0.005065110977739096, + 0.004612743854522705, + 0.006577800028026104, + 0.003303423523902893, + 0.007097139023244381, + 0.006049702875316143, + 0.006760552991181612, + 0.003797573037445545, + 0.003837971482425928, + 0.009525054134428501, + 0.0056738476268947124, + 0.006106655579060316, + 0.024904679507017136, + 0.007466792594641447, + 0.005854460410773754, + 0.00803978368639946, + 0.008880754001438618, + 0.007631838321685791, + 0.01590825617313385, + 0.006611588876694441, + 0.014318183064460754, + 0.037591055035591125, + 0.00897228717803955, + 0.009816802106797695, + 0.027333645150065422, + 0.008000809699296951, + 0.014775502495467663, + 0.01281652133911848, + 0.022467680275440216, + 0.016979549080133438, + 0.014433814212679863, + 0.015130347572267056, + 0.008369277231395245, + 0.0026606169994920492, + 0.004913181532174349, + 0.0017135548405349255, + 0.003784838132560253, + 0.0017804320668801665, + 0.004588297568261623, + 0.002191213658079505, + 0.003414330305531621, + 0.003199121681973338, + 0.004692752379924059, + 0.0036940746940672398, + 0.005177649669349194, + 0.004239633679389954, + 0.004387401510030031, + 0.003050727304071188, + 0.0026204066816717386, + 0.0017608857015147805, + 0.0023121752310544252, + 0.0016380705637857318, + 0.0031388842035084963, + 0.002142026089131832, + 0.002151664812117815, + 0.00220415648072958, + 0.001753801479935646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.01817513071000576 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a slightly low-angle medium shot, the camera tracking steadily alongside a magnificent golden retriever as it parades across a lush, sun-drenched backyard lawn. The dog's thick, honey-gold fur gleams in the warm afternoon sun as it proudly holds a sturdy, white rubber bone firmly but gently in its mouth, the object showing faint tooth marks from previous play. It moves with a brisk, trotting gait, head held high, its dark, intelligent eyes gleaming with happiness. Central to its joyful expression is the rhythmic, sweeping motion of its long, feathered tail, which sways enthusiastically from side to side. In the soft-focus background, a weathered wooden fence surrounds the yard. The low-hanging sunlight casts long, soft shadows across the vibrant green grass, bathing the entire scene in a golden glow and creating a warm atmosphere of pure, simple canine joy.", + "seed": 19056, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_7/index_5156_seed_19056_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.024913879064843058, + "vlm_logprob": -0.024913879064843058, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.024913879064843058, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00433386629447341 + ], + "opposite_logprob_tokens": [ + -6.504333972930908 + ], + "logprob_reward": -0.00433386629447341, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.00433386629447341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17887189984321594, + 0.17745353281497955, + 0.1061541885137558, + 0.05243583023548126, + 0.10904264450073242, + 0.09054259955883026, + 0.08387714624404907, + 0.04713628441095352, + 0.08974960446357727, + 0.0647362470626831 + ], + "credit_map": [ + 0.023536430671811104, + 0.02338169701397419, + 0.0188765749335289, + 0.025715351104736328, + 0.014526809565722942, + 0.020900854840874672, + 0.02691800333559513, + 0.0250161811709404, + 0.02479649893939495, + 0.00959756225347519, + 0.033028632402420044, + 0.015340723097324371, + 0.012109067291021347, + 0.03969018533825874, + 0.014288771897554398, + 0.028602104634046555, + 0.019515100866556168, + 0.01310431957244873, + 0.014792967587709427, + 0.012031754478812218, + 0.012013630010187626, + 0.012292650528252125, + 0.00987998303025961, + 0.012523782439529896, + 0.0038408110849559307, + 0.006913075223565102, + 0.005754794925451279, + 0.006067813374102116, + 0.008763635531067848, + 0.00833126250654459, + 0.005722577217966318, + 0.007041861303150654, + 0.011876972392201424, + 0.016375338658690453, + 0.011300554499030113, + 0.019347025081515312, + 0.017537357285618782, + 0.010643139481544495, + 0.014157387427985668, + 0.007804867345839739, + 0.017928695306181908, + 0.010280836373567581, + 0.012315562926232815, + 0.014089837670326233, + 0.007644921541213989, + 0.008868122473359108, + 0.009158805944025517, + 0.010255818255245686, + 0.011807408183813095, + 0.014870951883494854, + 0.009295808151364326, + 0.011409413069486618, + 0.011749043129384518, + 0.006631387863308191, + 0.010713114403188229, + 0.007400023750960827, + 0.005052599590271711, + 0.005328174214810133, + 0.0038443359080702066, + 0.005846159998327494, + 0.009744344279170036, + 0.007031206972897053, + 0.005783573258668184, + 0.004505888558924198, + 0.013244019821286201, + 0.008327468298375607, + 0.013434091582894325, + 0.011047882959246635, + 0.006483213510364294, + 0.013946644961833954, + 0.006810546852648258, + 0.016455741599202156, + 0.008453473448753357, + 0.0070274206809699535, + 0.007647702470421791, + 0.0086205517873168, + 0.007626927923411131, + 0.008197526447474957, + 0.007804377935826778, + 0.009358260780572891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.00433386629447341 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905494958162308 + ], + "opposite_logprob_tokens": [ + -2.464055061340332 + ], + "logprob_reward": -0.08905494958162308, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905494958162308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26627781987190247, + 0.13928864896297455, + 0.11109432578086853, + 0.0452331118285656, + 0.1259728968143463, + 0.09887759387493134, + 0.05155238136649132, + 0.027828196063637733, + 0.04866562783718109, + 0.0852094516158104 + ], + "credit_map": [ + 0.05208898335695267, + 0.03247258812189102, + 0.03268985077738762, + 0.03330747410655022, + 0.029217608273029327, + 0.033403247594833374, + 0.02407171204686165, + 0.029026325792074203, + 0.014463956467807293, + 0.010263001546263695, + 0.0278802327811718, + 0.013180439360439777, + 0.009787560440599918, + 0.028993597254157066, + 0.013562793843448162, + 0.021157052367925644, + 0.027509693056344986, + 0.014318957924842834, + 0.015874939039349556, + 0.008599568158388138, + 0.009221125394105911, + 0.016284793615341187, + 0.00701369671151042, + 0.012271544896066189, + 0.006107708904892206, + 0.005980344954878092, + 0.005448679439723492, + 0.006036700215190649, + 0.006701828446239233, + 0.005142054986208677, + 0.0057936133816838264, + 0.004022177308797836, + 0.013454223051667213, + 0.01187991164624691, + 0.014495453797280788, + 0.015009064227342606, + 0.009631880559027195, + 0.030380623415112495, + 0.01173184160143137, + 0.019389890134334564, + 0.01945931650698185, + 0.012005972675979137, + 0.012960296124219894, + 0.012521539814770222, + 0.010422848165035248, + 0.010944308713078499, + 0.010529184713959694, + 0.01003411877900362, + 0.008031024597585201, + 0.007768053095787764, + 0.00627669133245945, + 0.006311593111604452, + 0.0081541882827878, + 0.004326110240072012, + 0.0064702993258833885, + 0.004214414861053228, + 0.0030346610583364964, + 0.0038913842290639877, + 0.00202135369181633, + 0.006896063685417175, + 0.0033277508337050676, + 0.0029710859525948763, + 0.0023413191083818674, + 0.0033445735462009907, + 0.006478740368038416, + 0.005744305904954672, + 0.005706554278731346, + 0.0069936951622366905, + 0.004160646349191666, + 0.005554179195314646, + 0.005485270172357559, + 0.008542229421436787, + 0.012050875462591648, + 0.0073393117636442184, + 0.010962537489831448, + 0.018366912379860878, + 0.008942844346165657, + 0.008950705640017986, + 0.00785661581903696, + 0.010739644058048725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.08905494958162308 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071204207837582 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011071204207837582, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071204207837582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1788865625858307, + 0.05973595380783081, + 0.0970592349767685, + 0.06733441352844238, + 0.06083502247929573, + 0.17193062603473663, + 0.07356049120426178, + 0.12935896217823029, + 0.11571327596902847, + 0.045585498213768005 + ], + "credit_map": [ + 0.0492146760225296, + 0.024170296266674995, + 0.02051066793501377, + 0.009485076181590557, + 0.012566029094159603, + 0.0248249638825655, + 0.015171651728451252, + 0.02294320985674858, + 0.006911103148013353, + 0.0044687166810035706, + 0.010289774276316166, + 0.0072339726611971855, + 0.006396275479346514, + 0.011273745447397232, + 0.005387365818023682, + 0.007775001227855682, + 0.02062860131263733, + 0.01212647557258606, + 0.012286153621971607, + 0.009013798087835312, + 0.010127685964107513, + 0.014634175226092339, + 0.008830228820443153, + 0.009412119165062904, + 0.008245714008808136, + 0.007451663725078106, + 0.009179223328828812, + 0.009094693697988987, + 0.009144448675215244, + 0.00957812275737524, + 0.008785879239439964, + 0.005854668095707893, + 0.0074866036884486675, + 0.005501126404851675, + 0.007508066017180681, + 0.0061304583214223385, + 0.006669214460998774, + 0.012296409346163273, + 0.004853012040257454, + 0.010390135459601879, + 0.026856957003474236, + 0.01774849370121956, + 0.024049930274486542, + 0.027770057320594788, + 0.022553792223334312, + 0.01959805190563202, + 0.016224978491663933, + 0.017128366976976395, + 0.010319162160158157, + 0.015657491981983185, + 0.007639976684004068, + 0.010882895439863205, + 0.010345150716602802, + 0.004939513746649027, + 0.00683483574539423, + 0.006941462866961956, + 0.01393395196646452, + 0.018155887722969055, + 0.012719869613647461, + 0.023944808170199394, + 0.01741870492696762, + 0.016102872788906097, + 0.014072512276470661, + 0.013010364025831223, + 0.010015567764639854, + 0.01216716505587101, + 0.018522894009947777, + 0.014622672460973263, + 0.012988517060875893, + 0.013362075202167034, + 0.014938963577151299, + 0.019095420837402344, + 0.004091661889106035, + 0.0048416960053145885, + 0.004742812365293503, + 0.0062352316454052925, + 0.005688427947461605, + 0.007682652212679386, + 0.0031865700148046017, + 0.009116445668041706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.011071204207837582 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019365858752280474 + ], + "opposite_logprob_tokens": [ + -6.251936435699463 + ], + "logprob_reward": -0.0019365858752280474, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019365858752280474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1333313286304474, + 0.2773616313934326, + 0.12939071655273438, + 0.08862044662237167, + 0.03716961294412613, + 0.06793209910392761, + 0.062004994601011276, + 0.050518833100795746, + 0.1150788813829422, + 0.038591451942920685 + ], + "credit_map": [ + 0.021562034264206886, + 0.011669713072478771, + 0.0225641168653965, + 0.011476094834506512, + 0.006838424131274223, + 0.02415323816239834, + 0.011609370820224285, + 0.023458357900381088, + 0.05505514517426491, + 0.024923423305153847, + 0.03740784898400307, + 0.0264982171356678, + 0.015912769362330437, + 0.05097316578030586, + 0.028170298784971237, + 0.03842080011963844, + 0.03134791553020477, + 0.017103543505072594, + 0.017526444047689438, + 0.010993785224854946, + 0.01267226692289114, + 0.017282748594880104, + 0.008317135274410248, + 0.014146891422569752, + 0.022438716143369675, + 0.00824022013694048, + 0.011825180612504482, + 0.01000022329390049, + 0.0062544085085392, + 0.013782444410026073, + 0.007939626462757587, + 0.008139630779623985, + 0.004278621636331081, + 0.004163527861237526, + 0.0062814680859446526, + 0.005924329627305269, + 0.005968624260276556, + 0.0040150838904082775, + 0.0036711201537400484, + 0.0028668399900197983, + 0.01171048078685999, + 0.00692436657845974, + 0.00806642509996891, + 0.008158208802342415, + 0.007957668974995613, + 0.008439332246780396, + 0.008048996329307556, + 0.008626621216535568, + 0.00999694224447012, + 0.011755647137761116, + 0.005772709380835295, + 0.006311304401606321, + 0.008642805740237236, + 0.005720166023820639, + 0.006641556043177843, + 0.007163869217038155, + 0.004596955608576536, + 0.0050114537589251995, + 0.011193495243787766, + 0.009261755272746086, + 0.004363821353763342, + 0.006391141563653946, + 0.003548904787749052, + 0.006151311565190554, + 0.012120562605559826, + 0.008246250450611115, + 0.014665508642792702, + 0.009043538011610508, + 0.008709685876965523, + 0.016118459403514862, + 0.019214868545532227, + 0.026960022747516632, + 0.004069270100444555, + 0.0031634762417525053, + 0.005976871587336063, + 0.005314128939062357, + 0.004480750299990177, + 0.004296706989407539, + 0.006444121245294809, + 0.004846132360398769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0019365858752280474 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018172789365053177 + ], + "opposite_logprob_tokens": [ + -4.018172740936279 + ], + "logprob_reward": -0.018172789365053177, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018172789365053177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11588204652070999, + 0.09639199078083038, + 0.04585043340921402, + 0.04388219490647316, + 0.06518544256687164, + 0.06654605269432068, + 0.10965123027563095, + 0.3184488117694855, + 0.08456261456012726, + 0.05359918996691704 + ], + "credit_map": [ + 0.017957858741283417, + 0.014489374123513699, + 0.014428784139454365, + 0.012693866156041622, + 0.008482089266180992, + 0.016232986003160477, + 0.015039684250950813, + 0.01655741222202778, + 0.018754752352833748, + 0.00622445298358798, + 0.012020776979625225, + 0.009533585980534554, + 0.008767915889620781, + 0.020680660381913185, + 0.009036902338266373, + 0.011372944340109825, + 0.013588765636086464, + 0.0039552426896989346, + 0.005991033744066954, + 0.003578541334718466, + 0.0038162809796631336, + 0.007200554944574833, + 0.003094329498708248, + 0.004625686444342136, + 0.006284396164119244, + 0.004679921083152294, + 0.005007637199014425, + 0.00563715910539031, + 0.006380555219948292, + 0.006141710095107555, + 0.005775846540927887, + 0.003974969033151865, + 0.00789251271635294, + 0.006246955133974552, + 0.010506962426006794, + 0.007786832749843597, + 0.007421226240694523, + 0.009753316640853882, + 0.0076731182634830475, + 0.007904515601694584, + 0.00947677530348301, + 0.0067873308435082436, + 0.008656877093017101, + 0.011191368103027344, + 0.006794619373977184, + 0.007585166487842798, + 0.008787254802882671, + 0.007266664411872625, + 0.01957176998257637, + 0.02151765674352646, + 0.01359592191874981, + 0.010589232668280602, + 0.011949975974857807, + 0.011591245420277119, + 0.009977233596146107, + 0.010858197696506977, + 0.02751043438911438, + 0.03308689594268799, + 0.0669124647974968, + 0.06583669781684875, + 0.027843372896313667, + 0.027575086802244186, + 0.023307638242840767, + 0.046376217156648636, + 0.007132168393582106, + 0.011565078981220722, + 0.012523348443210125, + 0.013129840604960918, + 0.008704662322998047, + 0.010607434436678886, + 0.00808017048984766, + 0.012819912284612656, + 0.0033796303905546665, + 0.00387565023265779, + 0.005147795658558607, + 0.0066154696978628635, + 0.005046729929745197, + 0.00945322960615158, + 0.005931972060352564, + 0.014148714952170849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.018172789365053177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a slightly low-angle medium shot, the camera tracking steadily alongside a magnificent golden retriever as it parades across a lush, sun-drenched backyard lawn. The dog's thick, honey-gold fur gleams in the warm afternoon sun as it proudly holds a sturdy, white rubber bone firmly but gently in its mouth, the object showing faint tooth marks from previous play. It moves with a brisk, trotting gait, head held high, its dark, intelligent eyes gleaming with happiness. Central to its joyful expression is the rhythmic, sweeping motion of its long, feathered tail, which sways enthusiastically from side to side. In the soft-focus background, a weathered wooden fence surrounds the yard. The low-hanging sunlight casts long, soft shadows across the vibrant green grass, bathing the entire scene in a golden glow and creating a warm atmosphere of pure, simple canine joy.", + "seed": 19057, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_7/index_5156_seed_19057_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.034939119359478354, + "vlm_logprob": -0.034939119359478354, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.034939119359478354, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004141089040786028 + ], + "opposite_logprob_tokens": [ + -6.254140853881836 + ], + "logprob_reward": -0.004141089040786028, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.004141089040786028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2886035442352295, + 0.11000403016805649, + 0.04004811868071556, + 0.11711090803146362, + 0.05827539041638374, + 0.06457633525133133, + 0.11365548521280289, + 0.05063225328922272, + 0.07912394404411316, + 0.07797001302242279 + ], + "credit_map": [ + 0.01565391570329666, + 0.019595282152295113, + 0.024255605414509773, + 0.027969039976596832, + 0.011103097349405289, + 0.08614496886730194, + 0.015451558865606785, + 0.08843008428812027, + 0.01063678041100502, + 0.005920552648603916, + 0.03473607823252678, + 0.012490874156355858, + 0.007709624711424112, + 0.011666341684758663, + 0.00830228254199028, + 0.018541492521762848, + 0.007408155128359795, + 0.005152945406734943, + 0.005458432249724865, + 0.005853988695889711, + 0.0033769814763218164, + 0.0037538905162364244, + 0.005011285189539194, + 0.0040324414148926735, + 0.014064141549170017, + 0.015732256695628166, + 0.012525134719908237, + 0.01311175711452961, + 0.017938774079084396, + 0.01754363626241684, + 0.00991117488592863, + 0.01628403179347515, + 0.005201758351176977, + 0.005283671896904707, + 0.010252946056425571, + 0.009686972014605999, + 0.005802580155432224, + 0.005985253490507603, + 0.004735069815069437, + 0.011327137239277363, + 0.007656491361558437, + 0.00761300977319479, + 0.006163421552628279, + 0.006661987863481045, + 0.008220176212489605, + 0.010919329710304737, + 0.01009585615247488, + 0.007246064487844706, + 0.009088974446058273, + 0.025005754083395004, + 0.011832334101200104, + 0.015921680256724358, + 0.011733678169548512, + 0.014542530290782452, + 0.014650271274149418, + 0.010880260728299618, + 0.005581604316830635, + 0.005982100032269955, + 0.008005373179912567, + 0.006451707798987627, + 0.00558933662250638, + 0.007362773176282644, + 0.006637813989073038, + 0.005021547432988882, + 0.009711491875350475, + 0.007035385817289352, + 0.0098035279661417, + 0.008243769407272339, + 0.007261342369019985, + 0.006884059868752956, + 0.014392992481589317, + 0.015791377052664757, + 0.010486451908946037, + 0.007836353033781052, + 0.009110932238399982, + 0.010583561845123768, + 0.006975412368774414, + 0.010798533447086811, + 0.010108288377523422, + 0.012070472352206707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.004141089040786028 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283104240894318 + ], + "opposite_logprob_tokens": [ + -2.0178310871124268 + ], + "logprob_reward": -0.14283104240894318, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283104240894318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19255582988262177, + 0.09946935623884201, + 0.028901565819978714, + 0.21166864037513733, + 0.11103323101997375, + 0.03299354761838913, + 0.15194900333881378, + 0.03058747388422489, + 0.08099585026502609, + 0.059845540672540665 + ], + "credit_map": [ + 0.009911682456731796, + 0.010457641445100307, + 0.01849316619336605, + 0.013695458881556988, + 0.008000927045941353, + 0.05683840066194534, + 0.00820858683437109, + 0.0669499859213829, + 0.008732286281883717, + 0.006090782582759857, + 0.03096538409590721, + 0.01402448769658804, + 0.005211791954934597, + 0.008090442977845669, + 0.007716794963926077, + 0.018637385219335556, + 0.0030803699046373367, + 0.003966759890317917, + 0.003938064910471439, + 0.005481804255396128, + 0.0025885857176035643, + 0.002721247263252735, + 0.004349992144852877, + 0.0027747417334467173, + 0.02712262235581875, + 0.02596331387758255, + 0.02427542768418789, + 0.03922946751117706, + 0.031847696751356125, + 0.020876551046967506, + 0.021402668207883835, + 0.02095089666545391, + 0.00875156745314598, + 0.007697911933064461, + 0.036616433411836624, + 0.012134603224694729, + 0.006435532122850418, + 0.009072470478713512, + 0.007499748840928078, + 0.022824963554739952, + 0.004936763551086187, + 0.0037733109202235937, + 0.003479723120108247, + 0.0033294393215328455, + 0.004114880692213774, + 0.005550304893404245, + 0.004187156446278095, + 0.003621965181082487, + 0.012479284778237343, + 0.027658360078930855, + 0.010092213749885559, + 0.0194409117102623, + 0.011585992760956287, + 0.030725907534360886, + 0.023350318893790245, + 0.016616003587841988, + 0.0023946312721818686, + 0.0031041139736771584, + 0.002800371963530779, + 0.004480263218283653, + 0.005178757011890411, + 0.004146001301705837, + 0.004245080519467592, + 0.004238254390656948, + 0.0069189658388495445, + 0.006046493072062731, + 0.01046048104763031, + 0.010302758775651455, + 0.007950646802783012, + 0.006378582213073969, + 0.01635635830461979, + 0.016581561416387558, + 0.006578091997653246, + 0.006386687513440847, + 0.0060662562027573586, + 0.009103959426283836, + 0.010679647326469421, + 0.003955046180635691, + 0.009813995100557804, + 0.007261855062097311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.14283104240894318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734886672347784 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.006734886672347784, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734886672347784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14956240355968475, + 0.053703706711530685, + 0.016205793246626854, + 0.24044308066368103, + 0.0552835687994957, + 0.047018907964229584, + 0.238146111369133, + 0.030748406425118446, + 0.09325598925352097, + 0.07563208043575287 + ], + "credit_map": [ + 0.008577123284339905, + 0.009337091818451881, + 0.013660983182489872, + 0.014099163003265858, + 0.004434948321431875, + 0.04323919117450714, + 0.006973172537982464, + 0.049240726977586746, + 0.006743862293660641, + 0.004263497423380613, + 0.0098088588565588, + 0.010624853894114494, + 0.004918835591524839, + 0.006033978890627623, + 0.0062898606993258, + 0.005019959062337875, + 0.001810433343052864, + 0.002151170978322625, + 0.001901817275211215, + 0.0023030627053231, + 0.002244364470243454, + 0.0015567113878205419, + 0.0023612170480191708, + 0.0018770149908959866, + 0.026634112000465393, + 0.03140176832675934, + 0.02924214117228985, + 0.046150777488946915, + 0.020188746973872185, + 0.019606053829193115, + 0.02731504663825035, + 0.03990443795919418, + 0.006046708673238754, + 0.006585081107914448, + 0.006116326432675123, + 0.010498756542801857, + 0.00864744745194912, + 0.005042234901338816, + 0.007321944925934076, + 0.005025071557611227, + 0.00861248467117548, + 0.005854182876646519, + 0.004601932130753994, + 0.00643139099702239, + 0.005648571532219648, + 0.004331051371991634, + 0.007179616950452328, + 0.004359679762274027, + 0.016945118084549904, + 0.023041408509016037, + 0.026065275073051453, + 0.05462656170129776, + 0.021427614614367485, + 0.035922739654779434, + 0.03372836112976074, + 0.026389023289084435, + 0.0031868566293269396, + 0.0028133289888501167, + 0.004779730923473835, + 0.004888132680207491, + 0.003225343069061637, + 0.003829507390037179, + 0.0032769348472356796, + 0.004748570267111063, + 0.008106779307126999, + 0.006399029400199652, + 0.021739982068538666, + 0.008814153261482716, + 0.00762642826884985, + 0.005344556178897619, + 0.012445257976651192, + 0.022779803723096848, + 0.012334873899817467, + 0.0036605859640985727, + 0.027214093133807182, + 0.006797502748668194, + 0.004192016087472439, + 0.003672955557703972, + 0.006974112242460251, + 0.010785943828523159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.006734886672347784 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00281461956910789 + ], + "opposite_logprob_tokens": [ + -5.877814769744873 + ], + "logprob_reward": -0.00281461956910789, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00281461956910789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2101457715034485, + 0.28659752011299133, + 0.028623951599001884, + 0.13002237677574158, + 0.029681717976927757, + 0.02970450557768345, + 0.05722922086715698, + 0.0499504879117012, + 0.04489440470933914, + 0.13315004110336304 + ], + "credit_map": [ + 0.01858513429760933, + 0.010122401639819145, + 0.038100067526102066, + 0.014867518097162247, + 0.006685732863843441, + 0.05917169526219368, + 0.007808652240782976, + 0.05480458214879036, + 0.042829521000385284, + 0.04305332154035568, + 0.03405581787228584, + 0.0291266031563282, + 0.025981642305850983, + 0.05325688421726227, + 0.0294201597571373, + 0.028873568400740623, + 0.0029761737678200006, + 0.004659456666558981, + 0.0030131477396935225, + 0.00556902727112174, + 0.0023201764561235905, + 0.0021534052211791277, + 0.005040011368691921, + 0.002892553573474288, + 0.010884786024689674, + 0.015968455001711845, + 0.0180512722581625, + 0.026536379009485245, + 0.009379489347338676, + 0.015714647248387337, + 0.013588271103799343, + 0.019899077713489532, + 0.0030422499403357506, + 0.003122800961136818, + 0.007111530750989914, + 0.0034063407219946384, + 0.0033783013932406902, + 0.002841173205524683, + 0.0029450624715536833, + 0.0038342576008290052, + 0.005006609950214624, + 0.004175375681370497, + 0.004534773528575897, + 0.0023855818435549736, + 0.003380263689905405, + 0.0039713457226753235, + 0.003289332613348961, + 0.0029612232465296984, + 0.005166586022824049, + 0.015693217515945435, + 0.00421714223921299, + 0.007349324878305197, + 0.005418133456259966, + 0.007416162174195051, + 0.007685852702707052, + 0.00428279722109437, + 0.009125680662691593, + 0.004645772278308868, + 0.010443528182804585, + 0.007171463221311569, + 0.004964585416018963, + 0.004070626106113195, + 0.005128960125148296, + 0.004399875178933144, + 0.0038393063005059958, + 0.004113597329705954, + 0.00938294269144535, + 0.005417329724878073, + 0.003723764792084694, + 0.003314337460324168, + 0.00820245686918497, + 0.0069006700068712234, + 0.01923076994717121, + 0.011034521274268627, + 0.04115966707468033, + 0.013833465985953808, + 0.011423625983297825, + 0.008186967112123966, + 0.011035015806555748, + 0.017246011644601822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.00281461956910789 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a golden retriever with thick, honey-gold fur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the dog holding a sturdy, white rubber bone with faint tooth marks in its mouth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the dog moving with a brisk, trotting gait while its long, feathered tail sways side to side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a lush, sun-drenched backyard lawn with a weathered wooden fence in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a slightly low-angle medium shot with the camera tracking steadily alongside the dog?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018173959106206894 + ], + "opposite_logprob_tokens": [ + -4.018174171447754 + ], + "logprob_reward": -0.018173959106206894, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018173959106206894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10218311101198196, + 0.08129888772964478, + 0.03403082862496376, + 0.17141862213611603, + 0.10424406826496124, + 0.05018295720219612, + 0.14989499747753143, + 0.029506711289286613, + 0.1274663805961609, + 0.14977344870567322 + ], + "credit_map": [ + 0.011711201630532742, + 0.008676985278725624, + 0.014095058664679527, + 0.016534145921468735, + 0.007429973687976599, + 0.01904585398733616, + 0.00985240750014782, + 0.014837482944130898, + 0.007632760796695948, + 0.007248397916555405, + 0.014957155101001263, + 0.01737925596535206, + 0.008198044262826443, + 0.008462315425276756, + 0.009669954888522625, + 0.007750999182462692, + 0.005042446311563253, + 0.004326954949647188, + 0.004186385776847601, + 0.004903134889900684, + 0.003848931984975934, + 0.002170213032513857, + 0.004928207024931908, + 0.004624554887413979, + 0.023025624454021454, + 0.028889138251543045, + 0.014960316941142082, + 0.02009429968893528, + 0.016950903460383415, + 0.017729798331856728, + 0.01809590682387352, + 0.03167262673377991, + 0.008460083976387978, + 0.011111952364444733, + 0.016448520123958588, + 0.026779523119330406, + 0.007450463715940714, + 0.010623020119965076, + 0.01037498190999031, + 0.012995520606637001, + 0.00994568970054388, + 0.004371603485196829, + 0.006872096564620733, + 0.005386833567172289, + 0.005586662795394659, + 0.007061766926199198, + 0.0055174208246171474, + 0.0054408833384513855, + 0.011474218219518661, + 0.0290285125374794, + 0.013175111263990402, + 0.026929421350359917, + 0.011408005841076374, + 0.023285236209630966, + 0.019201578572392464, + 0.015392913483083248, + 0.004070314113050699, + 0.0021415846422314644, + 0.0052275280468165874, + 0.00458909198641777, + 0.00314419181086123, + 0.0037693982012569904, + 0.0028251109179109335, + 0.003739492269232869, + 0.013561495579779148, + 0.005720560438930988, + 0.038708027452230453, + 0.008850962854921818, + 0.0073361448012292385, + 0.005782820750027895, + 0.011542189866304398, + 0.035964176058769226, + 0.022179029881954193, + 0.006979756522923708, + 0.05297143757343292, + 0.01626540534198284, + 0.007582315243780613, + 0.009994265623390675, + 0.011956386268138885, + 0.02184486575424671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 9 + ], + "selected_reward": -0.018173959106206894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_8/index_tensor([5534, 5534])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_8/index_tensor([5534, 5534])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0456e9a023dfa64a985b9827bc4de28f08326909 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_8/index_tensor([5534, 5534])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "High-frame-rate, static extreme close-ups capture a dramatic moment of natural energy in breathtaking super-slow motion. A gnarled, ancient olive branch, its bark riddled with deep fissures and weathered, gray-brown veins, snaps upward with violent yet graceful force. This powerful rebound sends shockwaves through its length, causing clusters of plump, deep purple olives, their outer skins tinged with a dull, waxy sheen, to tremble on the stem. A succession of silver-green leaves is shaken loose, each one detached and beginning a slow, mesmerizing descent. They twist and turn in the air, their dusty green tops and gleaming silver undersides catching the light in a dance of ballet. The entire scene is bathed in the warm golden glow of the afternoon sun, which sculpts the intricate grain of the wood and casts brilliant rimlight on the dark fruit. All of this is set against the soft, scorched, sun-soaked blur of the olive grove in the distance, creating a serene and breathtaking atmosphere, a moment frozen in time.", + "seed": 14998, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_8/index_5534_seed_14998_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.18804682064801453, + "vlm_logprob": -0.18804682064801453, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.18804682064801453, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024208424612879753 + ], + "opposite_logprob_tokens": [ + -3.8992083072662354 + ], + "logprob_reward": -0.024208424612879753, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024208424612879753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1857379972934723, + 0.02823682129383087, + 0.05520908907055855, + 0.05052019655704498, + 0.10024597495794296, + 0.04015478491783142, + 0.0925428569316864, + 0.16655626893043518, + 0.21172723174095154, + 0.0690687745809555 + ], + "credit_map": [ + 0.021763814613223076, + 0.030362997204065323, + 0.02048240602016449, + 0.02842942252755165, + 0.02740134857594967, + 0.018541550263762474, + 0.019855588674545288, + 0.018900878727436066, + 0.0030925043392926455, + 0.004683691076934338, + 0.0030116098932921886, + 0.0033189086243510246, + 0.0022263743449002504, + 0.005225971341133118, + 0.002168622799217701, + 0.0045091379433870316, + 0.009692555293440819, + 0.0062246182933449745, + 0.006605456117540598, + 0.011470779776573181, + 0.004761067684739828, + 0.0060717700980603695, + 0.004956564866006374, + 0.005426276940852404, + 0.011083786375820637, + 0.005050776060670614, + 0.005024036858230829, + 0.004940556362271309, + 0.011927261017262936, + 0.005127443466335535, + 0.0024827844463288784, + 0.004883551504462957, + 0.015696322545409203, + 0.01146684680134058, + 0.015053537674248219, + 0.008541923947632313, + 0.01719565875828266, + 0.008714783005416393, + 0.010532278567552567, + 0.013044623658061028, + 0.013436825014650822, + 0.00255091255530715, + 0.0028804114554077387, + 0.0026457232888787985, + 0.012111742049455643, + 0.002730309497565031, + 0.0017289374954998493, + 0.002069923561066389, + 0.02366119995713234, + 0.0059304204769432545, + 0.007309018634259701, + 0.008102955296635628, + 0.027883265167474747, + 0.008026660420000553, + 0.005837128031998873, + 0.00579221174120903, + 0.06185504049062729, + 0.007418086752295494, + 0.006873589474707842, + 0.009634314104914665, + 0.057180147618055344, + 0.009464206174015999, + 0.005936817266047001, + 0.008194057270884514, + 0.038801468908786774, + 0.013510236516594887, + 0.010450855828821659, + 0.07592758536338806, + 0.02456841617822647, + 0.02028615027666092, + 0.01477315928786993, + 0.013409361243247986, + 0.012985529378056526, + 0.006790417246520519, + 0.007231299765408039, + 0.016620231792330742, + 0.007521296851336956, + 0.006292041856795549, + 0.005785524379462004, + 0.0058424342423677444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.024208424612879753 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743106961250305 + ], + "opposite_logprob_tokens": [ + -0.9743106961250305 + ], + "logprob_reward": -0.4743106961250305, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743106961250305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25321856141090393, + 0.04825855419039726, + 0.07228689640760422, + 0.05300811678171158, + 0.07956096529960632, + 0.2035670429468155, + 0.08462502062320709, + 0.1309424340724945, + 0.04663114994764328, + 0.02790120057761669 + ], + "credit_map": [ + 0.02543184906244278, + 0.04145732522010803, + 0.0331135131418705, + 0.030122483149170876, + 0.04671662300825119, + 0.033003468066453934, + 0.021175241097807884, + 0.022198092192411423, + 0.006627548951655626, + 0.004434175323694944, + 0.005214584991335869, + 0.008163487538695335, + 0.005444265902042389, + 0.007392072584480047, + 0.004805260803550482, + 0.00617716321721673, + 0.013609741814434528, + 0.013031675480306149, + 0.009098177775740623, + 0.016070876270532608, + 0.006409728899598122, + 0.0060370671562850475, + 0.003368600970134139, + 0.004661026876419783, + 0.010877971537411213, + 0.004823470953851938, + 0.005456001963466406, + 0.0048482585698366165, + 0.014144590124487877, + 0.004222076386213303, + 0.0031368255149573088, + 0.005498928017914295, + 0.012628190219402313, + 0.0073843118734657764, + 0.009876429103314877, + 0.006643384695053101, + 0.015764569863677025, + 0.00876759272068739, + 0.00931942742317915, + 0.009177068248391151, + 0.07421866059303284, + 0.008995050564408302, + 0.009466742165386677, + 0.011145352385938168, + 0.0706859827041626, + 0.011923379264771938, + 0.008471737615764141, + 0.008660152554512024, + 0.017898565158247948, + 0.00799970980733633, + 0.010958246886730194, + 0.009164275601506233, + 0.021471675485372543, + 0.00644040247425437, + 0.004527471959590912, + 0.006164682097733021, + 0.049204301089048386, + 0.004951806273311377, + 0.005758266896009445, + 0.007358719129115343, + 0.04685913398861885, + 0.007220103871077299, + 0.003839224809780717, + 0.00575088569894433, + 0.004983157385140657, + 0.004959761630743742, + 0.006889947224408388, + 0.008244899101555347, + 0.00500150490552187, + 0.0071111503057181835, + 0.004831318277865648, + 0.004609417635947466, + 0.00524144247174263, + 0.002307258313521743, + 0.0018231256399303675, + 0.00961137842386961, + 0.002966908738017082, + 0.0018860658165067434, + 0.0020269297529011965, + 0.0020380946807563305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.4743106961250305 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289820194244385 + ], + "opposite_logprob_tokens": [ + -1.0539820194244385 + ], + "logprob_reward": -0.4289820194244385, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289820194244385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17047056555747986, + 0.07587985694408417, + 0.06832391023635864, + 0.08136027306318283, + 0.07881617546081543, + 0.10975147783756256, + 0.045254871249198914, + 0.3053678274154663, + 0.030408130958676338, + 0.03436693921685219 + ], + "credit_map": [ + 0.020991800352931023, + 0.020053591579198837, + 0.01952856406569481, + 0.022177809849381447, + 0.014697249978780746, + 0.029245587065815926, + 0.023537753149867058, + 0.020238198339939117, + 0.008509576320648193, + 0.010021466761827469, + 0.007716208230704069, + 0.010741589590907097, + 0.007684464566409588, + 0.011582664214074612, + 0.008506912738084793, + 0.011116967536509037, + 0.01096708420664072, + 0.006377648562192917, + 0.004950609523802996, + 0.009811601601541042, + 0.007976029068231583, + 0.009367558173835278, + 0.009502795524895191, + 0.009370573796331882, + 0.014894337393343449, + 0.008714144118130207, + 0.007456392049789429, + 0.010843364521861076, + 0.01492936722934246, + 0.009650337509810925, + 0.006747574545443058, + 0.008124741725623608, + 0.009529637172818184, + 0.008723900653421879, + 0.01180195901542902, + 0.007102541625499725, + 0.015157499350607395, + 0.01016785204410553, + 0.0050803255289793015, + 0.01125245075672865, + 0.039930570870637894, + 0.004652955569326878, + 0.004402220249176025, + 0.005868234671652317, + 0.03754575923085213, + 0.007755531929433346, + 0.004396486096084118, + 0.005199707578867674, + 0.007429075427353382, + 0.004815246444195509, + 0.00737830251455307, + 0.004799704998731613, + 0.006837169639766216, + 0.004152984824031591, + 0.003665646305307746, + 0.006176739931106567, + 0.11712084710597992, + 0.010757933370769024, + 0.012529789470136166, + 0.019888732582330704, + 0.10759754478931427, + 0.01620846800506115, + 0.008950191549956799, + 0.012314281426370144, + 0.003126685041934252, + 0.003797221463173628, + 0.003936408087611198, + 0.0050654783844947815, + 0.0037147633265703917, + 0.0043711899779737, + 0.002587436232715845, + 0.0038089437875896692, + 0.0038853269070386887, + 0.003221749095246196, + 0.004632452968508005, + 0.006246166303753853, + 0.003248002612963319, + 0.00555068626999855, + 0.003637337824329734, + 0.003945215605199337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.4289820194244385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098350182175636 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.004098350182175636, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098350182175636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3271072208881378, + 0.11237844079732895, + 0.046589840203523636, + 0.04026968404650688, + 0.10296420007944107, + 0.0562250092625618, + 0.0526047945022583, + 0.1430741250514984, + 0.03844346106052399, + 0.08034329861402512 + ], + "credit_map": [ + 0.0361216701567173, + 0.043792277574539185, + 0.027860283851623535, + 0.04822831600904465, + 0.044805992394685745, + 0.044177792966365814, + 0.04088783636689186, + 0.04123305529356003, + 0.011944136582314968, + 0.02308131940662861, + 0.012917082756757736, + 0.014263452030718327, + 0.008855137974023819, + 0.020975124090909958, + 0.008420804515480995, + 0.011921389028429985, + 0.00532209686934948, + 0.008203573524951935, + 0.006360351573675871, + 0.007221203297376633, + 0.004309260752052069, + 0.003925029654055834, + 0.005177713930606842, + 0.006070609204471111, + 0.006739230826497078, + 0.003685424802824855, + 0.0030930594075471163, + 0.005203965585678816, + 0.0056648519821465015, + 0.006308217532932758, + 0.005850794725120068, + 0.003724138019606471, + 0.014713393524289131, + 0.011787104420363903, + 0.013638056814670563, + 0.008135998621582985, + 0.01719542220234871, + 0.00896601565182209, + 0.013208210468292236, + 0.015319996513426304, + 0.0190579816699028, + 0.0027059183921664953, + 0.00283528259024024, + 0.0031895712018013, + 0.018695922568440437, + 0.00409894622862339, + 0.002185957506299019, + 0.003455430269241333, + 0.012611711397767067, + 0.003734485013410449, + 0.004143856931477785, + 0.0038409417029470205, + 0.01510144118219614, + 0.00599291967228055, + 0.002851631725206971, + 0.004327806178480387, + 0.05367645248770714, + 0.004915033001452684, + 0.00632734689861536, + 0.009175984188914299, + 0.04976736381649971, + 0.008022583089768887, + 0.004241361748427153, + 0.00694799330085516, + 0.004622612148523331, + 0.003279098542407155, + 0.004078918602317572, + 0.005849888548254967, + 0.004187856800854206, + 0.0083687212318182, + 0.002853341167792678, + 0.005203024484217167, + 0.015902159735560417, + 0.00433615455403924, + 0.003772933967411518, + 0.029534291476011276, + 0.0075979698449373245, + 0.007213169243186712, + 0.005340884905308485, + 0.006645735818892717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.004098350182175636 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634612895548344 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.008634612895548344, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634612895548344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1537213921546936, + 0.0837232768535614, + 0.028186403214931488, + 0.02558244578540325, + 0.0977371335029602, + 0.15883342921733856, + 0.04928775131702423, + 0.2799459397792816, + 0.06911090761423111, + 0.053871333599090576 + ], + "credit_map": [ + 0.016492821276187897, + 0.02249596081674099, + 0.019421540200710297, + 0.021365180611610413, + 0.024384059011936188, + 0.01855693757534027, + 0.015850862488150597, + 0.015154029242694378, + 0.008990606293082237, + 0.012202844955027103, + 0.008419975638389587, + 0.00986363273113966, + 0.0064191799610853195, + 0.01737886108458042, + 0.00806964561343193, + 0.01237853430211544, + 0.0025184599217027426, + 0.004829542711377144, + 0.0040054526180028915, + 0.003360904986038804, + 0.002705532358959317, + 0.003497698809951544, + 0.0025118438061326742, + 0.004756969399750233, + 0.004563041962683201, + 0.0025875607971102, + 0.0027516705449670553, + 0.003548288019374013, + 0.00290897605009377, + 0.0036162491887807846, + 0.002096444834023714, + 0.0035102132242172956, + 0.01661154255270958, + 0.009698836132884026, + 0.01094540674239397, + 0.007799883373081684, + 0.01718718186020851, + 0.010425573214888573, + 0.011672661639750004, + 0.013396046124398708, + 0.053906042128801346, + 0.008071308955550194, + 0.006825998425483704, + 0.006822974421083927, + 0.05634113401174545, + 0.011702867224812508, + 0.0065017216838896275, + 0.008661382831633091, + 0.009648341685533524, + 0.003208008361980319, + 0.0032721790485084057, + 0.004531417042016983, + 0.011707031168043613, + 0.007309244014322758, + 0.0036605438217520714, + 0.00595098827034235, + 0.1072869822382927, + 0.009717843495309353, + 0.01142007764428854, + 0.018093489110469818, + 0.09863507002592087, + 0.015008469112217426, + 0.008222002536058426, + 0.011561977677047253, + 0.011349582113325596, + 0.0037816353142261505, + 0.0034056981094181538, + 0.023293519392609596, + 0.010039672255516052, + 0.007976316846907139, + 0.0044936551712453365, + 0.004770833533257246, + 0.006777262780815363, + 0.003601041389629245, + 0.003214566269889474, + 0.01555024366825819, + 0.007027014158666134, + 0.009559093974530697, + 0.0037895950954407454, + 0.004352516960352659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.008634612895548344 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "High-frame-rate, static extreme close-ups capture a dramatic moment of natural energy in breathtaking super-slow motion. A gnarled, ancient olive branch, its bark riddled with deep fissures and weathered, gray-brown veins, snaps upward with violent yet graceful force. This powerful rebound sends shockwaves through its length, causing clusters of plump, deep purple olives, their outer skins tinged with a dull, waxy sheen, to tremble on the stem. A succession of silver-green leaves is shaken loose, each one detached and beginning a slow, mesmerizing descent. They twist and turn in the air, their dusty green tops and gleaming silver undersides catching the light in a dance of ballet. The entire scene is bathed in the warm golden glow of the afternoon sun, which sculpts the intricate grain of the wood and casts brilliant rimlight on the dark fruit. All of this is set against the soft, scorched, sun-soaked blur of the olive grove in the distance, creating a serene and breathtaking atmosphere, a moment frozen in time.", + "seed": 14999, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_8/index_5534_seed_14999_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06807687883265316, + "vlm_logprob": -0.06807687883265316, + "vlm_yes_no_margin": 4.024999976158142, + "vlm_token_credit": -0.06807687883265316, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010518587194383144 + ], + "opposite_logprob_tokens": [ + -5.260518550872803 + ], + "logprob_reward": -0.010518587194383144, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.010518587194383144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18239682912826538, + 0.05427388846874237, + 0.05244262143969536, + 0.06162194907665253, + 0.10398335009813309, + 0.05796629935503006, + 0.13552169501781464, + 0.14103467762470245, + 0.14853394031524658, + 0.06222481653094292 + ], + "credit_map": [ + 0.01481778547167778, + 0.026036955416202545, + 0.025912083685398102, + 0.026546407490968704, + 0.02107682265341282, + 0.02285992167890072, + 0.024515599012374878, + 0.020631225779652596, + 0.00480737816542387, + 0.004379299934953451, + 0.008545616641640663, + 0.0070325215347111225, + 0.0043227109126746655, + 0.004738295450806618, + 0.011205289512872696, + 0.009242773987352848, + 0.004264935851097107, + 0.006042632274329662, + 0.007202340755611658, + 0.007066791877150536, + 0.006615674588829279, + 0.0056444560177624226, + 0.009179404005408287, + 0.006426381412893534, + 0.007710134610533714, + 0.016572242602705956, + 0.007329736836254597, + 0.004927119240164757, + 0.006630616262555122, + 0.0055968547239899635, + 0.004628194961696863, + 0.008227045647799969, + 0.004866534378379583, + 0.0067508346401154995, + 0.017940016463398933, + 0.017599007114768028, + 0.00823778472840786, + 0.010007585398852825, + 0.017452102154493332, + 0.02112947590649128, + 0.0033183954656124115, + 0.008019186556339264, + 0.006684453226625919, + 0.012423479929566383, + 0.0051756263710558414, + 0.004192964639514685, + 0.00977668259292841, + 0.008375507779419422, + 0.01118390541523695, + 0.02614625357091427, + 0.02570204623043537, + 0.01255780179053545, + 0.012938763946294785, + 0.013150951825082302, + 0.01630222052335739, + 0.017539748921990395, + 0.011219969019293785, + 0.03310816362500191, + 0.018630314618349075, + 0.012708179652690887, + 0.01809338666498661, + 0.01344915758818388, + 0.016478167846798897, + 0.017347324639558792, + 0.008742326870560646, + 0.01104031689465046, + 0.020955123007297516, + 0.02202891930937767, + 0.022801607847213745, + 0.012424474582076073, + 0.029537340626120567, + 0.02100382372736931, + 0.0048994687385857105, + 0.012038243934512138, + 0.009082920849323273, + 0.006788887549191713, + 0.007544390391558409, + 0.005024473182857037, + 0.010126152075827122, + 0.006720276549458504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.010518587194383144 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281709492206573 + ], + "opposite_logprob_tokens": [ + -2.0178170204162598 + ], + "logprob_reward": -0.14281709492206573, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14281709492206573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22180300951004028, + 0.07802414149045944, + 0.0539858378469944, + 0.06351540237665176, + 0.3201277554035187, + 0.03884502500295639, + 0.048276618123054504, + 0.06127075105905533, + 0.07381559908390045, + 0.04033586010336876 + ], + "credit_map": [ + 0.013975804671645164, + 0.030266446992754936, + 0.023097572848200798, + 0.024299664422869682, + 0.012665585614740849, + 0.046414196491241455, + 0.019033778458833694, + 0.05204997584223747, + 0.005039782728999853, + 0.0047315615229308605, + 0.011858789250254631, + 0.016883062198758125, + 0.005701122805476189, + 0.007363658398389816, + 0.013767988421022892, + 0.012678178027272224, + 0.003134383587166667, + 0.004482842981815338, + 0.005804430227726698, + 0.008537029847502708, + 0.008448712527751923, + 0.004443543963134289, + 0.01205475628376007, + 0.007080139592289925, + 0.005910604260861874, + 0.009917047806084156, + 0.008124331012368202, + 0.009428761899471283, + 0.00875761080533266, + 0.0060750157572329044, + 0.008635945618152618, + 0.006666089408099651, + 0.019125180318951607, + 0.021793408319354057, + 0.04882470518350601, + 0.06698433309793472, + 0.02783942222595215, + 0.03392492234706879, + 0.04845598340034485, + 0.0531797930598259, + 0.002686556661501527, + 0.005213604308664799, + 0.006387590430676937, + 0.004353320691734552, + 0.006776920519769192, + 0.0030086145270615816, + 0.006082306616008282, + 0.00433611124753952, + 0.002818165346980095, + 0.0038793268613517284, + 0.010073460638523102, + 0.014301293529570103, + 0.003282975871115923, + 0.0023741440381854773, + 0.006272281054407358, + 0.005274972412735224, + 0.004516340792179108, + 0.004993205890059471, + 0.00944519229233265, + 0.008053211495280266, + 0.005989928264170885, + 0.006315333768725395, + 0.013290977105498314, + 0.008666560985147953, + 0.004908575210720301, + 0.0053567285649478436, + 0.011619224213063717, + 0.010949474759399891, + 0.010753823444247246, + 0.007576001808047295, + 0.01208438165485859, + 0.010567390359938145, + 0.0032989555038511753, + 0.0032011510338634253, + 0.008607028052210808, + 0.004484815523028374, + 0.0057318201288580894, + 0.006081812083721161, + 0.004669475369155407, + 0.004260803107172251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.14281709492206573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992669343948364 + ], + "opposite_logprob_tokens": [ + -1.8049266338348389 + ], + "logprob_reward": -0.17992669343948364, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992669343948364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2109440267086029, + 0.08199736475944519, + 0.08616286516189575, + 0.03601740300655365, + 0.04188885539770126, + 0.07951679080724716, + 0.09662692993879318, + 0.10084719210863113, + 0.2282695472240448, + 0.037729062139987946 + ], + "credit_map": [ + 0.01866217888891697, + 0.029681865125894547, + 0.022163566201925278, + 0.026889117434620857, + 0.014707157388329506, + 0.04213843494653702, + 0.016307849436998367, + 0.04039383679628372, + 0.008393384516239166, + 0.005418913904577494, + 0.009885421022772789, + 0.011984909884631634, + 0.0075206877663731575, + 0.011843173764646053, + 0.015994999557733536, + 0.010955859906971455, + 0.007297296077013016, + 0.010082258842885494, + 0.010165390558540821, + 0.011071362532675266, + 0.013728153891861439, + 0.008194748312234879, + 0.014819302596151829, + 0.010804339312016964, + 0.006047610659152269, + 0.006943455431610346, + 0.003894156077876687, + 0.002645535860210657, + 0.00477882893756032, + 0.0034960941411554813, + 0.0030478360131382942, + 0.005163883324712515, + 0.0027875700034201145, + 0.0036883170250803232, + 0.005600951611995697, + 0.0064522093161940575, + 0.007566388696432114, + 0.006504158955067396, + 0.0048356931656599045, + 0.004453563597053289, + 0.004136556759476662, + 0.004197306465357542, + 0.012899577617645264, + 0.021449273452162743, + 0.009030804969370365, + 0.005178180988878012, + 0.013294585049152374, + 0.009330499917268753, + 0.007210967130959034, + 0.006883546710014343, + 0.018681980669498444, + 0.016981611028313637, + 0.009373391978442669, + 0.010024220682680607, + 0.016828037798404694, + 0.010643159970641136, + 0.007955818437039852, + 0.01878879964351654, + 0.012549365870654583, + 0.011935100890696049, + 0.011739881709218025, + 0.010227670893073082, + 0.015614017844200134, + 0.012036520056426525, + 0.0143866753205657, + 0.03708581626415253, + 0.031566888093948364, + 0.025081079453229904, + 0.039855264127254486, + 0.019183754920959473, + 0.028609197586774826, + 0.032500866800546646, + 0.0033285513054579496, + 0.006578356958925724, + 0.005039548501372337, + 0.0053866831585764885, + 0.003669236321002245, + 0.0026239329017698765, + 0.006495412439107895, + 0.004607335198670626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.17992669343948364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024888513144105673 + ], + "opposite_logprob_tokens": [ + -6.002488613128662 + ], + "logprob_reward": -0.0024888513144105673, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024888513144105673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23499388992786407, + 0.12676869332790375, + 0.1004154235124588, + 0.07593557983636856, + 0.09628558903932571, + 0.07484842091798782, + 0.056680116802453995, + 0.0723796933889389, + 0.04931316524744034, + 0.11237942427396774 + ], + "credit_map": [ + 0.019038476049900055, + 0.030389904975891113, + 0.03406577184796333, + 0.026017896831035614, + 0.017432507127523422, + 0.029296372085809708, + 0.035089291632175446, + 0.04366367310285568, + 0.009603888727724552, + 0.00974218174815178, + 0.021947583183646202, + 0.022595735266804695, + 0.008988549932837486, + 0.01098848506808281, + 0.019576488062739372, + 0.023325782269239426, + 0.0068165273405611515, + 0.009512688033282757, + 0.011329174973070621, + 0.016372691839933395, + 0.011987730860710144, + 0.015672922134399414, + 0.014325729571282864, + 0.014397963881492615, + 0.005511119030416012, + 0.011897630989551544, + 0.009982351213693619, + 0.011641202494502068, + 0.009504163637757301, + 0.006804711185395718, + 0.011939149349927902, + 0.008655259385704994, + 0.007512170355767012, + 0.009317651391029358, + 0.0160523634403944, + 0.011048397049307823, + 0.012135747820138931, + 0.015608168207108974, + 0.013172514736652374, + 0.011438573710620403, + 0.00708335917443037, + 0.012900873087346554, + 0.00980244018137455, + 0.01058440376073122, + 0.006829489953815937, + 0.006111269351094961, + 0.012196114286780357, + 0.009340468794107437, + 0.005672919098287821, + 0.00666040601208806, + 0.011862323619425297, + 0.006869190372526646, + 0.006422275677323341, + 0.00639736745506525, + 0.00694566685706377, + 0.0058499653823673725, + 0.004299327731132507, + 0.005390504375100136, + 0.013421349227428436, + 0.011177564971148968, + 0.007936729118227959, + 0.0039918995462358, + 0.01566079817712307, + 0.010501524433493614, + 0.0038470844738185406, + 0.00982686411589384, + 0.005129091441631317, + 0.005603857338428497, + 0.008513658307492733, + 0.0033700314816087484, + 0.006991563830524683, + 0.006031012628227472, + 0.015296456404030323, + 0.028126133605837822, + 0.012850476428866386, + 0.009775032289326191, + 0.015164821408689022, + 0.009219743311405182, + 0.01007837150245905, + 0.011868387460708618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0024888513144105673 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633167292922735 + ], + "opposite_logprob_tokens": [ + -5.379632949829102 + ], + "logprob_reward": -0.004633167292922735, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633167292922735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35325053334236145, + 0.08893579244613647, + 0.04061457887291908, + 0.02725937031209469, + 0.10873643308877945, + 0.13629581034183502, + 0.03993058577179909, + 0.0711965337395668, + 0.026725711300969124, + 0.10705457627773285 + ], + "credit_map": [ + 0.029223976656794548, + 0.04063789173960686, + 0.0487908311188221, + 0.029607659205794334, + 0.030823007225990295, + 0.06806372851133347, + 0.05081456899642944, + 0.05528886243700981, + 0.009332126006484032, + 0.007181514520198107, + 0.017417747527360916, + 0.015337592922151089, + 0.006662657950073481, + 0.0068838405422866344, + 0.013769567012786865, + 0.012350751087069511, + 0.003617712762206793, + 0.0033412303309887648, + 0.003749753814190626, + 0.006667895242571831, + 0.0040327925235033035, + 0.004819824360311031, + 0.01003287360072136, + 0.004352495074272156, + 0.0032429404091089964, + 0.004692526068538427, + 0.0037691874895244837, + 0.002849032636731863, + 0.0037807761691510677, + 0.00233492162078619, + 0.0033015005756169558, + 0.003288485575467348, + 0.006665682420134544, + 0.00960603728890419, + 0.01832818239927292, + 0.012346956878900528, + 0.013660927303135395, + 0.019461680203676224, + 0.01704818196594715, + 0.011618778109550476, + 0.00672015780583024, + 0.012394856661558151, + 0.02059347741305828, + 0.03785816952586174, + 0.0109349824488163, + 0.009789390489459038, + 0.021389007568359375, + 0.016615765169262886, + 0.0031153825111687183, + 0.0033842071425169706, + 0.005123440641909838, + 0.005418722052127123, + 0.007143772207200527, + 0.005298202857375145, + 0.005163031630218029, + 0.005283826030790806, + 0.004542591981589794, + 0.006689045578241348, + 0.012919414788484573, + 0.00938352756202221, + 0.010606765747070312, + 0.00560821732506156, + 0.011459252797067165, + 0.009987722150981426, + 0.0017713045235723257, + 0.0022774466779083014, + 0.0037685283459722996, + 0.003969951067119837, + 0.0038541024550795555, + 0.002304369118064642, + 0.004977135453373194, + 0.0038028741255402565, + 0.008199148811399937, + 0.015606753528118134, + 0.010856670327484608, + 0.01392630860209465, + 0.011109979823231697, + 0.009890052489936352, + 0.022107267752289772, + 0.01535839680582285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.004633167292922735 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "High-frame-rate, static extreme close-ups capture a dramatic moment of natural energy in breathtaking super-slow motion. A gnarled, ancient olive branch, its bark riddled with deep fissures and weathered, gray-brown veins, snaps upward with violent yet graceful force. This powerful rebound sends shockwaves through its length, causing clusters of plump, deep purple olives, their outer skins tinged with a dull, waxy sheen, to tremble on the stem. A succession of silver-green leaves is shaken loose, each one detached and beginning a slow, mesmerizing descent. They twist and turn in the air, their dusty green tops and gleaming silver undersides catching the light in a dance of ballet. The entire scene is bathed in the warm golden glow of the afternoon sun, which sculpts the intricate grain of the wood and casts brilliant rimlight on the dark fruit. All of this is set against the soft, scorched, sun-soaked blur of the olive grove in the distance, creating a serene and breathtaking atmosphere, a moment frozen in time.", + "seed": 15000, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_8/index_5534_seed_15000_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06307392166927457, + "vlm_logprob": -0.06307392166927457, + "vlm_yes_no_margin": 4.4249999761581424, + "vlm_token_credit": -0.06307392166927457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0072029666043818 + ], + "opposite_logprob_tokens": [ + -6.257203102111816 + ], + "logprob_reward": -0.0072029666043818, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0072029666043818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20174872875213623, + 0.05282190069556236, + 0.06283609569072723, + 0.028960643336176872, + 0.05819878354668617, + 0.12564553320407867, + 0.2681650221347809, + 0.03896452859044075, + 0.07352831959724426, + 0.08913056552410126 + ], + "credit_map": [ + 0.026138270273804665, + 0.025752125307917595, + 0.01385572925209999, + 0.015731919556856155, + 0.016176022589206696, + 0.018076734617352486, + 0.04520304873585701, + 0.040814850479364395, + 0.0065306248143315315, + 0.006788972299546003, + 0.006681318860501051, + 0.007693938445299864, + 0.006928500719368458, + 0.005820444785058498, + 0.005345011595636606, + 0.00703308405354619, + 0.0050308797508478165, + 0.016166653484106064, + 0.003403343027457595, + 0.007974687963724136, + 0.008605442009866238, + 0.01340293139219284, + 0.003913598600775003, + 0.004338553640991449, + 0.003715501632541418, + 0.002455862471833825, + 0.003714928636327386, + 0.003119409317150712, + 0.004471759777516127, + 0.004159740172326565, + 0.003752314718440175, + 0.0035711261443793774, + 0.0069147213362157345, + 0.006382837425917387, + 0.005582797806710005, + 0.0028869532980024815, + 0.011969214305281639, + 0.011437393724918365, + 0.006673150695860386, + 0.00635170703753829, + 0.016292991116642952, + 0.020176272839307785, + 0.008619788102805614, + 0.014526435174047947, + 0.01827945187687874, + 0.01155405305325985, + 0.018846232444047928, + 0.017350295558571815, + 0.03510577976703644, + 0.04774974659085274, + 0.020364832133054733, + 0.03428428992629051, + 0.02679113857448101, + 0.025739753618836403, + 0.0315389558672905, + 0.04659049212932587, + 0.003505825297906995, + 0.006908494513481855, + 0.0031247222796082497, + 0.005110512021929026, + 0.005639829207211733, + 0.0038700478617101908, + 0.0062668342143297195, + 0.004538259003311396, + 0.010680558159947395, + 0.012076297774910927, + 0.007237953133881092, + 0.006846935022622347, + 0.007123894523829222, + 0.010723724961280823, + 0.00656616548076272, + 0.012272781692445278, + 0.011889124289155006, + 0.013269183225929737, + 0.008022439666092396, + 0.007844020612537861, + 0.007171259261667728, + 0.018221575766801834, + 0.008458494208753109, + 0.014254461042582989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0072029666043818 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603575497865677 + ], + "opposite_logprob_tokens": [ + -1.9103575944900513 + ], + "logprob_reward": -0.1603575497865677, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603575497865677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6730050444602966, + 0.016594063490629196, + 0.03524157404899597, + 0.028634125366806984, + 0.050593819469213486, + 0.03861882537603378, + 0.1130090057849884, + 0.013786339201033115, + 0.016159191727638245, + 0.014358032494783401 + ], + "credit_map": [ + 0.05311546102166176, + 0.024510851129889488, + 0.02519974671304226, + 0.027077386155724525, + 0.030491691082715988, + 0.25472119450569153, + 0.2135886698961258, + 0.04430006071925163, + 0.0013304303865879774, + 0.002126049716025591, + 0.0017663708422333002, + 0.002685863059014082, + 0.0022985809482634068, + 0.00207587075419724, + 0.0022947757970541716, + 0.002016122918576002, + 0.0027716252952814102, + 0.010088933631777763, + 0.0014005384873598814, + 0.003534639487043023, + 0.002051902236416936, + 0.010770776309072971, + 0.002037002006545663, + 0.0025861584581434727, + 0.002687794389203191, + 0.0024459771811962128, + 0.0018047833582386374, + 0.003827821696177125, + 0.004980759229511023, + 0.006305940914899111, + 0.0032116458751261234, + 0.0033694016747176647, + 0.0042211515828967094, + 0.007730200886726379, + 0.0032330267131328583, + 0.002993155037984252, + 0.012565722689032555, + 0.011542065069079399, + 0.004657966550439596, + 0.0036505304742604494, + 0.004561073146760464, + 0.006241456139832735, + 0.003088463796302676, + 0.004862690810114145, + 0.003384660929441452, + 0.005744655150920153, + 0.005266190972179174, + 0.005469633266329765, + 0.014810445718467236, + 0.022443385794758797, + 0.007891290821135044, + 0.013825387693941593, + 0.011202854104340076, + 0.010150794871151447, + 0.014383452013134956, + 0.018301399424672127, + 0.0016746956389397383, + 0.002590655582025647, + 0.0012796358205378056, + 0.0014070449396967888, + 0.0018971145618706942, + 0.0017010767478495836, + 0.0016595328925177455, + 0.0015765827847644687, + 0.0020248284563422203, + 0.0032073718030005693, + 0.0011877943761646748, + 0.0011569643393158913, + 0.0038649209309369326, + 0.0017813288141041994, + 0.0016938832122832537, + 0.0012421002611517906, + 0.0014706263318657875, + 0.001992560923099518, + 0.0013793855905532837, + 0.0014802152290940285, + 0.0026014274917542934, + 0.0017386536346748471, + 0.0013550636358559132, + 0.002340099774301052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.1603575497865677 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428588479757309 + ], + "opposite_logprob_tokens": [ + -2.0178587436676025 + ], + "logprob_reward": -0.1428588479757309, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1428588479757309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15364137291908264, + 0.03422889485955238, + 0.049078404903411865, + 0.01842007040977478, + 0.02387853153049946, + 0.042082879692316055, + 0.5997229218482971, + 0.026182087138295174, + 0.016206087544560432, + 0.036558717489242554 + ], + "credit_map": [ + 0.02086920663714409, + 0.012170998379588127, + 0.011104123666882515, + 0.009442681446671486, + 0.011107626371085644, + 0.052127815783023834, + 0.019084744155406952, + 0.017734184861183167, + 0.0033351799938827753, + 0.0029191095381975174, + 0.004002690315246582, + 0.005403677932918072, + 0.005596253555268049, + 0.003792345989495516, + 0.004441578406840563, + 0.004738060291856527, + 0.00509644253179431, + 0.01114374678581953, + 0.004325630608946085, + 0.004343971144407988, + 0.00631544878706336, + 0.006394650787115097, + 0.006171423941850662, + 0.00528709776699543, + 0.002447693143039942, + 0.0018051268998533487, + 0.0015865173190832138, + 0.0033312218729406595, + 0.001795744406990707, + 0.0024407031014561653, + 0.0022986470721662045, + 0.00271441787481308, + 0.004168054088950157, + 0.0029888611752539873, + 0.0037287704180926085, + 0.0016106638358905911, + 0.002397222677245736, + 0.0040238467045128345, + 0.0025876148138195276, + 0.0023735016584396362, + 0.005134719889611006, + 0.0105317747220397, + 0.0029777807649224997, + 0.0034561047796159983, + 0.004435378592461348, + 0.0057206242345273495, + 0.0050740791484713554, + 0.004752419423311949, + 0.07873792946338654, + 0.10245022177696228, + 0.04472788795828819, + 0.07921908795833588, + 0.05595879256725311, + 0.06429934501647949, + 0.06989798694849014, + 0.10443171113729477, + 0.002444211160764098, + 0.006216689012944698, + 0.001723015564493835, + 0.0022340419236570597, + 0.0053063915111124516, + 0.002048791153356433, + 0.003260846948251128, + 0.0029481016099452972, + 0.0021851591300219297, + 0.0032821435015648603, + 0.0013733217492699623, + 0.0013962636003270745, + 0.0027487834449857473, + 0.0018668808043003082, + 0.0016341708833351731, + 0.0017193652456626296, + 0.004316262900829315, + 0.00574920978397131, + 0.002568852389231324, + 0.0027948515489697456, + 0.004950680769979954, + 0.008484926074743271, + 0.00384178152307868, + 0.003852155525237322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.1428588479757309 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013360392767935991 + ], + "opposite_logprob_tokens": [ + -6.626336097717285 + ], + "logprob_reward": -0.0013360392767935991, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013360392767935991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2684863805770874, + 0.024511439725756645, + 0.0491621196269989, + 0.009882801212370396, + 0.0813124030828476, + 0.02861206978559494, + 0.49843961000442505, + 0.020319297909736633, + 0.011250306852161884, + 0.00802355632185936 + ], + "credit_map": [ + 0.0198591947555542, + 0.009430049918591976, + 0.009110561572015285, + 0.01005209144204855, + 0.011745062656700611, + 0.10622116178274155, + 0.07743380963802338, + 0.024634437635540962, + 0.002607621718198061, + 0.002175902482122183, + 0.00311694061383605, + 0.004017354920506477, + 0.004106312990188599, + 0.0023327034432440996, + 0.003239530371502042, + 0.0029150741174817085, + 0.003776808734983206, + 0.0146212512627244, + 0.0019119298085570335, + 0.004084576852619648, + 0.0040783812291920185, + 0.015131570398807526, + 0.0023379528429359198, + 0.0032196473330259323, + 0.0011353602167218924, + 0.0011525923619046807, + 0.0007684211595915258, + 0.001333999214693904, + 0.002213971922174096, + 0.0009155244915746152, + 0.0011472682235762477, + 0.0012156636221334338, + 0.009436608292162418, + 0.010722433216869831, + 0.007973658852279186, + 0.006098932586610317, + 0.013830340467393398, + 0.014147644862532616, + 0.011095401830971241, + 0.008007379248738289, + 0.003156344871968031, + 0.004987822379916906, + 0.0020214274991303682, + 0.00298159453086555, + 0.0030399428214877844, + 0.0035109033342450857, + 0.004493616055697203, + 0.004420420154929161, + 0.0655146911740303, + 0.08622530102729797, + 0.03732757270336151, + 0.06541787087917328, + 0.046832941472530365, + 0.05262934789061546, + 0.057917602360248566, + 0.08657427877187729, + 0.002190094906836748, + 0.0036890979390591383, + 0.0017722145421430469, + 0.0019411372486501932, + 0.0035824975930154324, + 0.0016038070898503065, + 0.002425028709694743, + 0.0031154185999184847, + 0.0016322493320330977, + 0.001713115954771638, + 0.0009899114957079291, + 0.0010285015450790524, + 0.0013881896156817675, + 0.0018284624675288796, + 0.001050202758051455, + 0.0016196741489693522, + 0.0008991688955575228, + 0.0011259710881859064, + 0.0007869818364270031, + 0.0010508561972528696, + 0.0006278363871388137, + 0.001431829179637134, + 0.0008778584888204932, + 0.0012230537831783295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.0013360392767935991 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00361420470289886 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.00361420470289886, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00361420470289886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1306326538324356, + 0.04252419248223305, + 0.11146091669797897, + 0.033894218504428864, + 0.03114839643239975, + 0.04162927716970444, + 0.5112507343292236, + 0.06890325248241425, + 0.018104052171111107, + 0.01045229658484459 + ], + "credit_map": [ + 0.02080509066581726, + 0.02089810185134411, + 0.008423673920333385, + 0.011695705354213715, + 0.014258721843361855, + 0.027579402551054955, + 0.013727502897381783, + 0.013244460336863995, + 0.004384241532534361, + 0.0044406019151210785, + 0.00427637342363596, + 0.007300358731299639, + 0.00815211609005928, + 0.004501151852309704, + 0.005396194290369749, + 0.00407315231859684, + 0.00865243561565876, + 0.03345193713903427, + 0.004771217703819275, + 0.009001745842397213, + 0.007131930440664291, + 0.036451078951358795, + 0.006137125194072723, + 0.005863445810973644, + 0.0035335044376552105, + 0.002909200731664896, + 0.0030484902672469616, + 0.004002920817583799, + 0.005356968380510807, + 0.006734768860042095, + 0.004743031691759825, + 0.0035653335507959127, + 0.005256054922938347, + 0.0031692851334810257, + 0.0029049154836684465, + 0.0026999269612133503, + 0.0051523358561098576, + 0.004437240771949291, + 0.004039916209876537, + 0.0034887229558080435, + 0.006357233971357346, + 0.01105889305472374, + 0.003926804754883051, + 0.003831924870610237, + 0.005045242141932249, + 0.0033507489133626223, + 0.004808351863175631, + 0.0032500773668289185, + 0.06804530322551727, + 0.08975430577993393, + 0.03782559186220169, + 0.06696615368127823, + 0.0478927344083786, + 0.053904809057712555, + 0.058810558170080185, + 0.08805125951766968, + 0.007000688463449478, + 0.017543362453579903, + 0.005158886779099703, + 0.006979054305702448, + 0.00831300113350153, + 0.005307568237185478, + 0.008765293285250664, + 0.009835396893322468, + 0.0018706545233726501, + 0.002694131573662162, + 0.0014110313495621085, + 0.0014785248786211014, + 0.005494543816894293, + 0.002014016266912222, + 0.0017099945107474923, + 0.0014311550185084343, + 0.0011276170844212174, + 0.0016061401693150401, + 0.0012023525778204203, + 0.0012499273288995028, + 0.0010614593047648668, + 0.0016932840226218104, + 0.001039479742757976, + 0.0014720362378284335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.00361420470289886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "High-frame-rate, static extreme close-ups capture a dramatic moment of natural energy in breathtaking super-slow motion. A gnarled, ancient olive branch, its bark riddled with deep fissures and weathered, gray-brown veins, snaps upward with violent yet graceful force. This powerful rebound sends shockwaves through its length, causing clusters of plump, deep purple olives, their outer skins tinged with a dull, waxy sheen, to tremble on the stem. A succession of silver-green leaves is shaken loose, each one detached and beginning a slow, mesmerizing descent. They twist and turn in the air, their dusty green tops and gleaming silver undersides catching the light in a dance of ballet. The entire scene is bathed in the warm golden glow of the afternoon sun, which sculpts the intricate grain of the wood and casts brilliant rimlight on the dark fruit. All of this is set against the soft, scorched, sun-soaked blur of the olive grove in the distance, creating a serene and breathtaking atmosphere, a moment frozen in time.", + "seed": 15001, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_8/index_5534_seed_15001_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10158896541688592, + "vlm_logprob": -0.10158896541688592, + "vlm_yes_no_margin": 4.050000011920929, + "vlm_token_credit": -0.10158896541688592, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011210666969418526 + ], + "opposite_logprob_tokens": [ + -5.2612104415893555 + ], + "logprob_reward": -0.011210666969418526, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.011210666969418526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.144060418009758, + 0.16135194897651672, + 0.09688957780599594, + 0.0454922690987587, + 0.1046295315027237, + 0.12287092208862305, + 0.12835556268692017, + 0.051035571843385696, + 0.06730321794748306, + 0.07801096141338348 + ], + "credit_map": [ + 0.011831081472337246, + 0.011357076466083527, + 0.018565485253930092, + 0.02039169892668724, + 0.013790908269584179, + 0.021634532138705254, + 0.02696870267391205, + 0.019520942121744156, + 0.012025473639369011, + 0.018505152314901352, + 0.027360331267118454, + 0.018837980926036835, + 0.018142811954021454, + 0.0250963494181633, + 0.021390806883573532, + 0.019993042573332787, + 0.008629708550870419, + 0.0117397615686059, + 0.011672141961753368, + 0.016915395855903625, + 0.009520902298390865, + 0.01330268383026123, + 0.01344313845038414, + 0.01166584063321352, + 0.004041170235723257, + 0.005209417548030615, + 0.009102750569581985, + 0.0071029458194971085, + 0.0030589618254452944, + 0.0058780875988304615, + 0.005583398975431919, + 0.005515534430742264, + 0.006338546983897686, + 0.009465064853429794, + 0.021981043741106987, + 0.017011309042572975, + 0.005968673620373011, + 0.011917824856936932, + 0.023006249219179153, + 0.00894082523882389, + 0.012135866098105907, + 0.016211550682783127, + 0.01858171634376049, + 0.018103482201695442, + 0.009195210412144661, + 0.015805508941411972, + 0.021232997998595238, + 0.011604588478803635, + 0.007277519907802343, + 0.013648401945829391, + 0.01767820678651333, + 0.024350978434085846, + 0.01016156654804945, + 0.01519152894616127, + 0.028744637966156006, + 0.011302730068564415, + 0.003740833140909672, + 0.0061499890871346, + 0.007374284323304892, + 0.007100807037204504, + 0.004505745135247707, + 0.009747798554599285, + 0.007284609600901604, + 0.005131505895406008, + 0.003960225731134415, + 0.004276967141777277, + 0.014487489126622677, + 0.006896682549268007, + 0.004122870974242687, + 0.00767028983682394, + 0.01802189089357853, + 0.007866805419325829, + 0.005942898336797953, + 0.011653410270810127, + 0.012235029600560665, + 0.013267956674098969, + 0.008401486091315746, + 0.009362624026834965, + 0.009626424871385098, + 0.007521129213273525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.011210666969418526 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486267924308777 + ], + "opposite_logprob_tokens": [ + -1.2236268520355225 + ], + "logprob_reward": -0.3486267924308777, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.3486267924308777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10284094512462616, + 0.13246333599090576, + 0.09603971987962723, + 0.0870201587677002, + 0.19038273394107819, + 0.09044693410396576, + 0.053604818880558014, + 0.06973013281822205, + 0.14260950684547424, + 0.034861717373132706 + ], + "credit_map": [ + 0.01121517550200224, + 0.007169447373598814, + 0.013708723708987236, + 0.01976950652897358, + 0.01208828017115593, + 0.01151167694479227, + 0.014238584786653519, + 0.013139553368091583, + 0.008499843999743462, + 0.016322143375873566, + 0.022019775584340096, + 0.01777637004852295, + 0.00989975593984127, + 0.020889783278107643, + 0.021940410137176514, + 0.015115257352590561, + 0.008604904636740685, + 0.011406650766730309, + 0.011758195236325264, + 0.011927803978323936, + 0.009721025824546814, + 0.01690252311527729, + 0.013717475347220898, + 0.012001142837107182, + 0.01540553942322731, + 0.0077459667809307575, + 0.012728551402688026, + 0.0092079509049654, + 0.005362155381590128, + 0.012906387448310852, + 0.009908653795719147, + 0.01375496108084917, + 0.012550500221550465, + 0.013240368105471134, + 0.034570205956697464, + 0.03111371397972107, + 0.014682121574878693, + 0.0174599289894104, + 0.04638959839940071, + 0.020376285538077354, + 0.010215823538601398, + 0.00826902873814106, + 0.019734015688300133, + 0.011820759624242783, + 0.005070367828011513, + 0.00871361419558525, + 0.017665021121501923, + 0.008958298712968826, + 0.0034308999311178923, + 0.006091034505516291, + 0.009164337068796158, + 0.007663405034691095, + 0.0035837728064507246, + 0.008666160516440868, + 0.008980265818536282, + 0.006024940405040979, + 0.007553397212177515, + 0.006238460075110197, + 0.013228117488324642, + 0.01082885917276144, + 0.004705030936747789, + 0.009625034406781197, + 0.010333894751966, + 0.007217335514724255, + 0.009222632274031639, + 0.02365594543516636, + 0.030352182686328888, + 0.019270308315753937, + 0.011643882840871811, + 0.012568559497594833, + 0.02185763046145439, + 0.014038369990885258, + 0.0024745743721723557, + 0.005728097166866064, + 0.006987781263887882, + 0.005876568611711264, + 0.0032329997047781944, + 0.0035194302909076214, + 0.004815659020096064, + 0.0022266076412051916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.3486267924308777 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283797144889832 + ], + "opposite_logprob_tokens": [ + -2.0178380012512207 + ], + "logprob_reward": -0.14283797144889832, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283797144889832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12147529423236847, + 0.1055496409535408, + 0.07743291556835175, + 0.05617281794548035, + 0.18023182451725006, + 0.1600797325372696, + 0.10252062976360321, + 0.05629216507077217, + 0.10539379715919495, + 0.03485117480158806 + ], + "credit_map": [ + 0.010868215933442116, + 0.010451204143464565, + 0.02094697393476963, + 0.016690967604517937, + 0.011476353742182255, + 0.014444824308156967, + 0.01864463835954666, + 0.01795210689306259, + 0.006891836412250996, + 0.014701634645462036, + 0.019312173128128052, + 0.013605327345430851, + 0.0070152487605810165, + 0.014832168817520142, + 0.018071943894028664, + 0.011119299568235874, + 0.007356537040323019, + 0.008936753496527672, + 0.008655359968543053, + 0.011446764692664146, + 0.008849961683154106, + 0.011348637752234936, + 0.01327596791088581, + 0.007562924176454544, + 0.009853665716946125, + 0.00684297876432538, + 0.008429269306361675, + 0.006665816530585289, + 0.0051377806812524796, + 0.007086774799972773, + 0.006887627299875021, + 0.005268898792564869, + 0.011245209723711014, + 0.015066948719322681, + 0.0336248017847538, + 0.0303022600710392, + 0.01435618195682764, + 0.017544936388731003, + 0.03751964122056961, + 0.020571842789649963, + 0.01753859594464302, + 0.011517947539687157, + 0.025571245700120926, + 0.024177858605980873, + 0.01207053568214178, + 0.015003644861280918, + 0.038236115127801895, + 0.015963774174451828, + 0.005594051443040371, + 0.013184933923184872, + 0.011963224038481712, + 0.016492366790771484, + 0.009815788827836514, + 0.014869449660182, + 0.01655515655875206, + 0.014045645482838154, + 0.004745473153889179, + 0.005715153180062771, + 0.012171878479421139, + 0.008015958592295647, + 0.004356767982244492, + 0.008924302645027637, + 0.007587206084281206, + 0.004775417037308216, + 0.0058914171531796455, + 0.015646012499928474, + 0.025245629251003265, + 0.012810866348445415, + 0.007127733901143074, + 0.007633418310433626, + 0.021491415798664093, + 0.009547294117510319, + 0.0021520142909139395, + 0.003906628582626581, + 0.0063923196867108345, + 0.004320880398154259, + 0.004061406012624502, + 0.004865926690399647, + 0.0064553553238511086, + 0.0026966382283717394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.14283797144889832 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011792379664257169 + ], + "opposite_logprob_tokens": [ + -6.751179218292236 + ], + "logprob_reward": -0.0011792379664257169, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011792379664257169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2951635420322418, + 0.10164553672075272, + 0.045761194080114365, + 0.11655861139297485, + 0.07262860983610153, + 0.04177968204021454, + 0.11188347637653351, + 0.07300274074077606, + 0.09203917533159256, + 0.049537450075149536 + ], + "credit_map": [ + 0.017753850668668747, + 0.013850796967744827, + 0.050701528787612915, + 0.0729668140411377, + 0.01856408827006817, + 0.03724250569939613, + 0.05266408249735832, + 0.03141988813877106, + 0.007650506217032671, + 0.012183543294668198, + 0.014567098580300808, + 0.011603561230003834, + 0.013393981382250786, + 0.017168713733553886, + 0.012176614254713058, + 0.012901517562568188, + 0.0037319546099752188, + 0.009601599536836147, + 0.005595007911324501, + 0.006325846537947655, + 0.0037678948137909174, + 0.006771764252334833, + 0.004720868077129126, + 0.005246256943792105, + 0.023100024089217186, + 0.01144590973854065, + 0.016081655398011208, + 0.01066567748785019, + 0.007907302118837833, + 0.01688958890736103, + 0.012782403267920017, + 0.01768605411052704, + 0.004997251555323601, + 0.012961557134985924, + 0.010466894134879112, + 0.009547248482704163, + 0.00727548124268651, + 0.007530610077083111, + 0.012334441766142845, + 0.007515129633247852, + 0.003364589298143983, + 0.0035579754039645195, + 0.010270307771861553, + 0.007985330186784267, + 0.0027648541145026684, + 0.004751195665448904, + 0.005872898269444704, + 0.0032125317957252264, + 0.006298863794654608, + 0.00832211971282959, + 0.020003855228424072, + 0.014794320799410343, + 0.007366083562374115, + 0.015365841798484325, + 0.0229179784655571, + 0.016814416274428368, + 0.007117828819900751, + 0.007711566984653473, + 0.011202262714505196, + 0.008161040022969246, + 0.005332549102604389, + 0.012442557141184807, + 0.011236528865993023, + 0.009798412211239338, + 0.005552646238356829, + 0.013039379380643368, + 0.020751863718032837, + 0.01071953121572733, + 0.0066347080282866955, + 0.01204602885991335, + 0.01405242271721363, + 0.0092425886541605, + 0.003678083186969161, + 0.007117079570889473, + 0.006959260441362858, + 0.008671063929796219, + 0.004790906794369221, + 0.006123847793787718, + 0.007665726356208324, + 0.004531483631581068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.0011792379664257169 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled, ancient olive branch with deep fissures and gray-brown bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the branch snap upward with force, causing clusters of deep purple olives to tremble?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are silver-green leaves shown detaching and twisting in a slow descent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene bathed in warm golden afternoon sunlight with visible rimlight on the fruit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a soft, scorched, sun-soaked blur of an olive grove?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040901582688093185 + ], + "opposite_logprob_tokens": [ + -5.504090309143066 + ], + "logprob_reward": -0.0040901582688093185, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040901582688093185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16944201290607452, + 0.22141090035438538, + 0.04888944327831268, + 0.060486141592264175, + 0.10756342858076096, + 0.06347949057817459, + 0.14476461708545685, + 0.07333603501319885, + 0.06279009580612183, + 0.04783787205815315 + ], + "credit_map": [ + 0.009927215054631233, + 0.00825879443436861, + 0.03293688967823982, + 0.03908698633313179, + 0.011127597652375698, + 0.023280063644051552, + 0.02949940226972103, + 0.015325063839554787, + 0.016308613121509552, + 0.024101803079247475, + 0.04460999742150307, + 0.03623548522591591, + 0.022525561973452568, + 0.027985410764813423, + 0.031203852966427803, + 0.018440162762999535, + 0.005837968550622463, + 0.008744657970964909, + 0.006475144065916538, + 0.005880176555365324, + 0.005372834391891956, + 0.005951272323727608, + 0.004801109898835421, + 0.005826283246278763, + 0.010372216813266277, + 0.005074992775917053, + 0.00877510104328394, + 0.0055603645741939545, + 0.003868082305416465, + 0.010776260867714882, + 0.007010763511061668, + 0.009048360399901867, + 0.005767724476754665, + 0.016596410423517227, + 0.01938970945775509, + 0.012644201517105103, + 0.009655267931520939, + 0.008824006654322147, + 0.021488599479198456, + 0.013197507709264755, + 0.004995768889784813, + 0.006112190429121256, + 0.011420614086091518, + 0.014107877388596535, + 0.004153744783252478, + 0.007491156924515963, + 0.010235171765089035, + 0.00496297050267458, + 0.008755182847380638, + 0.010692758485674858, + 0.020331362262368202, + 0.02301464043557644, + 0.014108552597463131, + 0.020258858799934387, + 0.03044106997549534, + 0.01716219261288643, + 0.004693341441452503, + 0.008278864435851574, + 0.009745136834681034, + 0.010266805067658424, + 0.006495099049061537, + 0.013817213475704193, + 0.01066715270280838, + 0.00937242154031992, + 0.003449907759204507, + 0.005969376768916845, + 0.013731330633163452, + 0.006590707693248987, + 0.004182219970971346, + 0.010782266966998577, + 0.010088820941746235, + 0.007995462976396084, + 0.0031647118739783764, + 0.006551328580826521, + 0.007870039902627468, + 0.008173048496246338, + 0.004528277553617954, + 0.0064318738877773285, + 0.006724867038428783, + 0.004393727984279394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0040901582688093185 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_9/index_tensor([8930, 8930])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_9/index_tensor([8930, 8930])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3342def94fc4e11f6cf28202446487f2ee08da86 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_9/index_tensor([8930, 8930])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme macro lens performs a slow, graceful 360-degree orbital pan around a gnarled tree branch, its dark, richly textured bark contrasting with the scene's metamorphosis. Initially, the branch is completely covered in an intricate web of needle-like hoarfrost crystals, each delicate spire capturing the chilly ambient light of the predawn forest, rendered in a soft, out-of-focus bokeh with deep blues and gentle purples in the background. As the camera continues its circular path, the first warm golden rays of rising sun catch the branch, and the action begins: the sharp edges of the frost imperceptibly soften and begin to liquefy. This is no simple melt; the ice directly coalesces into distinctive, heavy droplets with high surface tension and the brilliant silvery sheen of liquid mercury. These shimmering, metallic-looking spheres tremble and expand, clinging to the crystalline structures before detaching and beginning to deliberately glide along the branch's contours, capturing and distorting the golden sunlight with their reflective surfaces, creating a surreal, magically quiet, elementally transformed atmosphere.", + "seed": 15772, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_9/index_8930_seed_15772_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09267686605453491, + "vlm_logprob": -0.09267686605453491, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.09267686605453491, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012249006889760494 + ], + "opposite_logprob_tokens": [ + -5.637248992919922 + ], + "logprob_reward": -0.012249006889760494, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.012249006889760494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16169656813144684, + 0.060031019151210785, + 0.0918375700712204, + 0.07733669877052307, + 0.21573269367218018, + 0.13712166249752045, + 0.038374003022909164, + 0.06929397583007812, + 0.039682503789663315, + 0.10889335721731186 + ], + "credit_map": [ + 0.02553677000105381, + 0.03747066855430603, + 0.009696024470031261, + 0.018382059410214424, + 0.02109929732978344, + 0.02745479717850685, + 0.011243488639593124, + 0.010813465341925621, + 0.01100130658596754, + 0.008181855082511902, + 0.005464114248752594, + 0.007629316300153732, + 0.006485197693109512, + 0.007311156019568443, + 0.006315239239484072, + 0.007642833050340414, + 0.011092642322182655, + 0.017271019518375397, + 0.009120938368141651, + 0.00853876955807209, + 0.011227769777178764, + 0.014934316277503967, + 0.011725078336894512, + 0.007927040569484234, + 0.011530064977705479, + 0.014421607367694378, + 0.004939214792102575, + 0.0036937135737389326, + 0.020498421043157578, + 0.009201673790812492, + 0.0068959626369178295, + 0.0061560324393212795, + 0.06179969012737274, + 0.07004069536924362, + 0.010637566447257996, + 0.009633393958210945, + 0.017568519338965416, + 0.01777253858745098, + 0.011373789049685001, + 0.016906507313251495, + 0.015627454966306686, + 0.01567089930176735, + 0.011017818935215473, + 0.00704248808324337, + 0.016332868486642838, + 0.025324704125523567, + 0.019177639856934547, + 0.026927797123789787, + 0.007601728197187185, + 0.005988894961774349, + 0.004475104156881571, + 0.002621613210067153, + 0.004002001136541367, + 0.004577249754220247, + 0.002539300126954913, + 0.006568110082298517, + 0.012195589020848274, + 0.014646606519818306, + 0.004680505953729153, + 0.005575215443968773, + 0.009073232300579548, + 0.011408599093556404, + 0.004550413694232702, + 0.0071638161316514015, + 0.004117067437618971, + 0.0032552520278841257, + 0.005010138265788555, + 0.004153622314333916, + 0.005777993239462376, + 0.005552948918193579, + 0.0047360933385789394, + 0.007079384755343199, + 0.009601296856999397, + 0.015110335312783718, + 0.007449095603078604, + 0.007068850100040436, + 0.03046766296029091, + 0.00889201182872057, + 0.00983954407274723, + 0.02046455629169941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.012249006889760494 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370732069015503 + ], + "opposite_logprob_tokens": [ + -3.4087073802948 + ], + "logprob_reward": -0.03370732069015503, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370732069015503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1051267683506012, + 0.04550928249955177, + 0.051962509751319885, + 0.05854180082678795, + 0.38616636395454407, + 0.061769288033246994, + 0.06374628096818924, + 0.13903683423995972, + 0.05560826510190964, + 0.03253268450498581 + ], + "credit_map": [ + 0.013056871481239796, + 0.019138073548674583, + 0.007899091579020023, + 0.013160545378923416, + 0.013543346896767616, + 0.020761484280228615, + 0.01048422884196043, + 0.007083128206431866, + 0.006308127660304308, + 0.006069050636142492, + 0.004023383371531963, + 0.004899848718196154, + 0.00554178049787879, + 0.005852744914591312, + 0.006898591760545969, + 0.005915749352425337, + 0.007410286460071802, + 0.005597085691988468, + 0.005745190195739269, + 0.005032571963965893, + 0.0085183996707201, + 0.007023541256785393, + 0.007752535864710808, + 0.0048828949220478535, + 0.010400582104921341, + 0.009265358559787273, + 0.0038389465771615505, + 0.0035569530446082354, + 0.01604558154940605, + 0.007202659733593464, + 0.003798059420660138, + 0.004433660302311182, + 0.0957983136177063, + 0.09921614825725555, + 0.02983318828046322, + 0.029126722365617752, + 0.03663179650902748, + 0.049383148550987244, + 0.02288992330431938, + 0.023287121206521988, + 0.006447250954806805, + 0.005396484863013029, + 0.003739959327504039, + 0.00410130200907588, + 0.012108489871025085, + 0.009201360866427422, + 0.006223293021321297, + 0.01455114409327507, + 0.010272151790559292, + 0.009701631963253021, + 0.005147678777575493, + 0.0056641544215381145, + 0.01161533035337925, + 0.009200110100209713, + 0.004099475219845772, + 0.00804575253278017, + 0.04196420684456825, + 0.019873574376106262, + 0.009466418996453285, + 0.008522182703018188, + 0.012960013002157211, + 0.0249103382229805, + 0.007976296357810497, + 0.013363801874220371, + 0.005959667731076479, + 0.00485619343817234, + 0.004453765694051981, + 0.004657748155295849, + 0.01760704629123211, + 0.010299216955900192, + 0.003853799542412162, + 0.003920829389244318, + 0.004354385659098625, + 0.004500872455537319, + 0.003864711383357644, + 0.003272206522524357, + 0.004270769655704498, + 0.005380250513553619, + 0.0033715013414621353, + 0.0035179837141186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.03370732069015503 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585225895047188 + ], + "opposite_logprob_tokens": [ + -3.895585298538208 + ], + "logprob_reward": -0.020585225895047188, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585225895047188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15574094653129578, + 0.062418386340141296, + 0.07016954571008682, + 0.07997292280197144, + 0.27448025345802307, + 0.07706649601459503, + 0.05327219516038895, + 0.05676468089222908, + 0.0859939306974411, + 0.08412058651447296 + ], + "credit_map": [ + 0.021136755123734474, + 0.021258367225527763, + 0.013338272459805012, + 0.021674053743481636, + 0.014233535155653954, + 0.01886235736310482, + 0.02915603667497635, + 0.016081565991044044, + 0.008233639411628246, + 0.007736906409263611, + 0.0065069058910012245, + 0.007284628227353096, + 0.006279791705310345, + 0.005869712680578232, + 0.012091832235455513, + 0.00841496791690588, + 0.012362268753349781, + 0.008768502622842789, + 0.009060157462954521, + 0.006316924002021551, + 0.010295748710632324, + 0.008818859234452248, + 0.008896494284272194, + 0.005650587845593691, + 0.018106693401932716, + 0.0098705580458045, + 0.003689706092700362, + 0.003680594265460968, + 0.02761729806661606, + 0.008008425123989582, + 0.0048325685784220695, + 0.0041670785285532475, + 0.08383532613515854, + 0.08997005969285965, + 0.01786688342690468, + 0.012061486952006817, + 0.021297791972756386, + 0.02155221626162529, + 0.013652581721544266, + 0.014243945479393005, + 0.009307037107646465, + 0.010026597417891026, + 0.005180492531508207, + 0.008283027447760105, + 0.010397990234196186, + 0.0162340234965086, + 0.005843580700457096, + 0.01179374847561121, + 0.009629987180233002, + 0.010764775797724724, + 0.004538864362984896, + 0.004743917845189571, + 0.006395767908543348, + 0.005931079853326082, + 0.004620130639523268, + 0.006647671107202768, + 0.01172034814953804, + 0.01246636826545, + 0.003802079474553466, + 0.005592191591858864, + 0.007477689068764448, + 0.0071328552439808846, + 0.004155534785240889, + 0.004417614080011845, + 0.006951986346393824, + 0.007463688496500254, + 0.007842780090868473, + 0.0051973494701087475, + 0.028727712109684944, + 0.013219269923865795, + 0.005991989281028509, + 0.010599159635603428, + 0.007417980115860701, + 0.010953936725854874, + 0.006535366643220186, + 0.006378495134413242, + 0.017379309982061386, + 0.01255048532038927, + 0.0105519387871027, + 0.012353070080280304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.020585225895047188 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783296845853329 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783296845853329, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783296845853329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3291268050670624, + 0.048115186393260956, + 0.03226766362786293, + 0.11383895576000214, + 0.06721881031990051, + 0.11389878392219543, + 0.09603019803762436, + 0.0358930267393589, + 0.07501830160617828, + 0.0885922983288765 + ], + "credit_map": [ + 0.0477278009057045, + 0.04419394209980965, + 0.02961394563317299, + 0.05208747833967209, + 0.028807150200009346, + 0.032281771302223206, + 0.06065624952316284, + 0.033758439123630524, + 0.006306931842118502, + 0.007402186281979084, + 0.004880402237176895, + 0.005199247971177101, + 0.0038589739706367254, + 0.006220971699804068, + 0.006757519207894802, + 0.007488952483981848, + 0.005611588712781668, + 0.005742581095546484, + 0.0028118763584643602, + 0.0028373717796057463, + 0.006201409734785557, + 0.0030039153061807156, + 0.00407507224008441, + 0.0019838481675833464, + 0.02727777510881424, + 0.019490547478199005, + 0.00682305870577693, + 0.007596895098686218, + 0.02092253975570202, + 0.012610145844519138, + 0.008251720108091831, + 0.010866272263228893, + 0.011570080183446407, + 0.009741517715156078, + 0.0087567875161767, + 0.0069234357215464115, + 0.00792722124606371, + 0.009450671263039112, + 0.006385703571140766, + 0.006463395431637764, + 0.015027150511741638, + 0.013603984378278255, + 0.00867066066712141, + 0.009291961789131165, + 0.011982908472418785, + 0.011715903878211975, + 0.011386309750378132, + 0.03221990540623665, + 0.015490061603486538, + 0.01642351783812046, + 0.007264331914484501, + 0.0069917943328619, + 0.019842607900500298, + 0.010233337059617043, + 0.008037774823606014, + 0.011746766977012157, + 0.006595872342586517, + 0.007432010490447283, + 0.002307719085365534, + 0.002796592889353633, + 0.004923038184642792, + 0.004476166795939207, + 0.002912544645369053, + 0.00444908207282424, + 0.0064995684660971165, + 0.005901040509343147, + 0.006019915454089642, + 0.011057347059249878, + 0.018376724794507027, + 0.010467258282005787, + 0.00849562045186758, + 0.008200823329389095, + 0.015513468533754349, + 0.0135919488966465, + 0.011329388245940208, + 0.0057259248569607735, + 0.011824790388345718, + 0.011077797040343285, + 0.009113699197769165, + 0.010415276512503624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.009783296845853329 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870594799518585 + ], + "opposite_logprob_tokens": [ + -1.1370594501495361 + ], + "logprob_reward": -0.3870594799518585, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870594799518585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07857133448123932, + 0.03362816944718361, + 0.045800525695085526, + 0.0946655422449112, + 0.2708373963832855, + 0.08497681468725204, + 0.021425362676382065, + 0.09485862404108047, + 0.16891731321811676, + 0.10631883889436722 + ], + "credit_map": [ + 0.011301923543214798, + 0.01638246886432171, + 0.005168924108147621, + 0.00997068639844656, + 0.010100137442350388, + 0.012278033420443535, + 0.00753485644236207, + 0.005834301933646202, + 0.004830487072467804, + 0.005295664072036743, + 0.0035158914979547262, + 0.0027581476606428623, + 0.0033349613659083843, + 0.004385479260236025, + 0.004558212123811245, + 0.004949324764311314, + 0.005608470179140568, + 0.009137448854744434, + 0.00454069534316659, + 0.00540779996663332, + 0.00396676454693079, + 0.008988860063254833, + 0.005151594989001751, + 0.0029988905880600214, + 0.02110859751701355, + 0.016787724569439888, + 0.006444217637181282, + 0.008017444983124733, + 0.014283830299973488, + 0.010510933585464954, + 0.007167687173932791, + 0.01034510601311922, + 0.06699561327695847, + 0.06577825546264648, + 0.01895149238407612, + 0.024253785610198975, + 0.026700273156166077, + 0.018352573737502098, + 0.028189795091748238, + 0.021615618839859962, + 0.01095091924071312, + 0.010682777501642704, + 0.007633914705365896, + 0.009072119370102882, + 0.010652028024196625, + 0.008174263872206211, + 0.018004726618528366, + 0.009806067682802677, + 0.004465920850634575, + 0.0028124696109443903, + 0.0021119152661412954, + 0.002038254402577877, + 0.002781383926048875, + 0.0030530705116689205, + 0.002042980631813407, + 0.002119366778060794, + 0.02243305929005146, + 0.019896049052476883, + 0.006375256460160017, + 0.005149866919964552, + 0.012944771908223629, + 0.012865559197962284, + 0.005691852420568466, + 0.009502215310931206, + 0.013418362475931644, + 0.015281747095286846, + 0.009021160192787647, + 0.007574248593300581, + 0.05947515740990639, + 0.03326629847288132, + 0.010027972050011158, + 0.020852360874414444, + 0.02077360637485981, + 0.01747320406138897, + 0.011492901481688023, + 0.007333171553909779, + 0.014935067854821682, + 0.012129535898566246, + 0.011616828851401806, + 0.010564529336988926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.3870594799518585 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro lens performs a slow, graceful 360-degree orbital pan around a gnarled tree branch, its dark, richly textured bark contrasting with the scene's metamorphosis. Initially, the branch is completely covered in an intricate web of needle-like hoarfrost crystals, each delicate spire capturing the chilly ambient light of the predawn forest, rendered in a soft, out-of-focus bokeh with deep blues and gentle purples in the background. As the camera continues its circular path, the first warm golden rays of rising sun catch the branch, and the action begins: the sharp edges of the frost imperceptibly soften and begin to liquefy. This is no simple melt; the ice directly coalesces into distinctive, heavy droplets with high surface tension and the brilliant silvery sheen of liquid mercury. These shimmering, metallic-looking spheres tremble and expand, clinging to the crystalline structures before detaching and beginning to deliberately glide along the branch's contours, capturing and distorting the golden sunlight with their reflective surfaces, creating a surreal, magically quiet, elementally transformed atmosphere.", + "seed": 15773, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_9/index_8930_seed_15773_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09698464889079332, + "vlm_logprob": -0.09698464889079332, + "vlm_yes_no_margin": 3.749999988079071, + "vlm_token_credit": -0.09698464889079332, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010575443506240845 + ], + "opposite_logprob_tokens": [ + -5.385575294494629 + ], + "logprob_reward": -0.010575443506240845, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.010575443506240845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05686076730489731, + 0.09138941019773483, + 0.09746947139501572, + 0.48070526123046875, + 0.02867773361504078, + 0.014162912033498287, + 0.130901500582695, + 0.02884678542613983, + 0.03676539286971092, + 0.03422076255083084 + ], + "credit_map": [ + 0.0063903252594172955, + 0.01099871750921011, + 0.004355937242507935, + 0.005562303587794304, + 0.004573512822389603, + 0.004626980982720852, + 0.0073238336481153965, + 0.013029160909354687, + 0.012981856241822243, + 0.02094906009733677, + 0.012127850204706192, + 0.00712974788621068, + 0.007007034029811621, + 0.005967934615910053, + 0.011727792210876942, + 0.013498129323124886, + 0.015105433762073517, + 0.015719763934612274, + 0.006122426595538855, + 0.005563313141465187, + 0.02005244605243206, + 0.011517221108078957, + 0.010368190705776215, + 0.013020675629377365, + 0.048686373978853226, + 0.03364946320652962, + 0.05642584338784218, + 0.05659583583474159, + 0.04604368284344673, + 0.03535320982336998, + 0.1114291101694107, + 0.09252175688743591, + 0.0032693014945834875, + 0.0048274933360517025, + 0.002547645242884755, + 0.0030089521314948797, + 0.0021753765176981688, + 0.002952287904918194, + 0.003998871892690659, + 0.005897806026041508, + 0.0012809342006221414, + 0.0012228057021275163, + 0.001034258515574038, + 0.0016615372151136398, + 0.00135145359672606, + 0.0016984861576929688, + 0.002404486760497093, + 0.0035089508164674044, + 0.005839818622916937, + 0.00996797438710928, + 0.005918616428971291, + 0.006734323687851429, + 0.00990239717066288, + 0.006759663112461567, + 0.04409118369221687, + 0.04168752580881119, + 0.0017274045385420322, + 0.0018705311231315136, + 0.001783937681466341, + 0.009125558659434319, + 0.003432384692132473, + 0.0024828186724334955, + 0.002168825827538967, + 0.006255324464291334, + 0.0034492439590394497, + 0.004897365812212229, + 0.002673245035111904, + 0.003469310700893402, + 0.0077440268360078335, + 0.0027925583999603987, + 0.00697906082496047, + 0.0047605810686945915, + 0.0021164570935070515, + 0.0019785258919000626, + 0.0015570423565804958, + 0.0031344317831099033, + 0.005939938127994537, + 0.0022982629016041756, + 0.010995456017553806, + 0.006200647912919521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.010575443506240845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297041654586792 + ], + "opposite_logprob_tokens": [ + -3.773297071456909 + ], + "logprob_reward": -0.023297041654586792, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297041654586792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042994946241378784, + 0.06916328519582748, + 0.0315687358379364, + 0.5129699110984802, + 0.018557926639914513, + 0.012728748843073845, + 0.16546864807605743, + 0.056177496910095215, + 0.06162639707326889, + 0.02874397486448288 + ], + "credit_map": [ + 0.006956937722861767, + 0.008284807205200195, + 0.004677452612668276, + 0.003342836396768689, + 0.005366353318095207, + 0.00416909996420145, + 0.005391798447817564, + 0.00480565894395113, + 0.008395491167902946, + 0.017752304673194885, + 0.009544278495013714, + 0.005218290723860264, + 0.005453657824546099, + 0.004552464000880718, + 0.009807179681956768, + 0.008439616300165653, + 0.003762737847864628, + 0.005759104620665312, + 0.0036813393235206604, + 0.003015163354575634, + 0.0039051754865795374, + 0.0029458014760166407, + 0.0038479738868772984, + 0.004651441238820553, + 0.04783930256962776, + 0.035280849784612656, + 0.055104441940784454, + 0.056724995374679565, + 0.048084557056427, + 0.03736847639083862, + 0.12449333071708679, + 0.10807392746210098, + 0.0014904545387253165, + 0.002712331945076585, + 0.0011813407763838768, + 0.0015227546682581306, + 0.0022818814031779766, + 0.002501121489331126, + 0.0026963348500430584, + 0.004171706736087799, + 0.0019127319101244211, + 0.0014287499943748116, + 0.001586745958775282, + 0.0015537010040134192, + 0.0009555495926178992, + 0.0012779745738953352, + 0.0017907596193253994, + 0.00222253636457026, + 0.011825961992144585, + 0.023802155628800392, + 0.01399585697799921, + 0.012330221943557262, + 0.02268705517053604, + 0.01588987559080124, + 0.032705266028642654, + 0.03223225474357605, + 0.0032314921263605356, + 0.004795044660568237, + 0.003206772729754448, + 0.01201673038303852, + 0.019114596769213676, + 0.003729067509993911, + 0.003905236255377531, + 0.00617855554446578, + 0.003913252614438534, + 0.006041291635483503, + 0.0035549872554838657, + 0.0067090499214828014, + 0.012914923951029778, + 0.005005231127142906, + 0.014376581646502018, + 0.009111077524721622, + 0.002121893921867013, + 0.0019880952313542366, + 0.0017415968468412757, + 0.002348418580368161, + 0.002835992258042097, + 0.004216525703668594, + 0.007428213953971863, + 0.0060632373206317425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.023297041654586792 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011083230376243591 + ], + "opposite_logprob_tokens": [ + -4.511083126068115 + ], + "logprob_reward": -0.011083230376243591, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011083230376243591, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045605942606925964, + 0.048594970256090164, + 0.053053900599479675, + 0.11495700478553772, + 0.03884897008538246, + 0.04989371821284294, + 0.3752112090587616, + 0.10661857575178146, + 0.08428119122982025, + 0.08293453603982925 + ], + "credit_map": [ + 0.0050521125085651875, + 0.005848584230989218, + 0.004467587918043137, + 0.004899287596344948, + 0.005072847940027714, + 0.0036467837635427713, + 0.005831747781485319, + 0.010786989703774452, + 0.005818353500217199, + 0.010506154038012028, + 0.006915534846484661, + 0.003925593104213476, + 0.005114791914820671, + 0.0033907215110957623, + 0.007409891579300165, + 0.005513928364962339, + 0.007254407275468111, + 0.011988718993961811, + 0.00400051474571228, + 0.004290338139981031, + 0.00605635903775692, + 0.0047419373877346516, + 0.007102820090949535, + 0.007618804927915335, + 0.01205056719481945, + 0.009116527624428272, + 0.010628432035446167, + 0.013662988319993019, + 0.013515236787497997, + 0.00962380412966013, + 0.025395214557647705, + 0.020964236930012703, + 0.003615101333707571, + 0.0052123842760920525, + 0.002267963020130992, + 0.0037178746424615383, + 0.003343586577102542, + 0.0045626903884112835, + 0.005826038308441639, + 0.010303334333002567, + 0.005426943767815828, + 0.0063436501659452915, + 0.0042194887064397335, + 0.00497348653152585, + 0.004486044868826866, + 0.00442954758182168, + 0.008165931329131126, + 0.011848625726997852, + 0.01761806569993496, + 0.03222142159938812, + 0.01857770048081875, + 0.020459992811083794, + 0.029212063178420067, + 0.021029746159911156, + 0.12203188985586166, + 0.11406032741069794, + 0.005363714415580034, + 0.007679336704313755, + 0.004893174394965172, + 0.035955820232629776, + 0.031095100566744804, + 0.005914729088544846, + 0.006727792322635651, + 0.008988908492028713, + 0.011559217236936092, + 0.013436785899102688, + 0.009191179648041725, + 0.007690787315368652, + 0.013296674937009811, + 0.00783834420144558, + 0.011348380707204342, + 0.009919813834130764, + 0.004733660724014044, + 0.006537547335028648, + 0.006700246594846249, + 0.005894533824175596, + 0.020176900550723076, + 0.009815003722906113, + 0.02051847241818905, + 0.008558169938623905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.011083230376243591 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075095273554325 + ], + "opposite_logprob_tokens": [ + -4.511075019836426 + ], + "logprob_reward": -0.011075095273554325, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075095273554325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03517483174800873, + 0.043714240193367004, + 0.028834786266088486, + 0.18631316721439362, + 0.1077795922756195, + 0.032768070697784424, + 0.30266374349594116, + 0.06200705096125603, + 0.041506558656692505, + 0.15923792123794556 + ], + "credit_map": [ + 0.005635668523609638, + 0.006466774735599756, + 0.004318661522120237, + 0.003762730397284031, + 0.0035629761405289173, + 0.003087467048317194, + 0.0041629113256931305, + 0.004177640657871962, + 0.006028621923178434, + 0.009869775734841824, + 0.006507283542305231, + 0.0034575301688164473, + 0.0034877366852015257, + 0.002672831527888775, + 0.005631560925394297, + 0.006058899685740471, + 0.0034157983027398586, + 0.004999184049665928, + 0.0030801764223724604, + 0.0031169787980616093, + 0.0028514398727566004, + 0.0025930593255907297, + 0.0037957036402076483, + 0.004982445854693651, + 0.021842600777745247, + 0.014063127338886261, + 0.02053176797926426, + 0.023886142298579216, + 0.02039523795247078, + 0.014606783166527748, + 0.03978918865323067, + 0.031198324635624886, + 0.009310864843428135, + 0.014245443046092987, + 0.006037174258381128, + 0.011045928113162518, + 0.0065462118946015835, + 0.010014571249485016, + 0.017716199159622192, + 0.032863203436136246, + 0.003991652280092239, + 0.0037748736795037985, + 0.003943671938031912, + 0.00383862410672009, + 0.0032216254621744156, + 0.0032826419919729233, + 0.004891769494861364, + 0.005823210347443819, + 0.02237175777554512, + 0.036957163363695145, + 0.021909382194280624, + 0.02224595658481121, + 0.031872399151325226, + 0.020157108083367348, + 0.07271706312894821, + 0.07443290948867798, + 0.004211134742945433, + 0.006277389358729124, + 0.004315040539950132, + 0.014643128030002117, + 0.01878947578370571, + 0.005096700042486191, + 0.004025844857096672, + 0.004648336209356785, + 0.0041487314738333225, + 0.003807901870459318, + 0.003582151373848319, + 0.004231965634971857, + 0.007772891782224178, + 0.0031169112771749496, + 0.01036113128066063, + 0.004484873730689287, + 0.011601592414081097, + 0.007494232151657343, + 0.008669593371450901, + 0.010658658109605312, + 0.016278138384222984, + 0.01388157345354557, + 0.05309673026204109, + 0.037557411938905716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.011075095273554325 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42889243364334106 + ], + "opposite_logprob_tokens": [ + -1.0538923740386963 + ], + "logprob_reward": -0.42889243364334106, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42889243364334106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0214496199041605, + 0.022888237610459328, + 0.0282465647906065, + 0.46184277534484863, + 0.033845607191324234, + 0.015912573784589767, + 0.10904419422149658, + 0.08711986243724823, + 0.06473397463560104, + 0.1549166589975357 + ], + "credit_map": [ + 0.002933258656412363, + 0.0048318360932171345, + 0.0025682731065899134, + 0.002442582044750452, + 0.00231828517280519, + 0.001741169486194849, + 0.002602681517601013, + 0.002011534757912159, + 0.0029784447979182005, + 0.005590954329818487, + 0.0029130871407687664, + 0.0018411322962492704, + 0.0022625187411904335, + 0.0013554490869864821, + 0.0027110327500849962, + 0.0032356197480112314, + 0.006869512144476175, + 0.005681448150426149, + 0.0022231489419937134, + 0.002334711840376258, + 0.002238266170024872, + 0.001724669593386352, + 0.0035835218150168657, + 0.0035912867169827223, + 0.04598132148385048, + 0.0325804129242897, + 0.05095847323536873, + 0.05366892367601395, + 0.04540957883000374, + 0.03419768437743187, + 0.10804639756679535, + 0.09099997580051422, + 0.003508098190650344, + 0.005932875443249941, + 0.0021628986578434706, + 0.003156218910589814, + 0.0039023433346301317, + 0.0024733985774219036, + 0.005786879453808069, + 0.006922894157469273, + 0.0015436828834936023, + 0.0019259381806477904, + 0.0018787195440381765, + 0.001776109915226698, + 0.0019637178629636765, + 0.00199430831708014, + 0.00220008403994143, + 0.0026300139725208282, + 0.006917118560522795, + 0.014472899958491325, + 0.008278118446469307, + 0.00789066031575203, + 0.013466437347233295, + 0.009171850048005581, + 0.027146760374307632, + 0.02170034684240818, + 0.00595626188442111, + 0.0075837392359972, + 0.006537232547998428, + 0.010023768059909344, + 0.02126779779791832, + 0.0067964536137878895, + 0.006676127668470144, + 0.02227848395705223, + 0.005869607906788588, + 0.007195326499640942, + 0.004041973501443863, + 0.006309677846729755, + 0.01153022050857544, + 0.004816111177206039, + 0.01680167391896248, + 0.0081693846732378, + 0.011128159239888191, + 0.00625914940610528, + 0.007652901578694582, + 0.009862639009952545, + 0.014531200751662254, + 0.010790382511913776, + 0.05457692593336105, + 0.040115296840667725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.42889243364334106 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro lens performs a slow, graceful 360-degree orbital pan around a gnarled tree branch, its dark, richly textured bark contrasting with the scene's metamorphosis. Initially, the branch is completely covered in an intricate web of needle-like hoarfrost crystals, each delicate spire capturing the chilly ambient light of the predawn forest, rendered in a soft, out-of-focus bokeh with deep blues and gentle purples in the background. As the camera continues its circular path, the first warm golden rays of rising sun catch the branch, and the action begins: the sharp edges of the frost imperceptibly soften and begin to liquefy. This is no simple melt; the ice directly coalesces into distinctive, heavy droplets with high surface tension and the brilliant silvery sheen of liquid mercury. These shimmering, metallic-looking spheres tremble and expand, clinging to the crystalline structures before detaching and beginning to deliberately glide along the branch's contours, capturing and distorting the golden sunlight with their reflective surfaces, creating a surreal, magically quiet, elementally transformed atmosphere.", + "seed": 15774, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_9/index_8930_seed_15774_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21180806569755078, + "vlm_logprob": -0.21180806569755078, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.21180806569755078, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014355714432895184 + ], + "opposite_logprob_tokens": [ + -4.889355659484863 + ], + "logprob_reward": -0.014355714432895184, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.014355714432895184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21683210134506226, + 0.07523217797279358, + 0.07485847175121307, + 0.06841862946748734, + 0.04695438593626022, + 0.040858715772628784, + 0.059942249208688736, + 0.18359707295894623, + 0.04836249351501465, + 0.18494370579719543 + ], + "credit_map": [ + 0.024937449023127556, + 0.02443336695432663, + 0.02420208230614662, + 0.021303651854395866, + 0.03675592690706253, + 0.03544803708791733, + 0.020865658298134804, + 0.02888590656220913, + 0.011280737817287445, + 0.008712242357432842, + 0.008358635008335114, + 0.009294593706727028, + 0.009100680239498615, + 0.008756092749536037, + 0.007580497767776251, + 0.01214869786053896, + 0.01030424702912569, + 0.006775314453989267, + 0.011799704283475876, + 0.007886099629104137, + 0.010922725312411785, + 0.007201784756034613, + 0.010920575819909573, + 0.00904801394790411, + 0.01711486279964447, + 0.010571485385298729, + 0.010495978407561779, + 0.007012159563601017, + 0.0064199757762253284, + 0.006911578122526407, + 0.0058636791072785854, + 0.004028903786092997, + 0.005319403018802404, + 0.004745578859001398, + 0.007505523972213268, + 0.005767586641013622, + 0.006532297469675541, + 0.005894436500966549, + 0.0061883265152573586, + 0.005001229699701071, + 0.003723355708643794, + 0.0043363673612475395, + 0.008030148223042488, + 0.005454178899526596, + 0.006230488885194063, + 0.005178998224437237, + 0.005061783362179995, + 0.002843392314389348, + 0.007109584752470255, + 0.007787192240357399, + 0.008127506822347641, + 0.006800311151891947, + 0.01319175772368908, + 0.006320632062852383, + 0.006138731725513935, + 0.0044665285386145115, + 0.017202043905854225, + 0.023190272971987724, + 0.020801851525902748, + 0.03984735533595085, + 0.012326502241194248, + 0.019870363175868988, + 0.029685484245419502, + 0.020673174411058426, + 0.006984005682170391, + 0.007453486323356628, + 0.006120316218584776, + 0.006746189668774605, + 0.005452280398458242, + 0.004306428600102663, + 0.006950443610548973, + 0.004349337425082922, + 0.012234718538820744, + 0.05776926130056381, + 0.011402036063373089, + 0.009068700484931469, + 0.05919899046421051, + 0.015606113709509373, + 0.010712855495512486, + 0.008951027877628803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.014355714432895184 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906170725822449 + ], + "opposite_logprob_tokens": [ + -2.464061737060547 + ], + "logprob_reward": -0.08906170725822449, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906170725822449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29894861578941345, + 0.07161842286586761, + 0.025540055707097054, + 0.04115176945924759, + 0.03635016828775406, + 0.030639098957180977, + 0.04661818593740463, + 0.1169988363981247, + 0.05082133784890175, + 0.2813134789466858 + ], + "credit_map": [ + 0.03313660994172096, + 0.032335445284843445, + 0.03177318722009659, + 0.028601843863725662, + 0.05528128519654274, + 0.04612063989043236, + 0.030488375574350357, + 0.04121124744415283, + 0.009152264334261417, + 0.007604519370943308, + 0.006345062050968409, + 0.0096003208309412, + 0.011028734035789967, + 0.010781466960906982, + 0.007836665026843548, + 0.009269394911825657, + 0.003730227006599307, + 0.0019928868860006332, + 0.0032013561576604843, + 0.0026354161091148853, + 0.002853562356904149, + 0.004578223451972008, + 0.0022714489605277777, + 0.0042769345454871655, + 0.010244794189929962, + 0.007589607499539852, + 0.005266114603728056, + 0.0034864689223468304, + 0.004319012630730867, + 0.0035184493754059076, + 0.0033276863396167755, + 0.0033996407873928547, + 0.003936486784368753, + 0.006692794151604176, + 0.004122445825487375, + 0.003520561847835779, + 0.0032769099343568087, + 0.0033380785025656223, + 0.0038463196251541376, + 0.007616573479026556, + 0.002963925478979945, + 0.003422340378165245, + 0.003540694247931242, + 0.006162288598716259, + 0.003876147558912635, + 0.003575510112568736, + 0.004860797896981239, + 0.0022373958490788937, + 0.005255366675555706, + 0.006606955546885729, + 0.005218306090682745, + 0.00895214919000864, + 0.008551910519599915, + 0.0038003879599273205, + 0.004772381391376257, + 0.003460729494690895, + 0.014394099824130535, + 0.01311138179153204, + 0.019279232248663902, + 0.02374623715877533, + 0.01618342287838459, + 0.008737520314753056, + 0.013231867924332619, + 0.008315082639455795, + 0.00768011249601841, + 0.006429024040699005, + 0.004280796740204096, + 0.012177566066384315, + 0.004590384662151337, + 0.004378890618681908, + 0.007306176237761974, + 0.003978388849645853, + 0.04641770198941231, + 0.05377672240138054, + 0.03905443474650383, + 0.02208046242594719, + 0.04631264507770538, + 0.023775314912199974, + 0.02844543568789959, + 0.02145078033208847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.08906170725822449 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430726483464241 + ], + "opposite_logprob_tokens": [ + -3.1680727005004883 + ], + "logprob_reward": -0.0430726483464241, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430726483464241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1994381546974182, + 0.11033700406551361, + 0.0412546843290329, + 0.055791206657886505, + 0.04335486888885498, + 0.0270476583391428, + 0.03539983555674553, + 0.16903309524059296, + 0.056625742465257645, + 0.2617177665233612 + ], + "credit_map": [ + 0.02064821869134903, + 0.022821135818958282, + 0.03015747293829918, + 0.01872285269200802, + 0.03606928139925003, + 0.027456795796751976, + 0.02049642987549305, + 0.023065980523824692, + 0.014796082861721516, + 0.013013741932809353, + 0.01562268752604723, + 0.010846414603292942, + 0.017270969226956367, + 0.01587638258934021, + 0.009989619255065918, + 0.012921107932925224, + 0.00792376697063446, + 0.0036435462534427643, + 0.006332959979772568, + 0.0032303412444889545, + 0.0050156512297689915, + 0.005008208565413952, + 0.006511050276458263, + 0.0035891609732061625, + 0.011827362701296806, + 0.010073966346681118, + 0.006069785449653864, + 0.005085982847958803, + 0.006048805546015501, + 0.0039873807691037655, + 0.00737285241484642, + 0.005325069185346365, + 0.004523808136582375, + 0.004725380800664425, + 0.006209179759025574, + 0.005765746347606182, + 0.0049766795709729195, + 0.003926085773855448, + 0.006810174323618412, + 0.006417809519916773, + 0.0022190609015524387, + 0.0027782463002949953, + 0.003547642845660448, + 0.00561492471024394, + 0.0035825511440634727, + 0.002323702210560441, + 0.004986278247088194, + 0.0019952526781708, + 0.004134125541895628, + 0.004381276201456785, + 0.003395724343135953, + 0.0037502527702599764, + 0.007608032785356045, + 0.004790085833519697, + 0.003291812026873231, + 0.004048526752740145, + 0.021165631711483, + 0.019972357898950577, + 0.024816425517201424, + 0.034683384001255035, + 0.02287925034761429, + 0.011953816749155521, + 0.019827665761113167, + 0.013734561391174793, + 0.007803040556609631, + 0.010278645902872086, + 0.005402808543294668, + 0.007463976740837097, + 0.006667113397270441, + 0.0066129835322499275, + 0.0062681143172085285, + 0.006129059940576553, + 0.03957855701446533, + 0.06584115326404572, + 0.022165944799780846, + 0.011800113134086132, + 0.060962602496147156, + 0.021637538447976112, + 0.017149459570646286, + 0.022582409903407097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.0430726483464241 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207052998244762 + ], + "opposite_logprob_tokens": [ + -4.264206886291504 + ], + "logprob_reward": -0.014207052998244762, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207052998244762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23313654959201813, + 0.045119281858205795, + 0.04937991499900818, + 0.07554333657026291, + 0.0335424430668354, + 0.023886526003479958, + 0.08723749220371246, + 0.06804876029491425, + 0.02879582718014717, + 0.3553098738193512 + ], + "credit_map": [ + 0.02583862654864788, + 0.040595632046461105, + 0.02790476754307747, + 0.02652668207883835, + 0.044978003948926926, + 0.02425423450767994, + 0.023438183590769768, + 0.01960042491555214, + 0.005582389887422323, + 0.004975018557161093, + 0.00519654480740428, + 0.0045869192108511925, + 0.004666426219046116, + 0.007137907203286886, + 0.005368023179471493, + 0.00760605139657855, + 0.005969048943370581, + 0.004450046457350254, + 0.009758645668625832, + 0.004688997287303209, + 0.008633172139525414, + 0.0066613503731787205, + 0.004370012786239386, + 0.0048486399464309216, + 0.026400728151202202, + 0.018799280747771263, + 0.006448057480156422, + 0.004155834671109915, + 0.0066176182590425014, + 0.0045695118606090546, + 0.004975923337042332, + 0.003576379967853427, + 0.0036686304956674576, + 0.00469425180926919, + 0.004936139099299908, + 0.004774283152073622, + 0.0041327704675495625, + 0.0041793109849095345, + 0.0037173328455537558, + 0.0034397251438349485, + 0.0035083612892776728, + 0.002811784390360117, + 0.0037231838796287775, + 0.0025981201324611902, + 0.003266814397647977, + 0.002738435985520482, + 0.002805556170642376, + 0.0024342695251107216, + 0.012490971013903618, + 0.012067759409546852, + 0.010263313539326191, + 0.019939199090003967, + 0.009120691567659378, + 0.005139532033354044, + 0.011692619882524014, + 0.006523407530039549, + 0.008081458508968353, + 0.007115177810192108, + 0.012927812524139881, + 0.014817499555647373, + 0.006346153095364571, + 0.00599629245698452, + 0.007344408892095089, + 0.00541995232924819, + 0.004967167507857084, + 0.005159706808626652, + 0.002560951514169574, + 0.002381247468292713, + 0.003756950842216611, + 0.0032898206263780594, + 0.003094540676102042, + 0.003585442202165723, + 0.0258100014179945, + 0.12053173035383224, + 0.0224728025496006, + 0.01408811192959547, + 0.11436467617750168, + 0.026378631591796875, + 0.017821630463004112, + 0.013842293061316013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.014207052998244762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983432054519653 + ], + "opposite_logprob_tokens": [ + -0.5233432054519653 + ], + "logprob_reward": -0.8983432054519653, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983432054519653, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20050527155399323, + 0.11198338866233826, + 0.05675327777862549, + 0.05507533997297287, + 0.02724991738796234, + 0.02300330065190792, + 0.11968494951725006, + 0.07985804975032806, + 0.0342085026204586, + 0.29167798161506653 + ], + "credit_map": [ + 0.022427478805184364, + 0.03207707405090332, + 0.0244703758507967, + 0.02049623429775238, + 0.040766410529613495, + 0.01665579527616501, + 0.02011386677622795, + 0.023498037829995155, + 0.015540733002126217, + 0.013786855153739452, + 0.008996591903269291, + 0.011236974969506264, + 0.021793538704514503, + 0.01863807998597622, + 0.01072426326572895, + 0.011266343295574188, + 0.008191517554223537, + 0.004368885885924101, + 0.009791126474738121, + 0.005972062703222036, + 0.008164306171238422, + 0.007919440045952797, + 0.006343408953398466, + 0.006002526730298996, + 0.014322483912110329, + 0.011403290554881096, + 0.006266893818974495, + 0.005905480589717627, + 0.005090640392154455, + 0.00457432447001338, + 0.004310884978622198, + 0.0032013379968702793, + 0.0022072670981287956, + 0.0031783776357769966, + 0.004351340234279633, + 0.004124676343053579, + 0.003491785144433379, + 0.0034253757912665606, + 0.0032059489749372005, + 0.0032651452347636223, + 0.0030853128992021084, + 0.0023296354338526726, + 0.002877857768908143, + 0.004948830232024193, + 0.0022240593098104, + 0.002353980438783765, + 0.0031551604624837637, + 0.0020284648053348064, + 0.017810413613915443, + 0.017896711826324463, + 0.01402641274034977, + 0.02537974715232849, + 0.012609850615262985, + 0.007235516328364611, + 0.015007095411419868, + 0.009719204157590866, + 0.009366772137582302, + 0.00918496958911419, + 0.013177251443266869, + 0.02008829638361931, + 0.00842560175806284, + 0.004970547277480364, + 0.008862880058586597, + 0.005781729239970446, + 0.005760669242590666, + 0.005531101021915674, + 0.0038751689717173576, + 0.002478111069649458, + 0.003915401175618172, + 0.0037318940740078688, + 0.0038239555433392525, + 0.005092201288789511, + 0.031984079629182816, + 0.0763181820511818, + 0.03189053386449814, + 0.017999127507209778, + 0.06942339986562729, + 0.01931019499897957, + 0.02392839640378952, + 0.020824044942855835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.8983432054519653 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "An extreme macro lens performs a slow, graceful 360-degree orbital pan around a gnarled tree branch, its dark, richly textured bark contrasting with the scene's metamorphosis. Initially, the branch is completely covered in an intricate web of needle-like hoarfrost crystals, each delicate spire capturing the chilly ambient light of the predawn forest, rendered in a soft, out-of-focus bokeh with deep blues and gentle purples in the background. As the camera continues its circular path, the first warm golden rays of rising sun catch the branch, and the action begins: the sharp edges of the frost imperceptibly soften and begin to liquefy. This is no simple melt; the ice directly coalesces into distinctive, heavy droplets with high surface tension and the brilliant silvery sheen of liquid mercury. These shimmering, metallic-looking spheres tremble and expand, clinging to the crystalline structures before detaching and beginning to deliberately glide along the branch's contours, capturing and distorting the golden sunlight with their reflective surfaces, creating a surreal, magically quiet, elementally transformed atmosphere.", + "seed": 15775, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000031/group_9/index_8930_seed_15775_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12898167334496974, + "vlm_logprob": -0.12898167334496974, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.12898167334496974, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012875283136963844 + ], + "opposite_logprob_tokens": [ + -5.262875080108643 + ], + "logprob_reward": -0.012875283136963844, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.012875283136963844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038285814225673676, + 0.055215317755937576, + 0.1883898675441742, + 0.03138066455721855, + 0.019235240295529366, + 0.3812311291694641, + 0.040872640907764435, + 0.08348633348941803, + 0.0468965545296669, + 0.1150064542889595 + ], + "credit_map": [ + 0.007605069316923618, + 0.004242082592099905, + 0.005086594261229038, + 0.004006400238722563, + 0.006468956358730793, + 0.0033033895306289196, + 0.003728947602212429, + 0.0038443775847554207, + 0.005546044558286667, + 0.01873094029724598, + 0.009051977656781673, + 0.005718207452446222, + 0.006460804957896471, + 0.004731717519462109, + 0.0025783537421375513, + 0.002397270640358329, + 0.008900139480829239, + 0.004858342465013266, + 0.07354598492383957, + 0.02146107517182827, + 0.007061136420816183, + 0.0033123758621513844, + 0.060361869633197784, + 0.008888940326869488, + 0.0024535278789699078, + 0.0028074271976947784, + 0.004612089600414038, + 0.0092666270211339, + 0.0015875933459028602, + 0.0019469545222818851, + 0.0051243784837424755, + 0.0035820621997117996, + 0.0027542272582650185, + 0.0029771090485155582, + 0.0023343581706285477, + 0.0021944702602922916, + 0.001862746081314981, + 0.001841706340201199, + 0.0024275798350572586, + 0.0028430423699319363, + 0.03334863483905792, + 0.03666916862130165, + 0.06565000116825104, + 0.07382301986217499, + 0.03392501547932625, + 0.030412251129746437, + 0.077422134578228, + 0.02998090721666813, + 0.006714798044413328, + 0.0035180088598281145, + 0.006760819349437952, + 0.005510176997631788, + 0.004895708989351988, + 0.004677327815443277, + 0.004946503788232803, + 0.0038492963649332523, + 0.007248071953654289, + 0.006566368509083986, + 0.013263755477964878, + 0.015839995816349983, + 0.009099651128053665, + 0.010859663598239422, + 0.0074239568784832954, + 0.013184873387217522, + 0.0022911601699888706, + 0.008998504839837551, + 0.0037865552585572004, + 0.004331819247454405, + 0.016923410817980766, + 0.004989122971892357, + 0.0031633919570595026, + 0.00241259066388011, + 0.010929171927273273, + 0.018181931227445602, + 0.00858923327177763, + 0.00780813954770565, + 0.03081713616847992, + 0.023486090824007988, + 0.007231715135276318, + 0.007963037118315697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.012875283136963844 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865909367799759 + ], + "opposite_logprob_tokens": [ + -3.048659086227417 + ], + "logprob_reward": -0.04865909367799759, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865909367799759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08066985011100769, + 0.047421835362911224, + 0.039571769535541534, + 0.020942701026797295, + 0.049054037779569626, + 0.4949575960636139, + 0.07989346235990524, + 0.10291022807359695, + 0.05211249366402626, + 0.03246600925922394 + ], + "credit_map": [ + 0.010950954630970955, + 0.005714363418519497, + 0.007979165762662888, + 0.009507331997156143, + 0.01036672480404377, + 0.011642822995781898, + 0.013035980984568596, + 0.011472516693174839, + 0.007772390265017748, + 0.007235264405608177, + 0.00657730083912611, + 0.00904051586985588, + 0.005567178595811129, + 0.00382458814419806, + 0.004331567324697971, + 0.0030730334110558033, + 0.0024381843395531178, + 0.00439325999468565, + 0.008255254477262497, + 0.006226338446140289, + 0.004071047995239496, + 0.0041054547764360905, + 0.00555396405979991, + 0.004528269171714783, + 0.0015827525639906526, + 0.001726814778521657, + 0.0026334146969020367, + 0.004166312050074339, + 0.0017599384300410748, + 0.0027447370812296867, + 0.0038468653801828623, + 0.002481866627931595, + 0.004143150988966227, + 0.006838467437773943, + 0.00593507569283247, + 0.005615505389869213, + 0.0054116337560117245, + 0.004384677391499281, + 0.006173893343657255, + 0.010551635175943375, + 0.05018055438995361, + 0.05492117628455162, + 0.07570762932300568, + 0.06155218183994293, + 0.05115383490920067, + 0.04245448485016823, + 0.1184169203042984, + 0.04057084769010544, + 0.010491647757589817, + 0.00917982216924429, + 0.008785543031990528, + 0.008677089586853981, + 0.016161613166332245, + 0.008853541687130928, + 0.007365603931248188, + 0.010378602892160416, + 0.009223779663443565, + 0.008671469986438751, + 0.016114994883537292, + 0.027545683085918427, + 0.00655274698510766, + 0.008825155906379223, + 0.010774918831884861, + 0.015201487578451633, + 0.0019725423771888018, + 0.013268597424030304, + 0.004257541615515947, + 0.003986380062997341, + 0.017993509769439697, + 0.004543682094663382, + 0.0032693403773009777, + 0.0028209059964865446, + 0.002656621392816305, + 0.004494884051382542, + 0.002846412593498826, + 0.004587452858686447, + 0.007350275758653879, + 0.004755090922117233, + 0.003244752297177911, + 0.002530517755076289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.04865909367799759 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348982006311417 + ], + "opposite_logprob_tokens": [ + -3.6513490676879883 + ], + "logprob_reward": -0.026348982006311417, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348982006311417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048542726784944534, + 0.03548255190253258, + 0.14734511077404022, + 0.026697710156440735, + 0.01794862188398838, + 0.29539749026298523, + 0.08006956428289413, + 0.1555410772562027, + 0.12221211194992065, + 0.070762999355793 + ], + "credit_map": [ + 0.008067229762673378, + 0.003997335210442543, + 0.006216842215508223, + 0.006013925652951002, + 0.0068711889907717705, + 0.0063973781652748585, + 0.004008107353001833, + 0.0069707236252725124, + 0.004267985932528973, + 0.009831150062382221, + 0.005319239106029272, + 0.0037704536225646734, + 0.003379870904609561, + 0.0037443675100803375, + 0.00236120936460793, + 0.002808277728036046, + 0.007497408427298069, + 0.00776676507666707, + 0.054025016725063324, + 0.015322151593863964, + 0.005540198180824518, + 0.003569420427083969, + 0.046081677079200745, + 0.007542475592344999, + 0.0024184470530599356, + 0.0023384513333439827, + 0.004664957523345947, + 0.005589349195361137, + 0.001921981223858893, + 0.002364332787692547, + 0.004054337274283171, + 0.003345854114741087, + 0.0019455791916698217, + 0.0018646084936335683, + 0.0023519592359662056, + 0.002187622245401144, + 0.0019279280677437782, + 0.002081760670989752, + 0.002631574869155884, + 0.0029575887601822615, + 0.026433156803250313, + 0.028264127671718597, + 0.049607858061790466, + 0.06090189889073372, + 0.025142693892121315, + 0.028264418244361877, + 0.05315955728292465, + 0.02362377755343914, + 0.010035507380962372, + 0.01072829868644476, + 0.009938232600688934, + 0.011206841096282005, + 0.009730840101838112, + 0.008956745266914368, + 0.010422106832265854, + 0.009050990454852581, + 0.013026461936533451, + 0.01191096194088459, + 0.02378600649535656, + 0.0332682840526104, + 0.013246465474367142, + 0.018484167754650116, + 0.01565065234899521, + 0.026168081909418106, + 0.006199894472956657, + 0.03484667092561722, + 0.011592021211981773, + 0.007741403765976429, + 0.03712812066078186, + 0.007555922493338585, + 0.008657203055918217, + 0.008490880019962788, + 0.009640838950872421, + 0.011718765832483768, + 0.007045099511742592, + 0.008194332011044025, + 0.012586025521159172, + 0.010993740521371365, + 0.005122057627886534, + 0.0054621449671685696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 2 + ], + "selected_reward": -0.026348982006311417 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370409458875656 + ], + "opposite_logprob_tokens": [ + -3.4087040424346924 + ], + "logprob_reward": -0.03370409458875656, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370409458875656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0320868156850338, + 0.04062953591346741, + 0.07473643869161606, + 0.03276783600449562, + 0.0464712493121624, + 0.1321253478527069, + 0.054641008377075195, + 0.2784166634082794, + 0.23085369169712067, + 0.07727143913507462 + ], + "credit_map": [ + 0.0043497150763869286, + 0.0032342253252863884, + 0.0042739445343613625, + 0.004463183227926493, + 0.0034874090924859047, + 0.0024655850138515234, + 0.006013966165482998, + 0.0037987835239619017, + 0.006770852487534285, + 0.009678689762949944, + 0.005400208756327629, + 0.004007645882666111, + 0.005047797225415707, + 0.0031684143468737602, + 0.002771681174635887, + 0.00378424278460443, + 0.0035058993380516768, + 0.0031387258786708117, + 0.025757232680916786, + 0.007759752217680216, + 0.003387495642527938, + 0.0023070957977324724, + 0.0249860268086195, + 0.0038942089304327965, + 0.002175819594413042, + 0.0025653906632214785, + 0.004852854646742344, + 0.010262271389365196, + 0.0015764230629429221, + 0.0019907753448933363, + 0.005711213685572147, + 0.003633086336776614, + 0.004304788075387478, + 0.0024882410652935505, + 0.005739426240324974, + 0.006221539340913296, + 0.0035596645902842283, + 0.005090784281492233, + 0.00711307767778635, + 0.011953726410865784, + 0.010674559511244297, + 0.012064622715115547, + 0.018828174099326134, + 0.019779693335294724, + 0.011963370256125927, + 0.012494096532464027, + 0.03285890072584152, + 0.013461929745972157, + 0.010341892018914223, + 0.006071153096854687, + 0.00984728243201971, + 0.007008551619946957, + 0.004045103210955858, + 0.005810304079204798, + 0.0077674975618720055, + 0.003749219933524728, + 0.024677671492099762, + 0.022521743550896645, + 0.04325677454471588, + 0.06668233871459961, + 0.02434827946126461, + 0.029472684487700462, + 0.028416506946086884, + 0.03904065117239952, + 0.013103730976581573, + 0.07434386014938354, + 0.01318443100899458, + 0.014691956341266632, + 0.06846875697374344, + 0.015529114753007889, + 0.01590907759964466, + 0.015622761100530624, + 0.013213320635259151, + 0.011293926276266575, + 0.009451122023165226, + 0.01020494382828474, + 0.009773719124495983, + 0.010704156942665577, + 0.006806009449064732, + 0.00582423759624362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.03370409458875656 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a gnarled tree branch with dark, richly textured bark?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial state of the branch covered in an intricate web of needle-like hoarfrost crystals?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the frost transform into heavy, silvery droplets with the appearance of liquid mercury?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the mercury droplets detach and glide along the branch's contours while reflecting golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, graceful 360-degree orbital pan around the branch?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233209133148193 + ], + "opposite_logprob_tokens": [ + -0.8983209133148193 + ], + "logprob_reward": -0.5233209133148193, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233209133148193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022860271856188774, + 0.012353996746242046, + 0.19121053814888, + 0.03799573704600334, + 0.01211251225322485, + 0.15800747275352478, + 0.04786583408713341, + 0.14223341643810272, + 0.32261279225349426, + 0.05274742469191551 + ], + "credit_map": [ + 0.0027363221161067486, + 0.001994900405406952, + 0.002597676357254386, + 0.002571654273197055, + 0.0030694594606757164, + 0.0022303664591163397, + 0.003714291611686349, + 0.003945601172745228, + 0.0010277241235598922, + 0.003059424227103591, + 0.001673632301390171, + 0.0015784946735948324, + 0.0021408931352198124, + 0.0010734341340139508, + 0.0009335238137282431, + 0.0008668706868775189, + 0.00883115641772747, + 0.006366183049976826, + 0.0733131617307663, + 0.020842568948864937, + 0.006447144318372011, + 0.002775222295895219, + 0.06404922902584076, + 0.008585887029767036, + 0.0024660269264131784, + 0.0033738110214471817, + 0.005571178160607815, + 0.007504465989768505, + 0.002649185946211219, + 0.004048983566462994, + 0.00789245218038559, + 0.004489636514335871, + 0.0011188400676473975, + 0.0021218585316091776, + 0.001320375013165176, + 0.001474961405619979, + 0.0010950597934424877, + 0.0015253906603902578, + 0.001566444057971239, + 0.001889582839794457, + 0.015084882266819477, + 0.017162809148430824, + 0.022613177075982094, + 0.019508270546793938, + 0.015963837504386902, + 0.011725441552698612, + 0.0409921370446682, + 0.01495691854506731, + 0.005481489934027195, + 0.005151695571839809, + 0.00528165465220809, + 0.005833092611283064, + 0.008604941889643669, + 0.005767653230577707, + 0.0049277557991445065, + 0.006817550864070654, + 0.014144872315227985, + 0.012768439948558807, + 0.022619374096393585, + 0.03933994472026825, + 0.00873636081814766, + 0.010911617428064346, + 0.01759711280465126, + 0.016115698963403702, + 0.01757357083261013, + 0.10490448027849197, + 0.025063086301088333, + 0.029949894174933434, + 0.07185728847980499, + 0.024605751037597656, + 0.025102877989411354, + 0.023555848747491837, + 0.008412019349634647, + 0.006912613287568092, + 0.005410470068454742, + 0.008465067483484745, + 0.00906050018966198, + 0.0063047707080841064, + 0.0038046562112867832, + 0.004377332981675863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.5233209133148193 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_3/index_tensor([8539, 8539])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_3/index_tensor([8539, 8539])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d2f6c6c8acff6a70aba48a3495b8b93ea85cc910 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_3/index_tensor([8539, 8539])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a fluid and elegant 360-degree panning shot, the image begins with a close-up of the muzzle of a Beretta 687 hunting rifle, a thin, curling plume of bluish-gray smoke rising upward. The gun's metal receiver, engraved with an intricate European scroll pattern, reflects a restrained silver luster in the warm, slanting golden sunlight of an autumn afternoon. The stock and handguard are crafted from deep-grained walnut, finished with a smooth, oil-waxed finish that imparts a warm reddish-brown hue. As the camera steadily pans backward and around the subject, we see the gun resting on an unseen stand against a vast, pastoral backdrop: golden grass undulating in the wind, and fallen leaves in ochre and crimson scattered across the ground. At that moment, a freshly fired projectile traces a nearly invisible, shimmering metallic trail through the air. It doesn't fly in a straight line, but rather, mid-flight, strikes an invisible, hard object at high speed, rebounding and changing direction. It ultimately strikes a weathered, round metal bull's eye in the distance at a treacherous angle, leaving a tiny, gleaming dent in the impact. The entire scene is bathed in a soft, warm light, with elongated shadows stretching across the ground, creating an atmosphere of tranquility, concentration, and a rustic feel, perfectly capturing the power and precision of this moment.", + "seed": 27546, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_3/index_8539_seed_27546_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04163965415209532, + "vlm_logprob": -0.04163965415209532, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.04163965415209532, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0323246531188488 + ], + "opposite_logprob_tokens": [ + -3.5323245525360107 + ], + "logprob_reward": -0.0323246531188488, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0323246531188488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22496908903121948, + 0.06980428099632263, + 0.08019908517599106, + 0.047400377690792084, + 0.1409166008234024, + 0.040504489094018936, + 0.059082962572574615, + 0.1643064171075821, + 0.09146711230278015, + 0.08134953677654266 + ], + "credit_map": [ + 0.026986265555024147, + 0.03328399360179901, + 0.02012520842254162, + 0.035671964287757874, + 0.02701527625322342, + 0.030364342033863068, + 0.026527902111411095, + 0.024994147941470146, + 0.005508446134626865, + 0.008655640296638012, + 0.009695387445390224, + 0.011179584078490734, + 0.007574775256216526, + 0.010857454501092434, + 0.009822404943406582, + 0.00651058554649353, + 0.007516879588365555, + 0.011636563576757908, + 0.005823835264891386, + 0.008081202395260334, + 0.012721825391054153, + 0.013880033977329731, + 0.00688044261187315, + 0.013658303767442703, + 0.0057381996884942055, + 0.005877564195543528, + 0.003594453213736415, + 0.0032898809295147657, + 0.005756362807005644, + 0.012270844541490078, + 0.004697785247117281, + 0.006175287999212742, + 0.015013044700026512, + 0.018231244757771492, + 0.018220825120806694, + 0.011426392942667007, + 0.01807529106736183, + 0.027408363297581673, + 0.011444329284131527, + 0.021097103133797646, + 0.003623265540227294, + 0.0046059489250183105, + 0.005766973830759525, + 0.005298036616295576, + 0.0049997735768556595, + 0.004880161490291357, + 0.006452989298850298, + 0.004877341911196709, + 0.0035413666628301144, + 0.0071546416729688644, + 0.004851445555686951, + 0.005524009931832552, + 0.010114558972418308, + 0.009607150219380856, + 0.011109680868685246, + 0.007180109154433012, + 0.008483723737299442, + 0.013173212297260761, + 0.006189107894897461, + 0.008371014147996902, + 0.01002419926226139, + 0.006954126060009003, + 0.06204593554139137, + 0.04906509816646576, + 0.00734120886772871, + 0.015460462309420109, + 0.0072696628049016, + 0.008419784717261791, + 0.009155686013400555, + 0.007293646689504385, + 0.022159310057759285, + 0.014367345720529556, + 0.00545266130939126, + 0.011065893806517124, + 0.008259623311460018, + 0.005571044981479645, + 0.020850826054811478, + 0.010400038212537766, + 0.011184093542397022, + 0.008565356023609638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.0323246531188488 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018204161897301674 + ], + "opposite_logprob_tokens": [ + -4.018204212188721 + ], + "logprob_reward": -0.018204161897301674, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018204161897301674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07724836468696594, + 0.0692211464047432, + 0.05302168428897858, + 0.061229076236486435, + 0.08174445480108261, + 0.04769301414489746, + 0.0684947595000267, + 0.0740244910120964, + 0.3726142644882202, + 0.09470880031585693 + ], + "credit_map": [ + 0.007510670460760593, + 0.009075501933693886, + 0.008153825998306274, + 0.01592499203979969, + 0.006907823961228132, + 0.009404242970049381, + 0.010480647906661034, + 0.009790642186999321, + 0.0052832369692623615, + 0.010762529447674751, + 0.008056179620325565, + 0.012898082844913006, + 0.007362923584878445, + 0.01219792291522026, + 0.007175040431320667, + 0.00548521988093853, + 0.004760946612805128, + 0.007485632319003344, + 0.002993121510371566, + 0.005473786033689976, + 0.008490036241710186, + 0.008336280472576618, + 0.007045047357678413, + 0.00843682512640953, + 0.006974183954298496, + 0.006812110543251038, + 0.0060910568572580814, + 0.004090407397598028, + 0.007035933434963226, + 0.015779394656419754, + 0.005751278717070818, + 0.008694701828062534, + 0.008553877472877502, + 0.009575597010552883, + 0.010751564055681229, + 0.006459258031100035, + 0.011445589363574982, + 0.01638755388557911, + 0.006838641595095396, + 0.01173236034810543, + 0.002939264290034771, + 0.005091870203614235, + 0.003021949203684926, + 0.004044611006975174, + 0.009172679856419563, + 0.005340311676263809, + 0.012018593028187752, + 0.006063730455935001, + 0.00611311849206686, + 0.00998470839112997, + 0.006593512371182442, + 0.006035782862454653, + 0.00643477588891983, + 0.0098953228443861, + 0.009844648651778698, + 0.013592878356575966, + 0.005286474712193012, + 0.012539351359009743, + 0.0071756113320589066, + 0.007632658816874027, + 0.008755761198699474, + 0.008037499152123928, + 0.015973418951034546, + 0.008623708039522171, + 0.025096822530031204, + 0.08396212011575699, + 0.022409124299883842, + 0.032145094126462936, + 0.039854995906353, + 0.023854726925492287, + 0.08850802481174469, + 0.05678333342075348, + 0.005852102767676115, + 0.012466445565223694, + 0.009117607958614826, + 0.006478847004473209, + 0.025093628093600273, + 0.01041542086750269, + 0.014136498793959618, + 0.011148232035338879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.018204161897301674 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700213760137558 + ], + "opposite_logprob_tokens": [ + -2.695021390914917 + ], + "logprob_reward": -0.0700213760137558, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700213760137558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11029766499996185, + 0.04628743231296539, + 0.0368671640753746, + 0.0324946790933609, + 0.04697573557496071, + 0.06569210439920425, + 0.1176423504948616, + 0.31241631507873535, + 0.17483632266521454, + 0.056490208953619 + ], + "credit_map": [ + 0.008580132387578487, + 0.009936317801475525, + 0.010059191845357418, + 0.02319778874516487, + 0.014617524109780788, + 0.010112856514751911, + 0.015051665715873241, + 0.018742185086011887, + 0.005052010528743267, + 0.009441068395972252, + 0.006599976681172848, + 0.005157213192433119, + 0.004080097656697035, + 0.006272262427955866, + 0.004975547548383474, + 0.004709254484623671, + 0.0032697354909032583, + 0.0061585367657244205, + 0.004376546014100313, + 0.0029681874439120293, + 0.004370329435914755, + 0.005834884010255337, + 0.005476731341332197, + 0.004412216134369373, + 0.0038202698342502117, + 0.003157719736918807, + 0.005170135758817196, + 0.0031121219508349895, + 0.0033678989857435226, + 0.0058759222738444805, + 0.003200203413143754, + 0.0047904071398079395, + 0.005656662862747908, + 0.006866443436592817, + 0.005099317524582148, + 0.00569317489862442, + 0.0055960360914468765, + 0.007218997925519943, + 0.004672734998166561, + 0.006172365043312311, + 0.005372712388634682, + 0.013962304219603539, + 0.007750729098916054, + 0.005344612058252096, + 0.006485138088464737, + 0.0055964128114283085, + 0.01323334313929081, + 0.007946851663291454, + 0.008055469952523708, + 0.012657946906983852, + 0.009966420009732246, + 0.01340580079704523, + 0.015742484480142593, + 0.020866259932518005, + 0.016657380387187004, + 0.020290594547986984, + 0.018179049715399742, + 0.035959891974925995, + 0.0146711440756917, + 0.01992599479854107, + 0.025764262303709984, + 0.016975663602352142, + 0.09439844638109207, + 0.08654186874628067, + 0.012671742588281631, + 0.03859838843345642, + 0.020714420825242996, + 0.016569336876273155, + 0.01650192029774189, + 0.013075297698378563, + 0.035392262041568756, + 0.021312963217496872, + 0.0044141667895019054, + 0.005728310439735651, + 0.007051251828670502, + 0.00447878148406744, + 0.012146095745265484, + 0.006385320331901312, + 0.007821948267519474, + 0.008464330807328224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.0700213760137558 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790116935968399 + ], + "opposite_logprob_tokens": [ + -2.5790116786956787 + ], + "logprob_reward": -0.0790116935968399, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790116935968399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09850771725177765, + 0.03749747574329376, + 0.12868307530879974, + 0.11669379472732544, + 0.08930299431085587, + 0.07273853570222855, + 0.05310191214084625, + 0.11321063339710236, + 0.08977767080068588, + 0.2004862278699875 + ], + "credit_map": [ + 0.008196279406547546, + 0.011299536563456059, + 0.00750785181298852, + 0.01765916682779789, + 0.016890110448002815, + 0.014945476315915585, + 0.011160195805132389, + 0.010849093087017536, + 0.003901899326592684, + 0.005322007928043604, + 0.0032736652065068483, + 0.004586480092257261, + 0.005132255144417286, + 0.009135173633694649, + 0.002975097158923745, + 0.0031708967871963978, + 0.01016785204410553, + 0.017418548464775085, + 0.007623796351253986, + 0.014537883922457695, + 0.022400936111807823, + 0.019513893872499466, + 0.019992509856820107, + 0.017027653753757477, + 0.0132821686565876, + 0.012094421312212944, + 0.0153726851567626, + 0.009140378795564175, + 0.012354213744401932, + 0.026723628863692284, + 0.01051029097288847, + 0.017216002568602562, + 0.009793810546398163, + 0.01032156590372324, + 0.013256444595754147, + 0.007767520844936371, + 0.01071957964450121, + 0.016920655965805054, + 0.007369224447757006, + 0.013154192827641964, + 0.005588262341916561, + 0.006708013825118542, + 0.006126175168901682, + 0.01046136673539877, + 0.008579542860388756, + 0.007729388773441315, + 0.016837026923894882, + 0.010708760470151901, + 0.004476397763937712, + 0.00885867141187191, + 0.00620616041123867, + 0.005787898786365986, + 0.006596420891582966, + 0.007853778079152107, + 0.005224745254963636, + 0.00809784047305584, + 0.005743104964494705, + 0.02121579274535179, + 0.005561851896345615, + 0.0061989519745111465, + 0.007435694802552462, + 0.00554881663993001, + 0.0393485352396965, + 0.022157886996865273, + 0.006550315301865339, + 0.009706468321383, + 0.008947945199906826, + 0.008759482763707638, + 0.008089919574558735, + 0.009613318368792534, + 0.0217587873339653, + 0.016351431608200073, + 0.012707694433629513, + 0.028239281848073006, + 0.018162261694669724, + 0.012545342557132244, + 0.048900626599788666, + 0.025452399626374245, + 0.02944680117070675, + 0.025031818076968193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.0790116935968399 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636386133730412 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636386133730412, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636386133730412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06202187016606331, + 0.07165438681840897, + 0.02017216384410858, + 0.042311351746320724, + 0.09869787096977234, + 0.04296937584877014, + 0.20148862898349762, + 0.05130099505186081, + 0.3102998435497284, + 0.09908352792263031 + ], + "credit_map": [ + 0.005407397169619799, + 0.00579716544598341, + 0.005107762757688761, + 0.011977290734648705, + 0.010664871893823147, + 0.008305985480546951, + 0.0075519694946706295, + 0.007209424395114183, + 0.006620161700993776, + 0.012822293676435947, + 0.007613389752805233, + 0.009099823422729969, + 0.009380443021655083, + 0.012964603491127491, + 0.006609551142901182, + 0.006544116418808699, + 0.002221983391791582, + 0.0037665609270334244, + 0.0012312856270000339, + 0.0016434094868600368, + 0.0026818583719432354, + 0.004124551545828581, + 0.0015628067776560783, + 0.0029397080652415752, + 0.004557630512863398, + 0.005492975935339928, + 0.00513838417828083, + 0.0034383025486022234, + 0.005087631288915873, + 0.009739591740071774, + 0.0030903532169759274, + 0.005766482558101416, + 0.010312429629266262, + 0.013757198117673397, + 0.012287664227187634, + 0.007562353741377592, + 0.013715344481170177, + 0.01948765479028225, + 0.00747193768620491, + 0.014103284105658531, + 0.0038803957868367434, + 0.004617257975041866, + 0.0033828297164291143, + 0.0015683380188420415, + 0.00412027956917882, + 0.006110414396971464, + 0.009795219637453556, + 0.009494645521044731, + 0.015998734161257744, + 0.018449505791068077, + 0.019575076177716255, + 0.012955999933183193, + 0.028182649984955788, + 0.027601024135947227, + 0.03739583492279053, + 0.04132979363203049, + 0.003650808706879616, + 0.013556852005422115, + 0.0031014455016702414, + 0.0032870119903236628, + 0.006630524527281523, + 0.007788856513798237, + 0.005298227071762085, + 0.007987270131707191, + 0.015469862148165703, + 0.01380064245313406, + 0.007040746510028839, + 0.010295499116182327, + 0.014291093684732914, + 0.016639837995171547, + 0.1258154809474945, + 0.1069466695189476, + 0.00781750213354826, + 0.008932553231716156, + 0.010067837312817574, + 0.00814717449247837, + 0.011657815426588058, + 0.0152369299903512, + 0.014830182306468487, + 0.022393541410565376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.008636386133730412 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a fluid and elegant 360-degree panning shot, the image begins with a close-up of the muzzle of a Beretta 687 hunting rifle, a thin, curling plume of bluish-gray smoke rising upward. The gun's metal receiver, engraved with an intricate European scroll pattern, reflects a restrained silver luster in the warm, slanting golden sunlight of an autumn afternoon. The stock and handguard are crafted from deep-grained walnut, finished with a smooth, oil-waxed finish that imparts a warm reddish-brown hue. As the camera steadily pans backward and around the subject, we see the gun resting on an unseen stand against a vast, pastoral backdrop: golden grass undulating in the wind, and fallen leaves in ochre and crimson scattered across the ground. At that moment, a freshly fired projectile traces a nearly invisible, shimmering metallic trail through the air. It doesn't fly in a straight line, but rather, mid-flight, strikes an invisible, hard object at high speed, rebounding and changing direction. It ultimately strikes a weathered, round metal bull's eye in the distance at a treacherous angle, leaving a tiny, gleaming dent in the impact. The entire scene is bathed in a soft, warm light, with elongated shadows stretching across the ground, creating an atmosphere of tranquility, concentration, and a rustic feel, perfectly capturing the power and precision of this moment.", + "seed": 27547, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_3/index_8539_seed_27547_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06482709757983685, + "vlm_logprob": -0.06482709757983685, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.06482709757983685, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05146760493516922 + ], + "opposite_logprob_tokens": [ + -3.0514676570892334 + ], + "logprob_reward": -0.05146760493516922, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05146760493516922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1329239457845688, + 0.05117854103446007, + 0.17930364608764648, + 0.1008630022406578, + 0.1922546774148941, + 0.05875958129763603, + 0.06571024656295776, + 0.050126854330301285, + 0.12057891488075256, + 0.048300545662641525 + ], + "credit_map": [ + 0.010917452163994312, + 0.01611345261335373, + 0.024930782616138458, + 0.018785301595926285, + 0.013425171375274658, + 0.014520910568535328, + 0.01783302053809166, + 0.01639787293970585, + 0.004304058849811554, + 0.004015494138002396, + 0.006022158544510603, + 0.007401082199066877, + 0.006577734369784594, + 0.00895464513450861, + 0.007212022785097361, + 0.006691343151032925, + 0.046320993453264236, + 0.012912807054817677, + 0.009427385404706001, + 0.0055247945711016655, + 0.06133238598704338, + 0.017790537327528, + 0.011338906362652779, + 0.014655832201242447, + 0.015662938356399536, + 0.011599713936448097, + 0.010093322023749352, + 0.008385156281292439, + 0.02851417101919651, + 0.007755942177027464, + 0.010817206464707851, + 0.00803455151617527, + 0.011463835835456848, + 0.008482268080115318, + 0.06192290410399437, + 0.012898451648652554, + 0.028152577579021454, + 0.013913430273532867, + 0.005944336298853159, + 0.04947686567902565, + 0.004913636017590761, + 0.005267445929348469, + 0.011696959845721722, + 0.007480219937860966, + 0.006423811428248882, + 0.00810773391276598, + 0.006328030023723841, + 0.008541745133697987, + 0.00985508132725954, + 0.005375332664698362, + 0.016930172219872475, + 0.005969400983303785, + 0.007161477115005255, + 0.0050519416108727455, + 0.00916242878884077, + 0.006204414181411266, + 0.0052983881905674934, + 0.006001502741128206, + 0.006143385544419289, + 0.0068363891914486885, + 0.005208825226873159, + 0.008712163195014, + 0.006940737832337618, + 0.004985463805496693, + 0.02336583472788334, + 0.007269308436661959, + 0.016305560246109962, + 0.00913266558200121, + 0.008982185274362564, + 0.012452111579477787, + 0.01190905924886465, + 0.03116217628121376, + 0.013243253342807293, + 0.0030660578049719334, + 0.016777029260993004, + 0.003916846588253975, + 0.0025504594668745995, + 0.004007059149444103, + 0.0023347693495452404, + 0.002405071398243308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.05146760493516922 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067544239573180676 + ], + "opposite_logprob_tokens": [ + -5.006754398345947 + ], + "logprob_reward": -0.0067544239573180676, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067544239573180676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030818866565823555, + 0.032917045056819916, + 0.0807437151670456, + 0.09253286570310593, + 0.10671564191579819, + 0.09485139697790146, + 0.03613734245300293, + 0.25167152285575867, + 0.081659696996212, + 0.19195197522640228 + ], + "credit_map": [ + 0.0021979964803904295, + 0.004309587646275759, + 0.0030270610004663467, + 0.003528710687533021, + 0.003985403571277857, + 0.0032461481168866158, + 0.00646382849663496, + 0.00406013336032629, + 0.0032296094577759504, + 0.0028677412774413824, + 0.003398974658921361, + 0.003969049546867609, + 0.005970372818410397, + 0.003946513403207064, + 0.00566178560256958, + 0.0038730010855942965, + 0.011723075062036514, + 0.007293299771845341, + 0.009770417585968971, + 0.00424095056951046, + 0.008809505961835384, + 0.010588740929961205, + 0.010208561085164547, + 0.018109165132045746, + 0.013602541759610176, + 0.011373721994459629, + 0.009496434591710567, + 0.008000270463526249, + 0.024842865765094757, + 0.00680239824578166, + 0.010595026426017284, + 0.007819603197276592, + 0.01738029345870018, + 0.00540336687117815, + 0.024807412177324295, + 0.006427912972867489, + 0.014960473403334618, + 0.011268251575529575, + 0.0043548643589019775, + 0.022113056853413582, + 0.01025179959833622, + 0.005998203065246344, + 0.014846039935946465, + 0.008382837288081646, + 0.014257021248340607, + 0.0180569626390934, + 0.009148528799414635, + 0.013910005800426006, + 0.003946671262383461, + 0.002753305248916149, + 0.004707500338554382, + 0.0037083642091602087, + 0.0038773033302277327, + 0.006038161925971508, + 0.006452643312513828, + 0.004653391893953085, + 0.033035118132829666, + 0.02411186508834362, + 0.03179552033543587, + 0.02763601392507553, + 0.040554486215114594, + 0.05115275830030441, + 0.019325178116559982, + 0.024060573428869247, + 0.007713476661592722, + 0.006612780503928661, + 0.017529645934700966, + 0.007503311615437269, + 0.007837988436222076, + 0.009290573187172413, + 0.007280166260898113, + 0.01789175532758236, + 0.0648643895983696, + 0.008019915781915188, + 0.0787188932299614, + 0.014740590937435627, + 0.005619281902909279, + 0.010973126627504826, + 0.004148807842284441, + 0.0048669748939573765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.0067544239573180676 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034221649169922 + ], + "opposite_logprob_tokens": [ + -1.9103422164916992 + ], + "logprob_reward": -0.16034221649169922, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034221649169922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04125020653009415, + 0.04324258491396904, + 0.1751309484243393, + 0.08189717680215836, + 0.05457155406475067, + 0.06669577956199646, + 0.05686166137456894, + 0.10853759199380875, + 0.12352125346660614, + 0.2482912540435791 + ], + "credit_map": [ + 0.0037061022594571114, + 0.0040190755389630795, + 0.007098113186657429, + 0.005604971665889025, + 0.005260353907942772, + 0.00471684243530035, + 0.004737901035696268, + 0.006106847431510687, + 0.005113476421684027, + 0.00381205091252923, + 0.004944397136569023, + 0.005049647763371468, + 0.008330748416483402, + 0.005703373346477747, + 0.004618131555616856, + 0.005670758429914713, + 0.04593095928430557, + 0.012623813934624195, + 0.00868625845760107, + 0.006014243699610233, + 0.061420589685440063, + 0.017053062096238136, + 0.008010849356651306, + 0.015391158871352673, + 0.011864490807056427, + 0.008506235666573048, + 0.011226235888898373, + 0.0076698730699718, + 0.02218000777065754, + 0.00444391556084156, + 0.007719420362263918, + 0.008286995813250542, + 0.003977403976023197, + 0.0015926960622891784, + 0.01989409700036049, + 0.004545558709651232, + 0.0033683369401842356, + 0.0030228302348405123, + 0.0014953460777178407, + 0.01667528599500656, + 0.006856652442365885, + 0.007092153187841177, + 0.007844367064535618, + 0.005642411764711142, + 0.014345630072057247, + 0.013564431108534336, + 0.006438524462282658, + 0.004911606200039387, + 0.006771990563720465, + 0.006897918879985809, + 0.012753356248140335, + 0.006205555517226458, + 0.0055379425175487995, + 0.006345421075820923, + 0.006526134442538023, + 0.005823343992233276, + 0.014687527902424335, + 0.009345253929495811, + 0.01557806134223938, + 0.012597731314599514, + 0.01762498915195465, + 0.021076340228319168, + 0.007703711744397879, + 0.009923974983394146, + 0.023092025890946388, + 0.009081699885427952, + 0.02156815119087696, + 0.012355986982584, + 0.010379752144217491, + 0.012198502197861671, + 0.01163789164274931, + 0.023207245394587517, + 0.08882355690002441, + 0.009744284674525261, + 0.11029782891273499, + 0.010490957647562027, + 0.005421765614300966, + 0.013259898871183395, + 0.0040360428392887115, + 0.006216921377927065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.16034221649169922 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032444447278976 + ], + "opposite_logprob_tokens": [ + -2.3503243923187256 + ], + "logprob_reward": -0.10032444447278976, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032444447278976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022930188104510307, + 0.022387947887182236, + 0.29315805435180664, + 0.11788304150104523, + 0.0901264175772667, + 0.07546648383140564, + 0.02709224261343479, + 0.04491286724805832, + 0.12247075885534286, + 0.1835719645023346 + ], + "credit_map": [ + 0.0020958080422133207, + 0.0023767175152897835, + 0.003905689809471369, + 0.0022654396016150713, + 0.002934709656983614, + 0.002761641750112176, + 0.003039164002984762, + 0.003551015630364418, + 0.002494336571544409, + 0.0019844069611281157, + 0.002057526493445039, + 0.002751380205154419, + 0.004418106749653816, + 0.003316732821986079, + 0.0031232167966663837, + 0.002242239424958825, + 0.0779365822672844, + 0.02470291033387184, + 0.010850346647202969, + 0.008932553231716156, + 0.11576142907142639, + 0.02624061331152916, + 0.013610176742076874, + 0.015123453922569752, + 0.01759430579841137, + 0.013448909856379032, + 0.013291258364915848, + 0.009242992848157883, + 0.03224881365895271, + 0.008973868563771248, + 0.013519851490855217, + 0.009563030675053596, + 0.015134623274207115, + 0.004159580450505018, + 0.022137900814414024, + 0.004960713442414999, + 0.01349196583032608, + 0.00744714867323637, + 0.003605291247367859, + 0.019189193844795227, + 0.006543382536619902, + 0.00811088364571333, + 0.00821597222238779, + 0.007725350558757782, + 0.013215368613600731, + 0.014230827800929546, + 0.006908399052917957, + 0.010516302660107613, + 0.004463768098503351, + 0.001802198588848114, + 0.0069082798436284065, + 0.0027138825971633196, + 0.0030428441241383553, + 0.0023354089353233576, + 0.0031201946549117565, + 0.002705663675442338, + 0.006011660676449537, + 0.003697063308209181, + 0.008305258117616177, + 0.007277754135429859, + 0.005359284114092588, + 0.006316591519862413, + 0.0036955454852432013, + 0.004249707330018282, + 0.017066016793251038, + 0.007146814838051796, + 0.019772270694375038, + 0.009586547501385212, + 0.010554080829024315, + 0.013343251310288906, + 0.011846770532429218, + 0.03315499797463417, + 0.06413416564464569, + 0.007373514585196972, + 0.08102280646562576, + 0.008226980455219746, + 0.0043106465600430965, + 0.010483457706868649, + 0.0032315903808921576, + 0.004788809921592474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 8 + ], + "selected_reward": -0.10032444447278976 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005246798042207956 + ], + "opposite_logprob_tokens": [ + -5.255246639251709 + ], + "logprob_reward": -0.005246798042207956, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005246798042207956, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05385517328977585, + 0.024092338979244232, + 0.21868684887886047, + 0.047992534935474396, + 0.10772310942411423, + 0.05381911247968674, + 0.03291939198970795, + 0.032431382685899734, + 0.15227548778057098, + 0.2762046456336975 + ], + "credit_map": [ + 0.004900102969259024, + 0.005859740078449249, + 0.01009247824549675, + 0.004862435162067413, + 0.00654127961024642, + 0.00659561762586236, + 0.007008668500930071, + 0.007994852028787136, + 0.00394322257488966, + 0.0016235352959483862, + 0.002266431227326393, + 0.0028449026867747307, + 0.0028816848061978817, + 0.004928630776703358, + 0.002678083721548319, + 0.002925848588347435, + 0.030045228078961372, + 0.022758884355425835, + 0.019319111481308937, + 0.012119712308049202, + 0.05901676416397095, + 0.022049980238080025, + 0.026782281696796417, + 0.02659488469362259, + 0.004458435345441103, + 0.006208977196365595, + 0.008214229717850685, + 0.004200412891805172, + 0.004781218711286783, + 0.006365087814629078, + 0.004999520257115364, + 0.008764653466641903, + 0.01750512793660164, + 0.006499359849840403, + 0.0120377317070961, + 0.00534784933552146, + 0.02983917109668255, + 0.02063179388642311, + 0.0064763352274894714, + 0.009385739453136921, + 0.0073722731322050095, + 0.00539390416815877, + 0.007830167189240456, + 0.005832641385495663, + 0.010736520402133465, + 0.00776117667555809, + 0.004823894239962101, + 0.004068533889949322, + 0.004558969289064407, + 0.0029720072634518147, + 0.00545661011710763, + 0.0037731570191681385, + 0.00414742948487401, + 0.004877380095422268, + 0.0037296810187399387, + 0.003404156770557165, + 0.0029348931275308132, + 0.0030911436770111322, + 0.005059527698904276, + 0.008399221114814281, + 0.002669863635674119, + 0.003770801704376936, + 0.0035967298317700624, + 0.0029092009644955397, + 0.028065966442227364, + 0.008775752037763596, + 0.02016194351017475, + 0.011176549829542637, + 0.007486199960112572, + 0.01575983688235283, + 0.014833653345704079, + 0.04601558297872543, + 0.0803256705403328, + 0.013002761639654636, + 0.08857716619968414, + 0.02404698356986046, + 0.0165559034794569, + 0.03480193391442299, + 0.008545728400349617, + 0.010348491370677948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 8 + ], + "selected_reward": -0.005246798042207956 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a fluid and elegant 360-degree panning shot, the image begins with a close-up of the muzzle of a Beretta 687 hunting rifle, a thin, curling plume of bluish-gray smoke rising upward. The gun's metal receiver, engraved with an intricate European scroll pattern, reflects a restrained silver luster in the warm, slanting golden sunlight of an autumn afternoon. The stock and handguard are crafted from deep-grained walnut, finished with a smooth, oil-waxed finish that imparts a warm reddish-brown hue. As the camera steadily pans backward and around the subject, we see the gun resting on an unseen stand against a vast, pastoral backdrop: golden grass undulating in the wind, and fallen leaves in ochre and crimson scattered across the ground. At that moment, a freshly fired projectile traces a nearly invisible, shimmering metallic trail through the air. It doesn't fly in a straight line, but rather, mid-flight, strikes an invisible, hard object at high speed, rebounding and changing direction. It ultimately strikes a weathered, round metal bull's eye in the distance at a treacherous angle, leaving a tiny, gleaming dent in the impact. The entire scene is bathed in a soft, warm light, with elongated shadows stretching across the ground, creating an atmosphere of tranquility, concentration, and a rustic feel, perfectly capturing the power and precision of this moment.", + "seed": 27548, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_3/index_8539_seed_27548_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11086550131440162, + "vlm_logprob": -0.11086550131440162, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.11086550131440162, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044789865612983704 + ], + "opposite_logprob_tokens": [ + -3.1697897911071777 + ], + "logprob_reward": -0.044789865612983704, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.044789865612983704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21306456625461578, + 0.03620710223913193, + 0.03767354413866997, + 0.054258670657873154, + 0.26914140582084656, + 0.21000193059444427, + 0.030913930386304855, + 0.040985990315675735, + 0.02703864872455597, + 0.08071417361497879 + ], + "credit_map": [ + 0.025660185143351555, + 0.04343453049659729, + 0.007750604767352343, + 0.022527607157826424, + 0.08735252171754837, + 0.00907923374325037, + 0.01047961600124836, + 0.006780252326279879, + 0.004478916060179472, + 0.004091828595846891, + 0.00537261413410306, + 0.004614979028701782, + 0.005562926176935434, + 0.005701927002519369, + 0.0032428426202386618, + 0.003141070483252406, + 0.0047966474667191505, + 0.006369303446263075, + 0.002645529806613922, + 0.0020300066098570824, + 0.006658545695245266, + 0.007760655600577593, + 0.0035181683488190174, + 0.0038946829736232758, + 0.006273095961660147, + 0.006398550234735012, + 0.008506841957569122, + 0.005326059181243181, + 0.0077284714207053185, + 0.009748970158398151, + 0.005317771341651678, + 0.004958911798894405, + 0.023375622928142548, + 0.0360235720872879, + 0.028918137773871422, + 0.036444708704948425, + 0.036680448800325394, + 0.04886115342378616, + 0.024893470108509064, + 0.033944278955459595, + 0.006280756555497646, + 0.01082176249474287, + 0.004791337065398693, + 0.003607490099966526, + 0.08490510284900665, + 0.08923749625682831, + 0.0035589118488132954, + 0.006799084134399891, + 0.003310247091576457, + 0.006596606690436602, + 0.0030128133948892355, + 0.002300298772752285, + 0.006093168631196022, + 0.006210912484675646, + 0.0018367565935477614, + 0.001553129404783249, + 0.0028062048368155956, + 0.0051993862725794315, + 0.004569983575493097, + 0.003997092135250568, + 0.007499359082430601, + 0.007129254285246134, + 0.003108775243163109, + 0.0066759358160197735, + 0.0019142841920256615, + 0.003284367034211755, + 0.002597169252112508, + 0.0026447915006428957, + 0.004441441036760807, + 0.003935229033231735, + 0.003949664067476988, + 0.004271702375262976, + 0.006013067439198494, + 0.008624848909676075, + 0.01070304587483406, + 0.0110360998660326, + 0.009355674497783184, + 0.00864317361265421, + 0.017792358994483948, + 0.008545905351638794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.044789865612983704 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011090068146586418 + ], + "opposite_logprob_tokens": [ + -4.511090278625488 + ], + "logprob_reward": -0.011090068146586418, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011090068146586418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1816045045852661, + 0.08362454921007156, + 0.10081733763217926, + 0.03307728469371796, + 0.07812928408384323, + 0.19658204913139343, + 0.07968908548355103, + 0.04906940087676048, + 0.016184363514184952, + 0.18122223019599915 + ], + "credit_map": [ + 0.014403674751520157, + 0.05149032548069954, + 0.012059305794537067, + 0.013498890213668346, + 0.05577552691102028, + 0.011386474594473839, + 0.011771664954721928, + 0.011218658648431301, + 0.011819237843155861, + 0.00904381088912487, + 0.013875830918550491, + 0.010849629528820515, + 0.009638194926083088, + 0.013508640229701996, + 0.006627826951444149, + 0.008261379785835743, + 0.015607429668307304, + 0.02332005836069584, + 0.011214233934879303, + 0.006755387876182795, + 0.013148225843906403, + 0.01507422886788845, + 0.008064758032560349, + 0.007633020170032978, + 0.0036286816466599703, + 0.0033495486713945866, + 0.005129192490130663, + 0.003837515600025654, + 0.004670105874538422, + 0.0047976491041481495, + 0.0032103725243359804, + 0.004454217851161957, + 0.008730954490602016, + 0.011432084254920483, + 0.00859607383608818, + 0.008788169361650944, + 0.010453339666128159, + 0.016592660918831825, + 0.005292592570185661, + 0.008243412710726261, + 0.007245667744427919, + 0.013128324411809444, + 0.008599624037742615, + 0.006728324573487043, + 0.07184099406003952, + 0.07642713934183121, + 0.0048373048193752766, + 0.007774658966809511, + 0.008255627006292343, + 0.017706451937556267, + 0.006273295264691114, + 0.005472844932228327, + 0.01668737828731537, + 0.01655406877398491, + 0.004598317667841911, + 0.004141099750995636, + 0.004000811371952295, + 0.0065556601621210575, + 0.005209142342209816, + 0.004313233774155378, + 0.008144787512719631, + 0.00781338382512331, + 0.003753668162971735, + 0.009278710000216961, + 0.0019996443297713995, + 0.0015778745291754603, + 0.0016076563624665141, + 0.0013669095933437347, + 0.002727821236476302, + 0.0024485825560986996, + 0.002177906921133399, + 0.0022779684513807297, + 0.014175225980579853, + 0.018973333761096, + 0.024782169610261917, + 0.024940496310591698, + 0.01962290331721306, + 0.0176401250064373, + 0.04192974418401718, + 0.019158227369189262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.011090068146586418 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38704338669776917 + ], + "opposite_logprob_tokens": [ + -1.1370433568954468 + ], + "logprob_reward": -0.38704338669776917, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38704338669776917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.355137437582016, + 0.07147087901830673, + 0.03423816338181496, + 0.15409912168979645, + 0.07558025419712067, + 0.14779219031333923, + 0.04605470970273018, + 0.025582116097211838, + 0.04140055552124977, + 0.04864463582634926 + ], + "credit_map": [ + 0.047981347888708115, + 0.07252118736505508, + 0.011220457032322884, + 0.03847949579358101, + 0.14726762473583221, + 0.01299446914345026, + 0.012937355786561966, + 0.01173546351492405, + 0.010052547790110111, + 0.008211693726480007, + 0.00914330966770649, + 0.008323060348629951, + 0.013916604220867157, + 0.00811702013015747, + 0.005786475725471973, + 0.007920156233012676, + 0.005275385919958353, + 0.006770613603293896, + 0.0038885835092514753, + 0.0029748131055384874, + 0.0037723586428910494, + 0.004344386979937553, + 0.004089373629540205, + 0.0031226444989442825, + 0.02007846161723137, + 0.01867934875190258, + 0.020332450047135353, + 0.012226040475070477, + 0.022702524438500404, + 0.03329072147607803, + 0.01649019494652748, + 0.01029935386031866, + 0.009314315393567085, + 0.01062684040516615, + 0.009503904730081558, + 0.007181728724390268, + 0.010346288792788982, + 0.016875559464097023, + 0.004431843291968107, + 0.0072997622191905975, + 0.006243762094527483, + 0.008893284946680069, + 0.005184423644095659, + 0.0038664236199110746, + 0.05464157462120056, + 0.059634897857904434, + 0.003449684474617243, + 0.005878125317394733, + 0.006830667611211538, + 0.007119601592421532, + 0.005584839265793562, + 0.006053805351257324, + 0.005590992048382759, + 0.005641130730509758, + 0.004553806036710739, + 0.0046798596158623695, + 0.0023204032331705093, + 0.0030583469197154045, + 0.0019657297525554895, + 0.0015555326826870441, + 0.00561893917620182, + 0.005170009098947048, + 0.0013677532551810145, + 0.0045253983698785305, + 0.003378656692802906, + 0.004692045971751213, + 0.005104736424982548, + 0.004532206803560257, + 0.005364629440009594, + 0.006114067975431681, + 0.005032241810113192, + 0.007181967608630657, + 0.0036969950888305902, + 0.005017008166760206, + 0.006921439431607723, + 0.006558120250701904, + 0.004993254318833351, + 0.006132859271019697, + 0.010353527031838894, + 0.004971426911652088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.38704338669776917 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033156722784042 + ], + "opposite_logprob_tokens": [ + -2.3503315448760986 + ], + "logprob_reward": -0.10033156722784042, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033156722784042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06962106376886368, + 0.051188528537750244, + 0.05759849771857262, + 0.06637833267450333, + 0.051268018782138824, + 0.25557300448417664, + 0.11890748143196106, + 0.0488465391099453, + 0.025412725284695625, + 0.25520578026771545 + ], + "credit_map": [ + 0.006753613241016865, + 0.019501972943544388, + 0.004401964135468006, + 0.005443926434963942, + 0.016222098842263222, + 0.004876823630183935, + 0.006099842023104429, + 0.00632081413641572, + 0.0061500840820372105, + 0.005875769536942244, + 0.010917210020124912, + 0.006668310146778822, + 0.004770995117723942, + 0.007729331497102976, + 0.0044413297437131405, + 0.004635498858988285, + 0.009138302877545357, + 0.01511706318706274, + 0.006766199134290218, + 0.0037274626083672047, + 0.007924051024019718, + 0.006074164528399706, + 0.005042470060288906, + 0.0038087833672761917, + 0.007525445893406868, + 0.007581131998449564, + 0.006923333276063204, + 0.006210508290678263, + 0.010378079488873482, + 0.012074724771082401, + 0.007652593310922384, + 0.008032518438994884, + 0.006134977098554373, + 0.007833357900381088, + 0.005522662773728371, + 0.005056178197264671, + 0.006701958365738392, + 0.011011172086000443, + 0.003750738687813282, + 0.005256970878690481, + 0.006993879564106464, + 0.012449763715267181, + 0.006660658400505781, + 0.005319346208125353, + 0.10277517884969711, + 0.10950375348329544, + 0.0036579896695911884, + 0.008212435990571976, + 0.011130527593195438, + 0.025336623191833496, + 0.009683240205049515, + 0.01009396556764841, + 0.024581817910075188, + 0.02347588539123535, + 0.008226185105741024, + 0.006379241589456797, + 0.004696239717304707, + 0.007065372075885534, + 0.006281869485974312, + 0.008104811422526836, + 0.005525058601051569, + 0.005934190470725298, + 0.004435799550265074, + 0.006803195923566818, + 0.0019701437558978796, + 0.0016400318127125502, + 0.002266420517116785, + 0.00195704004727304, + 0.006169392727315426, + 0.0021180687472224236, + 0.0022927820682525635, + 0.006998846307396889, + 0.020083751529455185, + 0.026890328153967857, + 0.03431512415409088, + 0.03322397544980049, + 0.028833571821451187, + 0.02607700787484646, + 0.05869124084711075, + 0.0270907711237669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.10033156722784042 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072618886828423 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072618886828423, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072618886828423, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12731176614761353, + 0.0638912245631218, + 0.049344439059495926, + 0.11945807188749313, + 0.07025810331106186, + 0.1393212378025055, + 0.037086643278598785, + 0.23800063133239746, + 0.05179087817668915, + 0.10353706777095795 + ], + "credit_map": [ + 0.020556824281811714, + 0.031213214620947838, + 0.008930893614888191, + 0.01272575743496418, + 0.020607715472579002, + 0.00824731681495905, + 0.014265577308833599, + 0.010764449834823608, + 0.010600283741950989, + 0.00838166382163763, + 0.013305173255503178, + 0.008008284494280815, + 0.008969267830252647, + 0.0060904622077941895, + 0.004382377956062555, + 0.00415370799601078, + 0.007194347679615021, + 0.01087083388119936, + 0.004971684888005257, + 0.0038257944397628307, + 0.005922734271734953, + 0.008107226341962814, + 0.0042610871605575085, + 0.0041907248087227345, + 0.01799261011183262, + 0.01650608517229557, + 0.01401558332145214, + 0.008010839112102985, + 0.019475283101201057, + 0.026470322161912918, + 0.00935974158346653, + 0.0076275900937616825, + 0.008444663137197495, + 0.009891157038509846, + 0.006763128563761711, + 0.007826260291039944, + 0.010036997497081757, + 0.01489783450961113, + 0.004647883120924234, + 0.007750173565000296, + 0.011223112232983112, + 0.01887270249426365, + 0.013112313114106655, + 0.009408984333276749, + 0.0348493717610836, + 0.03598926588892937, + 0.006105856504291296, + 0.00975960772484541, + 0.006741903722286224, + 0.007576454430818558, + 0.003979185596108437, + 0.002607855014503002, + 0.005762407090514898, + 0.005627727601677179, + 0.002399986842647195, + 0.002391116227954626, + 0.015670733526349068, + 0.02227606065571308, + 0.010721977800130844, + 0.010414157062768936, + 0.06946627795696259, + 0.06398668885231018, + 0.010001792572438717, + 0.03546291962265968, + 0.003120194189250469, + 0.005127826705574989, + 0.0031346397008746862, + 0.003022841876372695, + 0.012277236208319664, + 0.006301135756075382, + 0.004697458818554878, + 0.014109542593359947, + 0.008864962495863438, + 0.009643685072660446, + 0.01927466131746769, + 0.013369616121053696, + 0.01287997979670763, + 0.01175429206341505, + 0.01645461842417717, + 0.011295245960354805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.011072618886828423 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a fluid and elegant 360-degree panning shot, the image begins with a close-up of the muzzle of a Beretta 687 hunting rifle, a thin, curling plume of bluish-gray smoke rising upward. The gun's metal receiver, engraved with an intricate European scroll pattern, reflects a restrained silver luster in the warm, slanting golden sunlight of an autumn afternoon. The stock and handguard are crafted from deep-grained walnut, finished with a smooth, oil-waxed finish that imparts a warm reddish-brown hue. As the camera steadily pans backward and around the subject, we see the gun resting on an unseen stand against a vast, pastoral backdrop: golden grass undulating in the wind, and fallen leaves in ochre and crimson scattered across the ground. At that moment, a freshly fired projectile traces a nearly invisible, shimmering metallic trail through the air. It doesn't fly in a straight line, but rather, mid-flight, strikes an invisible, hard object at high speed, rebounding and changing direction. It ultimately strikes a weathered, round metal bull's eye in the distance at a treacherous angle, leaving a tiny, gleaming dent in the impact. The entire scene is bathed in a soft, warm light, with elongated shadows stretching across the ground, creating an atmosphere of tranquility, concentration, and a rustic feel, perfectly capturing the power and precision of this moment.", + "seed": 27549, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_3/index_8539_seed_27549_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03811879344284534, + "vlm_logprob": -0.03811879344284534, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.03811879344284534, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04092986881732941 + ], + "opposite_logprob_tokens": [ + -3.2909297943115234 + ], + "logprob_reward": -0.04092986881732941, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04092986881732941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36112692952156067, + 0.2538631558418274, + 0.09027467668056488, + 0.08400222659111023, + 0.027618294581770897, + 0.029510490596294403, + 0.028197363018989563, + 0.05633768066763878, + 0.03928269445896149, + 0.029786499217152596 + ], + "credit_map": [ + 0.03561365604400635, + 0.055636316537857056, + 0.015617634169757366, + 0.012629687786102295, + 0.03379529342055321, + 0.0694592297077179, + 0.029163600876927376, + 0.1092115193605423, + 0.036445219069719315, + 0.047016728669404984, + 0.026103204116225243, + 0.038352012634277344, + 0.017824582755565643, + 0.03066815808415413, + 0.02959258295595646, + 0.027860643342137337, + 0.009899935685098171, + 0.008179612457752228, + 0.014296622015535831, + 0.011068408377468586, + 0.00822079461067915, + 0.017640868201851845, + 0.01014102715998888, + 0.010827409103512764, + 0.009502390399575233, + 0.007105791009962559, + 0.011005247011780739, + 0.012117480859160423, + 0.012545815669000149, + 0.010034408420324326, + 0.011037672869861126, + 0.01065342128276825, + 0.001993180252611637, + 0.0035123967099934816, + 0.002917798236012459, + 0.0032384993974119425, + 0.002411405323073268, + 0.0034111402928829193, + 0.004037763923406601, + 0.006096112076193094, + 0.002524563344195485, + 0.004977855831384659, + 0.006742747500538826, + 0.0023955858778208494, + 0.0016735383542254567, + 0.005239698104560375, + 0.003531624795868993, + 0.0024248778354376554, + 0.002129203639924526, + 0.001827226486057043, + 0.004057781305164099, + 0.007027244195342064, + 0.0027562554460018873, + 0.0019275035010650754, + 0.0035460880026221275, + 0.004926059860736132, + 0.004294480197131634, + 0.005964219570159912, + 0.007346272934228182, + 0.004880065098404884, + 0.00606019189581275, + 0.011612850241363049, + 0.006071067880839109, + 0.010108535178005695, + 0.003278502030298114, + 0.004596443381160498, + 0.004402311984449625, + 0.005264983046799898, + 0.0068303002044558525, + 0.005772066302597523, + 0.00472218869253993, + 0.004415902309119701, + 0.004568761680275202, + 0.002588990842923522, + 0.004697620403021574, + 0.0032845467794686556, + 0.0043466659262776375, + 0.002334943041205406, + 0.002201103838160634, + 0.005763870198279619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04092986881732941 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059445157647133 + ], + "opposite_logprob_tokens": [ + -3.895594358444214 + ], + "logprob_reward": -0.02059445157647133, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059445157647133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21868877112865448, + 0.1024053767323494, + 0.06490156054496765, + 0.11515390872955322, + 0.06838049739599228, + 0.08873401582241058, + 0.13144473731517792, + 0.06400904804468155, + 0.08320876210927963, + 0.06307332217693329 + ], + "credit_map": [ + 0.03300481289625168, + 0.04204569384455681, + 0.012973751872777939, + 0.016053561121225357, + 0.019592033699154854, + 0.031333018094301224, + 0.01980138011276722, + 0.0438845269382, + 0.00926119927316904, + 0.016429925337433815, + 0.008704145438969135, + 0.013793472200632095, + 0.008203141391277313, + 0.013077709823846817, + 0.019962869584560394, + 0.012972919270396233, + 0.005618926603347063, + 0.008229321800172329, + 0.013486355543136597, + 0.006635110359638929, + 0.004660832230001688, + 0.008716760203242302, + 0.009296623058617115, + 0.008257630281150341, + 0.01723138429224491, + 0.010291804559528828, + 0.01416902244091034, + 0.015711821615695953, + 0.014873295091092587, + 0.015382574871182442, + 0.013169179670512676, + 0.01432482898235321, + 0.004696269519627094, + 0.00766550051048398, + 0.008655219338834286, + 0.009296627715229988, + 0.005447728559374809, + 0.009002692997455597, + 0.011402552016079426, + 0.012213901616632938, + 0.007127296645194292, + 0.014196302741765976, + 0.020735228434205055, + 0.008250693790614605, + 0.00592616805806756, + 0.009735509753227234, + 0.012897457927465439, + 0.009865366853773594, + 0.010906221345067024, + 0.010822408832609653, + 0.018295060843229294, + 0.029564060270786285, + 0.009011203423142433, + 0.009735440835356712, + 0.017915619537234306, + 0.025194721296429634, + 0.005360457580536604, + 0.006954368203878403, + 0.008587094023823738, + 0.010314795188605785, + 0.006382959429174662, + 0.00732513889670372, + 0.008375018835067749, + 0.010709214024245739, + 0.005901336669921875, + 0.009760944172739983, + 0.011086287908256054, + 0.013826346956193447, + 0.012215889059007168, + 0.0074861738830804825, + 0.014550494961440563, + 0.008381287567317486, + 0.009147492237389088, + 0.0064341542311012745, + 0.010153576731681824, + 0.006180221680551767, + 0.009572946466505527, + 0.004080071114003658, + 0.005967552773654461, + 0.011537306010723114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.02059445157647133 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001058757305145 + ], + "opposite_logprob_tokens": [ + -2.6950106620788574 + ], + "logprob_reward": -0.07001058757305145, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001058757305145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2623106837272644, + 0.08541517704725266, + 0.10275441408157349, + 0.072770856320858, + 0.042289990931749344, + 0.14169621467590332, + 0.08552148938179016, + 0.07811921834945679, + 0.09212809801101685, + 0.0369938500225544 + ], + "credit_map": [ + 0.023968836292624474, + 0.06274339556694031, + 0.010110077448189259, + 0.0073014600202441216, + 0.025037385523319244, + 0.0335085429251194, + 0.019939756020903587, + 0.0797012448310852, + 0.014971163123846054, + 0.01556948572397232, + 0.008344646543264389, + 0.007989120669662952, + 0.007106291130185127, + 0.011548563838005066, + 0.0103828189894557, + 0.009503082372248173, + 0.011166625656187534, + 0.009275306016206741, + 0.013454490341246128, + 0.01152859814465046, + 0.009813771583139896, + 0.01960841752588749, + 0.015398393385112286, + 0.012508806772530079, + 0.009892765432596207, + 0.007216526195406914, + 0.007399630267173052, + 0.00728990463539958, + 0.00689523434266448, + 0.014694079756736755, + 0.011050187982618809, + 0.008332528173923492, + 0.004847904201596975, + 0.0053926329128444195, + 0.005016316659748554, + 0.005902912002056837, + 0.0041009411215782166, + 0.006603753659874201, + 0.005165026057511568, + 0.005260503385215998, + 0.010866723954677582, + 0.01518158707767725, + 0.022932272404432297, + 0.015349796041846275, + 0.012270193547010422, + 0.031398992985486984, + 0.018201997503638268, + 0.015494653955101967, + 0.009118923917412758, + 0.009571843780577183, + 0.01088490430265665, + 0.017483148723840714, + 0.007845887914299965, + 0.006380104925483465, + 0.014374998398125172, + 0.009861680679023266, + 0.006141724996268749, + 0.009665091522037983, + 0.009820998646318913, + 0.014746662229299545, + 0.01048479788005352, + 0.008443891070783138, + 0.009774377569556236, + 0.009041679091751575, + 0.007738116662949324, + 0.008471780456602573, + 0.017201071605086327, + 0.01275421492755413, + 0.011668670922517776, + 0.008521486073732376, + 0.01100360881537199, + 0.014769150875508785, + 0.0052572754211723804, + 0.00498063862323761, + 0.004636381287127733, + 0.005031357053667307, + 0.005091035272926092, + 0.0038248116616159678, + 0.0035587064921855927, + 0.00461364584043622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.07001058757305145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496795102953911 + ], + "opposite_logprob_tokens": [ + -2.9299678802490234 + ], + "logprob_reward": -0.05496795102953911, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496795102953911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2786734700202942, + 0.10670329630374908, + 0.12108716368675232, + 0.06104755401611328, + 0.08388925343751907, + 0.11642000824213028, + 0.06814438849687576, + 0.048731714487075806, + 0.05409552901983261, + 0.061207689344882965 + ], + "credit_map": [ + 0.029110077768564224, + 0.0847523957490921, + 0.012495866976678371, + 0.011194663122296333, + 0.024167776107788086, + 0.07604797929525375, + 0.011821499094367027, + 0.029083214700222015, + 0.011668382212519646, + 0.020318450406193733, + 0.00891958735883236, + 0.015046589076519012, + 0.006842455826699734, + 0.012127124704420567, + 0.018705016002058983, + 0.0130756925791502, + 0.014762385748326778, + 0.010976262390613556, + 0.019797207787632942, + 0.013008179143071175, + 0.01040908508002758, + 0.018713325262069702, + 0.01834098808467388, + 0.015079726465046406, + 0.00912772212177515, + 0.005921827629208565, + 0.008188145235180855, + 0.007089576683938503, + 0.004611155483871698, + 0.01020786352455616, + 0.006889566779136658, + 0.00901169702410698, + 0.0058815437369048595, + 0.009991600178182125, + 0.00868789479136467, + 0.013028288260102272, + 0.0077643441036343575, + 0.009868663735687733, + 0.010492924600839615, + 0.01817399635910988, + 0.01084899716079235, + 0.02077663131058216, + 0.022385306656360626, + 0.007465170230716467, + 0.006735565606504679, + 0.024208666756749153, + 0.012878569774329662, + 0.011121106334030628, + 0.00632805610075593, + 0.006404026877135038, + 0.010312359780073166, + 0.00838293693959713, + 0.009055765345692635, + 0.006494392640888691, + 0.008671270683407784, + 0.012495579198002815, + 0.0050801164470613, + 0.005747512448579073, + 0.005461299791932106, + 0.007117476314306259, + 0.0063259899616241455, + 0.0058025275357067585, + 0.00549037242308259, + 0.007706419564783573, + 0.004333628807216883, + 0.005183493252843618, + 0.007324771489948034, + 0.007009882479906082, + 0.01174586545675993, + 0.005493642762303352, + 0.006863746326416731, + 0.006140499841421843, + 0.007919852621853352, + 0.008263888768851757, + 0.008877597749233246, + 0.0052400799468159676, + 0.009258273988962173, + 0.007018554490059614, + 0.004061069805175066, + 0.01056837011128664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.05496795102953911 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Beretta 687 hunting rifle with a muzzle, metal receiver, and walnut stock visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a thin, curling plume of bluish-gray smoke rising from the rifle's muzzle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a 360-degree pan that starts with a close-up and moves backward and around the subject?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a projectile visible tracing a shimmering metallic trail that strikes an object and rebounds before hitting a distant metal bull's eye?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a vast pastoral scene with golden grass and fallen leaves in ochre and crimson under warm sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091108217835426 + ], + "opposite_logprob_tokens": [ + -5.504091262817383 + ], + "logprob_reward": -0.004091108217835426, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091108217835426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26929134130477905, + 0.11175015568733215, + 0.07164464145898819, + 0.19694483280181885, + 0.050130169838666916, + 0.04328613728284836, + 0.02783031389117241, + 0.050006963312625885, + 0.09785732626914978, + 0.08125808835029602 + ], + "credit_map": [ + 0.041420094668865204, + 0.03928738087415695, + 0.01500205509364605, + 0.01917272061109543, + 0.028440287336707115, + 0.05083540827035904, + 0.021918270736932755, + 0.05321512371301651, + 0.012224560603499413, + 0.020770207047462463, + 0.013428626582026482, + 0.018459081649780273, + 0.008033758960664272, + 0.014321205206215382, + 0.012548350729048252, + 0.01196436770260334, + 0.009935877285897732, + 0.007640270050615072, + 0.010096701793372631, + 0.009271961636841297, + 0.005551514215767384, + 0.011460769921541214, + 0.009469067677855492, + 0.00821848027408123, + 0.02055574394762516, + 0.019529661163687706, + 0.033883221447467804, + 0.016631167382001877, + 0.014297399669885635, + 0.035140641033649445, + 0.02854377217590809, + 0.02836320735514164, + 0.004613429307937622, + 0.005296585615724325, + 0.006909988354891539, + 0.006250807084143162, + 0.004105014260858297, + 0.006235967855900526, + 0.008661240339279175, + 0.008057140745222569, + 0.0034338622353971004, + 0.005296030081808567, + 0.007985607720911503, + 0.0036320448853075504, + 0.0022265035659074783, + 0.00941020529717207, + 0.005703672766685486, + 0.005598211660981178, + 0.004535900428891182, + 0.0035062201786786318, + 0.0038676138501614332, + 0.00373921450227499, + 0.00327678513713181, + 0.002266770927235484, + 0.0032759502064436674, + 0.003361858893185854, + 0.003605486359447241, + 0.005766195245087147, + 0.006587882526218891, + 0.003336386289447546, + 0.005947201047092676, + 0.012431824579834938, + 0.00482903141528368, + 0.007502959109842777, + 0.007690418977290392, + 0.009574146009981632, + 0.017162291333079338, + 0.010884124785661697, + 0.012270916253328323, + 0.015848763287067413, + 0.009743858128786087, + 0.014682804234325886, + 0.018295103684067726, + 0.00941407959908247, + 0.012584126554429531, + 0.006107880733907223, + 0.007860307581722736, + 0.010262293741106987, + 0.006046738009899855, + 0.010687565430998802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.004091108217835426 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_4/index_tensor([8058, 8058])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_4/index_tensor([8058, 8058])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ee1af19adaa9bd9d9ce7fe91ff3aa1dc90639675 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_4/index_tensor([8058, 8058])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a warm, Studio Ghibli-inspired style, a steady, eye-level, medium tracking shot follows a young woman as she carefully crosses a sun-dappled cobblestone street. Her dark brown hair is pulled back in a simple, loose braid, resting on the shoulders of a cream cotton shirt. With gentle focus, she pedals steadily and deliberately on a classic Dutch freight tricycle, its light blue frame showing subtle, charming signs of wear. The large wooden cargo box out front is filled with a vibrant, picturesque display of fresh produce and flowers; vibrant yellow sunflowers and purple lavender stalks jut out from between heads of crisp green lettuce, bright red tomatoes, and round onions. As the tricycle moves, its wheels clattering gently against the uneven, mossy cobblestones, the camera sliding parallel to it, capturing the shifting patterns of golden light filtering through the leaves of the trees overhead. The setting shows a quaint storybook town square with half-timbered buildings, terracotta roofs, and window boxes blooming with geraniums, while the other gentle townspeople—an elderly couple walking arm in arm, a baker dusting flour from his apron—add to the scene’s serene, painterly atmosphere of peaceful daily life.", + "seed": 7663, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_4/index_8058_seed_7663_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03412394523620606, + "vlm_logprob": -0.03412394523620606, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03412394523620606, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05577179044485092 + ], + "opposite_logprob_tokens": [ + -2.930771827697754 + ], + "logprob_reward": -0.05577179044485092, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05577179044485092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3621313273906708, + 0.13312901556491852, + 0.1165374219417572, + 0.06755685806274414, + 0.03211911395192146, + 0.05945030227303505, + 0.07075975090265274, + 0.0639331042766571, + 0.06534082442522049, + 0.029042281210422516 + ], + "credit_map": [ + 0.07500433176755905, + 0.06716390699148178, + 0.029811562970280647, + 0.049948468804359436, + 0.01963852532207966, + 0.019952785223722458, + 0.030318086966872215, + 0.07029365003108978, + 0.02137152850627899, + 0.016515037044882774, + 0.01217514369636774, + 0.01882275566458702, + 0.01113711204379797, + 0.018431005999445915, + 0.013640480116009712, + 0.021035954356193542, + 0.01104824710637331, + 0.013587972149252892, + 0.016625095158815384, + 0.013704527169466019, + 0.016531934961676598, + 0.012649036012589931, + 0.02034471184015274, + 0.012045889161527157, + 0.005511531140655279, + 0.012587768025696278, + 0.008304482325911522, + 0.0075997416861355305, + 0.008091721683740616, + 0.010088388808071613, + 0.006673468742519617, + 0.008699753321707249, + 0.0025505449157208204, + 0.003413778031244874, + 0.004764694720506668, + 0.004691170994192362, + 0.004578587599098682, + 0.004757973365485668, + 0.004259171430021524, + 0.00310319010168314, + 0.006061160936951637, + 0.00477064261212945, + 0.004988464526832104, + 0.011266903951764107, + 0.007010811008512974, + 0.010305442847311497, + 0.008401594124734402, + 0.006645285990089178, + 0.00765709625557065, + 0.006656507961452007, + 0.010512594133615494, + 0.010392851196229458, + 0.009035453200340271, + 0.0073793185874819756, + 0.011160573922097683, + 0.007965349592268467, + 0.005285853054374456, + 0.006573123391717672, + 0.009766480885446072, + 0.011217175051569939, + 0.006664311047643423, + 0.004336056765168905, + 0.011691447347402573, + 0.00839865580201149, + 0.007754696067422628, + 0.008215049281716347, + 0.01198554690927267, + 0.005974607076495886, + 0.00807109847664833, + 0.007128119468688965, + 0.006779660936444998, + 0.009432049468159676, + 0.0032544112764298916, + 0.0043995995074510574, + 0.003181030508130789, + 0.003998837899416685, + 0.00319785438477993, + 0.004251545760780573, + 0.003878990886732936, + 0.002880012383684516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.05577179044485092 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097922682762146 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.0097922682762146, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097922682762146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2610834240913391, + 0.06709136068820953, + 0.04401851445436478, + 0.08438941836357117, + 0.06403873860836029, + 0.10869685560464859, + 0.08518602699041367, + 0.10534972697496414, + 0.1425561010837555, + 0.037589799612760544 + ], + "credit_map": [ + 0.05208798125386238, + 0.04173656925559044, + 0.023247459903359413, + 0.0360591784119606, + 0.01862875185906887, + 0.016696887090802193, + 0.02118772082030773, + 0.051438890397548676, + 0.012439959682524204, + 0.005892021581530571, + 0.006852020043879747, + 0.008733801543712616, + 0.004707934334874153, + 0.009244455955922604, + 0.004925878252834082, + 0.014295287430286407, + 0.005468765739351511, + 0.0038107361178845167, + 0.007163005881011486, + 0.004887500312179327, + 0.006476413458585739, + 0.005039104726165533, + 0.006465023383498192, + 0.004707964137196541, + 0.007610791362822056, + 0.011286352761089802, + 0.015729844570159912, + 0.009407488629221916, + 0.009021292440593243, + 0.013203267939388752, + 0.008010606281459332, + 0.01011977531015873, + 0.004823083057999611, + 0.005955271422863007, + 0.007203808519989252, + 0.006114202551543713, + 0.00894597265869379, + 0.011935019865632057, + 0.008665541186928749, + 0.010395841673016548, + 0.011736888438463211, + 0.008457044139504433, + 0.010618268512189388, + 0.021320084109902382, + 0.012812631204724312, + 0.018899869173765182, + 0.01254335418343544, + 0.01230872143059969, + 0.009952746331691742, + 0.009199379943311214, + 0.0075914207845926285, + 0.01545415073633194, + 0.010487969033420086, + 0.008313501253724098, + 0.016372578218579292, + 0.00781428162008524, + 0.010096056386828423, + 0.015135487541556358, + 0.01110038161277771, + 0.021676285192370415, + 0.013707233592867851, + 0.008920530788600445, + 0.013718741945922375, + 0.010995014570653439, + 0.01614796370267868, + 0.020771611481904984, + 0.02057054452598095, + 0.01572476513683796, + 0.013547180220484734, + 0.017525946721434593, + 0.022018158808350563, + 0.016249923035502434, + 0.004285491071641445, + 0.005913231056183577, + 0.004418774042278528, + 0.005297038704156876, + 0.004279993008822203, + 0.006073363590985537, + 0.004061567131429911, + 0.003260342637076974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.0097922682762146 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250474438071251 + ], + "opposite_logprob_tokens": [ + -5.255250453948975 + ], + "logprob_reward": -0.005250474438071251, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250474438071251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12492676824331284, + 0.07448423653841019, + 0.04181746393442154, + 0.06166839972138405, + 0.03896957263350487, + 0.1344117820262909, + 0.09634963423013687, + 0.07604122906923294, + 0.30348753929138184, + 0.0478433258831501 + ], + "credit_map": [ + 0.012217736802995205, + 0.022754503414034843, + 0.014440476894378662, + 0.024624431505799294, + 0.011276958510279655, + 0.011635432951152325, + 0.017015712335705757, + 0.010961524210870266, + 0.011119168251752853, + 0.009408830665051937, + 0.005815363954752684, + 0.012074582278728485, + 0.005751345306634903, + 0.008080818690359592, + 0.00807817094027996, + 0.01415595505386591, + 0.005513943266123533, + 0.004231373313814402, + 0.006487097125500441, + 0.004730473272502422, + 0.006571829319000244, + 0.004824171774089336, + 0.004696202464401722, + 0.004762370605021715, + 0.006287155672907829, + 0.007014847360551357, + 0.007954240776598454, + 0.005553594324737787, + 0.00659359572455287, + 0.010673175565898418, + 0.007482816465198994, + 0.010108974762260914, + 0.004497820045799017, + 0.0030625429935753345, + 0.003841431811451912, + 0.003268128726631403, + 0.006529474630951881, + 0.006477526854723692, + 0.005830054171383381, + 0.00546259293332696, + 0.0097417701035738, + 0.00982613768428564, + 0.008313486352562904, + 0.016783181577920914, + 0.017440972849726677, + 0.03657595068216324, + 0.013257023878395557, + 0.02247326262295246, + 0.016758976504206657, + 0.008160313591361046, + 0.010325294919312, + 0.009711137972772121, + 0.014247396029531956, + 0.015561413019895554, + 0.011444720439612865, + 0.010140384547412395, + 0.006628008559346199, + 0.007563613820821047, + 0.008841636590659618, + 0.008417148143053055, + 0.012801016680896282, + 0.008752714842557907, + 0.014352693222463131, + 0.008684404194355011, + 0.0350811667740345, + 0.03273102641105652, + 0.0419730618596077, + 0.04747204855084419, + 0.0244239903986454, + 0.04267808049917221, + 0.043724168092012405, + 0.035403985530138016, + 0.005399847403168678, + 0.005162978544831276, + 0.0059617008082568645, + 0.007281774654984474, + 0.005241024773567915, + 0.008178439922630787, + 0.005086742807179689, + 0.005530820228159428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.005250474438071251 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002515345811844 + ], + "opposite_logprob_tokens": [ + -2.6950252056121826 + ], + "logprob_reward": -0.07002515345811844, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002515345811844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23812490701675415, + 0.0627962052822113, + 0.04157641530036926, + 0.04379129409790039, + 0.06081143021583557, + 0.06024513393640518, + 0.08357499539852142, + 0.08278687298297882, + 0.26496419310569763, + 0.061328474432229996 + ], + "credit_map": [ + 0.04654674604535103, + 0.03684331476688385, + 0.020134350284934044, + 0.03240642696619034, + 0.019234975799918175, + 0.014483892358839512, + 0.022216880694031715, + 0.04625832289457321, + 0.004932214971631765, + 0.011207366362214088, + 0.005353136453777552, + 0.012166941538453102, + 0.006542900111526251, + 0.007218176499009132, + 0.008956321515142918, + 0.006419155281037092, + 0.004470175132155418, + 0.005628699902445078, + 0.007182943634688854, + 0.005457503721117973, + 0.006318128202110529, + 0.004681374877691269, + 0.004419355653226376, + 0.003418239764869213, + 0.006579244043678045, + 0.004323997534811497, + 0.005642182193696499, + 0.004485032521188259, + 0.006015475373715162, + 0.0057405224069952965, + 0.005150360055267811, + 0.005854483228176832, + 0.004565320909023285, + 0.0057404120452702045, + 0.007125936448574066, + 0.010290846228599548, + 0.01097389217466116, + 0.00782705843448639, + 0.00948505848646164, + 0.0048029134050011635, + 0.0074716112576425076, + 0.004294398706406355, + 0.00420868257060647, + 0.006523707415908575, + 0.011069145053625107, + 0.011682738550007343, + 0.006796333473175764, + 0.00819852203130722, + 0.011141967959702015, + 0.005578214768320322, + 0.011286468245089054, + 0.01273982971906662, + 0.010606341995298862, + 0.008232834748923779, + 0.015296515077352524, + 0.008692828938364983, + 0.007451516576111317, + 0.007380045484751463, + 0.009220513515174389, + 0.009898095391690731, + 0.009756302461028099, + 0.010293177329003811, + 0.01963200233876705, + 0.009155231527984142, + 0.026180142536759377, + 0.029117653146386147, + 0.03640313446521759, + 0.030561745166778564, + 0.0309754591435194, + 0.031243516132235527, + 0.035890188068151474, + 0.04459238424897194, + 0.006709866225719452, + 0.011408843100070953, + 0.00907136034220457, + 0.008296474814414978, + 0.006123983766883612, + 0.007192729506641626, + 0.005773452110588551, + 0.006751773413270712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.07002515345811844 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029780039563775063 + ], + "opposite_logprob_tokens": [ + -3.529780149459839 + ], + "logprob_reward": -0.029780039563775063, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029780039563775063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.277505099773407, + 0.07933855056762695, + 0.02867300435900688, + 0.08694175630807877, + 0.07890518754720688, + 0.1393827646970749, + 0.054366014897823334, + 0.10115111619234085, + 0.046539612114429474, + 0.10719695687294006 + ], + "credit_map": [ + 0.05975952744483948, + 0.05246717855334282, + 0.020828215405344963, + 0.03581518307328224, + 0.018607279285788536, + 0.018433742225170135, + 0.021971717476844788, + 0.049622271209955215, + 0.006472432054579258, + 0.009556915611028671, + 0.007475248072296381, + 0.013985542580485344, + 0.010836637578904629, + 0.012479103170335293, + 0.01130341924726963, + 0.007229247596114874, + 0.0034134394954890013, + 0.0029655625112354755, + 0.004433060064911842, + 0.004402178805321455, + 0.003929519560188055, + 0.003506998298689723, + 0.0033834525384008884, + 0.002638793084770441, + 0.008210903033614159, + 0.013321049511432648, + 0.011817960999906063, + 0.009540234692394733, + 0.010565970093011856, + 0.010530458763241768, + 0.010988245718181133, + 0.011966932564973831, + 0.011300274170935154, + 0.008017378859221935, + 0.013121268711984158, + 0.0073735578916966915, + 0.01116909645497799, + 0.01019988115876913, + 0.00912438239902258, + 0.008599342778325081, + 0.010767138563096523, + 0.013956262730062008, + 0.013952601701021194, + 0.02795000933110714, + 0.017555082216858864, + 0.023445960134267807, + 0.014042720198631287, + 0.017712995409965515, + 0.005372136831283569, + 0.005093897692859173, + 0.007372226566076279, + 0.007196352351456881, + 0.007149364333599806, + 0.0062037268653512, + 0.010034862905740738, + 0.005943445023149252, + 0.010269511491060257, + 0.00974392518401146, + 0.01877029798924923, + 0.016197234392166138, + 0.014016365632414818, + 0.009270097129046917, + 0.012695446610450745, + 0.010188236832618713, + 0.005722395610064268, + 0.004557015839964151, + 0.0062061636708676815, + 0.006850135046988726, + 0.0055327522568404675, + 0.005852740723639727, + 0.0057375747710466385, + 0.006080834195017815, + 0.010725934989750385, + 0.023149540647864342, + 0.009861372411251068, + 0.016660397872328758, + 0.010840332135558128, + 0.009187238290905952, + 0.014745510183274746, + 0.012026633135974407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.029780039563775063 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a warm, Studio Ghibli-inspired style, a steady, eye-level, medium tracking shot follows a young woman as she carefully crosses a sun-dappled cobblestone street. Her dark brown hair is pulled back in a simple, loose braid, resting on the shoulders of a cream cotton shirt. With gentle focus, she pedals steadily and deliberately on a classic Dutch freight tricycle, its light blue frame showing subtle, charming signs of wear. The large wooden cargo box out front is filled with a vibrant, picturesque display of fresh produce and flowers; vibrant yellow sunflowers and purple lavender stalks jut out from between heads of crisp green lettuce, bright red tomatoes, and round onions. As the tricycle moves, its wheels clattering gently against the uneven, mossy cobblestones, the camera sliding parallel to it, capturing the shifting patterns of golden light filtering through the leaves of the trees overhead. The setting shows a quaint storybook town square with half-timbered buildings, terracotta roofs, and window boxes blooming with geraniums, while the other gentle townspeople—an elderly couple walking arm in arm, a baker dusting flour from his apron—add to the scene’s serene, painterly atmosphere of peaceful daily life.", + "seed": 7664, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_4/index_8058_seed_7664_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03627109257504344, + "vlm_logprob": -0.03627109257504344, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.03627109257504344, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04984836280345917 + ], + "opposite_logprob_tokens": [ + -3.0498483180999756 + ], + "logprob_reward": -0.04984836280345917, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04984836280345917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15338364243507385, + 0.06231981888413429, + 0.08623506873846054, + 0.06537012755870819, + 0.029969658702611923, + 0.07995616644620895, + 0.060056500136852264, + 0.06311576068401337, + 0.13090816140174866, + 0.26868510246276855 + ], + "credit_map": [ + 0.019000183790922165, + 0.020566824823617935, + 0.023051638156175613, + 0.020647786557674408, + 0.015038219280540943, + 0.014993103221058846, + 0.01871010847389698, + 0.021375777199864388, + 0.009413516148924828, + 0.0050900643691420555, + 0.009743333794176579, + 0.006245432421565056, + 0.006719842553138733, + 0.004137557931244373, + 0.014048177748918533, + 0.0069218953140079975, + 0.009663749486207962, + 0.008509447798132896, + 0.009662800468504429, + 0.013752569444477558, + 0.010727286338806152, + 0.007058301474899054, + 0.015749603509902954, + 0.011111312545835972, + 0.0056756408885121346, + 0.007547457702457905, + 0.009128029458224773, + 0.009715406224131584, + 0.006182070821523666, + 0.007490879390388727, + 0.010692110285162926, + 0.008938541635870934, + 0.005097696091979742, + 0.003279627999290824, + 0.0054911058396101, + 0.003267622785642743, + 0.0026376680471003056, + 0.0028969510458409786, + 0.0040566264651715755, + 0.0032423618249595165, + 0.007152662146836519, + 0.007825841195881367, + 0.012082505039870739, + 0.010989870876073837, + 0.015108209103345871, + 0.008841745555400848, + 0.008100900799036026, + 0.009854425676167011, + 0.00523220794275403, + 0.004684419371187687, + 0.007873233407735825, + 0.006539307534694672, + 0.008444118313491344, + 0.005096794106066227, + 0.006139048840850592, + 0.016047367826104164, + 0.004334594123065472, + 0.0064833504147827625, + 0.01003403402864933, + 0.007700161077082157, + 0.008957646787166595, + 0.006790070794522762, + 0.009104137308895588, + 0.00971176102757454, + 0.013439735397696495, + 0.012675853446125984, + 0.012204490602016449, + 0.019593317061662674, + 0.019525766372680664, + 0.01957813650369644, + 0.015860646963119507, + 0.01803021878004074, + 0.018650783225893974, + 0.022546982392668724, + 0.06897386163473129, + 0.016406381502747536, + 0.018823374062776566, + 0.08746156096458435, + 0.010427521541714668, + 0.02539464645087719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.04984836280345917 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009796281345188618 + ], + "opposite_logprob_tokens": [ + -4.634796142578125 + ], + "logprob_reward": -0.009796281345188618, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009796281345188618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21396712958812714, + 0.10039345175027847, + 0.07215643674135208, + 0.08965596556663513, + 0.041517119854688644, + 0.05634106323122978, + 0.0673980638384819, + 0.04848235473036766, + 0.2133333683013916, + 0.09675505012273788 + ], + "credit_map": [ + 0.027744049206376076, + 0.028958890587091446, + 0.03408052772283554, + 0.029284140095114708, + 0.016892611980438232, + 0.017786091193556786, + 0.02846531942486763, + 0.030755488201975822, + 0.010136246681213379, + 0.007018571719527245, + 0.016310740262269974, + 0.012278841808438301, + 0.008378288708627224, + 0.0074129956774413586, + 0.02499418333172798, + 0.013863588683307171, + 0.008357644081115723, + 0.006827450357377529, + 0.009992407634854317, + 0.010349315591156483, + 0.007423504255712032, + 0.009261762723326683, + 0.010873790830373764, + 0.009070556610822678, + 0.009367261081933975, + 0.011530326679348946, + 0.013152073137462139, + 0.01582009345293045, + 0.007808193098753691, + 0.008131378330290318, + 0.012370140291750431, + 0.011476502753794193, + 0.006362785119563341, + 0.0047943261452019215, + 0.008635681122541428, + 0.004682114347815514, + 0.0031958508770912886, + 0.003170862328261137, + 0.005624954588711262, + 0.005050547420978546, + 0.004343015141785145, + 0.00504175154492259, + 0.010457571595907211, + 0.010860857553780079, + 0.006379485130310059, + 0.006691527087241411, + 0.004910489544272423, + 0.0076563600450754166, + 0.00950151402503252, + 0.00920530129224062, + 0.009071329608559608, + 0.00891053769737482, + 0.00936894677579403, + 0.006137572694569826, + 0.009817922487854958, + 0.005384942051023245, + 0.004977962467819452, + 0.006139705888926983, + 0.007121099624782801, + 0.008113444782793522, + 0.005209926050156355, + 0.006323506124317646, + 0.004541535396128893, + 0.006055171135812998, + 0.02635994926095009, + 0.022105664014816284, + 0.026821551844477654, + 0.040919575840234756, + 0.013854271732270718, + 0.02703588828444481, + 0.02056722156703472, + 0.03566925227642059, + 0.008001497015357018, + 0.011880057863891125, + 0.020370539277791977, + 0.006898059975355864, + 0.006170358508825302, + 0.025253837928175926, + 0.00594742177054286, + 0.012233288027346134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.009796281345188618 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636608529835939 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636608529835939, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636608529835939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30762386322021484, + 0.05931659787893295, + 0.11530890315771103, + 0.06945452094078064, + 0.032949499785900116, + 0.05891906097531319, + 0.06180150434374809, + 0.041368111968040466, + 0.06588955223560333, + 0.18736831843852997 + ], + "credit_map": [ + 0.04710421711206436, + 0.03600630909204483, + 0.052444715052843094, + 0.03604061156511307, + 0.027028176933526993, + 0.026329366490244865, + 0.03862619772553444, + 0.04404427856206894, + 0.0073927175253629684, + 0.005501617677509785, + 0.009783696383237839, + 0.006431975401937962, + 0.006746819708496332, + 0.0050015803426504135, + 0.011680054478347301, + 0.006778138689696789, + 0.013591551221907139, + 0.016457628458738327, + 0.008816419169306755, + 0.009440099820494652, + 0.027660878375172615, + 0.01310416404157877, + 0.014911588281393051, + 0.011326578445732594, + 0.008006459102034569, + 0.007580915931612253, + 0.010040460154414177, + 0.00869846623390913, + 0.006621405482292175, + 0.009077023714780807, + 0.010449651628732681, + 0.008980141021311283, + 0.00678408145904541, + 0.004174010828137398, + 0.0044814469292759895, + 0.002920580329373479, + 0.003203587606549263, + 0.0033308244310319424, + 0.004999361000955105, + 0.0030556079000234604, + 0.005921922158449888, + 0.004607229493558407, + 0.0072950622998178005, + 0.005274749360978603, + 0.0116655882447958, + 0.009340686723589897, + 0.006383497733622789, + 0.00843032542616129, + 0.007384531199932098, + 0.005274286027997732, + 0.008465390652418137, + 0.009369195438921452, + 0.006591839250177145, + 0.0063191684894263744, + 0.006494974717497826, + 0.011902122758328915, + 0.0028755732346326113, + 0.0071551185101270676, + 0.0062752231024205685, + 0.006118792574852705, + 0.004899953957647085, + 0.005180174484848976, + 0.00448587303981185, + 0.004377404693514109, + 0.009622643701732159, + 0.007614858914166689, + 0.007042727433145046, + 0.010228382423520088, + 0.004974301438778639, + 0.009850273840129375, + 0.006222425494343042, + 0.010333938524127007, + 0.013813963159918785, + 0.017442960292100906, + 0.04630095139145851, + 0.012130863964557648, + 0.01184500940144062, + 0.05797438323497772, + 0.009114478714764118, + 0.018745727837085724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.004636608529835939 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899561524391174 + ], + "opposite_logprob_tokens": [ + -2.578995704650879 + ], + "logprob_reward": -0.07899561524391174, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899561524391174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14328482747077942, + 0.04100814834237099, + 0.034677814692258835, + 0.10042495280504227, + 0.024542201310396194, + 0.09880859404802322, + 0.04134682938456535, + 0.05109400674700737, + 0.06308309733867645, + 0.40172943472862244 + ], + "credit_map": [ + 0.01651701331138611, + 0.01814831979572773, + 0.02277165837585926, + 0.021195480599999428, + 0.010181639343500137, + 0.012546617537736893, + 0.019943993538618088, + 0.02198011800646782, + 0.006187766790390015, + 0.003630200633779168, + 0.005195592995733023, + 0.006638285703957081, + 0.004654749762266874, + 0.0026097216177731752, + 0.0069152433425188065, + 0.00517659168690443, + 0.003791955765336752, + 0.004191417247056961, + 0.0033033546060323715, + 0.0036161032039672136, + 0.00829694326967001, + 0.0033292225562036037, + 0.004230007063597441, + 0.003918811678886414, + 0.010936307720839977, + 0.008558657020330429, + 0.014336039312183857, + 0.012212911620736122, + 0.009331751614809036, + 0.011163152754306793, + 0.01958269253373146, + 0.014303446747362614, + 0.0047346740029752254, + 0.003536779200658202, + 0.0033691017888486385, + 0.002219598973169923, + 0.001897524925880134, + 0.0024643088690936565, + 0.003333005355671048, + 0.002987211337313056, + 0.010360031388700008, + 0.00992530956864357, + 0.013577080331742764, + 0.009331402368843555, + 0.021437235176563263, + 0.009635411202907562, + 0.011714364401996136, + 0.012827767990529537, + 0.003931399900466204, + 0.004381976556032896, + 0.004912397358566523, + 0.006849451456218958, + 0.005165823735296726, + 0.004186094738543034, + 0.004689895547926426, + 0.007229792419821024, + 0.003916394431143999, + 0.008247233927249908, + 0.006714961491525173, + 0.009078271687030792, + 0.004971747286617756, + 0.005700760055333376, + 0.004974845796823502, + 0.007489793933928013, + 0.006900472100824118, + 0.006866761948913336, + 0.008178605698049068, + 0.009274134412407875, + 0.005607958417385817, + 0.010213636793196201, + 0.007263634353876114, + 0.008777898736298084, + 0.02820940501987934, + 0.02910074032843113, + 0.10678078234195709, + 0.022864893078804016, + 0.028114035725593567, + 0.13820256292819977, + 0.012452680617570877, + 0.03600437939167023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.07899561524391174 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03807859495282173 + ], + "opposite_logprob_tokens": [ + -3.288078546524048 + ], + "logprob_reward": -0.03807859495282173, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03807859495282173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17189869284629822, + 0.05520780012011528, + 0.10883767902851105, + 0.062005192041397095, + 0.026036452502012253, + 0.033465348184108734, + 0.055816199630498886, + 0.021940013393759727, + 0.03916550055146217, + 0.42562708258628845 + ], + "credit_map": [ + 0.013772835023701191, + 0.019017506390810013, + 0.02449028752744198, + 0.02540299855172634, + 0.010791763663291931, + 0.013146926648914814, + 0.03908843547105789, + 0.026187939569354057, + 0.005781552288681269, + 0.004552335478365421, + 0.007186488248407841, + 0.006479804404079914, + 0.004603422712534666, + 0.004986354615539312, + 0.01361559797078371, + 0.008002249523997307, + 0.011825582943856716, + 0.011742809787392616, + 0.012838482856750488, + 0.014751115813851357, + 0.020256852731108665, + 0.013678355142474174, + 0.012516072019934654, + 0.011228407733142376, + 0.0076925796456635, + 0.005751210730522871, + 0.009488827548921108, + 0.006435595452785492, + 0.005681298673152924, + 0.008429694920778275, + 0.009139783680438995, + 0.009386198595166206, + 0.0031211036257445812, + 0.003127473406493664, + 0.003546238411217928, + 0.00316121825017035, + 0.002624351065605879, + 0.0034731344785541296, + 0.003483532229438424, + 0.003499401267617941, + 0.005064006894826889, + 0.002777897985652089, + 0.005291670095175505, + 0.005038341972976923, + 0.0034053632989525795, + 0.0035798405297100544, + 0.002921291859820485, + 0.005386939737945795, + 0.004845588468015194, + 0.006080018822103739, + 0.006502281408756971, + 0.005872644484043121, + 0.008638539351522923, + 0.004390195943415165, + 0.012220409698784351, + 0.007266523782163858, + 0.0015332080656662583, + 0.0023571751080453396, + 0.0033211933914572, + 0.0033162129111588, + 0.00278992741368711, + 0.0024080034345388412, + 0.002722100354731083, + 0.003492193529382348, + 0.002539740176871419, + 0.004623467568308115, + 0.004148343577980995, + 0.005285854917019606, + 0.004946951288729906, + 0.006958149839192629, + 0.0050787776708602905, + 0.005584217142313719, + 0.030938759446144104, + 0.035912707448005676, + 0.11288052797317505, + 0.027101615443825722, + 0.026903187856078148, + 0.14028306305408478, + 0.015266227535903454, + 0.03634100407361984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.03807859495282173 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a warm, Studio Ghibli-inspired style, a steady, eye-level, medium tracking shot follows a young woman as she carefully crosses a sun-dappled cobblestone street. Her dark brown hair is pulled back in a simple, loose braid, resting on the shoulders of a cream cotton shirt. With gentle focus, she pedals steadily and deliberately on a classic Dutch freight tricycle, its light blue frame showing subtle, charming signs of wear. The large wooden cargo box out front is filled with a vibrant, picturesque display of fresh produce and flowers; vibrant yellow sunflowers and purple lavender stalks jut out from between heads of crisp green lettuce, bright red tomatoes, and round onions. As the tricycle moves, its wheels clattering gently against the uneven, mossy cobblestones, the camera sliding parallel to it, capturing the shifting patterns of golden light filtering through the leaves of the trees overhead. The setting shows a quaint storybook town square with half-timbered buildings, terracotta roofs, and window boxes blooming with geraniums, while the other gentle townspeople—an elderly couple walking arm in arm, a baker dusting flour from his apron—add to the scene’s serene, painterly atmosphere of peaceful daily life.", + "seed": 7665, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_4/index_8058_seed_7665_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0320593249052763, + "vlm_logprob": -0.0320593249052763, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.0320593249052763, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0392896793782711 + ], + "opposite_logprob_tokens": [ + -3.289289712905884 + ], + "logprob_reward": -0.0392896793782711, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0392896793782711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14773187041282654, + 0.06501255929470062, + 0.10217893868684769, + 0.040566716343164444, + 0.07963227480649948, + 0.1699710488319397, + 0.07156763225793839, + 0.1261482834815979, + 0.09788080304861069, + 0.09930985420942307 + ], + "credit_map": [ + 0.014021996408700943, + 0.0123760299757123, + 0.029517458751797676, + 0.007841464132070541, + 0.012645508162677288, + 0.015801167115569115, + 0.010733864270150661, + 0.04479437321424484, + 0.008642138913273811, + 0.0075715165585279465, + 0.0055389790795743465, + 0.006177125032991171, + 0.008612844161689281, + 0.011856697499752045, + 0.009612110443413258, + 0.0070011499337852, + 0.010083582252264023, + 0.019556978717446327, + 0.009251313284039497, + 0.012703842483460903, + 0.010124430060386658, + 0.006580257322639227, + 0.02069920301437378, + 0.013179331086575985, + 0.0035208850167691708, + 0.004011296201497316, + 0.00257935025729239, + 0.004218773916363716, + 0.009084313176572323, + 0.005040566436946392, + 0.006284876726567745, + 0.005826654378324747, + 0.006532719824463129, + 0.00925295427441597, + 0.0057930792681872845, + 0.00816483236849308, + 0.008322881534695625, + 0.008945168927311897, + 0.02203086018562317, + 0.010589784942567348, + 0.00863033626228571, + 0.007302852813154459, + 0.007364154793322086, + 0.008491420187056065, + 0.040113236755132675, + 0.056017085909843445, + 0.016129115596413612, + 0.025922855362296104, + 0.008353094570338726, + 0.010108182206749916, + 0.007939986884593964, + 0.008168080821633339, + 0.008660280145704746, + 0.009606754407286644, + 0.00952262245118618, + 0.009208637289702892, + 0.01153173390775919, + 0.015076708048582077, + 0.01948656514286995, + 0.009029937908053398, + 0.017117582261562347, + 0.020867248997092247, + 0.01678214780986309, + 0.016256358474493027, + 0.014036159962415695, + 0.014212746173143387, + 0.011243006214499474, + 0.011832534335553646, + 0.009712696075439453, + 0.009764394722878933, + 0.016722634434700012, + 0.010356630198657513, + 0.01992369443178177, + 0.018652955070137978, + 0.016419168561697006, + 0.007409208919852972, + 0.009375114925205708, + 0.008425821550190449, + 0.007806819397956133, + 0.01129707507789135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.0392896793782711 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553953565657139 + ], + "opposite_logprob_tokens": [ + -4.387554168701172 + ], + "logprob_reward": -0.012553953565657139, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553953565657139, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1409354954957962, + 0.06905288249254227, + 0.10856457054615021, + 0.035857029259204865, + 0.10248368233442307, + 0.1482580453157425, + 0.03858325257897377, + 0.15450124442577362, + 0.09269291907548904, + 0.10907084494829178 + ], + "credit_map": [ + 0.014570679515600204, + 0.019022369757294655, + 0.016010962426662445, + 0.011837883852422237, + 0.02121027559041977, + 0.022510318085551262, + 0.009921411983668804, + 0.025851599872112274, + 0.008052852936089039, + 0.00717192143201828, + 0.004461200442165136, + 0.005942320451140404, + 0.010148073546588421, + 0.0137077197432518, + 0.010177469812333584, + 0.009391330182552338, + 0.013400117866694927, + 0.010948262177407742, + 0.010075355879962444, + 0.01409817859530449, + 0.009910895489156246, + 0.011633886024355888, + 0.023771982640028, + 0.01472589559853077, + 0.003837328404188156, + 0.0028150160796940327, + 0.0029926227871328592, + 0.003876960137858987, + 0.007887601852416992, + 0.004408702719956636, + 0.005221336614340544, + 0.004817463457584381, + 0.009159401059150696, + 0.01243056170642376, + 0.00860986765474081, + 0.00984905194491148, + 0.010848384350538254, + 0.013248666189610958, + 0.0250238124281168, + 0.013313944451510906, + 0.007599299773573875, + 0.008387316949665546, + 0.007038044277578592, + 0.007285293657332659, + 0.052560705691576004, + 0.03386137634515762, + 0.012591635808348656, + 0.018934383988380432, + 0.005141356028616428, + 0.004282338544726372, + 0.004125996958464384, + 0.005238224286586046, + 0.005058591719716787, + 0.004982739221304655, + 0.005345767363905907, + 0.004408242180943489, + 0.01708154007792473, + 0.021319227293133736, + 0.029782138764858246, + 0.011039286851882935, + 0.015530637465417385, + 0.028859684243798256, + 0.014054693281650543, + 0.016834048554301262, + 0.01220135297626257, + 0.013729189522564411, + 0.011636152863502502, + 0.009205466136336327, + 0.012414205819368362, + 0.01035806629806757, + 0.014085827395319939, + 0.009062670171260834, + 0.021232377737760544, + 0.018450036644935608, + 0.014387884177267551, + 0.009683500044047832, + 0.010364742018282413, + 0.010658815503120422, + 0.012796896509826183, + 0.011496593244373798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.012553953565657139 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638276718556881 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638276718556881, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638276718556881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10162468254566193, + 0.1147620677947998, + 0.08238942176103592, + 0.03764205053448677, + 0.08610672503709793, + 0.29640692472457886, + 0.03172104433178902, + 0.07426222413778305, + 0.0553680956363678, + 0.11971674114465714 + ], + "credit_map": [ + 0.008663628250360489, + 0.006009361706674099, + 0.02414103038609028, + 0.005767241585999727, + 0.004147086292505264, + 0.008998294360935688, + 0.006905101705342531, + 0.03699294105172157, + 0.016275644302368164, + 0.013933131471276283, + 0.008691614493727684, + 0.009697631001472473, + 0.014429074712097645, + 0.023701945319771767, + 0.012482492253184319, + 0.015550535172224045, + 0.009546768851578236, + 0.007886472158133984, + 0.007428474724292755, + 0.009950087405741215, + 0.009046880528330803, + 0.010631948709487915, + 0.016667863354086876, + 0.011230921372771263, + 0.003618444548919797, + 0.003985478077083826, + 0.0027485345490276814, + 0.0049703484401106834, + 0.00909577775746584, + 0.004443801008164883, + 0.0046239206567406654, + 0.004155744798481464, + 0.007182523608207703, + 0.016327688470482826, + 0.006463317200541496, + 0.00783486943691969, + 0.010096501559019089, + 0.009339850395917892, + 0.01875658705830574, + 0.010105390101671219, + 0.014515858143568039, + 0.011782879009842873, + 0.008977442979812622, + 0.011678284034132957, + 0.1112048402428627, + 0.10072662681341171, + 0.022433407604694366, + 0.015087590552866459, + 0.0032372099813073874, + 0.004075830802321434, + 0.003514774376526475, + 0.003431756515055895, + 0.004585887771099806, + 0.004081607796251774, + 0.0042922524735331535, + 0.004501726478338242, + 0.007587414234876633, + 0.008809130638837814, + 0.012454874813556671, + 0.005307901185005903, + 0.008696517907083035, + 0.012906878255307674, + 0.007461438421159983, + 0.011038069613277912, + 0.007531273178756237, + 0.008182495832443237, + 0.007075151894241571, + 0.006544870790094137, + 0.006852495484054089, + 0.005106702446937561, + 0.008743823505938053, + 0.005331282503902912, + 0.023682191967964172, + 0.020674698054790497, + 0.018280353397130966, + 0.01237205509096384, + 0.011503703892230988, + 0.01105577778071165, + 0.009007995016872883, + 0.013139968737959862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.008638276718556881 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003270834684372 + ], + "opposite_logprob_tokens": [ + -2.6950325965881348 + ], + "logprob_reward": -0.07003270834684372, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003270834684372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11730305105447769, + 0.0706566870212555, + 0.057764485478401184, + 0.02232269011437893, + 0.06367392838001251, + 0.339979350566864, + 0.039371225982904434, + 0.058042969554662704, + 0.129071906208992, + 0.10181374847888947 + ], + "credit_map": [ + 0.009500217624008656, + 0.007292318157851696, + 0.0279834121465683, + 0.0071395812556147575, + 0.004894868936389685, + 0.011921744793653488, + 0.007046971470117569, + 0.04152393341064453, + 0.00926561001688242, + 0.006625684443861246, + 0.00477988924831152, + 0.006379454862326384, + 0.009398148395121098, + 0.014785571955144405, + 0.009797273203730583, + 0.009625047445297241, + 0.006174798123538494, + 0.006556064821779728, + 0.005032285116612911, + 0.006307236384600401, + 0.006467156577855349, + 0.007531417533755302, + 0.012028977274894714, + 0.007666545920073986, + 0.0020746146328747272, + 0.0023478753864765167, + 0.0016996704507619143, + 0.0023475114721804857, + 0.005589469335973263, + 0.0029150680638849735, + 0.002691378351300955, + 0.002657100325450301, + 0.005651902873069048, + 0.012389285489916801, + 0.005151802208274603, + 0.0057543134316802025, + 0.006848766468465328, + 0.006896699778735638, + 0.013704153709113598, + 0.007276996970176697, + 0.01545071043074131, + 0.010852974839508533, + 0.008722876198589802, + 0.013042818754911423, + 0.13470856845378876, + 0.11342247575521469, + 0.02571665309369564, + 0.018062254413962364, + 0.007776367478072643, + 0.004719610791653395, + 0.00495485682040453, + 0.005974502302706242, + 0.003414895385503769, + 0.0053681256249547005, + 0.0036849980242550373, + 0.003477866528555751, + 0.004986514337360859, + 0.006399668753147125, + 0.010194352827966213, + 0.005576878786087036, + 0.005482643377035856, + 0.007694591768085957, + 0.00900190882384777, + 0.008706405758857727, + 0.017664087936282158, + 0.018495582044124603, + 0.015822822228074074, + 0.015204611234366894, + 0.0159713514149189, + 0.011528617702424526, + 0.02116858959197998, + 0.013216236606240273, + 0.020329201593995094, + 0.018796060234308243, + 0.014364785514771938, + 0.010193283669650555, + 0.010403812862932682, + 0.009361237287521362, + 0.0077624088153243065, + 0.010602951981127262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.07003270834684372 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978200651705265 + ], + "opposite_logprob_tokens": [ + -3.5297820568084717 + ], + "logprob_reward": -0.02978200651705265, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978200651705265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1719478815793991, + 0.08983761072158813, + 0.16618315875530243, + 0.0356149896979332, + 0.07860720902681351, + 0.06439214944839478, + 0.09005114436149597, + 0.11709003150463104, + 0.025140056386590004, + 0.16113576292991638 + ], + "credit_map": [ + 0.015277618542313576, + 0.012454193085432053, + 0.03666481003165245, + 0.012702946551144123, + 0.011509304866194725, + 0.011582642793655396, + 0.011193371377885342, + 0.06056298688054085, + 0.007846670225262642, + 0.008127174340188503, + 0.009457939304411411, + 0.007409489247947931, + 0.010080193169414997, + 0.023428039625287056, + 0.009360265918076038, + 0.014127840287983418, + 0.018700838088989258, + 0.02049332670867443, + 0.016277173534035683, + 0.0222280565649271, + 0.014356878586113453, + 0.014229986816644669, + 0.03785831853747368, + 0.02203858271241188, + 0.003006849903613329, + 0.0026831082068383694, + 0.0036136889830231667, + 0.004319552797824144, + 0.004524081014096737, + 0.004109647590667009, + 0.006905770860612392, + 0.006452291272580624, + 0.004495576955378056, + 0.009842569939792156, + 0.006535365246236324, + 0.006106001324951649, + 0.007655362132936716, + 0.010600650683045387, + 0.023801812902092934, + 0.009569872170686722, + 0.004649672191590071, + 0.005174320191144943, + 0.004521941300481558, + 0.0046316408552229404, + 0.016111323609948158, + 0.012309138663113117, + 0.010221855714917183, + 0.006772254128009081, + 0.01694166101515293, + 0.010295460931956768, + 0.009945300407707691, + 0.009539954364299774, + 0.010763676837086678, + 0.011147289536893368, + 0.010750197805464268, + 0.01066760253161192, + 0.009409637190401554, + 0.015750965103507042, + 0.019886083900928497, + 0.009217730723321438, + 0.017804140225052834, + 0.01917569898068905, + 0.014613746665418148, + 0.011232023127377033, + 0.0027481818106025457, + 0.004755516070872545, + 0.003547989297658205, + 0.0030508767813444138, + 0.0026988887693732977, + 0.0028717010281980038, + 0.0021132580004632473, + 0.003353644860908389, + 0.02817310206592083, + 0.025862960144877434, + 0.026301342993974686, + 0.02056925557553768, + 0.014494901522994041, + 0.012328521348536015, + 0.018642300739884377, + 0.014763385988771915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.02978200651705265 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a warm, Studio Ghibli-inspired style, a steady, eye-level, medium tracking shot follows a young woman as she carefully crosses a sun-dappled cobblestone street. Her dark brown hair is pulled back in a simple, loose braid, resting on the shoulders of a cream cotton shirt. With gentle focus, she pedals steadily and deliberately on a classic Dutch freight tricycle, its light blue frame showing subtle, charming signs of wear. The large wooden cargo box out front is filled with a vibrant, picturesque display of fresh produce and flowers; vibrant yellow sunflowers and purple lavender stalks jut out from between heads of crisp green lettuce, bright red tomatoes, and round onions. As the tricycle moves, its wheels clattering gently against the uneven, mossy cobblestones, the camera sliding parallel to it, capturing the shifting patterns of golden light filtering through the leaves of the trees overhead. The setting shows a quaint storybook town square with half-timbered buildings, terracotta roofs, and window boxes blooming with geraniums, while the other gentle townspeople—an elderly couple walking arm in arm, a baker dusting flour from his apron—add to the scene’s serene, painterly atmosphere of peaceful daily life.", + "seed": 7666, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_4/index_8058_seed_7666_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.022801010124385356, + "vlm_logprob": -0.022801010124385356, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.022801010124385356, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021916687488555908 + ], + "opposite_logprob_tokens": [ + -3.896916627883911 + ], + "logprob_reward": -0.021916687488555908, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021916687488555908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09121565520763397, + 0.050157416611909866, + 0.08132439851760864, + 0.10819719731807709, + 0.11447152495384216, + 0.15983937680721283, + 0.11506164819002151, + 0.11528783291578293, + 0.05758444592356682, + 0.10686048120260239 + ], + "credit_map": [ + 0.006342724896967411, + 0.011384162120521069, + 0.0120089715346694, + 0.013371649198234081, + 0.007925031706690788, + 0.007661664392799139, + 0.0206571314483881, + 0.011864320375025272, + 0.004091017879545689, + 0.003061875933781266, + 0.006191618274897337, + 0.00886497087776661, + 0.004548835568130016, + 0.0046429201029241085, + 0.012526368722319603, + 0.006229809019714594, + 0.010277918539941311, + 0.0062574599869549274, + 0.011798898689448833, + 0.009698975831270218, + 0.007687968201935291, + 0.008504766039550304, + 0.014150463044643402, + 0.012947949580848217, + 0.009199787862598896, + 0.008619535714387894, + 0.023514555767178535, + 0.01995876058936119, + 0.012560470029711723, + 0.01170412078499794, + 0.010141543112695217, + 0.01249842718243599, + 0.010071055963635445, + 0.010561315342783928, + 0.007640989497303963, + 0.012635758146643639, + 0.031815581023693085, + 0.015872979536652565, + 0.010846971534192562, + 0.015026870183646679, + 0.011702670715749264, + 0.018643908202648163, + 0.023123743012547493, + 0.027366438880562782, + 0.027503665536642075, + 0.01949172466993332, + 0.016041051596403122, + 0.015966180711984634, + 0.010036581195890903, + 0.02404273860156536, + 0.008383015170693398, + 0.008292690850794315, + 0.020585816353559494, + 0.023441266268491745, + 0.010120345279574394, + 0.010159195400774479, + 0.009529078379273415, + 0.01823292300105095, + 0.012340922839939594, + 0.025707662105560303, + 0.012382944114506245, + 0.010468619875609875, + 0.019595224410295486, + 0.007030460052192211, + 0.007583671249449253, + 0.00601236242800951, + 0.00840227771550417, + 0.006295622792094946, + 0.007809022441506386, + 0.007090807426720858, + 0.006340064574033022, + 0.008050620555877686, + 0.011505966074764729, + 0.01672954484820366, + 0.013508348725736141, + 0.023511189967393875, + 0.010095074772834778, + 0.011060049757361412, + 0.010749039240181446, + 0.009701276198029518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.021916687488555908 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079042106866837 + ], + "opposite_logprob_tokens": [ + -4.141078948974609 + ], + "logprob_reward": -0.016079042106866837, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079042106866837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07575830072164536, + 0.11253972351551056, + 0.04409004747867584, + 0.03840302675962448, + 0.03062928467988968, + 0.07778257131576538, + 0.05826503783464432, + 0.27673086524009705, + 0.03637038543820381, + 0.24943076074123383 + ], + "credit_map": [ + 0.005048730410635471, + 0.00957477930933237, + 0.010390035808086395, + 0.01151898130774498, + 0.007007238455116749, + 0.007359125651419163, + 0.012903477065265179, + 0.011955935508012772, + 0.01729419454932213, + 0.005094445310533047, + 0.03730490803718567, + 0.0131464758887887, + 0.006460154429078102, + 0.006092913448810577, + 0.017913516610860825, + 0.009233108721673489, + 0.0039693694561719894, + 0.003771215910091996, + 0.006087653338909149, + 0.007194660138338804, + 0.006303722038865089, + 0.004051967058330774, + 0.007484743371605873, + 0.005226718727499247, + 0.0022195216733962297, + 0.0026539305690675974, + 0.004674512892961502, + 0.005289026070386171, + 0.006017453968524933, + 0.008205595426261425, + 0.002939145779237151, + 0.006403839681297541, + 0.0023628000635653734, + 0.0023179491981863976, + 0.0022423877380788326, + 0.004215680528432131, + 0.005843289662152529, + 0.004900816828012466, + 0.004182247444987297, + 0.004564114846289158, + 0.006050080992281437, + 0.012258194386959076, + 0.011332998052239418, + 0.009058532305061817, + 0.0113962572067976, + 0.011002073995769024, + 0.009190254844725132, + 0.0074941786006093025, + 0.004811230581253767, + 0.007780987769365311, + 0.00494007021188736, + 0.005212432239204645, + 0.00973973236978054, + 0.01191179919987917, + 0.005488356109708548, + 0.008380433544516563, + 0.021295910701155663, + 0.032920952886343, + 0.03313767537474632, + 0.05210014805197716, + 0.03455232083797455, + 0.0316789448261261, + 0.05345863103866577, + 0.017586300149559975, + 0.0035433017183095217, + 0.004855124279856682, + 0.004120153840631247, + 0.00495758606120944, + 0.004929466638714075, + 0.005448616575449705, + 0.005022821016609669, + 0.003493314841762185, + 0.020362146198749542, + 0.034347131848335266, + 0.03315852954983711, + 0.04694721847772598, + 0.027138328179717064, + 0.0317452996969223, + 0.030511444434523582, + 0.02522064745426178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 1 + ], + "selected_reward": -0.016079042106866837 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091820679605007 + ], + "opposite_logprob_tokens": [ + -5.504091739654541 + ], + "logprob_reward": -0.004091820679605007, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091820679605007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1401955485343933, + 0.11041166633367538, + 0.12551558017730713, + 0.12844206392765045, + 0.06082472205162048, + 0.03764636442065239, + 0.1215166300535202, + 0.13290369510650635, + 0.08604303747415543, + 0.05650070682168007 + ], + "credit_map": [ + 0.010115863755345345, + 0.01584290713071823, + 0.017079142853617668, + 0.01484075840562582, + 0.018994690850377083, + 0.033381156623363495, + 0.013226093724370003, + 0.016714923083782196, + 0.024501346051692963, + 0.0059849899262189865, + 0.04040098935365677, + 0.004387336317449808, + 0.005030794534832239, + 0.0073293945752084255, + 0.013782592490315437, + 0.008994223549962044, + 0.01627608947455883, + 0.008573896251618862, + 0.015662526711821556, + 0.012140967883169651, + 0.023804910480976105, + 0.015277586877346039, + 0.014299934729933739, + 0.019479675218462944, + 0.011625198647379875, + 0.013509630225598812, + 0.013568795286118984, + 0.013486496172845364, + 0.02153480239212513, + 0.02439163438975811, + 0.014473249204456806, + 0.015852253884077072, + 0.0063152308575809, + 0.0046397107653319836, + 0.004474164918065071, + 0.006629567593336105, + 0.007445304188877344, + 0.01221492886543274, + 0.01045407447963953, + 0.008651737123727798, + 0.0037958533503115177, + 0.00674405787140131, + 0.005311476998031139, + 0.004020329564809799, + 0.004522451665252447, + 0.004812033846974373, + 0.00384013750590384, + 0.0046000247821211815, + 0.008737889118492603, + 0.042404234409332275, + 0.007301670964807272, + 0.010767394676804543, + 0.00693500554189086, + 0.01674790494143963, + 0.007325539365410805, + 0.021296989172697067, + 0.009923416189849377, + 0.014141278341412544, + 0.023292094469070435, + 0.02549402229487896, + 0.011799423955380917, + 0.010819103568792343, + 0.02685200609266758, + 0.010582346469163895, + 0.011808209121227264, + 0.0096387704834342, + 0.011979037895798683, + 0.012077117338776588, + 0.01285864319652319, + 0.009561545215547085, + 0.009092437103390694, + 0.009027278050780296, + 0.004488042555749416, + 0.006788332015275955, + 0.0073744612745940685, + 0.006634822580963373, + 0.010731136426329613, + 0.00896618701517582, + 0.005315171554684639, + 0.006202553864568472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.004091820679605007 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048648759722709656 + ], + "opposite_logprob_tokens": [ + -3.0486488342285156 + ], + "logprob_reward": -0.048648759722709656, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048648759722709656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12553785741329193, + 0.046056631952524185, + 0.13388246297836304, + 0.13313910365104675, + 0.12929539382457733, + 0.060420144349336624, + 0.15179884433746338, + 0.10763894766569138, + 0.06009349599480629, + 0.05213712528347969 + ], + "credit_map": [ + 0.009360846132040024, + 0.012816697359085083, + 0.015537788160145283, + 0.015104922465980053, + 0.015640389174222946, + 0.01762343756854534, + 0.015477956272661686, + 0.02397581934928894, + 0.0063367015682160854, + 0.0024384979624301195, + 0.011141409166157246, + 0.004766072146594524, + 0.0031391200609505177, + 0.003961172420531511, + 0.009770134463906288, + 0.004503525793552399, + 0.017499880865216255, + 0.008193298242986202, + 0.02256985567510128, + 0.013337393291294575, + 0.020018070936203003, + 0.014950553886592388, + 0.015644192695617676, + 0.021669231355190277, + 0.009116006083786488, + 0.011455054394900799, + 0.010993296280503273, + 0.016825737431645393, + 0.03369034081697464, + 0.022235102951526642, + 0.013927622698247433, + 0.014895952306687832, + 0.010657639242708683, + 0.012225097976624966, + 0.007950027473270893, + 0.012742106802761555, + 0.03547518700361252, + 0.02039773017168045, + 0.012637214735150337, + 0.01721038855612278, + 0.003991750068962574, + 0.005564293824136257, + 0.008470234461128712, + 0.009146845899522305, + 0.010724615305662155, + 0.008298041298985481, + 0.008071282878518105, + 0.006153083872050047, + 0.011393969878554344, + 0.03755873814225197, + 0.010677465237677097, + 0.01388478372246027, + 0.014740267768502235, + 0.020219868049025536, + 0.01470316480845213, + 0.028620582073926926, + 0.008574907667934895, + 0.010616864077746868, + 0.017093658447265625, + 0.022981813177466393, + 0.010402211919426918, + 0.007591582369059324, + 0.02210182510316372, + 0.008276086300611496, + 0.008120154961943626, + 0.00791605468839407, + 0.007054750807583332, + 0.009387307800352573, + 0.008156636729836464, + 0.006338334176689386, + 0.00816338136792183, + 0.004956880118697882, + 0.003932082559913397, + 0.005804126616567373, + 0.006988866254687309, + 0.005992858204990625, + 0.008515438064932823, + 0.008329695090651512, + 0.006560367997735739, + 0.0060136932879686356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.048648759722709656 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with dark brown hair in a loose braid wearing a cream cotton shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman riding a classic Dutch freight tricycle with a light blue frame and a large wooden cargo box filled with fresh produce and flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a sun-dappled cobblestone street in a quaint town square featuring half-timbered buildings with terracotta roofs and window boxes blooming with geraniums?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are other townspeople visible, specifically an elderly couple walking arm in arm and a baker dusting flour from his apron?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera executing a steady, eye-level, medium tracking shot that slides parallel to the tricycle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023268740624189377 + ], + "opposite_logprob_tokens": [ + -3.773268699645996 + ], + "logprob_reward": -0.023268740624189377, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023268740624189377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12741954624652863, + 0.09136318415403366, + 0.06449802219867706, + 0.1299622654914856, + 0.05887088552117348, + 0.061100322753190994, + 0.14262183010578156, + 0.039367906749248505, + 0.21928247809410095, + 0.06551358848810196 + ], + "credit_map": [ + 0.008065118454396725, + 0.016649004071950912, + 0.017645059153437614, + 0.02058667689561844, + 0.013345973566174507, + 0.009419879876077175, + 0.025677233934402466, + 0.016030600294470787, + 0.02096310444176197, + 0.004178795497864485, + 0.0348396860063076, + 0.004504913464188576, + 0.004076012875884771, + 0.0045501296408474445, + 0.011756932362914085, + 0.006493601948022842, + 0.0075295655988156796, + 0.006482809782028198, + 0.007886731997132301, + 0.005838380660861731, + 0.01148796547204256, + 0.009153783321380615, + 0.006708146538585424, + 0.009410633705556393, + 0.012448935769498348, + 0.01623734086751938, + 0.014441853389143944, + 0.017039312049746513, + 0.012566139921545982, + 0.021973833441734314, + 0.013379302807152271, + 0.021875541657209396, + 0.0033239261247217655, + 0.004980637226253748, + 0.0034957348834723234, + 0.005188132636249065, + 0.017539460211992264, + 0.00792992115020752, + 0.006608744151890278, + 0.009804324246942997, + 0.005539603531360626, + 0.013606230728328228, + 0.0062082684598863125, + 0.008710576221346855, + 0.007699695881456137, + 0.008494673296809196, + 0.005326067563146353, + 0.005515204276889563, + 0.01012304238975048, + 0.05000787600874901, + 0.007195853162556887, + 0.009053381159901619, + 0.014148110523819923, + 0.029292341321706772, + 0.008541848510503769, + 0.014259365387260914, + 0.003680728841573, + 0.0055310106836259365, + 0.004658898804336786, + 0.005896556656807661, + 0.005401217844337225, + 0.004187287762761116, + 0.00606863247230649, + 0.00394357368350029, + 0.02621612139046192, + 0.03150549158453941, + 0.025356700643897057, + 0.04137339070439339, + 0.03920194134116173, + 0.019530799239873886, + 0.018552755936980247, + 0.017545271664857864, + 0.005777045153081417, + 0.008535570465028286, + 0.007993052713572979, + 0.009349400177598, + 0.01107022538781166, + 0.006195503752678633, + 0.007819939404726028, + 0.00877284910529852 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.023268740624189377 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_9/index_tensor([1139, 1139])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_9/index_tensor([1139, 1139])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2e892e471d1be22a89fae2cb159cbe2bb45e083c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_9/index_tensor([1139, 1139])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long shot, shot from a low, upward angle, the camera tilts steadily upward, tracking the paths of countless light trails ascending across the deep, inky, velvety midnight sky. Hundreds of dazzling golden arcs and brilliant silver beams, trailing long, shimmering trails of particles, shoot upward from the invisible bottom of the frame at astonishing speed, carving graceful, powerful parabolas through the sky, tearing through the night. As the camera tilts to its highest point, all the paths converge at a precise, breathtaking focal point at the very center of the frame, at the apex of their ascent. After a brief pause, they brilliantly merge and instantly explode into a vast, symmetrical, spherical corolla of pure light, illuminating the inky black sky with its intense radiance. The intertwining, yet distinct, strands of gold and silver create a visual climax of grandeur, celebration, and wonder.", + "seed": 28711, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_9/index_1139_seed_28711_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06043259087018669, + "vlm_logprob": -0.06043259087018669, + "vlm_yes_no_margin": 3.549999976158142, + "vlm_token_credit": -0.06043259087018669, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1422475129365921 + ], + "opposite_logprob_tokens": [ + -2.142247438430786 + ], + "logprob_reward": -0.1422475129365921, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1422475129365921, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05040964484214783, + 0.03479943051934242, + 0.1512303501367569, + 0.25636571645736694, + 0.054810430854558945, + 0.034554727375507355, + 0.1999303698539734, + 0.0803675651550293, + 0.09053661674261093, + 0.046995073556900024 + ], + "credit_map": [ + 0.0038269630167633295, + 0.004437460098415613, + 0.004473249893635511, + 0.010103068314492702, + 0.00843862071633339, + 0.005917360074818134, + 0.006036314181983471, + 0.007176612038165331, + 0.004227607045322657, + 0.003252989612519741, + 0.0022145414259284735, + 0.004085988737642765, + 0.008817985653877258, + 0.004340826999396086, + 0.003881637705489993, + 0.003977857995778322, + 0.02574850805103779, + 0.024702733382582664, + 0.019942495971918106, + 0.01723586954176426, + 0.020164135843515396, + 0.015553676523268223, + 0.010057426057755947, + 0.01782553642988205, + 0.02787172794342041, + 0.03389805555343628, + 0.038037050515413284, + 0.030037639662623405, + 0.05349747836589813, + 0.030463041737675667, + 0.023098068311810493, + 0.019462689757347107, + 0.008264217525720596, + 0.00963464193046093, + 0.007906017825007439, + 0.005828602239489555, + 0.00781253445893526, + 0.007849350571632385, + 0.0022866937797516584, + 0.0052283755503594875, + 0.0031287570018321276, + 0.005598469637334347, + 0.004200911149382591, + 0.004982899408787489, + 0.004706239327788353, + 0.004381630569696426, + 0.0033633941784501076, + 0.0041924286633729935, + 0.05037451535463333, + 0.01777632348239422, + 0.03953966125845909, + 0.012188573367893696, + 0.016343308612704277, + 0.0117707634344697, + 0.03578760474920273, + 0.016149641945958138, + 0.012775563634932041, + 0.008078191429376602, + 0.009790447540581226, + 0.01352536492049694, + 0.008838921785354614, + 0.0057765585370361805, + 0.012300320900976658, + 0.009282197803258896, + 0.009984886273741722, + 0.009029211476445198, + 0.01414938922971487, + 0.02073593996465206, + 0.005927977152168751, + 0.008757773786783218, + 0.010510408319532871, + 0.011441032402217388, + 0.006387459114193916, + 0.004014777485281229, + 0.008486210368573666, + 0.007884707301855087, + 0.004339324776083231, + 0.004037925507873297, + 0.005406289827078581, + 0.006438381504267454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.1422475129365921 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903008908033371 + ], + "opposite_logprob_tokens": [ + -2.5790300369262695 + ], + "logprob_reward": -0.07903008908033371, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903008908033371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026840640231966972, + 0.045991118997335434, + 0.15336525440216064, + 0.4036495089530945, + 0.039531320333480835, + 0.061955150216817856, + 0.16823247075080872, + 0.04542989656329155, + 0.03965523838996887, + 0.015349416062235832 + ], + "credit_map": [ + 0.003464687382802367, + 0.0027799198869615793, + 0.00392896356061101, + 0.004441374447196722, + 0.0027551488019526005, + 0.00280291261151433, + 0.003885773941874504, + 0.002781856106594205, + 0.004491032566875219, + 0.003519190475344658, + 0.0044835456646978855, + 0.006077172234654427, + 0.009393956512212753, + 0.0040744892321527, + 0.0063020214438438416, + 0.007649707142263651, + 0.048379555344581604, + 0.0496516227722168, + 0.008138677105307579, + 0.008752981200814247, + 0.015636717900633812, + 0.009899609722197056, + 0.005867937579751015, + 0.00703813461586833, + 0.045511916279792786, + 0.05302312225103378, + 0.06561785191297531, + 0.04281226173043251, + 0.08244733512401581, + 0.04581240937113762, + 0.0370158925652504, + 0.031408678740262985, + 0.007215291727334261, + 0.006561821326613426, + 0.006446746177971363, + 0.004385071340948343, + 0.003965600859373808, + 0.004456981550902128, + 0.0022065648809075356, + 0.004293239209800959, + 0.005563416983932257, + 0.014832091517746449, + 0.008411640301346779, + 0.006360657978802919, + 0.007164555601775646, + 0.007949857972562313, + 0.004075703211128712, + 0.007597225718200207, + 0.045096464455127716, + 0.012133697047829628, + 0.033192649483680725, + 0.008806939236819744, + 0.014438701793551445, + 0.00948680005967617, + 0.0325709767639637, + 0.012506227008998394, + 0.006926438771188259, + 0.005342313088476658, + 0.005548358429223299, + 0.01039329543709755, + 0.004455736838281155, + 0.003096827771514654, + 0.005811157636344433, + 0.0038557646330446005, + 0.003897811286151409, + 0.00432533910498023, + 0.005569001194089651, + 0.0077017382718622684, + 0.00299130380153656, + 0.005013816524296999, + 0.0042503997683525085, + 0.005905822850763798, + 0.0017257181461900473, + 0.0014816677430644631, + 0.0023341316264122725, + 0.0025715120136737823, + 0.001365453121252358, + 0.0015373174101114273, + 0.001851160661317408, + 0.0024824540596455336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.07903008908033371 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032069466542452574 + ], + "opposite_logprob_tokens": [ + -5.753206729888916 + ], + "logprob_reward": -0.0032069466542452574, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032069466542452574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07525084912776947, + 0.06820652633905411, + 0.09782290458679199, + 0.2343318611383438, + 0.04045490920543671, + 0.026254529133439064, + 0.3590600788593292, + 0.04980748891830444, + 0.018181955441832542, + 0.030628854408860207 + ], + "credit_map": [ + 0.007817475125193596, + 0.006914524827152491, + 0.007685075514018536, + 0.012424420565366745, + 0.011216524057090282, + 0.011527285911142826, + 0.00940090324729681, + 0.00826464593410492, + 0.008958697319030762, + 0.006968792527914047, + 0.004937987774610519, + 0.006994885392487049, + 0.022008314728736877, + 0.0061766416765749454, + 0.006111533381044865, + 0.006049678660929203, + 0.021677061915397644, + 0.02324446104466915, + 0.008163965307176113, + 0.012325081042945385, + 0.011249775998294353, + 0.005718911066651344, + 0.006862623151391745, + 0.00858103483915329, + 0.026637131348252296, + 0.030708106234669685, + 0.037810031324625015, + 0.025122882798314095, + 0.047437094151973724, + 0.02639690414071083, + 0.021300561726093292, + 0.018919168040156364, + 0.0059937238693237305, + 0.004985211882740259, + 0.005931857042014599, + 0.004174085799604654, + 0.004137871321290731, + 0.005087993107736111, + 0.0034224356058984995, + 0.006721733137965202, + 0.0026365392841398716, + 0.0047056470066308975, + 0.0033820054959505796, + 0.0030316829215735197, + 0.00427235709503293, + 0.003179301740601659, + 0.0020906468853354454, + 0.002956351265311241, + 0.09644613415002823, + 0.028013046830892563, + 0.0706571638584137, + 0.019993260502815247, + 0.030998123809695244, + 0.01836589351296425, + 0.06969805806875229, + 0.02488839067518711, + 0.007029309403151274, + 0.00474169198423624, + 0.007494593039155006, + 0.008041072636842728, + 0.0057365624234080315, + 0.005431834142655134, + 0.006138997618108988, + 0.005193428136408329, + 0.002115461975336075, + 0.002334811259061098, + 0.002365032210946083, + 0.0024408819153904915, + 0.0017530240584164858, + 0.0023954303469508886, + 0.002262591617181897, + 0.0025147232227027416, + 0.0032048961147665977, + 0.003606943879276514, + 0.0038170262705534697, + 0.004360162187367678, + 0.0029226448386907578, + 0.002981798257678747, + 0.004056091886013746, + 0.005679295398294926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.0032069466542452574 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004448771476746 + ], + "opposite_logprob_tokens": [ + -2.69504451751709 + ], + "logprob_reward": -0.07004448771476746, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004448771476746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07000032067298889, + 0.05844515934586525, + 0.24332498013973236, + 0.11047890037298203, + 0.04791878163814545, + 0.044069696217775345, + 0.29508984088897705, + 0.07576415687799454, + 0.028117945417761803, + 0.026790214702486992 + ], + "credit_map": [ + 0.005235905759036541, + 0.005791265983134508, + 0.005005925428122282, + 0.020631564781069756, + 0.0061417557299137115, + 0.005480250809341669, + 0.005891142878681421, + 0.015822499990463257, + 0.0067641837522387505, + 0.006429064553231001, + 0.006185148376971483, + 0.009815116412937641, + 0.011076748371124268, + 0.007420332171022892, + 0.005083122756332159, + 0.005671439692378044, + 0.06223604455590248, + 0.060475531965494156, + 0.020564228296279907, + 0.0179038867354393, + 0.02977622114121914, + 0.018713289871811867, + 0.01287537720054388, + 0.02078036218881607, + 0.012537231668829918, + 0.015607515349984169, + 0.01845322921872139, + 0.012176308780908585, + 0.018474210053682327, + 0.012202925980091095, + 0.010915272869169712, + 0.010112195275723934, + 0.007525703404098749, + 0.005782566033303738, + 0.009378288872539997, + 0.005018746946007013, + 0.005441764835268259, + 0.006881696172058582, + 0.00270438683219254, + 0.0051856255158782005, + 0.003573846537619829, + 0.007827718742191792, + 0.005559912417083979, + 0.0057626646012067795, + 0.00627981498837471, + 0.004890076816082001, + 0.003182048676535487, + 0.0069936104118824005, + 0.07674537599086761, + 0.022874444723129272, + 0.05408578738570213, + 0.014775614254176617, + 0.029085569083690643, + 0.017486216500401497, + 0.05874045938253403, + 0.021296340972185135, + 0.010442995466291904, + 0.007617809809744358, + 0.00730206910520792, + 0.008160263299942017, + 0.011271476745605469, + 0.008209332823753357, + 0.0112836305052042, + 0.011476571671664715, + 0.003098691813647747, + 0.0031796591356396675, + 0.003634171560406685, + 0.0048816311173141, + 0.002562482375651598, + 0.003145019058138132, + 0.0033586907666176558, + 0.004257596097886562, + 0.002926657674834132, + 0.00274377572350204, + 0.0031871565151959658, + 0.0027634648140519857, + 0.004320953041315079, + 0.003096870146691799, + 0.0032184894662350416, + 0.004532846622169018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.07004448771476746 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633917964994907 + ], + "opposite_logprob_tokens": [ + -4.882633686065674 + ], + "logprob_reward": -0.007633917964994907, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633917964994907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06409534066915512, + 0.03326532244682312, + 0.15594826638698578, + 0.062192343175411224, + 0.04604256525635719, + 0.043252475559711456, + 0.2944283187389374, + 0.14397622644901276, + 0.07566295564174652, + 0.08113611489534378 + ], + "credit_map": [ + 0.007431735750287771, + 0.0060721575282514095, + 0.007772131823003292, + 0.010602480731904507, + 0.0073624406941235065, + 0.006226920057088137, + 0.011468627490103245, + 0.007158853579312563, + 0.004266489762812853, + 0.0036069736815989017, + 0.0023747836239635944, + 0.0034573820885270834, + 0.00788716971874237, + 0.004093332681804895, + 0.003474211785942316, + 0.004104985389858484, + 0.023875948041677475, + 0.02461077831685543, + 0.019263746216893196, + 0.020503222942352295, + 0.023232566192746162, + 0.014149942435324192, + 0.012220881879329681, + 0.01809120364487171, + 0.007222762331366539, + 0.00799151137471199, + 0.012783320620656013, + 0.006466368678957224, + 0.009973030537366867, + 0.006374712567776442, + 0.004979817662388086, + 0.006400830578058958, + 0.009011855348944664, + 0.004931474104523659, + 0.012194370850920677, + 0.005710451398044825, + 0.003834802657365799, + 0.003991289995610714, + 0.0030088082421571016, + 0.003359514754265547, + 0.004071767441928387, + 0.005827507935464382, + 0.005509264767169952, + 0.004597796592861414, + 0.008092084899544716, + 0.00640077656134963, + 0.0035751352552324533, + 0.005178151186555624, + 0.07686817646026611, + 0.023729125037789345, + 0.05362304300069809, + 0.015269394963979721, + 0.028918014839291573, + 0.017751513049006462, + 0.05746719241142273, + 0.02080189436674118, + 0.017581885680556297, + 0.01924262009561062, + 0.01347668282687664, + 0.01294198539108038, + 0.025169778615236282, + 0.014075662940740585, + 0.01634996384382248, + 0.025137677788734436, + 0.007223049178719521, + 0.008025791496038437, + 0.010199728421866894, + 0.012406762689352036, + 0.009425932541489601, + 0.008871719241142273, + 0.00804829504340887, + 0.01146168913692236, + 0.008921858854591846, + 0.007313716225326061, + 0.010432815179228783, + 0.009306324645876884, + 0.011114723980426788, + 0.008781636133790016, + 0.010315671563148499, + 0.014949378557503223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.007633917964994907 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a low, upward angle, the camera tilts steadily upward, tracking the paths of countless light trails ascending across the deep, inky, velvety midnight sky. Hundreds of dazzling golden arcs and brilliant silver beams, trailing long, shimmering trails of particles, shoot upward from the invisible bottom of the frame at astonishing speed, carving graceful, powerful parabolas through the sky, tearing through the night. As the camera tilts to its highest point, all the paths converge at a precise, breathtaking focal point at the very center of the frame, at the apex of their ascent. After a brief pause, they brilliantly merge and instantly explode into a vast, symmetrical, spherical corolla of pure light, illuminating the inky black sky with its intense radiance. The intertwining, yet distinct, strands of gold and silver create a visual climax of grandeur, celebration, and wonder.", + "seed": 28712, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_9/index_1139_seed_28712_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.046919985208660364, + "vlm_logprob": -0.046919985208660364, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.046919985208660364, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06890660524368286 + ], + "opposite_logprob_tokens": [ + -2.943906545639038 + ], + "logprob_reward": -0.06890660524368286, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06890660524368286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06621327251195908, + 0.1438213437795639, + 0.056424498558044434, + 0.0742429569363594, + 0.07548864185810089, + 0.12207057327032089, + 0.16749995946884155, + 0.07732704281806946, + 0.08582659065723419, + 0.1310850828886032 + ], + "credit_map": [ + 0.00545953493565321, + 0.006790625862777233, + 0.009009257890284061, + 0.010984091088175774, + 0.006226761732250452, + 0.00877058319747448, + 0.00632887240499258, + 0.012643542140722275, + 0.010233955457806587, + 0.014149371534585953, + 0.010446831583976746, + 0.01709498092532158, + 0.010973288677632809, + 0.03179031237959862, + 0.028971362859010696, + 0.020161228254437447, + 0.008264757692813873, + 0.008597799576818943, + 0.009817074984312057, + 0.0057792807929217815, + 0.0044700331054627895, + 0.008886181749403477, + 0.004453381523489952, + 0.006155987735837698, + 0.008982568979263306, + 0.009338613599538803, + 0.013190411031246185, + 0.00874643586575985, + 0.006532256957143545, + 0.011185478419065475, + 0.005609085317701101, + 0.01065810862928629, + 0.0071412730030715466, + 0.011436874978244305, + 0.006771094165742397, + 0.01088731363415718, + 0.009519043378531933, + 0.009730849415063858, + 0.008226134814321995, + 0.011776057071983814, + 0.02394731529057026, + 0.011841783300042152, + 0.014924796298146248, + 0.014930163510143757, + 0.010503890924155712, + 0.019397782161831856, + 0.01158338226377964, + 0.014941452071070671, + 0.034995175898075104, + 0.014250931330025196, + 0.021665213629603386, + 0.016788683831691742, + 0.01703106239438057, + 0.03001643717288971, + 0.012387865222990513, + 0.02036459371447563, + 0.015047167427837849, + 0.00690974947065115, + 0.009109383448958397, + 0.006064617540687323, + 0.010397794656455517, + 0.011181536130607128, + 0.007481818553060293, + 0.011134974658489227, + 0.011753205209970474, + 0.010749327950179577, + 0.01093045249581337, + 0.012677493505179882, + 0.010317342355847359, + 0.00867723859846592, + 0.011357834562659264, + 0.009363693185150623, + 0.014132094569504261, + 0.008992789313197136, + 0.010889850556850433, + 0.036613836884498596, + 0.008761521428823471, + 0.00961042009294033, + 0.03186599165201187, + 0.01021857000887394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 9 + ], + "selected_reward": -0.06890660524368286 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813321888446808 + ], + "opposite_logprob_tokens": [ + -3.288133144378662 + ], + "logprob_reward": -0.03813321888446808, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813321888446808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.026600083336234093, + 0.020428303629159927, + 0.02498951181769371, + 0.019985487684607506, + 0.029583338648080826, + 0.023333366960287094, + 0.0868275836110115, + 0.04410899057984352, + 0.1423521488904953, + 0.5817911624908447 + ], + "credit_map": [ + 0.002222488634288311, + 0.0027972355019301176, + 0.003014253918081522, + 0.0024604678619652987, + 0.002277010353282094, + 0.0059565952979028225, + 0.003057426307350397, + 0.004814608488231897, + 0.0014240519376471639, + 0.0018504811450839043, + 0.002996050287038088, + 0.0023331213742494583, + 0.0021213192958384752, + 0.002885100431740284, + 0.002530523110181093, + 0.004287657793611288, + 0.002729282481595874, + 0.0036288758274167776, + 0.003578649368137121, + 0.0031417387071996927, + 0.002774527296423912, + 0.0034237485378980637, + 0.0024776379577815533, + 0.0032350546680390835, + 0.0029114841017872095, + 0.003648764453828335, + 0.002484062919393182, + 0.0023231850937008858, + 0.002584166591987014, + 0.002942201215773821, + 0.0012409492628648877, + 0.0018506764899939299, + 0.0037026898935437202, + 0.0035362825728952885, + 0.0033496543765068054, + 0.004196965601295233, + 0.0029289659578353167, + 0.0034751826897263527, + 0.004286667797714472, + 0.004106935579329729, + 0.0036138815339654684, + 0.002432420151308179, + 0.0025520287454128265, + 0.0028594594914466143, + 0.0023312363773584366, + 0.003882378339767456, + 0.0027866587042808533, + 0.0028753066435456276, + 0.010283390991389751, + 0.008122327737510204, + 0.0084689324721694, + 0.009463219903409481, + 0.012362738139927387, + 0.011262496933341026, + 0.012850320897996426, + 0.014014159329235554, + 0.010411759838461876, + 0.003769237780943513, + 0.003999697044491768, + 0.004585762042552233, + 0.004223409574478865, + 0.006392960902303457, + 0.0047075082547962666, + 0.00601866003125906, + 0.033192310482263565, + 0.007469731383025646, + 0.006405151914805174, + 0.03325324133038521, + 0.005608225706964731, + 0.0267077274620533, + 0.02441198006272316, + 0.0053037856705486774, + 0.07871168106794357, + 0.02512381412088871, + 0.01588013768196106, + 0.1959768533706665, + 0.016741331666707993, + 0.06085377186536789, + 0.16219092905521393, + 0.02631264552474022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.03813321888446808 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005960783455520868 + ], + "opposite_logprob_tokens": [ + -5.130960941314697 + ], + "logprob_reward": -0.005960783455520868, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005960783455520868, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0852944552898407, + 0.09787817299365997, + 0.02254526875913143, + 0.03653400391340256, + 0.04448836296796799, + 0.03969639539718628, + 0.0752367228269577, + 0.11722986400127411, + 0.40147000551223755, + 0.07962672412395477 + ], + "credit_map": [ + 0.00892315898090601, + 0.006673848256468773, + 0.00977862998843193, + 0.009705821052193642, + 0.007712238468229771, + 0.017407219856977463, + 0.0124165378510952, + 0.012677006423473358, + 0.007228398229926825, + 0.00977142434567213, + 0.008733137510716915, + 0.01010468602180481, + 0.0064560105092823505, + 0.021253131330013275, + 0.020328406244516373, + 0.014002993702888489, + 0.0027459750417619944, + 0.003058140631765127, + 0.0031850754749029875, + 0.0024438735563308, + 0.0029375210870057344, + 0.0030186129733920097, + 0.0021820124238729477, + 0.0029740571044385433, + 0.004445756785571575, + 0.008406165055930614, + 0.003633461892604828, + 0.004625850822776556, + 0.003328347811475396, + 0.00574911804869771, + 0.0027978732250630856, + 0.003547431668266654, + 0.005195970647037029, + 0.005418726243078709, + 0.004786094184964895, + 0.004558064043521881, + 0.006172099616378546, + 0.005884505808353424, + 0.006806490011513233, + 0.005666412878781557, + 0.0064170160330832005, + 0.004972991533577442, + 0.0032187895849347115, + 0.005268194247037172, + 0.0044846488162875175, + 0.0061344425193965435, + 0.004030948504805565, + 0.005169368814677, + 0.009750116616487503, + 0.007960250601172447, + 0.009170080535113811, + 0.00787518173456192, + 0.009882218204438686, + 0.009689074009656906, + 0.010781696066260338, + 0.010128114372491837, + 0.01830437406897545, + 0.011929570697247982, + 0.014165740460157394, + 0.008485834114253521, + 0.019394787028431892, + 0.020872335880994797, + 0.012584289535880089, + 0.011492935940623283, + 0.09643363952636719, + 0.017986653372645378, + 0.017553478479385376, + 0.09710240364074707, + 0.013229887932538986, + 0.07391399145126343, + 0.07079721242189407, + 0.014452751725912094, + 0.009650925174355507, + 0.005439881235361099, + 0.007345644757151604, + 0.021335920318961143, + 0.005189474672079086, + 0.00785281602293253, + 0.017838185653090477, + 0.004973888397216797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.005960783455520868 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295619606971741 + ], + "opposite_logprob_tokens": [ + -2.2379562854766846 + ], + "logprob_reward": -0.11295619606971741, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295619606971741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.081361785531044, + 0.2840788662433624, + 0.026786351576447487, + 0.023133231326937675, + 0.02467968501150608, + 0.036159515380859375, + 0.07003089040517807, + 0.059440404176712036, + 0.19334667921066284, + 0.20098255574703217 + ], + "credit_map": [ + 0.006963175255805254, + 0.00990832969546318, + 0.007733702659606934, + 0.00965464860200882, + 0.008135599084198475, + 0.011455454863607883, + 0.009948170743882656, + 0.017562713474035263, + 0.021319163963198662, + 0.02516302466392517, + 0.036442771553993225, + 0.027650875970721245, + 0.025057120248675346, + 0.05225071310997009, + 0.040459539741277695, + 0.055735670030117035, + 0.0027348557487130165, + 0.004328495357185602, + 0.004490171559154987, + 0.0028010269161313772, + 0.0031605137046426535, + 0.0036427380982786417, + 0.002577491570264101, + 0.0030510604847222567, + 0.002450281521305442, + 0.0035656208638101816, + 0.004209093749523163, + 0.002828675089403987, + 0.002248934004455805, + 0.0031045174691826105, + 0.0018515025731176138, + 0.0028746062889695168, + 0.0029266546480357647, + 0.003025433048605919, + 0.0030198937747627497, + 0.0032390279229730368, + 0.0024273917078971863, + 0.003312975401058793, + 0.002955904696136713, + 0.00377240520901978, + 0.006182951387017965, + 0.0038542861584573984, + 0.003442364512011409, + 0.004450412932783365, + 0.0031951458659023046, + 0.005700305569916964, + 0.004609187599271536, + 0.004724862053990364, + 0.009691166691482067, + 0.007088338490575552, + 0.007401592563837767, + 0.007417639717459679, + 0.008672419004142284, + 0.009336372837424278, + 0.009141217917203903, + 0.011282150633633137, + 0.010330839082598686, + 0.004585014656186104, + 0.006717987824231386, + 0.006758037023246288, + 0.006414657924324274, + 0.007298654410988092, + 0.0069786496460437775, + 0.010356565937399864, + 0.03818211704492569, + 0.010624945163726807, + 0.015772730112075806, + 0.04205809533596039, + 0.012865118682384491, + 0.02768702805042267, + 0.031317245215177536, + 0.014839417301118374, + 0.03398505970835686, + 0.020375775173306465, + 0.01608368754386902, + 0.024981359019875526, + 0.02462451346218586, + 0.03414418548345566, + 0.029546210542321205, + 0.017241790890693665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.11295619606971741 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643122389912605 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643122389912605, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643122389912605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05024374648928642, + 0.0773426964879036, + 0.027407122775912285, + 0.03556203097105026, + 0.027512051165103912, + 0.02153538540005684, + 0.03658260405063629, + 0.05146671459078789, + 0.3028305470943451, + 0.36951708793640137 + ], + "credit_map": [ + 0.004581707529723644, + 0.00479848263785243, + 0.006055419333279133, + 0.005791170056909323, + 0.0045746369287371635, + 0.00825788639485836, + 0.006924736779183149, + 0.009259711019694805, + 0.006684028077870607, + 0.006594250909984112, + 0.01102711632847786, + 0.007266377564519644, + 0.010111147537827492, + 0.0099486717954278, + 0.007754722144454718, + 0.017956385388970375, + 0.0030506961047649384, + 0.006030659656971693, + 0.003886997699737549, + 0.0031077098101377487, + 0.002157449023798108, + 0.004353371914476156, + 0.0019285213202238083, + 0.002891718177124858, + 0.003475958714261651, + 0.006834046915173531, + 0.005156383384019136, + 0.0034452166873961687, + 0.003922448493540287, + 0.004835067316889763, + 0.0025729425251483917, + 0.005319969728589058, + 0.002943364903330803, + 0.0031208365689963102, + 0.0037180923391133547, + 0.003389510326087475, + 0.002513082465156913, + 0.003618524642661214, + 0.0036423858255147934, + 0.004566254559904337, + 0.0031087142415344715, + 0.0020543953869491816, + 0.0031675344798713923, + 0.002488060388714075, + 0.0018341148970648646, + 0.003696757834404707, + 0.0024655486922711134, + 0.0027202614583075047, + 0.005675905384123325, + 0.0037034284323453903, + 0.003942208364605904, + 0.0031608466524630785, + 0.004895842168480158, + 0.0049577681347727776, + 0.0037975807208567858, + 0.006449026055634022, + 0.0078917620703578, + 0.0036352514289319515, + 0.005007991101592779, + 0.00513275945559144, + 0.009609498083591461, + 0.006294598802924156, + 0.0060326275415718555, + 0.007862234488129616, + 0.08639759570360184, + 0.01935098133981228, + 0.011703675612807274, + 0.04857737571001053, + 0.01517445407807827, + 0.07175013422966003, + 0.03825053572654724, + 0.011625816114246845, + 0.03210438787937164, + 0.021133221685886383, + 0.013298128731548786, + 0.13228200376033783, + 0.021011969074606895, + 0.01915479637682438, + 0.11335109919309616, + 0.017181510105729103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.008643122389912605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a low, upward angle, the camera tilts steadily upward, tracking the paths of countless light trails ascending across the deep, inky, velvety midnight sky. Hundreds of dazzling golden arcs and brilliant silver beams, trailing long, shimmering trails of particles, shoot upward from the invisible bottom of the frame at astonishing speed, carving graceful, powerful parabolas through the sky, tearing through the night. As the camera tilts to its highest point, all the paths converge at a precise, breathtaking focal point at the very center of the frame, at the apex of their ascent. After a brief pause, they brilliantly merge and instantly explode into a vast, symmetrical, spherical corolla of pure light, illuminating the inky black sky with its intense radiance. The intertwining, yet distinct, strands of gold and silver create a visual climax of grandeur, celebration, and wonder.", + "seed": 28713, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_9/index_1139_seed_28713_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0159269196446985, + "vlm_logprob": -0.0159269196446985, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.0159269196446985, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023671170696616173 + ], + "opposite_logprob_tokens": [ + -4.1486711502075195 + ], + "logprob_reward": -0.023671170696616173, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.023671170696616173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2131221741437912, + 0.060409776866436005, + 0.1523510068655014, + 0.03738556057214737, + 0.05156751722097397, + 0.05890578031539917, + 0.034177251160144806, + 0.056726325303316116, + 0.18402893841266632, + 0.15132573246955872 + ], + "credit_map": [ + 0.010725506581366062, + 0.006780081428587437, + 0.022833414375782013, + 0.015160239301621914, + 0.009261279366910458, + 0.06979156285524368, + 0.010140315629541874, + 0.06842977553606033, + 0.007878179661929607, + 0.003382341703400016, + 0.01437139417976141, + 0.0080641508102417, + 0.0035520654637366533, + 0.006895731668919325, + 0.006102581508457661, + 0.010163325816392899, + 0.017183426767587662, + 0.012441243976354599, + 0.02057100646197796, + 0.0179460309445858, + 0.012186700478196144, + 0.025676794350147247, + 0.017624901607632637, + 0.028720876201987267, + 0.003629954531788826, + 0.0034020415041595697, + 0.006334870588034391, + 0.003692949889227748, + 0.0033641804475337267, + 0.005355457309633493, + 0.0030565408524125814, + 0.008549564518034458, + 0.005555257201194763, + 0.0060999165289103985, + 0.008115447126328945, + 0.008171998895704746, + 0.003156311810016632, + 0.007242900785058737, + 0.005063952412456274, + 0.008161729201674461, + 0.006545482203364372, + 0.006168378051370382, + 0.009087923914194107, + 0.008114413358271122, + 0.004867544397711754, + 0.009351183660328388, + 0.006031275261193514, + 0.008739575743675232, + 0.004228996578603983, + 0.003900365438312292, + 0.004880664404481649, + 0.005499225575476885, + 0.004255419597029686, + 0.004942519590258598, + 0.0023260884918272495, + 0.004143970087170601, + 0.006892537698149681, + 0.009473138488829136, + 0.009263158775866032, + 0.005898734554648399, + 0.007440867368131876, + 0.007274979259818792, + 0.004994668066501617, + 0.005488235969096422, + 0.016939355060458183, + 0.04259531944990158, + 0.02083033137023449, + 0.02115297131240368, + 0.019217049703001976, + 0.021088149398565292, + 0.02276146411895752, + 0.01944429613649845, + 0.03461480885744095, + 0.058970604091882706, + 0.01148687582463026, + 0.007986187003552914, + 0.01193858589977026, + 0.014456270262598991, + 0.005461256019771099, + 0.0064111207611858845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.023671170696616173 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011108812876045704 + ], + "opposite_logprob_tokens": [ + -4.511108875274658 + ], + "logprob_reward": -0.011108812876045704, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011108812876045704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15940193831920624, + 0.11348141729831696, + 0.2581771910190582, + 0.05572857707738876, + 0.04221157357096672, + 0.09556310623884201, + 0.09028827399015427, + 0.0797131285071373, + 0.04020719602704048, + 0.06522762775421143 + ], + "credit_map": [ + 0.02049519680440426, + 0.00692457240074873, + 0.0322808101773262, + 0.017131196334958076, + 0.00658545782789588, + 0.031332243233919144, + 0.006978888995945454, + 0.03767356649041176, + 0.014587360434234142, + 0.007771001197397709, + 0.023589203134179115, + 0.015085947699844837, + 0.005917943082749844, + 0.01545138843357563, + 0.01347337756305933, + 0.017605189234018326, + 0.031073784455657005, + 0.0249649565666914, + 0.029109695926308632, + 0.03283102065324783, + 0.02229795791208744, + 0.04276886209845543, + 0.03191429004073143, + 0.04321662709116936, + 0.005938766524195671, + 0.00791736040264368, + 0.006815474480390549, + 0.006573755759745836, + 0.006233989726752043, + 0.007928935810923576, + 0.00589772779494524, + 0.008422561921179295, + 0.00707634910941124, + 0.006134277209639549, + 0.006807944271713495, + 0.004268353339284658, + 0.0038088734727352858, + 0.005791747942566872, + 0.0035743829794228077, + 0.004749641288071871, + 0.00941715482622385, + 0.009761891327798367, + 0.01821933127939701, + 0.015072474256157875, + 0.008676193654537201, + 0.013431398198008537, + 0.008341868408024311, + 0.01264279242604971, + 0.009215591475367546, + 0.009607288986444473, + 0.014405619353055954, + 0.010705862194299698, + 0.012417986057698727, + 0.01353232841938734, + 0.008172719739377499, + 0.012230878695845604, + 0.008642338216304779, + 0.01226981170475483, + 0.013722836039960384, + 0.007842261344194412, + 0.012702705338597298, + 0.009367282502353191, + 0.006860687397420406, + 0.008305205963551998, + 0.0032327163498848677, + 0.006393017712980509, + 0.005296748131513596, + 0.004981612786650658, + 0.006854934152215719, + 0.005285108927637339, + 0.0033760983496904373, + 0.00478696171194315, + 0.007804097607731819, + 0.01611650548875332, + 0.008556046523153782, + 0.007378182839602232, + 0.008233055472373962, + 0.008165428414940834, + 0.003647388191893697, + 0.005326926242560148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.011108812876045704 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549597769975662 + ], + "opposite_logprob_tokens": [ + -4.38754940032959 + ], + "logprob_reward": -0.012549597769975662, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549597769975662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21227994561195374, + 0.0945490300655365, + 0.271892786026001, + 0.05169447511434555, + 0.08317279070615768, + 0.022852174937725067, + 0.05395408719778061, + 0.12454664707183838, + 0.026243511587381363, + 0.05881454423069954 + ], + "credit_map": [ + 0.027476659044623375, + 0.011911457404494286, + 0.047378938645124435, + 0.0197086613625288, + 0.018049009144306183, + 0.028507506474852562, + 0.019667603075504303, + 0.039580103009939194, + 0.009293410927057266, + 0.00642914604395628, + 0.015140791423618793, + 0.009290418587625027, + 0.00941336527466774, + 0.017010189592838287, + 0.011339864693582058, + 0.0166318342089653, + 0.029713647440075874, + 0.025633351877331734, + 0.028538407757878304, + 0.0330088771879673, + 0.024483345448970795, + 0.04382021725177765, + 0.04113336652517319, + 0.04556157439947128, + 0.005252514500170946, + 0.007719298824667931, + 0.005536577198654413, + 0.00872337818145752, + 0.006240100134164095, + 0.006276911590248346, + 0.0051666367799043655, + 0.00677905697375536, + 0.007872109301388264, + 0.007832798175513744, + 0.014247041195631027, + 0.01152564212679863, + 0.007227386347949505, + 0.01212169136852026, + 0.009808998554944992, + 0.01253711711615324, + 0.002183807548135519, + 0.002205079421401024, + 0.0032453085295856, + 0.001977896550670266, + 0.0021006132010370493, + 0.004474030341953039, + 0.002064828760921955, + 0.004600612446665764, + 0.00762387877330184, + 0.005831330083310604, + 0.009014752693474293, + 0.0026636780239641666, + 0.007221308536827564, + 0.009002144448459148, + 0.004844069480895996, + 0.0077529242262244225, + 0.013630061410367489, + 0.020038215443491936, + 0.01689404807984829, + 0.0157371424138546, + 0.02142960950732231, + 0.015332013368606567, + 0.010487010702490807, + 0.01099854614585638, + 0.0026238728314638138, + 0.003461623564362526, + 0.0037963553331792355, + 0.004039643332362175, + 0.0036592544056475163, + 0.003049143822863698, + 0.0031051968690007925, + 0.0025084223598241806, + 0.005607496947050095, + 0.016491195186972618, + 0.007066552527248859, + 0.0077319517731666565, + 0.005032042507082224, + 0.0063988035544753075, + 0.0054476563818752766, + 0.005038848612457514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.012549597769975662 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029810933396220207 + ], + "opposite_logprob_tokens": [ + -3.529810905456543 + ], + "logprob_reward": -0.029810933396220207, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029810933396220207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14535346627235413, + 0.09965624660253525, + 0.03110470063984394, + 0.06378467381000519, + 0.05699886381626129, + 0.06048651412129402, + 0.12160545587539673, + 0.17242813110351562, + 0.10347138345241547, + 0.145110622048378 + ], + "credit_map": [ + 0.010278664529323578, + 0.005849855020642281, + 0.014204103499650955, + 0.011321939527988434, + 0.007074004039168358, + 0.04317883029580116, + 0.009324890561401844, + 0.04412117600440979, + 0.0158345066010952, + 0.007315841503441334, + 0.017461804673075676, + 0.016194496303796768, + 0.005493729840964079, + 0.012175953015685081, + 0.008827516809105873, + 0.016352389007806778, + 0.0025840539019554853, + 0.002338683232665062, + 0.0061700064688920975, + 0.00437797699123621, + 0.002445171121507883, + 0.002942523453384638, + 0.005098346155136824, + 0.005147937219589949, + 0.007470126263797283, + 0.006360616534948349, + 0.008936558850109577, + 0.009010065346956253, + 0.006459241732954979, + 0.007701924070715904, + 0.006308866664767265, + 0.011537270620465279, + 0.006201491691172123, + 0.005087449215352535, + 0.010898887179791927, + 0.007086783181875944, + 0.006002098321914673, + 0.007928386330604553, + 0.006110941991209984, + 0.007682818919420242, + 0.007254257332533598, + 0.00520280422642827, + 0.006630223710089922, + 0.004378902725875378, + 0.0044402810744941235, + 0.013351274654269218, + 0.004244530573487282, + 0.014984233304858208, + 0.014125008136034012, + 0.009724023751914501, + 0.01934812031686306, + 0.023874301463365555, + 0.018539614975452423, + 0.011244304478168488, + 0.010953187011182308, + 0.013796878978610039, + 0.0302213616669178, + 0.015011519193649292, + 0.04402874410152435, + 0.017687415704131126, + 0.02204550802707672, + 0.013078605756163597, + 0.013070439919829369, + 0.01728452369570732, + 0.008648384362459183, + 0.01740490272641182, + 0.013941909186542034, + 0.014819568023085594, + 0.017658764496445656, + 0.012799151241779327, + 0.008284456096589565, + 0.009914247319102287, + 0.021222682669758797, + 0.04684289172291756, + 0.012994970194995403, + 0.015269206836819649, + 0.010636952705681324, + 0.016748176887631416, + 0.011697439476847649, + 0.009698279201984406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.029810933396220207 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002494083484634757 + ], + "opposite_logprob_tokens": [ + -6.002493858337402 + ], + "logprob_reward": -0.002494083484634757, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002494083484634757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13707205653190613, + 0.06220793351531029, + 0.038109660148620605, + 0.032033197581768036, + 0.04240688681602478, + 0.07529699057340622, + 0.1362362951040268, + 0.14476443827152252, + 0.10219357162714005, + 0.2296789437532425 + ], + "credit_map": [ + 0.016565609723329544, + 0.007585383020341396, + 0.031040173023939133, + 0.015284251421689987, + 0.006774737499654293, + 0.024379000067710876, + 0.009673986583948135, + 0.02576890029013157, + 0.013472095131874084, + 0.0035465487744659185, + 0.013926561921834946, + 0.007239116821438074, + 0.0024730598088353872, + 0.005700075998902321, + 0.005192167591303587, + 0.010658311657607555, + 0.004975805524736643, + 0.003939665853977203, + 0.006628239061683416, + 0.004330590832978487, + 0.002641837578266859, + 0.003830366302281618, + 0.004808336962014437, + 0.0069548203609883785, + 0.0025647177826613188, + 0.003771220799535513, + 0.005120960064232349, + 0.004942233208566904, + 0.0024549481458961964, + 0.004098410252481699, + 0.0027497566770762205, + 0.006330951116979122, + 0.004698654171079397, + 0.004632619209587574, + 0.007463333196938038, + 0.005509209353476763, + 0.004463654477149248, + 0.005438886117190123, + 0.004761210177093744, + 0.005439319182187319, + 0.008929464966058731, + 0.006601911038160324, + 0.010976098477840424, + 0.007059907540678978, + 0.00581959867849946, + 0.01480680238455534, + 0.00530753331258893, + 0.015795685350894928, + 0.01796756125986576, + 0.01595432497560978, + 0.014600000344216824, + 0.015093828551471233, + 0.019480912014842033, + 0.02267599105834961, + 0.01006242074072361, + 0.0204012468457222, + 0.01624218560755253, + 0.024205949157476425, + 0.01713859662413597, + 0.014972087927162647, + 0.027014045044779778, + 0.01596912369132042, + 0.01634376309812069, + 0.012878688052296638, + 0.007972300052642822, + 0.03081870824098587, + 0.01047083456069231, + 0.00960620492696762, + 0.010469786822795868, + 0.01390120293945074, + 0.011110836640000343, + 0.007843692786991596, + 0.04464620724320412, + 0.06558220833539963, + 0.02053934708237648, + 0.024210773408412933, + 0.018508128821849823, + 0.024574950337409973, + 0.014587314799427986, + 0.017030032351613045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.002494083484634757 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long shot, shot from a low, upward angle, the camera tilts steadily upward, tracking the paths of countless light trails ascending across the deep, inky, velvety midnight sky. Hundreds of dazzling golden arcs and brilliant silver beams, trailing long, shimmering trails of particles, shoot upward from the invisible bottom of the frame at astonishing speed, carving graceful, powerful parabolas through the sky, tearing through the night. As the camera tilts to its highest point, all the paths converge at a precise, breathtaking focal point at the very center of the frame, at the apex of their ascent. After a brief pause, they brilliantly merge and instantly explode into a vast, symmetrical, spherical corolla of pure light, illuminating the inky black sky with its intense radiance. The intertwining, yet distinct, strands of gold and silver create a visual climax of grandeur, celebration, and wonder.", + "seed": 28714, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000046/group_9/index_1139_seed_28714_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.023288099095225334, + "vlm_logprob": -0.023288099095225334, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.023288099095225334, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035699643194675446 + ], + "opposite_logprob_tokens": [ + -3.7856996059417725 + ], + "logprob_reward": -0.035699643194675446, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.035699643194675446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24300535023212433, + 0.05911047384142876, + 0.045046813786029816, + 0.11908610165119171, + 0.04774366691708565, + 0.04515669867396355, + 0.08639530092477798, + 0.05762355774641037, + 0.16947200894355774, + 0.1273600459098816 + ], + "credit_map": [ + 0.019820861518383026, + 0.02164786495268345, + 0.03583512827754021, + 0.025463351979851723, + 0.019267169758677483, + 0.054978564381599426, + 0.024001730605959892, + 0.041990652680397034, + 0.0038750218227505684, + 0.003424630034714937, + 0.0037438878789544106, + 0.016471704468131065, + 0.006661299616098404, + 0.005035554524511099, + 0.015858257189393044, + 0.004040110856294632, + 0.005898016970604658, + 0.006237839814275503, + 0.004970569629222155, + 0.00974893569946289, + 0.002901424653828144, + 0.004804742522537708, + 0.006617601960897446, + 0.003867678577080369, + 0.015621635131537914, + 0.016689427196979523, + 0.011001288890838623, + 0.025920620188117027, + 0.008109765127301216, + 0.010296243242919445, + 0.020028652623295784, + 0.011418458074331284, + 0.006350411567837, + 0.006594937294721603, + 0.006539758760482073, + 0.0043491520918905735, + 0.006715536117553711, + 0.006476872600615025, + 0.0033832406625151634, + 0.0073337554931640625, + 0.006263107061386108, + 0.0064881788566708565, + 0.006353166885674, + 0.004423219244927168, + 0.005623251665383577, + 0.006978367455303669, + 0.00369714992120862, + 0.005330251529812813, + 0.013890057802200317, + 0.01098176185041666, + 0.00862307008355856, + 0.009846876375377178, + 0.011414696462452412, + 0.012542732991278172, + 0.007362544070929289, + 0.011733553372323513, + 0.009271720424294472, + 0.006717092823237181, + 0.007437891326844692, + 0.0051755113527178764, + 0.008512993343174458, + 0.007379340939223766, + 0.005808309651911259, + 0.007320690900087357, + 0.01791536994278431, + 0.01749792881309986, + 0.025638531893491745, + 0.017459377646446228, + 0.022268807515501976, + 0.018106510862708092, + 0.031343936920166016, + 0.019241537898778915, + 0.015332006849348545, + 0.013905785046517849, + 0.027027402073144913, + 0.009162480011582375, + 0.016324307769536972, + 0.014974887482821941, + 0.014818280003964901, + 0.015814896672964096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.035699643194675446 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026383813470602036 + ], + "opposite_logprob_tokens": [ + -3.651383876800537 + ], + "logprob_reward": -0.026383813470602036, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026383813470602036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13573673367500305, + 0.049032602459192276, + 0.02750890515744686, + 0.0767403095960617, + 0.1553385704755783, + 0.04710235819220543, + 0.05481649562716484, + 0.1709076315164566, + 0.1171165183186531, + 0.16569983959197998 + ], + "credit_map": [ + 0.010573907755315304, + 0.013712155632674694, + 0.01441130694001913, + 0.01317495759576559, + 0.01546567678451538, + 0.028653260320425034, + 0.013814954087138176, + 0.025930527597665787, + 0.006376341916620731, + 0.004863296169787645, + 0.006055494770407677, + 0.00629575876519084, + 0.006054627243429422, + 0.0075177038088440895, + 0.006847530137747526, + 0.005021856166422367, + 0.003389661433175206, + 0.003560679266229272, + 0.0029641331639140844, + 0.0055276937782764435, + 0.0015870750648900867, + 0.002629438415169716, + 0.004916283302009106, + 0.00293394410982728, + 0.007688039913773537, + 0.006372366566210985, + 0.00846356712281704, + 0.020869525149464607, + 0.004409871529787779, + 0.004728879779577255, + 0.01854097470641136, + 0.005667088553309441, + 0.01626548357307911, + 0.021730830892920494, + 0.026426568627357483, + 0.013279546983540058, + 0.020504271611571312, + 0.02529667504131794, + 0.01434448640793562, + 0.017490725964307785, + 0.005374390631914139, + 0.006383193656802177, + 0.007592654321342707, + 0.005917130503803492, + 0.006650297902524471, + 0.006874618586152792, + 0.004428194370120764, + 0.0038818831089884043, + 0.007616784889250994, + 0.00904100202023983, + 0.0043930490501224995, + 0.007147783879190683, + 0.0070091551169753075, + 0.007253054529428482, + 0.005310664419084787, + 0.0070450082421302795, + 0.022639727219939232, + 0.023980991914868355, + 0.016709452494978905, + 0.01617087610065937, + 0.029753098264336586, + 0.021408813074231148, + 0.01408290397375822, + 0.026161786168813705, + 0.01623866893351078, + 0.012071816250681877, + 0.019193103536963463, + 0.01074484083801508, + 0.013118311762809753, + 0.014708577655255795, + 0.015406920574605465, + 0.01563429646193981, + 0.020044580101966858, + 0.019354606047272682, + 0.027985699474811554, + 0.011956466361880302, + 0.02535359002649784, + 0.020544834434986115, + 0.021169139072299004, + 0.019290940836071968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 4 + ], + "selected_reward": -0.026383813470602036 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024986020289361477 + ], + "opposite_logprob_tokens": [ + -6.002498626708984 + ], + "logprob_reward": -0.0024986020289361477, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024986020289361477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2761750817298889, + 0.05969063192605972, + 0.036273933947086334, + 0.09171918034553528, + 0.056465961039066315, + 0.050423700362443924, + 0.09117712080478668, + 0.06579023599624634, + 0.07989205420017242, + 0.19239211082458496 + ], + "credit_map": [ + 0.021416211500763893, + 0.031821493059396744, + 0.023653220385313034, + 0.030254671350121498, + 0.028958003968000412, + 0.056923288851976395, + 0.033343080431222916, + 0.049805108457803726, + 0.006322786211967468, + 0.009255904704332352, + 0.007002450525760651, + 0.007636686787009239, + 0.009518621489405632, + 0.007081398740410805, + 0.00686492957174778, + 0.006007852498441935, + 0.00500073004513979, + 0.006030290853232145, + 0.002833537757396698, + 0.007391905412077904, + 0.002682451857253909, + 0.0028926131781190634, + 0.005693034268915653, + 0.0037493736017495394, + 0.009287355467677116, + 0.009199731983244419, + 0.00945513416081667, + 0.022259706631302834, + 0.0052668750286102295, + 0.0074441260658204556, + 0.020022602751851082, + 0.008783653378486633, + 0.005199274979531765, + 0.006740131881088018, + 0.006965760141611099, + 0.0048328121192753315, + 0.011388092301785946, + 0.00946885533630848, + 0.006821182556450367, + 0.00504984799772501, + 0.004888684023171663, + 0.0073058572597801685, + 0.006263916380703449, + 0.00488209817558527, + 0.009150105528533459, + 0.006780719850212336, + 0.004743853583931923, + 0.006408466957509518, + 0.011198931373655796, + 0.015568112954497337, + 0.010177654214203358, + 0.010421203449368477, + 0.011823623441159725, + 0.013502382673323154, + 0.007719376590102911, + 0.010765839368104935, + 0.009663674049079418, + 0.0074516902677714825, + 0.006672506686300039, + 0.007137599866837263, + 0.007242259569466114, + 0.009284709580242634, + 0.005839222576469183, + 0.012498571537435055, + 0.009176723659038544, + 0.005406985525041819, + 0.011044206097722054, + 0.00814307015389204, + 0.008192643523216248, + 0.011434935964643955, + 0.01590736396610737, + 0.010586126707494259, + 0.019807880744338036, + 0.025310931727290154, + 0.038410767912864685, + 0.015264078043401241, + 0.025946790352463722, + 0.028447862714529037, + 0.020298071205615997, + 0.018905730918049812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.0024986020289361477 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866420477628708 + ], + "opposite_logprob_tokens": [ + -3.048664093017578 + ], + "logprob_reward": -0.04866420477628708, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866420477628708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11339477449655533, + 0.09699077904224396, + 0.06464721262454987, + 0.07008983194828033, + 0.0679171234369278, + 0.08291766792535782, + 0.07610391080379486, + 0.1904262751340866, + 0.09954094141721725, + 0.13797149062156677 + ], + "credit_map": [ + 0.009369811043143272, + 0.013673059642314911, + 0.011385231278836727, + 0.013127024285495281, + 0.01094481535255909, + 0.017215333878993988, + 0.013677043840289116, + 0.024002451449632645, + 0.00754784420132637, + 0.00888240709900856, + 0.01139913871884346, + 0.015916937962174416, + 0.013826395384967327, + 0.010848660953342915, + 0.01552234124392271, + 0.013047059066593647, + 0.008512944914400578, + 0.010815835557878017, + 0.011000305414199829, + 0.007027777843177319, + 0.004664032254368067, + 0.005646266974508762, + 0.009824473410844803, + 0.007155579514801502, + 0.006634753197431564, + 0.008358929306268692, + 0.007383657153695822, + 0.01823493279516697, + 0.004305966664105654, + 0.004421802703291178, + 0.01563280262053013, + 0.005116990301758051, + 0.007095042150467634, + 0.008391877636313438, + 0.0147960614413023, + 0.004978989716619253, + 0.009061706252396107, + 0.011653658002614975, + 0.00435802573338151, + 0.007581758312880993, + 0.00885038822889328, + 0.012035265564918518, + 0.006980033591389656, + 0.005353650078177452, + 0.023251255974173546, + 0.013101494871079922, + 0.007153294049203396, + 0.006192290224134922, + 0.010443811304867268, + 0.014575607143342495, + 0.006886385381221771, + 0.006250942125916481, + 0.013678042218089104, + 0.010458127595484257, + 0.006103716790676117, + 0.007707284763455391, + 0.02857913263142109, + 0.01914847083389759, + 0.022524770349264145, + 0.01623094268143177, + 0.024796565994620323, + 0.026128603145480156, + 0.02002166025340557, + 0.032996125519275665, + 0.01081869751214981, + 0.008217540569603443, + 0.014527237974107265, + 0.00979702640324831, + 0.012417790479958057, + 0.011604796163737774, + 0.018402399495244026, + 0.013755451887845993, + 0.015243085101246834, + 0.02146359719336033, + 0.01405553799122572, + 0.0126863494515419, + 0.022990677505731583, + 0.016119128093123436, + 0.016322750598192215, + 0.019090373069047928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.04866420477628708 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a long shot featuring a low, upward camera angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera tilt steadily upward while tracking light trails ascending across a deep, inky, velvety midnight sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are hundreds of dazzling golden arcs and brilliant silver beams shooting upward from the bottom of the frame at astonishing speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the light trails carve graceful parabolas and converge at a precise focal point at the very center of the frame at the apex of their ascent?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the light trails merge and instantly explode into a vast, symmetrical, spherical corolla of pure light that illuminates the sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003194232005625963 + ], + "opposite_logprob_tokens": [ + -5.753194332122803 + ], + "logprob_reward": -0.003194232005625963, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003194232005625963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09464333951473236, + 0.047956641763448715, + 0.029863886535167694, + 0.039705634117126465, + 0.0316089391708374, + 0.05258296802639961, + 0.0990401953458786, + 0.4426521956920624, + 0.08380453288555145, + 0.07814167439937592 + ], + "credit_map": [ + 0.010144789703190327, + 0.009203820489346981, + 0.010944928042590618, + 0.009441203437745571, + 0.007629218976944685, + 0.014094291254878044, + 0.01116272434592247, + 0.022022360935807228, + 0.0031461413018405437, + 0.0032684989273548126, + 0.004029904026538134, + 0.010710371658205986, + 0.006349783390760422, + 0.0042740348726511, + 0.010769610293209553, + 0.005408294033259153, + 0.0036144587211310863, + 0.003409202676266432, + 0.0027178809978067875, + 0.007881486788392067, + 0.0018601323245093226, + 0.002068620640784502, + 0.0052157300524413586, + 0.0030963746830821037, + 0.0031810328364372253, + 0.004214635118842125, + 0.0045554423704743385, + 0.011102869175374508, + 0.002527941484004259, + 0.0023479682859033346, + 0.009114282205700874, + 0.002661460544914007, + 0.002478652400895953, + 0.002695034956559539, + 0.008770340122282505, + 0.00337031576782465, + 0.003072673687711358, + 0.0055322651751339436, + 0.0027648108080029488, + 0.002924844389781356, + 0.0052267140708863735, + 0.00658030342310667, + 0.00741388276219368, + 0.0052002123557031155, + 0.0077518438920378685, + 0.007557503879070282, + 0.004867387004196644, + 0.007985115982592106, + 0.015814702957868576, + 0.015910621732473373, + 0.008327268995344639, + 0.00805890467017889, + 0.01694389060139656, + 0.013082536868751049, + 0.007583151571452618, + 0.013319114223122597, + 0.05790162831544876, + 0.05666714161634445, + 0.04743389040231705, + 0.0401504822075367, + 0.06985527276992798, + 0.05668698623776436, + 0.03970314934849739, + 0.07425365597009659, + 0.009332765825092793, + 0.008977411314845085, + 0.010041921399533749, + 0.008313803002238274, + 0.008401929400861263, + 0.008766386657953262, + 0.01809110678732395, + 0.011879201978445053, + 0.008467534556984901, + 0.007885551080107689, + 0.017989521846175194, + 0.005555194802582264, + 0.006912394426763058, + 0.009809297509491444, + 0.008622247725725174, + 0.012899928726255894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.003194232005625963 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_0/index_tensor([2055, 2055])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_0/index_tensor([2055, 2055])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f75bba71f91e28f3c4b4089986dc7c1f5a2db21e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_0/index_tensor([2055, 2055])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "highly detailed documentary medium shot, shot from a level angle, slowly and steadily moves forward, gradually deepening into a family dining room filled with laughter and chatter. In the center of a dark, round wooden table, a stainless steel hot pot bubbling, half filled with a vibrant red spicy broth swirling with dried chilies and Sichuan peppercorns, the other half a milky white bone broth simmering with red dates and goji berries. Thick, white steam rises continuously from the pot, like a soft filter, imbuing the entire scene with a warm, fuzzy quality. A grandmother with graying hair and a kind smile carefully uses chopsticks to transfer a boiled fish ball to her granddaughter's bowl. The little girl, with pigtails and a pink sweater, gazes expectantly at the food in her bowl. Across from them, a middle-aged father in a dark blue V-neck sweater chats with his wife, a bright smile on his face. His mother, in a beige turtleneck, responds gently, expertly picking up a piece of green vegetable with her chopsticks and preparing to add it to the pot. The table was covered with a variety of ingredients: neatly arranged beef rolls, emerald greens, crisp white tofu slices, and fresh mushrooms. In front of each person was a bowl of prepared dipping sauce. A warm yellow chandelier hung above the table provided the sole source of light, its glow illuminating the steaming air and framing the happy expressions on everyone's faces, creating a warm, lively, and vibrant atmosphere for a family dinner.", + "seed": 18410, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_0/index_2055_seed_18410_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0695539548061788, + "vlm_logprob": -0.0695539548061788, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.0695539548061788, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041428081691265106 + ], + "opposite_logprob_tokens": [ + -3.7914280891418457 + ], + "logprob_reward": -0.041428081691265106, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.041428081691265106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048603493720293045, + 0.03768591582775116, + 0.07921181619167328, + 0.12991301715373993, + 0.15898942947387695, + 0.07875915616750717, + 0.08292100578546524, + 0.059002138674259186, + 0.014891856350004673, + 0.310022234916687 + ], + "credit_map": [ + 0.008882558904588223, + 0.007587239146232605, + 0.00631130812689662, + 0.004529870580881834, + 0.006240631453692913, + 0.006398941855877638, + 0.004815582651644945, + 0.003837354714050889, + 0.0035133794881403446, + 0.004088662564754486, + 0.006804453209042549, + 0.0026853582821786404, + 0.005616142880171537, + 0.005592689383774996, + 0.00612975936383009, + 0.0032554692588746548, + 0.006936077028512955, + 0.020756419748067856, + 0.007938981987535954, + 0.00829455815255642, + 0.010695182718336582, + 0.009940092451870441, + 0.008845513686537743, + 0.005804982967674732, + 0.009301516227424145, + 0.026473823934793472, + 0.017637915909290314, + 0.012780743651092052, + 0.020599817857146263, + 0.01289780717343092, + 0.021751344203948975, + 0.008470033295452595, + 0.014348084107041359, + 0.01581890881061554, + 0.03575022891163826, + 0.008341165259480476, + 0.015244708396494389, + 0.048951659351587296, + 0.013727877289056778, + 0.006806791760027409, + 0.009053659625351429, + 0.01609162986278534, + 0.005930874031037092, + 0.006043886765837669, + 0.011962131597101688, + 0.009181217290461063, + 0.014760885387659073, + 0.005734862759709358, + 0.006308692507445812, + 0.031310778111219406, + 0.0066444450058043, + 0.00605121860280633, + 0.01614786870777607, + 0.005119323264807463, + 0.008540762588381767, + 0.002797910710796714, + 0.006301378831267357, + 0.0060378191992640495, + 0.00435301149263978, + 0.006581808440387249, + 0.018279435113072395, + 0.006697104312479496, + 0.006966388784348965, + 0.0037851885426789522, + 0.0014765772502869368, + 0.001831227564252913, + 0.001403478323481977, + 0.0021217872854322195, + 0.002423305297270417, + 0.0020283646881580353, + 0.002226767595857382, + 0.0013803478796035051, + 0.022764816880226135, + 0.03584128990769386, + 0.06265603750944138, + 0.10101679712533951, + 0.03386315330862999, + 0.025117333978414536, + 0.017550529912114143, + 0.011212272569537163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 3 + ], + "selected_reward": -0.041428081691265106 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005968248937278986 + ], + "opposite_logprob_tokens": [ + -5.13096809387207 + ], + "logprob_reward": -0.005968248937278986, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005968248937278986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1926434487104416, + 0.06187031418085098, + 0.15908347070217133, + 0.15010106563568115, + 0.04521942138671875, + 0.08333998173475266, + 0.02892703004181385, + 0.044085580855607986, + 0.022391069680452347, + 0.21233859658241272 + ], + "credit_map": [ + 0.018824951723217964, + 0.02693925052881241, + 0.017936434596776962, + 0.015120444819331169, + 0.04196741804480553, + 0.02995217591524124, + 0.027882298454642296, + 0.014020484872162342, + 0.00676367012783885, + 0.007787420880049467, + 0.007811722345650196, + 0.0040633780881762505, + 0.008534845896065235, + 0.014096030034124851, + 0.00796259380877018, + 0.004850655794143677, + 0.010651187971234322, + 0.02350345440208912, + 0.029393868520855904, + 0.020701555535197258, + 0.014680328778922558, + 0.024169806391000748, + 0.025329990312457085, + 0.01065328624099493, + 0.015157139860093594, + 0.017557863146066666, + 0.02428094483911991, + 0.017921941354870796, + 0.022586511448025703, + 0.03093004785478115, + 0.01220569759607315, + 0.009460912086069584, + 0.004840836860239506, + 0.006039726547896862, + 0.00488497456535697, + 0.0035181869752705097, + 0.004720171447843313, + 0.0097736194729805, + 0.00741440849378705, + 0.004027501214295626, + 0.015592241659760475, + 0.011985061690211296, + 0.006609317846596241, + 0.005896105896681547, + 0.01164516806602478, + 0.014723069965839386, + 0.008464672602713108, + 0.008424350991845131, + 0.002947478322312236, + 0.007150373421609402, + 0.0014286538353189826, + 0.0021676141768693924, + 0.004851624369621277, + 0.0042817615903913975, + 0.004614057019352913, + 0.0014854689361527562, + 0.009003156796097755, + 0.005232821684330702, + 0.003374528605490923, + 0.004897695500403643, + 0.008976373821496964, + 0.00579569349065423, + 0.0034364040475338697, + 0.0033689108677208424, + 0.0028556559700518847, + 0.0038232379592955112, + 0.001476001343689859, + 0.003497267374768853, + 0.004295808263123035, + 0.0019074355950579047, + 0.0024336022324860096, + 0.0021020628046244383, + 0.030993429943919182, + 0.021383486688137054, + 0.042896904051303864, + 0.06548494845628738, + 0.01599203795194626, + 0.01394693460315466, + 0.012580364011228085, + 0.009060501120984554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.005968248937278986 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903488516807556 + ], + "opposite_logprob_tokens": [ + -2.4640347957611084 + ], + "logprob_reward": -0.08903488516807556, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903488516807556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054345499724149704, + 0.034006278961896896, + 0.05445316433906555, + 0.044900503009557724, + 0.046403612941503525, + 0.09020894765853882, + 0.04335509613156319, + 0.029790343716740608, + 0.014542259275913239, + 0.5879943370819092 + ], + "credit_map": [ + 0.005379193928092718, + 0.008157243952155113, + 0.0036787919234484434, + 0.004429378546774387, + 0.012445883825421333, + 0.008410747162997723, + 0.008358986116945744, + 0.003485272405669093, + 0.002926252316683531, + 0.0038355018477886915, + 0.004030134063214064, + 0.0028749946504831314, + 0.00654565030708909, + 0.005206369329243898, + 0.00594074372202158, + 0.002646633656695485, + 0.006581202149391174, + 0.015386978164315224, + 0.006455030757933855, + 0.005281422752887011, + 0.005427560769021511, + 0.004737279377877712, + 0.005410551559180021, + 0.005173137411475182, + 0.004568392410874367, + 0.0055104810744524, + 0.006747562438249588, + 0.005082395859062672, + 0.0074796658009290695, + 0.008785497397184372, + 0.003649801481515169, + 0.0030767093412578106, + 0.005154138430953026, + 0.005821833852678537, + 0.0074959066696465015, + 0.004337437450885773, + 0.004439235664904118, + 0.00785081461071968, + 0.007431033067405224, + 0.0038732143584638834, + 0.009964794851839542, + 0.02195412665605545, + 0.004855486564338207, + 0.00701015442609787, + 0.01674267090857029, + 0.01267362479120493, + 0.012000543996691704, + 0.005007550120353699, + 0.00302565167658031, + 0.013608061708509922, + 0.004750571679323912, + 0.0026609227061271667, + 0.010366237722337246, + 0.004225863143801689, + 0.003284808713942766, + 0.0014329747064039111, + 0.005983556620776653, + 0.0037496553268283606, + 0.002454821253195405, + 0.0031703291460871696, + 0.004840657114982605, + 0.005019051022827625, + 0.002430890453979373, + 0.002141382545232773, + 0.0017836922779679298, + 0.002473533386364579, + 0.0008185020415112376, + 0.0013219182146713138, + 0.0039394269697368145, + 0.0013264393201097846, + 0.0013789061922580004, + 0.0014998409897089005, + 0.05148392170667648, + 0.06729769706726074, + 0.1259063184261322, + 0.2000218778848648, + 0.05708915740251541, + 0.03156778961420059, + 0.03440401330590248, + 0.020223531872034073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.08903488516807556 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776213206350803 + ], + "opposite_logprob_tokens": [ + -4.6347761154174805 + ], + "logprob_reward": -0.009776213206350803, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776213206350803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03850049525499344, + 0.027435295283794403, + 0.11356491595506668, + 0.09805590659379959, + 0.05921784043312073, + 0.17041587829589844, + 0.04032352194190025, + 0.047215141355991364, + 0.02949022874236107, + 0.3757808208465576 + ], + "credit_map": [ + 0.005871348548680544, + 0.005008753389120102, + 0.004935374949127436, + 0.004185363184660673, + 0.0037036051508039236, + 0.004386580549180508, + 0.007233747281134129, + 0.0031757205724716187, + 0.0022554330062121153, + 0.0032074672635644674, + 0.004809701815247536, + 0.0025769209023565054, + 0.0035938795190304518, + 0.004354280419647694, + 0.004565509967505932, + 0.002072100294753909, + 0.009879444725811481, + 0.02815008908510208, + 0.014035910367965698, + 0.014576157554984093, + 0.014143621549010277, + 0.010917644947767258, + 0.01400822028517723, + 0.007853815332055092, + 0.008869467303156853, + 0.010819460265338421, + 0.016326095908880234, + 0.010871920734643936, + 0.02073998935520649, + 0.016317591071128845, + 0.008849849924445152, + 0.005261522252112627, + 0.004511374048888683, + 0.007038966752588749, + 0.010482598096132278, + 0.004852938000112772, + 0.004983299411833286, + 0.01796901598572731, + 0.005837449803948402, + 0.003542194375768304, + 0.029783571138978004, + 0.03371019288897514, + 0.009690837934613228, + 0.015113702975213528, + 0.01824648678302765, + 0.03145371004939079, + 0.01826789788901806, + 0.014149464666843414, + 0.0041700308211147785, + 0.009780017659068108, + 0.002567622810602188, + 0.004132130183279514, + 0.005069100297987461, + 0.007839161902666092, + 0.004654343705624342, + 0.0021111113019287586, + 0.009689610451459885, + 0.005823191720992327, + 0.00431655440479517, + 0.004278792534023523, + 0.01010897010564804, + 0.00586465559899807, + 0.0034203501418232918, + 0.0037130124401301146, + 0.003823918290436268, + 0.0034609525464475155, + 0.0016003026394173503, + 0.0032450442668050528, + 0.009981817565858364, + 0.003417876549065113, + 0.0020725633949041367, + 0.001887751161120832, + 0.03374086320400238, + 0.04290449246764183, + 0.07921014726161957, + 0.13575582206249237, + 0.03132006898522377, + 0.01463239174336195, + 0.023838404566049576, + 0.014378626830875874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.009776213206350803 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156234502792358 + ], + "opposite_logprob_tokens": [ + -1.7015624046325684 + ], + "logprob_reward": -0.20156234502792358, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156234502792358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09158288687467575, + 0.03701665624976158, + 0.1905149668455124, + 0.1292261928319931, + 0.03253849595785141, + 0.03277342766523361, + 0.06939136236906052, + 0.03560865297913551, + 0.024656163528561592, + 0.3566911518573761 + ], + "credit_map": [ + 0.008375601842999458, + 0.009177190251648426, + 0.014593244530260563, + 0.006481992546468973, + 0.01980021595954895, + 0.015866035595536232, + 0.010629047639667988, + 0.006659562699496746, + 0.004825753625482321, + 0.005354849621653557, + 0.0033641746267676353, + 0.0031952988356351852, + 0.006644774693995714, + 0.005042806267738342, + 0.005364217795431614, + 0.003224777989089489, + 0.01815686747431755, + 0.058395300060510635, + 0.018259771168231964, + 0.021036483347415924, + 0.022024178877472878, + 0.019311124458909035, + 0.020162831991910934, + 0.013168406672775745, + 0.008214001543819904, + 0.01004242803901434, + 0.030294271185994148, + 0.0077413152903318405, + 0.020823197439312935, + 0.035239577293395996, + 0.010659448802471161, + 0.0062119560316205025, + 0.0027095251716673374, + 0.00441306596621871, + 0.00303688389249146, + 0.002826487412676215, + 0.003879665629938245, + 0.005963355302810669, + 0.007108987774699926, + 0.002600528299808502, + 0.004152543377131224, + 0.006738321390002966, + 0.0025636397767812014, + 0.0035418530460447073, + 0.005515753291547298, + 0.0035829422995448112, + 0.004195117391645908, + 0.002483259653672576, + 0.004608152434229851, + 0.02089677006006241, + 0.003746427595615387, + 0.00295344740152359, + 0.020517677068710327, + 0.009185346774756908, + 0.005486126057803631, + 0.001997415442019701, + 0.004372468218207359, + 0.004142871592193842, + 0.0021164773497730494, + 0.003127929987385869, + 0.009967353194952011, + 0.005529611371457577, + 0.003844269784167409, + 0.002507670782506466, + 0.003171105869114399, + 0.00330224703066051, + 0.001170925796031952, + 0.002602861961349845, + 0.007846055552363396, + 0.002283141016960144, + 0.002088621025905013, + 0.0021912050433456898, + 0.029842812567949295, + 0.043511763215065, + 0.07163256406784058, + 0.13343575596809387, + 0.030141014605760574, + 0.011723051778972149, + 0.02328832633793354, + 0.013115867972373962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.20156234502792358 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly detailed documentary medium shot, shot from a level angle, slowly and steadily moves forward, gradually deepening into a family dining room filled with laughter and chatter. In the center of a dark, round wooden table, a stainless steel hot pot bubbling, half filled with a vibrant red spicy broth swirling with dried chilies and Sichuan peppercorns, the other half a milky white bone broth simmering with red dates and goji berries. Thick, white steam rises continuously from the pot, like a soft filter, imbuing the entire scene with a warm, fuzzy quality. A grandmother with graying hair and a kind smile carefully uses chopsticks to transfer a boiled fish ball to her granddaughter's bowl. The little girl, with pigtails and a pink sweater, gazes expectantly at the food in her bowl. Across from them, a middle-aged father in a dark blue V-neck sweater chats with his wife, a bright smile on his face. His mother, in a beige turtleneck, responds gently, expertly picking up a piece of green vegetable with her chopsticks and preparing to add it to the pot. The table was covered with a variety of ingredients: neatly arranged beef rolls, emerald greens, crisp white tofu slices, and fresh mushrooms. In front of each person was a bowl of prepared dipping sauce. A warm yellow chandelier hung above the table provided the sole source of light, its glow illuminating the steaming air and framing the happy expressions on everyone's faces, creating a warm, lively, and vibrant atmosphere for a family dinner.", + "seed": 18411, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_0/index_2055_seed_18411_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04167031170800328, + "vlm_logprob": -0.04167031170800328, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.04167031170800328, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054139748215675354 + ], + "opposite_logprob_tokens": [ + -3.4291398525238037 + ], + "logprob_reward": -0.054139748215675354, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.054139748215675354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10876471549272537, + 0.4111437201499939, + 0.03205443173646927, + 0.03515704348683357, + 0.16837914288043976, + 0.039748452603816986, + 0.020190654322504997, + 0.02684750035405159, + 0.06098220497369766, + 0.09673210233449936 + ], + "credit_map": [ + 0.00869621429592371, + 0.011741796508431435, + 0.008227081038057804, + 0.009136606007814407, + 0.023335043340921402, + 0.026626914739608765, + 0.012776507996022701, + 0.008224551565945148, + 0.1341780424118042, + 0.026899246498942375, + 0.04095270484685898, + 0.01680547557771206, + 0.04896198585629463, + 0.09539002180099487, + 0.01455891877412796, + 0.03339731693267822, + 0.005355117842555046, + 0.0046976725570857525, + 0.002451193518936634, + 0.003728027455508709, + 0.0035666823387145996, + 0.006846352946013212, + 0.00307158799842, + 0.002337794750928879, + 0.005900798365473747, + 0.005105944350361824, + 0.0028140738140791655, + 0.003251889953389764, + 0.0051489779725670815, + 0.004995426163077354, + 0.0024068120401352644, + 0.005533118732273579, + 0.018862999975681305, + 0.06537462770938873, + 0.008520116098225117, + 0.009552580304443836, + 0.030390875414013863, + 0.01838165707886219, + 0.006492717657238245, + 0.010803548619151115, + 0.005398075561970472, + 0.0072316057048738, + 0.003934585023671389, + 0.006452889181673527, + 0.005888105370104313, + 0.004440331365913153, + 0.0038453235756605864, + 0.002557538915425539, + 0.0025462752673774958, + 0.0028759727720171213, + 0.002166134538128972, + 0.0034798625856637955, + 0.0031942855566740036, + 0.0028564773965626955, + 0.001457309816032648, + 0.0016143372049555182, + 0.004058543127030134, + 0.0040375529788434505, + 0.0034114178270101547, + 0.00422927550971508, + 0.003126027062535286, + 0.0030846772715449333, + 0.0024787448346614838, + 0.002421261975541711, + 0.006900226231664419, + 0.016654985025525093, + 0.005341030191630125, + 0.007942854426801205, + 0.008678989484906197, + 0.0067632622085511684, + 0.005421383306384087, + 0.0032794764265418053, + 0.011732062324881554, + 0.014356536790728569, + 0.004048321861773729, + 0.004527243319898844, + 0.03592560067772865, + 0.017936749383807182, + 0.004825449548661709, + 0.0033801414538174868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.054139748215675354 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006761291529983282 + ], + "opposite_logprob_tokens": [ + -5.006761074066162 + ], + "logprob_reward": -0.006761291529983282, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006761291529983282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3251379132270813, + 0.096258245408535, + 0.18705125153064728, + 0.026462672278285027, + 0.09668511897325516, + 0.040451355278491974, + 0.01337437517940998, + 0.02421538345515728, + 0.07487383484840393, + 0.11548992246389389 + ], + "credit_map": [ + 0.02294672653079033, + 0.016434868797659874, + 0.01629503443837166, + 0.01817416399717331, + 0.0575600229203701, + 0.11171426624059677, + 0.05862812325358391, + 0.023384666070342064, + 0.012355991639196873, + 0.01392012182623148, + 0.007884257473051548, + 0.007826016284525394, + 0.014475730247795582, + 0.01609371416270733, + 0.00862421840429306, + 0.01507819164544344, + 0.04040580242872238, + 0.01738845929503441, + 0.009835545904934406, + 0.011767033487558365, + 0.01716724783182144, + 0.0591750405728817, + 0.01767541840672493, + 0.013636691495776176, + 0.004406863357871771, + 0.004674831405282021, + 0.002167775295674801, + 0.0021651419810950756, + 0.0038415181916207075, + 0.0035814393777400255, + 0.0028253798373043537, + 0.0027997188735753298, + 0.015617098659276962, + 0.03237118571996689, + 0.0063025979325175285, + 0.006273204460740089, + 0.01126284059137106, + 0.012443140149116516, + 0.005671938881278038, + 0.006743109319359064, + 0.004989840555936098, + 0.007880372926592827, + 0.0037780238781124353, + 0.006916367914527655, + 0.004978841636329889, + 0.0036765066906809807, + 0.0055189901031553745, + 0.0027124076150357723, + 0.0018791394541040063, + 0.0025631156750023365, + 0.001338996458798647, + 0.001388115924783051, + 0.0019384965999051929, + 0.0016894033178687096, + 0.001253100112080574, + 0.0013240075204521418, + 0.002966106403619051, + 0.0027458746917545795, + 0.0034403405152261257, + 0.003566934261471033, + 0.0029153754003345966, + 0.0029721674509346485, + 0.002194514498114586, + 0.003414069302380085, + 0.00881369598209858, + 0.018552280962467194, + 0.006660157348960638, + 0.007395679131150246, + 0.01269943080842495, + 0.0074968524277210236, + 0.0075056045316159725, + 0.005750127602368593, + 0.012963474728167057, + 0.035087134689092636, + 0.0071124606765806675, + 0.006870027165859938, + 0.03312620148062706, + 0.011282501742243767, + 0.0052251131273806095, + 0.0038230002392083406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.006761291529983282 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048660796135663986 + ], + "opposite_logprob_tokens": [ + -3.0486607551574707 + ], + "logprob_reward": -0.048660796135663986, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048660796135663986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09259310364723206, + 0.4837666153907776, + 0.05040967836976051, + 0.058609046041965485, + 0.09982386231422424, + 0.04163730517029762, + 0.02694321796298027, + 0.02530432865023613, + 0.03433715179562569, + 0.0865757167339325 + ], + "credit_map": [ + 0.007970640435814857, + 0.010129665024578571, + 0.0066865780390799046, + 0.008037111721932888, + 0.011672953143715858, + 0.012491540983319283, + 0.023438870906829834, + 0.01216574665158987, + 0.15834078192710876, + 0.032114796340465546, + 0.049750033766031265, + 0.019701723009347916, + 0.057756755501031876, + 0.11092214286327362, + 0.01571590267121792, + 0.03946450352668762, + 0.009453504346311092, + 0.008559152483940125, + 0.004599348176270723, + 0.004489688668400049, + 0.0062896981835365295, + 0.008516748435795307, + 0.004005869384855032, + 0.004495670553296804, + 0.015123891644179821, + 0.010236606933176517, + 0.006868343334645033, + 0.0038903341628611088, + 0.005982309114187956, + 0.00831501092761755, + 0.0044013806618750095, + 0.003791175317019224, + 0.017965279519557953, + 0.037894781678915024, + 0.005178286228328943, + 0.004107114393264055, + 0.011345945298671722, + 0.013339531607925892, + 0.004185859113931656, + 0.005807068198919296, + 0.008602631278336048, + 0.006894438061863184, + 0.0068279653787612915, + 0.00376254809089005, + 0.005171703640371561, + 0.0034185240510851145, + 0.0037684557028114796, + 0.003191041760146618, + 0.003414428560063243, + 0.004371264949440956, + 0.0026712226681411266, + 0.003825770691037178, + 0.0037652328610420227, + 0.0035790279507637024, + 0.002995892660692334, + 0.002320381347090006, + 0.0027495131362229586, + 0.0034592687152326107, + 0.0038884179666638374, + 0.0036674838047474623, + 0.00429321825504303, + 0.002506035380065441, + 0.0019737218972295523, + 0.002766669960692525, + 0.006734678987413645, + 0.005259055644273758, + 0.003159807063639164, + 0.00420193187892437, + 0.004989220295101404, + 0.003887756261974573, + 0.0021973149850964546, + 0.0039073871448636055, + 0.010612359270453453, + 0.008447086438536644, + 0.0055343396961688995, + 0.004609028343111277, + 0.0366191528737545, + 0.01515424158424139, + 0.0033510990906506777, + 0.002248415257781744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.048660796135663986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776213206350803 + ], + "opposite_logprob_tokens": [ + -4.6347761154174805 + ], + "logprob_reward": -0.009776213206350803, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776213206350803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40900343656539917, + 0.18544656038284302, + 0.025235429406166077, + 0.03202115744352341, + 0.1837061643600464, + 0.02355959825217724, + 0.017232006415724754, + 0.012341305613517761, + 0.035512059926986694, + 0.0759422704577446 + ], + "credit_map": [ + 0.025922659784555435, + 0.019014200195670128, + 0.01637408137321472, + 0.01922774314880371, + 0.08983432501554489, + 0.16753651201725006, + 0.0474521778523922, + 0.023641759529709816, + 0.058730632066726685, + 0.013293049298226833, + 0.01952817104756832, + 0.007837619632482529, + 0.02299157902598381, + 0.042004141956567764, + 0.0058554974384605885, + 0.015205868519842625, + 0.003955683670938015, + 0.004095079842954874, + 0.002447791863232851, + 0.00259435479529202, + 0.003269667038694024, + 0.004394794814288616, + 0.0026559557300060987, + 0.0018221022328361869, + 0.0037430860102176666, + 0.00489429384469986, + 0.0023419747594743967, + 0.0030113549437373877, + 0.00415201997384429, + 0.008613567799329758, + 0.002596416976302862, + 0.0026684459298849106, + 0.030361993238329887, + 0.06840843707323074, + 0.009940467774868011, + 0.012291239574551582, + 0.02040918916463852, + 0.023205464705824852, + 0.008158637210726738, + 0.01093073096126318, + 0.003001619828864932, + 0.0037148138508200645, + 0.002653274917975068, + 0.0033334195613861084, + 0.0026812178548425436, + 0.0020442840177565813, + 0.0037152806762605906, + 0.0024156877771019936, + 0.0019644538406282663, + 0.005180745851248503, + 0.0016566670965403318, + 0.0012662034714594483, + 0.0023132823407649994, + 0.0025196040514856577, + 0.0013157190987840295, + 0.0010153308976441622, + 0.0017447028076276183, + 0.0025795488618314266, + 0.0013766450574621558, + 0.0011801498476415873, + 0.001848988002166152, + 0.0016931419959291816, + 0.000986739294603467, + 0.0009313899208791554, + 0.00297345407307148, + 0.006069611292332411, + 0.003677781205624342, + 0.004424284212291241, + 0.007624280173331499, + 0.003781462786719203, + 0.0044717323035001755, + 0.002489451551809907, + 0.00856937374919653, + 0.018538936972618103, + 0.005401293747127056, + 0.006199994124472141, + 0.020916419103741646, + 0.008321958594024181, + 0.005348165053874254, + 0.0026461330708116293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.009776213206350803 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08901350945234299 + ], + "opposite_logprob_tokens": [ + -2.4640135765075684 + ], + "logprob_reward": -0.08901350945234299, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08901350945234299, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16055431962013245, + 0.17584866285324097, + 0.022136788815259933, + 0.027065228670835495, + 0.533074140548706, + 0.008497820235788822, + 0.009114349260926247, + 0.017251236364245415, + 0.00736160296946764, + 0.03909577801823616 + ], + "credit_map": [ + 0.01011977344751358, + 0.007800218649208546, + 0.006217002868652344, + 0.008577900938689709, + 0.0351724848151207, + 0.06437961012125015, + 0.01856151968240738, + 0.009725822135806084, + 0.05626337602734566, + 0.011582900770008564, + 0.017847223207354546, + 0.007173366378992796, + 0.021871423348784447, + 0.04018938168883324, + 0.00602493854239583, + 0.014896074309945107, + 0.0026970631442964077, + 0.00321793626062572, + 0.0029259570874273777, + 0.002586327027529478, + 0.002608610549941659, + 0.003962565213441849, + 0.0024713550228625536, + 0.0016669763717800379, + 0.002994938986375928, + 0.003915177658200264, + 0.0030365590937435627, + 0.0018082932801917195, + 0.003595370799303055, + 0.007850036956369877, + 0.0017551751807332039, + 0.0021096784621477127, + 0.0779026597738266, + 0.22980545461177826, + 0.021077554672956467, + 0.019045241177082062, + 0.06689370423555374, + 0.08129970729351044, + 0.01513606496155262, + 0.021913805976510048, + 0.001972772181034088, + 0.000989602180197835, + 0.0010398437734693289, + 0.0013399808667600155, + 0.0007986833225004375, + 0.0007486875983886421, + 0.0009484911570325494, + 0.0006597595638595521, + 0.0021535472478717566, + 0.0014180082362145185, + 0.0009893459500744939, + 0.0010818705195561051, + 0.0013431049883365631, + 0.0008979786653071642, + 0.0006728159496560693, + 0.0005576781113632023, + 0.0032578622922301292, + 0.004567102529108524, + 0.001809158711694181, + 0.0015589003451168537, + 0.0017396635375916958, + 0.002063891151919961, + 0.0012546912766993046, + 0.0009999671019613743, + 0.0010223084827885032, + 0.0009275127085857093, + 0.0008769562118686736, + 0.0010604987619444728, + 0.0008647868526168168, + 0.0010318141430616379, + 0.0008943593129515648, + 0.0006833670195192099, + 0.005222111940383911, + 0.005228884983807802, + 0.0013635437935590744, + 0.001757478341460228, + 0.016730165109038353, + 0.006031636614352465, + 0.0016002069460228086, + 0.001161751919426024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.08901350945234299 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly detailed documentary medium shot, shot from a level angle, slowly and steadily moves forward, gradually deepening into a family dining room filled with laughter and chatter. In the center of a dark, round wooden table, a stainless steel hot pot bubbling, half filled with a vibrant red spicy broth swirling with dried chilies and Sichuan peppercorns, the other half a milky white bone broth simmering with red dates and goji berries. Thick, white steam rises continuously from the pot, like a soft filter, imbuing the entire scene with a warm, fuzzy quality. A grandmother with graying hair and a kind smile carefully uses chopsticks to transfer a boiled fish ball to her granddaughter's bowl. The little girl, with pigtails and a pink sweater, gazes expectantly at the food in her bowl. Across from them, a middle-aged father in a dark blue V-neck sweater chats with his wife, a bright smile on his face. His mother, in a beige turtleneck, responds gently, expertly picking up a piece of green vegetable with her chopsticks and preparing to add it to the pot. The table was covered with a variety of ingredients: neatly arranged beef rolls, emerald greens, crisp white tofu slices, and fresh mushrooms. In front of each person was a bowl of prepared dipping sauce. A warm yellow chandelier hung above the table provided the sole source of light, its glow illuminating the steaming air and framing the happy expressions on everyone's faces, creating a warm, lively, and vibrant atmosphere for a family dinner.", + "seed": 18412, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_0/index_2055_seed_18412_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09753855485469103, + "vlm_logprob": -0.09753855485469103, + "vlm_yes_no_margin": 3.074999988079071, + "vlm_token_credit": -0.09753855485469103, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07293763011693954 + ], + "opposite_logprob_tokens": [ + -2.9479377269744873 + ], + "logprob_reward": -0.07293763011693954, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07293763011693954, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4150737524032593, + 0.05117695406079292, + 0.05419354513287544, + 0.022550886496901512, + 0.13514062762260437, + 0.058478403836488724, + 0.027273859828710556, + 0.13967293500900269, + 0.05875770002603531, + 0.03768144175410271 + ], + "credit_map": [ + 0.036962687969207764, + 0.07149383425712585, + 0.04455485939979553, + 0.008936609141528606, + 0.14355765283107758, + 0.022895382717251778, + 0.015556368976831436, + 0.07111626863479614, + 0.014955241233110428, + 0.007287176791578531, + 0.005548298358917236, + 0.004077116027474403, + 0.00890390481799841, + 0.005116519518196583, + 0.0025375301484018564, + 0.002751154825091362, + 0.010237336158752441, + 0.011238284409046173, + 0.003799617290496826, + 0.007791101932525635, + 0.00500941276550293, + 0.005476315040141344, + 0.007188800722360611, + 0.0034526644740253687, + 0.004606977570801973, + 0.004568289965391159, + 0.0026914123445749283, + 0.0023216481786221266, + 0.0025407446082681417, + 0.0023482840042561293, + 0.001908418838866055, + 0.0015651059802621603, + 0.0259409137070179, + 0.04384998977184296, + 0.008609776385128498, + 0.0073644304648041725, + 0.03083735518157482, + 0.007128506898880005, + 0.005766537971794605, + 0.005643086042255163, + 0.009752169251441956, + 0.010114840231835842, + 0.0062300050631165504, + 0.005060199648141861, + 0.008190786466002464, + 0.004875210579484701, + 0.00431007519364357, + 0.009945103898644447, + 0.004575046710669994, + 0.008144517429172993, + 0.001901663257740438, + 0.002553358441218734, + 0.0039329370483756065, + 0.002605065004900098, + 0.0020415703766047955, + 0.0015196953900158405, + 0.025862708687782288, + 0.019087698310613632, + 0.018421249464154243, + 0.015217036940157413, + 0.019075676798820496, + 0.01670960709452629, + 0.01623128168284893, + 0.009067638777196407, + 0.0118493577465415, + 0.008193222805857658, + 0.006548197008669376, + 0.010969433933496475, + 0.005648072808980942, + 0.00521698547527194, + 0.006459277123212814, + 0.0038731375243514776, + 0.0072876326739788055, + 0.004587720613926649, + 0.004318594932556152, + 0.007267570123076439, + 0.004976844415068626, + 0.003326946636661887, + 0.0038106879219412804, + 0.002105437684804201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.07293763011693954 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012561723589897156 + ], + "opposite_logprob_tokens": [ + -4.387561798095703 + ], + "logprob_reward": -0.012561723589897156, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012561723589897156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34060797095298767, + 0.07346922904253006, + 0.11912268400192261, + 0.03197286278009415, + 0.15919770300388336, + 0.04173155501484871, + 0.06330718845129013, + 0.08380872756242752, + 0.052570026367902756, + 0.03421204909682274 + ], + "credit_map": [ + 0.0395989827811718, + 0.05646046623587608, + 0.036607448011636734, + 0.01738944835960865, + 0.10379710048437119, + 0.022831398993730545, + 0.01591392606496811, + 0.04800920560956001, + 0.019422734156250954, + 0.01008216105401516, + 0.009057705290615559, + 0.005278953351080418, + 0.00920818280428648, + 0.008670436218380928, + 0.0049155764281749725, + 0.006833487655967474, + 0.01809694617986679, + 0.02023468166589737, + 0.011396222747862339, + 0.014903570525348186, + 0.010296660475432873, + 0.015969324856996536, + 0.019770847633481026, + 0.008454423397779465, + 0.004208728205412626, + 0.00512481015175581, + 0.00360991177149117, + 0.004271927289664745, + 0.003890827065333724, + 0.004964083898812532, + 0.003760572988539934, + 0.0021420044358819723, + 0.029453473165631294, + 0.05400650203227997, + 0.010193491354584694, + 0.009007236920297146, + 0.0351991280913353, + 0.008034906350076199, + 0.007802383042871952, + 0.005500570870935917, + 0.00774120120331645, + 0.007028201129287481, + 0.006135167554020882, + 0.00403499836102128, + 0.005339193157851696, + 0.004582156892865896, + 0.0039109839126467705, + 0.002959650941193104, + 0.01171835232526064, + 0.014219786040484905, + 0.00786677747964859, + 0.005587080493569374, + 0.008132626302540302, + 0.00479656970128417, + 0.005282877944409847, + 0.0057031153701245785, + 0.014018656685948372, + 0.013985429890453815, + 0.011422151699662209, + 0.009109380654990673, + 0.011445365846157074, + 0.00900590792298317, + 0.008392265066504478, + 0.006429566536098719, + 0.008761468343436718, + 0.007856806740164757, + 0.006051164120435715, + 0.007811809424310923, + 0.006785518955439329, + 0.0043974583968520164, + 0.006840047426521778, + 0.004065751563757658, + 0.004792619030922651, + 0.0036872727796435356, + 0.005456984508782625, + 0.004525511525571346, + 0.004748614504933357, + 0.0037054158747196198, + 0.005098940338939428, + 0.002196691930294037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.012561723589897156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897886633872986 + ], + "opposite_logprob_tokens": [ + -2.5789787769317627 + ], + "logprob_reward": -0.07897886633872986, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897886633872986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21602536737918854, + 0.13131265342235565, + 0.06700114160776138, + 0.044497329741716385, + 0.059492327272892, + 0.096983402967453, + 0.037284690886735916, + 0.0831494852900505, + 0.06088891252875328, + 0.20336468517780304 + ], + "credit_map": [ + 0.03217438980937004, + 0.03187215328216553, + 0.030071815475821495, + 0.026074109598994255, + 0.02388518862426281, + 0.03133923187851906, + 0.021043315529823303, + 0.0195651613175869, + 0.019558176398277283, + 0.018203653395175934, + 0.0156186418607831, + 0.014621014706790447, + 0.020768219605088234, + 0.020679304376244545, + 0.010608562268316746, + 0.011255082674324512, + 0.008789989165961742, + 0.009881221689283848, + 0.007996102795004845, + 0.005985653959214687, + 0.005938922055065632, + 0.011502501554787159, + 0.010678625665605068, + 0.006228118669241667, + 0.006078171078115702, + 0.005863087717443705, + 0.007925102487206459, + 0.0047130645252764225, + 0.004588960204273462, + 0.007179005537182093, + 0.004827028140425682, + 0.003322913311421871, + 0.010068228468298912, + 0.012902275659143925, + 0.005757552571594715, + 0.004176821559667587, + 0.010272505693137646, + 0.007198522333055735, + 0.005586361512541771, + 0.0035300576128065586, + 0.013445453718304634, + 0.013595298863947392, + 0.011223849840462208, + 0.009809838607907295, + 0.01447991281747818, + 0.011028267443180084, + 0.007896844297647476, + 0.01550393458455801, + 0.005037863273173571, + 0.005855703726410866, + 0.00453977519646287, + 0.0035868033301085234, + 0.006638188846409321, + 0.005035997834056616, + 0.003655578475445509, + 0.0029347797390073538, + 0.015734555199742317, + 0.012122792191803455, + 0.011201904155313969, + 0.009156009182333946, + 0.010092828422784805, + 0.008927383460104465, + 0.00991777889430523, + 0.005996237508952618, + 0.009115380235016346, + 0.00931251049041748, + 0.00683938805013895, + 0.00910775177180767, + 0.008988648653030396, + 0.006231465842574835, + 0.007890695706009865, + 0.0034030757378786802, + 0.025821059942245483, + 0.023651650175452232, + 0.042249325662851334, + 0.028818760067224503, + 0.01755695603787899, + 0.021463816985487938, + 0.03057331219315529, + 0.013229826465249062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.07897886633872986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977774802595377 + ], + "opposite_logprob_tokens": [ + -4.634777545928955 + ], + "logprob_reward": -0.00977774802595377, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977774802595377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3508991003036499, + 0.11170537024736404, + 0.041140176355838776, + 0.025627946481108665, + 0.034701455384492874, + 0.04626006260514259, + 0.023144260048866272, + 0.28324952721595764, + 0.06201007962226868, + 0.021262025460600853 + ], + "credit_map": [ + 0.038508202880620956, + 0.05656905099749565, + 0.0375884510576725, + 0.010495796799659729, + 0.11632397770881653, + 0.023259684443473816, + 0.0144829535856843, + 0.0536709688603878, + 0.03815530240535736, + 0.01409811619669199, + 0.01231513824313879, + 0.004870497155934572, + 0.027454867959022522, + 0.005274154245853424, + 0.004723130725324154, + 0.004814161919057369, + 0.006220505107194185, + 0.005866801366209984, + 0.005730783566832542, + 0.004396488890051842, + 0.004298028536140919, + 0.004962102044373751, + 0.006473235320299864, + 0.003192228963598609, + 0.004240483045578003, + 0.004330729134380817, + 0.0022457579616457224, + 0.00333020300604403, + 0.0029260246083140373, + 0.0031096478924155235, + 0.003266220213845372, + 0.0021788824815303087, + 0.004197558853775263, + 0.008859919384121895, + 0.004196214955300093, + 0.003175129648298025, + 0.0055278572253882885, + 0.0031432537361979485, + 0.0028387331403791904, + 0.0027627889066934586, + 0.007223833817988634, + 0.008221829310059547, + 0.007178930565714836, + 0.00422449829056859, + 0.005235899705439806, + 0.0029465926345437765, + 0.0040370626375079155, + 0.007191414479166269, + 0.003715437836945057, + 0.0030645988881587982, + 0.0033422340638935566, + 0.0036501772701740265, + 0.0038860884960740805, + 0.0017502362607046962, + 0.0018569539533928037, + 0.001878532930277288, + 0.05006786808371544, + 0.03808942809700966, + 0.041448988020420074, + 0.035349033772945404, + 0.035562485456466675, + 0.029738666489720345, + 0.03350777179002762, + 0.019485285505652428, + 0.01081005297601223, + 0.00890178233385086, + 0.007293988019227982, + 0.009993002749979496, + 0.007368129678070545, + 0.004920387640595436, + 0.008063950575888157, + 0.004658784717321396, + 0.003866306273266673, + 0.0024609474930912256, + 0.002307639690116048, + 0.002864058595150709, + 0.0026455565821379423, + 0.0027725209947675467, + 0.002946805441752076, + 0.001398189808242023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.00977774802595377 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134368062019348 + ], + "opposite_logprob_tokens": [ + -1.31343674659729 + ], + "logprob_reward": -0.3134368062019348, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.3134368062019348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22071924805641174, + 0.05177769064903259, + 0.23219773173332214, + 0.09559037536382675, + 0.22576504945755005, + 0.030447831377387047, + 0.016225047409534454, + 0.08623591810464859, + 0.028056398034095764, + 0.012984723784029484 + ], + "credit_map": [ + 0.03294997662305832, + 0.034257154911756516, + 0.02774745039641857, + 0.0112868407741189, + 0.06225549802184105, + 0.014405708760023117, + 0.012858674861490726, + 0.024957941845059395, + 0.007161527406424284, + 0.006814718712121248, + 0.005420925095677376, + 0.004454261157661676, + 0.01290837861597538, + 0.0046063619665801525, + 0.0043313829228281975, + 0.006080132909119129, + 0.03832343593239784, + 0.05420494079589844, + 0.0180138498544693, + 0.027176253497600555, + 0.024929139763116837, + 0.024293655529618263, + 0.030555203557014465, + 0.014701261185109615, + 0.017328936606645584, + 0.019926918670535088, + 0.014682128094136715, + 0.006654394790530205, + 0.01048166025429964, + 0.014164043590426445, + 0.007056097500026226, + 0.005296194460242987, + 0.04409492015838623, + 0.08065670728683472, + 0.013268561102449894, + 0.012217214331030846, + 0.046424925327301025, + 0.010356437414884567, + 0.008907325565814972, + 0.00983893871307373, + 0.004561229608952999, + 0.0035084066912531853, + 0.006453828886151314, + 0.0032615705858916044, + 0.003657704219222069, + 0.002802587114274502, + 0.002789023332297802, + 0.003413477446883917, + 0.0026036619674414396, + 0.0026793351862579584, + 0.001645672251470387, + 0.0012368389870971441, + 0.0028896238654851913, + 0.002645484171807766, + 0.0014092887286096811, + 0.0011151416692882776, + 0.01689108833670616, + 0.010782039724290371, + 0.010806706734001637, + 0.009437764063477516, + 0.012590928003191948, + 0.010224212892353535, + 0.009833221323788166, + 0.005669962614774704, + 0.005330157466232777, + 0.003909450490027666, + 0.0026640379801392555, + 0.004474501125514507, + 0.0033133020624518394, + 0.003360350616276264, + 0.00300477328710258, + 0.0019998238421976566, + 0.0023964783176779747, + 0.0016567736165598035, + 0.0012722959509119391, + 0.0015732607571408153, + 0.0021223051007837057, + 0.0018727524438872933, + 0.0012355181388556957, + 0.000855339749250561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.3134368062019348 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly detailed documentary medium shot, shot from a level angle, slowly and steadily moves forward, gradually deepening into a family dining room filled with laughter and chatter. In the center of a dark, round wooden table, a stainless steel hot pot bubbling, half filled with a vibrant red spicy broth swirling with dried chilies and Sichuan peppercorns, the other half a milky white bone broth simmering with red dates and goji berries. Thick, white steam rises continuously from the pot, like a soft filter, imbuing the entire scene with a warm, fuzzy quality. A grandmother with graying hair and a kind smile carefully uses chopsticks to transfer a boiled fish ball to her granddaughter's bowl. The little girl, with pigtails and a pink sweater, gazes expectantly at the food in her bowl. Across from them, a middle-aged father in a dark blue V-neck sweater chats with his wife, a bright smile on his face. His mother, in a beige turtleneck, responds gently, expertly picking up a piece of green vegetable with her chopsticks and preparing to add it to the pot. The table was covered with a variety of ingredients: neatly arranged beef rolls, emerald greens, crisp white tofu slices, and fresh mushrooms. In front of each person was a bowl of prepared dipping sauce. A warm yellow chandelier hung above the table provided the sole source of light, its glow illuminating the steaming air and framing the happy expressions on everyone's faces, creating a warm, lively, and vibrant atmosphere for a family dinner.", + "seed": 18413, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_0/index_2055_seed_18413_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05072403401136398, + "vlm_logprob": -0.05072403401136398, + "vlm_yes_no_margin": 3.600000023841858, + "vlm_token_credit": -0.05072403401136398, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04376630112528801 + ], + "opposite_logprob_tokens": [ + -3.7937662601470947 + ], + "logprob_reward": -0.04376630112528801, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.04376630112528801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1655929833650589, + 0.07688351720571518, + 0.051699455827474594, + 0.06699033081531525, + 0.09200963377952576, + 0.14754319190979004, + 0.23920351266860962, + 0.05258313566446304, + 0.02868344821035862, + 0.07881072908639908 + ], + "credit_map": [ + 0.026438841596245766, + 0.014353829436004162, + 0.02528144046664238, + 0.011278602294623852, + 0.02177826687693596, + 0.040733080357313156, + 0.012654487043619156, + 0.01307443343102932, + 0.012894637882709503, + 0.010015593841671944, + 0.01546423975378275, + 0.005831045098602772, + 0.01093317661434412, + 0.009165742434561253, + 0.006632652133703232, + 0.0059464252553880215, + 0.004804673604667187, + 0.008187186904251575, + 0.006832915358245373, + 0.006467114668339491, + 0.006516461726278067, + 0.005613492801785469, + 0.007174490485340357, + 0.006103115156292915, + 0.006734911818057299, + 0.008413662202656269, + 0.006382383406162262, + 0.006595919374376535, + 0.007364625111222267, + 0.011065702885389328, + 0.011786229908466339, + 0.008646892383694649, + 0.00924144871532917, + 0.0051653338596224785, + 0.0063473982736468315, + 0.01902187056839466, + 0.031355120241642, + 0.00783590879291296, + 0.0062199668027460575, + 0.006822588853538036, + 0.014368805103003979, + 0.019265446811914444, + 0.02631727233529091, + 0.020851150155067444, + 0.02927597612142563, + 0.011657988652586937, + 0.015304485335946083, + 0.010502072051167488, + 0.018697159364819527, + 0.029860451817512512, + 0.019525019451975822, + 0.055593546479940414, + 0.04774988815188408, + 0.013809191063046455, + 0.03712007403373718, + 0.016848187893629074, + 0.005118162836879492, + 0.007598699536174536, + 0.005273694638162851, + 0.012438051402568817, + 0.005546849686652422, + 0.00353619153611362, + 0.007919857278466225, + 0.0051516275852918625, + 0.002887028967961669, + 0.003487433772534132, + 0.0022942223586142063, + 0.005226870998740196, + 0.004780011251568794, + 0.0028069799300283194, + 0.0037043746560811996, + 0.003496523480862379, + 0.008326043374836445, + 0.014158078469336033, + 0.007698650471866131, + 0.006741129327565432, + 0.01773134060204029, + 0.006343938875943422, + 0.010707342997193336, + 0.007104202173650265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.04376630112528801 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009797816164791584 + ], + "opposite_logprob_tokens": [ + -4.634798049926758 + ], + "logprob_reward": -0.009797816164791584, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009797816164791584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16220985352993011, + 0.06416872143745422, + 0.038364604115486145, + 0.016922449693083763, + 0.06609681248664856, + 0.048739124089479446, + 0.0756523460149765, + 0.0824524387717247, + 0.06015656143426895, + 0.38523709774017334 + ], + "credit_map": [ + 0.029004648327827454, + 0.0214480459690094, + 0.013962565921247005, + 0.010382943786680698, + 0.025179650634527206, + 0.04124977067112923, + 0.00998492818325758, + 0.010997300036251545, + 0.006526723969727755, + 0.009433611296117306, + 0.006833200808614492, + 0.0053214081563055515, + 0.013571268878877163, + 0.009372546337544918, + 0.007790220435708761, + 0.0053197420202195644, + 0.0047446973621845245, + 0.0036374933551996946, + 0.005173767451196909, + 0.0029763076454401016, + 0.004741029813885689, + 0.005446572322398424, + 0.006280943285673857, + 0.005363794509321451, + 0.003171206684783101, + 0.00181767123285681, + 0.002493920736014843, + 0.0014273974811658263, + 0.0020858978386968374, + 0.0024414402432739735, + 0.0016253601061180234, + 0.0018595547880977392, + 0.009265978820621967, + 0.004396618343889713, + 0.004831762984395027, + 0.010670133866369724, + 0.022859901189804077, + 0.005766104906797409, + 0.0036169663071632385, + 0.004689344670623541, + 0.003924388438463211, + 0.004519870039075613, + 0.011411205865442753, + 0.008287214674055576, + 0.0077032120898365974, + 0.0038392250426113605, + 0.0057560959830880165, + 0.0032979112584143877, + 0.005892104934900999, + 0.008726767264306545, + 0.006826603785157204, + 0.017392735928297043, + 0.014258071780204773, + 0.004114828538149595, + 0.012549183331429958, + 0.005892047192901373, + 0.00749715743586421, + 0.009630071930587292, + 0.012879127636551857, + 0.01566203311085701, + 0.008771468885242939, + 0.0062388586811721325, + 0.01416905876249075, + 0.007604660000652075, + 0.006311256438493729, + 0.006836562417447567, + 0.005978723987936974, + 0.008330033160746098, + 0.009573540650308132, + 0.004195396322757006, + 0.007382849231362343, + 0.011548195965588093, + 0.04588218033313751, + 0.039616383612155914, + 0.08979911357164383, + 0.02825402468442917, + 0.061318594962358475, + 0.034908343106508255, + 0.06411982327699661, + 0.02133861742913723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.009797816164791584 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864705726504326 + ], + "opposite_logprob_tokens": [ + -3.048647165298462 + ], + "logprob_reward": -0.04864705726504326, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864705726504326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08116146922111511, + 0.14324137568473816, + 0.0656040608882904, + 0.06978970021009445, + 0.07828277349472046, + 0.07737373560667038, + 0.08324957638978958, + 0.14867255091667175, + 0.12257351726293564, + 0.1300511509180069 + ], + "credit_map": [ + 0.014381465502083302, + 0.008497479371726513, + 0.011919121257960796, + 0.006807034369558096, + 0.009783387184143066, + 0.013034443370997906, + 0.006326823029667139, + 0.01041172631084919, + 0.022944379597902298, + 0.016194075345993042, + 0.026730738580226898, + 0.014776446856558323, + 0.018489334732294083, + 0.01601393334567547, + 0.018179049715399742, + 0.009913432411849499, + 0.007130639627575874, + 0.008502365089952946, + 0.00964890606701374, + 0.008573087863624096, + 0.008126890286803246, + 0.007440634537488222, + 0.009982801042497158, + 0.006198746617883444, + 0.007594493683427572, + 0.006502723321318626, + 0.00973525457084179, + 0.00588735518977046, + 0.007584304548799992, + 0.010480745695531368, + 0.015941159799695015, + 0.0060636731795966625, + 0.017136767506599426, + 0.009318229742348194, + 0.009958784095942974, + 0.006553099025040865, + 0.011180558241903782, + 0.008876841515302658, + 0.005849621258676052, + 0.009408880956470966, + 0.006261568050831556, + 0.006328181829303503, + 0.019563764333724976, + 0.009084093384444714, + 0.01205441728234291, + 0.007796043995767832, + 0.010134772397577763, + 0.006150899920612574, + 0.006515997461974621, + 0.010984998196363449, + 0.006799246184527874, + 0.010100914165377617, + 0.012513892725110054, + 0.006157265976071358, + 0.02048552967607975, + 0.00969174038618803, + 0.012052226811647415, + 0.01565987803041935, + 0.020889077335596085, + 0.0223833154886961, + 0.016986573114991188, + 0.01097667682915926, + 0.03652641549706459, + 0.013198403641581535, + 0.012256311252713203, + 0.019880035892128944, + 0.0091170072555542, + 0.018963845446705818, + 0.026100564748048782, + 0.01087876595556736, + 0.013456999324262142, + 0.011919990181922913, + 0.014942667447030544, + 0.012313203886151314, + 0.033963944762945175, + 0.010525450110435486, + 0.016084788367152214, + 0.011643952690064907, + 0.02129903994500637, + 0.009278110228478909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 9 + ], + "selected_reward": -0.04864705726504326 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008633549325168133 + ], + "opposite_logprob_tokens": [ + -4.758633613586426 + ], + "logprob_reward": -0.008633549325168133, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008633549325168133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27082929015159607, + 0.03591086342930794, + 0.03583332151174545, + 0.025484487414360046, + 0.04665548726916313, + 0.038949236273765564, + 0.05990878865122795, + 0.0523260161280632, + 0.09943252801895142, + 0.3346700966358185 + ], + "credit_map": [ + 0.06050075963139534, + 0.025688255205750465, + 0.04702115058898926, + 0.01807120256125927, + 0.027633029967546463, + 0.05038946121931076, + 0.020777765661478043, + 0.020747611299157143, + 0.003980020061135292, + 0.004637021571397781, + 0.006420022342354059, + 0.003251772839576006, + 0.006178571842610836, + 0.004880961030721664, + 0.0036001878324896097, + 0.002962301019579172, + 0.004407424945384264, + 0.00418009702116251, + 0.0051825507543981075, + 0.002704909536987543, + 0.005223499611020088, + 0.005434699822217226, + 0.004549899138510227, + 0.00415023835375905, + 0.003990312106907368, + 0.003897874616086483, + 0.0025614402256906033, + 0.0017266330542042851, + 0.0035505902487784624, + 0.0037731335032731295, + 0.003064850578084588, + 0.0029196483083069324, + 0.005027162842452526, + 0.0026097234804183245, + 0.0030932393856346607, + 0.007754754740744829, + 0.01684412732720375, + 0.00372141576372087, + 0.003099446650594473, + 0.004505614284425974, + 0.0032751543913036585, + 0.004411840811371803, + 0.0052014607936143875, + 0.005269618704915047, + 0.005299342330545187, + 0.004817818757146597, + 0.005634985864162445, + 0.005039010662585497, + 0.004400609992444515, + 0.0055110822431743145, + 0.006215064320713282, + 0.008493160828948021, + 0.00914399791508913, + 0.0035262571182101965, + 0.016735613346099854, + 0.0058829933404922485, + 0.007019333075731993, + 0.006752749904990196, + 0.00523769948631525, + 0.012255992740392685, + 0.006441565230488777, + 0.0048490348272025585, + 0.0053956154733896255, + 0.004374020267277956, + 0.008927632123231888, + 0.009005645290017128, + 0.009225633926689625, + 0.011323097161948681, + 0.010232399217784405, + 0.009230357594788074, + 0.010842871852219105, + 0.03064488247036934, + 0.041241392493247986, + 0.03498300909996033, + 0.07623626291751862, + 0.02447071112692356, + 0.05352121591567993, + 0.02898682840168476, + 0.05440680682659149, + 0.020823810249567032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.008633549325168133 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot taken from a level angle that slowly moves forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a dark round wooden table with a stainless steel hot pot containing half red spicy broth and half milky white bone broth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a grandmother with graying hair using chopsticks to transfer a boiled fish ball to a little girl with pigtails wearing a pink sweater?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are there visible ingredients including beef rolls, emerald greens, white tofu slices, mushrooms, and bowls of dipping sauce on the table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a warm yellow chandelier hanging above the table?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277544617652893 + ], + "opposite_logprob_tokens": [ + -2.017775535583496 + ], + "logprob_reward": -0.14277544617652893, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14277544617652893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6374615430831909, + 0.10713538527488708, + 0.019917918369174004, + 0.019247382879257202, + 0.020549990236759186, + 0.038137584924697876, + 0.04837898910045624, + 0.03550955280661583, + 0.03156261891126633, + 0.042099062353372574 + ], + "credit_map": [ + 0.1727467179298401, + 0.0665716677904129, + 0.06646663695573807, + 0.024438444525003433, + 0.07687976956367493, + 0.10057899355888367, + 0.044959843158721924, + 0.08481939882040024, + 0.008674667216837406, + 0.016942286863923073, + 0.00842783972620964, + 0.007878171280026436, + 0.0321093387901783, + 0.013052579015493393, + 0.010451515205204487, + 0.009598996490240097, + 0.00233655096963048, + 0.0019506494281813502, + 0.0027351207099854946, + 0.0017579715931788087, + 0.002852432895451784, + 0.0024089848157018423, + 0.0033085988834500313, + 0.0025676090735942125, + 0.001967421267181635, + 0.0022950496058911085, + 0.002608111361041665, + 0.0021683997474610806, + 0.002376073971390724, + 0.0030638317111879587, + 0.002997312229126692, + 0.0017711828695610166, + 0.001968835946172476, + 0.002752438187599182, + 0.0019427560037001967, + 0.002426381688565016, + 0.0051649585366249084, + 0.0019035787554457784, + 0.001927029574289918, + 0.002464012475684285, + 0.0025266623124480247, + 0.004580693319439888, + 0.008088081143796444, + 0.004792718216776848, + 0.007290228735655546, + 0.003847267013043165, + 0.004438220523297787, + 0.002573713194578886, + 0.004432820249348879, + 0.006101036909967661, + 0.004227872937917709, + 0.012571008875966072, + 0.008318864740431309, + 0.0026371118146926165, + 0.006529863923788071, + 0.0035604119766503572, + 0.004081939812749624, + 0.004478324670344591, + 0.0031422674655914307, + 0.007219529710710049, + 0.0030381265096366405, + 0.0031752658542245626, + 0.0050564175471663475, + 0.005317682400345802, + 0.0024801520630717278, + 0.004713813774287701, + 0.0029619927518069744, + 0.0041783456690609455, + 0.006605198606848717, + 0.0019263331778347492, + 0.003193838754668832, + 0.00550294341519475, + 0.005520273931324482, + 0.005798882804811001, + 0.004682903178036213, + 0.0037984317168593407, + 0.005445980001240969, + 0.004552572965621948, + 0.0075798192992806435, + 0.004720199387520552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.14277544617652893 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_1/index_tensor([451, 451])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_1/index_tensor([451, 451])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cd39ce2a2ca62a353655711b8dfa2ce71986d923 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_1/index_tensor([451, 451])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium shot tracks a powerful four-stroke snowmobile, the camera shaking violently to simulate intense vibrations as the machine accelerates explosively across pristine snow. The entire scene is rendered in a detailed, artistic style reminiscent of traditional oil painting, with all surfaces possessing visible, thick, mural-like textures. The snowmobile, painted a deep red that contrasts sharply with its surroundings, charges forward aggressively, its rider, clad in a black thermal suit, leaning into a turn, gripping the chrome handlebars. Bright, painterly sunlight gleams from the handlebars and the rider's fully reflective visor, reflecting a distorted, vibrant image of the landscape. As the machine's tracks dig deep into the icy terrain, it kicks up a massive, billowing plume of fresh powder—a chaotic explosion of white and crystalline particles that catch the light. In the background, a majestic panorama of jagged, snow-capped peaks stands starkly against a clear, deep blue sky, creating an aura of thrilling power and sublime, rugged beauty.", + "seed": 11836, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_1/index_451_seed_11836_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07157129887491465, + "vlm_logprob": -0.07157129887491465, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.07157129887491465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08086143434047699 + ], + "opposite_logprob_tokens": [ + -2.9558613300323486 + ], + "logprob_reward": -0.08086143434047699, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.08086143434047699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22058284282684326, + 0.1024702712893486, + 0.04093409702181816, + 0.0787571594119072, + 0.1886233389377594, + 0.07828424125909805, + 0.1527738720178604, + 0.04959335923194885, + 0.04167834669351578, + 0.04630248248577118 + ], + "credit_map": [ + 0.04736103489995003, + 0.02780713513493538, + 0.02698427252471447, + 0.030747218057513237, + 0.0231862124055624, + 0.02361406572163105, + 0.025903670117259026, + 0.01497923955321312, + 0.01006296370178461, + 0.017197607085108757, + 0.013461999595165253, + 0.01891089603304863, + 0.008035366423428059, + 0.016896996647119522, + 0.009682534262537956, + 0.008221902884542942, + 0.003976656123995781, + 0.009356300346553326, + 0.005718696396797895, + 0.006449909880757332, + 0.003999779932200909, + 0.004938806407153606, + 0.0038962545804679394, + 0.002597692422568798, + 0.009448738768696785, + 0.016025828197598457, + 0.012223748490214348, + 0.008730783127248287, + 0.01075747236609459, + 0.011179443448781967, + 0.006053570657968521, + 0.004337569698691368, + 0.027332760393619537, + 0.022591929882764816, + 0.023551560938358307, + 0.02885543555021286, + 0.01718948408961296, + 0.017137493938207626, + 0.0281929150223732, + 0.02377176657319069, + 0.00874002743512392, + 0.009568390436470509, + 0.01888369955122471, + 0.007388237398117781, + 0.0057890270836651325, + 0.009361873380839825, + 0.007496315520256758, + 0.011056669056415558, + 0.013577320612967014, + 0.01886613853275776, + 0.014628041535615921, + 0.015224841423332691, + 0.016982199624180794, + 0.032196082174777985, + 0.022346336394548416, + 0.018952907994389534, + 0.004953284282237291, + 0.006470168009400368, + 0.00598332891240716, + 0.005612449254840612, + 0.0047920700162649155, + 0.0077606006525456905, + 0.007189292926341295, + 0.006832164246588945, + 0.008805743418633938, + 0.00509086437523365, + 0.005071946885436773, + 0.0038101377431303263, + 0.004255114588886499, + 0.008177871815860271, + 0.0036406395956873894, + 0.0028260292019695044, + 0.007905537262558937, + 0.005022673401981592, + 0.005995665676891804, + 0.006223457865417004, + 0.0060244230553507805, + 0.006124148145318031, + 0.003721141954883933, + 0.005285434890538454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.08086143434047699 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990688979625702 + ], + "opposite_logprob_tokens": [ + -1.8049068450927734 + ], + "logprob_reward": -0.17990688979625702, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990688979625702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17464978992938995, + 0.09938371181488037, + 0.05378245934844017, + 0.29510796070098877, + 0.21255217492580414, + 0.04342443868517876, + 0.03745908662676811, + 0.0276949442923069, + 0.03113641031086445, + 0.024808980524539948 + ], + "credit_map": [ + 0.02608289197087288, + 0.02838551439344883, + 0.016713455319404602, + 0.026288693770766258, + 0.01839158684015274, + 0.023944631218910217, + 0.021768497303128242, + 0.013074520975351334, + 0.007655717898160219, + 0.016647255048155785, + 0.012180891819298267, + 0.018407229334115982, + 0.008836994878947735, + 0.01838098093867302, + 0.009371746331453323, + 0.00790291279554367, + 0.005050323437899351, + 0.007112284190952778, + 0.0072656054981052876, + 0.008558331057429314, + 0.006649411283433437, + 0.007638263516128063, + 0.006961182225495577, + 0.004547064192593098, + 0.043876513838768005, + 0.026412807404994965, + 0.0510198175907135, + 0.033207640051841736, + 0.020332127809524536, + 0.08578917384147644, + 0.01412108913064003, + 0.02034882642328739, + 0.028858259320259094, + 0.026015814393758774, + 0.026412757113575935, + 0.03543917089700699, + 0.01646464318037033, + 0.017302339896559715, + 0.035577286034822464, + 0.02648191899061203, + 0.005227127578109503, + 0.0049414304085075855, + 0.009857303462922573, + 0.003282008459791541, + 0.0035331966355443, + 0.006347973830997944, + 0.003783603198826313, + 0.006451800465583801, + 0.0028928860556334257, + 0.00445227837190032, + 0.0032554659992456436, + 0.003932476043701172, + 0.0032653736416250467, + 0.009370830841362476, + 0.005184654146432877, + 0.005105127114802599, + 0.0020683349575847387, + 0.003913123160600662, + 0.0031532146967947483, + 0.0039497725665569305, + 0.0023005008697509766, + 0.003448768984526396, + 0.0036952337250113487, + 0.005165998358279467, + 0.0043207998387515545, + 0.003987371455878019, + 0.0051896884106099606, + 0.002288431627675891, + 0.004569243174046278, + 0.007246800232678652, + 0.0016793763497844338, + 0.0018547028303146362, + 0.0027973975520581007, + 0.0032654404640197754, + 0.0025212126784026623, + 0.0037167086265981197, + 0.004779586102813482, + 0.003096537897363305, + 0.0019280891865491867, + 0.0027040105778723955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.17990688979625702 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052522532641887665 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.0052522532641887665, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052522532641887665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05210559070110321, + 0.07314503192901611, + 0.032794710248708725, + 0.09216631203889847, + 0.45635801553726196, + 0.10060160607099533, + 0.08690062165260315, + 0.040891047567129135, + 0.03728977590799332, + 0.02774730511009693 + ], + "credit_map": [ + 0.005751839838922024, + 0.006657876539975405, + 0.005661841481924057, + 0.007142902817577124, + 0.005865329410880804, + 0.008426527492702007, + 0.008973575197160244, + 0.003625693265348673, + 0.006438566837459803, + 0.010131759569048882, + 0.00829489715397358, + 0.011108004488050938, + 0.008601340465247631, + 0.014477226883172989, + 0.007905181497335434, + 0.006188046652823687, + 0.002284885151311755, + 0.003354906104505062, + 0.004505798686295748, + 0.004297696053981781, + 0.005183716770261526, + 0.005659247282892466, + 0.004792880732566118, + 0.0027155783027410507, + 0.013211828656494617, + 0.00808136910200119, + 0.016298359259963036, + 0.012402178719639778, + 0.005566684063524008, + 0.02604590728878975, + 0.004036722704768181, + 0.0065232496708631516, + 0.07047293335199356, + 0.05481121316552162, + 0.05675835162401199, + 0.06455890834331512, + 0.040041979402303696, + 0.042403560131788254, + 0.0723133459687233, + 0.054997678846120834, + 0.011814644560217857, + 0.007437099236994982, + 0.02644379250705242, + 0.009927758947014809, + 0.007561062928289175, + 0.012598331086337566, + 0.008897032588720322, + 0.015921873971819878, + 0.005766233894973993, + 0.009958268143236637, + 0.0048344931565225124, + 0.007990996353328228, + 0.011915997602045536, + 0.023459387943148613, + 0.010612204670906067, + 0.01236303523182869, + 0.003058129455894232, + 0.005157811567187309, + 0.004818357061594725, + 0.003492278978228569, + 0.003884174395352602, + 0.00566903967410326, + 0.004994956776499748, + 0.009816297329962254, + 0.006692629307508469, + 0.003231146140024066, + 0.007722756825387478, + 0.00198852620087564, + 0.0033597892615944147, + 0.009507722221314907, + 0.002892873017117381, + 0.001894330489449203, + 0.004957233089953661, + 0.002774889813736081, + 0.0034542048815637827, + 0.0033107963390648365, + 0.0037649760488420725, + 0.004321807064116001, + 0.002585234586149454, + 0.002578161656856537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.0052522532641887665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902201056480408 + ], + "opposite_logprob_tokens": [ + -2.464021921157837 + ], + "logprob_reward": -0.08902201056480408, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902201056480408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11626114696264267, + 0.06812621653079987, + 0.04532388970255852, + 0.13627363741397858, + 0.33728349208831787, + 0.13318780064582825, + 0.050731707364320755, + 0.01731775514781475, + 0.0673120766878128, + 0.02818233147263527 + ], + "credit_map": [ + 0.023652730509638786, + 0.012677783146500587, + 0.014543904922902584, + 0.01702672429382801, + 0.011205490678548813, + 0.01890086568892002, + 0.009914636611938477, + 0.00833901483565569, + 0.0053182183764874935, + 0.00827272143214941, + 0.01142133865505457, + 0.014316165819764137, + 0.0054018497467041016, + 0.008859029039740562, + 0.007764573674649, + 0.006772319320589304, + 0.003531381953507662, + 0.007748527452349663, + 0.006809836253523827, + 0.009065822698175907, + 0.003208864713087678, + 0.007596525363624096, + 0.004134477116167545, + 0.0032284539192914963, + 0.019023843109607697, + 0.01374958548694849, + 0.024158308282494545, + 0.016160577535629272, + 0.008096275851130486, + 0.04019146040081978, + 0.005340345203876495, + 0.00955323874950409, + 0.05915410444140434, + 0.0402192696928978, + 0.04363133758306503, + 0.03914084658026695, + 0.033963728696107864, + 0.03356329724192619, + 0.04650658741593361, + 0.04110429063439369, + 0.015729909762740135, + 0.012795984745025635, + 0.029056735336780548, + 0.011556671932339668, + 0.011618484742939472, + 0.021679289638996124, + 0.010760713368654251, + 0.019990017637610435, + 0.0035933980252593756, + 0.004108164459466934, + 0.003595085348933935, + 0.004427225794643164, + 0.006556199863553047, + 0.014869132079184055, + 0.005764254834502935, + 0.007818248122930527, + 0.0018636478343978524, + 0.002380198333412409, + 0.002275308594107628, + 0.002116748131811619, + 0.0020712031982839108, + 0.002447768347337842, + 0.002482197480276227, + 0.001680682529695332, + 0.01015401165932417, + 0.005709779914468527, + 0.016463439911603928, + 0.0041332729160785675, + 0.004677808843553066, + 0.020138351246714592, + 0.002859849249944091, + 0.0031755613163113594, + 0.004631366580724716, + 0.003410537028685212, + 0.004343112930655479, + 0.004344517830759287, + 0.0024915747344493866, + 0.004167042672634125, + 0.0024734167382121086, + 0.0023207624908536673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.08902201056480408 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028139064088463783 + ], + "opposite_logprob_tokens": [ + -5.877813816070557 + ], + "logprob_reward": -0.0028139064088463783, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028139064088463783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40054190158843994, + 0.12490231543779373, + 0.0416770838201046, + 0.08328939229249954, + 0.1282462328672409, + 0.039711300283670425, + 0.061688438057899475, + 0.022573744878172874, + 0.06887143105268478, + 0.028498077765107155 + ], + "credit_map": [ + 0.08532349020242691, + 0.09440621733665466, + 0.04545563831925392, + 0.05624973401427269, + 0.0324104018509388, + 0.03004790097475052, + 0.031580451875925064, + 0.025068076327443123, + 0.011154391802847385, + 0.019895408302545547, + 0.01541944406926632, + 0.02021554298698902, + 0.011982138268649578, + 0.020612312480807304, + 0.01451075542718172, + 0.011112330481410027, + 0.004199440125375986, + 0.007656923495233059, + 0.006423180922865868, + 0.008383616805076599, + 0.003225666703656316, + 0.004821326583623886, + 0.003642406314611435, + 0.003324525896459818, + 0.012311466038227081, + 0.00919131189584732, + 0.013924735598266125, + 0.010567297227680683, + 0.005348751321434975, + 0.022535420954227448, + 0.003956666216254234, + 0.005453753285109997, + 0.018949449062347412, + 0.01728609763085842, + 0.016357824206352234, + 0.018092025071382523, + 0.010579045861959457, + 0.011251603253185749, + 0.01977163553237915, + 0.01595856249332428, + 0.003985550720244646, + 0.0033931653015315533, + 0.008510549552738667, + 0.005041919182986021, + 0.0037721660919487476, + 0.004886118695139885, + 0.004205025732517242, + 0.0059168110601603985, + 0.006132262758910656, + 0.0069116088561713696, + 0.0062663620337843895, + 0.0055425893515348434, + 0.008016352541744709, + 0.013612979091703892, + 0.0077124182134866714, + 0.0074938698671758175, + 0.0026613834779709578, + 0.004862037021666765, + 0.003682065289467573, + 0.0018013275694102049, + 0.0023061060346663, + 0.0027670313138514757, + 0.0019667602609843016, + 0.0025270353071391582, + 0.00742778554558754, + 0.006780486553907394, + 0.018684975802898407, + 0.004285972565412521, + 0.00668251421302557, + 0.01849348656833172, + 0.003376852022483945, + 0.003139359876513481, + 0.004626310430467129, + 0.0038218230474740267, + 0.0025713853538036346, + 0.004925816785544157, + 0.004682423546910286, + 0.003827136941254139, + 0.001976169878616929, + 0.0020670148078352213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0028139064088463783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot tracks a powerful four-stroke snowmobile, the camera shaking violently to simulate intense vibrations as the machine accelerates explosively across pristine snow. The entire scene is rendered in a detailed, artistic style reminiscent of traditional oil painting, with all surfaces possessing visible, thick, mural-like textures. The snowmobile, painted a deep red that contrasts sharply with its surroundings, charges forward aggressively, its rider, clad in a black thermal suit, leaning into a turn, gripping the chrome handlebars. Bright, painterly sunlight gleams from the handlebars and the rider's fully reflective visor, reflecting a distorted, vibrant image of the landscape. As the machine's tracks dig deep into the icy terrain, it kicks up a massive, billowing plume of fresh powder—a chaotic explosion of white and crystalline particles that catch the light. In the background, a majestic panorama of jagged, snow-capped peaks stands starkly against a clear, deep blue sky, creating an aura of thrilling power and sublime, rugged beauty.", + "seed": 11837, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_1/index_451_seed_11837_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.08156484253704548, + "vlm_logprob": -0.08156484253704548, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.08156484253704548, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11449434608221054 + ], + "opposite_logprob_tokens": [ + -2.4894943237304688 + ], + "logprob_reward": -0.11449434608221054, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.11449434608221054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07015500962734222, + 0.04471695050597191, + 0.09023090451955795, + 0.10523010045289993, + 0.199000284075737, + 0.36754047870635986, + 0.029126180335879326, + 0.01375522930175066, + 0.021380487829446793, + 0.05886435881257057 + ], + "credit_map": [ + 0.006932494230568409, + 0.010542469099164009, + 0.009260788559913635, + 0.010239597409963608, + 0.006970575544983149, + 0.00779760954901576, + 0.009671013802289963, + 0.008740459568798542, + 0.004278956446796656, + 0.005461342632770538, + 0.0050437236204743385, + 0.007096961140632629, + 0.005398217588663101, + 0.0052154576405882835, + 0.006674978416413069, + 0.005547313950955868, + 0.00990927591919899, + 0.012481068260967731, + 0.010537514463067055, + 0.012556529603898525, + 0.008142314851284027, + 0.02202923223376274, + 0.006846443749964237, + 0.007728526368737221, + 0.009027465246617794, + 0.016800010576844215, + 0.01393580250442028, + 0.018764164298772812, + 0.011278871446847916, + 0.01295530516654253, + 0.009431402198970318, + 0.013037089258432388, + 0.01958981715142727, + 0.023158499971032143, + 0.03479797765612602, + 0.04705336317420006, + 0.013908748514950275, + 0.01731155440211296, + 0.017827820032835007, + 0.02535250596702099, + 0.0357234925031662, + 0.04989634081721306, + 0.056774113327264786, + 0.09024648368358612, + 0.027709273621439934, + 0.03542246297001839, + 0.02844144217669964, + 0.04332686960697174, + 0.0024991645477712154, + 0.005811640061438084, + 0.004483065102249384, + 0.005047757178544998, + 0.0029041224624961615, + 0.002747915219515562, + 0.002735473681241274, + 0.0028970418497920036, + 0.0008920608670450747, + 0.0035385771188884974, + 0.0023638177663087845, + 0.002239964436739683, + 0.0014239373849704862, + 0.0010839856695383787, + 0.001113109989091754, + 0.0010997757781296968, + 0.0019373147515580058, + 0.003283743280917406, + 0.002513797953724861, + 0.00199599820189178, + 0.004986352752894163, + 0.0019908910617232323, + 0.002298539038747549, + 0.0023738492745906115, + 0.004517809953540564, + 0.010756173171103, + 0.009236510843038559, + 0.009854855947196484, + 0.005239515565335751, + 0.006359915714710951, + 0.007067510392516851, + 0.005832068156450987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.11449434608221054 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280955493450165 + ], + "opposite_logprob_tokens": [ + -2.0178096294403076 + ], + "logprob_reward": -0.14280955493450165, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14280955493450165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16638711094856262, + 0.09515582025051117, + 0.11787015944719315, + 0.1265515238046646, + 0.13779911398887634, + 0.20197643339633942, + 0.029516087844967842, + 0.018853584304451942, + 0.023883068934082985, + 0.08200709521770477 + ], + "credit_map": [ + 0.01644068770110607, + 0.04470507800579071, + 0.025062348693609238, + 0.02051066979765892, + 0.016023743897676468, + 0.010921063832938671, + 0.019778944551944733, + 0.012944569811224937, + 0.0073770503513514996, + 0.009281645528972149, + 0.011422726325690746, + 0.01910616271197796, + 0.011951626278460026, + 0.011693604290485382, + 0.013829092495143414, + 0.010493908077478409, + 0.00937158614397049, + 0.014306828379631042, + 0.013407830148935318, + 0.017005525529384613, + 0.011288821697235107, + 0.022483207285404205, + 0.01323855109512806, + 0.01676781475543976, + 0.00916109886020422, + 0.017844391986727715, + 0.01645052060484886, + 0.023912042379379272, + 0.011792529374361038, + 0.01605966128408909, + 0.013234472833573818, + 0.018096817657351494, + 0.014563478529453278, + 0.01836986094713211, + 0.022755736485123634, + 0.02991730161011219, + 0.011253908276557922, + 0.011008144356310368, + 0.01364806666970253, + 0.01628262922167778, + 0.019566187635064125, + 0.025963839143514633, + 0.0307527594268322, + 0.05089813843369484, + 0.015319486148655415, + 0.016718406230211258, + 0.01907356083393097, + 0.02368406020104885, + 0.0031075715087354183, + 0.005610113497823477, + 0.004347797948867083, + 0.0051433430053293705, + 0.002109774388372898, + 0.003441276028752327, + 0.002915425691753626, + 0.002840786473825574, + 0.001729159033857286, + 0.0045964037999510765, + 0.003059100592508912, + 0.002957567572593689, + 0.001562556135468185, + 0.001649211859330535, + 0.0016717901453375816, + 0.0016277952818199992, + 0.002502252347767353, + 0.003985484596341848, + 0.0029191344510763884, + 0.0017397227929905057, + 0.0047739543952047825, + 0.002336009405553341, + 0.003058121306821704, + 0.0025683902204036713, + 0.005746726877987385, + 0.013857015408575535, + 0.013584010303020477, + 0.014844344928860664, + 0.0077979289926588535, + 0.010813756845891476, + 0.00900490116328001, + 0.006358408369123936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.14280955493450165 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253320559859276 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253320559859276, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253320559859276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2511349320411682, + 0.06737835705280304, + 0.08684607595205307, + 0.155823215842247, + 0.11083264648914337, + 0.09066646546125412, + 0.05504494160413742, + 0.039804574102163315, + 0.04818452149629593, + 0.09428428113460541 + ], + "credit_map": [ + 0.02515716291964054, + 0.03777472674846649, + 0.029519805684685707, + 0.030568409711122513, + 0.03155786544084549, + 0.027690647169947624, + 0.03784361109137535, + 0.031022725626826286, + 0.005658082198351622, + 0.006974299903959036, + 0.006466097664088011, + 0.009830265305936337, + 0.010565627366304398, + 0.008362228982150555, + 0.010516926646232605, + 0.00900483038276434, + 0.009065767750144005, + 0.009897743351757526, + 0.013301224447786808, + 0.018953585997223854, + 0.00961086992174387, + 0.010986417531967163, + 0.007664161268621683, + 0.00736630754545331, + 0.013127370737493038, + 0.026847461238503456, + 0.01948440633714199, + 0.030078433454036713, + 0.01201522909104824, + 0.01845681667327881, + 0.014850427396595478, + 0.020963087677955627, + 0.011170045472681522, + 0.014641433954238892, + 0.019350312650203705, + 0.028582416474819183, + 0.006414457224309444, + 0.010904904454946518, + 0.009888521395623684, + 0.009880557656288147, + 0.00799611397087574, + 0.008804875425994396, + 0.011858264915645123, + 0.01786721684038639, + 0.011972914449870586, + 0.010018919594585896, + 0.008613565005362034, + 0.013534601777791977, + 0.006496873218566179, + 0.009664610959589481, + 0.008370905183255672, + 0.00825977697968483, + 0.006809729617089033, + 0.006051722913980484, + 0.0050037759356200695, + 0.004387552849948406, + 0.002621854655444622, + 0.004728286527097225, + 0.004975497722625732, + 0.006385798566043377, + 0.004628477618098259, + 0.005016690120100975, + 0.006026626564562321, + 0.005421349313110113, + 0.004483520053327084, + 0.00990956649184227, + 0.006944945547729731, + 0.0052196490578353405, + 0.007131156511604786, + 0.004797577857971191, + 0.004462527111172676, + 0.005235582590103149, + 0.006052161566913128, + 0.015279987826943398, + 0.01630566269159317, + 0.02170330472290516, + 0.007087375037372112, + 0.008494487032294273, + 0.010126940906047821, + 0.00923437811434269 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.005253320559859276 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427827775478363 + ], + "opposite_logprob_tokens": [ + -2.017782688140869 + ], + "logprob_reward": -0.1427827775478363, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1427827775478363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20060515403747559, + 0.06314610689878464, + 0.10605113208293915, + 0.07481559365987778, + 0.07472424954175949, + 0.1298014223575592, + 0.033473215997219086, + 0.0695403590798378, + 0.024170519784092903, + 0.2236722856760025 + ], + "credit_map": [ + 0.020719487220048904, + 0.029668597504496574, + 0.026477806270122528, + 0.02809455804526806, + 0.023885196074843407, + 0.02076607383787632, + 0.02951093576848507, + 0.021482501178979874, + 0.007276673801243305, + 0.006560669280588627, + 0.007874168455600739, + 0.012973935343325138, + 0.0077886986546218395, + 0.0057747806422412395, + 0.009239204227924347, + 0.0056579760275781155, + 0.01140742376446724, + 0.013105160556733608, + 0.012645457871258259, + 0.01564101316034794, + 0.010423509404063225, + 0.02599896490573883, + 0.008296620100736618, + 0.008532984182238579, + 0.0069026583805680275, + 0.013006908819079399, + 0.010782384313642979, + 0.014717092737555504, + 0.007257521152496338, + 0.010091996751725674, + 0.006031835917383432, + 0.006025201641023159, + 0.007541963364928961, + 0.008806563913822174, + 0.014008704572916031, + 0.017750069499015808, + 0.004733784589916468, + 0.006009596865624189, + 0.007064186502248049, + 0.008809378370642662, + 0.014193186536431313, + 0.022340506315231323, + 0.017783042043447495, + 0.02646099589765072, + 0.009171918034553528, + 0.014139319770038128, + 0.009297583252191544, + 0.01641487143933773, + 0.0037551342975348234, + 0.008190971799194813, + 0.004902905318886042, + 0.004920804873108864, + 0.0029746145009994507, + 0.0024883965961635113, + 0.0030770855955779552, + 0.0031633013859391212, + 0.0056464783847332, + 0.015808189287781715, + 0.010870405472815037, + 0.011215939186513424, + 0.006280544679611921, + 0.007668269332498312, + 0.006264856085181236, + 0.005785681772977114, + 0.0023188346531242132, + 0.004379304125905037, + 0.0029904453549534082, + 0.002024588640779257, + 0.0040895058773458, + 0.0028500864282250404, + 0.0029425369575619698, + 0.0025752161163836718, + 0.02456580474972725, + 0.053654130548238754, + 0.03453042358160019, + 0.03010442480444908, + 0.017848221585154533, + 0.02254226990044117, + 0.022819845005869865, + 0.017607158049941063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.1427827775478363 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002484213560819626 + ], + "opposite_logprob_tokens": [ + -6.002484321594238 + ], + "logprob_reward": -0.002484213560819626, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002484213560819626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13585014641284943, + 0.06994125992059708, + 0.3261859118938446, + 0.03585004806518555, + 0.08271339535713196, + 0.11722078919410706, + 0.02281663566827774, + 0.05533742532134056, + 0.07418300956487656, + 0.07990138977766037 + ], + "credit_map": [ + 0.013434208929538727, + 0.022046957165002823, + 0.018828529864549637, + 0.017070366069674492, + 0.015910008922219276, + 0.012604427523911, + 0.023028690367937088, + 0.012926957570016384, + 0.0072658671997487545, + 0.009761196561157703, + 0.008604495786130428, + 0.012857153080403805, + 0.0074083758518099785, + 0.007938414812088013, + 0.008361629210412502, + 0.007744125556200743, + 0.03840094059705734, + 0.036097049713134766, + 0.04174324497580528, + 0.05108143761754036, + 0.03834885358810425, + 0.07320012152194977, + 0.021681049838662148, + 0.02563321962952614, + 0.0038065980188548565, + 0.007475723512470722, + 0.00494013587012887, + 0.006158350966870785, + 0.003102006623521447, + 0.004811203107237816, + 0.0028144298121333122, + 0.0027415987569838762, + 0.008132887072861195, + 0.014146744273602962, + 0.012277323752641678, + 0.014337483793497086, + 0.007082290016114712, + 0.009136658161878586, + 0.008020391687750816, + 0.009579619392752647, + 0.013409308157861233, + 0.02183137647807598, + 0.014507570303976536, + 0.01787835732102394, + 0.009014267474412918, + 0.0170480664819479, + 0.007733688689768314, + 0.015798160806298256, + 0.002130842301994562, + 0.0027149070519953966, + 0.0031373214442282915, + 0.0034989304840564728, + 0.002546320203691721, + 0.0038886084221303463, + 0.0028133599553257227, + 0.002086346037685871, + 0.004521781112998724, + 0.017704090103507042, + 0.01042037084698677, + 0.006268505472689867, + 0.006717203184962273, + 0.003639624221250415, + 0.0032425532117486, + 0.0028232999611645937, + 0.00866302102804184, + 0.02007398195564747, + 0.011425263248383999, + 0.007006468717008829, + 0.00910879485309124, + 0.0061939675360918045, + 0.00536587368696928, + 0.006345640402287245, + 0.007298030890524387, + 0.016789846122264862, + 0.011959481053054333, + 0.01223235484212637, + 0.006397145334631205, + 0.008494815789163113, + 0.010557298548519611, + 0.006172419060021639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.002484213560819626 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot tracks a powerful four-stroke snowmobile, the camera shaking violently to simulate intense vibrations as the machine accelerates explosively across pristine snow. The entire scene is rendered in a detailed, artistic style reminiscent of traditional oil painting, with all surfaces possessing visible, thick, mural-like textures. The snowmobile, painted a deep red that contrasts sharply with its surroundings, charges forward aggressively, its rider, clad in a black thermal suit, leaning into a turn, gripping the chrome handlebars. Bright, painterly sunlight gleams from the handlebars and the rider's fully reflective visor, reflecting a distorted, vibrant image of the landscape. As the machine's tracks dig deep into the icy terrain, it kicks up a massive, billowing plume of fresh powder—a chaotic explosion of white and crystalline particles that catch the light. In the background, a majestic panorama of jagged, snow-capped peaks stands starkly against a clear, deep blue sky, creating an aura of thrilling power and sublime, rugged beauty.", + "seed": 11838, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_1/index_451_seed_11838_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14404328325763344, + "vlm_logprob": -0.14404328325763344, + "vlm_yes_no_margin": 2.974999988079071, + "vlm_token_credit": -0.14404328325763344, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2062443643808365 + ], + "opposite_logprob_tokens": [ + -1.9562443494796753 + ], + "logprob_reward": -0.2062443643808365, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.2062443643808365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.372901052236557, + 0.05180595442652702, + 0.06164627894759178, + 0.05076398327946663, + 0.10959993302822113, + 0.12108127027750015, + 0.12011632323265076, + 0.04515967518091202, + 0.05024067685008049, + 0.016684798523783684 + ], + "credit_map": [ + 0.03874016925692558, + 0.06658200919628143, + 0.062093403190374374, + 0.05629672482609749, + 0.03731152415275574, + 0.036229245364665985, + 0.04884814843535423, + 0.026799894869327545, + 0.005642685107886791, + 0.007292044814676046, + 0.007395319174975157, + 0.009652952663600445, + 0.0040547908283770084, + 0.006964302156120539, + 0.004709907807409763, + 0.0060939546674489975, + 0.005390678998082876, + 0.008604319766163826, + 0.009423695504665375, + 0.00550620025023818, + 0.00493575306609273, + 0.00956922397017479, + 0.00372042553499341, + 0.014495988376438618, + 0.0068662771955132484, + 0.0046510035172104836, + 0.008197558112442493, + 0.005706660449504852, + 0.005810892675071955, + 0.010145129635930061, + 0.004509460646659136, + 0.004877003841102123, + 0.012117653153836727, + 0.015477180480957031, + 0.016285380348563194, + 0.010680153965950012, + 0.012448855675756931, + 0.02047591097652912, + 0.012552195228636265, + 0.009562618099153042, + 0.011063321493566036, + 0.012920595705509186, + 0.016810230910778046, + 0.023191183805465698, + 0.014517353847622871, + 0.011600072495639324, + 0.016082359477877617, + 0.014896158128976822, + 0.011830808594822884, + 0.015669120475649834, + 0.015643620863556862, + 0.012568512931466103, + 0.014892756007611752, + 0.015056273899972439, + 0.01576291024684906, + 0.01869233325123787, + 0.005456859245896339, + 0.005569431930780411, + 0.005029316060245037, + 0.005189066287130117, + 0.004289431031793356, + 0.006652807351201773, + 0.005918498151004314, + 0.007054269779473543, + 0.005334053188562393, + 0.007752724923193455, + 0.007185779977589846, + 0.005589460488408804, + 0.004499398171901703, + 0.008878093212842941, + 0.0046910555101931095, + 0.006310114171355963, + 0.001691215205937624, + 0.0019882943015545607, + 0.002233024686574936, + 0.0024164069909602404, + 0.0013018125900998712, + 0.002307626185938716, + 0.002442876808345318, + 0.0023035428021103144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.2062443643808365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289253354072571 + ], + "opposite_logprob_tokens": [ + -1.0539252758026123 + ], + "logprob_reward": -0.4289253354072571, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.4289253354072571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3793504238128662, + 0.17869317531585693, + 0.07531073689460754, + 0.04577254503965378, + 0.14961299300193787, + 0.04953336715698242, + 0.04702931270003319, + 0.02907285839319229, + 0.02360052429139614, + 0.02202409692108631 + ], + "credit_map": [ + 0.0394580103456974, + 0.04418681934475899, + 0.06816408783197403, + 0.07347439229488373, + 0.03695186600089073, + 0.043781619518995285, + 0.03830762952566147, + 0.03502598777413368, + 0.01726491190493107, + 0.03472835198044777, + 0.023025725036859512, + 0.026789488270878792, + 0.016222070902585983, + 0.021311365067958832, + 0.020966632291674614, + 0.018384616822004318, + 0.006144124083220959, + 0.010320755653083324, + 0.009699571877717972, + 0.007415444124490023, + 0.007357208523899317, + 0.016341745853424072, + 0.006554564461112022, + 0.011477314867079258, + 0.006257164292037487, + 0.006726326420903206, + 0.006467251572757959, + 0.006961645092815161, + 0.003605472156777978, + 0.004675000440329313, + 0.006152474321424961, + 0.004927203990519047, + 0.016573233529925346, + 0.02012283354997635, + 0.021367959678173065, + 0.015722453594207764, + 0.016513995826244354, + 0.031293708831071854, + 0.014622519724071026, + 0.013396280817687511, + 0.004884496331214905, + 0.005323856603354216, + 0.008395612239837646, + 0.011330434121191502, + 0.0036854087375104427, + 0.006510361097753048, + 0.004395043011754751, + 0.005008149426430464, + 0.005179084837436676, + 0.006804954260587692, + 0.005341487005352974, + 0.005687644239515066, + 0.00498654879629612, + 0.007470068521797657, + 0.0047315931878983974, + 0.006827928125858307, + 0.002998133422806859, + 0.0037988577969372272, + 0.003565048100426793, + 0.004097369499504566, + 0.0026854737661778927, + 0.0033610130194574594, + 0.004405568819493055, + 0.004161391872912645, + 0.0028061093762516975, + 0.0020892920438200235, + 0.0038965120911598206, + 0.002238011686131358, + 0.0022903410717844963, + 0.0056517417542636395, + 0.001766473869793117, + 0.002862038556486368, + 0.002267559990286827, + 0.00209213187918067, + 0.0032908201683312654, + 0.0030348720028996468, + 0.002181818475946784, + 0.003973809070885181, + 0.002290928503498435, + 0.0028921542689204216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.4289253354072571 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525747099891305 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.00525747099891305, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525747099891305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30538418889045715, + 0.07361587136983871, + 0.050074901431798935, + 0.05405691638588905, + 0.135066956281662, + 0.12314412742853165, + 0.15885668992996216, + 0.04431040957570076, + 0.039068713784217834, + 0.016421178355813026 + ], + "credit_map": [ + 0.025494758039712906, + 0.03995969519019127, + 0.05047579109668732, + 0.06604254990816116, + 0.03791629895567894, + 0.024572933092713356, + 0.03584301471710205, + 0.02507914789021015, + 0.007910260930657387, + 0.010556324385106564, + 0.009970544837415218, + 0.009300090372562408, + 0.005777839105576277, + 0.014059064909815788, + 0.006784465629607439, + 0.009257275611162186, + 0.004630550276488066, + 0.006925053894519806, + 0.0059956880286335945, + 0.005272687878459692, + 0.005045696627348661, + 0.009902482852339745, + 0.0057005807757377625, + 0.006602164823561907, + 0.007870715111494064, + 0.0058588506653904915, + 0.012531022541224957, + 0.004774461500346661, + 0.005386678036302328, + 0.006642803084105253, + 0.0053373947739601135, + 0.005654985550791025, + 0.013686935417354107, + 0.0182542335242033, + 0.01910756342113018, + 0.016251418739557266, + 0.017657514661550522, + 0.02345097064971924, + 0.0161200612783432, + 0.010538247413933277, + 0.00797831080853939, + 0.01240663230419159, + 0.016392773017287254, + 0.02360861748456955, + 0.01659321039915085, + 0.010639281012117863, + 0.01810755580663681, + 0.017417743802070618, + 0.014460071921348572, + 0.02083408087491989, + 0.020316189154982567, + 0.017756251618266106, + 0.02110009640455246, + 0.017877761274576187, + 0.019373252987861633, + 0.02713898941874504, + 0.0048732091672718525, + 0.0065128132700920105, + 0.005800333805382252, + 0.005343502853065729, + 0.0038440462667495012, + 0.004441301338374615, + 0.005775518249720335, + 0.007719683926552534, + 0.003918992821127176, + 0.005250032991170883, + 0.005229168105870485, + 0.005036411341279745, + 0.004497169051319361, + 0.005029102321714163, + 0.004738436546176672, + 0.005369399208575487, + 0.0015507674543187022, + 0.00229079183191061, + 0.002404289785772562, + 0.00199038814753294, + 0.0018372626509517431, + 0.0023682231549173594, + 0.001863323152065277, + 0.0021161329932510853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.00525747099891305 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001692801713943 + ], + "opposite_logprob_tokens": [ + -2.695016860961914 + ], + "logprob_reward": -0.07001692801713943, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001692801713943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27800193428993225, + 0.11739462614059448, + 0.057521212846040726, + 0.11300381273031235, + 0.2027934491634369, + 0.0426299162209034, + 0.12019406259059906, + 0.020809385925531387, + 0.028963590040802956, + 0.018688037991523743 + ], + "credit_map": [ + 0.02873884327709675, + 0.04755323752760887, + 0.038734953850507736, + 0.044109780341386795, + 0.04104411229491234, + 0.028597526252269745, + 0.028052158653736115, + 0.021171292290091515, + 0.011491796001791954, + 0.019864412024617195, + 0.016014713793992996, + 0.018821071833372116, + 0.011867543682456017, + 0.015600750222802162, + 0.01310334075242281, + 0.010630980134010315, + 0.006544129922986031, + 0.00708546070381999, + 0.007288120687007904, + 0.004673232790082693, + 0.00557311438024044, + 0.01360798254609108, + 0.005210842005908489, + 0.007538325153291225, + 0.015407873317599297, + 0.01323946937918663, + 0.01868920773267746, + 0.012784837745130062, + 0.012882675975561142, + 0.014800709672272205, + 0.012767217122018337, + 0.012431801296770573, + 0.019955065101385117, + 0.02658134698867798, + 0.02999839559197426, + 0.03281567990779877, + 0.020883841440081596, + 0.03639255464076996, + 0.019856436178088188, + 0.016310114413499832, + 0.0036242036148905754, + 0.004056572914123535, + 0.005636352580040693, + 0.007668065372854471, + 0.006211498752236366, + 0.0051741646602749825, + 0.004833607468754053, + 0.005425446201115847, + 0.013003542087972164, + 0.013723347336053848, + 0.017490411177277565, + 0.0107015585526824, + 0.015146637335419655, + 0.016287826001644135, + 0.010499521158635616, + 0.023341210559010506, + 0.002244071802124381, + 0.0021203700453042984, + 0.002999653574079275, + 0.002281478838995099, + 0.001823551021516323, + 0.00420546717941761, + 0.002145446138456464, + 0.002989345695823431, + 0.0025777285918593407, + 0.0036456575617194176, + 0.004274723585695028, + 0.0031812924426048994, + 0.0030402818229049444, + 0.00562283955514431, + 0.00295809842646122, + 0.003662965726107359, + 0.0020332089625298977, + 0.0019794367253780365, + 0.0025150892324745655, + 0.0027930033393204212, + 0.0016031061531975865, + 0.003405954223126173, + 0.002535655628889799, + 0.0018225815147161484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.07001692801713943 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009772317484021187 + ], + "opposite_logprob_tokens": [ + -4.634772300720215 + ], + "logprob_reward": -0.009772317484021187, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009772317484021187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2134837955236435, + 0.09552661329507828, + 0.04126247763633728, + 0.02851158007979393, + 0.238213449716568, + 0.23752450942993164, + 0.022425375878810883, + 0.0414113849401474, + 0.04206322133541107, + 0.03957756236195564 + ], + "credit_map": [ + 0.026697028428316116, + 0.030461343005299568, + 0.030028466135263443, + 0.027201255783438683, + 0.030597256496548653, + 0.02787250652909279, + 0.02107345312833786, + 0.01955248974263668, + 0.008463418111205101, + 0.015102636069059372, + 0.014544552192091942, + 0.011344172060489655, + 0.008268550969660282, + 0.01536515261977911, + 0.010305230505764484, + 0.012132907286286354, + 0.0037193249445408583, + 0.004082884639501572, + 0.0064729307778179646, + 0.0024446274619549513, + 0.0028990707360208035, + 0.010464453138411045, + 0.0025140896905213594, + 0.008665098808705807, + 0.0031009921804070473, + 0.002669957000762224, + 0.005237544886767864, + 0.00307414005510509, + 0.0029397716280072927, + 0.004725305829197168, + 0.0032806687522679567, + 0.0034832023084163666, + 0.03021668642759323, + 0.029940158128738403, + 0.03576095029711723, + 0.029243171215057373, + 0.022943709045648575, + 0.04670228064060211, + 0.023867059499025345, + 0.019539449363946915, + 0.028797324746847153, + 0.034566693007946014, + 0.038598861545324326, + 0.04913279041647911, + 0.017851093783974648, + 0.02742668427526951, + 0.020298315212130547, + 0.02085273526608944, + 0.00283951498568058, + 0.003153052646666765, + 0.0032413797453045845, + 0.0021551321260631084, + 0.0020583367440849543, + 0.0028571973089128733, + 0.002414233051240444, + 0.0037065288051962852, + 0.0043825628235936165, + 0.0034823131281882524, + 0.005952519830316305, + 0.004013004247099161, + 0.004062448162585497, + 0.00727342814207077, + 0.0047927298583090305, + 0.007452379446476698, + 0.004109859466552734, + 0.0037332302890717983, + 0.007276194170117378, + 0.003829968860372901, + 0.004162321798503399, + 0.008902503177523613, + 0.0032047319691628218, + 0.006844410672783852, + 0.00560412835329771, + 0.003753145458176732, + 0.0062190028838813305, + 0.003763347864151001, + 0.0035379326436668634, + 0.0070359548553824425, + 0.0034037188161164522, + 0.006260330323129892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.009772317484021187 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot tracks a powerful four-stroke snowmobile, the camera shaking violently to simulate intense vibrations as the machine accelerates explosively across pristine snow. The entire scene is rendered in a detailed, artistic style reminiscent of traditional oil painting, with all surfaces possessing visible, thick, mural-like textures. The snowmobile, painted a deep red that contrasts sharply with its surroundings, charges forward aggressively, its rider, clad in a black thermal suit, leaning into a turn, gripping the chrome handlebars. Bright, painterly sunlight gleams from the handlebars and the rider's fully reflective visor, reflecting a distorted, vibrant image of the landscape. As the machine's tracks dig deep into the icy terrain, it kicks up a massive, billowing plume of fresh powder—a chaotic explosion of white and crystalline particles that catch the light. In the background, a majestic panorama of jagged, snow-capped peaks stands starkly against a clear, deep blue sky, creating an aura of thrilling power and sublime, rugged beauty.", + "seed": 11839, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_1/index_451_seed_11839_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05516368141397834, + "vlm_logprob": -0.05516368141397834, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.05516368141397834, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07159964740276337 + ], + "opposite_logprob_tokens": [ + -3.0715997219085693 + ], + "logprob_reward": -0.07159964740276337, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.07159964740276337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1638956218957901, + 0.06056831404566765, + 0.05480888485908508, + 0.06373583525419235, + 0.15339787304401398, + 0.18786078691482544, + 0.11791227757930756, + 0.058155979961156845, + 0.12085787206888199, + 0.018806541338562965 + ], + "credit_map": [ + 0.020374655723571777, + 0.012519885785877705, + 0.014015588909387589, + 0.01784883439540863, + 0.027569791302084923, + 0.010507139377295971, + 0.030340587720274925, + 0.03071914240717888, + 0.004995646420866251, + 0.00918559916317463, + 0.005403608549386263, + 0.00703443493694067, + 0.008292782120406628, + 0.008152453228831291, + 0.010214434005320072, + 0.007289355155080557, + 0.006074701435863972, + 0.007820279337465763, + 0.00635378435254097, + 0.0059409551322460175, + 0.007934157736599445, + 0.009625616483390331, + 0.005373440682888031, + 0.005685948301106691, + 0.010866750963032246, + 0.008604773320257664, + 0.006675239652395248, + 0.0046485308557748795, + 0.008747304789721966, + 0.012012479826807976, + 0.005317922681570053, + 0.006862832698971033, + 0.024184677749872208, + 0.012394269928336143, + 0.016983790323138237, + 0.022980883717536926, + 0.013421286828815937, + 0.029774324968457222, + 0.017567897215485573, + 0.016090748831629753, + 0.024235118180513382, + 0.0225947555154562, + 0.024396440014243126, + 0.014059706591069698, + 0.023338811472058296, + 0.03942986577749252, + 0.018225857987999916, + 0.021580221131443977, + 0.013255976140499115, + 0.01585129275918007, + 0.011048741638660431, + 0.017407895997166634, + 0.0128802340477705, + 0.019366787746548653, + 0.015051223337650299, + 0.013050127774477005, + 0.005570387002080679, + 0.009866737760603428, + 0.005943698342889547, + 0.006619539577513933, + 0.006732981652021408, + 0.006087569519877434, + 0.010014318861067295, + 0.007320743054151535, + 0.012823731638491154, + 0.013921577483415604, + 0.01655641384422779, + 0.02089180052280426, + 0.010852660983800888, + 0.01268412172794342, + 0.018109766766428947, + 0.0150177963078022, + 0.0023007337003946304, + 0.002406913787126541, + 0.001833760761655867, + 0.0017734721768647432, + 0.0030158134177327156, + 0.00297269388101995, + 0.0022764462046325207, + 0.002226707059890032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.07159964740276337 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270471066236496 + ], + "opposite_logprob_tokens": [ + -2.127047061920166 + ], + "logprob_reward": -0.1270471066236496, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270471066236496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2829715311527252, + 0.0804339125752449, + 0.08101366460323334, + 0.09431198239326477, + 0.06136564165353775, + 0.07620327174663544, + 0.06126250699162483, + 0.03997568413615227, + 0.19585616886615753, + 0.02660559117794037 + ], + "credit_map": [ + 0.026982519775629044, + 0.031121445819735527, + 0.027574418112635612, + 0.03771175816655159, + 0.05768462270498276, + 0.02900296077132225, + 0.04757314547896385, + 0.025320684537291527, + 0.005664507858455181, + 0.008907375857234001, + 0.007364498917013407, + 0.013129587285220623, + 0.007941431365907192, + 0.011500832624733448, + 0.015462514944374561, + 0.010463177226483822, + 0.008675975725054741, + 0.012078138068318367, + 0.00991759728640318, + 0.007557911332696676, + 0.009963717311620712, + 0.018536364659667015, + 0.006299791857600212, + 0.00798417441546917, + 0.015537348575890064, + 0.014215084724128246, + 0.010295053943991661, + 0.00623831944540143, + 0.012792795896530151, + 0.01939980871975422, + 0.009305943734943867, + 0.0065276408568024635, + 0.009658163413405418, + 0.004599356092512608, + 0.007074710913002491, + 0.008833091706037521, + 0.005757912993431091, + 0.011822232976555824, + 0.006654008291661739, + 0.006966174114495516, + 0.008915780112147331, + 0.008986163884401321, + 0.009618903510272503, + 0.0061706299893558025, + 0.009876075200736523, + 0.015666261315345764, + 0.009116812609136105, + 0.007852647453546524, + 0.005418900866061449, + 0.009961959905922413, + 0.007517984602600336, + 0.009290643967688084, + 0.007772794924676418, + 0.0060744029469788074, + 0.00823903363198042, + 0.006986793130636215, + 0.004049076698720455, + 0.005841286387294531, + 0.0034030266106128693, + 0.0034317336976528168, + 0.006258074659854174, + 0.006976928561925888, + 0.005354295019060373, + 0.004661267157644033, + 0.023211713880300522, + 0.0227724090218544, + 0.026372313499450684, + 0.02500498853623867, + 0.023176753893494606, + 0.026126399636268616, + 0.02718769572675228, + 0.022003918886184692, + 0.004303286783397198, + 0.0027993277180939913, + 0.0026858169585466385, + 0.0027277872432023287, + 0.003830929286777973, + 0.0038103556726127863, + 0.002957116113975644, + 0.003490975359454751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.1270471066236496 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023284463211894035 + ], + "opposite_logprob_tokens": [ + -3.773284435272217 + ], + "logprob_reward": -0.023284463211894035, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023284463211894035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23183700442314148, + 0.09053907543420792, + 0.09153680503368378, + 0.08383432030677795, + 0.05338667705655098, + 0.07007007300853729, + 0.09063728898763657, + 0.16094431281089783, + 0.0870891660451889, + 0.04012526944279671 + ], + "credit_map": [ + 0.01752340979874134, + 0.019720222800970078, + 0.025421584025025368, + 0.03181488439440727, + 0.044371336698532104, + 0.017088498920202255, + 0.04537957161664963, + 0.03051750734448433, + 0.01323727983981371, + 0.012773163616657257, + 0.008205033838748932, + 0.013859418220818043, + 0.011049763299524784, + 0.008365864865481853, + 0.013250740244984627, + 0.009797808714210987, + 0.008199072442948818, + 0.012386126443743706, + 0.01493485365062952, + 0.011366677470505238, + 0.012757042422890663, + 0.011279357597231865, + 0.011180871166288853, + 0.009432803839445114, + 0.013086115941405296, + 0.008473473601043224, + 0.011815708130598068, + 0.005561248864978552, + 0.01165749505162239, + 0.017868883907794952, + 0.0060221184976398945, + 0.009349283762276173, + 0.008941173553466797, + 0.005498491693288088, + 0.006920509971678257, + 0.004741388838738203, + 0.007494641933590174, + 0.009789489209651947, + 0.005484950263053179, + 0.004516038112342358, + 0.005576409865170717, + 0.0070838965475559235, + 0.01034885086119175, + 0.010387483984231949, + 0.009603646583855152, + 0.005711605306714773, + 0.009987378492951393, + 0.011370805092155933, + 0.007638143375515938, + 0.01608704775571823, + 0.011633963324129581, + 0.013319060206413269, + 0.009507188573479652, + 0.007167026400566101, + 0.01488969475030899, + 0.0103951720520854, + 0.01914464496076107, + 0.018023621290922165, + 0.025332655757665634, + 0.02169804647564888, + 0.019606322050094604, + 0.020299406722187996, + 0.021013595163822174, + 0.015826035290956497, + 0.006037080194801092, + 0.014172408729791641, + 0.014140340499579906, + 0.013385379686951637, + 0.007530247326940298, + 0.0064505524933338165, + 0.014616320841014385, + 0.010756838135421276, + 0.005946821067482233, + 0.002971147419884801, + 0.0032108507584780455, + 0.003552940906956792, + 0.005662919487804174, + 0.00973210297524929, + 0.0031848824582993984, + 0.0058636050671339035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.023284463211894035 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864228889346123 + ], + "opposite_logprob_tokens": [ + -3.04864239692688 + ], + "logprob_reward": -0.04864228889346123, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864228889346123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36841875314712524, + 0.04251166433095932, + 0.0630861148238182, + 0.08701654523611069, + 0.059457067400217056, + 0.04245724156498909, + 0.05575739964842796, + 0.1688358038663864, + 0.07818285375833511, + 0.03427654132246971 + ], + "credit_map": [ + 0.03981905058026314, + 0.03130422905087471, + 0.021310430020093918, + 0.0428728312253952, + 0.0727490782737732, + 0.02731439098715782, + 0.06407701224088669, + 0.06897171586751938, + 0.0050661321729421616, + 0.005915987305343151, + 0.0036590159870684147, + 0.0063689034432172775, + 0.005626347381621599, + 0.00495738023892045, + 0.0058773173950612545, + 0.005040577147156, + 0.005932851694524288, + 0.008837281726300716, + 0.007730955258011818, + 0.007250855211168528, + 0.008043312467634678, + 0.01142862904816866, + 0.0070152441039681435, + 0.006846989039331675, + 0.012065397575497627, + 0.010252626612782478, + 0.011824019253253937, + 0.005224886815994978, + 0.013674535788595676, + 0.017985785380005836, + 0.007469811011105776, + 0.008519480004906654, + 0.010538620874285698, + 0.005041369237005711, + 0.00714945187792182, + 0.00728140864521265, + 0.00749389361590147, + 0.011834967881441116, + 0.004797847010195255, + 0.005319509655237198, + 0.00611065374687314, + 0.0049651265144348145, + 0.00614688266068697, + 0.0035211164504289627, + 0.00526424078270793, + 0.007256293669342995, + 0.0032397210597991943, + 0.005953206215053797, + 0.005200543440878391, + 0.008709982968866825, + 0.006181912496685982, + 0.008433527313172817, + 0.0075028203427791595, + 0.005473609082400799, + 0.008032284677028656, + 0.00622271653264761, + 0.019081326201558113, + 0.026401912793517113, + 0.020677009597420692, + 0.017833327874541283, + 0.021784868091344833, + 0.02318982221186161, + 0.020638514310121536, + 0.01922902651131153, + 0.008957592770457268, + 0.009147639386355877, + 0.009293859824538231, + 0.01074861828237772, + 0.006559403613209724, + 0.008463358506560326, + 0.015037724748253822, + 0.00997465755790472, + 0.005763770081102848, + 0.003972531296312809, + 0.0036164706107228994, + 0.0037224269472062588, + 0.003096259431913495, + 0.0063512916676700115, + 0.003231719834730029, + 0.004522069822996855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.04864228889346123 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in a detailed artistic style reminiscent of traditional oil painting with thick, mural-like textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a deep red snowmobile accelerating aggressively across pristine snow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera shaking violently to simulate intense vibrations while the snowmobile charges forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the background feature a panorama of jagged, snow-capped peaks against a clear, deep blue sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005244900938123465 + ], + "opposite_logprob_tokens": [ + -5.255244731903076 + ], + "logprob_reward": -0.005244900938123465, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005244900938123465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.253383070230484, + 0.06585143506526947, + 0.07407140731811523, + 0.14850781857967377, + 0.059739649295806885, + 0.03459833934903145, + 0.03071352280676365, + 0.09962480515241623, + 0.17956142127513885, + 0.05394849181175232 + ], + "credit_map": [ + 0.038554929196834564, + 0.020816460251808167, + 0.046440932899713516, + 0.022991225123405457, + 0.02003011293709278, + 0.02227659523487091, + 0.028674552217125893, + 0.053598277270793915, + 0.004392899107187986, + 0.0035934909246861935, + 0.0034837471321225166, + 0.006185115780681372, + 0.0075697991997003555, + 0.01816508360207081, + 0.011603821069002151, + 0.010857474990189075, + 0.008342807181179523, + 0.007404509000480175, + 0.01037302240729332, + 0.007396903354674578, + 0.009760093875229359, + 0.017820965498685837, + 0.005592432804405689, + 0.007380670867860317, + 0.020834732800722122, + 0.0207511056214571, + 0.017078986391425133, + 0.008124860934913158, + 0.02321641519665718, + 0.03075200878083706, + 0.012867449782788754, + 0.014882270246744156, + 0.009299892000854015, + 0.005950459744781256, + 0.007470249198377132, + 0.005743345245718956, + 0.005277141463011503, + 0.014906457625329494, + 0.004676555283367634, + 0.006415547337383032, + 0.004292881581932306, + 0.004060573410242796, + 0.0047063943929970264, + 0.0035248654894530773, + 0.004627013113349676, + 0.005791862960904837, + 0.003320635762065649, + 0.004274113569408655, + 0.00292807980440557, + 0.004477561451494694, + 0.004686613567173481, + 0.0037467600777745247, + 0.0026473328471183777, + 0.004431467037647963, + 0.003927986137568951, + 0.003867720253765583, + 0.010173391550779343, + 0.011324985884130001, + 0.012218055315315723, + 0.01422712579369545, + 0.011573459953069687, + 0.013306614942848682, + 0.011528383009135723, + 0.01527278684079647, + 0.01952621340751648, + 0.02125629410147667, + 0.02779119275510311, + 0.027179844677448273, + 0.018246851861476898, + 0.02168862149119377, + 0.023095818236470222, + 0.02077658101916313, + 0.00693080760538578, + 0.006070184987038374, + 0.005441045854240656, + 0.004787186626344919, + 0.008505670353770256, + 0.009041566401720047, + 0.0056608873419463634, + 0.007511140778660774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.005244900938123465 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_10/index_tensor([9133, 9133])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_10/index_tensor([9133, 9133])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4e5c1f29a2ecbeb9926d123b4e8d5f3eca80f5ef --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_10/index_tensor([9133, 9133])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, right-facing, long shot transports us into a deep-sea world reminiscent of a richly colored classical oil painting. A massive humpback whale swims by with a majestic and deliberate rhythm, its massive frame practically dominating the frame. Its deep blue, almost black skin, stained with white scratches and barnacles from the passage of time, reflects the rich luster of oil paint in the faint, divinely inspired beams of light that filter down from the surface. Its massive, powerful tail fin rises and falls with almost imperceptible grace and stability, each movement propelling it effortlessly through the deep, almost inky Prussian blue waters. Light filters through the water, creating soft, rich gradations of light and dark, reminiscent of the chiaroscuro in a Rembrandt painting. Countless tiny plankton, like suspended golden dust, imbue the gloom with a shimmer of life. The background is an endless abyss, and the whole picture is full of tranquility, solemnity and an epic feeling that transcends time. The texture is heavy, as if every frame is a carefully painted ocean-themed masterpiece.", + "seed": 23136, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_10/index_9133_seed_23136_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.2734656870365143, + "vlm_logprob": -0.2734656870365143, + "vlm_yes_no_margin": 1.375, + "vlm_token_credit": -0.2734656870365143, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06422582268714905 + ], + "opposite_logprob_tokens": [ + -2.814225912094116 + ], + "logprob_reward": -0.06422582268714905, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06422582268714905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3018113374710083, + 0.06462564319372177, + 0.11587735265493393, + 0.08708252757787704, + 0.14357618987560272, + 0.031203461810946465, + 0.03807086870074272, + 0.10465894639492035, + 0.054973967373371124, + 0.05811968073248863 + ], + "credit_map": [ + 0.027206337079405785, + 0.028404507786035538, + 0.022053861990571022, + 0.0154085299000144, + 0.11014677584171295, + 0.06018320843577385, + 0.022338900715112686, + 0.016069214791059494, + 0.00418071448802948, + 0.006754746660590172, + 0.010358133353292942, + 0.009131251834332943, + 0.009098094888031483, + 0.00918120052665472, + 0.010001248680055141, + 0.005920251365751028, + 0.011779421009123325, + 0.0047013601288199425, + 0.04301958903670311, + 0.01403130404651165, + 0.008674257434904575, + 0.007258498575538397, + 0.009664403274655342, + 0.016748519614338875, + 0.008954309858381748, + 0.007034218404442072, + 0.025967370718717575, + 0.012164229527115822, + 0.005905300844460726, + 0.004835017956793308, + 0.009170026518404484, + 0.013052063062787056, + 0.01314496248960495, + 0.019100040197372437, + 0.018311651423573494, + 0.021384084597229958, + 0.013715624809265137, + 0.012817531824111938, + 0.02335183322429657, + 0.02175046317279339, + 0.002746314974501729, + 0.0030034128576517105, + 0.007179454900324345, + 0.004243252333253622, + 0.00275148986838758, + 0.0039378502406179905, + 0.0037013322580605745, + 0.0036403522826731205, + 0.002954325405880809, + 0.0042955996468663216, + 0.005310700740665197, + 0.005705493967980146, + 0.0032836487516760826, + 0.006555515807121992, + 0.005582287907600403, + 0.0043832953087985516, + 0.00628116587176919, + 0.011642646975815296, + 0.01889084279537201, + 0.017918149009346962, + 0.010915580205619335, + 0.016560055315494537, + 0.011877148412168026, + 0.010573362000286579, + 0.004387005232274532, + 0.00936645083129406, + 0.008731676265597343, + 0.00807389710098505, + 0.004737135488539934, + 0.005363101605325937, + 0.008001631125807762, + 0.006313068326562643, + 0.0037903771735727787, + 0.009902825579047203, + 0.009402593597769737, + 0.010144001804292202, + 0.006570357363671064, + 0.004803884774446487, + 0.007942724041640759, + 0.005562916398048401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.06422582268714905 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813894748687744 + ], + "opposite_logprob_tokens": [ + -1.4063894748687744 + ], + "logprob_reward": -0.2813894748687744, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813894748687744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27422359585762024, + 0.09035281091928482, + 0.19956079125404358, + 0.11808092147111893, + 0.07042734324932098, + 0.052110154181718826, + 0.04958476126194, + 0.03270924836397171, + 0.05738084018230438, + 0.055569544434547424 + ], + "credit_map": [ + 0.015751928091049194, + 0.022633275017142296, + 0.025363611057400703, + 0.016009915620088577, + 0.0875144749879837, + 0.04665020480751991, + 0.02238442935049534, + 0.03791573643684387, + 0.0065145548433065414, + 0.008558436296880245, + 0.013119244016706944, + 0.00816383771598339, + 0.017917253077030182, + 0.016139673069119453, + 0.012449042871594429, + 0.0074907708913087845, + 0.016889942809939384, + 0.013922363519668579, + 0.05594060942530632, + 0.025790056213736534, + 0.020888160914182663, + 0.017930924892425537, + 0.021596956998109818, + 0.026601793244481087, + 0.01527788583189249, + 0.006084947846829891, + 0.03985866531729698, + 0.01362018845975399, + 0.007377735339105129, + 0.005931449122726917, + 0.011511248536407948, + 0.018418801948428154, + 0.006669263355433941, + 0.008128941059112549, + 0.00858860183507204, + 0.011125154793262482, + 0.00641136709600687, + 0.00748298317193985, + 0.01049050223082304, + 0.011530532501637936, + 0.0045604826882481575, + 0.006916350685060024, + 0.01331108808517456, + 0.007712278515100479, + 0.0038491683080792427, + 0.0049941763281822205, + 0.005324241705238819, + 0.005442371126264334, + 0.0031010715756565332, + 0.005461978726089001, + 0.008738328702747822, + 0.011262118816375732, + 0.0031916548032313585, + 0.004654544871300459, + 0.007939434610307217, + 0.005235630087554455, + 0.0029568513855338097, + 0.003959581721574068, + 0.005003905855119228, + 0.005680210888385773, + 0.0028185511473566294, + 0.004084194079041481, + 0.004248437471687794, + 0.003957516048103571, + 0.003918119240552187, + 0.0071716830134391785, + 0.016721311956644058, + 0.008461328223347664, + 0.003929806873202324, + 0.004559110384434462, + 0.007966962642967701, + 0.004652516450732946, + 0.002963207196444273, + 0.005549573805183172, + 0.012037127278745174, + 0.008665239438414574, + 0.00475182244554162, + 0.007137260865420103, + 0.009661831893026829, + 0.004803481977432966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.2813894748687744 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871055543422699 + ], + "opposite_logprob_tokens": [ + -1.1371055841445923 + ], + "logprob_reward": -0.3871055543422699, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3871055543422699, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4798526465892792, + 0.04142574220895767, + 0.085693359375, + 0.025594158098101616, + 0.04585397616028786, + 0.04528963938355446, + 0.10492867231369019, + 0.05138786882162094, + 0.09319537878036499, + 0.026778561994433403 + ], + "credit_map": [ + 0.046998005360364914, + 0.04524289071559906, + 0.03454829007387161, + 0.020481042563915253, + 0.17792324721813202, + 0.09400846809148788, + 0.03535301983356476, + 0.025297684594988823, + 0.0023849206045269966, + 0.004038387909531593, + 0.0069020818918943405, + 0.005931094288825989, + 0.005950935184955597, + 0.005523992236703634, + 0.006005050148814917, + 0.00468928087502718, + 0.00819824356585741, + 0.005388750694692135, + 0.02918955869972706, + 0.010675812140107155, + 0.007962914183735847, + 0.0049922289326786995, + 0.00823348667472601, + 0.01105236541479826, + 0.002647863235324621, + 0.002267104107886553, + 0.007423111703246832, + 0.003334824927151203, + 0.002365896012634039, + 0.001498773111961782, + 0.0029435420874506235, + 0.0031130448915064335, + 0.0037415060214698315, + 0.00573717150837183, + 0.00766328489407897, + 0.006417850963771343, + 0.004307515453547239, + 0.006665110122412443, + 0.005892536137253046, + 0.005429006647318602, + 0.0037855692207813263, + 0.005502237007021904, + 0.006739575881510973, + 0.005965511314570904, + 0.0038742355536669493, + 0.004918626509606838, + 0.007387102115899324, + 0.007116783410310745, + 0.007177378516644239, + 0.015595503151416779, + 0.015578676015138626, + 0.019331147894263268, + 0.00837625376880169, + 0.013629992492496967, + 0.015464321710169315, + 0.009775406681001186, + 0.004241593182086945, + 0.007741415873169899, + 0.010755599476397038, + 0.008015834726393223, + 0.004081326071172953, + 0.004772515501827002, + 0.005427131894975901, + 0.006352453026920557, + 0.007822683081030846, + 0.01563359424471855, + 0.01603386364877224, + 0.013054952956736088, + 0.007323698606342077, + 0.007751565892249346, + 0.01434376835823059, + 0.011231256648898125, + 0.0017812344012781978, + 0.0034099523909389973, + 0.006159666460007429, + 0.005171848926693201, + 0.0015511546516790986, + 0.0019422076875343919, + 0.004338534083217382, + 0.0024239635095000267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.3871055543422699 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603488028049469 + ], + "opposite_logprob_tokens": [ + -1.9103487730026245 + ], + "logprob_reward": -0.1603488028049469, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603488028049469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21813873946666718, + 0.06691715121269226, + 0.17384135723114014, + 0.055008478462696075, + 0.09100162982940674, + 0.05639449506998062, + 0.047986529767513275, + 0.09405333548784256, + 0.08574307709932327, + 0.11091519147157669 + ], + "credit_map": [ + 0.022719858214259148, + 0.0145519208163023, + 0.031460441648960114, + 0.020158134400844574, + 0.036379776895046234, + 0.029636740684509277, + 0.025266064330935478, + 0.0379658043384552, + 0.004467880353331566, + 0.006370642688125372, + 0.015432237647473812, + 0.00895051658153534, + 0.007511843927204609, + 0.008028551936149597, + 0.007825113832950592, + 0.008330360986292362, + 0.018940983340144157, + 0.007957088761031628, + 0.06639906764030457, + 0.01934727095067501, + 0.013126016594469547, + 0.009143329225480556, + 0.014304758049547672, + 0.024622851982712746, + 0.006400682497769594, + 0.004374072887003422, + 0.015575222671031952, + 0.0065415832214057446, + 0.005487867631018162, + 0.003383914241567254, + 0.005889467895030975, + 0.007355669047683477, + 0.006534428335726261, + 0.012300566770136356, + 0.018091553822159767, + 0.014745009131729603, + 0.007526902481913567, + 0.008297225460410118, + 0.012530612759292126, + 0.01097533293068409, + 0.004395866300910711, + 0.00884594302624464, + 0.01081226859241724, + 0.008844810537993908, + 0.004985304549336433, + 0.0055124652571976185, + 0.006534797605127096, + 0.006463041994720697, + 0.0025017326697707176, + 0.007933065295219421, + 0.005952590145170689, + 0.011183380149304867, + 0.0037488946691155434, + 0.0035071144811809063, + 0.008564159274101257, + 0.004595595877617598, + 0.007137312088161707, + 0.012886863201856613, + 0.014319531619548798, + 0.01450709905475378, + 0.010946692898869514, + 0.012984178960323334, + 0.010492082685232162, + 0.010779576376080513, + 0.005650936625897884, + 0.010365718975663185, + 0.01610887609422207, + 0.015770776197314262, + 0.01121511124074459, + 0.010697169229388237, + 0.010117069818079472, + 0.005817413330078125, + 0.006364576518535614, + 0.013768638484179974, + 0.022274410352110863, + 0.023173993453383446, + 0.007021979428827763, + 0.013255883939564228, + 0.016032416373491287, + 0.009023289196193218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.1603488028049469 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47425878047943115 + ], + "opposite_logprob_tokens": [ + -0.9742587804794312 + ], + "logprob_reward": -0.47425878047943115, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47425878047943115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14611433446407318, + 0.08942665904760361, + 0.17882892489433289, + 0.0984005406498909, + 0.09954685717821121, + 0.07375049591064453, + 0.053720325231552124, + 0.07896672934293747, + 0.08767106384038925, + 0.09357404708862305 + ], + "credit_map": [ + 0.008615583181381226, + 0.012959344312548637, + 0.019064366817474365, + 0.014836438000202179, + 0.02677825838327408, + 0.014382997527718544, + 0.020366661250591278, + 0.02911069244146347, + 0.008269217796623707, + 0.005104037933051586, + 0.026850270107388496, + 0.01147386059165001, + 0.010308541357517242, + 0.006954843178391457, + 0.010372346267104149, + 0.010093544609844685, + 0.021846096962690353, + 0.0064688255079090595, + 0.07008232921361923, + 0.020442871376872063, + 0.011304428800940514, + 0.008979735895991325, + 0.013372265733778477, + 0.026332370936870575, + 0.012615622021257877, + 0.005025481339544058, + 0.03621778264641762, + 0.01097017154097557, + 0.006095563061535358, + 0.004551352001726627, + 0.007864885963499546, + 0.015059683471918106, + 0.00786632765084505, + 0.01083638146519661, + 0.019884126260876656, + 0.015326382592320442, + 0.009107865393161774, + 0.009082735516130924, + 0.01349632441997528, + 0.013946712017059326, + 0.0060916561633348465, + 0.007379025686532259, + 0.01853630319237709, + 0.00977329257875681, + 0.0065633393824100494, + 0.008500023745000362, + 0.008353282697498798, + 0.00855356827378273, + 0.004614253994077444, + 0.0068674939684569836, + 0.00719135906547308, + 0.009687017649412155, + 0.00421861233189702, + 0.006230559665709734, + 0.007852382957935333, + 0.007058647461235523, + 0.008353840559720993, + 0.0069662840105593204, + 0.011639135889708996, + 0.011436251923441887, + 0.008647755719721317, + 0.01388411596417427, + 0.009191987104713917, + 0.008847354911267757, + 0.005884977988898754, + 0.009258989244699478, + 0.02281266264617443, + 0.017434455454349518, + 0.00560910115018487, + 0.005173893645405769, + 0.013038783334195614, + 0.008458196185529232, + 0.006820950191468, + 0.00990927405655384, + 0.01881999708712101, + 0.01677233725786209, + 0.008889630436897278, + 0.00901536550372839, + 0.014772933907806873, + 0.00857354886829853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.47425878047943115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, right-facing, long shot transports us into a deep-sea world reminiscent of a richly colored classical oil painting. A massive humpback whale swims by with a majestic and deliberate rhythm, its massive frame practically dominating the frame. Its deep blue, almost black skin, stained with white scratches and barnacles from the passage of time, reflects the rich luster of oil paint in the faint, divinely inspired beams of light that filter down from the surface. Its massive, powerful tail fin rises and falls with almost imperceptible grace and stability, each movement propelling it effortlessly through the deep, almost inky Prussian blue waters. Light filters through the water, creating soft, rich gradations of light and dark, reminiscent of the chiaroscuro in a Rembrandt painting. Countless tiny plankton, like suspended golden dust, imbue the gloom with a shimmer of life. The background is an endless abyss, and the whole picture is full of tranquility, solemnity and an epic feeling that transcends time. The texture is heavy, as if every frame is a carefully painted ocean-themed masterpiece.", + "seed": 23137, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_10/index_9133_seed_23137_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.30994212739169597, + "vlm_logprob": -0.30994212739169597, + "vlm_yes_no_margin": 1.8749999761581422, + "vlm_token_credit": -0.30994212739169597, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01572013460099697 + ], + "opposite_logprob_tokens": [ + -4.265720367431641 + ], + "logprob_reward": -0.01572013460099697, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01572013460099697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030426785349845886, + 0.8070766925811768, + 0.02018401399254799, + 0.021823132410645485, + 0.06912857294082642, + 0.00719339307397604, + 0.011144866235554218, + 0.012063240632414818, + 0.007868037559092045, + 0.013091310858726501 + ], + "credit_map": [ + 0.002355551114305854, + 0.0038454600144177675, + 0.0061617386527359486, + 0.003801158629357815, + 0.003737366059795022, + 0.003399172332137823, + 0.004094438627362251, + 0.003031899919733405, + 0.16940896213054657, + 0.043085455894470215, + 0.07829560339450836, + 0.03395208716392517, + 0.15363089740276337, + 0.14435651898384094, + 0.04875894635915756, + 0.13558822870254517, + 0.001927573117427528, + 0.003600951749831438, + 0.0017511177575215697, + 0.0017656321870163083, + 0.0034666520077735186, + 0.0033219484612345695, + 0.002650016453117132, + 0.0017001223750412464, + 0.003420052817091346, + 0.003408248769119382, + 0.0022878823801875114, + 0.0017663592007011175, + 0.0025452093686908484, + 0.0026526416186243296, + 0.0026685071643441916, + 0.003074231557548046, + 0.009800574742257595, + 0.011570703238248825, + 0.004937455989420414, + 0.006316689308732748, + 0.008223719894886017, + 0.011610670015215874, + 0.010385114699602127, + 0.00628364784643054, + 0.0008269799873232841, + 0.0008729919791221619, + 0.0012148291571065784, + 0.0006469098152592778, + 0.0009560981416143477, + 0.000864142959471792, + 0.0009019599528983235, + 0.0009094811393879354, + 0.0012359812390059233, + 0.001382658607326448, + 0.0012109967647120357, + 0.00212480197660625, + 0.0009259277721866965, + 0.0013568061403930187, + 0.0013332223752513528, + 0.0015744718257337809, + 0.0011226929491385818, + 0.0016888576792553067, + 0.0016362963942810893, + 0.0011294662253931165, + 0.002248926553875208, + 0.0012344716815277934, + 0.001504000392742455, + 0.001498528989031911, + 0.0006222960655577481, + 0.001044855103828013, + 0.001110562589019537, + 0.0014316556043922901, + 0.0009597963071428239, + 0.0007826908258721232, + 0.0013676310190930963, + 0.0005485511501319706, + 0.000820752524305135, + 0.00207146187312901, + 0.0021722144447267056, + 0.002133829053491354, + 0.0013782292371615767, + 0.0015488772187381983, + 0.00187794235534966, + 0.001088003977201879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.01572013460099697 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14286576211452484 + ], + "opposite_logprob_tokens": [ + -2.0178656578063965 + ], + "logprob_reward": -0.14286576211452484, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14286576211452484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04517807438969612, + 0.4147968590259552, + 0.034733012318611145, + 0.02321661077439785, + 0.13195914030075073, + 0.013723836280405521, + 0.0437345877289772, + 0.25361451506614685, + 0.020535435527563095, + 0.018507923930883408 + ], + "credit_map": [ + 0.004125279374420643, + 0.004894251003861427, + 0.009362147189676762, + 0.006167800165712833, + 0.0036685247905552387, + 0.004796823486685753, + 0.007664389442652464, + 0.0044988528825342655, + 0.08646757900714874, + 0.020056504756212234, + 0.041011661291122437, + 0.016717420890927315, + 0.07632199674844742, + 0.07683274894952774, + 0.023849673569202423, + 0.07353929430246353, + 0.005480034742504358, + 0.004157050978392363, + 0.0025293920189142227, + 0.0023972350172698498, + 0.008214613422751427, + 0.0058046062476933, + 0.0035940620582550764, + 0.0025560189969837666, + 0.0021890681236982346, + 0.002882974920794368, + 0.002487044082954526, + 0.00198324304074049, + 0.002906763693317771, + 0.0030403633136302233, + 0.003776532830670476, + 0.003950619138777256, + 0.016445184126496315, + 0.02527576871216297, + 0.008179578930139542, + 0.014898676425218582, + 0.015059931203722954, + 0.01883113756775856, + 0.021209169179201126, + 0.012059695087373257, + 0.0013133531901985407, + 0.0022600656375288963, + 0.0018807287560775876, + 0.001457696664147079, + 0.0019550591241568327, + 0.0017153830267488956, + 0.0014838236384093761, + 0.0016577267087996006, + 0.0037126813549548388, + 0.006085200235247612, + 0.005447455681860447, + 0.007443396840244532, + 0.004943574778735638, + 0.005530654452741146, + 0.004045069683343172, + 0.006526554003357887, + 0.018977001309394836, + 0.03501052036881447, + 0.028792476281523705, + 0.03027196228504181, + 0.03790697455406189, + 0.025930631905794144, + 0.021179720759391785, + 0.05554521828889847, + 0.0016033550491556525, + 0.004962638486176729, + 0.0024622264318168163, + 0.002466482575982809, + 0.002911760937422514, + 0.0018871872453019023, + 0.0028496847953647375, + 0.0013921002391725779, + 0.0017783709336072206, + 0.0019342419691383839, + 0.0037525526713579893, + 0.0025061885826289654, + 0.00300434953533113, + 0.0019275075756013393, + 0.0021513982210308313, + 0.0014533151406794786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.14286576211452484 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705792486667633 + ], + "opposite_logprob_tokens": [ + -2.1270580291748047 + ], + "logprob_reward": -0.12705792486667633, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705792486667633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04666783660650253, + 0.4218074679374695, + 0.053439486771821976, + 0.043966468423604965, + 0.026619285345077515, + 0.02967432513833046, + 0.032402656972408295, + 0.12369377166032791, + 0.07262619584798813, + 0.14910241961479187 + ], + "credit_map": [ + 0.003409293945878744, + 0.007121279835700989, + 0.007320265285670757, + 0.006257894914597273, + 0.005298970267176628, + 0.004799810238182545, + 0.008071528747677803, + 0.004388798028230667, + 0.08873084932565689, + 0.022504592314362526, + 0.04079298675060272, + 0.01666724868118763, + 0.08152494579553604, + 0.07580563426017761, + 0.024731218814849854, + 0.07105006277561188, + 0.0043596806935966015, + 0.009657534770667553, + 0.0053563788533210754, + 0.006128096953034401, + 0.0072565325535833836, + 0.007077672053128481, + 0.008573711849749088, + 0.005029883235692978, + 0.005935732740908861, + 0.005793284624814987, + 0.005506289191544056, + 0.004052694886922836, + 0.00467635365203023, + 0.005467401351779699, + 0.006590118166059256, + 0.005944602657109499, + 0.002391410991549492, + 0.0028493115678429604, + 0.001963738352060318, + 0.0028412139508873224, + 0.003954463638365269, + 0.004913461394608021, + 0.004587107803672552, + 0.0031185804400593042, + 0.0032920516096055508, + 0.003585746744647622, + 0.0037630682345479727, + 0.0024562212638556957, + 0.004113204777240753, + 0.005533393006771803, + 0.003619245020672679, + 0.0033113984391093254, + 0.002654408570379019, + 0.004129419103264809, + 0.0033492620568722486, + 0.004899695049971342, + 0.0039374735206365585, + 0.003793156472966075, + 0.0030708706472069025, + 0.006568376440554857, + 0.008115213364362717, + 0.018048157915472984, + 0.013661254197359085, + 0.0160696879029274, + 0.01792498119175434, + 0.012973635457456112, + 0.010217652656137943, + 0.026683207601308823, + 0.007491919212043285, + 0.014275547116994858, + 0.011399423703551292, + 0.01094807218760252, + 0.007604166865348816, + 0.0069689564406871796, + 0.007777406834065914, + 0.006160715129226446, + 0.009731265716254711, + 0.028057141229510307, + 0.016574902459979057, + 0.02063041552901268, + 0.021050238981842995, + 0.016295669600367546, + 0.023687755689024925, + 0.013075049966573715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.12705792486667633 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270359754562378 + ], + "opposite_logprob_tokens": [ + -2.1270360946655273 + ], + "logprob_reward": -0.1270359754562378, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270359754562378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029740091413259506, + 0.35295408964157104, + 0.025752900168299675, + 0.045474324375391006, + 0.1314007043838501, + 0.018177838996052742, + 0.017068877816200256, + 0.34854239225387573, + 0.0140569768846035, + 0.01683177798986435 + ], + "credit_map": [ + 0.0024586492218077183, + 0.004205751232802868, + 0.0045897685922682285, + 0.0043102651834487915, + 0.0029426536057144403, + 0.0026785547379404306, + 0.005225656554102898, + 0.003328795311972499, + 0.07335687428712845, + 0.017558235675096512, + 0.03483637049794197, + 0.014395146630704403, + 0.06631585955619812, + 0.06484825164079666, + 0.020213814452290535, + 0.06142956763505936, + 0.0024987312499433756, + 0.003364281263202429, + 0.002065283479169011, + 0.0020422230008989573, + 0.006726181600242853, + 0.004538414999842644, + 0.002288615331053734, + 0.0022291692439466715, + 0.0032616103999316692, + 0.005739604122936726, + 0.005128959659487009, + 0.004478445276618004, + 0.004427350591868162, + 0.005215803161263466, + 0.005603224039077759, + 0.011619326658546925, + 0.007573015987873077, + 0.020639153197407722, + 0.013861543498933315, + 0.012437812983989716, + 0.014042694121599197, + 0.0157377477735281, + 0.01891455054283142, + 0.028194203972816467, + 0.001606276840902865, + 0.001985058421269059, + 0.0035608778707683086, + 0.0015877613332122564, + 0.0030738231725990772, + 0.00221709324978292, + 0.0017555283848196268, + 0.002391418209299445, + 0.00125921203289181, + 0.0024040231946855783, + 0.0021465341560542583, + 0.0018133677076548338, + 0.0028262732084840536, + 0.0022811181843280792, + 0.0015221047215163708, + 0.0028162463568150997, + 0.02467981167137623, + 0.04621008411049843, + 0.04463021084666252, + 0.047102149575948715, + 0.03630410507321358, + 0.04131622985005379, + 0.02823742851614952, + 0.0800623744726181, + 0.0013004894135519862, + 0.002793473657220602, + 0.001934674452058971, + 0.0018839177209883928, + 0.0018717263592407107, + 0.0013098778435960412, + 0.0018974317936226726, + 0.001065386226400733, + 0.0013908410910516977, + 0.0021507125347852707, + 0.002439526841044426, + 0.003323059529066086, + 0.002263587899506092, + 0.0015318694058805704, + 0.0019778849091380835, + 0.0017542969435453415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.1270359754562378 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.137030839920044 + ], + "opposite_logprob_tokens": [ + -0.38703083992004395 + ], + "logprob_reward": -1.137030839920044, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.137030839920044, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04644714295864105, + 0.4712992310523987, + 0.026978693902492523, + 0.01020418107509613, + 0.021191906183958054, + 0.024159034714102745, + 0.049338843673467636, + 0.31902846693992615, + 0.018346348777413368, + 0.013006121851503849 + ], + "credit_map": [ + 0.0050118789076805115, + 0.003104205010458827, + 0.007623955141752958, + 0.004938275087624788, + 0.004244422540068626, + 0.005931945517659187, + 0.006634972058236599, + 0.008957487531006336, + 0.09892341494560242, + 0.024866895750164986, + 0.04538431391119957, + 0.020529793575406075, + 0.09057807177305222, + 0.0849907398223877, + 0.027586327865719795, + 0.07843969017267227, + 0.0033861605916172266, + 0.0033857575617730618, + 0.0025526597164571285, + 0.002378704259172082, + 0.005422120448201895, + 0.003907209727913141, + 0.0030077206902205944, + 0.0029383597429841757, + 0.0017040862003341317, + 0.0011544288136065006, + 0.0013798789586871862, + 0.0009348325547762215, + 0.0012546856887638569, + 0.0012193189468234777, + 0.0011511440388858318, + 0.001405805116519332, + 0.0026435505133122206, + 0.0019189424347132444, + 0.0027390068862587214, + 0.0026491673197597265, + 0.002547050593420863, + 0.0034483338240534067, + 0.002776900539174676, + 0.002468954771757126, + 0.0022365443874150515, + 0.0034746478777378798, + 0.0036089764907956123, + 0.0023417214397341013, + 0.0034908077213913202, + 0.00274058454670012, + 0.0031977626495063305, + 0.0030679889023303986, + 0.004878150299191475, + 0.007021887693554163, + 0.005899690091609955, + 0.007945764809846878, + 0.0051674447022378445, + 0.005691174417734146, + 0.003958731424063444, + 0.008775996044278145, + 0.022854136303067207, + 0.04230174049735069, + 0.03618461266160011, + 0.04215215891599655, + 0.039790935814380646, + 0.03469249606132507, + 0.02480264939367771, + 0.07624972611665726, + 0.0015237982152029872, + 0.0023536989465355873, + 0.0029647587798535824, + 0.0035469054710119963, + 0.0014287916710600257, + 0.0015703766839578748, + 0.002249121433123946, + 0.002708897227421403, + 0.0010577506618574262, + 0.0010648634051904082, + 0.002681142184883356, + 0.0019185537239536643, + 0.001621329691261053, + 0.0012870922219008207, + 0.0017318533500656486, + 0.0016435361467301846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -1.137030839920044 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "slow, right-facing, long shot transports us into a deep-sea world reminiscent of a richly colored classical oil painting. A massive humpback whale swims by with a majestic and deliberate rhythm, its massive frame practically dominating the frame. Its deep blue, almost black skin, stained with white scratches and barnacles from the passage of time, reflects the rich luster of oil paint in the faint, divinely inspired beams of light that filter down from the surface. Its massive, powerful tail fin rises and falls with almost imperceptible grace and stability, each movement propelling it effortlessly through the deep, almost inky Prussian blue waters. Light filters through the water, creating soft, rich gradations of light and dark, reminiscent of the chiaroscuro in a Rembrandt painting. Countless tiny plankton, like suspended golden dust, imbue the gloom with a shimmer of life. The background is an endless abyss, and the whole picture is full of tranquility, solemnity and an epic feeling that transcends time. The texture is heavy, as if every frame is a carefully painted ocean-themed masterpiece.", + "seed": 23138, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_10/index_9133_seed_23138_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.5155632801353931, + "vlm_logprob": -0.5155632801353931, + "vlm_yes_no_margin": 0.8499999880790711, + "vlm_token_credit": -0.5155632801353931, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.057142142206430435 + ], + "opposite_logprob_tokens": [ + -2.9321422576904297 + ], + "logprob_reward": -0.057142142206430435, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.057142142206430435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2993400990962982, + 0.09308213740587234, + 0.11466584354639053, + 0.045612990856170654, + 0.14719489216804504, + 0.08531031012535095, + 0.1101057380437851, + 0.0548955574631691, + 0.026079844683408737, + 0.023712623864412308 + ], + "credit_map": [ + 0.03865453600883484, + 0.04193120822310448, + 0.02589249610900879, + 0.05264788493514061, + 0.04740134999155998, + 0.030029648914933205, + 0.026684816926717758, + 0.03609813377261162, + 0.006156042218208313, + 0.009451918303966522, + 0.006789592560380697, + 0.011107709258794785, + 0.018927184864878654, + 0.009627366438508034, + 0.016494043171405792, + 0.014528272673487663, + 0.010321161709725857, + 0.009774449281394482, + 0.016843551769852638, + 0.015492701902985573, + 0.01955871656537056, + 0.013519112020730972, + 0.011408226564526558, + 0.017747912555933, + 0.00486770598217845, + 0.005292537622153759, + 0.004435205366462469, + 0.00786223541945219, + 0.007744122762233019, + 0.005076105240732431, + 0.005369230639189482, + 0.004965844564139843, + 0.008933309465646744, + 0.010406107641756535, + 0.014477266930043697, + 0.015199173241853714, + 0.016772745177149773, + 0.01711631380021572, + 0.0353463739156723, + 0.028943590819835663, + 0.008989629335701466, + 0.008556685410439968, + 0.006675729528069496, + 0.012254090048372746, + 0.011518992483615875, + 0.010792659595608711, + 0.01687530428171158, + 0.009647209197282791, + 0.006661978084594011, + 0.012136437930166721, + 0.011561917141079903, + 0.008668545633554459, + 0.01619008742272854, + 0.014687156304717064, + 0.015064382925629616, + 0.0251352246850729, + 0.006037912331521511, + 0.0072483704425394535, + 0.006897404324263334, + 0.00544616999104619, + 0.007289353758096695, + 0.0069940160028636456, + 0.00628699641674757, + 0.008695331402122974, + 0.0036009554751217365, + 0.0034257303923368454, + 0.00320226582698524, + 0.0026544765569269657, + 0.00395282544195652, + 0.0032995352521538734, + 0.002714170841500163, + 0.0032298830337822437, + 0.0028520971536636353, + 0.0017007424030452967, + 0.0026241957675665617, + 0.0016361633315682411, + 0.005500488448888063, + 0.003918151371181011, + 0.0017232476966455579, + 0.0037575361784547567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.057142142206430435 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31353455781936646 + ], + "opposite_logprob_tokens": [ + -1.3135344982147217 + ], + "logprob_reward": -0.31353455781936646, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31353455781936646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19479331374168396, + 0.1379697322845459, + 0.1437745988368988, + 0.09194126725196838, + 0.12097755074501038, + 0.0612533800303936, + 0.023743927478790283, + 0.13089434802532196, + 0.04170555621385574, + 0.05294634401798248 + ], + "credit_map": [ + 0.031875740736722946, + 0.02367822639644146, + 0.019958874210715294, + 0.03498227521777153, + 0.02191912941634655, + 0.01914656162261963, + 0.019982730969786644, + 0.023249758407473564, + 0.01542860921472311, + 0.012989831157028675, + 0.010089030489325523, + 0.011119731701910496, + 0.029819410294294357, + 0.018883317708969116, + 0.02181725576519966, + 0.017822539433836937, + 0.013265014626085758, + 0.01167797576636076, + 0.013953267596662045, + 0.01515553891658783, + 0.031994737684726715, + 0.016243917867541313, + 0.018612723797559738, + 0.022871410474181175, + 0.010383605025708675, + 0.01155547983944416, + 0.00937765184789896, + 0.01092104148119688, + 0.01467917300760746, + 0.011089893989264965, + 0.008868754841387272, + 0.015065668150782585, + 0.008393223397433758, + 0.009150858037173748, + 0.010577062144875526, + 0.011892183683812618, + 0.014436337165534496, + 0.014450457878410816, + 0.026596324518322945, + 0.025481102988123894, + 0.004867202136665583, + 0.0060187350027263165, + 0.004010586068034172, + 0.004997233394533396, + 0.01436584908515215, + 0.008495926856994629, + 0.0089256064966321, + 0.009572234936058521, + 0.0019696615636348724, + 0.0034549066331237555, + 0.0027128178626298904, + 0.002326125046238303, + 0.003975406754761934, + 0.0029575498774647713, + 0.00356534356251359, + 0.0027821133844554424, + 0.014273200184106827, + 0.013852229341864586, + 0.009711298160254955, + 0.009438740089535713, + 0.017590807750821114, + 0.017282377928495407, + 0.012305499985814095, + 0.036440182477235794, + 0.006545880809426308, + 0.006284418515861034, + 0.004914579913020134, + 0.004013280849903822, + 0.005126821808516979, + 0.004561012610793114, + 0.004438421223312616, + 0.0058211381547153, + 0.002345004351809621, + 0.0029027198906987906, + 0.005580356810241938, + 0.003981032408773899, + 0.012779111973941326, + 0.008297323249280453, + 0.004711890127509832, + 0.0123489024117589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.31353455781936646 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7579282522201538 + ], + "opposite_logprob_tokens": [ + -0.6329282522201538 + ], + "logprob_reward": -0.7579282522201538, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7579282522201538, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05899371579289436, + 0.317468523979187, + 0.07924062013626099, + 0.04548565298318863, + 0.12041855603456497, + 0.09397689998149872, + 0.11186434328556061, + 0.1172611191868782, + 0.0321597121655941, + 0.023130852729082108 + ], + "credit_map": [ + 0.00926519650965929, + 0.00523885665461421, + 0.005721873603761196, + 0.008381008170545101, + 0.0065598576329648495, + 0.005623540375381708, + 0.009832942858338356, + 0.008370441384613514, + 0.03093692474067211, + 0.0399763360619545, + 0.02053021639585495, + 0.026625750586390495, + 0.08040567487478256, + 0.04306483268737793, + 0.030433528125286102, + 0.045495253056287766, + 0.009619210846722126, + 0.007307491265237331, + 0.007833638228476048, + 0.009245690889656544, + 0.01624377630650997, + 0.009419414214789867, + 0.008719107136130333, + 0.010852289386093616, + 0.005354434251785278, + 0.004792514722794294, + 0.004717049188911915, + 0.0071237049996852875, + 0.00589315639808774, + 0.005082840099930763, + 0.0052459402941167355, + 0.00727600883692503, + 0.009872970171272755, + 0.010131134651601315, + 0.009912711568176746, + 0.013149124570190907, + 0.013698704540729523, + 0.012274441309273243, + 0.02941048890352249, + 0.021968992426991463, + 0.007436720654368401, + 0.009202500805258751, + 0.006966118700802326, + 0.008403202518820763, + 0.019827086478471756, + 0.011945770122110844, + 0.01275319792330265, + 0.017442304641008377, + 0.004914230201393366, + 0.01274290680885315, + 0.011088459752500057, + 0.009746679104864597, + 0.018352625891566277, + 0.01535054761916399, + 0.015538801439106464, + 0.024130096659064293, + 0.011843958869576454, + 0.01357696671038866, + 0.01387950498610735, + 0.011008437722921371, + 0.01385090034455061, + 0.015425527468323708, + 0.012346436269581318, + 0.025329383090138435, + 0.002574954414740205, + 0.0031267176382243633, + 0.004333258606493473, + 0.0031324054580181837, + 0.0037239445373415947, + 0.0041799526661634445, + 0.004655003547668457, + 0.006433472968637943, + 0.0037922326009720564, + 0.002362634288147092, + 0.0032634951639920473, + 0.001957574626430869, + 0.002056780271232128, + 0.002625530119985342, + 0.002424051985144615, + 0.004648555535823107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.7579282522201538 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559377551078796 + ], + "opposite_logprob_tokens": [ + -1.6005938053131104 + ], + "logprob_reward": -0.22559377551078796, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559377551078796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09908981621265411, + 0.06208342686295509, + 0.18950560688972473, + 0.1208239421248436, + 0.041330158710479736, + 0.09367028623819351, + 0.03610369563102722, + 0.1574549376964569, + 0.04548615589737892, + 0.15445201098918915 + ], + "credit_map": [ + 0.013846537098288536, + 0.01064069103449583, + 0.01077998336404562, + 0.014873763546347618, + 0.014840196818113327, + 0.010318778455257416, + 0.010465793311595917, + 0.013324075378477573, + 0.006234331056475639, + 0.007461392320692539, + 0.005726194474846125, + 0.0072527797892689705, + 0.010096888989210129, + 0.006366108078509569, + 0.010010605677962303, + 0.008935119025409222, + 0.01586681418120861, + 0.014209671877324581, + 0.015551401302218437, + 0.014021351933479309, + 0.04665990546345711, + 0.025118369609117508, + 0.018957212567329407, + 0.0391208678483963, + 0.01406936813145876, + 0.014908073469996452, + 0.013847021386027336, + 0.01491507887840271, + 0.018550246953964233, + 0.014281217940151691, + 0.011748448945581913, + 0.01850447803735733, + 0.004642445594072342, + 0.00482004601508379, + 0.004284133668988943, + 0.004455334972590208, + 0.0054216752760112286, + 0.004799396265298128, + 0.007167524192482233, + 0.005739599000662565, + 0.005091740284115076, + 0.00908517837524414, + 0.007374800741672516, + 0.007074314635246992, + 0.02287914976477623, + 0.012739365920424461, + 0.013371286913752556, + 0.01605444774031639, + 0.0025231458712369204, + 0.005279026925563812, + 0.0038878226187080145, + 0.00250078528188169, + 0.006602872163057327, + 0.004749831277877092, + 0.0042089736089110374, + 0.006351232063025236, + 0.015207287855446339, + 0.013743233866989613, + 0.010135713964700699, + 0.011173166334629059, + 0.026502326130867004, + 0.021312981843948364, + 0.0153165552765131, + 0.04406367242336273, + 0.003351711668074131, + 0.003363276133313775, + 0.004829990211874247, + 0.003929778467863798, + 0.011729346588253975, + 0.007414115127176046, + 0.003568523097783327, + 0.007299412507563829, + 0.0075633954256772995, + 0.009931274689733982, + 0.016336005181074142, + 0.011284870095551014, + 0.040024105459451675, + 0.025129517540335655, + 0.014719021506607533, + 0.029463812708854675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 9 + ], + "selected_reward": -0.22559377551078796 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.223617672920227 + ], + "opposite_logprob_tokens": [ + -0.34861764311790466 + ], + "logprob_reward": -1.223617672920227, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.223617672920227, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1383674293756485, + 0.12871432304382324, + 0.11776050180196762, + 0.08187929540872574, + 0.3180811405181885, + 0.09273936599493027, + 0.0411989688873291, + 0.036298155784606934, + 0.02041594497859478, + 0.024544868618249893 + ], + "credit_map": [ + 0.01081240363419056, + 0.015145055949687958, + 0.012537679634988308, + 0.015157115645706654, + 0.03842007368803024, + 0.016504507511854172, + 0.012007581070065498, + 0.017783010378479958, + 0.014804888516664505, + 0.01294097863137722, + 0.011080863885581493, + 0.013442267663776875, + 0.025147443637251854, + 0.012989142909646034, + 0.02078351005911827, + 0.017525237053632736, + 0.010820094496011734, + 0.009931989945471287, + 0.008981321938335896, + 0.00866511557251215, + 0.028757784515619278, + 0.015597413294017315, + 0.011637908406555653, + 0.02336888015270233, + 0.009552747011184692, + 0.0101234819740057, + 0.008470533415675163, + 0.009323298931121826, + 0.012955812737345695, + 0.010315414518117905, + 0.007809508591890335, + 0.01332850195467472, + 0.02301599830389023, + 0.024647181853652, + 0.03384160250425339, + 0.03393440321087837, + 0.03835763409733772, + 0.03268495202064514, + 0.07652775198221207, + 0.055071622133255005, + 0.00496419845148921, + 0.008471477776765823, + 0.00831945426762104, + 0.00901669729501009, + 0.018108882009983063, + 0.010791178792715073, + 0.015152279287576675, + 0.01791520230472088, + 0.002669320208951831, + 0.006087312009185553, + 0.0046180072240531445, + 0.0036520089488476515, + 0.0065882448107004166, + 0.005244532134383917, + 0.004676944110542536, + 0.007662598509341478, + 0.003352582920342684, + 0.003630072111263871, + 0.004648177418857813, + 0.0038796644657850266, + 0.004085765220224857, + 0.004820544738322496, + 0.004357951693236828, + 0.007523398380726576, + 0.002708013402298093, + 0.0017001372762024403, + 0.002263157395645976, + 0.0018113133264705539, + 0.002577591687440872, + 0.0026931932661682367, + 0.0025563060771673918, + 0.004106233362108469, + 0.0021878976840525866, + 0.0018813415663316846, + 0.0031314913649111986, + 0.002026595175266266, + 0.006704402156174183, + 0.004320342093706131, + 0.002141904318705201, + 0.002150895306840539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -1.223617672920227 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "No" + ] + }, + { + "prompt": "slow, right-facing, long shot transports us into a deep-sea world reminiscent of a richly colored classical oil painting. A massive humpback whale swims by with a majestic and deliberate rhythm, its massive frame practically dominating the frame. Its deep blue, almost black skin, stained with white scratches and barnacles from the passage of time, reflects the rich luster of oil paint in the faint, divinely inspired beams of light that filter down from the surface. Its massive, powerful tail fin rises and falls with almost imperceptible grace and stability, each movement propelling it effortlessly through the deep, almost inky Prussian blue waters. Light filters through the water, creating soft, rich gradations of light and dark, reminiscent of the chiaroscuro in a Rembrandt painting. Countless tiny plankton, like suspended golden dust, imbue the gloom with a shimmer of life. The background is an endless abyss, and the whole picture is full of tranquility, solemnity and an epic feeling that transcends time. The texture is heavy, as if every frame is a carefully painted ocean-themed masterpiece.", + "seed": 23139, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_10/index_9133_seed_23139_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.4124704018235207, + "vlm_logprob": -0.4124704018235207, + "vlm_yes_no_margin": 1.025, + "vlm_token_credit": -0.4124704018235207, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11481461673974991 + ], + "opposite_logprob_tokens": [ + -2.239814519882202 + ], + "logprob_reward": -0.11481461673974991, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11481461673974991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3129303455352783, + 0.0273644607514143, + 0.049314599484205246, + 0.17017875611782074, + 0.08626385033130646, + 0.034429240971803665, + 0.04543270170688629, + 0.04240332916378975, + 0.1104697734117508, + 0.12121293693780899 + ], + "credit_map": [ + 0.02762967348098755, + 0.04682071879506111, + 0.03966328129172325, + 0.04102722927927971, + 0.050144072622060776, + 0.044538434594869614, + 0.03259478881955147, + 0.030512137338519096, + 0.002782798372209072, + 0.0030723426025360823, + 0.00329136336222291, + 0.002868295181542635, + 0.004087279085069895, + 0.0035705550108104944, + 0.003936965484172106, + 0.0037548618856817484, + 0.004152558743953705, + 0.006020326167345047, + 0.004858460742980242, + 0.005379638634622097, + 0.011567694135010242, + 0.007238509599119425, + 0.005170739721506834, + 0.004926672205328941, + 0.012231932953000069, + 0.023947110399603844, + 0.02617739886045456, + 0.022981682792305946, + 0.030495811253786087, + 0.02652587927877903, + 0.012219836935400963, + 0.015599108301103115, + 0.006799185182899237, + 0.009018310345709324, + 0.020537691190838814, + 0.01138695701956749, + 0.013828820548951626, + 0.008262407034635544, + 0.010461138561367989, + 0.005969340447336435, + 0.0032524392008781433, + 0.005351893603801727, + 0.004861781839281321, + 0.005220532417297363, + 0.004675230477005243, + 0.004748460836708546, + 0.0032857516780495644, + 0.0030331553425639868, + 0.004285378847271204, + 0.004282520152628422, + 0.015063048340380192, + 0.008065777830779552, + 0.0030582332983613014, + 0.0028034078422933817, + 0.004119377117604017, + 0.003754958976060152, + 0.005134853534400463, + 0.0059079513885080814, + 0.0074891624972224236, + 0.0059657772071659565, + 0.004253079649060965, + 0.004555802326649427, + 0.004777989350259304, + 0.0043187132105231285, + 0.013402161188423634, + 0.016540834680199623, + 0.013089153915643692, + 0.015339452773332596, + 0.014848296530544758, + 0.01023712009191513, + 0.013608690351247787, + 0.01340406946837902, + 0.013107603415846825, + 0.014424189925193787, + 0.018518513068556786, + 0.013621699064970016, + 0.014129270799458027, + 0.015867939218878746, + 0.01509624533355236, + 0.016447482630610466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.11481461673974991 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20164546370506287 + ], + "opposite_logprob_tokens": [ + -1.7016454935073853 + ], + "logprob_reward": -0.20164546370506287, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20164546370506287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33475813269615173, + 0.07365452498197556, + 0.06330574303865433, + 0.14529666304588318, + 0.2351454347372055, + 0.015891797840595245, + 0.019631456583738327, + 0.04789281263947487, + 0.028803233057260513, + 0.03562023118138313 + ], + "credit_map": [ + 0.032326482236385345, + 0.04968006908893585, + 0.04707582667469978, + 0.0378454215824604, + 0.05534099414944649, + 0.04714121297001839, + 0.033934880048036575, + 0.0314132422208786, + 0.007129915524274111, + 0.007495481986552477, + 0.008776836097240448, + 0.01437009871006012, + 0.007559863850474358, + 0.010071363300085068, + 0.006796914618462324, + 0.01145404577255249, + 0.004840932320803404, + 0.006639846600592136, + 0.006619331426918507, + 0.01107317116111517, + 0.007735853083431721, + 0.008155875839293003, + 0.006863209884613752, + 0.011377513408660889, + 0.011393263004720211, + 0.01864277571439743, + 0.02052600122988224, + 0.01724059134721756, + 0.02149835415184498, + 0.029365694150328636, + 0.011715013533830643, + 0.014914964325726032, + 0.017464004456996918, + 0.0255773663520813, + 0.053428806364536285, + 0.03310554102063179, + 0.0364769771695137, + 0.022243862971663475, + 0.028464285656809807, + 0.01838456466794014, + 0.0011510991025716066, + 0.0019676971714943647, + 0.002278151223435998, + 0.0026721719186753035, + 0.0020952625200152397, + 0.0017134018708020449, + 0.0022228448651731014, + 0.0017911677714437246, + 0.002524044830352068, + 0.0027121067978441715, + 0.003165746573358774, + 0.0033319902140647173, + 0.002039691898971796, + 0.0017310722032561898, + 0.0024429894983768463, + 0.0016838134033605456, + 0.005394447594881058, + 0.00599268265068531, + 0.009224840439856052, + 0.006486415863037109, + 0.006418284960091114, + 0.004533312749117613, + 0.0054098572582006454, + 0.004432970192283392, + 0.0020926513243466616, + 0.002612514654174447, + 0.0052971383556723595, + 0.004829860758036375, + 0.0028323198202997446, + 0.003661356633529067, + 0.004159771371632814, + 0.0033176203723996878, + 0.0032475353218615055, + 0.005099279806017876, + 0.006428874097764492, + 0.004544625524431467, + 0.003967810422182083, + 0.0043847947381436825, + 0.0035957114305347204, + 0.0043515958823263645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.20164546370506287 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289557635784149 + ], + "opposite_logprob_tokens": [ + -1.0539557933807373 + ], + "logprob_reward": -0.4289557635784149, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289557635784149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08875596523284912, + 0.11080174893140793, + 0.20829062163829803, + 0.06593813002109528, + 0.22135937213897705, + 0.047251246869564056, + 0.08008403331041336, + 0.05819712579250336, + 0.055779438465833664, + 0.06354238837957382 + ], + "credit_map": [ + 0.008624355308711529, + 0.0069425152614712715, + 0.012088848277926445, + 0.01191049162298441, + 0.010366504080593586, + 0.012596041895449162, + 0.012851567938923836, + 0.013375646434724331, + 0.009484941139817238, + 0.012291209772229195, + 0.01809840276837349, + 0.019043168053030968, + 0.013449734076857567, + 0.013085159473121166, + 0.01200942788273096, + 0.013339710421860218, + 0.017093196511268616, + 0.022461337968707085, + 0.023101864382624626, + 0.032161686569452286, + 0.03432055562734604, + 0.025029778480529785, + 0.019879542291164398, + 0.0342426598072052, + 0.005837780889123678, + 0.007668318692594767, + 0.012941516935825348, + 0.0065086111426353455, + 0.010345724411308765, + 0.010465657338500023, + 0.00544087216258049, + 0.006729649379849434, + 0.024787498638033867, + 0.03126274049282074, + 0.031192760914564133, + 0.030155787244439125, + 0.03444690257310867, + 0.020629538223147392, + 0.021867316216230392, + 0.027016818523406982, + 0.00455824937671423, + 0.005102976690977812, + 0.009262636303901672, + 0.0057874396443367004, + 0.008888281881809235, + 0.004858710337430239, + 0.005435916129499674, + 0.003357039764523506, + 0.011391899548470974, + 0.009518648497760296, + 0.014259416610002518, + 0.011625494807958603, + 0.007416995242238045, + 0.009435465559363365, + 0.008851893246173859, + 0.007584214210510254, + 0.0064215580932796, + 0.0075064171105623245, + 0.00885813869535923, + 0.006485056132078171, + 0.006483340170234442, + 0.009505219757556915, + 0.0057900212705135345, + 0.0071473740972578526, + 0.00533438241109252, + 0.008793284185230732, + 0.008391115814447403, + 0.008087311871349812, + 0.006254719104617834, + 0.006096351891756058, + 0.006406507454812527, + 0.0064157648012042046, + 0.005256114527583122, + 0.008947213180363178, + 0.010453975759446621, + 0.006962314248085022, + 0.007602449506521225, + 0.009238922037184238, + 0.008034586906433105, + 0.007046807557344437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.4289557635784149 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989304661750793 + ], + "opposite_logprob_tokens": [ + -1.8048930168151855 + ], + "logprob_reward": -0.17989304661750793, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989304661750793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1501869261264801, + 0.07581410557031631, + 0.11333717405796051, + 0.06933825463056564, + 0.12485992908477783, + 0.14276164770126343, + 0.05731470510363579, + 0.06723278760910034, + 0.11863743513822556, + 0.08051704615354538 + ], + "credit_map": [ + 0.01460726372897625, + 0.017278147861361504, + 0.024151302874088287, + 0.020889610052108765, + 0.018406162038445473, + 0.024239446967840195, + 0.015356142073869705, + 0.01525884959846735, + 0.007383633870631456, + 0.007239908911287785, + 0.012111100368201733, + 0.008497070521116257, + 0.01058128010481596, + 0.010842891409993172, + 0.0115519929677248, + 0.007606229744851589, + 0.008424353785812855, + 0.01044398546218872, + 0.011558731086552143, + 0.024803712964057922, + 0.013396709226071835, + 0.010475671850144863, + 0.011889209970831871, + 0.0223447997123003, + 0.006366862915456295, + 0.0077189235016703606, + 0.011516689322888851, + 0.008431156165897846, + 0.01046721637248993, + 0.009135637432336807, + 0.006649693474173546, + 0.00905207172036171, + 0.010514701716601849, + 0.016130076721310616, + 0.02124618925154209, + 0.023914022371172905, + 0.01761304959654808, + 0.011436847038567066, + 0.013465498574078083, + 0.01053953729569912, + 0.012863692827522755, + 0.015539780259132385, + 0.020234573632478714, + 0.021270547062158585, + 0.02757980115711689, + 0.018079174682497978, + 0.016196055337786674, + 0.010998022742569447, + 0.005805847700685263, + 0.005213397089391947, + 0.009451043792068958, + 0.008630373515188694, + 0.0068471492268145084, + 0.008463218808174133, + 0.008004339411854744, + 0.004899333696812391, + 0.007998387329280376, + 0.007648117374628782, + 0.011695361696183681, + 0.008778993971645832, + 0.008313124999403954, + 0.009030960500240326, + 0.006532733794301748, + 0.007235103752464056, + 0.009676910005509853, + 0.014021639712154865, + 0.018928902223706245, + 0.017194218933582306, + 0.011145067401230335, + 0.012770838104188442, + 0.017474263906478882, + 0.01742558367550373, + 0.007216684520244598, + 0.009491619653999805, + 0.01157599687576294, + 0.00850777979940176, + 0.010399792343378067, + 0.011514425277709961, + 0.009970974177122116, + 0.01183976512402296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.17989304661750793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive humpback whale swimming in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the whale's skin deep blue or black with visible white scratches and barnacles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's tail fin visible rising and falling to propel it through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny plankton visible as suspended golden dust in the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style that of a richly colored classical oil painting with heavy texture?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370431184768677 + ], + "opposite_logprob_tokens": [ + -0.38704314827919006 + ], + "logprob_reward": -1.1370431184768677, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370431184768677, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2486105114221573, + 0.1637202352285385, + 0.10974624752998352, + 0.08966365456581116, + 0.15346594154834747, + 0.050196241587400436, + 0.048121556639671326, + 0.04590016230940819, + 0.049276210367679596, + 0.04129919037222862 + ], + "credit_map": [ + 0.01819898933172226, + 0.029866835102438927, + 0.036159686744213104, + 0.03823915496468544, + 0.03479161486029625, + 0.04012228175997734, + 0.02276422269642353, + 0.028467737138271332, + 0.014752646908164024, + 0.018625779077410698, + 0.02646646462380886, + 0.017155742272734642, + 0.02632172964513302, + 0.025247910991311073, + 0.019456079229712486, + 0.015693897381424904, + 0.00893101841211319, + 0.008732480928301811, + 0.010238423012197018, + 0.016372935846447945, + 0.009220492094755173, + 0.014959797263145447, + 0.014247944578528404, + 0.027043171226978302, + 0.0058344826102256775, + 0.007886760868132114, + 0.015516120940446854, + 0.011294694617390633, + 0.010670771822333336, + 0.013209933415055275, + 0.0077979969792068005, + 0.0174529068171978, + 0.01402549259364605, + 0.021298112347722054, + 0.027699362486600876, + 0.01964747905731201, + 0.025578157976269722, + 0.014253339730203152, + 0.018453162163496017, + 0.01251085102558136, + 0.006313663441687822, + 0.005177115090191364, + 0.009733135811984539, + 0.007349728606641293, + 0.0064801499247550964, + 0.0040925368666648865, + 0.005658820271492004, + 0.005391094367951155, + 0.005654048174619675, + 0.003430925542488694, + 0.012413615360856056, + 0.008401051163673401, + 0.003348811063915491, + 0.0045128571800887585, + 0.006258364766836166, + 0.004101887810975313, + 0.004849019926041365, + 0.005732260644435883, + 0.007191132288426161, + 0.005817093420773745, + 0.006331659387797117, + 0.005823249462991953, + 0.00586191238835454, + 0.004293838981539011, + 0.004216945730149746, + 0.006699644029140472, + 0.007970238104462624, + 0.007123335264623165, + 0.004884170368313789, + 0.00586795574054122, + 0.006890894845128059, + 0.005623030476272106, + 0.005142170004546642, + 0.00453827902674675, + 0.005858723074197769, + 0.004392748232930899, + 0.005121944472193718, + 0.004543978720903397, + 0.006020473316311836, + 0.005680878181010485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -1.1370431184768677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_11/index_tensor([5175, 5175])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_11/index_tensor([5175, 5175])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d64f8e0caa70c1d712ee2f66d12a9d0291fa06f1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_11/index_tensor([5175, 5175])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "As the camera executes a slow, perfectly steady tracking motion, a majestic extreme wide shot unfolds, gliding parallel to a massive, whale-like, semi-rigid airship, drifting silently across a breathtakingly vast, pastel-colored sky. The airship's shape is organic and monumental, reminiscent of a giant blue whale, with a broad, rounded nose and elegantly tapered tail. Its skin is a seamless, high-tech fabric of pearlescent off-white, beneath which the subtle outlines of its internal structural ribs are faintly visible. The entire scene is bathed in the warm golden glow of a low, off-camera sun, which caresses the airship's massive, curved flanks, creating a soft sheen on its fabric and casting long, gentle shadows that emphasize its organic contours. The background is a masterpiece of soft colors, the sky melting into a gentle gradient of pale rose, creamy orange, and gentle lavender, dotted with wispy, feathery clouds. The airship floats forward with an almost imperceptible serene grace, its immense scale creating a profound sense of peace in the quiet, still air, while the camera’s synchronized, deliberate tracking movements enhance the feeling of a silent, dreamlike journey, creating an extremely soothing and awe-inspiring atmosphere.", + "seed": 428, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_11/index_5175_seed_428_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.047996940463781355, + "vlm_logprob": -0.047996940463781355, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.047996940463781355, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02522406354546547 + ], + "opposite_logprob_tokens": [ + -3.775223970413208 + ], + "logprob_reward": -0.02522406354546547, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02522406354546547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2148752510547638, + 0.09420251846313477, + 0.12278451025485992, + 0.08241578936576843, + 0.047391124069690704, + 0.04453060030937195, + 0.030401581898331642, + 0.04848679155111313, + 0.043004654347896576, + 0.2719072103500366 + ], + "credit_map": [ + 0.01586991176009178, + 0.04076625406742096, + 0.022407058626413345, + 0.03501386195421219, + 0.03416382148861885, + 0.02405366487801075, + 0.025518780574202538, + 0.017081884667277336, + 0.007373040076345205, + 0.021000245586037636, + 0.010137920267879963, + 0.012295966036617756, + 0.012936968356370926, + 0.015889203175902367, + 0.007424267940223217, + 0.007144896313548088, + 0.007024102378636599, + 0.017933446913957596, + 0.009523546323180199, + 0.008402818813920021, + 0.01829739287495613, + 0.03909066319465637, + 0.00560063635930419, + 0.016911892220377922, + 0.005861001554876566, + 0.009248130954802036, + 0.009005769155919552, + 0.008348803967237473, + 0.015426703728735447, + 0.014855609275400639, + 0.006582478526979685, + 0.013087284751236439, + 0.0034971782006323338, + 0.0057081240229308605, + 0.004107479937374592, + 0.004805366974323988, + 0.011055255308747292, + 0.00956748891621828, + 0.003609950188547373, + 0.005040278192609549, + 0.007636299356818199, + 0.007052398752421141, + 0.003393664723262191, + 0.004912506323307753, + 0.005053649190813303, + 0.009827195666730404, + 0.003382175462320447, + 0.003272706177085638, + 0.0034531725104898214, + 0.004278607200831175, + 0.0024364946875721216, + 0.004838107153773308, + 0.00479298597201705, + 0.005562751553952694, + 0.0022154750768095255, + 0.002823985880240798, + 0.011937716044485569, + 0.005080082919448614, + 0.0029216741677373648, + 0.004032660275697708, + 0.005761383567005396, + 0.011484947986900806, + 0.0026126322336494923, + 0.004655689466744661, + 0.004827504511922598, + 0.004056053701788187, + 0.003190648276358843, + 0.004590432159602642, + 0.006365879904478788, + 0.01044267788529396, + 0.003986382856965065, + 0.005545068997889757, + 0.010243557393550873, + 0.018283754587173462, + 0.004750851076096296, + 0.008131546899676323, + 0.029766295105218887, + 0.12245023995637894, + 0.005284236278384924, + 0.07299672812223434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.02522406354546547 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038109809160232544 + ], + "opposite_logprob_tokens": [ + -3.28810977935791 + ], + "logprob_reward": -0.038109809160232544, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038109809160232544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10806656628847122, + 0.06285154819488525, + 0.08062052726745605, + 0.076700858771801, + 0.04577753692865372, + 0.04917525872588158, + 0.03447559475898743, + 0.07417144626379013, + 0.0500962994992733, + 0.4180644154548645 + ], + "credit_map": [ + 0.008352176286280155, + 0.018336746841669083, + 0.013132822699844837, + 0.01043226569890976, + 0.017079267650842667, + 0.0192803293466568, + 0.01063278503715992, + 0.010820167139172554, + 0.003770845476537943, + 0.008368021808564663, + 0.00903876218944788, + 0.008040313608944416, + 0.009940543211996555, + 0.010464087128639221, + 0.006224740296602249, + 0.007004230748862028, + 0.0076842946000397205, + 0.013182750903069973, + 0.006547958590090275, + 0.009059480391442776, + 0.020908035337924957, + 0.011012745089828968, + 0.005681047216057777, + 0.006544212810695171, + 0.004957269877195358, + 0.00870436616241932, + 0.00896360445767641, + 0.006966623477637768, + 0.013562974520027637, + 0.014725049026310444, + 0.005684333853423595, + 0.013136623427271843, + 0.0030161128379404545, + 0.008176782168447971, + 0.004033809993416071, + 0.0020446472335606813, + 0.010234755463898182, + 0.012293263338506222, + 0.0018972400575876236, + 0.004080920480191708, + 0.0071194591000676155, + 0.00788833200931549, + 0.004363618791103363, + 0.0048063346184790134, + 0.006305784918367863, + 0.01050494983792305, + 0.004123246297240257, + 0.004063529893755913, + 0.006765910889953375, + 0.004955741111189127, + 0.0028385319747030735, + 0.0030199571046978235, + 0.0049425452016294, + 0.006673155352473259, + 0.0023942897096276283, + 0.0028854594565927982, + 0.009194274432957172, + 0.008689780719578266, + 0.0046881637535989285, + 0.004808325786143541, + 0.010337118990719318, + 0.022098330780863762, + 0.005862225312739611, + 0.008493214845657349, + 0.004312554839998484, + 0.004310193005949259, + 0.0031333635561168194, + 0.003655743785202503, + 0.008312838152050972, + 0.015708312392234802, + 0.0036275798920542, + 0.007035708986222744, + 0.015274629928171635, + 0.031722210347652435, + 0.006714285351336002, + 0.009764055721461773, + 0.043374963104724884, + 0.19090485572814941, + 0.007554823998361826, + 0.11275456100702286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.038109809160232544 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017128810286521912 + ], + "opposite_logprob_tokens": [ + -6.376712799072266 + ], + "logprob_reward": -0.0017128810286521912, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017128810286521912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10207737237215042, + 0.08931593596935272, + 0.22621223330497742, + 0.025655414909124374, + 0.03613472729921341, + 0.09706475585699081, + 0.05249449983239174, + 0.057235073298215866, + 0.02258608303964138, + 0.2912239134311676 + ], + "credit_map": [ + 0.008436773903667927, + 0.013673853129148483, + 0.011402110569179058, + 0.01273567695170641, + 0.019917724654078484, + 0.01362450048327446, + 0.01009769830852747, + 0.012189035303890705, + 0.007496255915611982, + 0.015929238870739937, + 0.009839529171586037, + 0.007046634331345558, + 0.014273091219365597, + 0.01959628239274025, + 0.00567850423976779, + 0.009456406347453594, + 0.018829738721251488, + 0.045302070677280426, + 0.01625385321676731, + 0.012530595995485783, + 0.060363270342350006, + 0.03456190973520279, + 0.013978139497339725, + 0.02439265325665474, + 0.0019209404708817601, + 0.0035300967283546925, + 0.0023819333873689175, + 0.0033017487730830908, + 0.004656058270484209, + 0.004129577893763781, + 0.0021940895821899176, + 0.003540969919413328, + 0.0024922993034124374, + 0.004689990542829037, + 0.0023696355056017637, + 0.0026371197309345007, + 0.011412501335144043, + 0.006879507564008236, + 0.002529588295146823, + 0.003124081762507558, + 0.016586262732744217, + 0.015130390413105488, + 0.007296173367649317, + 0.010953059419989586, + 0.009726077318191528, + 0.021822215989232063, + 0.00807992648333311, + 0.007470651529729366, + 0.010069399140775204, + 0.007180843036621809, + 0.0028964306693524122, + 0.004065453540533781, + 0.006836315616965294, + 0.013750278390944004, + 0.002909368136897683, + 0.004786413628607988, + 0.014245080761611462, + 0.007051121909171343, + 0.005013960879296064, + 0.0067764827981591225, + 0.009293944574892521, + 0.007064401637762785, + 0.004180311691015959, + 0.003609768580645323, + 0.002464017365127802, + 0.002204248681664467, + 0.001992077101022005, + 0.003979525528848171, + 0.002840798581019044, + 0.004280406050384045, + 0.002160118892788887, + 0.0026648903731256723, + 0.010857993736863136, + 0.021084003150463104, + 0.008186662569642067, + 0.013258676044642925, + 0.0337364487349987, + 0.12257836014032364, + 0.0069530135951936245, + 0.07456875592470169 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.0017128810286521912 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062039993703365326 + ], + "opposite_logprob_tokens": [ + -2.812040090560913 + ], + "logprob_reward": -0.062039993703365326, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062039993703365326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13413703441619873, + 0.1560421884059906, + 0.10028009116649628, + 0.19559267163276672, + 0.0642327144742012, + 0.10089918226003647, + 0.04322896525263786, + 0.04912612959742546, + 0.03050762414932251, + 0.12595339119434357 + ], + "credit_map": [ + 0.01107083447277546, + 0.021181585267186165, + 0.019228991121053696, + 0.019015712663531303, + 0.022933410480618477, + 0.014372654259204865, + 0.014707524329423904, + 0.011626319959759712, + 0.013199408538639545, + 0.047299981117248535, + 0.01284913718700409, + 0.0069563924334943295, + 0.027552947402000427, + 0.02596256509423256, + 0.008565830066800117, + 0.013655922375619411, + 0.007449453696608543, + 0.012878959067165852, + 0.008335703983902931, + 0.006970961578190327, + 0.021990960463881493, + 0.02450239658355713, + 0.005085777956992388, + 0.0130658820271492, + 0.011522932909429073, + 0.02264718897640705, + 0.019843226298689842, + 0.015606474131345749, + 0.038260623812675476, + 0.041500166058540344, + 0.014586846344172955, + 0.03162521496415138, + 0.003911998588591814, + 0.008880670182406902, + 0.004681124351918697, + 0.0032797229941934347, + 0.021115213632583618, + 0.012455800548195839, + 0.004128975328058004, + 0.005779210943728685, + 0.018195465207099915, + 0.014363699592649937, + 0.0077428738586604595, + 0.008314892649650574, + 0.011785396374762058, + 0.024158984422683716, + 0.008008846081793308, + 0.008329016156494617, + 0.007900413125753403, + 0.005464120302349329, + 0.0035626646131277084, + 0.005632663611322641, + 0.007683664094656706, + 0.005353446118533611, + 0.0037407446652650833, + 0.0038912477903068066, + 0.00783476885408163, + 0.00616019032895565, + 0.003891330910846591, + 0.004070155322551727, + 0.0064026666805148125, + 0.011854533106088638, + 0.0038638890255242586, + 0.005048598628491163, + 0.004148401319980621, + 0.0038417745381593704, + 0.003705347655341029, + 0.0033852141350507736, + 0.0044365497305989265, + 0.004273522179573774, + 0.00400953646749258, + 0.0027072764933109283, + 0.008181572891771793, + 0.013702580705285072, + 0.0058214678429067135, + 0.006187691353261471, + 0.013018731027841568, + 0.04666442051529884, + 0.005812698043882847, + 0.026564229279756546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.062039993703365326 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11289795488119125 + ], + "opposite_logprob_tokens": [ + -2.2378978729248047 + ], + "logprob_reward": -0.11289795488119125, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11289795488119125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10743599385023117, + 0.08473144471645355, + 0.10769306868314743, + 0.05675512179732323, + 0.02098281867802143, + 0.16864584386348724, + 0.08078846335411072, + 0.050589434802532196, + 0.10254465788602829, + 0.21983323991298676 + ], + "credit_map": [ + 0.009179798886179924, + 0.015595890581607819, + 0.010278863832354546, + 0.009061227552592754, + 0.02213122881948948, + 0.018486732617020607, + 0.009486980736255646, + 0.013215260580182076, + 0.006703030318021774, + 0.01515959482640028, + 0.008005683310329914, + 0.009533289819955826, + 0.011778982356190681, + 0.014943767338991165, + 0.008151878602802753, + 0.010455211624503136, + 0.007067132741212845, + 0.014881482347846031, + 0.008311139419674873, + 0.005478634964674711, + 0.021321503445506096, + 0.0318119078874588, + 0.0051618413999676704, + 0.01365942694246769, + 0.004074343014508486, + 0.0061151133850216866, + 0.00624024448916316, + 0.007913254201412201, + 0.008825224824249744, + 0.010457973927259445, + 0.004674938041716814, + 0.008454030379652977, + 0.001638739020563662, + 0.002765150275081396, + 0.0012973560951650143, + 0.0033255750313401222, + 0.004282121080905199, + 0.004060368053615093, + 0.001521423226222396, + 0.002092083916068077, + 0.0290826428681612, + 0.025143681094050407, + 0.011882234364748001, + 0.01632477529346943, + 0.01787690259516239, + 0.041933611035346985, + 0.013004393316805363, + 0.013397595845162868, + 0.012112182565033436, + 0.00907891895622015, + 0.004685740452259779, + 0.007023829035460949, + 0.011257190257310867, + 0.02294166572391987, + 0.005299877375364304, + 0.008389061316847801, + 0.00895966961979866, + 0.005641757510602474, + 0.003734258469194174, + 0.006804846692830324, + 0.006499225273728371, + 0.00987576600164175, + 0.004171833395957947, + 0.004902074579149485, + 0.01119917444884777, + 0.00981076993048191, + 0.00442104646936059, + 0.01669260673224926, + 0.013981114141643047, + 0.027405114844441414, + 0.008809174410998821, + 0.010225658304989338, + 0.012879172340035439, + 0.013675130903720856, + 0.0061935787089169025, + 0.0067554013803601265, + 0.025091297924518585, + 0.09357315301895142, + 0.006345052737742662, + 0.05532044544816017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.11289795488119125 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "As the camera executes a slow, perfectly steady tracking motion, a majestic extreme wide shot unfolds, gliding parallel to a massive, whale-like, semi-rigid airship, drifting silently across a breathtakingly vast, pastel-colored sky. The airship's shape is organic and monumental, reminiscent of a giant blue whale, with a broad, rounded nose and elegantly tapered tail. Its skin is a seamless, high-tech fabric of pearlescent off-white, beneath which the subtle outlines of its internal structural ribs are faintly visible. The entire scene is bathed in the warm golden glow of a low, off-camera sun, which caresses the airship's massive, curved flanks, creating a soft sheen on its fabric and casting long, gentle shadows that emphasize its organic contours. The background is a masterpiece of soft colors, the sky melting into a gentle gradient of pale rose, creamy orange, and gentle lavender, dotted with wispy, feathery clouds. The airship floats forward with an almost imperceptible serene grace, its immense scale creating a profound sense of peace in the quiet, still air, while the camera’s synchronized, deliberate tracking movements enhance the feeling of a silent, dreamlike journey, creating an extremely soothing and awe-inspiring atmosphere.", + "seed": 429, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_11/index_5175_seed_429_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06719439830631017, + "vlm_logprob": -0.06719439830631017, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.06719439830631017, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018848128616809845 + ], + "opposite_logprob_tokens": [ + -4.268847942352295 + ], + "logprob_reward": -0.018848128616809845, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.018848128616809845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08309833705425262, + 0.09905137121677399, + 0.048144128173589706, + 0.022288279607892036, + 0.04418836161494255, + 0.03121861256659031, + 0.06628785282373428, + 0.31003427505493164, + 0.24916668236255646, + 0.0465221181511879 + ], + "credit_map": [ + 0.007839212194085121, + 0.008820796385407448, + 0.017241593450307846, + 0.00868875253945589, + 0.00787880178540945, + 0.009192284196615219, + 0.010676650330424309, + 0.012760250829160213, + 0.010323761962354183, + 0.00861302949488163, + 0.030336154624819756, + 0.006973551120609045, + 0.008226375095546246, + 0.013393173925578594, + 0.00908732507377863, + 0.012098003178834915, + 0.005396646447479725, + 0.004972139839082956, + 0.014459013938903809, + 0.0038016606122255325, + 0.0035798593889921904, + 0.006210096646100283, + 0.0047653429210186005, + 0.004959368146955967, + 0.0021650483831763268, + 0.001835792325437069, + 0.0071094143204391, + 0.001526149921119213, + 0.001581471529789269, + 0.0024485543835908175, + 0.0020094728097319603, + 0.003612376982346177, + 0.003355660242959857, + 0.003638615133240819, + 0.0067489491775631905, + 0.003675494808703661, + 0.004120584111660719, + 0.0065603842958807945, + 0.006072790827602148, + 0.0100158816203475, + 0.003011904191225767, + 0.0031561758369207382, + 0.004627720918506384, + 0.002307124435901642, + 0.004976741969585419, + 0.004582515452057123, + 0.0038540265522897243, + 0.004702402278780937, + 0.004288855008780956, + 0.0025626530405133963, + 0.01822349615395069, + 0.002004193142056465, + 0.0051200855523347855, + 0.00860962551087141, + 0.003539965720847249, + 0.021938972175121307, + 0.03191777691245079, + 0.023255016654729843, + 0.05079188197851181, + 0.014614197425544262, + 0.07801182568073273, + 0.05182289704680443, + 0.016706883907318115, + 0.04291379824280739, + 0.030908966436982155, + 0.014506729319691658, + 0.05752374976873398, + 0.017255248501896858, + 0.021311042830348015, + 0.02590198628604412, + 0.03971438109874725, + 0.04204457998275757, + 0.005111544393002987, + 0.004524601623415947, + 0.0070152608677744865, + 0.003389256075024605, + 0.00473713967949152, + 0.005636081099510193, + 0.008626015856862068, + 0.007482217624783516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.018848128616809845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004149258136749 + ], + "opposite_logprob_tokens": [ + -2.6950414180755615 + ], + "logprob_reward": -0.07004149258136749, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004149258136749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11768830567598343, + 0.09833700954914093, + 0.06931682676076889, + 0.058389052748680115, + 0.015872549265623093, + 0.04627938196063042, + 0.18940642476081848, + 0.19642719626426697, + 0.190010204911232, + 0.018273109570145607 + ], + "credit_map": [ + 0.013566922396421432, + 0.01307609025388956, + 0.0426071435213089, + 0.008209963329136372, + 0.007115645334124565, + 0.010365220718085766, + 0.007437286898493767, + 0.01531002577394247, + 0.011081383563578129, + 0.011802281253039837, + 0.03560025244951248, + 0.006148966029286385, + 0.005418941844254732, + 0.010348163545131683, + 0.006213797256350517, + 0.011723223142325878, + 0.0060150898061692715, + 0.00688421493396163, + 0.024311812594532967, + 0.0037754755467176437, + 0.004242734517902136, + 0.008885596878826618, + 0.006689468398690224, + 0.008512431755661964, + 0.00631332490593195, + 0.004693359136581421, + 0.01797022670507431, + 0.0035221269354224205, + 0.0047700065188109875, + 0.006180396303534508, + 0.005611588712781668, + 0.009328019805252552, + 0.0012247281847521663, + 0.0018860420677810907, + 0.002727870363742113, + 0.0009547034278512001, + 0.0016658114036545157, + 0.002464795485138893, + 0.0018458106787875295, + 0.003102787770330906, + 0.004916682373732328, + 0.002830968704074621, + 0.008556581102311611, + 0.002649656031280756, + 0.006614356301724911, + 0.007864945568144321, + 0.0038198730908334255, + 0.009026316925883293, + 0.012736286967992783, + 0.006803465075790882, + 0.056576963514089584, + 0.004962962120771408, + 0.010789930820465088, + 0.02521013654768467, + 0.0063115619122982025, + 0.06601510941982269, + 0.020851267501711845, + 0.014306245371699333, + 0.038237106055021286, + 0.012754738330841064, + 0.02467631734907627, + 0.03568305820226669, + 0.014395171776413918, + 0.035523273050785065, + 0.022256216034293175, + 0.010773285292088985, + 0.04341363534331322, + 0.0127488998696208, + 0.016143731772899628, + 0.019332483410835266, + 0.030534878373146057, + 0.03480706736445427, + 0.0027631737757474184, + 0.001803801627829671, + 0.002371230162680149, + 0.0012247757986187935, + 0.002151590073481202, + 0.002839218359440565, + 0.0024961780291050673, + 0.002623139414936304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.07004149258136749 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024906350299715996 + ], + "opposite_logprob_tokens": [ + -6.002490520477295 + ], + "logprob_reward": -0.0024906350299715996, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024906350299715996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09366767108440399, + 0.04047428071498871, + 0.025091150775551796, + 0.015121095813810825, + 0.04892436042428017, + 0.04157102853059769, + 0.5686018466949463, + 0.11039239168167114, + 0.03660886362195015, + 0.01954735815525055 + ], + "credit_map": [ + 0.008932731114327908, + 0.009491292759776115, + 0.02274077944457531, + 0.007352554704993963, + 0.008355585858225822, + 0.010553435422480106, + 0.01037201564759016, + 0.015869265422225, + 0.0032541339751333, + 0.004976044874638319, + 0.007524855900555849, + 0.0041637783870100975, + 0.004350778181105852, + 0.004977645352482796, + 0.0045783547684550285, + 0.006648685783147812, + 0.002666878281161189, + 0.00219505256973207, + 0.004873787984251976, + 0.0014139822451397777, + 0.002699164440855384, + 0.004218968562781811, + 0.002133042085915804, + 0.004890270531177521, + 0.001124515780247748, + 0.0013528787530958652, + 0.0023994629736989737, + 0.0014065735740587115, + 0.0017286906950175762, + 0.002224886789917946, + 0.002072331728413701, + 0.002811754820868373, + 0.004934818018227816, + 0.003627088852226734, + 0.014517494477331638, + 0.00265790824778378, + 0.003626229241490364, + 0.006097407545894384, + 0.003693816950544715, + 0.009769595228135586, + 0.004377859178930521, + 0.0030019062105566263, + 0.008870167657732964, + 0.0022086091339588165, + 0.0062408740632236, + 0.006852225866168737, + 0.0031963190995156765, + 0.006823064759373665, + 0.037345435470342636, + 0.018384641036391258, + 0.17475847899913788, + 0.013974200002849102, + 0.028629347681999207, + 0.06606186181306839, + 0.017565131187438965, + 0.21188266575336456, + 0.013060826808214188, + 0.007753210142254829, + 0.012116537429392338, + 0.004312254022806883, + 0.0366048663854599, + 0.017649687826633453, + 0.005004384089261293, + 0.013890616595745087, + 0.0037652780301868916, + 0.0036849933676421642, + 0.007485989481210709, + 0.0027705649845302105, + 0.003471693489700556, + 0.005955635569989681, + 0.004013566765934229, + 0.0054611400701105595, + 0.002294026082381606, + 0.003121736226603389, + 0.0029187225736677647, + 0.001543120015412569, + 0.0021411727648228407, + 0.0024730772711336613, + 0.0023968140594661236, + 0.0026586896274238825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.0024906350299715996 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043054722249507904 + ], + "opposite_logprob_tokens": [ + -3.1680548191070557 + ], + "logprob_reward": -0.043054722249507904, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043054722249507904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14359620213508606, + 0.07287918776273727, + 0.06399980187416077, + 0.043295230716466904, + 0.04474024847149849, + 0.0661529079079628, + 0.1945933997631073, + 0.29022353887557983, + 0.05758209154009819, + 0.022937370464205742 + ], + "credit_map": [ + 0.015728281810879707, + 0.014839734882116318, + 0.05114030838012695, + 0.009282978251576424, + 0.011694644577801228, + 0.013310706242918968, + 0.010348545387387276, + 0.01725100539624691, + 0.006176323164254427, + 0.0068943193182349205, + 0.012841200456023216, + 0.00516577810049057, + 0.009697877801954746, + 0.011485063470900059, + 0.010036658495664597, + 0.010581967420876026, + 0.006022097542881966, + 0.005315450485795736, + 0.01934809237718582, + 0.003919356036931276, + 0.0045242696069180965, + 0.00987346563488245, + 0.00691758468747139, + 0.00807948037981987, + 0.003369629615917802, + 0.003675237763673067, + 0.010444972664117813, + 0.0019923888612538576, + 0.005239584483206272, + 0.006071798503398895, + 0.0037384498864412308, + 0.008763170801103115, + 0.0032829390838742256, + 0.004248894285410643, + 0.006268880795687437, + 0.003322000615298748, + 0.004989676643162966, + 0.007061121519654989, + 0.006384679116308689, + 0.009182052686810493, + 0.006562893278896809, + 0.003992668818682432, + 0.010523257777094841, + 0.0036175744608044624, + 0.009591935202479362, + 0.012825975194573402, + 0.005013861693441868, + 0.014024744741618633, + 0.01664033904671669, + 0.009724495932459831, + 0.05096077919006348, + 0.006677203346043825, + 0.014524267986416817, + 0.03091338649392128, + 0.009577549993991852, + 0.05557536706328392, + 0.03594435006380081, + 0.020831691101193428, + 0.03709195554256439, + 0.01101307850331068, + 0.09627203643321991, + 0.05208240821957588, + 0.01282372698187828, + 0.02416430041193962, + 0.007135712541639805, + 0.005768521688878536, + 0.01335957646369934, + 0.002783717354759574, + 0.00630394509062171, + 0.007279940415173769, + 0.005771832540631294, + 0.009178848005831242, + 0.0031679777894169092, + 0.001624871278181672, + 0.004435473587363958, + 0.0011918204836547375, + 0.0023497710935771465, + 0.0036666837986558676, + 0.002414640039205551, + 0.004086134023964405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.043054722249507904 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20153701305389404 + ], + "opposite_logprob_tokens": [ + -1.701537013053894 + ], + "logprob_reward": -0.20153701305389404, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20153701305389404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11611625552177429, + 0.06591910123825073, + 0.023023977875709534, + 0.02646629512310028, + 0.08140119910240173, + 0.09389962255954742, + 0.25964289903640747, + 0.10541205108165741, + 0.15658420324325562, + 0.07153439521789551 + ], + "credit_map": [ + 0.011503225192427635, + 0.013221788220107555, + 0.03804071620106697, + 0.009690775536000729, + 0.010735873132944107, + 0.011060206219553947, + 0.007724187336862087, + 0.01413948554545641, + 0.005115412641316652, + 0.008531668223440647, + 0.009295808151364326, + 0.005782523658126593, + 0.00847337394952774, + 0.009124727919697762, + 0.010518571361899376, + 0.009077019058167934, + 0.0021874967496842146, + 0.0023857434280216694, + 0.006145934108644724, + 0.0017233544494956732, + 0.0023785012308508158, + 0.00267051556147635, + 0.002201515482738614, + 0.0033309173304587603, + 0.00216223718598485, + 0.003054016036912799, + 0.004136581905186176, + 0.002313041826710105, + 0.0040645054541528225, + 0.003667179960757494, + 0.0036546962801367044, + 0.0034140348434448242, + 0.007796525489538908, + 0.006777678616344929, + 0.018132761120796204, + 0.008006295189261436, + 0.00805041752755642, + 0.010897675529122353, + 0.010586111806333065, + 0.01115372683852911, + 0.010367671959102154, + 0.007627258077263832, + 0.016787094995379448, + 0.005927578546106815, + 0.014066708274185658, + 0.015090515837073326, + 0.0070592379197478294, + 0.016973551362752914, + 0.016772054135799408, + 0.00797679927200079, + 0.08005499839782715, + 0.006423585116863251, + 0.014131428673863411, + 0.028886843472719193, + 0.0081202806904912, + 0.09727692604064941, + 0.012893114238977432, + 0.006858838256448507, + 0.016623549163341522, + 0.006892463192343712, + 0.022036613896489143, + 0.01818346232175827, + 0.008571144193410873, + 0.013352860696613789, + 0.01834268867969513, + 0.010564122349023819, + 0.03164638578891754, + 0.009444874711334705, + 0.013763559982180595, + 0.020909689366817474, + 0.014889949932694435, + 0.03702292591333389, + 0.010351481847465038, + 0.006126104854047298, + 0.013432850129902363, + 0.004112075548619032, + 0.006212160922586918, + 0.010239019058644772, + 0.006376760080456734, + 0.014683940447866917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.20153701305389404 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "As the camera executes a slow, perfectly steady tracking motion, a majestic extreme wide shot unfolds, gliding parallel to a massive, whale-like, semi-rigid airship, drifting silently across a breathtakingly vast, pastel-colored sky. The airship's shape is organic and monumental, reminiscent of a giant blue whale, with a broad, rounded nose and elegantly tapered tail. Its skin is a seamless, high-tech fabric of pearlescent off-white, beneath which the subtle outlines of its internal structural ribs are faintly visible. The entire scene is bathed in the warm golden glow of a low, off-camera sun, which caresses the airship's massive, curved flanks, creating a soft sheen on its fabric and casting long, gentle shadows that emphasize its organic contours. The background is a masterpiece of soft colors, the sky melting into a gentle gradient of pale rose, creamy orange, and gentle lavender, dotted with wispy, feathery clouds. The airship floats forward with an almost imperceptible serene grace, its immense scale creating a profound sense of peace in the quiet, still air, while the camera’s synchronized, deliberate tracking movements enhance the feeling of a silent, dreamlike journey, creating an extremely soothing and awe-inspiring atmosphere.", + "seed": 430, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_11/index_5175_seed_430_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01658291332423687, + "vlm_logprob": -0.01658291332423687, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.01658291332423687, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004375883378088474 + ], + "opposite_logprob_tokens": [ + -5.75437593460083 + ], + "logprob_reward": -0.004375883378088474, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004375883378088474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08547844737768173, + 0.031088585034012794, + 0.030093684792518616, + 0.019787119701504707, + 0.21569544076919556, + 0.22527165710926056, + 0.17720986902713776, + 0.09110736101865768, + 0.07013065367937088, + 0.05413716658949852 + ], + "credit_map": [ + 0.009375806897878647, + 0.011277382262051105, + 0.010120623745024204, + 0.01006906945258379, + 0.011316635645925999, + 0.011463746428489685, + 0.011017046868801117, + 0.010838142596185207, + 0.0032724731136113405, + 0.0035935789346694946, + 0.003869683714583516, + 0.00328222569078207, + 0.0034419801086187363, + 0.0054025473073124886, + 0.0037451006937772036, + 0.0044809975661337376, + 0.002841716632246971, + 0.0034255986101925373, + 0.0041184029541909695, + 0.003454863093793392, + 0.003478157566860318, + 0.005381160881370306, + 0.0035197369288653135, + 0.003874048823490739, + 0.001976095139980316, + 0.0018648866098374128, + 0.00243462435901165, + 0.00268027582205832, + 0.002282531000673771, + 0.003365284763276577, + 0.0019479667535051703, + 0.0032354556024074554, + 0.01699749194085598, + 0.014263572171330452, + 0.01845957525074482, + 0.012601839378476143, + 0.042497165501117706, + 0.05731285363435745, + 0.01143425889313221, + 0.04212868586182594, + 0.014736438170075417, + 0.016579117625951767, + 0.044552579522132874, + 0.011968434788286686, + 0.03737195208668709, + 0.07882637530565262, + 0.007757880724966526, + 0.013478885404765606, + 0.010936683975160122, + 0.013785065151751041, + 0.03701578825712204, + 0.01487649790942669, + 0.02341661974787712, + 0.050387464463710785, + 0.011814399622380733, + 0.014977368526160717, + 0.005549966357648373, + 0.008706655353307724, + 0.007939704693853855, + 0.007001824211329222, + 0.01524391956627369, + 0.019450943917036057, + 0.012113358825445175, + 0.015100989490747452, + 0.0066677057184278965, + 0.008474265225231647, + 0.008927972987294197, + 0.0071338023990392685, + 0.00927323754876852, + 0.006864812690764666, + 0.004570286720991135, + 0.018218565732240677, + 0.004306656774133444, + 0.007926742546260357, + 0.00793479010462761, + 0.007332052104175091, + 0.006181178614497185, + 0.0054230960085988045, + 0.004827913828194141, + 0.010204740799963474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.004375883378088474 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635757438838482 + ], + "opposite_logprob_tokens": [ + -3.651357650756836 + ], + "logprob_reward": -0.02635757438838482, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635757438838482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12916578352451324, + 0.06824172288179398, + 0.03223709017038345, + 0.025809207931160927, + 0.08259186893701553, + 0.12686654925346375, + 0.11238110065460205, + 0.11484574526548386, + 0.07495550066232681, + 0.23290549218654633 + ], + "credit_map": [ + 0.014068959280848503, + 0.016899047419428825, + 0.011938121169805527, + 0.01534687913954258, + 0.01958250254392624, + 0.01594303734600544, + 0.014802162535488605, + 0.02058507688343525, + 0.007043315563350916, + 0.010157032869756222, + 0.007694590371102095, + 0.006749106105417013, + 0.00900670699775219, + 0.01363626029342413, + 0.00646852096542716, + 0.007486191112548113, + 0.0036509803030639887, + 0.003043347503989935, + 0.004717916715890169, + 0.003486348781734705, + 0.0033999516163021326, + 0.005604081321507692, + 0.004782717674970627, + 0.0035517457872629166, + 0.003043076489120722, + 0.0025879002641886473, + 0.0028961258940398693, + 0.003116064704954624, + 0.002874210011214018, + 0.004056883510202169, + 0.0021546429488807917, + 0.00508030503988266, + 0.008688262663781643, + 0.006925571709871292, + 0.006025515962392092, + 0.005478816572576761, + 0.018139483407139778, + 0.022738276049494743, + 0.0068878247402608395, + 0.007708118297159672, + 0.005359831731766462, + 0.011281066574156284, + 0.015787273645401, + 0.006810791324824095, + 0.024797789752483368, + 0.04216749221086502, + 0.006997537799179554, + 0.013664761558175087, + 0.0077596502378582954, + 0.011239524930715561, + 0.01262865960597992, + 0.014406436122953892, + 0.015534988604485989, + 0.025384943932294846, + 0.010240111500024796, + 0.015186781994998455, + 0.006407474167644978, + 0.010106137953698635, + 0.011708913370966911, + 0.009164235554635525, + 0.02279103733599186, + 0.027474798262119293, + 0.014685309492051601, + 0.012507842853665352, + 0.00632626935839653, + 0.009752067737281322, + 0.010761150158941746, + 0.007694576866924763, + 0.012035496532917023, + 0.008153696544468403, + 0.005548845045268536, + 0.014683397486805916, + 0.022660940885543823, + 0.024421092122793198, + 0.042698830366134644, + 0.02632131054997444, + 0.03177712485194206, + 0.01748456433415413, + 0.016184907406568527, + 0.051356736570596695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.02635757438838482 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001512932125478983 + ], + "opposite_logprob_tokens": [ + -6.5015130043029785 + ], + "logprob_reward": -0.001512932125478983, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001512932125478983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08715194463729858, + 0.025973612442612648, + 0.021588034927845, + 0.03621842712163925, + 0.3051902651786804, + 0.11670924723148346, + 0.12920348346233368, + 0.040459439158439636, + 0.12338321655988693, + 0.11412233114242554 + ], + "credit_map": [ + 0.009026192128658295, + 0.011136863380670547, + 0.009958229959011078, + 0.009720949456095695, + 0.01447579637169838, + 0.012043162249028683, + 0.00899737048894167, + 0.011793392710387707, + 0.0036544683389365673, + 0.003313091117888689, + 0.003734285943210125, + 0.0031885800417512655, + 0.0027008764445781708, + 0.003993689548224211, + 0.0025812056846916676, + 0.002807416720315814, + 0.003709295764565468, + 0.0019288009498268366, + 0.0029808315448462963, + 0.0028898196760565042, + 0.00196963082998991, + 0.00338284601457417, + 0.0022111276630312204, + 0.002515683649107814, + 0.005421670153737068, + 0.0043699112720787525, + 0.0053939297795295715, + 0.004459671676158905, + 0.0035670928191393614, + 0.004850773606449366, + 0.003010787768289447, + 0.0051445867866277695, + 0.016697155311703682, + 0.021894261240959167, + 0.022999107837677002, + 0.015586571767926216, + 0.06545552611351013, + 0.08858682215213776, + 0.01685377024114132, + 0.05711705982685089, + 0.01794656366109848, + 0.011841649189591408, + 0.014018607325851917, + 0.011714228428900242, + 0.01228492334485054, + 0.013769415207207203, + 0.009534173645079136, + 0.025599690154194832, + 0.009775621816515923, + 0.01327165775001049, + 0.023393116891384125, + 0.007960733957588673, + 0.020258011296391487, + 0.04111245274543762, + 0.005907826591283083, + 0.00752407219260931, + 0.002480127150192857, + 0.00387268397025764, + 0.0042100208811461926, + 0.003754353616386652, + 0.006971992086619139, + 0.009548882953822613, + 0.0036102423910051584, + 0.006011140067130327, + 0.012000586837530136, + 0.015260451473295689, + 0.017896009609103203, + 0.018803758546710014, + 0.012913479469716549, + 0.011554551310837269, + 0.008170638233423233, + 0.026783747598528862, + 0.011691487394273281, + 0.010198920965194702, + 0.020313458517193794, + 0.012874896638095379, + 0.015254420228302479, + 0.009024967439472675, + 0.007820077240467072, + 0.02694411389529705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.001512932125478983 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626701612025499 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626701612025499, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626701612025499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0762500986456871, + 0.1116768941283226, + 0.017574597150087357, + 0.0221242755651474, + 0.2020930051803589, + 0.14119172096252441, + 0.1085989773273468, + 0.04415181651711464, + 0.04656820744276047, + 0.22977039217948914 + ], + "credit_map": [ + 0.007789121475070715, + 0.007720257621258497, + 0.00861360877752304, + 0.010220403783023357, + 0.009500059299170971, + 0.009438173845410347, + 0.00959481205791235, + 0.013373668305575848, + 0.010470593348145485, + 0.011630142107605934, + 0.015467117540538311, + 0.010603366419672966, + 0.01257310900837183, + 0.025129999965429306, + 0.011892496608197689, + 0.013910076580941677, + 0.002833003643900156, + 0.0026434266474097967, + 0.002452733227983117, + 0.00200158404186368, + 0.0017303272616118193, + 0.0021965252235531807, + 0.0015519948210567236, + 0.0021650034468621016, + 0.0028905312065035105, + 0.0023625430185347795, + 0.0027844414580613375, + 0.0036721904762089252, + 0.002817227505147457, + 0.0038979691453278065, + 0.0016400321619585156, + 0.0020593421068042517, + 0.013105801306664944, + 0.01662713848054409, + 0.015743086114525795, + 0.013012237846851349, + 0.03751472011208534, + 0.0482485294342041, + 0.013993298634886742, + 0.0438481867313385, + 0.011741308495402336, + 0.012360419146716595, + 0.014938549138605595, + 0.00991993397474289, + 0.024851646274328232, + 0.038903046399354935, + 0.007176227401942015, + 0.021300598978996277, + 0.016045965254306793, + 0.014129521325230598, + 0.02173747308552265, + 0.01842741295695305, + 0.007330668158829212, + 0.014776283875107765, + 0.006933451630175114, + 0.00921819917857647, + 0.0028126975521445274, + 0.004113816190510988, + 0.004821266047656536, + 0.003540273057296872, + 0.009064161218702793, + 0.011458701454102993, + 0.004413141403347254, + 0.003927756100893021, + 0.004073998425155878, + 0.007357003632932901, + 0.0048845987766981125, + 0.005517101380974054, + 0.006949518341571093, + 0.007173255551606417, + 0.004113667644560337, + 0.006499065551906824, + 0.022057564929127693, + 0.023556141182780266, + 0.043221279978752136, + 0.03041430562734604, + 0.028393473476171494, + 0.016933953389525414, + 0.015012123622000217, + 0.050181563943624496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.007626701612025499 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043041475117206573 + ], + "opposite_logprob_tokens": [ + -3.168041467666626 + ], + "logprob_reward": -0.043041475117206573, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043041475117206573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0733834058046341, + 0.016404850408434868, + 0.02532607689499855, + 0.03113424964249134, + 0.1510549932718277, + 0.18210826814174652, + 0.28212466835975647, + 0.06775360554456711, + 0.08556132763624191, + 0.08514858037233353 + ], + "credit_map": [ + 0.007631508633494377, + 0.007347290404140949, + 0.007509622722864151, + 0.008387329988181591, + 0.011601809412240982, + 0.00935006607323885, + 0.008945544250309467, + 0.012610236182808876, + 0.0016342410817742348, + 0.0023739016614854336, + 0.0017060891259461641, + 0.0017452194588258862, + 0.00244678626768291, + 0.0031441415194422007, + 0.0018087439239025116, + 0.001545728649944067, + 0.0023241417948156595, + 0.0023807871621102095, + 0.004008156247437, + 0.0029345282819122076, + 0.0030279732309281826, + 0.006235001143068075, + 0.0019185702549293637, + 0.002496918896213174, + 0.003919727634638548, + 0.005829808302223682, + 0.0036637603770941496, + 0.005504064727574587, + 0.003633256070315838, + 0.003571411594748497, + 0.002725077560171485, + 0.0022871403489261866, + 0.012930640950798988, + 0.009687373414635658, + 0.014553125016391277, + 0.009505094029009342, + 0.0257016159594059, + 0.03884809836745262, + 0.008111584931612015, + 0.03171747177839279, + 0.0064859031699597836, + 0.015007217414677143, + 0.02898833341896534, + 0.007914907298982143, + 0.0391852930188179, + 0.070960633456707, + 0.0048288386315107346, + 0.008737152442336082, + 0.024900346994400024, + 0.017463011667132378, + 0.07080066949129105, + 0.020783791318535805, + 0.03447989746928215, + 0.08271325379610062, + 0.015078527852892876, + 0.01590517722070217, + 0.0047227246686816216, + 0.005765052977949381, + 0.005479938816279173, + 0.006884095259010792, + 0.010542910546064377, + 0.010919624008238316, + 0.011740189045667648, + 0.011699071153998375, + 0.008185996674001217, + 0.008342613466084003, + 0.00923786498606205, + 0.008587374351918697, + 0.008924992755055428, + 0.009785636328160763, + 0.007198890205472708, + 0.02529795467853546, + 0.009074877947568893, + 0.006783149670809507, + 0.01689746230840683, + 0.010397772304713726, + 0.009381799027323723, + 0.005934198386967182, + 0.005810080096125603, + 0.020869240164756775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.043041475117206573 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "As the camera executes a slow, perfectly steady tracking motion, a majestic extreme wide shot unfolds, gliding parallel to a massive, whale-like, semi-rigid airship, drifting silently across a breathtakingly vast, pastel-colored sky. The airship's shape is organic and monumental, reminiscent of a giant blue whale, with a broad, rounded nose and elegantly tapered tail. Its skin is a seamless, high-tech fabric of pearlescent off-white, beneath which the subtle outlines of its internal structural ribs are faintly visible. The entire scene is bathed in the warm golden glow of a low, off-camera sun, which caresses the airship's massive, curved flanks, creating a soft sheen on its fabric and casting long, gentle shadows that emphasize its organic contours. The background is a masterpiece of soft colors, the sky melting into a gentle gradient of pale rose, creamy orange, and gentle lavender, dotted with wispy, feathery clouds. The airship floats forward with an almost imperceptible serene grace, its immense scale creating a profound sense of peace in the quiet, still air, while the camera’s synchronized, deliberate tracking movements enhance the feeling of a silent, dreamlike journey, creating an extremely soothing and awe-inspiring atmosphere.", + "seed": 431, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_11/index_5175_seed_431_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0481928903143853, + "vlm_logprob": -0.0481928903143853, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.0481928903143853, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008458266966044903 + ], + "opposite_logprob_tokens": [ + -5.008458137512207 + ], + "logprob_reward": -0.008458266966044903, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008458266966044903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13003940880298615, + 0.036540426313877106, + 0.4027354121208191, + 0.10373018682003021, + 0.03592493757605553, + 0.07624837011098862, + 0.0738716647028923, + 0.044805653393268585, + 0.06521013379096985, + 0.03089374490082264 + ], + "credit_map": [ + 0.019410189241170883, + 0.011562156490981579, + 0.014879746362566948, + 0.00920785777270794, + 0.019708754494786263, + 0.021787768229842186, + 0.014988060109317303, + 0.018494879826903343, + 0.0035845499951392412, + 0.0038341463077813387, + 0.003433777019381523, + 0.0031733305659145117, + 0.007709107361733913, + 0.005556359887123108, + 0.0043417480774223804, + 0.0049074082635343075, + 0.03293141350150108, + 0.019981253892183304, + 0.013787896372377872, + 0.009248468093574047, + 0.03038981929421425, + 0.18045403063297272, + 0.008101707324385643, + 0.10784083604812622, + 0.009579666890203953, + 0.006706232205033302, + 0.006180751137435436, + 0.0026741824112832546, + 0.005740417633205652, + 0.03623468801379204, + 0.003385561052709818, + 0.03322869539260864, + 0.005741461180150509, + 0.0038144427817314863, + 0.0036047555040568113, + 0.002269319724291563, + 0.0023123545106500387, + 0.005118216387927532, + 0.003786947811022401, + 0.009277445264160633, + 0.011762220412492752, + 0.005120161920785904, + 0.009367226622998714, + 0.00567658431828022, + 0.005425711628049612, + 0.010288144461810589, + 0.004751081578433514, + 0.023857245221734047, + 0.009776562452316284, + 0.006568740587681532, + 0.008182510733604431, + 0.009389122016727924, + 0.006834077648818493, + 0.00764687592163682, + 0.012583148665726185, + 0.012890633195638657, + 0.006627166643738747, + 0.005781908053904772, + 0.00767143489792943, + 0.005722085479646921, + 0.003869471373036504, + 0.004659496713429689, + 0.006158432923257351, + 0.004315660800784826, + 0.012369255535304546, + 0.007576616480946541, + 0.007593205664306879, + 0.00527142034843564, + 0.009631045162677765, + 0.007187334354966879, + 0.005853904411196709, + 0.009727361612021923, + 0.00543438084423542, + 0.003405253868550062, + 0.0029953112825751305, + 0.003489562077447772, + 0.003567259991541505, + 0.0047921352088451385, + 0.003955338150262833, + 0.0032545062713325024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.008458266966044903 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204492226243019 + ], + "opposite_logprob_tokens": [ + -2.812044858932495 + ], + "logprob_reward": -0.06204492226243019, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204492226243019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06790503859519958, + 0.040103230625391006, + 0.34443143010139465, + 0.2098972648382187, + 0.030243292450904846, + 0.07713521271944046, + 0.08962450176477432, + 0.10477375239133835, + 0.019189661368727684, + 0.016696687787771225 + ], + "credit_map": [ + 0.007947023026645184, + 0.005392515100538731, + 0.004767994862049818, + 0.0029301350004971027, + 0.007180360611528158, + 0.018208513036370277, + 0.004922803025692701, + 0.016555679962038994, + 0.0034690045285969973, + 0.003687860444188118, + 0.00272786314599216, + 0.0024027435574680567, + 0.006359262391924858, + 0.008462558500468731, + 0.0058173914439976215, + 0.007176538463681936, + 0.027088886126875877, + 0.01728113181889057, + 0.013251439668238163, + 0.00897170789539814, + 0.024445990100502968, + 0.14939700067043304, + 0.00807621143758297, + 0.09591902047395706, + 0.03086083196103573, + 0.013226951472461224, + 0.011797984130680561, + 0.008600259199738503, + 0.022622426971793175, + 0.09233500808477402, + 0.0055116936564445496, + 0.02494206838309765, + 0.0037547831889241934, + 0.004863288719207048, + 0.0033232211135327816, + 0.0024644581135362387, + 0.00444363709539175, + 0.004107895772904158, + 0.002750907326117158, + 0.004535097628831863, + 0.011838627979159355, + 0.004872907884418964, + 0.008823893964290619, + 0.005227647256106138, + 0.006268779281526804, + 0.01126739289611578, + 0.004553288221359253, + 0.024282662197947502, + 0.010852417908608913, + 0.007452188991010189, + 0.009581714868545532, + 0.011246608570218086, + 0.00806686095893383, + 0.01008578296750784, + 0.014103757217526436, + 0.01823515258729458, + 0.015431053005158901, + 0.012869983911514282, + 0.011181270703673363, + 0.011044422164559364, + 0.01266527734696865, + 0.013401929289102554, + 0.011790517717599869, + 0.016389282420277596, + 0.003485472407191992, + 0.0024515585973858833, + 0.0019409818341955543, + 0.002043850952759385, + 0.00255920784547925, + 0.0026397984474897385, + 0.0023192125372588634, + 0.0017495752545073628, + 0.002613807562738657, + 0.0020678548607975245, + 0.0021693571470677853, + 0.0017025049310177565, + 0.0020912010222673416, + 0.0023416883777827024, + 0.0017271974356845021, + 0.001983074704185128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.06204492226243019 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024876620154827833 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024876620154827833, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024876620154827833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23946566879749298, + 0.08510898053646088, + 0.10251903533935547, + 0.2309167981147766, + 0.047470055520534515, + 0.05336517095565796, + 0.06690175086259842, + 0.07468905299901962, + 0.06374043226242065, + 0.035823095589876175 + ], + "credit_map": [ + 0.034543853253126144, + 0.019099265336990356, + 0.025340348482131958, + 0.011481665074825287, + 0.01895226538181305, + 0.05814661458134651, + 0.015638725832104683, + 0.05626293271780014, + 0.015268521383404732, + 0.006265716627240181, + 0.015920260921120644, + 0.004756343550980091, + 0.008656037040054798, + 0.011142252013087273, + 0.008964880369603634, + 0.014134969562292099, + 0.007719699293375015, + 0.00875764712691307, + 0.009837892837822437, + 0.004132034722715616, + 0.006186814978718758, + 0.036956969648599625, + 0.003503313986584544, + 0.0254246573895216, + 0.02582579292356968, + 0.012460584752261639, + 0.011929621919989586, + 0.00612002843990922, + 0.009281564503908157, + 0.07755232602357864, + 0.008333365432918072, + 0.0794135257601738, + 0.005836365278810263, + 0.010604280978441238, + 0.006479036528617144, + 0.0037080191541463137, + 0.006512995343655348, + 0.005779636092483997, + 0.004795327316969633, + 0.0037543962243944407, + 0.009176288731396198, + 0.003612047526985407, + 0.008669066242873669, + 0.0069559174589812756, + 0.003882749006152153, + 0.005646042060106993, + 0.003428778378292918, + 0.01199428178369999, + 0.009251057170331478, + 0.007581024896353483, + 0.007028704509139061, + 0.009600563906133175, + 0.005950156599283218, + 0.006049419287592173, + 0.012791269458830357, + 0.008649559691548347, + 0.012569773942232132, + 0.00989295169711113, + 0.013291901908814907, + 0.009208512492477894, + 0.008409326896071434, + 0.0066429367288947105, + 0.008852199651300907, + 0.00582145107910037, + 0.013434906490147114, + 0.009671199135482311, + 0.007346224505454302, + 0.005079949274659157, + 0.007039661053568125, + 0.00810728408396244, + 0.006214375142008066, + 0.006846831180155277, + 0.005817395634949207, + 0.004051652736961842, + 0.004154604859650135, + 0.002844555303454399, + 0.005145877134054899, + 0.005879239644855261, + 0.004015674348920584, + 0.0039141010493040085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0024876620154827833 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896587252616882 + ], + "opposite_logprob_tokens": [ + -2.578965902328491 + ], + "logprob_reward": -0.07896587252616882, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896587252616882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19190633296966553, + 0.028163447976112366, + 0.269449383020401, + 0.06844376027584076, + 0.12299275398254395, + 0.14753492176532745, + 0.06098958104848862, + 0.050727590918540955, + 0.026713110506534576, + 0.03307914361357689 + ], + "credit_map": [ + 0.024390891194343567, + 0.011911610141396523, + 0.01586076058447361, + 0.00946848001331091, + 0.012442226521670818, + 0.06059669703245163, + 0.011092589236795902, + 0.0461430586874485, + 0.0029291417449712753, + 0.003032463602721691, + 0.003759252605959773, + 0.00229442585259676, + 0.004985779989510775, + 0.005627127829939127, + 0.0018818376120179892, + 0.0036534180399030447, + 0.021787727251648903, + 0.012357614003121853, + 0.00908651016652584, + 0.0058727264404296875, + 0.019980039447546005, + 0.11992429196834564, + 0.00543505372479558, + 0.07500540465116501, + 0.011984598822891712, + 0.008255591616034508, + 0.008685319684445858, + 0.007105719298124313, + 0.006166485138237476, + 0.007731697056442499, + 0.00421276967972517, + 0.014301575720310211, + 0.016004472970962524, + 0.009009533561766148, + 0.012385917827486992, + 0.007462858688086271, + 0.00876218918710947, + 0.025082169100642204, + 0.008451706729829311, + 0.035833898931741714, + 0.026958728209137917, + 0.010400122031569481, + 0.012154079042375088, + 0.009628956206142902, + 0.013356656767427921, + 0.020728394389152527, + 0.00796101801097393, + 0.04634695500135422, + 0.009439597837626934, + 0.007550798822194338, + 0.007179104257375002, + 0.006307380273938179, + 0.006116878241300583, + 0.005475692451000214, + 0.01025924552232027, + 0.008660882711410522, + 0.00921500101685524, + 0.006032979115843773, + 0.007889939472079277, + 0.0050599039532244205, + 0.005091371014714241, + 0.005630464758723974, + 0.0059950384311378, + 0.005812888965010643, + 0.00501751946285367, + 0.00438928185030818, + 0.002538357162848115, + 0.0022869424428790808, + 0.0036672584246844053, + 0.002851189346984029, + 0.0037190313450992107, + 0.002243529772385955, + 0.005114616360515356, + 0.004006828181445599, + 0.003016093512997031, + 0.004533074330538511, + 0.0041402243077754974, + 0.004385471343994141, + 0.004409511107951403, + 0.003473325865343213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.07896587252616882 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive airship resembling a giant blue whale with a rounded nose and tapered tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the airship's skin a seamless pearlescent off-white fabric with faintly visible internal structural ribs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a pastel sky with a gradient of pale rose, creamy orange, and lavender dotted with wispy clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a warm golden glow from a low off-camera sun creating long gentle shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, perfectly steady tracking motion parallel to the airship?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900772780179977 + ], + "opposite_logprob_tokens": [ + -2.464007616043091 + ], + "logprob_reward": -0.08900772780179977, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900772780179977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10932609438896179, + 0.03638841211795807, + 0.27579551935195923, + 0.09324058145284653, + 0.22640614211559296, + 0.025612065568566322, + 0.05058453977108002, + 0.054874856024980545, + 0.08326681703329086, + 0.044504981487989426 + ], + "credit_map": [ + 0.012079390697181225, + 0.007099137641489506, + 0.009880234487354755, + 0.008724980987608433, + 0.019332069903612137, + 0.025315549224615097, + 0.006883528083562851, + 0.020011207088828087, + 0.005670421291142702, + 0.0037019632291048765, + 0.0051470124162733555, + 0.002793740713968873, + 0.004982455167919397, + 0.004501576069742441, + 0.00430170726031065, + 0.005289534572511911, + 0.02272922731935978, + 0.012438524514436722, + 0.009870525449514389, + 0.007079632021486759, + 0.019363608211278915, + 0.12021367996931076, + 0.00644083833321929, + 0.077659472823143, + 0.019560161978006363, + 0.0060938564129173756, + 0.006402906030416489, + 0.004714228678494692, + 0.010473172180354595, + 0.0333646722137928, + 0.004196405876427889, + 0.008435172960162163, + 0.03698646277189255, + 0.01589617133140564, + 0.014731228351593018, + 0.010824468918144703, + 0.010256596840918064, + 0.04631216451525688, + 0.013702556490898132, + 0.07769649475812912, + 0.0038482716772705317, + 0.0021875554230064154, + 0.005042750854045153, + 0.0022049115505069494, + 0.0022789991926401854, + 0.002869912190362811, + 0.0017659730510786176, + 0.005413692444562912, + 0.006522596813738346, + 0.0065684993751347065, + 0.005958018358796835, + 0.006166852544993162, + 0.0054539949633181095, + 0.0046087303198874, + 0.009968332014977932, + 0.005337515380233526, + 0.0073563102632761, + 0.006217361427843571, + 0.0067022331058979034, + 0.00811275839805603, + 0.006820083595812321, + 0.005036689806729555, + 0.009745690040290356, + 0.004883732181042433, + 0.017260676249861717, + 0.012458369135856628, + 0.008429421111941338, + 0.006875097751617432, + 0.010026978328824043, + 0.010302468203008175, + 0.007795084733515978, + 0.010118719190359116, + 0.007508313283324242, + 0.006012633442878723, + 0.005181971937417984, + 0.004441277123987675, + 0.006446850020438433, + 0.004058938939124346, + 0.005072667263448238, + 0.005782329011708498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.08900772780179977 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_12/index_tensor([8237, 8237])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_12/index_tensor([8237, 8237])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1d5f6dcd59de2eec9adbd62aa14afbdd568081ee --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_12/index_tensor([8237, 8237])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle shot from near the horizon, the camera tilted steadily upward, capturing in slow motion the moment a massive humpback whale breaches the surface with awe-inspiring power and grace. Beneath the golden glow of a setting sun, the beast's deep, almost black skin is covered in white barnacles and scarred by age. As its massive body almost completely breaks the surface, tons of seawater cascade down its smooth hide, breaking into countless shimmering, iridescent droplets in mid-air. Reaching its peak in mid-air, it gracefully twists sideways, fully extending its massive, wing-like pectoral fins. The ridged, milky white skin of its belly stands out against the warm light. Against the backdrop lies a vast expanse of deep blue ocean, shimmering with golden ripples. The distant sky is tinted in a gradient of warm orange and soft purple. The entire scene is imbued with the raw power of life and the majestic beauty of nature, creating a breathtaking, epic tranquility and awe.", + "seed": 3081, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_12/index_8237_seed_3081_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06174675300717354, + "vlm_logprob": -0.06174675300717354, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.06174675300717354, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05288558453321457 + ], + "opposite_logprob_tokens": [ + -3.8028855323791504 + ], + "logprob_reward": -0.05288558453321457, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.05288558453321457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2460487335920334, + 0.09674608707427979, + 0.08071582764387131, + 0.06536572426557541, + 0.05106547474861145, + 0.05952759459614754, + 0.10210929811000824, + 0.18184907734394073, + 0.06889171153306961, + 0.047680433839559555 + ], + "credit_map": [ + 0.03489234670996666, + 0.025948986411094666, + 0.014334323816001415, + 0.026049042120575905, + 0.04448767006397247, + 0.06357733905315399, + 0.01745578646659851, + 0.0193032156676054, + 0.011436326429247856, + 0.019920486956834793, + 0.009393793530762196, + 0.01133319828659296, + 0.012985285371541977, + 0.012816378846764565, + 0.010763708502054214, + 0.008096905425190926, + 0.008491641841828823, + 0.013255146332085133, + 0.007466287352144718, + 0.00822540745139122, + 0.00888596661388874, + 0.010432644747197628, + 0.011832378804683685, + 0.012126347050070763, + 0.006771525368094444, + 0.007836122997105122, + 0.00861731544137001, + 0.00952582061290741, + 0.004033817909657955, + 0.00869588553905487, + 0.008155664429068565, + 0.011729576624929905, + 0.005781287793070078, + 0.005247932393103838, + 0.007703625597059727, + 0.007700871676206589, + 0.003558976110070944, + 0.006906409747898579, + 0.0063348631374537945, + 0.007831505499780178, + 0.010120452381670475, + 0.006830678321421146, + 0.010397627018392086, + 0.0044312370009720325, + 0.00542857963591814, + 0.010617469437420368, + 0.0059050302952528, + 0.005796521902084351, + 0.010720659978687763, + 0.010405323468148708, + 0.01939796283841133, + 0.016765117645263672, + 0.00622732425108552, + 0.017866821959614754, + 0.012454969808459282, + 0.00827112514525652, + 0.028909562155604362, + 0.019477514550089836, + 0.029090680181980133, + 0.02176150120794773, + 0.015782620757818222, + 0.026948224753141403, + 0.024845512583851814, + 0.015033459290862083, + 0.008226822130382061, + 0.007578746415674686, + 0.009806801564991474, + 0.00819865707308054, + 0.004536368418484926, + 0.00895729660987854, + 0.014256758615374565, + 0.007330261170864105, + 0.005122845061123371, + 0.00554991140961647, + 0.006963884923607111, + 0.008655522018671036, + 0.0035732456017285585, + 0.004512985236942768, + 0.00801678653806448, + 0.005285249557346106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.05288558453321457 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016086433082818985 + ], + "opposite_logprob_tokens": [ + -4.141086578369141 + ], + "logprob_reward": -0.016086433082818985, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016086433082818985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21799726784229279, + 0.20196959376335144, + 0.13796330988407135, + 0.06345418095588684, + 0.08691752701997757, + 0.06297008693218231, + 0.04937303438782692, + 0.054829973727464676, + 0.09620033949613571, + 0.028324641287326813 + ], + "credit_map": [ + 0.030105669051408768, + 0.027105262503027916, + 0.0144147714599967, + 0.02492613159120083, + 0.03561387583613396, + 0.05073227733373642, + 0.01850026659667492, + 0.0165990199893713, + 0.020531682297587395, + 0.037485793232917786, + 0.020929083228111267, + 0.0184346754103899, + 0.028926052153110504, + 0.028887972235679626, + 0.023540915921330452, + 0.023233426734805107, + 0.014582366682589054, + 0.018104752525687218, + 0.018333880230784416, + 0.013448510318994522, + 0.013228598050773144, + 0.0169148538261652, + 0.023899823427200317, + 0.019450535997748375, + 0.009382300078868866, + 0.008890400640666485, + 0.006826729979366064, + 0.009862752631306648, + 0.005916079971939325, + 0.007130637299269438, + 0.007487867027521133, + 0.00795741006731987, + 0.009194605052471161, + 0.008905249647796154, + 0.013606371358036995, + 0.013207180425524712, + 0.00724221533164382, + 0.011939396150410175, + 0.010583333671092987, + 0.012239177711308002, + 0.007850264199078083, + 0.009536142461001873, + 0.008945634588599205, + 0.005806366913020611, + 0.00608521094545722, + 0.00953591801226139, + 0.007456528954207897, + 0.007754017133265734, + 0.005365029443055391, + 0.0033126750495284796, + 0.01212275680154562, + 0.0069406237453222275, + 0.0025685583241283894, + 0.009901344776153564, + 0.004907798953354359, + 0.004254249390214682, + 0.008057696744799614, + 0.005749653559178114, + 0.007284157443791628, + 0.006049129646271467, + 0.004567868076264858, + 0.006955714430660009, + 0.009866288863122463, + 0.006299464497715235, + 0.0146713238209486, + 0.010484257712960243, + 0.01355290599167347, + 0.012133754789829254, + 0.007777069695293903, + 0.008552120067179203, + 0.01734413579106331, + 0.01168476976454258, + 0.0031961617060005665, + 0.004285592585802078, + 0.004121797159314156, + 0.002366286003962159, + 0.0025531237479299307, + 0.0022201985120773315, + 0.004281603265553713, + 0.005299877375364304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.016086433082818985 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12709666788578033 + ], + "opposite_logprob_tokens": [ + -2.127096652984619 + ], + "logprob_reward": -0.12709666788578033, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12709666788578033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20143431425094604, + 0.0756671205163002, + 0.05696318298578262, + 0.05667702108621597, + 0.08061542361974716, + 0.1090698167681694, + 0.03381967172026634, + 0.1473836451768875, + 0.079948790371418, + 0.15842099487781525 + ], + "credit_map": [ + 0.030333079397678375, + 0.020770510658621788, + 0.013698807917535305, + 0.019839800894260406, + 0.04307565838098526, + 0.04358788579702377, + 0.01635848544538021, + 0.013770082965493202, + 0.008137092925608158, + 0.01613665372133255, + 0.006247805897146463, + 0.0075093708001077175, + 0.012468719854950905, + 0.010240579023957253, + 0.00760563462972641, + 0.0073212627321481705, + 0.006099037826061249, + 0.00641608564183116, + 0.008672327734529972, + 0.006745063699781895, + 0.004087192006409168, + 0.007881715893745422, + 0.006142148282378912, + 0.010919613763689995, + 0.007018598262220621, + 0.0070348745211958885, + 0.008921685628592968, + 0.008879470638930798, + 0.004192808177322149, + 0.0068029011599719524, + 0.007197949569672346, + 0.006628734990954399, + 0.008091885596513748, + 0.009701281785964966, + 0.012390189804136753, + 0.012528532184660435, + 0.007503223605453968, + 0.011453317478299141, + 0.010009069927036762, + 0.008937929756939411, + 0.0176126379519701, + 0.014247287064790726, + 0.01698252744972706, + 0.008583898656070232, + 0.012102010659873486, + 0.019054308533668518, + 0.010886861942708492, + 0.009600295685231686, + 0.003622898831963539, + 0.004238543100655079, + 0.004831991158425808, + 0.004216274246573448, + 0.003237394616007805, + 0.005486908834427595, + 0.003305500838905573, + 0.004880161490291357, + 0.020469719544053078, + 0.01920405775308609, + 0.020154763013124466, + 0.01641015149652958, + 0.014311064034700394, + 0.019295118749141693, + 0.022808847948908806, + 0.014729918912053108, + 0.018141644075512886, + 0.007203530054539442, + 0.012815551832318306, + 0.008264799602329731, + 0.005860996898263693, + 0.008005487732589245, + 0.01253205444663763, + 0.0071247234009206295, + 0.018268000334501266, + 0.011293641291558743, + 0.020536720752716064, + 0.02530629187822342, + 0.011930984444916248, + 0.017289871349930763, + 0.029729772359132767, + 0.024065714329481125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.12709666788578033 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896928489208221 + ], + "opposite_logprob_tokens": [ + -2.5789692401885986 + ], + "logprob_reward": -0.07896928489208221, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896928489208221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24763503670692444, + 0.16515885293483734, + 0.14179782569408417, + 0.0962369441986084, + 0.13746494054794312, + 0.03717900440096855, + 0.060076408088207245, + 0.046087708324193954, + 0.022494487464427948, + 0.04586881026625633 + ], + "credit_map": [ + 0.03822542354464531, + 0.022669712081551552, + 0.0182191114872694, + 0.015869194641709328, + 0.04858649894595146, + 0.06478217989206314, + 0.019985273480415344, + 0.01929766684770584, + 0.019688980653882027, + 0.03174746036529541, + 0.013014992699027061, + 0.017591439187526703, + 0.02389899641275406, + 0.0223414096981287, + 0.022717012092471123, + 0.01415855810046196, + 0.01517331600189209, + 0.015510434284806252, + 0.02142324671149254, + 0.01847894862294197, + 0.013346885330975056, + 0.016095297411084175, + 0.022405803203582764, + 0.019363896921277046, + 0.01367987785488367, + 0.013347544707357883, + 0.011608552187681198, + 0.016043931245803833, + 0.006832228973507881, + 0.012170634232461452, + 0.011348205618560314, + 0.011205964721739292, + 0.015167972072958946, + 0.013219866901636124, + 0.025851961225271225, + 0.021981632336974144, + 0.010497335344552994, + 0.021470483392477036, + 0.013574250973761082, + 0.015701446682214737, + 0.004340808838605881, + 0.0057411640882492065, + 0.004273599479347467, + 0.00404510460793972, + 0.002916438737884164, + 0.005014361813664436, + 0.0057203504256904125, + 0.00512717617675662, + 0.005570016801357269, + 0.007115960121154785, + 0.008362364023923874, + 0.008274358697235584, + 0.0056761084124445915, + 0.008570411242544651, + 0.006930576171725988, + 0.00957661122083664, + 0.004802900366485119, + 0.0052555520087480545, + 0.006145345512777567, + 0.007717812433838844, + 0.0036822445690631866, + 0.004909489769488573, + 0.009118874557316303, + 0.004455490503460169, + 0.003130128839984536, + 0.002654805313795805, + 0.0030793920159339905, + 0.002892410149797797, + 0.001883718534372747, + 0.0027115640696138144, + 0.0037339439149945974, + 0.002408526372164488, + 0.00489148311316967, + 0.0062865144573152065, + 0.007205244153738022, + 0.006024752277880907, + 0.003693363396450877, + 0.003701768582686782, + 0.0075922878459095955, + 0.006473396439105272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07896928489208221 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369579464197159 + ], + "opposite_logprob_tokens": [ + -3.408695697784424 + ], + "logprob_reward": -0.03369579464197159, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369579464197159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8132559657096863, + 0.028837131336331367, + 0.03530218452215195, + 0.02116103284060955, + 0.019422804936766624, + 0.020723262801766396, + 0.01650305464863777, + 0.017709271982312202, + 0.016315743327140808, + 0.01076955534517765 + ], + "credit_map": [ + 0.12581205368041992, + 0.07312709838151932, + 0.044537510722875595, + 0.05896538868546486, + 0.1714693009853363, + 0.22865618765354156, + 0.050243914127349854, + 0.06044453755021095, + 0.003705082694068551, + 0.004162123892456293, + 0.0026734559796750546, + 0.0035795150324702263, + 0.004729068372398615, + 0.003910050727427006, + 0.0033090943470597267, + 0.0027687400579452515, + 0.0037197633646428585, + 0.004313645418733358, + 0.00399200851097703, + 0.003755526151508093, + 0.0031100090127438307, + 0.003982222639024258, + 0.004179959651082754, + 0.008249052800238132, + 0.0029332987032830715, + 0.0026810960844159126, + 0.003752154763787985, + 0.0037832630332559347, + 0.0013025695225223899, + 0.0027750609442591667, + 0.002083034720271826, + 0.0018505541374906898, + 0.0030270223505795, + 0.0014561606803908944, + 0.003928599879145622, + 0.003009616630151868, + 0.0012856608955189586, + 0.002946204040199518, + 0.0016679390100762248, + 0.002101602265611291, + 0.0025736461393535137, + 0.0016032207058742642, + 0.005214477889239788, + 0.0023246558848768473, + 0.0012374208308756351, + 0.004457659088075161, + 0.0016002418706193566, + 0.0017119384137913585, + 0.0024018401745706797, + 0.0016296305693686008, + 0.0037489940878003836, + 0.0015328573063015938, + 0.0009634989546611905, + 0.002820596331730485, + 0.0015307640423998237, + 0.0018748720176517963, + 0.0032712756656110287, + 0.0015518213622272015, + 0.0026515719946473837, + 0.002001515356823802, + 0.0012847721809521317, + 0.002662608865648508, + 0.0026537799276411533, + 0.0016319273272529244, + 0.0031592068262398243, + 0.001656644162721932, + 0.0026223943568766117, + 0.0018319633090868592, + 0.001162321656011045, + 0.0018617665627971292, + 0.0025868641678243876, + 0.0014345813542604446, + 0.0010793714318424463, + 0.0014808356063440442, + 0.0015225335955619812, + 0.001451657502911985, + 0.0009115823777392507, + 0.0008728746324777603, + 0.002015582751482725, + 0.0014351174468174577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.03369579464197159 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot from near the horizon, the camera tilted steadily upward, capturing in slow motion the moment a massive humpback whale breaches the surface with awe-inspiring power and grace. Beneath the golden glow of a setting sun, the beast's deep, almost black skin is covered in white barnacles and scarred by age. As its massive body almost completely breaks the surface, tons of seawater cascade down its smooth hide, breaking into countless shimmering, iridescent droplets in mid-air. Reaching its peak in mid-air, it gracefully twists sideways, fully extending its massive, wing-like pectoral fins. The ridged, milky white skin of its belly stands out against the warm light. Against the backdrop lies a vast expanse of deep blue ocean, shimmering with golden ripples. The distant sky is tinted in a gradient of warm orange and soft purple. The entire scene is imbued with the raw power of life and the majestic beauty of nature, creating a breathtaking, epic tranquility and awe.", + "seed": 3082, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_12/index_8237_seed_3082_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07387560168281197, + "vlm_logprob": -0.07387560168281197, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.07387560168281197, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03236343339085579 + ], + "opposite_logprob_tokens": [ + -4.032363414764404 + ], + "logprob_reward": -0.03236343339085579, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.03236343339085579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39167118072509766, + 0.03863770142197609, + 0.058015402406454086, + 0.04113030061125755, + 0.07354655116796494, + 0.11431567370891571, + 0.1549796164035797, + 0.027761230245232582, + 0.06170610710978508, + 0.03823617100715637 + ], + "credit_map": [ + 0.028134308755397797, + 0.057993438094854355, + 0.03434041142463684, + 0.01459075603634119, + 0.15522779524326324, + 0.06585673987865448, + 0.016280336305499077, + 0.019247401505708694, + 0.004504695534706116, + 0.00867663137614727, + 0.004152835346758366, + 0.004058235790580511, + 0.005518033634871244, + 0.005193301476538181, + 0.0030359697993844748, + 0.0034980003256350756, + 0.00919041596353054, + 0.005549561232328415, + 0.010453587397933006, + 0.005711641628295183, + 0.003659166395664215, + 0.011655045673251152, + 0.008095023222267628, + 0.0037009615916758776, + 0.004702207166701555, + 0.006207020487636328, + 0.004297340754419565, + 0.003975424449890852, + 0.005867426283657551, + 0.004611878655850887, + 0.0078036426566541195, + 0.0036653601564466953, + 0.011560783721506596, + 0.008286494761705399, + 0.00900304690003395, + 0.011931868270039558, + 0.006442561745643616, + 0.01095794141292572, + 0.008643064647912979, + 0.006720793899148703, + 0.024129854515194893, + 0.014140332117676735, + 0.01217834372073412, + 0.009511222131550312, + 0.009317349642515182, + 0.021880796179175377, + 0.014231325127184391, + 0.008926448412239552, + 0.021239569410681725, + 0.013212095946073532, + 0.03573715314269066, + 0.024285996332764626, + 0.013575293123722076, + 0.012639501132071018, + 0.017899487167596817, + 0.016390521079301834, + 0.004237718414515257, + 0.002474073087796569, + 0.006898981053382158, + 0.0035376690793782473, + 0.0027079060673713684, + 0.00260803266428411, + 0.002240808680653572, + 0.0030560404993593693, + 0.008614180609583855, + 0.0046225301921367645, + 0.014561140909790993, + 0.004892156925052404, + 0.00506963487714529, + 0.008030089549720287, + 0.004281311295926571, + 0.011635061353445053, + 0.0038884589448571205, + 0.006221144460141659, + 0.007776294369250536, + 0.004373051691800356, + 0.004379965830594301, + 0.005139906890690327, + 0.0027884815353900194, + 0.0036688665859401226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.03236343339085579 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005267313215881586 + ], + "opposite_logprob_tokens": [ + -5.255267143249512 + ], + "logprob_reward": -0.005267313215881586, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005267313215881586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2586432993412018, + 0.11678637564182281, + 0.08284065127372742, + 0.0383564718067646, + 0.08960310369729996, + 0.2040177285671234, + 0.07849837094545364, + 0.026980431750416756, + 0.027226030826568604, + 0.07704755663871765 + ], + "credit_map": [ + 0.06026897579431534, + 0.014999892562627792, + 0.017199547961354256, + 0.011885639280080795, + 0.08886279910802841, + 0.037987928837537766, + 0.014845547266304493, + 0.01259297039359808, + 0.012680213898420334, + 0.02487013302743435, + 0.011010222136974335, + 0.01299301628023386, + 0.018829846754670143, + 0.013474714010953903, + 0.011033588089048862, + 0.0118946498259902, + 0.014038506895303726, + 0.010055921040475368, + 0.010737755335867405, + 0.008511597290635109, + 0.005389669444411993, + 0.018485097214579582, + 0.008566154167056084, + 0.007055951748043299, + 0.00526911485940218, + 0.005849362351000309, + 0.003934739623218775, + 0.004753528628498316, + 0.003978027030825615, + 0.0056514129973948, + 0.005868218839168549, + 0.0030520688742399216, + 0.013738719746470451, + 0.013430331833660603, + 0.00899857934564352, + 0.01217482890933752, + 0.010873604565858841, + 0.011994087137281895, + 0.01108348648995161, + 0.007309464272111654, + 0.043592337518930435, + 0.017133401706814766, + 0.030776331201195717, + 0.022169848904013634, + 0.011531366035342216, + 0.03885020315647125, + 0.025527384132146835, + 0.014436866156756878, + 0.012319752015173435, + 0.0073832678608596325, + 0.018580259755253792, + 0.011621975339949131, + 0.006123997271060944, + 0.006960421800613403, + 0.008234888315200806, + 0.007273808121681213, + 0.0040739490650594234, + 0.0016968166455626488, + 0.00888748001307249, + 0.002299512503668666, + 0.0022599780932068825, + 0.0037290232721716166, + 0.0015142697375267744, + 0.002519400790333748, + 0.0034384047612547874, + 0.0024183413479477167, + 0.008378931321203709, + 0.003027933882549405, + 0.0031316024251282215, + 0.0021841847337782383, + 0.001629871898330748, + 0.0030167619697749615, + 0.011785951443016529, + 0.011090802028775215, + 0.014072335325181484, + 0.009784369729459286, + 0.008364344015717506, + 0.007834305055439472, + 0.006429584231227636, + 0.007685868535190821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.005267313215881586 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28137722611427307 + ], + "opposite_logprob_tokens": [ + -1.4063771963119507 + ], + "logprob_reward": -0.28137722611427307, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28137722611427307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20200997591018677, + 0.09015530347824097, + 0.03491373360157013, + 0.04753802344202995, + 0.07489624619483948, + 0.24285799264907837, + 0.11561010032892227, + 0.054848529398441315, + 0.06285470724105835, + 0.07431536912918091 + ], + "credit_map": [ + 0.055784985423088074, + 0.018565723672509193, + 0.02366890385746956, + 0.014086787588894367, + 0.044039998203516006, + 0.021698297932744026, + 0.013443248346447945, + 0.01072203740477562, + 0.0094603905454278, + 0.011520536616444588, + 0.011029423214495182, + 0.015491586178541183, + 0.011828843504190445, + 0.01102980226278305, + 0.009216978214681149, + 0.01057773269712925, + 0.004202975891530514, + 0.004533954430371523, + 0.0051074279472231865, + 0.004775889217853546, + 0.003550780238583684, + 0.005878570955246687, + 0.0037087020464241505, + 0.0031554324086755514, + 0.007524408865720034, + 0.0038414043374359608, + 0.007368035614490509, + 0.008911814540624619, + 0.0033098547719419003, + 0.007703548762947321, + 0.005833243951201439, + 0.0030457121320068836, + 0.011532705277204514, + 0.008357822895050049, + 0.009485005401074886, + 0.010137636214494705, + 0.006471497006714344, + 0.008740884251892567, + 0.011829315684735775, + 0.008341376669704914, + 0.05104179307818413, + 0.037626974284648895, + 0.026713868603110313, + 0.01935085467994213, + 0.01959439367055893, + 0.04616551846265793, + 0.027153119444847107, + 0.015211464837193489, + 0.01830836571753025, + 0.010027029551565647, + 0.020806320011615753, + 0.018914876505732536, + 0.011230637319386005, + 0.014662384055554867, + 0.009736219421029091, + 0.011924267746508121, + 0.007818752899765968, + 0.0037686359137296677, + 0.021713778376579285, + 0.004697723779827356, + 0.0035691235680133104, + 0.006512487772852182, + 0.0024226768873631954, + 0.004345355089753866, + 0.014638779684901237, + 0.005541511811316013, + 0.01540453266352415, + 0.004935320001095533, + 0.004670409485697746, + 0.004662270192056894, + 0.004679199308156967, + 0.008322683162987232, + 0.0075018941424787045, + 0.014340261928737164, + 0.010052906349301338, + 0.009014234878122807, + 0.007534665521234274, + 0.011071787215769291, + 0.007508995942771435, + 0.007290623150765896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.28137722611427307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794851318001747 + ], + "opposite_logprob_tokens": [ + -3.529794931411743 + ], + "logprob_reward": -0.029794851318001747, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794851318001747, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12355165183544159, + 0.13695552945137024, + 0.0873606950044632, + 0.06925750523805618, + 0.11346696317195892, + 0.1759215146303177, + 0.049263905733823776, + 0.0794961228966713, + 0.07343646883964539, + 0.09128965437412262 + ], + "credit_map": [ + 0.03487468510866165, + 0.013566878624260426, + 0.016621088609099388, + 0.009783503599464893, + 0.020580148324370384, + 0.012796662747859955, + 0.008367289789021015, + 0.006961398292332888, + 0.013900039717555046, + 0.03434548154473305, + 0.011226232163608074, + 0.016133632510900497, + 0.01961153745651245, + 0.016261370852589607, + 0.011371022090315819, + 0.014106214046478271, + 0.014157535508275032, + 0.014728833921253681, + 0.009700650349259377, + 0.00639166496694088, + 0.006548579316586256, + 0.021716326475143433, + 0.008530857972800732, + 0.0055862488225102425, + 0.0086135258898139, + 0.009910829365253448, + 0.008747050538659096, + 0.00849098339676857, + 0.005846360698342323, + 0.009359246119856834, + 0.011110314168035984, + 0.007179196923971176, + 0.01587294600903988, + 0.014886226505041122, + 0.01330037321895361, + 0.014876064844429493, + 0.00818604789674282, + 0.015103770419955254, + 0.02106684073805809, + 0.01017469447106123, + 0.03784320503473282, + 0.013789934106171131, + 0.03429952636361122, + 0.018956605345010757, + 0.007515440229326487, + 0.032866887748241425, + 0.018482934683561325, + 0.012166988104581833, + 0.008015082217752934, + 0.005339810624718666, + 0.008186490274965763, + 0.008017182350158691, + 0.004009464755654335, + 0.005660209804773331, + 0.0049436758272349834, + 0.005091990809887648, + 0.012006888166069984, + 0.005000600125640631, + 0.028252990916371346, + 0.006832558661699295, + 0.00543719157576561, + 0.010529457591474056, + 0.004859298933297396, + 0.006577143445611, + 0.009732840582728386, + 0.006304789334535599, + 0.017846031114459038, + 0.007854706607758999, + 0.009037136100232601, + 0.009347955696284771, + 0.0050500561483204365, + 0.008262949995696545, + 0.010522955097258091, + 0.013993424363434315, + 0.01656593382358551, + 0.01001049391925335, + 0.00892656296491623, + 0.011885306797921658, + 0.011034050956368446, + 0.008350924588739872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.029794851318001747 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575184375047684 + ], + "opposite_logprob_tokens": [ + -3.8955752849578857 + ], + "logprob_reward": -0.020575184375047684, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575184375047684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21517015993595123, + 0.0670223981142044, + 0.0694270133972168, + 0.051025163382291794, + 0.13511833548545837, + 0.03958522155880928, + 0.10901523381471634, + 0.031152909621596336, + 0.25496888160705566, + 0.027514666318893433 + ], + "credit_map": [ + 0.02180112898349762, + 0.03091481886804104, + 0.013816461898386478, + 0.018573403358459473, + 0.046151302754879, + 0.046805329620838165, + 0.0144173139706254, + 0.022690419107675552, + 0.007354690693318844, + 0.011627426370978355, + 0.0074506839737296104, + 0.008336336351931095, + 0.007297954987734556, + 0.008712153881788254, + 0.008468793705105782, + 0.007774357683956623, + 0.013129094615578651, + 0.009627939201891422, + 0.007838640362024307, + 0.009290609508752823, + 0.004614759236574173, + 0.013219069689512253, + 0.007701449561864138, + 0.004005448427051306, + 0.009495931677520275, + 0.004383912775665522, + 0.007523307576775551, + 0.006982876919209957, + 0.003228219458833337, + 0.009318615309894085, + 0.0061872247606515884, + 0.003905070712789893, + 0.022132473066449165, + 0.009451725520193577, + 0.02412566915154457, + 0.02587573230266571, + 0.007502779830247164, + 0.01955907605588436, + 0.013806264847517014, + 0.012664617970585823, + 0.007572764530777931, + 0.0029347105883061886, + 0.009269265457987785, + 0.0050171054899692535, + 0.0016388102667406201, + 0.005747584160417318, + 0.004593453835695982, + 0.0028115250170230865, + 0.01580827869474888, + 0.007177144289016724, + 0.028541546314954758, + 0.018522243946790695, + 0.007509246468544006, + 0.010081850923597813, + 0.010486757382750511, + 0.010888161137700081, + 0.005855794996023178, + 0.0021195746958255768, + 0.007470428477972746, + 0.00374489719979465, + 0.001889498671516776, + 0.006245692726224661, + 0.0017204704927280545, + 0.002106552943587303, + 0.04837258905172348, + 0.01824457384645939, + 0.06362035125494003, + 0.020030494779348373, + 0.01795145496726036, + 0.02409921959042549, + 0.016847427934408188, + 0.04580274596810341, + 0.003018595278263092, + 0.004088979680091143, + 0.005499221384525299, + 0.0037614849861711264, + 0.0026456774212419987, + 0.002805953612551093, + 0.0027700173668563366, + 0.002924733329564333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.020575184375047684 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot from near the horizon, the camera tilted steadily upward, capturing in slow motion the moment a massive humpback whale breaches the surface with awe-inspiring power and grace. Beneath the golden glow of a setting sun, the beast's deep, almost black skin is covered in white barnacles and scarred by age. As its massive body almost completely breaks the surface, tons of seawater cascade down its smooth hide, breaking into countless shimmering, iridescent droplets in mid-air. Reaching its peak in mid-air, it gracefully twists sideways, fully extending its massive, wing-like pectoral fins. The ridged, milky white skin of its belly stands out against the warm light. Against the backdrop lies a vast expanse of deep blue ocean, shimmering with golden ripples. The distant sky is tinted in a gradient of warm orange and soft purple. The entire scene is imbued with the raw power of life and the majestic beauty of nature, creating a breathtaking, epic tranquility and awe.", + "seed": 3083, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_12/index_8237_seed_3083_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04411726826801896, + "vlm_logprob": -0.04411726826801896, + "vlm_yes_no_margin": 3.7749999523162843, + "vlm_token_credit": -0.04411726826801896, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03634900227189064 + ], + "opposite_logprob_tokens": [ + -4.036348819732666 + ], + "logprob_reward": -0.03634900227189064, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.03634900227189064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05681033432483673, + 0.05227052792906761, + 0.051985643804073334, + 0.08003007620573044, + 0.3182491958141327, + 0.05193251743912697, + 0.2635727822780609, + 0.03632211312651634, + 0.02710532583296299, + 0.061721526086330414 + ], + "credit_map": [ + 0.006738059688359499, + 0.007389029487967491, + 0.0065079438500106335, + 0.007101481780409813, + 0.008819789625704288, + 0.00748860090970993, + 0.0062733907252550125, + 0.006492035463452339, + 0.0033579529263079166, + 0.008501533418893814, + 0.0063260337337851524, + 0.0072998590767383575, + 0.007484328467398882, + 0.008226890116930008, + 0.006464713253080845, + 0.004609213676303625, + 0.005160568282008171, + 0.004385020118206739, + 0.005426659248769283, + 0.007222763728350401, + 0.00593465706333518, + 0.008197501301765442, + 0.00804478581994772, + 0.007613687310367823, + 0.010240019299089909, + 0.012696066871285439, + 0.010147794149816036, + 0.008728543296456337, + 0.011298059485852718, + 0.008515638299286366, + 0.010440753772854805, + 0.007963187992572784, + 0.013043437153100967, + 0.00733494246378541, + 0.12520624697208405, + 0.016114700585603714, + 0.010452997870743275, + 0.11969935148954391, + 0.014571713283658028, + 0.01182576734572649, + 0.004379795864224434, + 0.003183643566444516, + 0.007480866741389036, + 0.00475379778072238, + 0.0040518357418477535, + 0.009407376870512962, + 0.01231515035033226, + 0.006360047962516546, + 0.030100172385573387, + 0.025297153741121292, + 0.0198668222874403, + 0.019012942910194397, + 0.032082438468933105, + 0.06081975996494293, + 0.030833572149276733, + 0.04555990919470787, + 0.003404556307941675, + 0.00475727766752243, + 0.005172498524188995, + 0.005174006801098585, + 0.004358069971203804, + 0.0064086648635566235, + 0.003074572654440999, + 0.003972463309764862, + 0.003142724744975567, + 0.002488110912963748, + 0.00466700317338109, + 0.003124528331682086, + 0.0036584483459591866, + 0.003053952008485794, + 0.0036824208218604326, + 0.0032881353981792927, + 0.008658003993332386, + 0.003595472779124975, + 0.005473634693771601, + 0.0029710764065384865, + 0.01414569467306137, + 0.01963152177631855, + 0.0038015502505004406, + 0.003444572212174535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 3 + ], + "selected_reward": -0.03634900227189064 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008650213479995728 + ], + "opposite_logprob_tokens": [ + -4.758650302886963 + ], + "logprob_reward": -0.008650213479995728, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008650213479995728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06914417445659637, + 0.08065225183963776, + 0.12347230315208435, + 0.15103712677955627, + 0.10771863162517548, + 0.12528476119041443, + 0.11332620680332184, + 0.10421159118413925, + 0.05318179354071617, + 0.07197112590074539 + ], + "credit_map": [ + 0.005035359878093004, + 0.008890443481504917, + 0.007857752032577991, + 0.011677536182105541, + 0.00856783427298069, + 0.008019939996302128, + 0.007960996590554714, + 0.011134311556816101, + 0.008880279958248138, + 0.009899519383907318, + 0.006809075828641653, + 0.01888556033372879, + 0.007522897329181433, + 0.008144964464008808, + 0.00927578005939722, + 0.011234178207814693, + 0.020832566544413567, + 0.009037059731781483, + 0.011176047846674919, + 0.011190506629645824, + 0.018610579892992973, + 0.02299654111266136, + 0.014681551605463028, + 0.014947453513741493, + 0.02246679551899433, + 0.020544348284602165, + 0.025586221367120743, + 0.014635094441473484, + 0.018620705232024193, + 0.017275793477892876, + 0.017096351832151413, + 0.014811811968684196, + 0.005104388576000929, + 0.003815817879512906, + 0.04012840613722801, + 0.007406230550259352, + 0.0050981417298316956, + 0.036671463400125504, + 0.004865600261837244, + 0.004628575406968594, + 0.006523567717522383, + 0.005733202211558819, + 0.022236328572034836, + 0.004793492611497641, + 0.010899871587753296, + 0.04222774878144264, + 0.019452299922704697, + 0.013418243266642094, + 0.013403449207544327, + 0.008917829021811485, + 0.008996693417429924, + 0.01031607948243618, + 0.013373035937547684, + 0.024633649736642838, + 0.014193888753652573, + 0.019491584971547127, + 0.009741156361997128, + 0.01365067157894373, + 0.02373455837368965, + 0.012139900587499142, + 0.011435788124799728, + 0.012425810098648071, + 0.009185973554849625, + 0.011897744610905647, + 0.006364407483488321, + 0.0038151636254042387, + 0.01195264421403408, + 0.007168178912252188, + 0.0038408921100199223, + 0.004673991817981005, + 0.00928951520472765, + 0.006076995749026537, + 0.010180397890508175, + 0.003292417386546731, + 0.004730719141662121, + 0.0026470981538295746, + 0.018703101202845573, + 0.02616339735686779, + 0.003374736290425062, + 0.0028792577795684338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.008650213479995728 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033340007066727 + ], + "opposite_logprob_tokens": [ + -2.3503334522247314 + ], + "logprob_reward": -0.10033340007066727, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033340007066727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07938393950462341, + 0.10461688041687012, + 0.08112221956253052, + 0.07363486289978027, + 0.09264063835144043, + 0.04943627864122391, + 0.10601112991571426, + 0.1369822472333908, + 0.07320728898048401, + 0.20296449959278107 + ], + "credit_map": [ + 0.009033986367285252, + 0.00766005739569664, + 0.007626124657690525, + 0.011436069384217262, + 0.012661510147154331, + 0.009324624203145504, + 0.009821215644478798, + 0.011820349842309952, + 0.011353774927556515, + 0.01403915323317051, + 0.010828990489244461, + 0.018461529165506363, + 0.011912995018064976, + 0.018794197589159012, + 0.010260158218443394, + 0.008966087363660336, + 0.01113987248390913, + 0.007665171287953854, + 0.006898844614624977, + 0.008655204437673092, + 0.010726537555456161, + 0.017362570390105247, + 0.00928468070924282, + 0.00938933901488781, + 0.012055306695401669, + 0.010855482891201973, + 0.008340075612068176, + 0.0060040815733373165, + 0.010277392342686653, + 0.008105892688035965, + 0.009574821218848228, + 0.008421813137829304, + 0.005684980191290379, + 0.0045280857011675835, + 0.0287180133163929, + 0.00701899453997612, + 0.007301448844373226, + 0.024077847599983215, + 0.006984169129282236, + 0.008327099494636059, + 0.00620712386444211, + 0.004840822424739599, + 0.00514656025916338, + 0.0038319865707308054, + 0.005046652164310217, + 0.010849365964531898, + 0.007791704498231411, + 0.005722061265259981, + 0.012213647365570068, + 0.01077840756624937, + 0.010574557818472385, + 0.016099540516734123, + 0.009970491752028465, + 0.017451662570238113, + 0.013002114370465279, + 0.01592070236802101, + 0.014223468489944935, + 0.016653474420309067, + 0.03210175782442093, + 0.018809663131833076, + 0.013515644706785679, + 0.012769593857228756, + 0.013738042674958706, + 0.015170594677329063, + 0.00955210905522108, + 0.004861027467995882, + 0.009039494208991528, + 0.010854806751012802, + 0.007631033193320036, + 0.006690219044685364, + 0.006609909702092409, + 0.017968690022826195, + 0.027147984132170677, + 0.010451248846948147, + 0.01676042750477791, + 0.012833783403038979, + 0.04801874980330467, + 0.06690002232789993, + 0.01107963640242815, + 0.009772646240890026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.10033340007066727 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700051486492157 + ], + "opposite_logprob_tokens": [ + -2.695005178451538 + ], + "logprob_reward": -0.0700051486492157, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700051486492157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06500361114740372, + 0.13354620337486267, + 0.07635320723056793, + 0.08398944139480591, + 0.38023021817207336, + 0.040341854095458984, + 0.07753006368875504, + 0.03765886276960373, + 0.055601004511117935, + 0.04974551871418953 + ], + "credit_map": [ + 0.004424092825502157, + 0.0072231777012348175, + 0.005974111147224903, + 0.009085163474082947, + 0.014324567280709743, + 0.009243245236575603, + 0.007091477047652006, + 0.007637779228389263, + 0.012178700417280197, + 0.019572420045733452, + 0.011751650832593441, + 0.028191793709993362, + 0.014588238671422005, + 0.021973999217152596, + 0.014156851917505264, + 0.011132556013762951, + 0.008170181885361671, + 0.005204319953918457, + 0.005232715047895908, + 0.009232149459421635, + 0.010202855803072453, + 0.022185025736689568, + 0.007527180947363377, + 0.008598783053457737, + 0.012208690866827965, + 0.010622586123645306, + 0.010689396411180496, + 0.011384121142327785, + 0.009800395928323269, + 0.008452024310827255, + 0.011392497457563877, + 0.009439733810722828, + 0.015791060402989388, + 0.00831506960093975, + 0.14920784533023834, + 0.02006877213716507, + 0.012963149696588516, + 0.14357157051563263, + 0.01620127260684967, + 0.014111523516476154, + 0.0030856169760227203, + 0.0027604203205555677, + 0.004770670551806688, + 0.00215704832226038, + 0.0038748455699533224, + 0.011733178980648518, + 0.0071835643611848354, + 0.004776515066623688, + 0.009269353933632374, + 0.009536073543131351, + 0.007075955159962177, + 0.00828902330249548, + 0.007630521431565285, + 0.013556471094489098, + 0.009180803783237934, + 0.012991874478757381, + 0.0036440580151975155, + 0.0035790344700217247, + 0.007219007238745689, + 0.005358911119401455, + 0.004761104471981525, + 0.00450359471142292, + 0.004560752771794796, + 0.004032404161989689, + 0.0063467673026025295, + 0.003764174645766616, + 0.014524991624057293, + 0.006137690041214228, + 0.004969372414052486, + 0.005157088860869408, + 0.007193573750555515, + 0.0075073521584272385, + 0.006890572141855955, + 0.002995437476783991, + 0.005615915171802044, + 0.0032878005877137184, + 0.01032599713653326, + 0.013425985351204872, + 0.004092507530003786, + 0.0031113086733967066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 3 + ], + "selected_reward": -0.0700051486492157 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248576868325472 + ], + "opposite_logprob_tokens": [ + -5.255248546600342 + ], + "logprob_reward": -0.005248576868325472, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248576868325472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17316770553588867, + 0.08788973838090897, + 0.09133558720350266, + 0.07245554774999619, + 0.049580179154872894, + 0.03574138507246971, + 0.3281010389328003, + 0.04658874124288559, + 0.03135610744357109, + 0.08378398418426514 + ], + "credit_map": [ + 0.018237516283988953, + 0.035501789301633835, + 0.017106637358665466, + 0.022271156311035156, + 0.024657946079969406, + 0.017457865178585052, + 0.01856452412903309, + 0.019370265305042267, + 0.008398843929171562, + 0.009634207002818584, + 0.006645621731877327, + 0.021003909409046173, + 0.009395716711878777, + 0.013829962350428104, + 0.009441176429390907, + 0.00954030267894268, + 0.011284918524324894, + 0.011398779228329659, + 0.007614905945956707, + 0.015573538839817047, + 0.010575261898338795, + 0.0124724255874753, + 0.00993322767317295, + 0.012482527643442154, + 0.00970727764070034, + 0.008170172572135925, + 0.007816746830940247, + 0.006009794771671295, + 0.011263453401625156, + 0.011436368338763714, + 0.00908682867884636, + 0.008964902721345425, + 0.005197515711188316, + 0.0035254312679171562, + 0.009884067811071873, + 0.005769821349531412, + 0.004245177377015352, + 0.006951386574655771, + 0.005360941868275404, + 0.00864583533257246, + 0.0059615070931613445, + 0.00389747298322618, + 0.005885164253413677, + 0.003981365822255611, + 0.0025793814565986395, + 0.0038631767965853214, + 0.004778004717081785, + 0.0047953128814697266, + 0.041281621903181076, + 0.03035101480782032, + 0.025455186143517494, + 0.027903558686375618, + 0.037189606577157974, + 0.07108892500400543, + 0.03946125879883766, + 0.05536986514925957, + 0.004938299767673016, + 0.006644684821367264, + 0.007323900703340769, + 0.007352584972977638, + 0.005014384165406227, + 0.006141482386738062, + 0.0038990702014416456, + 0.005274335388094187, + 0.0033835433423519135, + 0.0027216782327741385, + 0.0060304380021989346, + 0.0037406866904348135, + 0.003186403075233102, + 0.0030888444744050503, + 0.0029788489919155836, + 0.0062256669625639915, + 0.011960511095821857, + 0.006107956636697054, + 0.008440814912319183, + 0.005272507201880217, + 0.017923004925251007, + 0.025107188150286674, + 0.0044072652235627174, + 0.0045647332444787025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.005248576868325472 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot from near the horizon, the camera tilted steadily upward, capturing in slow motion the moment a massive humpback whale breaches the surface with awe-inspiring power and grace. Beneath the golden glow of a setting sun, the beast's deep, almost black skin is covered in white barnacles and scarred by age. As its massive body almost completely breaks the surface, tons of seawater cascade down its smooth hide, breaking into countless shimmering, iridescent droplets in mid-air. Reaching its peak in mid-air, it gracefully twists sideways, fully extending its massive, wing-like pectoral fins. The ridged, milky white skin of its belly stands out against the warm light. Against the backdrop lies a vast expanse of deep blue ocean, shimmering with golden ripples. The distant sky is tinted in a gradient of warm orange and soft purple. The entire scene is imbued with the raw power of life and the majestic beauty of nature, creating a breathtaking, epic tranquility and awe.", + "seed": 3084, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_12/index_8237_seed_3084_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.021290703630074858, + "vlm_logprob": -0.021290703630074858, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.021290703630074858, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02927899919450283 + ], + "opposite_logprob_tokens": [ + -4.5292792320251465 + ], + "logprob_reward": -0.02927899919450283, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.02927899919450283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.170668825507164, + 0.03790294751524925, + 0.10857364535331726, + 0.11613379418849945, + 0.09750230610370636, + 0.07671035081148148, + 0.26261645555496216, + 0.03724052011966705, + 0.030734488740563393, + 0.06191668286919594 + ], + "credit_map": [ + 0.020317398011684418, + 0.022085610777139664, + 0.017008548602461815, + 0.02237013727426529, + 0.029471950605511665, + 0.03277279809117317, + 0.014413350261747837, + 0.012229029089212418, + 0.00541139068081975, + 0.004898935090750456, + 0.005403328686952591, + 0.0043816291727125645, + 0.005609878338873386, + 0.004801257513463497, + 0.003448502393439412, + 0.00394802400842309, + 0.013919525779783726, + 0.012895181775093079, + 0.01799960993230343, + 0.014474843628704548, + 0.012927012518048286, + 0.011547008529305458, + 0.012829831801354885, + 0.011980629526078701, + 0.013806982897222042, + 0.010262539610266685, + 0.027967127040028572, + 0.010243876837193966, + 0.007654201239347458, + 0.02910219505429268, + 0.010164047591388226, + 0.006932820659130812, + 0.022749552503228188, + 0.013148813508450985, + 0.00918351300060749, + 0.008935358375310898, + 0.010064143687486649, + 0.011172682978212833, + 0.012609024532139301, + 0.009639213792979717, + 0.014767920598387718, + 0.005765278823673725, + 0.016345055773854256, + 0.011025029234588146, + 0.005057384259998798, + 0.009070264175534248, + 0.007447594776749611, + 0.007231825031340122, + 0.08231425285339355, + 0.016307400539517403, + 0.07806043326854706, + 0.01940644346177578, + 0.01141714584082365, + 0.026414718478918076, + 0.017063802108168602, + 0.011632253415882587, + 0.003770882962271571, + 0.0036831246688961983, + 0.005179283209145069, + 0.0040902066975831985, + 0.0057756551541388035, + 0.003796007251366973, + 0.005381689872592688, + 0.005563669838011265, + 0.0034000424202531576, + 0.004510113038122654, + 0.005902424454689026, + 0.004629716742783785, + 0.00486341118812561, + 0.0019671826157718897, + 0.003088151104748249, + 0.0023734471760690212, + 0.008846502751111984, + 0.009885886684060097, + 0.0093344422057271, + 0.008281954564154148, + 0.006889856420457363, + 0.005701814312487841, + 0.006647300440818071, + 0.006328926887363195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.02927899919450283 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003201955696567893 + ], + "opposite_logprob_tokens": [ + -5.753201961517334 + ], + "logprob_reward": -0.003201955696567893, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003201955696567893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3010621964931488, + 0.06891961395740509, + 0.11585492640733719, + 0.12564685940742493, + 0.053415168076753616, + 0.07470960915088654, + 0.09958293288946152, + 0.060484617948532104, + 0.05417414754629135, + 0.046149905771017075 + ], + "credit_map": [ + 0.05057303234934807, + 0.03555476292967796, + 0.01923920027911663, + 0.037507060915231705, + 0.036580346524715424, + 0.08022236824035645, + 0.02596208080649376, + 0.015423361212015152, + 0.011543347500264645, + 0.008462619967758656, + 0.007769028656184673, + 0.005673668812960386, + 0.008003138937056065, + 0.015283418819308281, + 0.006592891179025173, + 0.005591508466750383, + 0.01245358306914568, + 0.01757807843387127, + 0.016030000522732735, + 0.015065115876495838, + 0.013957235030829906, + 0.015735814347863197, + 0.013902806676924229, + 0.011132301762700081, + 0.03137660026550293, + 0.0076403385028243065, + 0.017032407224178314, + 0.012477790005505085, + 0.009814515709877014, + 0.02942190319299698, + 0.009341923519968987, + 0.008541390299797058, + 0.009642103686928749, + 0.006555749103426933, + 0.006936631631106138, + 0.004305281676352024, + 0.0048639727756381035, + 0.006674342323094606, + 0.009424211457371712, + 0.005012876354157925, + 0.016939319670200348, + 0.005767914466559887, + 0.016501346603035927, + 0.013990833424031734, + 0.003762603737413883, + 0.007643210235983133, + 0.005211149342358112, + 0.004893242847174406, + 0.027724962681531906, + 0.006901896558701992, + 0.02717403694987297, + 0.0082851005718112, + 0.007131210528314114, + 0.010042672045528889, + 0.0066277640871703625, + 0.005695291794836521, + 0.00864629540592432, + 0.0068192072212696075, + 0.009178060106933117, + 0.005449825432151556, + 0.0064256396144628525, + 0.011169149540364742, + 0.005543997045606375, + 0.007252448238432407, + 0.006404190789908171, + 0.010058309882879257, + 0.006138098891824484, + 0.006772603373974562, + 0.006220824550837278, + 0.0063273124396800995, + 0.006880923639982939, + 0.005371883511543274, + 0.005118114408105612, + 0.004314636345952749, + 0.006834226660430431, + 0.007709641009569168, + 0.005251540802419186, + 0.004017441533505917, + 0.006371395196765661, + 0.006532914936542511 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.003201955696567893 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043067965656518936 + ], + "opposite_logprob_tokens": [ + -3.1680679321289062 + ], + "logprob_reward": -0.043067965656518936, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067965656518936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13808764517307281, + 0.05593455582857132, + 0.07533884793519974, + 0.08317378908395767, + 0.09225514531135559, + 0.07587606459856033, + 0.2893243432044983, + 0.03050549328327179, + 0.025657257065176964, + 0.13384683430194855 + ], + "credit_map": [ + 0.022297168150544167, + 0.019022062420845032, + 0.014346293173730373, + 0.016804177314043045, + 0.02135646715760231, + 0.022897176444530487, + 0.01009389292448759, + 0.011270414106547832, + 0.007837393321096897, + 0.010385546833276749, + 0.006036181002855301, + 0.006176023278385401, + 0.006970886141061783, + 0.008855605497956276, + 0.003971609752625227, + 0.0057013072073459625, + 0.010658392682671547, + 0.01039916556328535, + 0.011871684342622757, + 0.008575902320444584, + 0.009002191945910454, + 0.010161170735955238, + 0.007017869967967272, + 0.007652468513697386, + 0.018823903053998947, + 0.005822949111461639, + 0.011080210097134113, + 0.007395542226731777, + 0.006842643953859806, + 0.015735594555735588, + 0.010118614882230759, + 0.00735432468354702, + 0.013038013130426407, + 0.011197484098374844, + 0.016092674806714058, + 0.009641844779253006, + 0.00619532773271203, + 0.008343031629920006, + 0.019510574638843536, + 0.008236200548708439, + 0.01758342608809471, + 0.005472314078360796, + 0.01557848695665598, + 0.015249667689204216, + 0.00342170731164515, + 0.008016332983970642, + 0.004939873702824116, + 0.005614257417619228, + 0.09616108238697052, + 0.014617200940847397, + 0.09584520757198334, + 0.019081125035881996, + 0.008489315398037434, + 0.0276570376008749, + 0.01701304130256176, + 0.010460354387760162, + 0.003937418106943369, + 0.004455708432942629, + 0.003353254636749625, + 0.004454043228179216, + 0.003488888032734394, + 0.0043155052699148655, + 0.002791717881336808, + 0.003708956530317664, + 0.003064981894567609, + 0.005165779031813145, + 0.0029782981146126986, + 0.0032050798181444407, + 0.0034028049558401108, + 0.0026632873341441154, + 0.0025225847493857145, + 0.002654440002515912, + 0.02025759406387806, + 0.022183740511536598, + 0.01870591752231121, + 0.019351065158843994, + 0.018408285453915596, + 0.011229440569877625, + 0.01328987255692482, + 0.010420924052596092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.043067965656518936 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280853405594826 + ], + "opposite_logprob_tokens": [ + -3.7732808589935303 + ], + "logprob_reward": -0.023280853405594826, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280853405594826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1008496955037117, + 0.10018996149301529, + 0.15695840120315552, + 0.16575028002262115, + 0.09319432079792023, + 0.055801503360271454, + 0.15604543685913086, + 0.033338360488414764, + 0.051433052867650986, + 0.08643898367881775 + ], + "credit_map": [ + 0.015159682370722294, + 0.011232233606278896, + 0.009187002666294575, + 0.011744944378733635, + 0.016801752150058746, + 0.02246103808283806, + 0.006879925727844238, + 0.0073831225745379925, + 0.011147916316986084, + 0.02026456408202648, + 0.011980085633695126, + 0.011913289315998554, + 0.014780114404857159, + 0.015091290697455406, + 0.007685989141464233, + 0.007326715625822544, + 0.019886581227183342, + 0.018963316455483437, + 0.018571801483631134, + 0.027365541085600853, + 0.014457530342042446, + 0.017929090186953545, + 0.019641904160380363, + 0.02014264278113842, + 0.03632502257823944, + 0.007997901178896427, + 0.02620512805879116, + 0.0112994983792305, + 0.008276271633803844, + 0.05675642937421799, + 0.009913012385368347, + 0.00897702295333147, + 0.01486902590841055, + 0.013134628534317017, + 0.012689601629972458, + 0.009910138323903084, + 0.008957288227975368, + 0.010531114414334297, + 0.012539359740912914, + 0.010563170537352562, + 0.012287171557545662, + 0.004236564505845308, + 0.011356176808476448, + 0.01151665486395359, + 0.003019473282620311, + 0.005597068462520838, + 0.0038407943211495876, + 0.003947598859667778, + 0.051016226410865784, + 0.008085415698587894, + 0.05186837166547775, + 0.010021117515861988, + 0.005508511792868376, + 0.014601421542465687, + 0.009058593772351742, + 0.005885788705199957, + 0.003981413319706917, + 0.005546091590076685, + 0.0030352675821632147, + 0.002952940296381712, + 0.004625569097697735, + 0.0037823680322617292, + 0.004409336484968662, + 0.0050053722225129604, + 0.005879824049770832, + 0.007050043437629938, + 0.006568826735019684, + 0.007159112021327019, + 0.008409252390265465, + 0.005389510653913021, + 0.006557615473866463, + 0.00441886717453599, + 0.011807235889136791, + 0.012366869486868382, + 0.01041349582374096, + 0.011709043756127357, + 0.012237399816513062, + 0.008341733366250992, + 0.011652219109237194, + 0.00791099015623331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.023280853405594826 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle near the horizon, tilted steadily upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video captured in slow motion showing a massive humpback whale breaching the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the whale's skin deep black with visible white barnacles and scars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale twist sideways in mid-air with its pectoral fins fully extended?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep blue ocean with a sky tinted in warm orange and soft purple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623744197189808 + ], + "opposite_logprob_tokens": [ + -4.882623672485352 + ], + "logprob_reward": -0.007623744197189808, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623744197189808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23648719489574432, + 0.08660826832056046, + 0.08451032638549805, + 0.1197577714920044, + 0.052045803517103195, + 0.0699189305305481, + 0.1518653780221939, + 0.043790437281131744, + 0.04313674196600914, + 0.11187911778688431 + ], + "credit_map": [ + 0.039275988936424255, + 0.029251130297780037, + 0.02250278927385807, + 0.021260449662804604, + 0.025768205523490906, + 0.0661487802863121, + 0.01620381884276867, + 0.016076043248176575, + 0.01218612864613533, + 0.014299234375357628, + 0.01048329472541809, + 0.0101785808801651, + 0.00837140902876854, + 0.01397977489978075, + 0.009644088335335255, + 0.007465756963938475, + 0.017523521557450294, + 0.0101900240406394, + 0.012209954671561718, + 0.008549944497644901, + 0.006203732918947935, + 0.016560150310397148, + 0.0061914012767374516, + 0.0070815980434417725, + 0.0325925387442112, + 0.004936285316944122, + 0.015085574239492416, + 0.005853533744812012, + 0.004738053306937218, + 0.045428019016981125, + 0.004834465216845274, + 0.00628930889070034, + 0.010761690326035023, + 0.004105422645807266, + 0.006873166188597679, + 0.005335534922778606, + 0.0035264072939753532, + 0.006549042649567127, + 0.011384540237486362, + 0.0035099994856864214, + 0.015774350613355637, + 0.005243151448667049, + 0.012680841609835625, + 0.009942932054400444, + 0.0036945028696209192, + 0.009150011464953423, + 0.006553411018103361, + 0.006879727356135845, + 0.045167598873376846, + 0.008829725906252861, + 0.04303561523556709, + 0.010100745595991611, + 0.006775987800210714, + 0.01975144073367119, + 0.009450086392462254, + 0.008754183538258076, + 0.006488758139312267, + 0.004156453534960747, + 0.006628717295825481, + 0.004669752903282642, + 0.005256799980998039, + 0.007680126931518316, + 0.004025931004434824, + 0.004883898422122002, + 0.005746320355683565, + 0.0056228917092084885, + 0.01493875216692686, + 0.005377547349780798, + 0.002921624109148979, + 0.002834788290783763, + 0.0031707414891570807, + 0.0025240746326744556, + 0.016093578189611435, + 0.018693052232265472, + 0.026794014498591423, + 0.015171483159065247, + 0.009935542941093445, + 0.007220351602882147, + 0.0075973463244736195, + 0.010373753495514393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.007623744197189808 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_13/index_tensor([5986, 5986])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_13/index_tensor([5986, 5986])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..76fffb9fc16d88eb095feae016f3e50631eede07 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_13/index_tensor([5986, 5986])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Filmed with a stunning high-speed macro camera, the camera slowly and steadily moves forward, revealing a vast, undulating landscape composed of countless densely packed structures resembling healthy intestinal villi. These finger-like projections, rendered in soft, warm pink and coral hues, possess a velvety, almost textile-like surface texture and gently pulsate with perfect consistency, a slow, rhythmic wave of contraction and expansion flowing throughout the frame. Countless particles of luminous golden light, suspended in a perfectly clear, pristine liquid, drift weightlessly, carried by invisible microcurrents. As these luminous specks approach the villi, their surface seems to deliberately draw them in; the particles make contact and are rhythmically absorbed, their golden light gently diffusing into the pink tissue and generating a faint internal glow that temporarily illuminates the structures from within. The scene is illuminated solely by the particles themselves, casting a warm, ethereal light that creates delicate highlights and soft shadows within the complex folds of the landscape, establishing an enchanting and serene atmosphere and showcasing a beautiful yet abstract display of biological absorption.", + "seed": 15061, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_13/index_5986_seed_15061_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02908051637932658, + "vlm_logprob": -0.02908051637932658, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.02908051637932658, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04903838410973549 + ], + "opposite_logprob_tokens": [ + -3.1740384101867676 + ], + "logprob_reward": -0.04903838410973549, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04903838410973549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.078374944627285, + 0.05527837201952934, + 0.021365756168961525, + 0.06289331614971161, + 0.0785805806517601, + 0.13550618290901184, + 0.31654655933380127, + 0.15373943746089935, + 0.04952187091112137, + 0.04819298908114433 + ], + "credit_map": [ + 0.007761049084365368, + 0.009176512248814106, + 0.016006076708436012, + 0.010887203738093376, + 0.0066567761823534966, + 0.007871977984905243, + 0.00994728785008192, + 0.010068069212138653, + 0.004507281351834536, + 0.008059479296207428, + 0.008905841037631035, + 0.006599269807338715, + 0.0050272648222744465, + 0.010328581556677818, + 0.005225528031587601, + 0.0066251251846551895, + 0.002496404340490699, + 0.002962724072858691, + 0.003129988443106413, + 0.0029480946250259876, + 0.0021051298826932907, + 0.002128438325598836, + 0.0029861489310860634, + 0.002608827082440257, + 0.004542031791061163, + 0.0032401992939412594, + 0.02014438435435295, + 0.013096688315272331, + 0.0063638463616371155, + 0.0030598884914070368, + 0.004372215364128351, + 0.008074055425822735, + 0.006021840963512659, + 0.009931064210832119, + 0.023627251386642456, + 0.012118245474994183, + 0.0039224773645401, + 0.005997800268232822, + 0.008530153892934322, + 0.008431742899119854, + 0.010052540339529514, + 0.010242423973977566, + 0.033810362219810486, + 0.02659604698419571, + 0.01877126470208168, + 0.010001711547374725, + 0.01844082400202751, + 0.007591008674353361, + 0.011192097328603268, + 0.01054361928254366, + 0.11827441304922104, + 0.13422231376171112, + 0.0029125569853931665, + 0.003403114853426814, + 0.028862103819847107, + 0.007136315573006868, + 0.009497580118477345, + 0.00981025118380785, + 0.01895206607878208, + 0.04074632376432419, + 0.008584018796682358, + 0.009476521983742714, + 0.033152323216199875, + 0.02352035976946354, + 0.006903331261128187, + 0.0053090788424015045, + 0.005314372014254332, + 0.009617756120860577, + 0.005081180017441511, + 0.006018112879246473, + 0.0053215534426271915, + 0.0059564863331615925, + 0.005718534346669912, + 0.005327393766492605, + 0.006771765183657408, + 0.008294453844428062, + 0.0053381952457129955, + 0.004941812250763178, + 0.006115172524005175, + 0.005685663316398859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.04903838410973549 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005968722980469465 + ], + "opposite_logprob_tokens": [ + -5.1309685707092285 + ], + "logprob_reward": -0.005968722980469465, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005968722980469465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14966824650764465, + 0.10025187581777573, + 0.03281676024198532, + 0.06757133454084396, + 0.03368905559182167, + 0.21228718757629395, + 0.15525412559509277, + 0.1660388559103012, + 0.03451590985059738, + 0.04790658876299858 + ], + "credit_map": [ + 0.030773017555475235, + 0.013462007977068424, + 0.04205573350191116, + 0.02341735176742077, + 0.008249576203525066, + 0.011893168091773987, + 0.007842213846743107, + 0.01197519339621067, + 0.010240591131150723, + 0.006564868148416281, + 0.020366881042718887, + 0.010216339491307735, + 0.009635292924940586, + 0.02202577330172062, + 0.009168539196252823, + 0.012033609673380852, + 0.003518531331792474, + 0.003535495139658451, + 0.007273384369909763, + 0.005385186988860369, + 0.0025011938996613026, + 0.003309410996735096, + 0.0029964044224470854, + 0.004297156818211079, + 0.004413696005940437, + 0.008028284646570683, + 0.01713971421122551, + 0.010954085737466812, + 0.014947288669645786, + 0.004603785462677479, + 0.0038022834341973066, + 0.0036821970716118813, + 0.0025026837829500437, + 0.006930602248758078, + 0.008804935961961746, + 0.004645338747650385, + 0.0021122009493410587, + 0.0017623892053961754, + 0.0031500719487667084, + 0.0037808355409651995, + 0.01410399004817009, + 0.013731495477259159, + 0.06770197302103043, + 0.0346926748752594, + 0.04278184473514557, + 0.016050908714532852, + 0.010051017627120018, + 0.0131732989102602, + 0.006020630244165659, + 0.005259523168206215, + 0.05533519387245178, + 0.06295478343963623, + 0.0030953458044677973, + 0.0026890491135418415, + 0.013516339473426342, + 0.006383274681866169, + 0.00836842879652977, + 0.008436599746346474, + 0.019738594070076942, + 0.054296720772981644, + 0.00728609599173069, + 0.007527223788201809, + 0.038777727633714676, + 0.02160746604204178, + 0.004543117713183165, + 0.003654599655419588, + 0.003388898679986596, + 0.007390837650746107, + 0.0032485215924680233, + 0.004195288754999638, + 0.003583917859941721, + 0.004510731901973486, + 0.0033829703461378813, + 0.004417053889483213, + 0.012683975510299206, + 0.006581154651939869, + 0.004015813581645489, + 0.0038957111537456512, + 0.009118715301156044, + 0.0038112003821879625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.005968722980469465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003624657168984413 + ], + "opposite_logprob_tokens": [ + -5.628624439239502 + ], + "logprob_reward": -0.003624657168984413, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003624657168984413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08931432664394379, + 0.0811188593506813, + 0.013220720924437046, + 0.17090493440628052, + 0.047088995575904846, + 0.09607091546058655, + 0.2527710497379303, + 0.19043654203414917, + 0.030003178864717484, + 0.029070425778627396 + ], + "credit_map": [ + 0.009913827292621136, + 0.029177023097872734, + 0.009562866762280464, + 0.0098414346575737, + 0.0063937813974916935, + 0.007322121411561966, + 0.007426295429468155, + 0.009676977060735226, + 0.009917786344885826, + 0.00852589588612318, + 0.022728247568011284, + 0.0064252689480781555, + 0.008555260486900806, + 0.016123181208968163, + 0.004798169247806072, + 0.004045057110488415, + 0.0013467803364619613, + 0.0012386825401335955, + 0.0024328737054020166, + 0.002392409835010767, + 0.0011488328455016017, + 0.0013885473599657416, + 0.0012413120130077004, + 0.0020312839187681675, + 0.010524751618504524, + 0.008003306575119495, + 0.06165425106883049, + 0.04003158211708069, + 0.02961592935025692, + 0.00737000023946166, + 0.005761212669312954, + 0.00794390682131052, + 0.0030959441792219877, + 0.006910099182277918, + 0.018264058977365494, + 0.007479794789105654, + 0.0015277828788384795, + 0.0016483142971992493, + 0.003726488910615444, + 0.004436518996953964, + 0.005853856448084116, + 0.005902113858610392, + 0.02787795662879944, + 0.018070951104164124, + 0.01757054217159748, + 0.007505621295422316, + 0.005567505955696106, + 0.007722379174083471, + 0.008840028196573257, + 0.007976794615387917, + 0.09579892456531525, + 0.1068931370973587, + 0.002310526091605425, + 0.003309873165562749, + 0.022033575922250748, + 0.005608196835964918, + 0.010497520677745342, + 0.009996715933084488, + 0.02410484291613102, + 0.05963628366589546, + 0.008321594446897507, + 0.009740925393998623, + 0.040722668170928955, + 0.02741602249443531, + 0.0043386672623455524, + 0.0034234970808029175, + 0.0035066180862486362, + 0.00435287319123745, + 0.003159844083711505, + 0.002712988993152976, + 0.004503886681050062, + 0.0040048048831522465, + 0.0026830527931451797, + 0.002541859168559313, + 0.006113803945481777, + 0.004282509442418814, + 0.0027851825580000877, + 0.0029439637437462807, + 0.004779825918376446, + 0.0029402319341897964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.003624657168984413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038108546286821365 + ], + "opposite_logprob_tokens": [ + -3.2881085872650146 + ], + "logprob_reward": -0.038108546286821365, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038108546286821365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06178274750709534, + 0.04525531455874443, + 0.008492201566696167, + 0.058866556733846664, + 0.05223146453499794, + 0.2367933690547943, + 0.22222711145877838, + 0.2444315403699875, + 0.04137406870722771, + 0.028545573353767395 + ], + "credit_map": [ + 0.008115822449326515, + 0.01947914808988571, + 0.007445774506777525, + 0.004562494345009327, + 0.00480511924251914, + 0.0044986288994550705, + 0.006827834527939558, + 0.006047928240150213, + 0.006757889874279499, + 0.007575165946036577, + 0.011072125285863876, + 0.004082638304680586, + 0.004046702291816473, + 0.006882253102958202, + 0.0022914865985512733, + 0.0025470538530498743, + 0.0009508795337751508, + 0.0009974332060664892, + 0.0011382992379367352, + 0.001550205866806209, + 0.0007991770398803055, + 0.0010582096874713898, + 0.0009329535532742739, + 0.0010650439653545618, + 0.003997422754764557, + 0.0033829407766461372, + 0.019323676824569702, + 0.012636042200028896, + 0.010352392680943012, + 0.0028132456354796886, + 0.002147429157048464, + 0.004213407170027494, + 0.0032825611997395754, + 0.009417103603482246, + 0.020443692803382874, + 0.007865394465625286, + 0.0017857131315395236, + 0.0019287775503471494, + 0.0037013846449553967, + 0.0038068389985710382, + 0.014122326858341694, + 0.013311938382685184, + 0.0699291080236435, + 0.04600208252668381, + 0.04672393202781677, + 0.01655396819114685, + 0.014725807122886181, + 0.015424208715558052, + 0.00794885866343975, + 0.0069973887875676155, + 0.08517426252365112, + 0.09277140349149704, + 0.002367940964177251, + 0.002493322594091296, + 0.018390966579318047, + 0.006082967855036259, + 0.01398210134357214, + 0.01287045981734991, + 0.029342278838157654, + 0.07329896092414856, + 0.011849984526634216, + 0.01152038387954235, + 0.05477308854460716, + 0.0367942713201046, + 0.006027888972312212, + 0.004225431475788355, + 0.004701205994933844, + 0.007621810305863619, + 0.003349537728354335, + 0.0037538649048656225, + 0.0064063966274261475, + 0.005287932697683573, + 0.0024191983975470066, + 0.0025589591823518276, + 0.006186976097524166, + 0.0046498775482177734, + 0.002514037536457181, + 0.0022994910832494497, + 0.005403828341513872, + 0.0025132051669061184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 6 + ], + "selected_reward": -0.038108546286821365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866227135062218 + ], + "opposite_logprob_tokens": [ + -3.0486621856689453 + ], + "logprob_reward": -0.04866227135062218, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866227135062218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13280139863491058, + 0.08275088667869568, + 0.013483691960573196, + 0.10253264755010605, + 0.02090136520564556, + 0.1547645926475525, + 0.19478893280029297, + 0.2041575014591217, + 0.041604407131671906, + 0.05221462994813919 + ], + "credit_map": [ + 0.017441734671592712, + 0.03361076861619949, + 0.038016900420188904, + 0.016125069931149483, + 0.007566666696220636, + 0.008441275916993618, + 0.004923293367028236, + 0.0066756876185536385, + 0.009116102941334248, + 0.005788395181298256, + 0.02456321381032467, + 0.006818403024226427, + 0.007573939394205809, + 0.020739009603857994, + 0.004066632594913244, + 0.004085191059857607, + 0.0013124513207003474, + 0.001541623380035162, + 0.0014585580211132765, + 0.002835995750501752, + 0.0012058126740157604, + 0.0014716081786900759, + 0.0014961533015593886, + 0.0021614893339574337, + 0.006792356725782156, + 0.0044714538380503654, + 0.03607064113020897, + 0.023393208160996437, + 0.01761532574892044, + 0.004897754639387131, + 0.003351910039782524, + 0.005940004251897335, + 0.0013100883224979043, + 0.0030213617719709873, + 0.004689169581979513, + 0.003090037265792489, + 0.001433566096238792, + 0.0020416888874024153, + 0.0029670302756130695, + 0.002348423469811678, + 0.00800749845802784, + 0.006998548749834299, + 0.05404981970787048, + 0.035569481551647186, + 0.025234650820493698, + 0.009018256328999996, + 0.008314717561006546, + 0.007571615744382143, + 0.006842559669166803, + 0.0064163547940552235, + 0.07283898442983627, + 0.07932987809181213, + 0.002936422126367688, + 0.0030871694907546043, + 0.015379938296973705, + 0.00795763824135065, + 0.013905740343034267, + 0.012308242730796337, + 0.025939829647541046, + 0.056893911212682724, + 0.010440317913889885, + 0.011527159251272678, + 0.039193227887153625, + 0.03394906595349312, + 0.005979681853204966, + 0.004407248925417662, + 0.004685040097683668, + 0.007401708513498306, + 0.0037901613395661116, + 0.003762553445994854, + 0.00609184755012393, + 0.005486161448061466, + 0.0058525060303509235, + 0.005784338805824518, + 0.006979493424296379, + 0.00861420203000307, + 0.005910789128392935, + 0.005478204227983952, + 0.0071465023793280125, + 0.0064485943876206875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.04866227135062218 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Filmed with a stunning high-speed macro camera, the camera slowly and steadily moves forward, revealing a vast, undulating landscape composed of countless densely packed structures resembling healthy intestinal villi. These finger-like projections, rendered in soft, warm pink and coral hues, possess a velvety, almost textile-like surface texture and gently pulsate with perfect consistency, a slow, rhythmic wave of contraction and expansion flowing throughout the frame. Countless particles of luminous golden light, suspended in a perfectly clear, pristine liquid, drift weightlessly, carried by invisible microcurrents. As these luminous specks approach the villi, their surface seems to deliberately draw them in; the particles make contact and are rhythmically absorbed, their golden light gently diffusing into the pink tissue and generating a faint internal glow that temporarily illuminates the structures from within. The scene is illuminated solely by the particles themselves, casting a warm, ethereal light that creates delicate highlights and soft shadows within the complex folds of the landscape, establishing an enchanting and serene atmosphere and showcasing a beautiful yet abstract display of biological absorption.", + "seed": 15062, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_13/index_5986_seed_15062_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10542981568723916, + "vlm_logprob": -0.10542981568723916, + "vlm_yes_no_margin": 2.9499999642372132, + "vlm_token_credit": -0.10542981568723916, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05833565443754196 + ], + "opposite_logprob_tokens": [ + -3.058335542678833 + ], + "logprob_reward": -0.05833565443754196, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05833565443754196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07232604920864105, + 0.04484482854604721, + 0.023781267926096916, + 0.06670445948839188, + 0.0438559390604496, + 0.04838244989514351, + 0.18022866547107697, + 0.14892582595348358, + 0.283554345369339, + 0.08739612996578217 + ], + "credit_map": [ + 0.007028242107480764, + 0.004612072836607695, + 0.02177887037396431, + 0.008322560228407383, + 0.00410732626914978, + 0.006861600559204817, + 0.011019148863852024, + 0.00859623122960329, + 0.003016032977029681, + 0.0023442881647497416, + 0.005183396395295858, + 0.004203669261187315, + 0.002710366155952215, + 0.0037459777668118477, + 0.012204869650304317, + 0.01143623050302267, + 0.0021387324668467045, + 0.002659819321706891, + 0.00258926535025239, + 0.0021960039157420397, + 0.003951911814510822, + 0.0021934318356215954, + 0.003972514532506466, + 0.004079587757587433, + 0.008233432658016682, + 0.020894890651106834, + 0.0037198655772954226, + 0.012741993181407452, + 0.006232210900634527, + 0.00418442115187645, + 0.005330810789018869, + 0.005366834811866283, + 0.005103674717247486, + 0.004422079771757126, + 0.007889128290116787, + 0.005051868502050638, + 0.003305408637970686, + 0.0041225021705031395, + 0.006440031807869673, + 0.007521242368966341, + 0.0038744972553104162, + 0.0036652658600360155, + 0.007315122988075018, + 0.006298599764704704, + 0.004292589612305164, + 0.004075509496033192, + 0.01155176106840372, + 0.0073091029189527035, + 0.014045518822968006, + 0.012939492240548134, + 0.028762642294168472, + 0.05463993921875954, + 0.015079790726304054, + 0.01303769089281559, + 0.026186328381299973, + 0.015537251718342304, + 0.007740981411188841, + 0.008820505812764168, + 0.04964615777134895, + 0.017592601478099823, + 0.012694158591330051, + 0.011485233902931213, + 0.01608811691403389, + 0.024858081713318825, + 0.05412105470895767, + 0.020886681973934174, + 0.03353342413902283, + 0.01865951530635357, + 0.09553305804729462, + 0.02451719343662262, + 0.01701718010008335, + 0.019286254420876503, + 0.010538658127188683, + 0.004539186600595713, + 0.0032935263589024544, + 0.002270742319524288, + 0.0034071982372552156, + 0.003382896538823843, + 0.028040282428264618, + 0.03192363679409027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.05833565443754196 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02061360329389572 + ], + "opposite_logprob_tokens": [ + -3.895613670349121 + ], + "logprob_reward": -0.02061360329389572, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02061360329389572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09725721180438995, + 0.06671687960624695, + 0.05566652491688728, + 0.22687803208827972, + 0.07574450969696045, + 0.03869476169347763, + 0.1089322492480278, + 0.0776195377111435, + 0.22243057191371918, + 0.03005967289209366 + ], + "credit_map": [ + 0.010862021706998348, + 0.007323997560888529, + 0.010058759711682796, + 0.011110242456197739, + 0.006293742451816797, + 0.010911766439676285, + 0.020704153925180435, + 0.019992535933852196, + 0.005734235979616642, + 0.006057292222976685, + 0.011898905038833618, + 0.010321000590920448, + 0.005308032967150211, + 0.006495540961623192, + 0.010291443206369877, + 0.010610434226691723, + 0.004654949530959129, + 0.004567606374621391, + 0.006407491862773895, + 0.006412912160158157, + 0.005006890278309584, + 0.006922455970197916, + 0.011530249379575253, + 0.010163971222937107, + 0.010488254949450493, + 0.03728233277797699, + 0.011718076653778553, + 0.0725182443857193, + 0.014974049292504787, + 0.010738172568380833, + 0.05569327250123024, + 0.01346563920378685, + 0.006122648250311613, + 0.005783020053058863, + 0.009972021915018559, + 0.009563063271343708, + 0.0055894977413117886, + 0.0072332569397985935, + 0.013756314292550087, + 0.017724698409438133, + 0.0030638144817203283, + 0.003012024564668536, + 0.006611617747694254, + 0.008071746677160263, + 0.0031831541564315557, + 0.0034119251649826765, + 0.0047607216984033585, + 0.006579761393368244, + 0.0077506257221102715, + 0.010451573878526688, + 0.016733622178435326, + 0.03396103158593178, + 0.007322372402995825, + 0.00859127938747406, + 0.014626219868659973, + 0.009495529346168041, + 0.005613977089524269, + 0.006257388740777969, + 0.01105726882815361, + 0.008487820625305176, + 0.012073715217411518, + 0.010035799816250801, + 0.01262052170932293, + 0.011473058722913265, + 0.05207286775112152, + 0.013765164650976658, + 0.01669975183904171, + 0.007974239066243172, + 0.08210090547800064, + 0.01854889839887619, + 0.017293186858296394, + 0.013975566253066063, + 0.005032040644437075, + 0.006189414765685797, + 0.003403069917112589, + 0.002558639971539378, + 0.0027457086835056543, + 0.0035264156758785248, + 0.0038058452773839235, + 0.002798538887873292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.02061360329389572 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076431455090641975 + ], + "opposite_logprob_tokens": [ + -4.882643222808838 + ], + "logprob_reward": -0.0076431455090641975, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076431455090641975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.124428890645504, + 0.07710196077823639, + 0.0533953420817852, + 0.2632097601890564, + 0.046523645520210266, + 0.08042299747467041, + 0.10949505120515823, + 0.07016662508249283, + 0.1421026587486267, + 0.03315296024084091 + ], + "credit_map": [ + 0.012909073382616043, + 0.011367376893758774, + 0.021086739376187325, + 0.011910251341760159, + 0.009846361353993416, + 0.015659185126423836, + 0.01761258952319622, + 0.024037327617406845, + 0.003307979553937912, + 0.002959972247481346, + 0.006473359186202288, + 0.005923392716795206, + 0.0030809903983026743, + 0.003921240568161011, + 0.02383900061249733, + 0.02759603038430214, + 0.004668009001761675, + 0.004697307478636503, + 0.0051833754405379295, + 0.00544143607839942, + 0.007340975571423769, + 0.005357466172426939, + 0.012999256141483784, + 0.007707524113357067, + 0.0356636680662632, + 0.08469187468290329, + 0.0123506560921669, + 0.04804070293903351, + 0.023890115320682526, + 0.017144987359642982, + 0.02579587884247303, + 0.015631914138793945, + 0.005199200939387083, + 0.006809928920120001, + 0.0061116269789636135, + 0.006178434006869793, + 0.004956331569701433, + 0.004712772089987993, + 0.006405688356608152, + 0.006149669643491507, + 0.005689227022230625, + 0.0050110043957829475, + 0.014972430653870106, + 0.017094766721129417, + 0.0061438255943357944, + 0.006430775858461857, + 0.012421797029674053, + 0.012659181840717793, + 0.007826585322618484, + 0.010950490832328796, + 0.014680526219308376, + 0.016891075298190117, + 0.012843957170844078, + 0.015992214903235435, + 0.016691531985998154, + 0.013618674129247665, + 0.005050755571573973, + 0.004576748237013817, + 0.017950408160686493, + 0.006308242212980986, + 0.007075706031173468, + 0.006720344070345163, + 0.009716915898025036, + 0.012767513282597065, + 0.03450267016887665, + 0.009273332543671131, + 0.01093232911080122, + 0.005168834701180458, + 0.053599484264850616, + 0.012523643672466278, + 0.008676751516759396, + 0.0074256365187466145, + 0.00459011597558856, + 0.0028492403216660023, + 0.002542841015383601, + 0.0019523033406585455, + 0.0023769917897880077, + 0.002213903469964862, + 0.007272019982337952, + 0.009355548769235611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.0076431455090641975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31347113847732544 + ], + "opposite_logprob_tokens": [ + -1.3134710788726807 + ], + "logprob_reward": -0.31347113847732544, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31347113847732544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21619489789009094, + 0.05200113728642464, + 0.030314365401864052, + 0.13420039415359497, + 0.05534151569008827, + 0.11582475155591965, + 0.08381404727697372, + 0.029279746115207672, + 0.1533626914024353, + 0.12966638803482056 + ], + "credit_map": [ + 0.023797068744897842, + 0.014115957543253899, + 0.05611232668161392, + 0.022380638867616653, + 0.013535507954657078, + 0.023053333163261414, + 0.03545941412448883, + 0.02774064801633358, + 0.00501022394746542, + 0.005978152621537447, + 0.007800048217177391, + 0.00769837386906147, + 0.005041185766458511, + 0.0054713026620447636, + 0.007976172491908073, + 0.007025677245110273, + 0.002988215768709779, + 0.0026303566992282867, + 0.003377144690603018, + 0.003803784493356943, + 0.0025467071682214737, + 0.003309173509478569, + 0.005916661117225885, + 0.005742322653532028, + 0.014227013103663921, + 0.03432782366871834, + 0.005671209655702114, + 0.030201097950339317, + 0.01125242654234171, + 0.0070091611705720425, + 0.02211717702448368, + 0.009394478984177113, + 0.005077202804386616, + 0.008778451010584831, + 0.006175329443067312, + 0.006019144784659147, + 0.005299087148159742, + 0.00480790575966239, + 0.00852204579859972, + 0.010662354528903961, + 0.008216935209929943, + 0.007890944369137287, + 0.019468126818537712, + 0.023856017738580704, + 0.008534662425518036, + 0.009109282866120338, + 0.020651819184422493, + 0.018096966668963432, + 0.0077278451062738895, + 0.007374566979706287, + 0.010749478824436665, + 0.014817888848483562, + 0.012275826185941696, + 0.00995729025453329, + 0.012491632252931595, + 0.008419519290328026, + 0.0031889809761196375, + 0.0026167957112193108, + 0.00668073957785964, + 0.0035269279032945633, + 0.003247263142839074, + 0.003471429692581296, + 0.00317571684718132, + 0.003371893661096692, + 0.033449068665504456, + 0.009248376823961735, + 0.013637162744998932, + 0.006873561069369316, + 0.05222926661372185, + 0.012315950356423855, + 0.013986327685415745, + 0.01162299420684576, + 0.016024475917220116, + 0.007124916650354862, + 0.004908647853881121, + 0.003853290807455778, + 0.005003304220736027, + 0.005034522153437138, + 0.04132881760597229, + 0.0463884100317955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.31347113847732544 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708553671836853 + ], + "opposite_logprob_tokens": [ + -2.1270854473114014 + ], + "logprob_reward": -0.12708553671836853, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12708553671836853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18971066176891327, + 0.056631702929735184, + 0.03439617529511452, + 0.06534933298826218, + 0.04269087314605713, + 0.06749522686004639, + 0.20078830420970917, + 0.062399376183748245, + 0.16595123708248138, + 0.11458716541528702 + ], + "credit_map": [ + 0.01912650465965271, + 0.012654629535973072, + 0.05359819531440735, + 0.0191380362957716, + 0.01130333635956049, + 0.020884351804852486, + 0.029266664758324623, + 0.023738931864500046, + 0.002259236527606845, + 0.001884804805740714, + 0.0034326089080423117, + 0.003882212098687887, + 0.0031384306494146585, + 0.004164548125118017, + 0.018373146653175354, + 0.01949671469628811, + 0.0030248998664319515, + 0.003734357189387083, + 0.00426176842302084, + 0.004177812021225691, + 0.005333192180842161, + 0.00331346457824111, + 0.005510742776095867, + 0.005039937794208527, + 0.008837981149554253, + 0.01779673434793949, + 0.004849453456699848, + 0.008947602473199368, + 0.006871932186186314, + 0.005066976882517338, + 0.00619879737496376, + 0.006779858376830816, + 0.0048772976733744144, + 0.007100380957126617, + 0.006035903003066778, + 0.0038656299002468586, + 0.004268365912139416, + 0.004223913419991732, + 0.005498670507222414, + 0.006820713635534048, + 0.004761105868965387, + 0.004014272708445787, + 0.0130156846717, + 0.014691876247525215, + 0.004514652770012617, + 0.005574124399572611, + 0.009537121281027794, + 0.011386385187506676, + 0.01828816346824169, + 0.013970663771033287, + 0.027582360431551933, + 0.0429191067814827, + 0.025879744440317154, + 0.019480561837553978, + 0.03212185949087143, + 0.020545847713947296, + 0.00505270529538393, + 0.005428683944046497, + 0.016122516244649887, + 0.007270114030689001, + 0.0067217242904007435, + 0.007928081788122654, + 0.0051604402251541615, + 0.008715108968317509, + 0.03457446023821831, + 0.009650015272200108, + 0.01316281221807003, + 0.007222005631774664, + 0.05506748706102371, + 0.013217858038842678, + 0.017817506566643715, + 0.015239091590046883, + 0.014907782897353172, + 0.007190912961959839, + 0.005642293021082878, + 0.00348867941647768, + 0.0049193487502634525, + 0.005090796854346991, + 0.03361599147319794, + 0.03973136842250824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.12708553671836853 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Filmed with a stunning high-speed macro camera, the camera slowly and steadily moves forward, revealing a vast, undulating landscape composed of countless densely packed structures resembling healthy intestinal villi. These finger-like projections, rendered in soft, warm pink and coral hues, possess a velvety, almost textile-like surface texture and gently pulsate with perfect consistency, a slow, rhythmic wave of contraction and expansion flowing throughout the frame. Countless particles of luminous golden light, suspended in a perfectly clear, pristine liquid, drift weightlessly, carried by invisible microcurrents. As these luminous specks approach the villi, their surface seems to deliberately draw them in; the particles make contact and are rhythmically absorbed, their golden light gently diffusing into the pink tissue and generating a faint internal glow that temporarily illuminates the structures from within. The scene is illuminated solely by the particles themselves, casting a warm, ethereal light that creates delicate highlights and soft shadows within the complex folds of the landscape, establishing an enchanting and serene atmosphere and showcasing a beautiful yet abstract display of biological absorption.", + "seed": 15063, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_13/index_5986_seed_15063_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21662406846880913, + "vlm_logprob": -0.21662406846880913, + "vlm_yes_no_margin": 2.249999976158142, + "vlm_token_credit": -0.21662406846880913, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1493014246225357 + ], + "opposite_logprob_tokens": [ + -2.024301528930664 + ], + "logprob_reward": -0.1493014246225357, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1493014246225357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4296919107437134, + 0.2533564269542694, + 0.037704139947891235, + 0.012650697492063046, + 0.027227304875850677, + 0.033484119921922684, + 0.11426644772291183, + 0.03314334899187088, + 0.021229978650808334, + 0.03724561631679535 + ], + "credit_map": [ + 0.02316051907837391, + 0.1556280553340912, + 0.011318053118884563, + 0.013356136158108711, + 0.14750248193740845, + 0.030362999066710472, + 0.01728826016187668, + 0.031075384467840195, + 0.011185829527676105, + 0.021780721843242645, + 0.006753033492714167, + 0.004364321008324623, + 0.10694404691457748, + 0.08878035843372345, + 0.006501962896436453, + 0.007046164479106665, + 0.004431885201483965, + 0.006717419717460871, + 0.0035996465012431145, + 0.0025830978993326426, + 0.007331048604100943, + 0.004956594668328762, + 0.003533761715516448, + 0.004550686571747065, + 0.0011628272477537394, + 0.0013405968202278018, + 0.0012919272994622588, + 0.0013022146886214614, + 0.0022206453140825033, + 0.0018028346821665764, + 0.001395478961057961, + 0.0021341729443520308, + 0.0026772578712552786, + 0.0030156010761857033, + 0.0019357928540557623, + 0.003662356175482273, + 0.003107578493654728, + 0.00345571874640882, + 0.003186878515407443, + 0.006186119746416807, + 0.005063854623585939, + 0.0036202531773597, + 0.0043097385205328465, + 0.002396156545728445, + 0.00519805820658803, + 0.004174315836280584, + 0.005076483357697725, + 0.003645259188488126, + 0.01703137345612049, + 0.006831615697592497, + 0.007419826928526163, + 0.006761045660823584, + 0.012300531379878521, + 0.018478447571396828, + 0.03649375960230827, + 0.008949854411184788, + 0.004428059794008732, + 0.0029376454185694456, + 0.003541225101798773, + 0.003840244375169277, + 0.007114713080227375, + 0.005026745609939098, + 0.0035117450170218945, + 0.0027429722249507904, + 0.0025240678805857897, + 0.0029106931760907173, + 0.002105000428855419, + 0.0034301297273486853, + 0.002466582925990224, + 0.0030377418734133244, + 0.0024436546955257654, + 0.0023121077101677656, + 0.0066227782517671585, + 0.0037773435469716787, + 0.00456705829128623, + 0.004371189512312412, + 0.005401722155511379, + 0.0038875818718224764, + 0.00402933731675148, + 0.004588605836033821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.1493014246225357 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018217740580439568 + ], + "opposite_logprob_tokens": [ + -4.01821756362915 + ], + "logprob_reward": -0.018217740580439568, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018217740580439568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4573679566383362, + 0.21758614480495453, + 0.04654364287853241, + 0.010413173586130142, + 0.020884491503238678, + 0.047889579087495804, + 0.1612541675567627, + 0.025407185778021812, + 0.007107815705239773, + 0.0055458154529333115 + ], + "credit_map": [ + 0.026736898347735405, + 0.17164066433906555, + 0.013562008738517761, + 0.013779476284980774, + 0.15231777727603912, + 0.02890050783753395, + 0.01826871931552887, + 0.03216187283396721, + 0.009082886390388012, + 0.013888776302337646, + 0.005680288188159466, + 0.004721668548882008, + 0.09140463918447495, + 0.08077386766672134, + 0.0059621911495924, + 0.006071812938898802, + 0.004329552408307791, + 0.008756882511079311, + 0.006080930586904287, + 0.005476091988384724, + 0.0066665359772741795, + 0.00404631020501256, + 0.007553632836788893, + 0.003633701940998435, + 0.001248929649591446, + 0.0011158344568684697, + 0.0007546512060798705, + 0.0009555748547427356, + 0.001970960758626461, + 0.0012983970809727907, + 0.0011243120534345508, + 0.0019445125944912434, + 0.0022590765729546547, + 0.0026683269534260035, + 0.002079859608784318, + 0.0016684752190485597, + 0.0024919973220676184, + 0.0035584389697760344, + 0.001550282584503293, + 0.004608035087585449, + 0.005333486944437027, + 0.004246634431183338, + 0.007755084428936243, + 0.0046910978853702545, + 0.00467562535777688, + 0.005974281579256058, + 0.005220961291342974, + 0.009992405772209167, + 0.008328521624207497, + 0.005911565385758877, + 0.006354143843054771, + 0.007654780987650156, + 0.013703792355954647, + 0.050046298652887344, + 0.056453682482242584, + 0.01280138548463583, + 0.0033335613552480936, + 0.0029086433351039886, + 0.0024423967115581036, + 0.0027162153273820877, + 0.005009439308196306, + 0.0037048314698040485, + 0.0024657328613102436, + 0.0028263640124350786, + 0.0008208282524719834, + 0.0009869495406746864, + 0.0007876978488638997, + 0.0013031732523813844, + 0.0008145423489622772, + 0.0010282627772539854, + 0.000631286296993494, + 0.0007350746891461313, + 0.0005900923279114068, + 0.000680167693644762, + 0.0006836405373178422, + 0.0006790522020310163, + 0.000944998930208385, + 0.0005975448875688016, + 0.000655441777780652, + 0.0007148768054321408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.018217740580439568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020597605034708977 + ], + "opposite_logprob_tokens": [ + -3.8955976963043213 + ], + "logprob_reward": -0.020597605034708977, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020597605034708977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3127082884311676, + 0.12293963879346848, + 0.028840072453022003, + 0.020219121128320694, + 0.029181070625782013, + 0.047111187130212784, + 0.37244078516960144, + 0.04308965429663658, + 0.010263779200613499, + 0.01320637483149767 + ], + "credit_map": [ + 0.014017799869179726, + 0.10540135949850082, + 0.012746328487992287, + 0.010395261459052563, + 0.11264965683221817, + 0.02443484589457512, + 0.014436895027756691, + 0.018626166507601738, + 0.00787961296737194, + 0.021989278495311737, + 0.004550294950604439, + 0.004432611633092165, + 0.04681101068854332, + 0.027475429698824883, + 0.004535780753940344, + 0.005265621468424797, + 0.0026360771153122187, + 0.0033605939242988825, + 0.0034323795698583126, + 0.0019810358062386513, + 0.006492785178124905, + 0.00491408072412014, + 0.003166084410622716, + 0.002857035491615534, + 0.0020598778501152992, + 0.0016757402336224914, + 0.0018026159377768636, + 0.001703658257611096, + 0.004653925076127052, + 0.0027216076850891113, + 0.0032061459496617317, + 0.002395549789071083, + 0.003034149529412389, + 0.003019447438418865, + 0.002085083397105336, + 0.0017355926102027297, + 0.004192078951746225, + 0.007137695327401161, + 0.0025718307588249445, + 0.005405191797763109, + 0.00420112581923604, + 0.0042154425755143166, + 0.005485810339450836, + 0.004363711457699537, + 0.005215320270508528, + 0.00825752504169941, + 0.007570821326225996, + 0.007801429368555546, + 0.01899818889796734, + 0.016026990488171577, + 0.0151381129398942, + 0.018672525882720947, + 0.029201284050941467, + 0.10506182163953781, + 0.13622690737247467, + 0.033114925026893616, + 0.007294963579624891, + 0.003482076106593013, + 0.005148005206137896, + 0.005630101077258587, + 0.006123574450612068, + 0.005484987981617451, + 0.0054570394568145275, + 0.004468907136470079, + 0.0014598140260204673, + 0.0015072383685037494, + 0.0009664632380008698, + 0.0012247953563928604, + 0.0010850521503016353, + 0.0014670657692477107, + 0.0012670682044699788, + 0.001286282204091549, + 0.001626749406568706, + 0.0013289814814925194, + 0.0019313025986775756, + 0.0017234045080840588, + 0.001827420317567885, + 0.0016701272688806057, + 0.001635467866435647, + 0.00146292126737535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.020597605034708977 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933908462524414 + ], + "opposite_logprob_tokens": [ + -0.6933908462524414 + ], + "logprob_reward": -0.6933908462524414, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933908462524414, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17317643761634827, + 0.20439213514328003, + 0.03803881257772446, + 0.014494325965642929, + 0.06844469904899597, + 0.04317193105816841, + 0.33673372864723206, + 0.03650020435452461, + 0.05813325196504593, + 0.02691449224948883 + ], + "credit_map": [ + 0.010861669667065144, + 0.04540387541055679, + 0.007853996008634567, + 0.006278793327510357, + 0.06141332909464836, + 0.01942416839301586, + 0.008870303630828857, + 0.013070307672023773, + 0.01391645334661007, + 0.0465698204934597, + 0.013393616303801537, + 0.00785271730273962, + 0.07609009742736816, + 0.030941244214773178, + 0.00828783493489027, + 0.007340362295508385, + 0.002888553310185671, + 0.006259896792471409, + 0.0021808138117194176, + 0.0031434795819222927, + 0.005009436048567295, + 0.013113725930452347, + 0.0027134320698678493, + 0.0027294729370623827, + 0.0011738952016457915, + 0.0012473500100895762, + 0.0017902514664456248, + 0.001604289747774601, + 0.001718783052638173, + 0.0020197106059640646, + 0.0034911592956632376, + 0.001448886119760573, + 0.0065181259997189045, + 0.005116351414471865, + 0.005370690952986479, + 0.0033601303584873676, + 0.010225193575024605, + 0.017322203144431114, + 0.006918085739016533, + 0.013613914139568806, + 0.005888065788894892, + 0.00481572188436985, + 0.006561313755810261, + 0.0034321441780775785, + 0.006315943785011768, + 0.004732091911137104, + 0.004289464559406042, + 0.00713719055056572, + 0.020312542095780373, + 0.015033423900604248, + 0.017464270815253258, + 0.015929048880934715, + 0.0301122535020113, + 0.09871897846460342, + 0.11661634594202042, + 0.022546878084540367, + 0.005089974030852318, + 0.002464591059833765, + 0.0036126398481428623, + 0.0026454131584614515, + 0.009009955450892448, + 0.006900371052324772, + 0.0036203584168106318, + 0.003156898310407996, + 0.007944883778691292, + 0.006034299731254578, + 0.006288127973675728, + 0.010963700711727142, + 0.006331795360893011, + 0.007266730535775423, + 0.005849059205502272, + 0.0074546546675264835, + 0.003521754639223218, + 0.0035597614478319883, + 0.003956084605306387, + 0.0028796892147511244, + 0.003598026465624571, + 0.002453291555866599, + 0.0028793979436159134, + 0.004066484048962593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.6933908462524414 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161272585391998 + ], + "opposite_logprob_tokens": [ + -1.7016127109527588 + ], + "logprob_reward": -0.20161272585391998, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161272585391998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3887511193752289, + 0.06003084406256676, + 0.019285526126623154, + 0.026894381269812584, + 0.07256193459033966, + 0.034201838076114655, + 0.30749303102493286, + 0.03711114451289177, + 0.031127721071243286, + 0.02254243567585945 + ], + "credit_map": [ + 0.017674528062343597, + 0.14092135429382324, + 0.009818436577916145, + 0.011678146198391914, + 0.14612558484077454, + 0.030018705874681473, + 0.015041660517454147, + 0.017472751438617706, + 0.007886335253715515, + 0.010122260078787804, + 0.006227642297744751, + 0.0049229278229177, + 0.014878574758768082, + 0.007563640363514423, + 0.004356030840426683, + 0.004073430318385363, + 0.001989304320886731, + 0.0030022854916751385, + 0.0017751854611560702, + 0.00159234378952533, + 0.0024418598040938377, + 0.00410475255921483, + 0.0022507328540086746, + 0.0021290627773851156, + 0.0022795707918703556, + 0.0019186231074854732, + 0.0031799592543393373, + 0.003261075122281909, + 0.003323022276163101, + 0.0036114819813519716, + 0.005977585446089506, + 0.0033430622424930334, + 0.006472882814705372, + 0.0047727362252771854, + 0.006240834016352892, + 0.003663594601675868, + 0.010077369399368763, + 0.015408284962177277, + 0.008677780628204346, + 0.017248451709747314, + 0.005515174008905888, + 0.003445884445682168, + 0.0057831634767353535, + 0.0032930050510913134, + 0.004440024960786104, + 0.004128357861191034, + 0.0035314217675477266, + 0.004064807202666998, + 0.023296846076846123, + 0.01334378682076931, + 0.014874744229018688, + 0.014860070310533047, + 0.03035767562687397, + 0.07853148132562637, + 0.10972774773836136, + 0.022500690072774887, + 0.005127019714564085, + 0.0030523682944476604, + 0.004210594575852156, + 0.0056893727742135525, + 0.0050550298765301704, + 0.004245712887495756, + 0.005457624793052673, + 0.004273420665413141, + 0.004146837629377842, + 0.00418409239500761, + 0.002904431661590934, + 0.00516071543097496, + 0.003522953949868679, + 0.0044625503942370415, + 0.0031636233907192945, + 0.003582516685128212, + 0.002698478288948536, + 0.002487804275006056, + 0.0029836136382073164, + 0.0028206815477460623, + 0.0025602562818676233, + 0.0021922749001532793, + 0.003615719499066472, + 0.0031836084090173244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.20161272585391998 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "Filmed with a stunning high-speed macro camera, the camera slowly and steadily moves forward, revealing a vast, undulating landscape composed of countless densely packed structures resembling healthy intestinal villi. These finger-like projections, rendered in soft, warm pink and coral hues, possess a velvety, almost textile-like surface texture and gently pulsate with perfect consistency, a slow, rhythmic wave of contraction and expansion flowing throughout the frame. Countless particles of luminous golden light, suspended in a perfectly clear, pristine liquid, drift weightlessly, carried by invisible microcurrents. As these luminous specks approach the villi, their surface seems to deliberately draw them in; the particles make contact and are rhythmically absorbed, their golden light gently diffusing into the pink tissue and generating a faint internal glow that temporarily illuminates the structures from within. The scene is illuminated solely by the particles themselves, casting a warm, ethereal light that creates delicate highlights and soft shadows within the complex folds of the landscape, establishing an enchanting and serene atmosphere and showcasing a beautiful yet abstract display of biological absorption.", + "seed": 15064, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_13/index_5986_seed_15064_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3310657409951091, + "vlm_logprob": -0.3310657409951091, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.3310657409951091, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08419765532016754 + ], + "opposite_logprob_tokens": [ + -2.584197759628296 + ], + "logprob_reward": -0.08419765532016754, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08419765532016754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03427143022418022, + 0.02448512241244316, + 0.02523118257522583, + 0.024661030620336533, + 0.03087296150624752, + 0.013299933634698391, + 0.1205964908003807, + 0.04181361198425293, + 0.5818864107131958, + 0.10288186371326447 + ], + "credit_map": [ + 0.004193290136754513, + 0.004654013551771641, + 0.004092034418135881, + 0.004390417132526636, + 0.00654883636161685, + 0.004235957283526659, + 0.0030986538622528315, + 0.0030582258477807045, + 0.003350106067955494, + 0.0025597463827580214, + 0.005134516395628452, + 0.0021624076180160046, + 0.0032363554928451777, + 0.0027955712284892797, + 0.00279439939185977, + 0.0024520219303667545, + 0.003080892376601696, + 0.0028667771257460117, + 0.0038335812278091908, + 0.0025686374865472317, + 0.0028438784647732973, + 0.004370507318526506, + 0.002519680419936776, + 0.0031472258269786835, + 0.0019115334143862128, + 0.0018247911939397454, + 0.0066682761535048485, + 0.0019226892618462443, + 0.00458311103284359, + 0.0027369277086108923, + 0.0024183487985283136, + 0.0025953538715839386, + 0.004109468311071396, + 0.004042916931211948, + 0.006060387939214706, + 0.003719743574038148, + 0.004210658371448517, + 0.003207368543371558, + 0.002619236009195447, + 0.002903181128203869, + 0.001814482850022614, + 0.0007137197535485029, + 0.0020462963730096817, + 0.001802360755391419, + 0.0010666020680218935, + 0.0018321310635656118, + 0.0013620129320770502, + 0.0026623273734003305, + 0.026089495047926903, + 0.009145104326307774, + 0.011379229836165905, + 0.038966547697782516, + 0.006974256597459316, + 0.008615863509476185, + 0.007703113369643688, + 0.01172288041561842, + 0.004376864992082119, + 0.0032568357419222593, + 0.007236844394356012, + 0.006337781902402639, + 0.004226204007863998, + 0.0049541424959897995, + 0.004870550241321325, + 0.006554385181516409, + 0.03412614017724991, + 0.013892566785216331, + 0.22193492949008942, + 0.03666127100586891, + 0.026762258261442184, + 0.027249949052929878, + 0.18564365804195404, + 0.03561563789844513, + 0.009009404107928276, + 0.0063743311911821365, + 0.02702287770807743, + 0.018947534263134003, + 0.010450787842273712, + 0.009310055524110794, + 0.009164019487798214, + 0.01260285172611475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.08419765532016754 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023325225338339806 + ], + "opposite_logprob_tokens": [ + -3.773325204849243 + ], + "logprob_reward": -0.023325225338339806, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023325225338339806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04746873676776886, + 0.043583355844020844, + 0.03983192518353462, + 0.026613524183630943, + 0.01896491087973118, + 0.022597048431634903, + 0.07375334948301315, + 0.008084463886916637, + 0.694614589214325, + 0.024488072842359543 + ], + "credit_map": [ + 0.004701725672930479, + 0.009347569197416306, + 0.004937727935612202, + 0.004791724495589733, + 0.008576147258281708, + 0.005346162710338831, + 0.005525326356291771, + 0.004242352209985256, + 0.009026295505464077, + 0.00476917764171958, + 0.00678985845297575, + 0.004755252972245216, + 0.004638795740902424, + 0.006475937087088823, + 0.004587405361235142, + 0.0025406337808817625, + 0.005034947767853737, + 0.004610639065504074, + 0.006694038864225149, + 0.004287270829081535, + 0.0035338453017175198, + 0.005000095348805189, + 0.006104778964072466, + 0.004566306713968515, + 0.0018604635260999203, + 0.001577219576574862, + 0.007373125292360783, + 0.0014039853122085333, + 0.007925919257104397, + 0.002993914531543851, + 0.001590221538208425, + 0.0018886752659454942, + 0.0022787204943597317, + 0.002178828464820981, + 0.0027550552040338516, + 0.002322963671758771, + 0.0022816623095422983, + 0.0023480700328946114, + 0.0028749590273946524, + 0.0019246528390794992, + 0.003051420208066702, + 0.0015911116497591138, + 0.004260048270225525, + 0.003311404725536704, + 0.0017082730773836374, + 0.002588045084849, + 0.002232893370091915, + 0.0038538523949682713, + 0.0142856165766716, + 0.003792417934164405, + 0.00491332495585084, + 0.029148826375603676, + 0.002876999555155635, + 0.002424253150820732, + 0.00481161754578352, + 0.011500297114253044, + 0.0010551800951361656, + 0.0005515381344594061, + 0.0011439700610935688, + 0.0010916199535131454, + 0.0008115776581689715, + 0.0009280230151489377, + 0.0008953301585279405, + 0.0016072243452072144, + 0.022379321977496147, + 0.02235305681824684, + 0.18876276910305023, + 0.29524385929107666, + 0.01988263987004757, + 0.01737910509109497, + 0.03817183896899223, + 0.09044203162193298, + 0.0015438769478350878, + 0.0013866279041394591, + 0.003117969026789069, + 0.004421323072165251, + 0.0019570093136280775, + 0.002941662445664406, + 0.005789760034531355, + 0.0033298442140221596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.023325225338339806 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008652695454657078 + ], + "opposite_logprob_tokens": [ + -4.758652687072754 + ], + "logprob_reward": -0.008652695454657078, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008652695454657078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06678338348865509, + 0.19314692914485931, + 0.04404473304748535, + 0.027713963761925697, + 0.02381323091685772, + 0.02217387594282627, + 0.1095055490732193, + 0.018817689269781113, + 0.457943856716156, + 0.03605677932500839 + ], + "credit_map": [ + 0.006296128034591675, + 0.010334687307476997, + 0.008268763311207294, + 0.008010300807654858, + 0.012090603820979595, + 0.008857037872076035, + 0.007590909022837877, + 0.005334951914846897, + 0.035297051072120667, + 0.021636908873915672, + 0.034655697643756866, + 0.019764499738812447, + 0.021347694098949432, + 0.026659870520234108, + 0.02349076420068741, + 0.010294447652995586, + 0.003999921027570963, + 0.004123480059206486, + 0.007476274389773607, + 0.006205128505825996, + 0.003458705497905612, + 0.005878849420696497, + 0.006214132998138666, + 0.006688243243843317, + 0.0036694591399282217, + 0.004050231538712978, + 0.003145734081044793, + 0.0022627401631325483, + 0.005843895021826029, + 0.003666795790195465, + 0.0028319822158664465, + 0.0022431258112192154, + 0.003485885914415121, + 0.0018904849421232939, + 0.0038800037000328302, + 0.002115837298333645, + 0.00402627931907773, + 0.003821096383035183, + 0.0024336737114936113, + 0.002159970346838236, + 0.0023531580809503794, + 0.001679810811765492, + 0.0034377898555248976, + 0.0034230994060635567, + 0.0016820656601339579, + 0.002506592310965061, + 0.002482510171830654, + 0.004608849063515663, + 0.009053991176187992, + 0.008301813155412674, + 0.013648119755089283, + 0.03345327451825142, + 0.008990069851279259, + 0.01080082543194294, + 0.009475228376686573, + 0.015782229602336884, + 0.0017743677599355578, + 0.0011661250609904528, + 0.0033535403199493885, + 0.0032006562687456608, + 0.0017043719999492168, + 0.0021297778002917767, + 0.0022977811750024557, + 0.0031910690013319254, + 0.014294687658548355, + 0.015903374180197716, + 0.09256139397621155, + 0.1968671679496765, + 0.011153988540172577, + 0.013210366480052471, + 0.042739566415548325, + 0.07121329754590988, + 0.003454161575064063, + 0.004133910406380892, + 0.005769111216068268, + 0.0038958783261477947, + 0.003733451012521982, + 0.006069396622478962, + 0.004696747753769159, + 0.004304121248424053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.008652695454657078 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.313456416130066 + ], + "opposite_logprob_tokens": [ + -0.3134564161300659 + ], + "logprob_reward": -1.313456416130066, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.313456416130066, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05215417221188545, + 0.023302972316741943, + 0.024487962946295738, + 0.05717603489756584, + 0.027837097644805908, + 0.0279666930437088, + 0.26076626777648926, + 0.02377232350409031, + 0.4452759027481079, + 0.05726052075624466 + ], + "credit_map": [ + 0.007506121881306171, + 0.007268878631293774, + 0.005079190246760845, + 0.007520013954490423, + 0.00749948900192976, + 0.007488228380680084, + 0.005500994622707367, + 0.004291255958378315, + 0.002551295096054673, + 0.0023720981553196907, + 0.006567368749529123, + 0.002709624357521534, + 0.002381515223532915, + 0.0024319239892065525, + 0.002663475228473544, + 0.0016256708186119795, + 0.004120849538594484, + 0.0031705081928521395, + 0.004241408780217171, + 0.0031520521733909845, + 0.0021128407679498196, + 0.0023282961919903755, + 0.0025377890560775995, + 0.002824218710884452, + 0.0030842560809105635, + 0.0024685657117515802, + 0.01627073623239994, + 0.0033609094098210335, + 0.017398012802004814, + 0.006235762964934111, + 0.0035966152790933847, + 0.004761175252497196, + 0.004031730350106955, + 0.002319329185411334, + 0.0050690011121332645, + 0.0035545395221561193, + 0.0032851763535290956, + 0.0031736355740576982, + 0.0034460104070603848, + 0.0029576763045042753, + 0.0038062084931880236, + 0.001676131971180439, + 0.0039537944830954075, + 0.0058638895861804485, + 0.0019913974683731794, + 0.003124890848994255, + 0.002807757118716836, + 0.004742623306810856, + 0.044786807149648666, + 0.02523197792470455, + 0.02464037388563156, + 0.0759289562702179, + 0.017656836658716202, + 0.017496807500720024, + 0.016413317993283272, + 0.03861118108034134, + 0.00256727053783834, + 0.0020047014113515615, + 0.003840584773570299, + 0.003712765406817198, + 0.0023912345059216022, + 0.002949685091152787, + 0.002075372263789177, + 0.004230709280818701, + 0.024112924933433533, + 0.011680273339152336, + 0.1871948540210724, + 0.09591197222471237, + 0.01852053962647915, + 0.015988675877451897, + 0.07146739959716797, + 0.02039928361773491, + 0.0038827427197247744, + 0.0029134685173630714, + 0.01791365258395672, + 0.006621816195547581, + 0.004113830626010895, + 0.004477594513446093, + 0.01190985832363367, + 0.005427559372037649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -1.313456416130066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera moving slowly and steadily forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the structures in the landscape shaped like densely packed, finger-like intestinal villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the villi possess a soft, warm pink and coral color with a velvety, textile-like surface texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless luminous golden light particles visible drifting in a clear liquid and being absorbed by the villi?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the particles generate a faint internal glow that illuminates the villi from within upon contact?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22569671273231506 + ], + "opposite_logprob_tokens": [ + -1.6006966829299927 + ], + "logprob_reward": -0.22569671273231506, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22569671273231506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03355100378394127, + 0.03159184008836746, + 0.025501487776637077, + 0.05029577761888504, + 0.028165282681584358, + 0.010773667134344578, + 0.39987272024154663, + 0.018539484590291977, + 0.3559863567352295, + 0.0457223504781723 + ], + "credit_map": [ + 0.004738468676805496, + 0.0031706872396171093, + 0.003222200321033597, + 0.005884774494916201, + 0.005206753965467215, + 0.005406465381383896, + 0.0035834689624607563, + 0.0023381835781037807, + 0.005780468694865704, + 0.00312935677357018, + 0.007567096035927534, + 0.00267426366917789, + 0.0027575844433158636, + 0.0035904739052057266, + 0.004025259055197239, + 0.0020673368126153946, + 0.004411676432937384, + 0.004339146893471479, + 0.004201534669846296, + 0.002373593859374523, + 0.0023587450850754976, + 0.0031958238687366247, + 0.0027793338522315025, + 0.0018416332313790917, + 0.002709265099838376, + 0.002464902587234974, + 0.01429829467087984, + 0.002098395489156246, + 0.017148518934845924, + 0.005671132355928421, + 0.0026178841944783926, + 0.0032873887103050947, + 0.003530368208885193, + 0.002768261358141899, + 0.004385158885270357, + 0.003346246900036931, + 0.0031789091881364584, + 0.003650018712505698, + 0.003956094849854708, + 0.0033502241130918264, + 0.0010038897162303329, + 0.0008656911086291075, + 0.0014099366962909698, + 0.001964960480108857, + 0.0009315989445894957, + 0.0013749210629612207, + 0.0015205995878204703, + 0.0017020691884681582, + 0.06753282994031906, + 0.028079181909561157, + 0.03148824721574783, + 0.14250299334526062, + 0.019031185656785965, + 0.02231704257428646, + 0.026120807975530624, + 0.06280042231082916, + 0.0017695139395073056, + 0.0012996011646464467, + 0.0032599964179098606, + 0.002810377161949873, + 0.0018924985779449344, + 0.0027755864430218935, + 0.0019404133781790733, + 0.0027914976235479116, + 0.018486278131604195, + 0.010406892746686935, + 0.14880983531475067, + 0.0934356302022934, + 0.014726785011589527, + 0.012580068781971931, + 0.040257856249809265, + 0.01728304848074913, + 0.003447085851803422, + 0.0023752900306135416, + 0.013596565462648869, + 0.006645885296165943, + 0.0036044945009052753, + 0.003831408219411969, + 0.006821953225880861, + 0.005399670917540789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.22569671273231506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_14/index_tensor([9878, 9878])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_14/index_tensor([9878, 9878])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ea38e300585c40857e98b10f60c4813b04e6ba0d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_14/index_tensor([9878, 9878])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from a slightly upward angle, the shot begins in medium shot, then slowly pulls back to reveal a vaster, more majestic vista, centered on a sacred phoenix. Enveloped in a shimmering golden flame, the flames not blazing, but flowing gracefully across its glossy feathers like liquid gold. Each feather appears meticulously woven from pure sunlight, and its massive, ornate wings flap with a majestic, deliberate rhythm. Each rise and fall of its wings is imbued with power and control, scattering the golden flames backward, carving brilliant trails across the deep indigo sky. Countless burning golden particles, like stardust, emanate from the wings' edges. The phoenix soars above a boundless sea of clouds, soft and fluffy like cotton candy. Its sole source of light, the phoenix's warm, sacred golden glow tints the surrounding cloudtops with a brilliant golden rim, creating an epic atmosphere of tranquility, grandeur, and awe.", + "seed": 28531, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_14/index_9878_seed_28531_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024275316670536996, + "vlm_logprob": -0.024275316670536996, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.024275316670536996, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05323253199458122 + ], + "opposite_logprob_tokens": [ + -3.053232431411743 + ], + "logprob_reward": -0.05323253199458122, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05323253199458122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03582229092717171, + 0.023404795676469803, + 0.06972825527191162, + 0.5216915011405945, + 0.026772087439894676, + 0.15623894333839417, + 0.030243204906582832, + 0.030230864882469177, + 0.07451719790697098, + 0.031350892037153244 + ], + "credit_map": [ + 0.004224848933517933, + 0.004013816826045513, + 0.005485232919454575, + 0.004971460439264774, + 0.004793031606823206, + 0.003721219254657626, + 0.004817724693566561, + 0.0037949520628899336, + 0.003148335497826338, + 0.0029011324513703585, + 0.004555440507829189, + 0.0030345250852406025, + 0.0021397240925580263, + 0.0020832244772464037, + 0.0030680601485073566, + 0.002474351553246379, + 0.014380175620317459, + 0.006084382068365812, + 0.007907586172223091, + 0.007555925752967596, + 0.00780285382643342, + 0.009023958817124367, + 0.006916964892297983, + 0.010056406259536743, + 0.07198707014322281, + 0.009594962932169437, + 0.014316492713987827, + 0.008610425516963005, + 0.014542845077812672, + 0.22821883857250214, + 0.017073586583137512, + 0.15734727680683136, + 0.0036591559182852507, + 0.002453095745295286, + 0.003842028556391597, + 0.0031371042132377625, + 0.001606989768333733, + 0.003247768385335803, + 0.0028029910754412413, + 0.00602295296266675, + 0.019058534875512123, + 0.028974181041121483, + 0.024411635473370552, + 0.018849555402994156, + 0.01151509303599596, + 0.010024591349065304, + 0.0173025019466877, + 0.026102837175130844, + 0.002870655618607998, + 0.0031885053031146526, + 0.0058841644786298275, + 0.005711250472813845, + 0.0023933129850775003, + 0.0017222583992406726, + 0.0038754662964493036, + 0.004597590304911137, + 0.0034631183370947838, + 0.0023842169903218746, + 0.004241046030074358, + 0.0036085075698792934, + 0.0032815830782055855, + 0.0020993107464164495, + 0.007035466376692057, + 0.00411761412397027, + 0.006426102947443724, + 0.0048673017881810665, + 0.005377370864152908, + 0.013002493418753147, + 0.0070190709084272385, + 0.003914056345820427, + 0.009618504904210567, + 0.024292290210723877, + 0.002435497008264065, + 0.0031517932657152414, + 0.004085157532244921, + 0.004189485218375921, + 0.004912927746772766, + 0.003057517111301422, + 0.006048834882676601, + 0.0034696736838668585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.05323253199458122 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005973344668745995 + ], + "opposite_logprob_tokens": [ + -5.1309733390808105 + ], + "logprob_reward": -0.005973344668745995, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005973344668745995, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10629609227180481, + 0.0440998412668705, + 0.07436857372522354, + 0.12577120959758759, + 0.0791415274143219, + 0.06759011745452881, + 0.030987000092864037, + 0.0808025449514389, + 0.2422429472208023, + 0.14870013296604156 + ], + "credit_map": [ + 0.01314515434205532, + 0.012642385438084602, + 0.015451714396476746, + 0.014600838534533978, + 0.013705778867006302, + 0.011486479081213474, + 0.012451251037418842, + 0.012812497094273567, + 0.005256366450339556, + 0.005661366041749716, + 0.00434865802526474, + 0.005293342284858227, + 0.004703794606029987, + 0.00694900332018733, + 0.006479877047240734, + 0.005407434422522783, + 0.013324798084795475, + 0.006759376730769873, + 0.009194835089147091, + 0.00947872456163168, + 0.00726218381896615, + 0.009413509629666805, + 0.009501319378614426, + 0.00943383201956749, + 0.013193941675126553, + 0.005654897540807724, + 0.013100774958729744, + 0.005648162681609392, + 0.005584650672972202, + 0.03908977285027504, + 0.006957103963941336, + 0.0365419015288353, + 0.007001593243330717, + 0.004706850741058588, + 0.022740155458450317, + 0.01190275326371193, + 0.00410126056522131, + 0.005286218132823706, + 0.007589477580040693, + 0.01581321656703949, + 0.008325036615133286, + 0.011956039816141129, + 0.00858266931027174, + 0.007877022959291935, + 0.0058982642367482185, + 0.007232269272208214, + 0.007170998957008123, + 0.01054782047867775, + 0.0027200246695429087, + 0.0022125286050140858, + 0.006970156449824572, + 0.007176860235631466, + 0.0023017849307507277, + 0.0016887213569134474, + 0.0034671605098992586, + 0.0044497656635940075, + 0.00887370016425848, + 0.008260956034064293, + 0.008764886297285557, + 0.013690878637135029, + 0.010218419134616852, + 0.006774105131626129, + 0.014620456844568253, + 0.00959914643317461, + 0.015058214776217937, + 0.00906146876513958, + 0.01538844220340252, + 0.04310601204633713, + 0.009843164123594761, + 0.009225869551301003, + 0.043594129383563995, + 0.09696567058563232, + 0.012300611473619938, + 0.011282430961728096, + 0.030040977522730827, + 0.023508114740252495, + 0.008350104093551636, + 0.013608760200440884, + 0.02738495171070099, + 0.022224200889468193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.005973344668745995 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799826443195343 + ], + "opposite_logprob_tokens": [ + -3.5297999382019043 + ], + "logprob_reward": -0.029799826443195343, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799826443195343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03228673338890076, + 0.057793889194726944, + 0.07899834960699081, + 0.2179667055606842, + 0.07836233079433441, + 0.2095562219619751, + 0.06968492269515991, + 0.07150379568338394, + 0.09345736354589462, + 0.09038973599672318 + ], + "credit_map": [ + 0.0045767370611429214, + 0.0031251332256942987, + 0.007425487507134676, + 0.004279694985598326, + 0.003334088483825326, + 0.003428064752370119, + 0.0028388025239109993, + 0.0032787236850708723, + 0.007133744191378355, + 0.008120881393551826, + 0.004359262064099312, + 0.015082331374287605, + 0.004808282479643822, + 0.006164087913930416, + 0.0047561367973685265, + 0.007369161583483219, + 0.016002334654331207, + 0.006415015086531639, + 0.009782697074115276, + 0.007678629830479622, + 0.007947171106934547, + 0.011103947646915913, + 0.007369123864918947, + 0.012699424289166927, + 0.025007039308547974, + 0.0057393768802285194, + 0.014747818000614643, + 0.007889881730079651, + 0.00713377445936203, + 0.08311773836612701, + 0.009096921421587467, + 0.06523416191339493, + 0.009324216283857822, + 0.00817296002060175, + 0.015304568223655224, + 0.007578042335808277, + 0.007399341091513634, + 0.008222495205700397, + 0.011562364175915718, + 0.01079834345728159, + 0.02575589343905449, + 0.03947218134999275, + 0.02892262116074562, + 0.024730417877435684, + 0.015303732827305794, + 0.014333082363009453, + 0.022190028801560402, + 0.03884825482964516, + 0.006951713468879461, + 0.0049904449842870235, + 0.016973571851849556, + 0.012004745192825794, + 0.006175383925437927, + 0.003071841085329652, + 0.007756008300930262, + 0.011761211790144444, + 0.007166816852986813, + 0.008875636383891106, + 0.009613404981791973, + 0.0138234943151474, + 0.008817771449685097, + 0.00461505027487874, + 0.009516092017292976, + 0.009075530804693699, + 0.006379323545843363, + 0.004431368317455053, + 0.007237844169139862, + 0.022253958508372307, + 0.007064016070216894, + 0.004490514285862446, + 0.010088254697620869, + 0.03151208534836769, + 0.006813081447035074, + 0.008561319671571255, + 0.008890083059668541, + 0.013008302077651024, + 0.014864569529891014, + 0.007077930960804224, + 0.019932720810174942, + 0.011241728439927101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 8 + ], + "selected_reward": -0.029799826443195343 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190716668963432 + ], + "opposite_logprob_tokens": [ + -4.264190673828125 + ], + "logprob_reward": -0.014190716668963432, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190716668963432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03543291240930557, + 0.04329318925738335, + 0.05784028768539429, + 0.0611518993973732, + 0.14251920580863953, + 0.11636032164096832, + 0.04789094254374504, + 0.10600538551807404, + 0.09929759800434113, + 0.29020822048187256 + ], + "credit_map": [ + 0.004093748517334461, + 0.0033225929364562035, + 0.005404761992394924, + 0.0041913967579603195, + 0.0046706716530025005, + 0.0047400896437466145, + 0.004495295695960522, + 0.0045143552124500275, + 0.0052175153978168964, + 0.0048737358301877975, + 0.0029834541492164135, + 0.012106090784072876, + 0.0029487034771591425, + 0.0051614767871797085, + 0.003645240096375346, + 0.006356972735375166, + 0.008520993404090405, + 0.004393113777041435, + 0.006710308603942394, + 0.005997870583087206, + 0.008134416304528713, + 0.009255895391106606, + 0.005005569662898779, + 0.00982211995869875, + 0.008125384338200092, + 0.0018697973573580384, + 0.005993668921291828, + 0.003071399638429284, + 0.0018966806819662452, + 0.022123908624053, + 0.0028445071075111628, + 0.015226553194224834, + 0.016359107568860054, + 0.008680008351802826, + 0.03294200077652931, + 0.01629091054201126, + 0.00816567987203598, + 0.008114028722047806, + 0.013430844992399216, + 0.03853662312030792, + 0.014162172563374043, + 0.021322399377822876, + 0.015059355646371841, + 0.0138787180185318, + 0.00908732507377863, + 0.007465822156518698, + 0.013939638622105122, + 0.0214448943734169, + 0.005265161395072937, + 0.004862518515437841, + 0.0073605915531516075, + 0.01023627445101738, + 0.004324684850871563, + 0.003779500024393201, + 0.005069923121482134, + 0.006992289796471596, + 0.009741034358739853, + 0.005064744036644697, + 0.0149796511977911, + 0.014188366010785103, + 0.0074838572181761265, + 0.006361730862408876, + 0.03348708152770996, + 0.014698919840157032, + 0.007026586681604385, + 0.005167167168110609, + 0.010753655806183815, + 0.02002648077905178, + 0.005060725845396519, + 0.004426438361406326, + 0.026376333087682724, + 0.0204602163285017, + 0.022672154009342194, + 0.01804635301232338, + 0.05467280000448227, + 0.05000171437859535, + 0.01720069907605648, + 0.02282349020242691, + 0.06409479677677155, + 0.040696196258068085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 5 + ], + "selected_reward": -0.014190716668963432 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180163577198982 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180163577198982, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180163577198982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04827331379055977, + 0.11237331479787827, + 0.05531136319041252, + 0.13366393744945526, + 0.10268353670835495, + 0.053825169801712036, + 0.042602598667144775, + 0.11161985248327255, + 0.2655317783355713, + 0.07411517947912216 + ], + "credit_map": [ + 0.006578928325325251, + 0.0049917190335690975, + 0.010283938609063625, + 0.0069434852339327335, + 0.0044148508459329605, + 0.005815751384943724, + 0.0042070746421813965, + 0.005037564784288406, + 0.012766101397573948, + 0.013172541745007038, + 0.012486187741160393, + 0.024238714948296547, + 0.007582618854939938, + 0.016041452065110207, + 0.009552360512316227, + 0.016533346846699715, + 0.00736619345843792, + 0.0029730573296546936, + 0.006583086680620909, + 0.006210686173290014, + 0.0065842121839523315, + 0.009400480426847935, + 0.005916516296565533, + 0.010277136228978634, + 0.014604129828512669, + 0.004049462731927633, + 0.011153792031109333, + 0.004984721541404724, + 0.004415265284478664, + 0.04884568974375725, + 0.006269076839089394, + 0.03934179246425629, + 0.011488579213619232, + 0.01502924133092165, + 0.015569129958748817, + 0.016632001847028732, + 0.0075737047009170055, + 0.0121329165995121, + 0.012946251779794693, + 0.011311715468764305, + 0.006314966361969709, + 0.009591839276254177, + 0.005752203520387411, + 0.004909684881567955, + 0.006995625328272581, + 0.00619974359869957, + 0.005424111615866423, + 0.008636998012661934, + 0.005086799617856741, + 0.00377639033831656, + 0.005534299649298191, + 0.006834541447460651, + 0.0047583975829184055, + 0.0031039046589285135, + 0.005576483439654112, + 0.007931781932711601, + 0.011858826503157616, + 0.00894059706479311, + 0.011494854465126991, + 0.016079772263765335, + 0.015113930217921734, + 0.008168838918209076, + 0.025879886001348495, + 0.014083141461014748, + 0.022100374102592468, + 0.011317425407469273, + 0.017476197332143784, + 0.05087241902947426, + 0.015847131609916687, + 0.009928522631525993, + 0.0370812714099884, + 0.10090843588113785, + 0.006335211452096701, + 0.0051961131393909454, + 0.01692180149257183, + 0.010709167458117008, + 0.0056114462204277515, + 0.00619969516992569, + 0.010259933769702911, + 0.01288180984556675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.018180163577198982 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly upward angle, the shot begins in medium shot, then slowly pulls back to reveal a vaster, more majestic vista, centered on a sacred phoenix. Enveloped in a shimmering golden flame, the flames not blazing, but flowing gracefully across its glossy feathers like liquid gold. Each feather appears meticulously woven from pure sunlight, and its massive, ornate wings flap with a majestic, deliberate rhythm. Each rise and fall of its wings is imbued with power and control, scattering the golden flames backward, carving brilliant trails across the deep indigo sky. Countless burning golden particles, like stardust, emanate from the wings' edges. The phoenix soars above a boundless sea of clouds, soft and fluffy like cotton candy. Its sole source of light, the phoenix's warm, sacred golden glow tints the surrounding cloudtops with a brilliant golden rim, creating an epic atmosphere of tranquility, grandeur, and awe.", + "seed": 28532, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_14/index_9878_seed_28532_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.036028752010315655, + "vlm_logprob": -0.036028752010315655, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.036028752010315655, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026515932753682137 + ], + "opposite_logprob_tokens": [ + -3.9015159606933594 + ], + "logprob_reward": -0.026515932753682137, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.026515932753682137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1242225393652916, + 0.15745893120765686, + 0.1727820336818695, + 0.12944619357585907, + 0.06733141839504242, + 0.05145953595638275, + 0.13298961520195007, + 0.042406704276800156, + 0.0717342272400856, + 0.05016878619790077 + ], + "credit_map": [ + 0.010565534234046936, + 0.00924846064299345, + 0.02349984087049961, + 0.015569671988487244, + 0.010405675508081913, + 0.009981279261410236, + 0.02599656768143177, + 0.018955513834953308, + 0.015034504234790802, + 0.012769436463713646, + 0.03656700253486633, + 0.019590748474001884, + 0.008940545842051506, + 0.008258135989308357, + 0.02742895483970642, + 0.028869593515992165, + 0.018311787396669388, + 0.030793214216828346, + 0.030248217284679413, + 0.018388187512755394, + 0.011977038346230984, + 0.011874962598085403, + 0.021554885432124138, + 0.02963373437523842, + 0.014442129991948605, + 0.020030155777931213, + 0.018666578456759453, + 0.015232515521347523, + 0.013588869012892246, + 0.00816289521753788, + 0.018601303920149803, + 0.02072174660861492, + 0.0068238480016589165, + 0.00991431437432766, + 0.01388639584183693, + 0.010455782525241375, + 0.006173913832753897, + 0.006155597046017647, + 0.007202579639852047, + 0.0067189885303378105, + 0.0057508922182023525, + 0.00614316388964653, + 0.006525631062686443, + 0.005776985082775354, + 0.004604395944625139, + 0.005829658824950457, + 0.007405159063637257, + 0.009423649869859219, + 0.016122331842780113, + 0.011913006193935871, + 0.017453178763389587, + 0.022430168464779854, + 0.011701389215886593, + 0.01667853631079197, + 0.01332845538854599, + 0.0233625415712595, + 0.006144797895103693, + 0.0033279252238571644, + 0.004721707198768854, + 0.006480318494141102, + 0.0038229492492973804, + 0.006138373166322708, + 0.004559572786092758, + 0.007211057003587484, + 0.008885042741894722, + 0.008047488518059254, + 0.00801461935043335, + 0.012451441027224064, + 0.006152537185698748, + 0.009237275458872318, + 0.0075840940698981285, + 0.01136173214763403, + 0.006416176911443472, + 0.007469266653060913, + 0.005435990169644356, + 0.005235590040683746, + 0.007861349731683731, + 0.005923914723098278, + 0.007471225690096617, + 0.004355272743850946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.026515932753682137 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007653201464563608 + ], + "opposite_logprob_tokens": [ + -4.88265323638916 + ], + "logprob_reward": -0.007653201464563608, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007653201464563608, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22566679120063782, + 0.1248752623796463, + 0.07020562887191772, + 0.2169262021780014, + 0.08689029514789581, + 0.05725448578596115, + 0.08631971478462219, + 0.03510351479053497, + 0.04889649152755737, + 0.04786160960793495 + ], + "credit_map": [ + 0.021167490631341934, + 0.01601593941450119, + 0.033995095640420914, + 0.030214395374059677, + 0.02908538095653057, + 0.03111495077610016, + 0.03852866217494011, + 0.02554488554596901, + 0.02271314151585102, + 0.012619530782103539, + 0.014143191277980804, + 0.014895174652338028, + 0.008772761560976505, + 0.004690059460699558, + 0.01999741792678833, + 0.027043987065553665, + 0.007817870005965233, + 0.007478631567209959, + 0.012181962840259075, + 0.008617793209850788, + 0.006327018607407808, + 0.00544934393838048, + 0.009030485525727272, + 0.013302522711455822, + 0.0301523357629776, + 0.030287643894553185, + 0.03629199415445328, + 0.023051133379340172, + 0.01864878088235855, + 0.017046896740794182, + 0.0279120784252882, + 0.03353533148765564, + 0.010209114290773869, + 0.0063247415237128735, + 0.01164640486240387, + 0.024223988875746727, + 0.00620408495888114, + 0.009052570909261703, + 0.010407671332359314, + 0.008821727707982063, + 0.0071516274474561214, + 0.007955287583172321, + 0.007609405089169741, + 0.007027350831776857, + 0.00571546470746398, + 0.007416271138936281, + 0.006783927790820599, + 0.007595148868858814, + 0.011432022787630558, + 0.009028533473610878, + 0.009944149293005466, + 0.010267074219882488, + 0.008245236240327358, + 0.01188487745821476, + 0.010276385582983494, + 0.015241429209709167, + 0.004118956625461578, + 0.002620553132146597, + 0.004548646975308657, + 0.007936164736747742, + 0.002688782988116145, + 0.004017722327262163, + 0.004743814934045076, + 0.00442887470126152, + 0.0071100397035479546, + 0.004799922462552786, + 0.006934594828635454, + 0.009654870256781578, + 0.004227343946695328, + 0.006053542252629995, + 0.004016941413283348, + 0.006099238526076078, + 0.008054791949689388, + 0.003527523949742317, + 0.006917974445968866, + 0.003982787486165762, + 0.004940217360854149, + 0.006562985014170408, + 0.007548831403255463, + 0.006326497998088598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.007653201464563608 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11291978508234024 + ], + "opposite_logprob_tokens": [ + -2.237919807434082 + ], + "logprob_reward": -0.11291978508234024, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11291978508234024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10503855347633362, + 0.17144760489463806, + 0.0890159085392952, + 0.1555713415145874, + 0.13744239509105682, + 0.07615617662668228, + 0.058668289333581924, + 0.09024661779403687, + 0.031864359974861145, + 0.084548719227314 + ], + "credit_map": [ + 0.00742693105712533, + 0.005999239161610603, + 0.017509030178189278, + 0.012087952345609665, + 0.015346694737672806, + 0.012492737732827663, + 0.02228989079594612, + 0.011886080726981163, + 0.017956916242837906, + 0.011830553412437439, + 0.023717058822512627, + 0.02179310843348503, + 0.008535731583833694, + 0.009854336269199848, + 0.0285529512912035, + 0.049206946045160294, + 0.014665071852505207, + 0.014910419471561909, + 0.01186459418386221, + 0.009099593386054039, + 0.007110710255801678, + 0.005206980276852846, + 0.01397071685642004, + 0.012187827378511429, + 0.024496033787727356, + 0.017420589923858643, + 0.023097364231944084, + 0.016771096736192703, + 0.013546421192586422, + 0.00943774078041315, + 0.022265102714300156, + 0.028537020087242126, + 0.009909044951200485, + 0.009359801188111305, + 0.04419584199786186, + 0.04209941253066063, + 0.0074032945558428764, + 0.005418751854449511, + 0.009758994914591312, + 0.009297259151935577, + 0.00892806239426136, + 0.010260144248604774, + 0.01032860018312931, + 0.010800150223076344, + 0.007244559936225414, + 0.00756490183994174, + 0.009493328630924225, + 0.01153643149882555, + 0.007122513372451067, + 0.006658261176198721, + 0.00632475595921278, + 0.006564868614077568, + 0.006729026325047016, + 0.008441921323537827, + 0.006722982972860336, + 0.010103964246809483, + 0.013057034462690353, + 0.004506254103034735, + 0.00792242307215929, + 0.015071910806000233, + 0.007353414315730333, + 0.014648922719061375, + 0.016561076045036316, + 0.011125590652227402, + 0.004187765531241894, + 0.003770024050027132, + 0.004553371574729681, + 0.003529709530994296, + 0.004769077990204096, + 0.003689053701236844, + 0.0037223882973194122, + 0.003642972791567445, + 0.01327600423246622, + 0.011972596868872643, + 0.0059949979186058044, + 0.007416090462356806, + 0.011820334941148758, + 0.011785359121859074, + 0.013459520414471626, + 0.0088238175958395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 4 + ], + "selected_reward": -0.11291978508234024 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009775268845260143 + ], + "opposite_logprob_tokens": [ + -4.634775161743164 + ], + "logprob_reward": -0.009775268845260143, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009775268845260143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02845681831240654, + 0.028049418702721596, + 0.029546594247221947, + 0.714896023273468, + 0.0240463986992836, + 0.015315861441195011, + 0.034061990678310394, + 0.026178162544965744, + 0.051663294434547424, + 0.04778546839952469 + ], + "credit_map": [ + 0.002980592427775264, + 0.001846906146965921, + 0.0035145748406648636, + 0.004241012502461672, + 0.0036637457087635994, + 0.0034172330051660538, + 0.004689772613346577, + 0.0041029807180166245, + 0.004615013487637043, + 0.0025177428033202887, + 0.005339144729077816, + 0.003573687979951501, + 0.002091421512886882, + 0.0014927649172022939, + 0.0054267412051558495, + 0.002992902183905244, + 0.0032626567408442497, + 0.003353249281644821, + 0.007433477323502302, + 0.004171822685748339, + 0.0016659136163070798, + 0.002335605211555958, + 0.003762665204703808, + 0.0035612036008387804, + 0.11096207797527313, + 0.09405840933322906, + 0.11660803109407425, + 0.07742774486541748, + 0.05813169479370117, + 0.05599169060587883, + 0.09356237947940826, + 0.10815402120351791, + 0.0017362169455736876, + 0.0015609476249665022, + 0.007832874543964863, + 0.00616404227912426, + 0.001560934237204492, + 0.0011925846338272095, + 0.001958497567102313, + 0.0020402988884598017, + 0.001720038359053433, + 0.002241160487756133, + 0.0019320801366120577, + 0.0020536924712359905, + 0.001662654336541891, + 0.0015531372046098113, + 0.001970307668671012, + 0.0021827907767146826, + 0.004761800169944763, + 0.0029148489702492952, + 0.00399051932618022, + 0.004172197077423334, + 0.004009206313639879, + 0.0041038719937205315, + 0.002871133154258132, + 0.007238413672894239, + 0.00474757794290781, + 0.001726682297885418, + 0.0030044650193303823, + 0.004522261675447226, + 0.0015551465330645442, + 0.003963547293096781, + 0.0024534936528652906, + 0.004204987082630396, + 0.007271607872098684, + 0.0028804128523916006, + 0.006129525601863861, + 0.009577001444995403, + 0.0028187758289277554, + 0.0067049977369606495, + 0.0043245842680335045, + 0.011956389993429184, + 0.007019710261374712, + 0.0030335576739162207, + 0.006268979050219059, + 0.00528658926486969, + 0.003549034707248211, + 0.005824603606015444, + 0.009579501114785671, + 0.007223488297313452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.009775268845260143 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279571905732155 + ], + "opposite_logprob_tokens": [ + -3.7732796669006348 + ], + "logprob_reward": -0.023279571905732155, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279571905732155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1402875781059265, + 0.09307947754859924, + 0.03681277483701706, + 0.21358810365200043, + 0.1437835991382599, + 0.031194057315587997, + 0.09515777230262756, + 0.05348170921206474, + 0.03499878942966461, + 0.15761607885360718 + ], + "credit_map": [ + 0.010148306377232075, + 0.008437054231762886, + 0.024356884881854057, + 0.02165943942964077, + 0.021247541531920433, + 0.01592876762151718, + 0.02228633500635624, + 0.016223235055804253, + 0.007945757359266281, + 0.008413958363234997, + 0.010959016159176826, + 0.012673611752688885, + 0.006111809518188238, + 0.004999963566660881, + 0.01608995907008648, + 0.025885405018925667, + 0.004978042095899582, + 0.0040136463940143585, + 0.0059828502126038074, + 0.0042421272955834866, + 0.0024674697779119015, + 0.002958002733066678, + 0.004498777911067009, + 0.007671857252717018, + 0.03333297744393349, + 0.02623745985329151, + 0.037930361926555634, + 0.025942591950297356, + 0.01687469147145748, + 0.01677553541958332, + 0.026775481179356575, + 0.029719017446041107, + 0.010805720463395119, + 0.009377548471093178, + 0.043941181153059006, + 0.04302769526839256, + 0.009181467816233635, + 0.005926367826759815, + 0.011106460355222225, + 0.010417165234684944, + 0.003692649770528078, + 0.004158938303589821, + 0.0028940439224243164, + 0.004439557436853647, + 0.0035322632174938917, + 0.0037514320574700832, + 0.004156953189522028, + 0.004568219650536776, + 0.01247651781886816, + 0.015593239106237888, + 0.008941694162786007, + 0.012610272504389286, + 0.009082337841391563, + 0.01149788685142994, + 0.01324145682156086, + 0.011714364401996136, + 0.006288767792284489, + 0.003910165745764971, + 0.006384425796568394, + 0.011107855476439, + 0.00452832318842411, + 0.006490916479378939, + 0.009066476486623287, + 0.005704781506210566, + 0.004153764806687832, + 0.0037350174970924854, + 0.005289745982736349, + 0.005520602222532034, + 0.005054232198745012, + 0.0042112129740417, + 0.0029821062926203012, + 0.004052106291055679, + 0.016418030485510826, + 0.02710607461631298, + 0.01160033792257309, + 0.012318010441958904, + 0.028001731261610985, + 0.02123730070888996, + 0.025924313813447952, + 0.015010281465947628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.023279571905732155 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly upward angle, the shot begins in medium shot, then slowly pulls back to reveal a vaster, more majestic vista, centered on a sacred phoenix. Enveloped in a shimmering golden flame, the flames not blazing, but flowing gracefully across its glossy feathers like liquid gold. Each feather appears meticulously woven from pure sunlight, and its massive, ornate wings flap with a majestic, deliberate rhythm. Each rise and fall of its wings is imbued with power and control, scattering the golden flames backward, carving brilliant trails across the deep indigo sky. Countless burning golden particles, like stardust, emanate from the wings' edges. The phoenix soars above a boundless sea of clouds, soft and fluffy like cotton candy. Its sole source of light, the phoenix's warm, sacred golden glow tints the surrounding cloudtops with a brilliant golden rim, creating an epic atmosphere of tranquility, grandeur, and awe.", + "seed": 28533, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_14/index_9878_seed_28533_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03536747042089701, + "vlm_logprob": -0.03536747042089701, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.03536747042089701, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03961387649178505 + ], + "opposite_logprob_tokens": [ + -3.414613962173462 + ], + "logprob_reward": -0.03961387649178505, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03961387649178505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09227647632360458, + 0.16852359473705292, + 0.03435133397579193, + 0.1647350937128067, + 0.04072783514857292, + 0.26883599162101746, + 0.06672071665525436, + 0.08105684071779251, + 0.042739298194646835, + 0.040032777935266495 + ], + "credit_map": [ + 0.014854616485536098, + 0.0073896609246730804, + 0.011830084957182407, + 0.00865314994007349, + 0.007354762405157089, + 0.014362791553139687, + 0.0072157918475568295, + 0.02061561867594719, + 0.01929504983127117, + 0.015159904956817627, + 0.03331702575087547, + 0.021873919293284416, + 0.007703426294028759, + 0.015388395637273788, + 0.01512879692018032, + 0.04065708443522453, + 0.004491083789616823, + 0.004187124315649271, + 0.0037828313652426004, + 0.005414887331426144, + 0.005327257327735424, + 0.0034457044675946236, + 0.0042261239141225815, + 0.0034763242583721876, + 0.027255656197667122, + 0.011179805733263493, + 0.019709056243300438, + 0.06381713598966599, + 0.010234742425382137, + 0.008822962641716003, + 0.013331788592040539, + 0.010383939370512962, + 0.004679636098444462, + 0.005150239914655685, + 0.007807937916368246, + 0.005106586497277021, + 0.0038330191746354103, + 0.002847864292562008, + 0.005263750907033682, + 0.006038801744580269, + 0.04958757758140564, + 0.02174302749335766, + 0.03486352786421776, + 0.022711794823408127, + 0.031285565346479416, + 0.03460632637143135, + 0.049073003232479095, + 0.02496517449617386, + 0.010289667174220085, + 0.006207332480698824, + 0.010598008520901203, + 0.007363216485828161, + 0.006914949510246515, + 0.004777644295245409, + 0.010630939155817032, + 0.009938967414200306, + 0.01827062852680683, + 0.004697490017861128, + 0.011224743910133839, + 0.025891397148370743, + 0.003472575917840004, + 0.0030631907284259796, + 0.006621385458856821, + 0.007815432734787464, + 0.004624457098543644, + 0.005096015520393848, + 0.006847570650279522, + 0.005387357901781797, + 0.006331318058073521, + 0.0030150189995765686, + 0.007008632645010948, + 0.004428931977599859, + 0.007450829725712538, + 0.0038929185830056667, + 0.004842645488679409, + 0.005365987773984671, + 0.0027744497638195753, + 0.003684388706460595, + 0.00657995417714119, + 0.005441606044769287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.03961387649178505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005980336107313633 + ], + "opposite_logprob_tokens": [ + -5.130980491638184 + ], + "logprob_reward": -0.005980336107313633, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005980336107313633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14161749184131622, + 0.29518622159957886, + 0.04679875075817108, + 0.09267932921648026, + 0.0780404731631279, + 0.07491056621074677, + 0.04506406560540199, + 0.09401185065507889, + 0.09892068058252335, + 0.03277056664228439 + ], + "credit_map": [ + 0.01858416199684143, + 0.0170974750071764, + 0.02490794099867344, + 0.021582553163170815, + 0.011757921427488327, + 0.013694905675947666, + 0.01395229622721672, + 0.020040249451994896, + 0.03007800504565239, + 0.024133387953042984, + 0.053722258657217026, + 0.033830784261226654, + 0.022733626887202263, + 0.02872214838862419, + 0.034777749329805374, + 0.06718827039003372, + 0.006993157789111137, + 0.004676097072660923, + 0.008759365417063236, + 0.007319370284676552, + 0.0039382330141961575, + 0.004526760894805193, + 0.005446846596896648, + 0.005138918291777372, + 0.009768055751919746, + 0.013385189697146416, + 0.012101643718779087, + 0.02278578095138073, + 0.011040108278393745, + 0.006606620270758867, + 0.009177017025649548, + 0.007814921438694, + 0.008702286519110203, + 0.004908318631350994, + 0.013641078025102615, + 0.02120700664818287, + 0.005066032987087965, + 0.0054677994921803474, + 0.007234939374029636, + 0.011813019402325153, + 0.009627090767025948, + 0.006889187730848789, + 0.012545697391033173, + 0.007991689257323742, + 0.008931521326303482, + 0.010318038053810596, + 0.009942479431629181, + 0.008664865978062153, + 0.005292345304042101, + 0.005285116843879223, + 0.008660421706736088, + 0.006158290430903435, + 0.00508841359987855, + 0.0033345757983624935, + 0.004999891389161348, + 0.006245016120374203, + 0.018606355413794518, + 0.0058589959517121315, + 0.010844497010111809, + 0.03385695442557335, + 0.005079476162791252, + 0.003541187848895788, + 0.008767811581492424, + 0.007456573192030191, + 0.02018519677221775, + 0.008265096694231033, + 0.019908742979168892, + 0.010632123798131943, + 0.0075027719140052795, + 0.009546618908643723, + 0.008941351436078548, + 0.0139387808740139, + 0.005568967200815678, + 0.0026586612220853567, + 0.004094351548701525, + 0.004156234674155712, + 0.0034749647602438927, + 0.0032549297902733088, + 0.005210381466895342, + 0.004352078307420015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.005980336107313633 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001892477273941 + ], + "opposite_logprob_tokens": [ + -2.695019006729126 + ], + "logprob_reward": -0.07001892477273941, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001892477273941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043807823210954666, + 0.0829799547791481, + 0.0502103827893734, + 0.546395480632782, + 0.055196166038513184, + 0.08799059689044952, + 0.0185391902923584, + 0.02499074675142765, + 0.01716560870409012, + 0.07272404432296753 + ], + "credit_map": [ + 0.0050807916559278965, + 0.0051332637667655945, + 0.007999680005013943, + 0.006975881289690733, + 0.00325214397162199, + 0.003660950344055891, + 0.00420979131013155, + 0.007495325058698654, + 0.017647450789809227, + 0.008477548137307167, + 0.007791634649038315, + 0.008704773150384426, + 0.006300259847193956, + 0.01941538229584694, + 0.005749275907874107, + 0.008893630467355251, + 0.007449187804013491, + 0.004732409492135048, + 0.0065363431349396706, + 0.007091425824910402, + 0.006593579426407814, + 0.005384182091802359, + 0.006998993456363678, + 0.005424263421446085, + 0.1008775532245636, + 0.029670007526874542, + 0.06416047364473343, + 0.21948324143886566, + 0.02855757810175419, + 0.025760246440768242, + 0.04212629050016403, + 0.03576009348034859, + 0.005384752992540598, + 0.004802376497536898, + 0.009059042669832706, + 0.012123958207666874, + 0.004335601814091206, + 0.004574724473059177, + 0.00517515791580081, + 0.009740553796291351, + 0.016210129484534264, + 0.006722138728946447, + 0.012682389467954636, + 0.00760893477126956, + 0.009523998945951462, + 0.011449064128100872, + 0.015532048419117928, + 0.008261887356638908, + 0.0020622271113097668, + 0.0020611847285181284, + 0.002982034580782056, + 0.0027782053221017122, + 0.0021162675693631172, + 0.0008961677085608244, + 0.0031442914623767138, + 0.0024988127406686544, + 0.0055087837390601635, + 0.00155642069876194, + 0.002236601896584034, + 0.008707033470273018, + 0.0017282346962019801, + 0.0006674728356301785, + 0.002506410935893655, + 0.0020797881297767162, + 0.002460758900269866, + 0.0017027293797582388, + 0.002488216385245323, + 0.0021573675330728292, + 0.0015108162770047784, + 0.0013642539270222187, + 0.0030090543441474438, + 0.0024724137037992477, + 0.008274620398879051, + 0.00739828497171402, + 0.010098958387970924, + 0.012107664719223976, + 0.009820724837481976, + 0.0051572867669165134, + 0.011753653176128864, + 0.00811284501105547 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 1 + ], + "selected_reward": -0.07001892477273941 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018178056925535202 + ], + "opposite_logprob_tokens": [ + -4.0181779861450195 + ], + "logprob_reward": -0.018178056925535202, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018178056925535202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10192248225212097, + 0.05982980877161026, + 0.03744346275925636, + 0.13048627972602844, + 0.09475604444742203, + 0.06460943073034286, + 0.07029131054878235, + 0.09885305166244507, + 0.20869404077529907, + 0.13311415910720825 + ], + "credit_map": [ + 0.019928274676203728, + 0.0070901778526604176, + 0.012238085269927979, + 0.01206995639950037, + 0.0056069595739245415, + 0.011988968588411808, + 0.008285037241876125, + 0.02471500262618065, + 0.007591054309159517, + 0.006688358727842569, + 0.007034484762698412, + 0.009578448720276356, + 0.0042149838991463184, + 0.005790499970316887, + 0.0069001102820038795, + 0.012031864374876022, + 0.005318840034306049, + 0.004236776381731033, + 0.004987059626728296, + 0.005771418567746878, + 0.005351829808205366, + 0.0035407766699790955, + 0.004405041690915823, + 0.0038317148573696613, + 0.023479536175727844, + 0.007986257784068584, + 0.01538660004734993, + 0.051199790090322495, + 0.007392244879156351, + 0.006024015601724386, + 0.010106871835887432, + 0.008910948410630226, + 0.011431104503571987, + 0.007348251063376665, + 0.01997065544128418, + 0.01981819048523903, + 0.007295444142073393, + 0.004990155342966318, + 0.00957823172211647, + 0.014324002899229527, + 0.008488894440233707, + 0.006301407236605883, + 0.009370251558721066, + 0.007244574371725321, + 0.008504347875714302, + 0.008380393497645855, + 0.008257274515926838, + 0.008062278851866722, + 0.010635318234562874, + 0.005927780643105507, + 0.013187674805521965, + 0.00864297617226839, + 0.004829278215765953, + 0.00527534494176507, + 0.008874853141605854, + 0.012918071821331978, + 0.017035091295838356, + 0.005632299464195967, + 0.013882630504667759, + 0.03175309672951698, + 0.005256575997918844, + 0.0054477727971971035, + 0.009210397489368916, + 0.010635174810886383, + 0.04172554239630699, + 0.01813744194805622, + 0.04207221791148186, + 0.02473743073642254, + 0.01571652479469776, + 0.018101077526807785, + 0.018322354182600975, + 0.029881440103054047, + 0.021545900031924248, + 0.008505207486450672, + 0.03372745215892792, + 0.018977459520101547, + 0.00830660481005907, + 0.007561114616692066, + 0.015828140079975128, + 0.018662262707948685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.018178056925535202 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304615780711174 + ], + "opposite_logprob_tokens": [ + -3.168046236038208 + ], + "logprob_reward": -0.04304615780711174, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304615780711174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09440140426158905, + 0.12860606610774994, + 0.04102169722318649, + 0.1441061645746231, + 0.1003449484705925, + 0.21273113787174225, + 0.05181203410029411, + 0.10295308381319046, + 0.04035875201225281, + 0.0836646780371666 + ], + "credit_map": [ + 0.014496926218271255, + 0.010314522311091423, + 0.011273931711912155, + 0.01570701040327549, + 0.006040018051862717, + 0.008246644400060177, + 0.009415749460458755, + 0.018906598910689354, + 0.017682095989584923, + 0.019671054556965828, + 0.0208022091537714, + 0.015462723560631275, + 0.010004268027842045, + 0.013487730175256729, + 0.011864619329571724, + 0.019631361588835716, + 0.004341779742389917, + 0.005566645413637161, + 0.006648763548582792, + 0.006691506132483482, + 0.005224673077464104, + 0.002892401535063982, + 0.005241330713033676, + 0.004414594732224941, + 0.025882015004754066, + 0.013365098275244236, + 0.015858935192227364, + 0.050604093819856644, + 0.009970255196094513, + 0.006137377582490444, + 0.012301531620323658, + 0.009986864402890205, + 0.011525120586156845, + 0.008224355056881905, + 0.020786307752132416, + 0.027115358039736748, + 0.005283705424517393, + 0.005161387380212545, + 0.008540118113160133, + 0.013708599843084812, + 0.03870045021176338, + 0.019401434808969498, + 0.02313748560845852, + 0.022786496207118034, + 0.01800849847495556, + 0.019111135974526405, + 0.051664870232343674, + 0.019920770078897476, + 0.007593117654323578, + 0.0055915419943630695, + 0.010257340967655182, + 0.00723870936781168, + 0.006013845559209585, + 0.0033837584778666496, + 0.005856095813214779, + 0.005877628922462463, + 0.017619045451283455, + 0.011640581302344799, + 0.0175238698720932, + 0.013306974433362484, + 0.00912526622414589, + 0.007439424283802509, + 0.010554693639278412, + 0.015743231400847435, + 0.006599768064916134, + 0.0037025813944637775, + 0.0036047534085810184, + 0.004136893432587385, + 0.002874450059607625, + 0.004890053998678923, + 0.006965597625821829, + 0.00758465426042676, + 0.012176882475614548, + 0.007367664482444525, + 0.019399186596274376, + 0.010268460027873516, + 0.00757818715646863, + 0.007265942171216011, + 0.007898970507085323, + 0.01170938741415739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.04304615780711174 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly upward angle, the shot begins in medium shot, then slowly pulls back to reveal a vaster, more majestic vista, centered on a sacred phoenix. Enveloped in a shimmering golden flame, the flames not blazing, but flowing gracefully across its glossy feathers like liquid gold. Each feather appears meticulously woven from pure sunlight, and its massive, ornate wings flap with a majestic, deliberate rhythm. Each rise and fall of its wings is imbued with power and control, scattering the golden flames backward, carving brilliant trails across the deep indigo sky. Countless burning golden particles, like stardust, emanate from the wings' edges. The phoenix soars above a boundless sea of clouds, soft and fluffy like cotton candy. Its sole source of light, the phoenix's warm, sacred golden glow tints the surrounding cloudtops with a brilliant golden rim, creating an epic atmosphere of tranquility, grandeur, and awe.", + "seed": 28534, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_14/index_9878_seed_28534_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.048976126313209536, + "vlm_logprob": -0.048976126313209536, + "vlm_yes_no_margin": 3.6249999523162844, + "vlm_token_credit": -0.048976126313209536, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02229424938559532 + ], + "opposite_logprob_tokens": [ + -4.022294044494629 + ], + "logprob_reward": -0.02229424938559532, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.02229424938559532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1936190128326416, + 0.065034419298172, + 0.1608314961194992, + 0.0453750304877758, + 0.06289950758218765, + 0.2800447344779968, + 0.04421069100499153, + 0.05313832312822342, + 0.033118005841970444, + 0.06172880902886391 + ], + "credit_map": [ + 0.010596276260912418, + 0.008942962624132633, + 0.01008078083395958, + 0.025387154892086983, + 0.011212409473955631, + 0.01313490979373455, + 0.06772397458553314, + 0.04654054716229439, + 0.005671082064509392, + 0.005271804053336382, + 0.007833192124962807, + 0.01272356417030096, + 0.005558387376368046, + 0.005069151520729065, + 0.006647381465882063, + 0.01625985838472843, + 0.013472309336066246, + 0.015996580943465233, + 0.018574334681034088, + 0.01892869919538498, + 0.016936831176280975, + 0.011135229840874672, + 0.04249796271324158, + 0.02328953705728054, + 0.003666296601295471, + 0.004507082514464855, + 0.011232767254114151, + 0.005816230084747076, + 0.006218296941369772, + 0.0037935900036245584, + 0.005720276851207018, + 0.004420492798089981, + 0.005075342021882534, + 0.004355335608124733, + 0.009361994452774525, + 0.007640562020242214, + 0.0064434465020895, + 0.005652164574712515, + 0.01054312102496624, + 0.013827545568346977, + 0.03999746963381767, + 0.030595853924751282, + 0.03317330777645111, + 0.028922202065587044, + 0.040286533534526825, + 0.02553870901465416, + 0.031976327300071716, + 0.04955434054136276, + 0.006734962575137615, + 0.005110177677124739, + 0.005775355733931065, + 0.005432054866105318, + 0.004651905037462711, + 0.004400456789880991, + 0.005651556886732578, + 0.006454219575971365, + 0.0055190022103488445, + 0.004128412343561649, + 0.004779115319252014, + 0.004969182424247265, + 0.011332537978887558, + 0.010200779885053635, + 0.007640677969902754, + 0.004568617325276136, + 0.0032030348666012287, + 0.0031422413885593414, + 0.004321218468248844, + 0.003980511799454689, + 0.006554621737450361, + 0.0033916349057108164, + 0.004602575674653053, + 0.003922165837138891, + 0.004739930387586355, + 0.0057907081209123135, + 0.005840294063091278, + 0.006437279284000397, + 0.013952921144664288, + 0.00606981385499239, + 0.013108872808516026, + 0.0057889907620847225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.02229424938559532 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00597393698990345 + ], + "opposite_logprob_tokens": [ + -5.130973815917969 + ], + "logprob_reward": -0.00597393698990345, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00597393698990345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2837192714214325, + 0.049227356910705566, + 0.1237017959356308, + 0.029406316578388214, + 0.027910679578781128, + 0.36434048414230347, + 0.03959021344780922, + 0.017400318756699562, + 0.026509754359722137, + 0.038193799555301666 + ], + "credit_map": [ + 0.012330378405749798, + 0.007984171621501446, + 0.012065141461789608, + 0.026859914883971214, + 0.011943655088543892, + 0.011181733570992947, + 0.11402728408575058, + 0.08732698857784271, + 0.0056974575854837894, + 0.005394116044044495, + 0.008594632148742676, + 0.008208641782402992, + 0.0035289402585476637, + 0.006247426848858595, + 0.006068251561373472, + 0.005487892311066389, + 0.010865776799619198, + 0.011480040848255157, + 0.013288082554936409, + 0.01350689958781004, + 0.018654393032193184, + 0.010070184245705605, + 0.026971586048603058, + 0.01886482909321785, + 0.002228535944595933, + 0.0029102549888193607, + 0.004260217305272818, + 0.002897700062021613, + 0.005995452404022217, + 0.003221557242795825, + 0.004089724738150835, + 0.003802876453846693, + 0.0024096141569316387, + 0.001648695208132267, + 0.00496230972930789, + 0.00334175955504179, + 0.00239315046928823, + 0.0021147867664694786, + 0.004479352384805679, + 0.00656101293861866, + 0.050438087433576584, + 0.039850831031799316, + 0.043125275522470474, + 0.03906787186861038, + 0.05406501516699791, + 0.03251947835087776, + 0.03703315556049347, + 0.06824079155921936, + 0.0055662174709141254, + 0.005242347717285156, + 0.0038634738884866238, + 0.004612913820892572, + 0.004518249537795782, + 0.0044512031599879265, + 0.00728429714217782, + 0.0040515088476240635, + 0.002025452209636569, + 0.0019218133529648185, + 0.0016352712409570813, + 0.002163340337574482, + 0.002342052524909377, + 0.0028140475042164326, + 0.0028613212052732706, + 0.0016370207304134965, + 0.0031828575301915407, + 0.0025204282719641924, + 0.003817928722128272, + 0.0027913174126297235, + 0.0043114167638123035, + 0.0037257648073136806, + 0.003505413420498371, + 0.0026546271983534098, + 0.004004629794508219, + 0.003123810049146414, + 0.005127473268657923, + 0.004854455590248108, + 0.0038903316017240286, + 0.004040752537548542, + 0.007623033132404089, + 0.005529311019927263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.00597393698990345 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16034434735774994 + ], + "opposite_logprob_tokens": [ + -1.9103443622589111 + ], + "logprob_reward": -0.16034434735774994, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16034434735774994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23834331333637238, + 0.0987546294927597, + 0.10755737125873566, + 0.06606867909431458, + 0.04266345873475075, + 0.15404409170150757, + 0.1255299299955368, + 0.05903426557779312, + 0.08379179984331131, + 0.024212460964918137 + ], + "credit_map": [ + 0.012709374539554119, + 0.01575934700667858, + 0.019150439649820328, + 0.02075684815645218, + 0.016782715916633606, + 0.019237414002418518, + 0.07060391455888748, + 0.06334323436021805, + 0.010279289446771145, + 0.009558125399053097, + 0.008558773435652256, + 0.020214637741446495, + 0.010128690861165524, + 0.007684864569455385, + 0.009568234905600548, + 0.022762015461921692, + 0.009303150698542595, + 0.01014180388301611, + 0.010329141281545162, + 0.01166016049683094, + 0.016732241958379745, + 0.010009669698774815, + 0.023608144372701645, + 0.015773067250847816, + 0.005008543375879526, + 0.0067552002146840096, + 0.0070508685894310474, + 0.008351203054189682, + 0.015247298404574394, + 0.008282549679279327, + 0.00874332431703806, + 0.0066296933218836784, + 0.0034204369876533747, + 0.00344105064868927, + 0.0060584512539207935, + 0.0058891731314361095, + 0.004887154325842857, + 0.0036197248846292496, + 0.008417925797402859, + 0.006929539609700441, + 0.014759400859475136, + 0.018446633592247963, + 0.015611735172569752, + 0.019336815923452377, + 0.021895533427596092, + 0.023778634145855904, + 0.021404767408967018, + 0.018810583278536797, + 0.015958700329065323, + 0.011832146905362606, + 0.015422718599438667, + 0.015175472013652325, + 0.01790943741798401, + 0.011421511881053448, + 0.024752961471676826, + 0.013056967407464981, + 0.006348294205963612, + 0.0048950426280498505, + 0.005065898410975933, + 0.0065286834724247456, + 0.013266644440591335, + 0.011059083975851536, + 0.00715925358235836, + 0.004711365792900324, + 0.006116964854300022, + 0.007957392372190952, + 0.00777010852470994, + 0.010543648153543472, + 0.016264310106635094, + 0.009374482557177544, + 0.016742076724767685, + 0.009022819809615612, + 0.0021062688902020454, + 0.0021030157804489136, + 0.002328519243746996, + 0.0035957195796072483, + 0.0040719700045883656, + 0.0023010242730379105, + 0.004971585236489773, + 0.0027343560941517353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.16034434735774994 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182387575507164 + ], + "opposite_logprob_tokens": [ + -4.018182277679443 + ], + "logprob_reward": -0.018182387575507164, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018182387575507164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1318458467721939, + 0.036540742963552475, + 0.16352154314517975, + 0.07955020666122437, + 0.042916569858789444, + 0.17072002589702606, + 0.08198583871126175, + 0.08679353445768356, + 0.0942624881863594, + 0.1118631660938263 + ], + "credit_map": [ + 0.006141019519418478, + 0.007090504746884108, + 0.00847938098013401, + 0.014723791740834713, + 0.008484731428325176, + 0.008131277747452259, + 0.04583825170993805, + 0.032956890761852264, + 0.003364080563187599, + 0.00546586187556386, + 0.004278586246073246, + 0.004840746987611055, + 0.002876905957236886, + 0.004469675477594137, + 0.005672901403158903, + 0.005571987479925156, + 0.014024723321199417, + 0.017221523448824883, + 0.021441049873828888, + 0.018560968339443207, + 0.024110939353704453, + 0.016917498782277107, + 0.026723412796854973, + 0.02452145330607891, + 0.008425583131611347, + 0.005752654746174812, + 0.02342362143099308, + 0.009895033203065395, + 0.004719523247331381, + 0.005146656651049852, + 0.007709796074777842, + 0.014477343298494816, + 0.004958718083798885, + 0.0036606749054044485, + 0.011201386339962482, + 0.005502380896359682, + 0.0033167810179293156, + 0.004619744140654802, + 0.003907883074134588, + 0.0057490067556500435, + 0.023178396746516228, + 0.01819577068090439, + 0.020684655755758286, + 0.01754174567759037, + 0.021588237956166267, + 0.014248890802264214, + 0.01771809719502926, + 0.0375642403960228, + 0.009072056040167809, + 0.008155939169228077, + 0.009219593368470669, + 0.01074485294520855, + 0.01004945207387209, + 0.009431561455130577, + 0.011249278672039509, + 0.014063107781112194, + 0.01278169360011816, + 0.007033273112028837, + 0.009401497431099415, + 0.009683177806437016, + 0.007637967821210623, + 0.021584926173090935, + 0.010337566025555134, + 0.00833344180136919, + 0.012380742467939854, + 0.006267961580306292, + 0.018628310412168503, + 0.00960446521639824, + 0.009131313301622868, + 0.014538614079356194, + 0.011562620289623737, + 0.012148469686508179, + 0.013002577237784863, + 0.008726291358470917, + 0.01999616250395775, + 0.01662902906537056, + 0.008770477958023548, + 0.01219774316996336, + 0.01712469756603241, + 0.01541619747877121 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.018182387575507164 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sacred phoenix with glossy feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers appearing woven from pure sunlight and enveloped in flowing golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the phoenix's wings flapping with a deliberate rhythm, scattering golden flames backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep indigo sky with a boundless sea of soft, fluffy clouds below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot from a slightly upward angle and slowly pull back to reveal a vast vista?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038085710257291794 + ], + "opposite_logprob_tokens": [ + -3.288085699081421 + ], + "logprob_reward": -0.038085710257291794, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038085710257291794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06346751749515533, + 0.05876320227980614, + 0.17676761746406555, + 0.028496071696281433, + 0.023666825145483017, + 0.5846098065376282, + 0.027221988886594772, + 0.014543221332132816, + 0.012331793084740639, + 0.0101320156827569 + ], + "credit_map": [ + 0.004379755817353725, + 0.004828975535929203, + 0.007971187122166157, + 0.00850712601095438, + 0.004478937014937401, + 0.005699755158275366, + 0.016841333359479904, + 0.010760444216430187, + 0.004111147951334715, + 0.007635415066033602, + 0.004982523620128632, + 0.00950499065220356, + 0.008567709475755692, + 0.00424192612990737, + 0.005428887903690338, + 0.01429060474038124, + 0.016685567796230316, + 0.01561782881617546, + 0.02570180967450142, + 0.021289626136422157, + 0.025152744725346565, + 0.011507967486977577, + 0.03814928978681564, + 0.022662781178951263, + 0.0026670722290873528, + 0.0032635240349918604, + 0.00464205164462328, + 0.0030738201458007097, + 0.005036143120378256, + 0.002879636362195015, + 0.003348548896610737, + 0.003585275262594223, + 0.0029140023980289698, + 0.0021573277190327644, + 0.0038346038199961185, + 0.003297003684565425, + 0.0022112829610705376, + 0.0018791230395436287, + 0.0025728377513587475, + 0.0048006447032094, + 0.07826665788888931, + 0.0678769052028656, + 0.07255352288484573, + 0.06473584473133087, + 0.086371511220932, + 0.05848345905542374, + 0.056485921144485474, + 0.09983598440885544, + 0.0035647789482027292, + 0.003250597510486841, + 0.0030532933305948973, + 0.0031189334113150835, + 0.0027821653056889772, + 0.003766626352444291, + 0.004572882317006588, + 0.003112712875008583, + 0.0016756175318732858, + 0.0012445530155673623, + 0.0018576725851744413, + 0.0016777811106294394, + 0.001531675923615694, + 0.0030077029950916767, + 0.0020035167690366507, + 0.0015447005862370133, + 0.001363211078569293, + 0.0009395746164955199, + 0.0018645047675818205, + 0.0014035720378160477, + 0.0023744211066514254, + 0.0016058958135545254, + 0.0013382572215050459, + 0.0014423562679439783, + 0.000884242937900126, + 0.0009152444545179605, + 0.0009819442639127374, + 0.0012752908514812589, + 0.0017762079369276762, + 0.0010024793446063995, + 0.0021753099281340837, + 0.0011212957324460149 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.038085710257291794 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_15/index_tensor([713, 713])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_15/index_tensor([713, 713])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3bba7369ba61adca0118625331769065a4ed5eb6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_15/index_tensor([713, 713])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning aerial drone shot, the camera held still, dominates the frame as a massive white tiger crouches on a rugged expanse of dark gray, almost black volcanic rock. Its thick, snow-white fur is streaked with deep black stripes, flowing like ink along the contours of its muscular body. At that moment, its massive head tilts upward, facing the camera overhead, its maw gaping to its fullest, revealing dagger-sharp canines and a curled, pink tongue, as if unleashing a silent yet earth-shattering roar. Though silent, the bulging muscles of its neck and shoulders, its eyes squeezed shut with intense exertion, and its flared nostrils convey a sense of devastating, explosive power. An intense, concentrated beam of overhead light shines vertically down, precisely outlining the undulations of the tiger's spine and the texture of every inch of its muscles, while casting deep, sharply defined shadows on the rough rocky ground. The dominant tones of the entire painting are extreme black and white, creating a primal, almost sacred sense of power and an awe-inspiring silence, imbued with unparalleled visual tension.", + "seed": 2941, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_15/index_713_seed_2941_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.5051767010241747, + "vlm_logprob": -0.5051767010241747, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.5051767010241747, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012008044868707657 + ], + "opposite_logprob_tokens": [ + -4.637008190155029 + ], + "logprob_reward": -0.012008044868707657, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012008044868707657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2381618171930313, + 0.2723688781261444, + 0.043031737208366394, + 0.06672261655330658, + 0.05057476833462715, + 0.08382196724414825, + 0.1016499474644661, + 0.04122396185994148, + 0.07136727124452591, + 0.031077049672603607 + ], + "credit_map": [ + 0.03970377892255783, + 0.01645912602543831, + 0.07463671267032623, + 0.018064914271235466, + 0.014447745867073536, + 0.014080570079386234, + 0.018812960013747215, + 0.041956011205911636, + 0.040806785225868225, + 0.017741763964295387, + 0.08691469579935074, + 0.018638869747519493, + 0.03973023220896721, + 0.01840927079319954, + 0.013857945799827576, + 0.036269303411245346, + 0.0062667191959917545, + 0.004196806810796261, + 0.006213100161403418, + 0.006138999946415424, + 0.004453221336007118, + 0.007064708974212408, + 0.00473752012476325, + 0.003960659261792898, + 0.00793673750013113, + 0.007867692969739437, + 0.007383027113974094, + 0.01249037217348814, + 0.008697094395756721, + 0.005388231948018074, + 0.008689284324645996, + 0.008270173333585262, + 0.00703459931537509, + 0.005032537039369345, + 0.006151512730866671, + 0.007679465692490339, + 0.005810264497995377, + 0.005772924516350031, + 0.00687671173363924, + 0.006216753739863634, + 0.01664658635854721, + 0.008587978780269623, + 0.011767999269068241, + 0.006782734766602516, + 0.008169211447238922, + 0.011534281075000763, + 0.006737601011991501, + 0.01359557081013918, + 0.02399642951786518, + 0.008062522858381271, + 0.017029287293553352, + 0.01282969769090414, + 0.0064904759638011456, + 0.010112534277141094, + 0.011183002032339573, + 0.011945999227464199, + 0.007300597615540028, + 0.005387172568589449, + 0.005138838663697243, + 0.00542242918163538, + 0.00495718652382493, + 0.00319930212572217, + 0.004899608436971903, + 0.004918825812637806, + 0.010060738772153854, + 0.008192132227122784, + 0.009815869852900505, + 0.010565249249339104, + 0.009243110194802284, + 0.00609154487028718, + 0.007523880340158939, + 0.009874743409454823, + 0.004969004541635513, + 0.0028598648495972157, + 0.0037149449344724417, + 0.004534929525107145, + 0.0030375767964869738, + 0.0035746567882597446, + 0.004254159051924944, + 0.004131915979087353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.012008044868707657 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330239862203598 + ], + "opposite_logprob_tokens": [ + -3.7733023166656494 + ], + "logprob_reward": -0.02330239862203598, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330239862203598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11924586445093155, + 0.25657838582992554, + 0.04511120170354843, + 0.0472165010869503, + 0.06209462136030197, + 0.113169364631176, + 0.07435519248247147, + 0.12864139676094055, + 0.08979716151952744, + 0.06379033625125885 + ], + "credit_map": [ + 0.02180432714521885, + 0.013127356767654419, + 0.03331605717539787, + 0.011208472773432732, + 0.009071661159396172, + 0.006975379306823015, + 0.00966974813491106, + 0.014072858728468418, + 0.038392551243305206, + 0.016607997938990593, + 0.08199511468410492, + 0.017807291820645332, + 0.037658412009477615, + 0.017411179840564728, + 0.011612010188400745, + 0.03509383648633957, + 0.007588006556034088, + 0.004299344960600138, + 0.006691206246614456, + 0.006976064760237932, + 0.004666158463805914, + 0.0052266656421124935, + 0.005033802706748247, + 0.004629950039088726, + 0.006816998589783907, + 0.005480040330439806, + 0.005416912958025932, + 0.007893814705312252, + 0.005112280137836933, + 0.003964677453041077, + 0.0057168323546648026, + 0.006814945954829454, + 0.0070631965063512325, + 0.007097779307514429, + 0.009021823294460773, + 0.007499272469431162, + 0.011825699359178543, + 0.005642084404826164, + 0.008573425933718681, + 0.005371340084820986, + 0.015536687336862087, + 0.011094527319073677, + 0.013936433009803295, + 0.014829938299953938, + 0.010786703787744045, + 0.016395309939980507, + 0.01621287502348423, + 0.014376896433532238, + 0.01194873359054327, + 0.007426057942211628, + 0.010506494902074337, + 0.00835293810814619, + 0.0065293703228235245, + 0.009819160215556622, + 0.006602572277188301, + 0.013169855810701847, + 0.01584339328110218, + 0.017447851598262787, + 0.017803367227315903, + 0.020656101405620575, + 0.015491748228669167, + 0.010804595425724983, + 0.012431668117642403, + 0.018162677064538002, + 0.013226098380982876, + 0.00890160258859396, + 0.012473882175981998, + 0.011262777261435986, + 0.012033914215862751, + 0.00844783615320921, + 0.009734943509101868, + 0.013716102577745914, + 0.010859199799597263, + 0.005738976411521435, + 0.007562466897070408, + 0.00818062387406826, + 0.009666780941188335, + 0.005917801521718502, + 0.009784512221813202, + 0.006079971790313721 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.02330239862203598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058218978345394 + ], + "opposite_logprob_tokens": [ + -3.8955821990966797 + ], + "logprob_reward": -0.02058218978345394, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058218978345394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41535893082618713, + 0.13631150126457214, + 0.07232781499624252, + 0.046277470886707306, + 0.0571281760931015, + 0.10834841430187225, + 0.04361592233181, + 0.04696893319487572, + 0.03595744073390961, + 0.03770536929368973 + ], + "credit_map": [ + 0.06700840592384338, + 0.03533042222261429, + 0.10631618648767471, + 0.044157642871141434, + 0.036950938403606415, + 0.028117351233959198, + 0.04223156347870827, + 0.055246468633413315, + 0.020215798169374466, + 0.009720737114548683, + 0.029786569997668266, + 0.01616310141980648, + 0.009910952299833298, + 0.010783309116959572, + 0.011177632957696915, + 0.028553415089845657, + 0.007958688773214817, + 0.009414700791239738, + 0.012201878242194653, + 0.010664055123925209, + 0.008798155002295971, + 0.007014717906713486, + 0.008115090429782867, + 0.008160530589520931, + 0.007431478705257177, + 0.006048383191227913, + 0.005594434216618538, + 0.005782789085060358, + 0.004228385630995035, + 0.006059907842427492, + 0.005544096697121859, + 0.0055880011059343815, + 0.009592155925929546, + 0.006592956371605396, + 0.009148775599896908, + 0.005830409005284309, + 0.007334756664931774, + 0.00604171771556139, + 0.0060144225135445595, + 0.006572984158992767, + 0.015850046649575233, + 0.01285415980964899, + 0.014172319322824478, + 0.014396167360246181, + 0.013190540485084057, + 0.011580698192119598, + 0.01500053983181715, + 0.011303961277008057, + 0.0071935574524104595, + 0.004521999042481184, + 0.006014718674123287, + 0.008088330738246441, + 0.0033017215318977833, + 0.004753554239869118, + 0.005060567054897547, + 0.0046814787201583385, + 0.006117439363151789, + 0.005416539963334799, + 0.0058447206392884254, + 0.005674571264535189, + 0.006632237695157528, + 0.004415004979819059, + 0.007125604432076216, + 0.005742822773754597, + 0.005209443159401417, + 0.0035809383261948824, + 0.004854568745940924, + 0.00682938564568758, + 0.004532011225819588, + 0.003513828618451953, + 0.003357220208272338, + 0.0040800501592457294, + 0.009117459878325462, + 0.003717902349308133, + 0.005526553839445114, + 0.004265565425157547, + 0.0038951721508055925, + 0.003930976614356041, + 0.003264607395976782, + 0.003987137693911791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.02058218978345394 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950592458248138 + ], + "opposite_logprob_tokens": [ + -5.130950450897217 + ], + "logprob_reward": -0.005950592458248138, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950592458248138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13173796236515045, + 0.3329010307788849, + 0.0599488690495491, + 0.05709260702133179, + 0.07455756515264511, + 0.12093199789524078, + 0.056913357228040695, + 0.06904920935630798, + 0.06394342333078384, + 0.032923925668001175 + ], + "credit_map": [ + 0.025123825296759605, + 0.011904318816959858, + 0.0309501513838768, + 0.014240222051739693, + 0.010089889168739319, + 0.012311465106904507, + 0.008638299070298672, + 0.018479786813259125, + 0.049446944147348404, + 0.020704790949821472, + 0.10953629016876221, + 0.01970011368393898, + 0.047032397240400314, + 0.022072304040193558, + 0.018415536731481552, + 0.04599263519048691, + 0.008075975812971592, + 0.006870643701404333, + 0.00935785286128521, + 0.008327562361955643, + 0.006423970218747854, + 0.005875738337635994, + 0.006443953141570091, + 0.008573174476623535, + 0.007377333473414183, + 0.006699494551867247, + 0.008763611316680908, + 0.009952377527952194, + 0.005580764729529619, + 0.005731572397053242, + 0.0050668553449213505, + 0.007920598611235619, + 0.01985386572778225, + 0.0058763353154063225, + 0.011666719801723957, + 0.0075406660325825214, + 0.004739138297736645, + 0.007946331985294819, + 0.0037719078827649355, + 0.01316259428858757, + 0.02023581601679325, + 0.013561723753809929, + 0.017047232016921043, + 0.01589885726571083, + 0.012251920998096466, + 0.012985346838831902, + 0.014586047269403934, + 0.01436506025493145, + 0.008863256312906742, + 0.007041088305413723, + 0.007554005831480026, + 0.0070446147583425045, + 0.0056913201697170734, + 0.008130567148327827, + 0.005162478424608707, + 0.007426026742905378, + 0.010320506989955902, + 0.00789681077003479, + 0.011056103743612766, + 0.011798238381743431, + 0.00595385767519474, + 0.005369709338992834, + 0.006002096924930811, + 0.010651887394487858, + 0.01828310824930668, + 0.005475039128214121, + 0.011089703999459743, + 0.009670085273683071, + 0.003723731730133295, + 0.0044181207194924355, + 0.004820225294679403, + 0.006463413592427969, + 0.008572129532694817, + 0.0031671326141804457, + 0.004987811204046011, + 0.0037182215601205826, + 0.003152405144646764, + 0.0034129777923226357, + 0.002656955039128661, + 0.0032562939450144768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.005950592458248138 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.4640402793884277 + ], + "opposite_logprob_tokens": [ + -0.08904033154249191 + ], + "logprob_reward": -2.4640402793884277, + "yes_no_margin_reward": -2.375, + "token_credit_reward": -2.4640402793884277, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2052464634180069, + 0.058580365031957626, + 0.0665048211812973, + 0.08543388545513153, + 0.08736290782690048, + 0.13631786406040192, + 0.12190669775009155, + 0.06050318852066994, + 0.12464305013418198, + 0.05350077152252197 + ], + "credit_map": [ + 0.027859553694725037, + 0.025654230266809464, + 0.04548321291804314, + 0.020029254257678986, + 0.018785350024700165, + 0.013424157164990902, + 0.015211598016321659, + 0.03879909962415695, + 0.013172613456845284, + 0.005047145765274763, + 0.013027217239141464, + 0.005601974669843912, + 0.004440974444150925, + 0.0038598845712840557, + 0.005333176348358393, + 0.008097381331026554, + 0.010601688176393509, + 0.0073399776592850685, + 0.008524123579263687, + 0.006487415172159672, + 0.007335711736232042, + 0.007586276158690453, + 0.008958479389548302, + 0.009671149775385857, + 0.01608867757022381, + 0.009743083268404007, + 0.012575386092066765, + 0.012757045216858387, + 0.005993995815515518, + 0.009153186343610287, + 0.007672629784792662, + 0.011449882760643959, + 0.01712374947965145, + 0.008767458610236645, + 0.012678722850978374, + 0.01323106698691845, + 0.006336923222988844, + 0.009504381567239761, + 0.00930019561201334, + 0.010420409962534904, + 0.025732358917593956, + 0.016989795491099358, + 0.01991095393896103, + 0.011916212737560272, + 0.01169002428650856, + 0.018746891990303993, + 0.011446462012827396, + 0.019885174930095673, + 0.025920448824763298, + 0.012646063230931759, + 0.017578519880771637, + 0.01690719649195671, + 0.007892348803579807, + 0.01224812213331461, + 0.01024638582020998, + 0.018467608839273453, + 0.014101524837315083, + 0.005171949043869972, + 0.009161975234746933, + 0.005447596777230501, + 0.005674428306519985, + 0.005311399232596159, + 0.006002604495733976, + 0.009631709195673466, + 0.024534963071346283, + 0.013551481068134308, + 0.018383830785751343, + 0.016663895919919014, + 0.012033323757350445, + 0.014443820342421532, + 0.011626439169049263, + 0.01340529602020979, + 0.01314183697104454, + 0.0045271869748830795, + 0.008950071409344673, + 0.0057716951705515385, + 0.004391615279018879, + 0.005220189690589905, + 0.004931305069476366, + 0.006566870491951704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -2.4640402793884277 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "stunning aerial drone shot, the camera held still, dominates the frame as a massive white tiger crouches on a rugged expanse of dark gray, almost black volcanic rock. Its thick, snow-white fur is streaked with deep black stripes, flowing like ink along the contours of its muscular body. At that moment, its massive head tilts upward, facing the camera overhead, its maw gaping to its fullest, revealing dagger-sharp canines and a curled, pink tongue, as if unleashing a silent yet earth-shattering roar. Though silent, the bulging muscles of its neck and shoulders, its eyes squeezed shut with intense exertion, and its flared nostrils convey a sense of devastating, explosive power. An intense, concentrated beam of overhead light shines vertically down, precisely outlining the undulations of the tiger's spine and the texture of every inch of its muscles, while casting deep, sharply defined shadows on the rough rocky ground. The dominant tones of the entire painting are extreme black and white, creating a primal, almost sacred sense of power and an awe-inspiring silence, imbued with unparalleled visual tension.", + "seed": 2942, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_15/index_713_seed_2942_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3421495479531586, + "vlm_logprob": -0.3421495479531586, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.3421495479531586, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.041339095681905746 + ], + "opposite_logprob_tokens": [ + -3.416339159011841 + ], + "logprob_reward": -0.041339095681905746, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.041339095681905746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47418391704559326, + 0.056576553732156754, + 0.033948078751564026, + 0.03397297486662865, + 0.020045679062604904, + 0.02620912715792656, + 0.14515073597431183, + 0.05444836616516113, + 0.10521985590457916, + 0.05024468153715134 + ], + "credit_map": [ + 0.04000084474682808, + 0.056597743183374405, + 0.11938901245594025, + 0.03335944935679436, + 0.04728325456380844, + 0.031043613329529762, + 0.031288258731365204, + 0.11522171646356583, + 0.005853745620697737, + 0.008382183499634266, + 0.010988274589180946, + 0.0041297124698758125, + 0.006105304695665836, + 0.0038419177290052176, + 0.004933382384479046, + 0.012342038564383984, + 0.002303375629708171, + 0.0020554657094180584, + 0.009792485274374485, + 0.001577900256961584, + 0.003216018434613943, + 0.0022664424031972885, + 0.0027144509367644787, + 0.010021939873695374, + 0.004058200400322676, + 0.004182681441307068, + 0.0070872739888727665, + 0.0017615716205909848, + 0.003313229186460376, + 0.003195814322680235, + 0.0028469020035117865, + 0.007527303881943226, + 0.001688665128313005, + 0.0035017149057239294, + 0.00332609168253839, + 0.001615558285266161, + 0.002404820639640093, + 0.0019624591805040836, + 0.0017367992550134659, + 0.003809569636359811, + 0.0020788072142750025, + 0.0020556843373924494, + 0.0050823502242565155, + 0.0011596903204917908, + 0.004516495391726494, + 0.002240169094875455, + 0.0028604946564882994, + 0.0062154363840818405, + 0.005840456113219261, + 0.00851485040038824, + 0.043757397681474686, + 0.006064021028578281, + 0.013118680566549301, + 0.01270492747426033, + 0.006918103434145451, + 0.0482323095202446, + 0.0034473969135433435, + 0.004934393335133791, + 0.013417727313935757, + 0.002200473565608263, + 0.00518497359007597, + 0.005435440689325333, + 0.0040082745254039764, + 0.01581968180835247, + 0.0036200990434736013, + 0.00461113266646862, + 0.037276800721883774, + 0.00326516292989254, + 0.003689203644171357, + 0.009199334308505058, + 0.002551503013819456, + 0.04100660979747772, + 0.004344030283391476, + 0.0031916596926748753, + 0.011390154249966145, + 0.0016382235335186124, + 0.003686796873807907, + 0.016004642471671104, + 0.0028194745536893606, + 0.007169701624661684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.041339095681905746 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982516586780548 + ], + "opposite_logprob_tokens": [ + -3.529825210571289 + ], + "logprob_reward": -0.02982516586780548, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982516586780548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2676987648010254, + 0.04592648893594742, + 0.055422380566596985, + 0.034382838755846024, + 0.04075317829847336, + 0.0534135140478611, + 0.07523864507675171, + 0.09066947549581528, + 0.15391188859939575, + 0.18258288502693176 + ], + "credit_map": [ + 0.026194684207439423, + 0.03979898616671562, + 0.04741499945521355, + 0.022032124921679497, + 0.03468453884124756, + 0.021289529278874397, + 0.02469165064394474, + 0.05159226432442665, + 0.004552808124572039, + 0.007145551964640617, + 0.008714447729289532, + 0.0035409354604780674, + 0.005548814311623573, + 0.003401699475944042, + 0.004675149451941252, + 0.008347081951797009, + 0.0026015795301645994, + 0.005717348773032427, + 0.01630096510052681, + 0.003001420060172677, + 0.00583450123667717, + 0.0033155009150505066, + 0.0024137587752193213, + 0.016237307339906693, + 0.0033446410670876503, + 0.0035086581483483315, + 0.007924743928015232, + 0.0018385323928669095, + 0.003838181495666504, + 0.0032852194271981716, + 0.0030091339722275734, + 0.0076337275095283985, + 0.003696235129609704, + 0.006266497075557709, + 0.00824189092963934, + 0.002270358381792903, + 0.004566427320241928, + 0.004309645388275385, + 0.004335164092481136, + 0.007066960446536541, + 0.004829296376556158, + 0.005814660806208849, + 0.01072217058390379, + 0.0029197330586612225, + 0.007754936348646879, + 0.0044960337691009045, + 0.004220306407660246, + 0.012656373903155327, + 0.003588048741221428, + 0.006009775213897228, + 0.01855628378689289, + 0.0032884087413549423, + 0.007606968749314547, + 0.007852577604353428, + 0.008049826137721539, + 0.02028675749897957, + 0.007356169167906046, + 0.012024972587823868, + 0.01442747563123703, + 0.0066613745875656605, + 0.018982701003551483, + 0.01027668733149767, + 0.007309846114367247, + 0.013630242086946964, + 0.004872975870966911, + 0.006618476007133722, + 0.05441617965698242, + 0.004061401821672916, + 0.0065827276557683945, + 0.01200451236218214, + 0.00535163376480341, + 0.060003966093063354, + 0.02116997353732586, + 0.015294752083718777, + 0.03295957297086716, + 0.012522569857537746, + 0.02446623519062996, + 0.03683561086654663, + 0.01377967931330204, + 0.0255544763058424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.02982516586780548 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371181711554527 + ], + "opposite_logprob_tokens": [ + -3.4087119102478027 + ], + "logprob_reward": -0.03371181711554527, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371181711554527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10978186875581741, + 0.11133038997650146, + 0.033532705157995224, + 0.05406605452299118, + 0.1404140740633011, + 0.06345769017934799, + 0.2278779149055481, + 0.037854086607694626, + 0.19708772003650665, + 0.024597562849521637 + ], + "credit_map": [ + 0.009140979498624802, + 0.010037306696176529, + 0.03181370347738266, + 0.005793901160359383, + 0.00943590048700571, + 0.007547818589955568, + 0.0076423874124884605, + 0.02836986631155014, + 0.0061845057643949986, + 0.006268478464335203, + 0.03701840713620186, + 0.004013767931610346, + 0.005685701966285706, + 0.005257315933704376, + 0.005885733291506767, + 0.041016463190317154, + 0.0030900691635906696, + 0.0039862883277237415, + 0.006314477883279324, + 0.00216218875721097, + 0.005062843672931194, + 0.003339961636811495, + 0.003772291587665677, + 0.005804579239338636, + 0.0037612945307046175, + 0.008906489238142967, + 0.009258284233510494, + 0.0036058693658560514, + 0.007398633752018213, + 0.006896383129060268, + 0.004279495216906071, + 0.009959600865840912, + 0.009064855985343456, + 0.013557501137256622, + 0.03670099377632141, + 0.005691606551408768, + 0.01400140393525362, + 0.008341114968061447, + 0.008388584479689598, + 0.04466800019145012, + 0.006610433105379343, + 0.00877175759524107, + 0.010500468313694, + 0.0041021304205060005, + 0.008837968111038208, + 0.008120331913232803, + 0.005722165107727051, + 0.01079243328422308, + 0.009681555442512035, + 0.015236703678965569, + 0.06484343856573105, + 0.009094166569411755, + 0.02280508354306221, + 0.018527012318372726, + 0.01429843157529831, + 0.07339151203632355, + 0.0035677191335707903, + 0.005768857430666685, + 0.0075483196415007114, + 0.002962953643873334, + 0.00445152074098587, + 0.003955144435167313, + 0.003071501152589917, + 0.006528067402541637, + 0.006164371035993099, + 0.005176427774131298, + 0.07295425981283188, + 0.0036194028798490763, + 0.005656030494719744, + 0.017100566998124123, + 0.004094964824616909, + 0.0823216661810875, + 0.0014963132562115788, + 0.002473491011187434, + 0.004683283623307943, + 0.0012336944928392768, + 0.0031232633627951145, + 0.006124972831457853, + 0.0018822579877451062, + 0.0035802836064249277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.03371181711554527 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525747099891305 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.00525747099891305, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525747099891305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060568057000637054, + 0.10631256550550461, + 0.05284615233540535, + 0.0530443973839283, + 0.026421567425131798, + 0.05354366824030876, + 0.4870198667049408, + 0.04599886015057564, + 0.08600391447544098, + 0.028241030871868134 + ], + "credit_map": [ + 0.0061575789004564285, + 0.01311548613011837, + 0.009033525362610817, + 0.004417626187205315, + 0.010896347463130951, + 0.0051574381068348885, + 0.0055561363697052, + 0.006233911495655775, + 0.005285982042551041, + 0.010982531122863293, + 0.02867220900952816, + 0.004144469741731882, + 0.009658229537308216, + 0.007208982016891241, + 0.008763554506003857, + 0.031596601009368896, + 0.0024884401354938745, + 0.0026925087440758944, + 0.017896732315421104, + 0.0017822079826146364, + 0.003009144915267825, + 0.002558268141001463, + 0.0022408082149922848, + 0.020178034901618958, + 0.004642052575945854, + 0.009156052954494953, + 0.009217644110321999, + 0.003440890461206436, + 0.005484080873429775, + 0.007378548849374056, + 0.0041245087049901485, + 0.00960061140358448, + 0.003379278350621462, + 0.0042711468413472176, + 0.0038461429066956043, + 0.001571446773596108, + 0.004782653413712978, + 0.002631389768794179, + 0.002179561648517847, + 0.003759942017495632, + 0.005370361264795065, + 0.006313474383205175, + 0.008968774229288101, + 0.0027921139262616634, + 0.008801762945950031, + 0.006398324854671955, + 0.004861967638134956, + 0.010036884807050228, + 0.021084073930978775, + 0.032387059181928635, + 0.14720980823040009, + 0.017700428143143654, + 0.03468889743089676, + 0.03941820561885834, + 0.025813709944486618, + 0.1687176674604416, + 0.0032203588634729385, + 0.0058860015124082565, + 0.010929419659078121, + 0.0025511286221444607, + 0.005314735695719719, + 0.00401289714500308, + 0.0026105400174856186, + 0.011473774909973145, + 0.005700522568076849, + 0.005683878902345896, + 0.02781274914741516, + 0.0020384304225444794, + 0.004060867242515087, + 0.007621040102094412, + 0.0038552263285964727, + 0.029231196269392967, + 0.0023261194583028555, + 0.0030311306472867727, + 0.005372985266149044, + 0.001992424950003624, + 0.0038541266694664955, + 0.005095504689961672, + 0.0028130074497312307, + 0.0037557303439825773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.00525747099891305 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6006141901016235 + ], + "opposite_logprob_tokens": [ + -0.22561423480510712 + ], + "logprob_reward": -1.6006141901016235, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6006141901016235, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1942426860332489, + 0.03830617666244507, + 0.030070431530475616, + 0.09160033613443375, + 0.12226759642362595, + 0.11162953078746796, + 0.3225804269313812, + 0.03612830489873886, + 0.0149619709700346, + 0.03821253404021263 + ], + "credit_map": [ + 0.018302302807569504, + 0.028183361515402794, + 0.04450974985957146, + 0.01534654013812542, + 0.019604355096817017, + 0.012744761072099209, + 0.014157525263726711, + 0.04139408841729164, + 0.003316954942420125, + 0.002690741326659918, + 0.009595230221748352, + 0.0017359990160912275, + 0.0027775911148637533, + 0.0025893354322761297, + 0.004593358840793371, + 0.011006963439285755, + 0.002636219607666135, + 0.003188920207321644, + 0.006574100814759731, + 0.0019241065019741654, + 0.0033115604892373085, + 0.0024067312479019165, + 0.003059754380956292, + 0.006969035603106022, + 0.006594842299818993, + 0.010292187333106995, + 0.02073112316429615, + 0.004949284251779318, + 0.009481715969741344, + 0.010323304682970047, + 0.006863907445222139, + 0.022363973781466484, + 0.008706612512469292, + 0.010895862244069576, + 0.03225176781415939, + 0.005780498031526804, + 0.013264707289636135, + 0.006461200304329395, + 0.007971026003360748, + 0.036935918033123016, + 0.010395652614533901, + 0.015980565920472145, + 0.018419910222291946, + 0.007411363534629345, + 0.017180155962705612, + 0.013618597760796547, + 0.009832557290792465, + 0.018790725618600845, + 0.01268548984080553, + 0.02256958931684494, + 0.09579222649335861, + 0.013259506784379482, + 0.02725346013903618, + 0.025049611926078796, + 0.016386592760682106, + 0.10958395153284073, + 0.0016420349711552262, + 0.0022770606447011232, + 0.011527256108820438, + 0.001486923429183662, + 0.002016717102378607, + 0.0021946774795651436, + 0.001422004890628159, + 0.013561628758907318, + 0.0011736026499420404, + 0.0018603741191327572, + 0.002751984167844057, + 0.0014101011911407113, + 0.0025061238557100296, + 0.0018573247361928225, + 0.0011698984308168292, + 0.0022325615864247084, + 0.002720566699281335, + 0.0027473927475512028, + 0.00792606920003891, + 0.0015947750071063638, + 0.0031096276361495256, + 0.012307221069931984, + 0.002095169387757778, + 0.0057117147371172905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -1.6006141901016235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "stunning aerial drone shot, the camera held still, dominates the frame as a massive white tiger crouches on a rugged expanse of dark gray, almost black volcanic rock. Its thick, snow-white fur is streaked with deep black stripes, flowing like ink along the contours of its muscular body. At that moment, its massive head tilts upward, facing the camera overhead, its maw gaping to its fullest, revealing dagger-sharp canines and a curled, pink tongue, as if unleashing a silent yet earth-shattering roar. Though silent, the bulging muscles of its neck and shoulders, its eyes squeezed shut with intense exertion, and its flared nostrils convey a sense of devastating, explosive power. An intense, concentrated beam of overhead light shines vertically down, precisely outlining the undulations of the tiger's spine and the texture of every inch of its muscles, while casting deep, sharply defined shadows on the rough rocky ground. The dominant tones of the entire painting are extreme black and white, creating a primal, almost sacred sense of power and an awe-inspiring silence, imbued with unparalleled visual tension.", + "seed": 2943, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_15/index_713_seed_2943_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.5613054861314595, + "vlm_logprob": -0.5613054861314595, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.5613054861314595, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03310007229447365 + ], + "opposite_logprob_tokens": [ + -3.908100128173828 + ], + "logprob_reward": -0.03310007229447365, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.03310007229447365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22712232172489166, + 0.17766140401363373, + 0.06364620476961136, + 0.07859549671411514, + 0.07473114877939224, + 0.05164100602269173, + 0.081454336643219, + 0.04473867639899254, + 0.11810354888439178, + 0.082305908203125 + ], + "credit_map": [ + 0.012125428766012192, + 0.017029378563165665, + 0.011665656231343746, + 0.01433459110558033, + 0.021690838038921356, + 0.05188195779919624, + 0.02010408230125904, + 0.07829038053750992, + 0.007156313396990299, + 0.009276465512812138, + 0.0071724108420312405, + 0.00903917383402586, + 0.010329896584153175, + 0.04419579356908798, + 0.014953849837183952, + 0.07553748786449432, + 0.004842514172196388, + 0.007191265933215618, + 0.0043677156791090965, + 0.005002086516469717, + 0.006353746633976698, + 0.012222358956933022, + 0.007004607934504747, + 0.016661902889609337, + 0.006285307463258505, + 0.011077098548412323, + 0.004837111569941044, + 0.004776948131620884, + 0.009648307226598263, + 0.01405555009841919, + 0.007393193896859884, + 0.020521972328424454, + 0.008060800842940807, + 0.006734876893460751, + 0.007021917030215263, + 0.005323857069015503, + 0.008990027941763401, + 0.013219586573541164, + 0.007993255741894245, + 0.017386822029948235, + 0.004537472501397133, + 0.003626675810664892, + 0.0036113981623202562, + 0.0030859315302222967, + 0.004674185533076525, + 0.011012056842446327, + 0.004373703617602587, + 0.016719581559300423, + 0.010905588045716286, + 0.008253888227045536, + 0.0057562231086194515, + 0.006603450980037451, + 0.009021444246172905, + 0.018543822690844536, + 0.009696846827864647, + 0.012673071585595608, + 0.005295933689922094, + 0.00383953214623034, + 0.005273902323096991, + 0.003753918455913663, + 0.005374849308282137, + 0.007674989756196737, + 0.006815601140260696, + 0.0067099458537995815, + 0.01152421347796917, + 0.010292273946106434, + 0.010146857239305973, + 0.011163106188178062, + 0.011055142618715763, + 0.0299945417791605, + 0.012045047245919704, + 0.021882357075810432, + 0.008433368988335133, + 0.010490423068404198, + 0.009547702968120575, + 0.005825615022331476, + 0.008035261183977127, + 0.017091229557991028, + 0.011151627637445927, + 0.011730678379535675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.03310007229447365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497686192393303 + ], + "opposite_logprob_tokens": [ + -2.9299769401550293 + ], + "logprob_reward": -0.05497686192393303, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497686192393303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1490517407655716, + 0.18222501873970032, + 0.13396668434143066, + 0.09367536753416061, + 0.07919260114431381, + 0.06891370564699173, + 0.07808540016412735, + 0.055259767919778824, + 0.09681122750043869, + 0.0628185048699379 + ], + "credit_map": [ + 0.011587061919271946, + 0.02041059359908104, + 0.01222223974764347, + 0.013416999951004982, + 0.018936263397336006, + 0.02293998934328556, + 0.020627273246645927, + 0.028911320492625237, + 0.013365982100367546, + 0.018848631531000137, + 0.012354846112430096, + 0.011772774159908295, + 0.01539146900177002, + 0.0353623703122139, + 0.02144300378859043, + 0.05368594452738762, + 0.016904467716813087, + 0.026688257232308388, + 0.00924030039459467, + 0.008235795423388481, + 0.015639295801520348, + 0.01779056340456009, + 0.013755845837295055, + 0.025712165981531143, + 0.008709960617125034, + 0.008464460261166096, + 0.006971105467528105, + 0.009490237571299076, + 0.010061107575893402, + 0.01597411558032036, + 0.010802472941577435, + 0.02320191264152527, + 0.008069167844951153, + 0.012187539599835873, + 0.006813068874180317, + 0.004960585851222277, + 0.009820192120969296, + 0.011587620712816715, + 0.008954663760960102, + 0.01679975911974907, + 0.00819381419569254, + 0.006657527294009924, + 0.0048917667008936405, + 0.003484910586848855, + 0.01698831096291542, + 0.009217645041644573, + 0.0068657295778393745, + 0.012613996863365173, + 0.007597016170620918, + 0.008623764850199223, + 0.008697767741978168, + 0.007972647435963154, + 0.01260684709995985, + 0.01156618818640709, + 0.009344464167952538, + 0.011676699854433537, + 0.005092926789075136, + 0.004226616118103266, + 0.004590048920363188, + 0.004156948998570442, + 0.006259850692003965, + 0.014094688929617405, + 0.007576503790915012, + 0.009262179024517536, + 0.009898575022816658, + 0.008160904981195927, + 0.008311879821121693, + 0.008878503926098347, + 0.00997826736420393, + 0.024515481665730476, + 0.009385757148265839, + 0.01768185757100582, + 0.006022257264703512, + 0.0064310128800570965, + 0.0069739436730742455, + 0.005837338510900736, + 0.007524194661527872, + 0.011849921196699142, + 0.00835705827921629, + 0.009822782129049301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.05497686192393303 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592818036675453 + ], + "opposite_logprob_tokens": [ + -3.8955929279327393 + ], + "logprob_reward": -0.020592818036675453, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592818036675453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08058399707078934, + 0.07108383625745773, + 0.0810810849070549, + 0.04387398436665535, + 0.10161488503217697, + 0.04331763833761215, + 0.13721834123134613, + 0.19429001212120056, + 0.09465845674276352, + 0.15227781236171722 + ], + "credit_map": [ + 0.008166410028934479, + 0.007533086463809013, + 0.006853163242340088, + 0.0074743786826729774, + 0.010409411042928696, + 0.01293849851936102, + 0.008681229315698147, + 0.018527815118432045, + 0.005114816129207611, + 0.006456272676587105, + 0.007840842939913273, + 0.00485799927264452, + 0.005785709712654352, + 0.013724572025239468, + 0.007805021479725838, + 0.01949859783053398, + 0.00643335934728384, + 0.006667534355074167, + 0.005658573471009731, + 0.004639980848878622, + 0.008260552771389484, + 0.017168696969747543, + 0.009808040224015713, + 0.0224443431943655, + 0.0033466240856796503, + 0.005675590597093105, + 0.0023193180095404387, + 0.003786079352721572, + 0.0037542316131293774, + 0.007793649565428495, + 0.005778859835118055, + 0.011419630609452724, + 0.01263485662639141, + 0.009432290680706501, + 0.006893934216350317, + 0.007788053248077631, + 0.014939405024051666, + 0.017011666670441628, + 0.011271928437054157, + 0.02164275012910366, + 0.00479592802003026, + 0.003113224869593978, + 0.0031226661521941423, + 0.0025526119861751795, + 0.007398755755275488, + 0.006929508410394192, + 0.005140939261764288, + 0.010263999924063683, + 0.010809458792209625, + 0.008144418708980083, + 0.01345602236688137, + 0.011050282046198845, + 0.01345544122159481, + 0.03994983807206154, + 0.01615753024816513, + 0.0241953544318676, + 0.01562528870999813, + 0.016850827261805534, + 0.01875637285411358, + 0.017446037381887436, + 0.024926435202360153, + 0.03794423118233681, + 0.03141563758254051, + 0.031325168907642365, + 0.010503049939870834, + 0.011170091107487679, + 0.00884371530264616, + 0.007684687618166208, + 0.012417565099895, + 0.01777421124279499, + 0.011017869226634502, + 0.01524727139621973, + 0.014420985244214535, + 0.016473961994051933, + 0.015766113996505737, + 0.015233688987791538, + 0.015324505977332592, + 0.03215765953063965, + 0.016770964488387108, + 0.02612992189824581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.020592818036675453 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028246049769222736 + ], + "opposite_logprob_tokens": [ + -5.877824783325195 + ], + "logprob_reward": -0.0028246049769222736, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028246049769222736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20325876772403717, + 0.08787520974874496, + 0.10444678366184235, + 0.07214970141649246, + 0.03492600470781326, + 0.07171034812927246, + 0.08232337236404419, + 0.08499658107757568, + 0.11995843797922134, + 0.1383548080921173 + ], + "credit_map": [ + 0.01269408781081438, + 0.020111113786697388, + 0.014403961598873138, + 0.018837254494428635, + 0.019886335358023643, + 0.037922073155641556, + 0.021812012419104576, + 0.05759192630648613, + 0.0049203005619347095, + 0.006825328804552555, + 0.006567631382495165, + 0.0055695162154734135, + 0.0060745589435100555, + 0.020189279690384865, + 0.006798694841563702, + 0.03092990256845951, + 0.008245636709034443, + 0.006301198620349169, + 0.008304741233587265, + 0.004611485172063112, + 0.008991935290396214, + 0.023242246359586716, + 0.008000800386071205, + 0.0367487408220768, + 0.0072359791956841946, + 0.011831614188849926, + 0.005787150468677282, + 0.006801158655434847, + 0.009873563423752785, + 0.010428558103740215, + 0.007124448195099831, + 0.013067228719592094, + 0.004141855053603649, + 0.0034132308792322874, + 0.0026017536874860525, + 0.0016023896168917418, + 0.0030168709345161915, + 0.00640834541991353, + 0.004170155618339777, + 0.009571405127644539, + 0.00577485840767622, + 0.00731446361169219, + 0.005340988282114267, + 0.004586442839354277, + 0.015259278938174248, + 0.011053197085857391, + 0.00818928424268961, + 0.014191831462085247, + 0.009352940134704113, + 0.005692266393452883, + 0.006891340017318726, + 0.0071093118749558926, + 0.010771011002361774, + 0.018586505204439163, + 0.009453443810343742, + 0.014466555789113045, + 0.00915660709142685, + 0.008173112757503986, + 0.010075541213154793, + 0.00892034824937582, + 0.010947167873382568, + 0.012002354487776756, + 0.01239552441984415, + 0.013325924053788185, + 0.012356373481452465, + 0.01264510303735733, + 0.014458393678069115, + 0.011713080108165741, + 0.014426167123019695, + 0.020529400557279587, + 0.015614064410328865, + 0.018215859308838844, + 0.014264210127294064, + 0.019636724144220352, + 0.015822919085621834, + 0.011478258296847343, + 0.014744430780410767, + 0.022801730781793594, + 0.019827427342534065, + 0.019779110327363014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0028246049769222736 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.695033073425293 + ], + "opposite_logprob_tokens": [ + -0.07003304362297058 + ], + "logprob_reward": -2.695033073425293, + "yes_no_margin_reward": -2.625, + "token_credit_reward": -2.695033073425293, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21463781595230103, + 0.13262127339839935, + 0.05952640250325203, + 0.03589572757482529, + 0.05089495703577995, + 0.0477934293448925, + 0.15917187929153442, + 0.05237620323896408, + 0.18009519577026367, + 0.06698713451623917 + ], + "credit_map": [ + 0.013407394289970398, + 0.018609903752803802, + 0.015115588903427124, + 0.011466179974377155, + 0.014946873299777508, + 0.047962721437215805, + 0.01606602966785431, + 0.07706312090158463, + 0.004877259023487568, + 0.006748650688678026, + 0.004476059693843126, + 0.0037249901797622442, + 0.00589476665481925, + 0.036137908697128296, + 0.008947486989200115, + 0.06181412562727928, + 0.002729898551478982, + 0.0029553391505032778, + 0.0032757672015577555, + 0.0030727104749530554, + 0.0034656624775379896, + 0.014417363330721855, + 0.007031205575913191, + 0.02257845178246498, + 0.003289892105385661, + 0.0037449724040925503, + 0.0024085957556962967, + 0.0029440559446811676, + 0.003904178971424699, + 0.00594846997410059, + 0.004550816956907511, + 0.009104742668569088, + 0.006221340503543615, + 0.00690535269677639, + 0.0038129258900880814, + 0.0038266354240477085, + 0.006244337186217308, + 0.007836203090846539, + 0.006157455965876579, + 0.009890701621770859, + 0.003971579018980265, + 0.003979154862463474, + 0.0041302647441625595, + 0.0033693648874759674, + 0.005185250658541918, + 0.009925764985382557, + 0.0045437198132276535, + 0.012688325718045235, + 0.017934922128915787, + 0.014198996126651764, + 0.01269582286477089, + 0.012565797194838524, + 0.019987963140010834, + 0.034640926867723465, + 0.01988055184483528, + 0.027266893535852432, + 0.005615190602838993, + 0.004462029784917831, + 0.004674108233302832, + 0.005186380352824926, + 0.006509755738079548, + 0.009688196703791618, + 0.008264187723398209, + 0.007976345717906952, + 0.017850790172815323, + 0.012804349884390831, + 0.0134614622220397, + 0.018302731215953827, + 0.019155094400048256, + 0.043474845588207245, + 0.022700199857354164, + 0.032345712184906006, + 0.006742483004927635, + 0.004726945422589779, + 0.005362721160054207, + 0.006541866809129715, + 0.006351460237056017, + 0.01705426722764969, + 0.008114892058074474, + 0.012092490680515766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -2.695033073425293 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "stunning aerial drone shot, the camera held still, dominates the frame as a massive white tiger crouches on a rugged expanse of dark gray, almost black volcanic rock. Its thick, snow-white fur is streaked with deep black stripes, flowing like ink along the contours of its muscular body. At that moment, its massive head tilts upward, facing the camera overhead, its maw gaping to its fullest, revealing dagger-sharp canines and a curled, pink tongue, as if unleashing a silent yet earth-shattering roar. Though silent, the bulging muscles of its neck and shoulders, its eyes squeezed shut with intense exertion, and its flared nostrils convey a sense of devastating, explosive power. An intense, concentrated beam of overhead light shines vertically down, precisely outlining the undulations of the tiger's spine and the texture of every inch of its muscles, while casting deep, sharply defined shadows on the rough rocky ground. The dominant tones of the entire painting are extreme black and white, creating a primal, almost sacred sense of power and an awe-inspiring silence, imbued with unparalleled visual tension.", + "seed": 2944, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_15/index_713_seed_2944_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3013111882843077, + "vlm_logprob": -0.3013111882843077, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.3013111882843077, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.042298413813114166 + ], + "opposite_logprob_tokens": [ + -3.5422983169555664 + ], + "logprob_reward": -0.042298413813114166, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.042298413813114166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22213859856128693, + 0.18710248172283173, + 0.03473503887653351, + 0.06028946861624718, + 0.06965514272451401, + 0.13712839782238007, + 0.04737921059131622, + 0.07513932883739471, + 0.08056725561618805, + 0.08586505055427551 + ], + "credit_map": [ + 0.013204541988670826, + 0.014311089180409908, + 0.03819744661450386, + 0.01794392056763172, + 0.028920486569404602, + 0.05211899057030678, + 0.020457284525036812, + 0.036984845995903015, + 0.017459863796830177, + 0.014609947800636292, + 0.02756299078464508, + 0.014046913012862206, + 0.03090556710958481, + 0.03120257891714573, + 0.027610858902335167, + 0.02370377816259861, + 0.004023231565952301, + 0.004454637877643108, + 0.005611299071460962, + 0.002915781456977129, + 0.003385222749784589, + 0.004864758811891079, + 0.0035809907130897045, + 0.005899122916162014, + 0.00534293195232749, + 0.007146233227103949, + 0.009748700074851513, + 0.0056207794696092606, + 0.004718750715255737, + 0.009529655799269676, + 0.006581936962902546, + 0.011600482277572155, + 0.010126905515789986, + 0.010525411926209927, + 0.010568865574896336, + 0.0073859491385519505, + 0.006348296534270048, + 0.008721469901502132, + 0.007060654927045107, + 0.008917595259845257, + 0.01608276553452015, + 0.02294236421585083, + 0.018583491444587708, + 0.014619121327996254, + 0.014224675484001637, + 0.019772032275795937, + 0.012412553653120995, + 0.018491409718990326, + 0.004640236031264067, + 0.007171210832893848, + 0.0074943178333342075, + 0.004809916019439697, + 0.00628442270681262, + 0.006463600322604179, + 0.00547671876847744, + 0.005038791336119175, + 0.0032743033953011036, + 0.004144526552408934, + 0.013676000759005547, + 0.010814428329467773, + 0.005022657103836536, + 0.020348263904452324, + 0.005968580488115549, + 0.011890574358403683, + 0.005564957857131958, + 0.005731448531150818, + 0.01270004641264677, + 0.008856118656694889, + 0.007053920533508062, + 0.019325654953718185, + 0.007264017127454281, + 0.014071092940866947, + 0.004761686082929373, + 0.003302071476355195, + 0.015431653708219528, + 0.013249244540929794, + 0.00505237327888608, + 0.0242637749761343, + 0.005962107330560684, + 0.013842141255736351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.042298413813114166 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890665054321289 + ], + "opposite_logprob_tokens": [ + -2.464066505432129 + ], + "logprob_reward": -0.0890665054321289, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890665054321289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1669030487537384, + 0.31157881021499634, + 0.06390514969825745, + 0.03977135941386223, + 0.028069129213690758, + 0.03900527581572533, + 0.10032123327255249, + 0.14624559879302979, + 0.027759555727243423, + 0.07644084095954895 + ], + "credit_map": [ + 0.01009746827185154, + 0.010536257177591324, + 0.028892062604427338, + 0.013025247491896152, + 0.021353071555495262, + 0.037710487842559814, + 0.017565926536917686, + 0.027722524479031563, + 0.024626359343528748, + 0.02940233051776886, + 0.0442776083946228, + 0.02082541212439537, + 0.0500183142721653, + 0.05463015288114548, + 0.0440041720867157, + 0.043794453144073486, + 0.005225547589361668, + 0.0060959868133068085, + 0.011162965558469296, + 0.006566121242940426, + 0.004397636279463768, + 0.009925690479576588, + 0.007489623501896858, + 0.013041581027209759, + 0.003431119257584214, + 0.005266143474727869, + 0.006057111546397209, + 0.00465246569365263, + 0.0039016450755298138, + 0.005857778247445822, + 0.0037924223579466343, + 0.0068126763217151165, + 0.0017459901282563806, + 0.0039106556214392185, + 0.004352980758994818, + 0.0030002195853739977, + 0.0021353671327233315, + 0.004390268586575985, + 0.00337962550111115, + 0.005154021549969912, + 0.003958272282034159, + 0.003934144042432308, + 0.007762678433209658, + 0.003341356758028269, + 0.003363578813150525, + 0.006630228366702795, + 0.004218911752104759, + 0.005796106066554785, + 0.010325807146728039, + 0.01683550700545311, + 0.015046683140099049, + 0.009176318533718586, + 0.013031366281211376, + 0.012384089641273022, + 0.011992032639682293, + 0.011529427953064442, + 0.00838722474873066, + 0.015969200059771538, + 0.02148696966469288, + 0.01942041888833046, + 0.013835630379617214, + 0.029695406556129456, + 0.011566066183149815, + 0.025884682312607765, + 0.0024177238810807467, + 0.002266640542075038, + 0.004537756089121103, + 0.003209589282050729, + 0.0020496740471571684, + 0.006480690091848373, + 0.003365845885127783, + 0.0034316349774599075, + 0.008032668381929398, + 0.010972159914672375, + 0.012313692830502987, + 0.006895876489579678, + 0.00887871254235506, + 0.01030462421476841, + 0.010466977953910828, + 0.008576126769185066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.0890665054321289 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496569350361824 + ], + "opposite_logprob_tokens": [ + -2.9299657344818115 + ], + "logprob_reward": -0.05496569350361824, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496569350361824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2924661934375763, + 0.15443146228790283, + 0.05563786253333092, + 0.08691539615392685, + 0.043105561286211014, + 0.05315526947379112, + 0.12188456952571869, + 0.06951253116130829, + 0.07794864475727081, + 0.04494250938296318 + ], + "credit_map": [ + 0.013880528509616852, + 0.021075118333101273, + 0.050207678228616714, + 0.02229878306388855, + 0.030629005283117294, + 0.07078461349010468, + 0.03194925934076309, + 0.05164123699069023, + 0.015599826350808144, + 0.012387732043862343, + 0.02149893343448639, + 0.011809193529188633, + 0.022803543135523796, + 0.02686295285820961, + 0.02140962705016136, + 0.02205965295433998, + 0.007221464067697525, + 0.005131070502102375, + 0.008167613297700882, + 0.005744661204516888, + 0.003863105084747076, + 0.00971557479351759, + 0.004859714303165674, + 0.010934661142528057, + 0.005477385129779577, + 0.01007538940757513, + 0.014731461182236671, + 0.007960415445268154, + 0.0065063913352787495, + 0.017006907612085342, + 0.009357097558677197, + 0.0158003568649292, + 0.002564861439168453, + 0.0057481927797198296, + 0.006604495923966169, + 0.004095820244401693, + 0.0029845910612493753, + 0.008466794155538082, + 0.005055893678218126, + 0.0075849127024412155, + 0.005431584548205137, + 0.006428997032344341, + 0.008412712253630161, + 0.00378827890381217, + 0.004123243503272533, + 0.009971396066248417, + 0.006716831587255001, + 0.00828222744166851, + 0.009341665543615818, + 0.014592050574719906, + 0.01975741609930992, + 0.0119782080873847, + 0.014157023280858994, + 0.020870361477136612, + 0.014148286543786526, + 0.017039554193615913, + 0.005776315927505493, + 0.008835975080728531, + 0.010578655637800694, + 0.005868804175406694, + 0.008177542127668858, + 0.015012197196483612, + 0.006666637025773525, + 0.008596409112215042, + 0.0052977055311203, + 0.006121547892689705, + 0.01238737627863884, + 0.00879973266273737, + 0.006152838934212923, + 0.018849968910217285, + 0.00800967775285244, + 0.012329798191785812, + 0.002943087834864855, + 0.004420265089720488, + 0.007293631322681904, + 0.0033105071634054184, + 0.00489450478926301, + 0.009159996174275875, + 0.00432030763477087, + 0.008600208908319473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.05496569350361824 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067449514754116535 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.0067449514754116535, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067449514754116535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24863168597221375, + 0.10816659033298492, + 0.07355539500713348, + 0.08847033232450485, + 0.07760468870401382, + 0.09194714576005936, + 0.08176825195550919, + 0.06712193042039871, + 0.06184364855289459, + 0.10089033842086792 + ], + "credit_map": [ + 0.01497930008918047, + 0.01913280040025711, + 0.0415772944688797, + 0.02087595872581005, + 0.029125623404979706, + 0.05668143555521965, + 0.02386341243982315, + 0.042395856231451035, + 0.00900504644960165, + 0.008788105100393295, + 0.01516794040799141, + 0.008482011035084724, + 0.017407651990652084, + 0.020491937175393105, + 0.014345026575028896, + 0.014478874392807484, + 0.007040803320705891, + 0.00611190777271986, + 0.011654569767415524, + 0.0076638017781078815, + 0.0068685151636600494, + 0.01302316039800644, + 0.008709337562322617, + 0.01248329970985651, + 0.005017715040594339, + 0.005778771359473467, + 0.015618269331753254, + 0.009116006083786488, + 0.007038618437945843, + 0.016660913825035095, + 0.012266608886420727, + 0.016973428428173065, + 0.0060104248113930225, + 0.006716412026435137, + 0.012565653771162033, + 0.009495118632912636, + 0.005770050920546055, + 0.01610633172094822, + 0.009313074871897697, + 0.0116276191547513, + 0.007004824001342058, + 0.011628739535808563, + 0.014943446032702923, + 0.011457596905529499, + 0.0060082366690039635, + 0.019878700375556946, + 0.00817056279629469, + 0.012855040840804577, + 0.00647056894376874, + 0.005171309690922499, + 0.013847832567989826, + 0.009377939626574516, + 0.00606788694858551, + 0.017958778887987137, + 0.008682076819241047, + 0.014191851019859314, + 0.0032076467759907246, + 0.004663069732487202, + 0.011806088499724865, + 0.0069560944102704525, + 0.004726074170321226, + 0.02058989368379116, + 0.007601040881127119, + 0.007572030648589134, + 0.004621563479304314, + 0.005991695448756218, + 0.010820402763783932, + 0.006464304868131876, + 0.005451734643429518, + 0.01521280501037836, + 0.005319961812347174, + 0.007961180061101913, + 0.004659495782107115, + 0.00699731195345521, + 0.018124466761946678, + 0.013663262128829956, + 0.007014256902039051, + 0.026797715574502945, + 0.0076911007054150105, + 0.015942728146910667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0067449514754116535 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive white tiger crouching on dark gray or black volcanic rock?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tiger's head tilted upward facing the camera with its mouth open, revealing sharp canines and a pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there an intense beam of overhead light shining vertically down that outlines the tiger's spine and muscles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the dominant tones of the scene extreme black and white with deep, sharply defined shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still in an aerial drone shot that dominates the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134803771972656 + ], + "opposite_logprob_tokens": [ + -0.3134803771972656 + ], + "logprob_reward": -1.3134803771972656, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134803771972656, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3486904203891754, + 0.07500835508108139, + 0.07859597355127335, + 0.0484679713845253, + 0.0692773163318634, + 0.04388989135622978, + 0.16469678282737732, + 0.03673558309674263, + 0.07989417016506195, + 0.05474357679486275 + ], + "credit_map": [ + 0.028921639546751976, + 0.021089617162942886, + 0.05609578639268875, + 0.025502033531665802, + 0.02514474093914032, + 0.08848348259925842, + 0.035689570009708405, + 0.06776352971792221, + 0.004258061293512583, + 0.005366835277527571, + 0.011816272512078285, + 0.006585256662219763, + 0.010129469446837902, + 0.015182551927864552, + 0.006620082538574934, + 0.015049817971885204, + 0.005690116435289383, + 0.006227102596312761, + 0.013657773844897747, + 0.006361232604831457, + 0.006330532021820545, + 0.016969412565231323, + 0.008132882416248322, + 0.015226906165480614, + 0.0033009443432092667, + 0.0035087571013718843, + 0.008072741329669952, + 0.002884883899241686, + 0.004300413653254509, + 0.01121245976537466, + 0.003789311507716775, + 0.011398456990718842, + 0.0041062720119953156, + 0.0049788155592978, + 0.01138631533831358, + 0.0069250124506652355, + 0.007126742508262396, + 0.01564907468855381, + 0.007492023520171642, + 0.011613057926297188, + 0.005621737334877253, + 0.006171143148094416, + 0.006238470785319805, + 0.003654194064438343, + 0.00464505422860384, + 0.00797666609287262, + 0.0033722994849085808, + 0.006210324354469776, + 0.014138856902718544, + 0.01600891537964344, + 0.02577342465519905, + 0.0159517303109169, + 0.016428040340542793, + 0.03428163006901741, + 0.015642384067177773, + 0.026471775025129318, + 0.003004770027473569, + 0.0024480256251990795, + 0.00606017978861928, + 0.005148249678313732, + 0.0023419801145792007, + 0.00809171050786972, + 0.0029224043246358633, + 0.006718259770423174, + 0.007352456916123629, + 0.006954059470444918, + 0.012558769434690475, + 0.0075027430430054665, + 0.006023528520017862, + 0.017496256157755852, + 0.00606898870319128, + 0.015937354415655136, + 0.0040639545768499374, + 0.0038181915879249573, + 0.0087600601837039, + 0.004297028761357069, + 0.004569645505398512, + 0.01127520203590393, + 0.005476463586091995, + 0.012483026832342148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -1.3134803771972656 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_2/index_tensor([9527, 9527])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_2/index_tensor([9527, 9527])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..50dd726e8c4df246cb06cd788e31728aef611973 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_2/index_tensor([9527, 9527])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up, the camera's slow, smooth orbiting motion orbits a high-precision e-reader suspended in a pure, endless void of darkness. The device performs complex, deliberate multi-axis flips in a physically precise gyroscope. Its ultra-thin body, crafted from matte, dark graphite magnesium alloy, reflects a subtle, refined quality under a single, soft, overhead light. As the device slowly rotates, this light creates crisp highlights along its sharp edges while casting soft shadows against the deep background. Its ultra-narrow, bezel-less E-Ink screen displays a solitary cover design: a richly detailed, withered tree stands alone on a barren hill, beneath a pale, empty sky rendered in delicate grayscale tones. As the e-reader rolls steadily and gracefully, the camera remains focused on the non-luminous, paper-like screen, capturing every moment of light flowing through the cover pattern. The entire scene is shrouded in a quiet, focused, minimalist atmosphere with a technological aesthetic, emphasizing the object itself and the lonely story it carries.", + "seed": 23855, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_2/index_9527_seed_23855_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.046999120712280275, + "vlm_logprob": -0.046999120712280275, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.046999120712280275, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0505407340824604 + ], + "opposite_logprob_tokens": [ + -3.3005406856536865 + ], + "logprob_reward": -0.0505407340824604, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0505407340824604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24660250544548035, + 0.051652807742357254, + 0.041565731167793274, + 0.20360328257083893, + 0.10513089597225189, + 0.04712167754769325, + 0.04154202714562416, + 0.08706242591142654, + 0.09981781989336014, + 0.0759008377790451 + ], + "credit_map": [ + 0.03058481216430664, + 0.01028758566826582, + 0.06865718960762024, + 0.02593945525586605, + 0.014209463261067867, + 0.058495983481407166, + 0.016572140157222748, + 0.02185584418475628, + 0.005719535052776337, + 0.0025333112571388483, + 0.012127049267292023, + 0.007687856908887625, + 0.0051396987400949, + 0.007222646381705999, + 0.0056095304898917675, + 0.005613177083432674, + 0.004342266824096441, + 0.003006214741617441, + 0.010431763716042042, + 0.006159469019621611, + 0.003176084952428937, + 0.007836227305233479, + 0.00354880397208035, + 0.003064897609874606, + 0.02584543637931347, + 0.005174805875867605, + 0.05158621072769165, + 0.006362667307257652, + 0.0041849659755826, + 0.06268391758203506, + 0.009897541254758835, + 0.0378677137196064, + 0.01701626181602478, + 0.00421513058245182, + 0.011794091202318668, + 0.0044691301882267, + 0.006059725768864155, + 0.028537141159176826, + 0.006882787216454744, + 0.026156624779105186, + 0.006785462144762278, + 0.006476833485066891, + 0.004074008669704199, + 0.0070978025905787945, + 0.003365529468283057, + 0.006858143489807844, + 0.006586086004972458, + 0.005877804011106491, + 0.006971156224608421, + 0.0036154724657535553, + 0.00568361533805728, + 0.005754431243985891, + 0.002512324834242463, + 0.0058196173049509525, + 0.005764991044998169, + 0.0054204161278903484, + 0.008453941904008389, + 0.010674319230020046, + 0.01120336726307869, + 0.011725331656634808, + 0.006241423077881336, + 0.012954809702932835, + 0.0165559109300375, + 0.009253312833607197, + 0.007686039432883263, + 0.011417746543884277, + 0.011790960095822811, + 0.010965481400489807, + 0.009540660306811333, + 0.018666982650756836, + 0.012154935859143734, + 0.017595000565052032, + 0.006738267373293638, + 0.008216622285544872, + 0.009687047451734543, + 0.00715015409514308, + 0.007059414871037006, + 0.01432991772890091, + 0.008357970044016838, + 0.014361442066729069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0505407340824604 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08909922093153 + ], + "opposite_logprob_tokens": [ + -2.464099168777466 + ], + "logprob_reward": -0.08909922093153, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08909922093153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2392292022705078, + 0.10575026273727417, + 0.056141193956136703, + 0.11071030795574188, + 0.15398114919662476, + 0.07005727291107178, + 0.03081616200506687, + 0.07767876982688904, + 0.1172456368803978, + 0.03839002922177315 + ], + "credit_map": [ + 0.026039909571409225, + 0.015725400298833847, + 0.045691415667533875, + 0.04387354478240013, + 0.021414954215288162, + 0.034050971269607544, + 0.028690222650766373, + 0.023742791265249252, + 0.014409148134291172, + 0.005289183929562569, + 0.02499138005077839, + 0.012421582825481892, + 0.007914411835372448, + 0.01715771295130253, + 0.011793979443609715, + 0.01177286822348833, + 0.006293654907494783, + 0.0048360456712543964, + 0.00859962310642004, + 0.008157908916473389, + 0.00582780921831727, + 0.009876674972474575, + 0.00597601430490613, + 0.006573464721441269, + 0.01123522687703371, + 0.006764295045286417, + 0.025960654020309448, + 0.008391405455768108, + 0.00594610208645463, + 0.029101649299263954, + 0.008247793652117252, + 0.015063178725540638, + 0.028778955340385437, + 0.00657067634165287, + 0.021324459463357925, + 0.00875916238874197, + 0.010110786184668541, + 0.03525211662054062, + 0.010359710082411766, + 0.0328252799808979, + 0.01479834970086813, + 0.005152761936187744, + 0.006490772124379873, + 0.00833375658839941, + 0.005337364040315151, + 0.011189810000360012, + 0.007243416737765074, + 0.011511041782796383, + 0.006404072977602482, + 0.003620176576077938, + 0.0033410033211112022, + 0.004186235833913088, + 0.0023970992770045996, + 0.004067859146744013, + 0.0033482543658465147, + 0.003451460739597678, + 0.007899951189756393, + 0.011565442197024822, + 0.008613861165940762, + 0.010973337106406689, + 0.01016632653772831, + 0.00972425565123558, + 0.010774324648082256, + 0.007961263880133629, + 0.009057658724486828, + 0.015670735388994217, + 0.015405317768454552, + 0.012998344376683235, + 0.012340910732746124, + 0.02086801826953888, + 0.013242295011878014, + 0.017662351951003075, + 0.0039543756283819675, + 0.0038635742384940386, + 0.0049107312224805355, + 0.005297541618347168, + 0.0028397925198078156, + 0.006789733190089464, + 0.005487123969942331, + 0.005247159395366907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.08909922093153 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043067168444395065 + ], + "opposite_logprob_tokens": [ + -3.168067216873169 + ], + "logprob_reward": -0.043067168444395065, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067168444395065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1406061053276062, + 0.13687719404697418, + 0.08181773126125336, + 0.4164051413536072, + 0.042645372450351715, + 0.0263160839676857, + 0.059820305556058884, + 0.016101108863949776, + 0.04705299064517021, + 0.03235797956585884 + ], + "credit_map": [ + 0.018356122076511383, + 0.00869643036276102, + 0.021796977147459984, + 0.030357815325260162, + 0.012159655801951885, + 0.01691039651632309, + 0.016345441341400146, + 0.015983259305357933, + 0.017437053844332695, + 0.011425589211285114, + 0.021264296025037766, + 0.022780021652579308, + 0.014228442683815956, + 0.013749646954238415, + 0.021442709490656853, + 0.01454942487180233, + 0.006509967148303986, + 0.008118481375277042, + 0.01741790771484375, + 0.012037447653710842, + 0.008696860633790493, + 0.012501934543251991, + 0.01002995390444994, + 0.006505179684609175, + 0.035732388496398926, + 0.01473103929311037, + 0.11335673928260803, + 0.019588839262723923, + 0.01503300853073597, + 0.1402023285627365, + 0.02194429375231266, + 0.05581649765372276, + 0.004939485341310501, + 0.004619898274540901, + 0.00532791530713439, + 0.004682878963649273, + 0.003975262399762869, + 0.008923213928937912, + 0.0036248723044991493, + 0.006551844533532858, + 0.004668615758419037, + 0.0034206591080874205, + 0.0024245157837867737, + 0.0035374853760004044, + 0.002685998333618045, + 0.0035810654517263174, + 0.002858650404959917, + 0.0031390958465635777, + 0.008571245707571507, + 0.008438319899141788, + 0.0054360488429665565, + 0.008621345274150372, + 0.00512714171782136, + 0.007873802445828915, + 0.00806799903512001, + 0.007684407290071249, + 0.0021571465767920017, + 0.0016698085237294436, + 0.0019809850491583347, + 0.0026584069710224867, + 0.001500640413723886, + 0.0020632073283195496, + 0.0023732322733849287, + 0.0016976805636659265, + 0.005365523509681225, + 0.005359409376978874, + 0.007403976283967495, + 0.006266851909458637, + 0.005284968297928572, + 0.006564540322870016, + 0.005374494474381208, + 0.005433227401226759, + 0.0037515682633966208, + 0.003804235253483057, + 0.0045133428648114204, + 0.0038429955020546913, + 0.0029793966095894575, + 0.004613033961504698, + 0.0035462374798953533, + 0.005307170562446117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.043067168444395065 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200706034898758 + ], + "opposite_logprob_tokens": [ + -4.264200687408447 + ], + "logprob_reward": -0.014200706034898758, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200706034898758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3318651616573334, + 0.04841063171625137, + 0.0778498500585556, + 0.1895483434200287, + 0.08637338876724243, + 0.051426757127046585, + 0.09994104504585266, + 0.04926034435629845, + 0.013533484190702438, + 0.05179101228713989 + ], + "credit_map": [ + 0.041686221957206726, + 0.01094499509781599, + 0.08830944448709488, + 0.03937912359833717, + 0.01601775735616684, + 0.08758307248353958, + 0.02316945232450962, + 0.024775102734565735, + 0.005964184179902077, + 0.0020798579789698124, + 0.008206416852772236, + 0.012156701646745205, + 0.004964139778167009, + 0.005931512452661991, + 0.005340001545846462, + 0.0037678161170333624, + 0.00894832145422697, + 0.008419450372457504, + 0.01282971166074276, + 0.014491555280983448, + 0.009951806627213955, + 0.007064930163323879, + 0.009089551866054535, + 0.007054523564875126, + 0.018076475709676743, + 0.00683588394895196, + 0.043962541967630386, + 0.006438559852540493, + 0.004821769427508116, + 0.06477084010839462, + 0.009394392371177673, + 0.03524787724018097, + 0.015090406872332096, + 0.008841169066727161, + 0.00995065551251173, + 0.00693456782028079, + 0.007835743017494678, + 0.01661943458020687, + 0.007094309665262699, + 0.014007099904119968, + 0.009228828363120556, + 0.0037957464810460806, + 0.0034938007593154907, + 0.005971807986497879, + 0.0041127074509859085, + 0.008887996897101402, + 0.007873726077377796, + 0.008062143810093403, + 0.012649141252040863, + 0.014318442903459072, + 0.008828149177134037, + 0.01441903691738844, + 0.009121039882302284, + 0.013763600029051304, + 0.01456325501203537, + 0.012278378941118717, + 0.004995718598365784, + 0.007008380722254515, + 0.006797207519412041, + 0.006126358173787594, + 0.004243034403771162, + 0.007022082339972258, + 0.0077828820794820786, + 0.005284680984914303, + 0.0010933057637885213, + 0.0015711549203842878, + 0.0020718369632959366, + 0.0015812655910849571, + 0.001357348752208054, + 0.0027717500925064087, + 0.0009661800577305257, + 0.0021206424571573734, + 0.0051813554018735886, + 0.00708064716309309, + 0.007270244415849447, + 0.005812923889607191, + 0.006183325313031673, + 0.006999506615102291, + 0.007936724461615086, + 0.0053262850269675255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.014200706034898758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808777406811714 + ], + "opposite_logprob_tokens": [ + -3.288087844848633 + ], + "logprob_reward": -0.03808777406811714, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808777406811714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14652368426322937, + 0.07191359251737595, + 0.09768030047416687, + 0.2925218343734741, + 0.14141424000263214, + 0.06253056973218918, + 0.05805151164531708, + 0.056566670536994934, + 0.01674136519432068, + 0.05605621635913849 + ], + "credit_map": [ + 0.03085862472653389, + 0.01015842892229557, + 0.023805886507034302, + 0.02157464064657688, + 0.010758491232991219, + 0.019520897418260574, + 0.01431009080260992, + 0.01553662121295929, + 0.009997429326176643, + 0.004689873196184635, + 0.012236393988132477, + 0.01408161036670208, + 0.007764451205730438, + 0.009210164658725262, + 0.00795304961502552, + 0.005980619229376316, + 0.009850563481450081, + 0.010128799825906754, + 0.01935487613081932, + 0.01476765051484108, + 0.0090620256960392, + 0.01298490259796381, + 0.00958434958010912, + 0.01194713357836008, + 0.03322497382760048, + 0.008913032710552216, + 0.06376329064369202, + 0.0126958591863513, + 0.007254862226545811, + 0.08648408204317093, + 0.016665926203131676, + 0.06351981312036514, + 0.018326735123991966, + 0.00799897126853466, + 0.012272139079868793, + 0.007617721799761057, + 0.007713370956480503, + 0.04188351333141327, + 0.00721409497782588, + 0.038387689739465714, + 0.010177611373364925, + 0.0066125462763011456, + 0.005000760778784752, + 0.009038537740707397, + 0.00569159397855401, + 0.009214576333761215, + 0.010915493592619896, + 0.005879450123757124, + 0.008885569870471954, + 0.006422462873160839, + 0.007817077450454235, + 0.008697395212948322, + 0.004073431249707937, + 0.00830149371176958, + 0.007380231283605099, + 0.006473848596215248, + 0.00564988749101758, + 0.005694637540727854, + 0.006864266935735941, + 0.007237918674945831, + 0.006056518759578466, + 0.008764044381678104, + 0.010538412258028984, + 0.005760984029620886, + 0.002532372949644923, + 0.001601748401299119, + 0.0026798914186656475, + 0.0013772485544905066, + 0.0012324390700086951, + 0.003200736828148365, + 0.0014951848424971104, + 0.0026217426639050245, + 0.005220360588282347, + 0.007720149587839842, + 0.009574036113917828, + 0.005468586925417185, + 0.006648655980825424, + 0.009297608397901058, + 0.006055875681340694, + 0.006070941686630249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.03808777406811714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, the camera's slow, smooth orbiting motion orbits a high-precision e-reader suspended in a pure, endless void of darkness. The device performs complex, deliberate multi-axis flips in a physically precise gyroscope. Its ultra-thin body, crafted from matte, dark graphite magnesium alloy, reflects a subtle, refined quality under a single, soft, overhead light. As the device slowly rotates, this light creates crisp highlights along its sharp edges while casting soft shadows against the deep background. Its ultra-narrow, bezel-less E-Ink screen displays a solitary cover design: a richly detailed, withered tree stands alone on a barren hill, beneath a pale, empty sky rendered in delicate grayscale tones. As the e-reader rolls steadily and gracefully, the camera remains focused on the non-luminous, paper-like screen, capturing every moment of light flowing through the cover pattern. The entire scene is shrouded in a quiet, focused, minimalist atmosphere with a technological aesthetic, emphasizing the object itself and the lonely story it carries.", + "seed": 23856, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_2/index_9527_seed_23856_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09659745767712594, + "vlm_logprob": -0.09659745767712594, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.09659745767712594, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05837377533316612 + ], + "opposite_logprob_tokens": [ + -3.0583736896514893 + ], + "logprob_reward": -0.05837377533316612, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05837377533316612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3282276690006256, + 0.042097873985767365, + 0.048398345708847046, + 0.03206798434257507, + 0.07147392630577087, + 0.05971319228410721, + 0.06781752407550812, + 0.15940499305725098, + 0.10590101778507233, + 0.0848974660038948 + ], + "credit_map": [ + 0.03595902770757675, + 0.03115512989461422, + 0.047471124678850174, + 0.024643393233418465, + 0.023479383438825607, + 0.07302873581647873, + 0.04553810507059097, + 0.046952761709690094, + 0.005977321416139603, + 0.0037974021397531033, + 0.004747788887470961, + 0.003960760775953531, + 0.00473286397755146, + 0.006294931750744581, + 0.005736878607422113, + 0.006849929690361023, + 0.005312982946634293, + 0.0024572040420025587, + 0.006129968445748091, + 0.0023732890840619802, + 0.002553769154474139, + 0.013247801922261715, + 0.004170237574726343, + 0.012153089977800846, + 0.004480992443859577, + 0.0031524645164608955, + 0.005241552367806435, + 0.003096575615927577, + 0.002055361634120345, + 0.004936687648296356, + 0.004025725647807121, + 0.005078625399619341, + 0.007306321058422327, + 0.007350513711571693, + 0.013001509010791779, + 0.011798239313066006, + 0.006924652960151434, + 0.007478709332644939, + 0.006474447902292013, + 0.011139538139104843, + 0.005527268163859844, + 0.006013341713696718, + 0.007629964966326952, + 0.007839451543986797, + 0.008967986330389977, + 0.006720308214426041, + 0.007578250020742416, + 0.00943661667406559, + 0.004133520182222128, + 0.009238922037184238, + 0.009936662390828133, + 0.01159072294831276, + 0.005517346318811178, + 0.01429906114935875, + 0.006268608383834362, + 0.006832679733633995, + 0.02922665700316429, + 0.017270172014832497, + 0.025093119591474533, + 0.014955599792301655, + 0.013944215141236782, + 0.023799994960427284, + 0.013326389715075493, + 0.02178884483873844, + 0.008944547735154629, + 0.009584191255271435, + 0.007177775725722313, + 0.012408477254211903, + 0.011883716098964214, + 0.018892209976911545, + 0.019919654354453087, + 0.017090443521738052, + 0.007931213825941086, + 0.006177860777825117, + 0.006742601282894611, + 0.008213469758629799, + 0.006110898684710264, + 0.020476430654525757, + 0.009686548262834549, + 0.019558440893888474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.05837377533316612 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12719428539276123 + ], + "opposite_logprob_tokens": [ + -2.127194404602051 + ], + "logprob_reward": -0.12719428539276123, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12719428539276123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2453642040491104, + 0.06339273601770401, + 0.09299775958061218, + 0.04151545464992523, + 0.11766722798347473, + 0.08769088238477707, + 0.028397060930728912, + 0.06954162567853928, + 0.06377431005239487, + 0.1896587312221527 + ], + "credit_map": [ + 0.030399128794670105, + 0.020270295441150665, + 0.03264136612415314, + 0.019552238285541534, + 0.015490308403968811, + 0.05481449142098427, + 0.02961435355246067, + 0.04258202761411667, + 0.008833982050418854, + 0.005472531076520681, + 0.008695967495441437, + 0.00525682745501399, + 0.0050812470726668835, + 0.00907666515558958, + 0.011952570639550686, + 0.009022949263453484, + 0.010699382051825523, + 0.0031651563476771116, + 0.005755286663770676, + 0.003652123501524329, + 0.002322572749108076, + 0.033145081251859665, + 0.004641187842935324, + 0.02961697429418564, + 0.004020635969936848, + 0.003542723599821329, + 0.0072562601417303085, + 0.004180419258773327, + 0.002576245693489909, + 0.007124852389097214, + 0.005725758150219917, + 0.0070885601453483105, + 0.011572333984076977, + 0.012236874550580978, + 0.022462790831923485, + 0.019840804859995842, + 0.010974102653563023, + 0.012123185209929943, + 0.01096352655440569, + 0.017493611201643944, + 0.00851859524846077, + 0.0080402921885252, + 0.012087120674550533, + 0.009766857139766216, + 0.011584349907934666, + 0.010621917434036732, + 0.008877611719071865, + 0.018194135278463364, + 0.002381587401032448, + 0.0035599807742983103, + 0.0037034405395388603, + 0.004962770733982325, + 0.002438891911879182, + 0.004737740382552147, + 0.003438113257288933, + 0.0031745340675115585, + 0.011328364722430706, + 0.006084348540753126, + 0.01012100838124752, + 0.0069656167179346085, + 0.006075400859117508, + 0.010355024598538876, + 0.008876853622496128, + 0.009735009633004665, + 0.005299705080688, + 0.008821716532111168, + 0.005675071384757757, + 0.011753701604902744, + 0.006397636607289314, + 0.008594815619289875, + 0.00978587381541729, + 0.007445790804922581, + 0.019785072654485703, + 0.009013686329126358, + 0.009622425772249699, + 0.012526487931609154, + 0.007738793734461069, + 0.054285380989313126, + 0.018971944227814674, + 0.05771492421627045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.12719428539276123 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270611733198166 + ], + "opposite_logprob_tokens": [ + -2.127061128616333 + ], + "logprob_reward": -0.1270611733198166, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270611733198166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1744452863931656, + 0.0989895835518837, + 0.09199101477861404, + 0.05245218798518181, + 0.09506787359714508, + 0.06728794425725937, + 0.1429530382156372, + 0.14864055812358856, + 0.09169808775186539, + 0.03647439554333687 + ], + "credit_map": [ + 0.022108444944024086, + 0.017073990777134895, + 0.028555626049637794, + 0.015137651935219765, + 0.015261893160641193, + 0.03215666487812996, + 0.02257600799202919, + 0.021575013175606728, + 0.014064817689359188, + 0.008685793727636337, + 0.007410269696265459, + 0.008799966424703598, + 0.010279042646288872, + 0.014899042434990406, + 0.013684659264981747, + 0.021165991201996803, + 0.011284569278359413, + 0.006340139079838991, + 0.018687918782234192, + 0.008095440454781055, + 0.007842537946999073, + 0.01614801585674286, + 0.009921137243509293, + 0.013671251945197582, + 0.0067304284311831, + 0.0069383964873850346, + 0.008197858929634094, + 0.005430908873677254, + 0.0035676658153533936, + 0.007839595898985863, + 0.005808431189507246, + 0.007938902825117111, + 0.00949636846780777, + 0.009676013141870499, + 0.01675531081855297, + 0.01724330335855484, + 0.009602823294699192, + 0.009967264719307423, + 0.010570697486400604, + 0.011756096966564655, + 0.00686879176646471, + 0.005404427647590637, + 0.012717826291918755, + 0.009222266264259815, + 0.0049110837280750275, + 0.011195619590580463, + 0.007877977564930916, + 0.009089954197406769, + 0.01006283424794674, + 0.018944159150123596, + 0.02208477444946766, + 0.019411804154515266, + 0.011498315259814262, + 0.03322243317961693, + 0.012647242285311222, + 0.015081467106938362, + 0.022346241399645805, + 0.017496952787041664, + 0.019770368933677673, + 0.015072430483996868, + 0.01827588491141796, + 0.02127734012901783, + 0.013812072575092316, + 0.020589273422956467, + 0.005767103284597397, + 0.007394309621304274, + 0.007402787916362286, + 0.009648957289755344, + 0.010664206929504871, + 0.017996085807681084, + 0.015025070868432522, + 0.017799563705921173, + 0.003516672644764185, + 0.0021047794725745916, + 0.0022105902899056673, + 0.002917403820902109, + 0.0017323448555544019, + 0.00997983943670988, + 0.0034447037614881992, + 0.010568059980869293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.1270611733198166 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005926966667175 + ], + "opposite_logprob_tokens": [ + -2.695059299468994 + ], + "logprob_reward": -0.07005926966667175, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005926966667175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42402005195617676, + 0.06843271106481552, + 0.07887960970401764, + 0.03789977729320526, + 0.03315732628107071, + 0.05427861213684082, + 0.039130114018917084, + 0.10128014534711838, + 0.10154354572296143, + 0.06137806177139282 + ], + "credit_map": [ + 0.04097574204206467, + 0.04247022420167923, + 0.04882471263408661, + 0.030382946133613586, + 0.03469409793615341, + 0.09791747480630875, + 0.05745505541563034, + 0.07129982113838196, + 0.0090034119784832, + 0.005008238833397627, + 0.006647392641752958, + 0.006387313362210989, + 0.006004699971526861, + 0.012766621075570583, + 0.006444369442760944, + 0.016170667484402657, + 0.011095857247710228, + 0.003861317876726389, + 0.010081079788506031, + 0.007260337006300688, + 0.005286431405693293, + 0.018761228770017624, + 0.006194621790200472, + 0.016338737681508064, + 0.004372176248580217, + 0.0038902985397726297, + 0.005769943352788687, + 0.004812182858586311, + 0.0023807804100215435, + 0.006184399127960205, + 0.004578862804919481, + 0.005911138374358416, + 0.003305708058178425, + 0.0035991761833429337, + 0.0061437711119651794, + 0.004527300130575895, + 0.0027694175951182842, + 0.004787355661392212, + 0.0039032502099871635, + 0.004121348261833191, + 0.005846015177667141, + 0.003140670945867896, + 0.006205439567565918, + 0.004951357841491699, + 0.004733594134449959, + 0.007142622023820877, + 0.007642167620360851, + 0.014616752043366432, + 0.0022562879603356123, + 0.008421885780990124, + 0.005258725490421057, + 0.00790414772927761, + 0.0031122644431889057, + 0.0044561950489878654, + 0.004787127021700144, + 0.002933480776846409, + 0.012432560324668884, + 0.011085166595876217, + 0.011596869677305222, + 0.012955393642187119, + 0.009774102829396725, + 0.014829698018729687, + 0.012222463265061378, + 0.01638389751315117, + 0.0069192806258797646, + 0.009781202301383018, + 0.007624547462910414, + 0.01110110618174076, + 0.008761661127209663, + 0.018973451107740402, + 0.020459411665797234, + 0.017922891303896904, + 0.00620585260912776, + 0.0052122329361736774, + 0.005141841247677803, + 0.007221665699034929, + 0.00601307163015008, + 0.011626280844211578, + 0.00906130950897932, + 0.01089581660926342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.07005926966667175 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029878467321396 + ], + "opposite_logprob_tokens": [ + -2.3502988815307617 + ], + "logprob_reward": -0.10029878467321396, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029878467321396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21345756947994232, + 0.0963318794965744, + 0.10343970358371735, + 0.10511790961027145, + 0.1320899873971939, + 0.11520661413669586, + 0.0678388923406601, + 0.0508456751704216, + 0.06469140946865082, + 0.050980400294065475 + ], + "credit_map": [ + 0.02163011208176613, + 0.017728226259350777, + 0.03239762783050537, + 0.016161169856786728, + 0.016998128965497017, + 0.04231621325016022, + 0.03601837530732155, + 0.030207710340619087, + 0.011434141546487808, + 0.009326670318841934, + 0.009660281240940094, + 0.01567797362804413, + 0.009026641957461834, + 0.01054415199905634, + 0.01656007207930088, + 0.014101941138505936, + 0.013842405751347542, + 0.006034483201801777, + 0.014029494486749172, + 0.010658849962055683, + 0.008892139419913292, + 0.018018165603280067, + 0.016331080347299576, + 0.01563308574259281, + 0.013979324139654636, + 0.011902056634426117, + 0.021908245980739594, + 0.008738173171877861, + 0.007102502975612879, + 0.014000153169035912, + 0.011997040361166, + 0.01549040898680687, + 0.012685473076999187, + 0.013259472325444221, + 0.01730763353407383, + 0.022205451503396034, + 0.012238716706633568, + 0.02063572034239769, + 0.014406207948923111, + 0.019351301714777946, + 0.009279982186853886, + 0.010250057093799114, + 0.014334657229483128, + 0.0194245558232069, + 0.009580238722264767, + 0.012441018596291542, + 0.02107580378651619, + 0.01882029138505459, + 0.004059681668877602, + 0.008236008696258068, + 0.009970328770577908, + 0.016493724659085274, + 0.007304014638066292, + 0.007963788695633411, + 0.008228776045143604, + 0.005582564044743776, + 0.00919113028794527, + 0.0043941643089056015, + 0.011170661076903343, + 0.004739874042570591, + 0.004685739520937204, + 0.0065653640776872635, + 0.004235752858221531, + 0.005862981081008911, + 0.003742511384189129, + 0.005772626493126154, + 0.006103276740759611, + 0.007763960398733616, + 0.006300501991063356, + 0.010797002352774143, + 0.01399141363799572, + 0.010220116004347801, + 0.005813558120280504, + 0.00393499992787838, + 0.003405018011108041, + 0.004803098272532225, + 0.0036811705213040113, + 0.011045103892683983, + 0.006020647939294577, + 0.01227679755538702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.10029878467321396 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, the camera's slow, smooth orbiting motion orbits a high-precision e-reader suspended in a pure, endless void of darkness. The device performs complex, deliberate multi-axis flips in a physically precise gyroscope. Its ultra-thin body, crafted from matte, dark graphite magnesium alloy, reflects a subtle, refined quality under a single, soft, overhead light. As the device slowly rotates, this light creates crisp highlights along its sharp edges while casting soft shadows against the deep background. Its ultra-narrow, bezel-less E-Ink screen displays a solitary cover design: a richly detailed, withered tree stands alone on a barren hill, beneath a pale, empty sky rendered in delicate grayscale tones. As the e-reader rolls steadily and gracefully, the camera remains focused on the non-luminous, paper-like screen, capturing every moment of light flowing through the cover pattern. The entire scene is shrouded in a quiet, focused, minimalist atmosphere with a technological aesthetic, emphasizing the object itself and the lonely story it carries.", + "seed": 23857, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_2/index_9527_seed_23857_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06943440809845924, + "vlm_logprob": -0.06943440809845924, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.06943440809845924, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.051611170172691345 + ], + "opposite_logprob_tokens": [ + -3.1766111850738525 + ], + "logprob_reward": -0.051611170172691345, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.051611170172691345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20889604091644287, + 0.16799786686897278, + 0.08677363395690918, + 0.06486440449953079, + 0.11388466507196426, + 0.06657882034778595, + 0.08795320242643356, + 0.11973586678504944, + 0.044032398611307144, + 0.03928311914205551 + ], + "credit_map": [ + 0.02739599347114563, + 0.026149533689022064, + 0.031344223767519, + 0.03111693076789379, + 0.015694405883550644, + 0.026852278038859367, + 0.01985517144203186, + 0.030487501993775368, + 0.015171398408710957, + 0.039127152413129807, + 0.02395671419799328, + 0.01699882000684738, + 0.015159428119659424, + 0.018447794020175934, + 0.016765091568231583, + 0.022371456027030945, + 0.00847333949059248, + 0.012062808498740196, + 0.011137076653540134, + 0.011180059984326363, + 0.007330083753913641, + 0.01115377340465784, + 0.01422529574483633, + 0.011211195029318333, + 0.005437441170215607, + 0.005559920333325863, + 0.006161119788885117, + 0.007682824041694403, + 0.004522974602878094, + 0.01153397187590599, + 0.011395883746445179, + 0.012570270337164402, + 0.01299167051911354, + 0.007809180300682783, + 0.008178453892469406, + 0.00848263967782259, + 0.006873383652418852, + 0.028293181210756302, + 0.012046783231198788, + 0.029209379106760025, + 0.0041844192892313, + 0.0058909147046506405, + 0.0073441616259515285, + 0.009730487130582333, + 0.005079000722616911, + 0.011245531029999256, + 0.011525810696184635, + 0.011578493751585484, + 0.006030064541846514, + 0.0065238215029239655, + 0.008517432026565075, + 0.01578102447092533, + 0.014552622102200985, + 0.01073600072413683, + 0.013479281216859818, + 0.012332960963249207, + 0.012649795971810818, + 0.005636805668473244, + 0.018416959792375565, + 0.015516672283411026, + 0.013001124374568462, + 0.019490111619234085, + 0.01425770204514265, + 0.020766692236065865, + 0.006290509831160307, + 0.0050685047172009945, + 0.007325981743633747, + 0.008834482170641422, + 0.003958831541240215, + 0.0033287815749645233, + 0.0038249820936471224, + 0.005400323774665594, + 0.0071069952100515366, + 0.003478327300399542, + 0.008246743120253086, + 0.007041921839118004, + 0.0021716661285609007, + 0.002073056995868683, + 0.0037227696739137173, + 0.005441637244075537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.051611170172691345 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07904429733753204 + ], + "opposite_logprob_tokens": [ + -2.5790443420410156 + ], + "logprob_reward": -0.07904429733753204, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07904429733753204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19373928010463715, + 0.13055866956710815, + 0.08045711368322372, + 0.03491390496492386, + 0.13193009793758392, + 0.08745212852954865, + 0.08582908660173416, + 0.06477020680904388, + 0.13546214997768402, + 0.05488738790154457 + ], + "credit_map": [ + 0.02608688734471798, + 0.026557795703411102, + 0.02911587804555893, + 0.02967025712132454, + 0.016268150880932808, + 0.02292259782552719, + 0.01950375735759735, + 0.0236139465123415, + 0.013418368063867092, + 0.020503465086221695, + 0.017126908525824547, + 0.014994264580309391, + 0.014381514862179756, + 0.015970297157764435, + 0.01660289615392685, + 0.01756095513701439, + 0.007900652475655079, + 0.008810123428702354, + 0.010449045337736607, + 0.007151556201279163, + 0.006492705550044775, + 0.010793696157634258, + 0.016629287973046303, + 0.012230047956109047, + 0.00294517376460135, + 0.004389242734760046, + 0.0048794224858284, + 0.00429096445441246, + 0.002577332779765129, + 0.005400561261922121, + 0.004618981387466192, + 0.005812220741063356, + 0.010193469002842903, + 0.005416565574705601, + 0.00806634221225977, + 0.00821311492472887, + 0.006825372576713562, + 0.04150551185011864, + 0.00950466189533472, + 0.04220505431294441, + 0.004891956225037575, + 0.01029124390333891, + 0.008288692682981491, + 0.009390337392687798, + 0.007327025290578604, + 0.017492644488811493, + 0.012250552885234356, + 0.017519665881991386, + 0.010324577800929546, + 0.006344147492200136, + 0.011270346119999886, + 0.013147610239684582, + 0.01074542198330164, + 0.011074543930590153, + 0.010888967663049698, + 0.012033474631607533, + 0.00912508275359869, + 0.003203257918357849, + 0.010428463108837605, + 0.00811096467077732, + 0.00711375055834651, + 0.009745175950229168, + 0.007718182634562254, + 0.00932532362639904, + 0.019748926162719727, + 0.014908766373991966, + 0.023699413985013962, + 0.020886041224002838, + 0.012353052385151386, + 0.009881931357085705, + 0.010570657439529896, + 0.023413347080349922, + 0.008052544668316841, + 0.007825127802789211, + 0.01011178083717823, + 0.008878467604517937, + 0.004062131978571415, + 0.0033232567366212606, + 0.00446936534717679, + 0.008164712227880955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.07904429733753204 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035957634449005 + ], + "opposite_logprob_tokens": [ + -1.9103596210479736 + ], + "logprob_reward": -0.16035957634449005, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035957634449005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17137674987316132, + 0.18790461122989655, + 0.07687847316265106, + 0.0787682980298996, + 0.07206377387046814, + 0.09398780018091202, + 0.0957517996430397, + 0.08525142073631287, + 0.0972575694322586, + 0.04075954109430313 + ], + "credit_map": [ + 0.020839471369981766, + 0.02554444782435894, + 0.028128093108534813, + 0.027163494378328323, + 0.01340574212372303, + 0.02076154761016369, + 0.011927719227969646, + 0.02360621467232704, + 0.015956688672304153, + 0.032454632222652435, + 0.029439831152558327, + 0.02335401065647602, + 0.014104174450039864, + 0.02937283366918564, + 0.022279763594269753, + 0.02094266749918461, + 0.007971225306391716, + 0.007557798642665148, + 0.011143545620143414, + 0.006046423688530922, + 0.004907194525003433, + 0.015478328801691532, + 0.011329780332744122, + 0.012444180436432362, + 0.011524914763867855, + 0.006661891937255859, + 0.010707912966609001, + 0.011293561197817326, + 0.005207763984799385, + 0.011129195801913738, + 0.011991245672106743, + 0.010251807048916817, + 0.006445296574383974, + 0.006044507957994938, + 0.0063654291443526745, + 0.0077060917392373085, + 0.00311815133318305, + 0.01772431470453739, + 0.006294742226600647, + 0.01836523972451687, + 0.004337521735578775, + 0.013731631450355053, + 0.007262476719915867, + 0.012511005625128746, + 0.008128222078084946, + 0.012362221255898476, + 0.02371065877377987, + 0.011944049969315529, + 0.005432451609522104, + 0.005422140005975962, + 0.009687519632279873, + 0.017697228118777275, + 0.011989776976406574, + 0.017297115176916122, + 0.009616248309612274, + 0.018609311431646347, + 0.008777698501944542, + 0.003917391411960125, + 0.011582630686461926, + 0.014971393160521984, + 0.011099220253527164, + 0.011699049733579159, + 0.01191775780171156, + 0.011286269873380661, + 0.011720181442797184, + 0.009920058771967888, + 0.017277352511882782, + 0.015317117795348167, + 0.009149488061666489, + 0.008532674051821232, + 0.008199003525078297, + 0.017141683027148247, + 0.005694261286407709, + 0.005744313355535269, + 0.0070543126203119755, + 0.008036640472710133, + 0.002465497702360153, + 0.0023607315961271524, + 0.00476126279681921, + 0.0046425205655395985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.16035957634449005 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029815562069416046 + ], + "opposite_logprob_tokens": [ + -3.529815673828125 + ], + "logprob_reward": -0.029815562069416046, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029815562069416046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12780174612998962, + 0.23358508944511414, + 0.11667396873235703, + 0.07432858645915985, + 0.1304142028093338, + 0.05737799406051636, + 0.10260811448097229, + 0.07308618724346161, + 0.04065944254398346, + 0.04346461221575737 + ], + "credit_map": [ + 0.016399085521697998, + 0.017065931111574173, + 0.01975150778889656, + 0.017091019079089165, + 0.0084958765655756, + 0.021517368033528328, + 0.009080751799046993, + 0.018400205299258232, + 0.019769687205553055, + 0.044107623398303986, + 0.03012133575975895, + 0.0277558583766222, + 0.019936366006731987, + 0.037101127207279205, + 0.022525396198034286, + 0.03226768970489502, + 0.013530079275369644, + 0.012402893975377083, + 0.012676612474024296, + 0.012601113878190517, + 0.008628261275589466, + 0.023141905665397644, + 0.018684085458517075, + 0.015009012073278427, + 0.0088431965559721, + 0.006800195202231407, + 0.011818182654678822, + 0.008843045681715012, + 0.006323386449366808, + 0.010758220218122005, + 0.010745139792561531, + 0.010197221301496029, + 0.007989389821887016, + 0.011177084408700466, + 0.009537441655993462, + 0.0103263258934021, + 0.006688376888632774, + 0.03868621215224266, + 0.011224932037293911, + 0.034784432500600815, + 0.003325922880321741, + 0.006475324742496014, + 0.005873410496860743, + 0.008566792123019695, + 0.004580758512020111, + 0.008082376793026924, + 0.012708848342299461, + 0.0077645620331168175, + 0.007361824624240398, + 0.0074983746744692326, + 0.009509299881756306, + 0.01875699684023857, + 0.018688516691327095, + 0.013097969815135002, + 0.013253591023385525, + 0.014441538602113724, + 0.007840365171432495, + 0.003552475944161415, + 0.01250905729830265, + 0.010449071414768696, + 0.007062390446662903, + 0.011614912189543247, + 0.008081999607384205, + 0.011975917033851147, + 0.007063466589897871, + 0.004334742669016123, + 0.0065117934718728065, + 0.006275208666920662, + 0.004375086631625891, + 0.0028597803320735693, + 0.003548035863786936, + 0.0056913308799266815, + 0.006842322647571564, + 0.006669546477496624, + 0.010707980021834373, + 0.0061517320573329926, + 0.0029580453410744667, + 0.0020899514202028513, + 0.002792966552078724, + 0.00525206932798028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.029815562069416046 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341434568166733 + ], + "opposite_logprob_tokens": [ + -3.651341438293457 + ], + "logprob_reward": -0.026341434568166733, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341434568166733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13920418918132782, + 0.29524123668670654, + 0.10186523199081421, + 0.038079749792814255, + 0.05252648890018463, + 0.05070348456501961, + 0.08630041033029556, + 0.09022309631109238, + 0.09993581473827362, + 0.045920271426439285 + ], + "credit_map": [ + 0.02192644774913788, + 0.014076519757509232, + 0.015676887705922127, + 0.01563769020140171, + 0.010842339135706425, + 0.019713154062628746, + 0.00872520636767149, + 0.03260594978928566, + 0.02860834449529648, + 0.05417921766638756, + 0.03908560052514076, + 0.033634163439273834, + 0.02969726175069809, + 0.03522024303674698, + 0.03914734348654747, + 0.035669051110744476, + 0.009446111507713795, + 0.013559844344854355, + 0.012537469156086445, + 0.013638065196573734, + 0.007731017656624317, + 0.013364333659410477, + 0.018565291538834572, + 0.01302309799939394, + 0.004221735056489706, + 0.004253735300153494, + 0.0034541403874754906, + 0.004754339344799519, + 0.0040359655395150185, + 0.005265402141958475, + 0.006289405282586813, + 0.0058050272054970264, + 0.005541538819670677, + 0.003663752693682909, + 0.004969524219632149, + 0.005811468698084354, + 0.00337046105414629, + 0.012186889536678791, + 0.00444592721760273, + 0.012536927126348019, + 0.002525336341932416, + 0.004976887721568346, + 0.0038878864143043756, + 0.005941597279161215, + 0.004529249854385853, + 0.00902283564209938, + 0.011971095576882362, + 0.007848595269024372, + 0.0073386868461966515, + 0.004423320759087801, + 0.009833033196628094, + 0.012243879958987236, + 0.010642812587320805, + 0.014988251961767673, + 0.009754977189004421, + 0.017075451090931892, + 0.010046492330729961, + 0.005124013405293226, + 0.013758866116404533, + 0.013731954619288445, + 0.01412185002118349, + 0.010134299285709858, + 0.012687284499406815, + 0.01061833743005991, + 0.015469455160200596, + 0.009657870046794415, + 0.02165168523788452, + 0.014721698127686977, + 0.007529326248914003, + 0.007260671351104975, + 0.00861072726547718, + 0.015034381300210953, + 0.006213739514350891, + 0.007806837558746338, + 0.007020139135420322, + 0.00731574697420001, + 0.003496844321489334, + 0.002868211129680276, + 0.0037887811195105314, + 0.007409973070025444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.026341434568166733 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, the camera's slow, smooth orbiting motion orbits a high-precision e-reader suspended in a pure, endless void of darkness. The device performs complex, deliberate multi-axis flips in a physically precise gyroscope. Its ultra-thin body, crafted from matte, dark graphite magnesium alloy, reflects a subtle, refined quality under a single, soft, overhead light. As the device slowly rotates, this light creates crisp highlights along its sharp edges while casting soft shadows against the deep background. Its ultra-narrow, bezel-less E-Ink screen displays a solitary cover design: a richly detailed, withered tree stands alone on a barren hill, beneath a pale, empty sky rendered in delicate grayscale tones. As the e-reader rolls steadily and gracefully, the camera remains focused on the non-luminous, paper-like screen, capturing every moment of light flowing through the cover pattern. The entire scene is shrouded in a quiet, focused, minimalist atmosphere with a technological aesthetic, emphasizing the object itself and the lonely story it carries.", + "seed": 23858, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_2/index_9527_seed_23858_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11285354532301425, + "vlm_logprob": -0.11285354532301425, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.11285354532301425, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0986545979976654 + ], + "opposite_logprob_tokens": [ + -2.4736545085906982 + ], + "logprob_reward": -0.0986545979976654, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0986545979976654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056829098612070084, + 0.06313520669937134, + 0.04237598553299904, + 0.6732363104820251, + 0.03537478297948837, + 0.055619146674871445, + 0.041210997849702835, + 0.011195098049938679, + 0.006049665156751871, + 0.014973675832152367 + ], + "credit_map": [ + 0.008175901137292385, + 0.007486641872674227, + 0.008991965092718601, + 0.006604467052966356, + 0.005692020058631897, + 0.0069215367548167706, + 0.0034454527776688337, + 0.009511114098131657, + 0.009602692909538746, + 0.006924211513251066, + 0.006810448132455349, + 0.008434779942035675, + 0.0053231543861329556, + 0.009366044774651527, + 0.006663821171969175, + 0.010010058991611004, + 0.005899387877434492, + 0.00388471526093781, + 0.005996379069983959, + 0.004517486318945885, + 0.005228583700954914, + 0.005196536425501108, + 0.004872554913163185, + 0.006780343595892191, + 0.21342505514621735, + 0.07983855903148651, + 0.034118618816137314, + 0.02325713075697422, + 0.03211547061800957, + 0.11590877920389175, + 0.023405589163303375, + 0.1511671245098114, + 0.004769935738295317, + 0.005445944610983133, + 0.0032193788792937994, + 0.004525845870375633, + 0.0033222700003534555, + 0.005651128012686968, + 0.00394895626232028, + 0.00449132826179266, + 0.0053584217093884945, + 0.006073512602597475, + 0.005634883418679237, + 0.004851917270570993, + 0.006002508103847504, + 0.00982288271188736, + 0.007922527380287647, + 0.009952498599886894, + 0.0042794570326805115, + 0.00306266313418746, + 0.008335784077644348, + 0.004422751720994711, + 0.0041951402090489864, + 0.005774065852165222, + 0.005946268793195486, + 0.005194868892431259, + 0.0012766238069161773, + 0.0007446283707395196, + 0.0012701790546998382, + 0.0022746571339666843, + 0.0013850993709638715, + 0.0005899684620089829, + 0.0020680741872638464, + 0.0015858691185712814, + 0.0007714434177614748, + 0.00048129254719242454, + 0.001073310268111527, + 0.000902234110981226, + 0.0008462920086458325, + 0.0003082396578975022, + 0.0006305964780040085, + 0.0010362565517425537, + 0.0018250603461638093, + 0.002022702246904373, + 0.00213837344199419, + 0.001262778416275978, + 0.002083597471937537, + 0.0009639131603762507, + 0.001854400965385139, + 0.0028228515293449163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.0986545979976654 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799718141555786 + ], + "opposite_logprob_tokens": [ + -1.8049718141555786 + ], + "logprob_reward": -0.1799718141555786, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799718141555786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11622433364391327, + 0.19848452508449554, + 0.09015198051929474, + 0.23419462144374847, + 0.12406739592552185, + 0.04921265318989754, + 0.06977995485067368, + 0.07473184168338776, + 0.010978753678500652, + 0.03217398002743721 + ], + "credit_map": [ + 0.020966337993741035, + 0.014557705260813236, + 0.01613582670688629, + 0.016209622845053673, + 0.011461920104920864, + 0.013763774186372757, + 0.006366009823977947, + 0.016763130202889442, + 0.04000326991081238, + 0.02124815806746483, + 0.020352791994810104, + 0.02085346356034279, + 0.01633206568658352, + 0.032070547342300415, + 0.01963752508163452, + 0.02798670530319214, + 0.012396548874676228, + 0.008285416290163994, + 0.009702161885797977, + 0.01140300277620554, + 0.008373832330107689, + 0.015424815006554127, + 0.010929914191365242, + 0.013636292889714241, + 0.06535323709249496, + 0.027532096952199936, + 0.01284758746623993, + 0.007789327763020992, + 0.011394660919904709, + 0.052183765918016434, + 0.010073726996779442, + 0.04702020809054375, + 0.01840152218937874, + 0.014613362960517406, + 0.009861881844699383, + 0.011834056116640568, + 0.0114369485527277, + 0.02949357219040394, + 0.012865366414189339, + 0.01556068193167448, + 0.0049947998486459255, + 0.007105154916644096, + 0.00583298783749342, + 0.004535940941423178, + 0.0057575819082558155, + 0.009345969185233116, + 0.004771082196384668, + 0.006869135890156031, + 0.008712942712008953, + 0.0056524137035012245, + 0.012129509821534157, + 0.007874270901083946, + 0.008171220310032368, + 0.00906448159366846, + 0.00946296937763691, + 0.008712152019143105, + 0.01111162081360817, + 0.0057072872295975685, + 0.0073049976490437984, + 0.01373149175196886, + 0.006313334684818983, + 0.004421014804393053, + 0.012920711189508438, + 0.013221384026110172, + 0.0016413641860708594, + 0.0011233752593398094, + 0.001887094578705728, + 0.0017881308449432254, + 0.0009238449274562299, + 0.0006284250994212925, + 0.0011876859935000539, + 0.0017988326726481318, + 0.00556124746799469, + 0.00329980137757957, + 0.004920000676065683, + 0.0034572493750602007, + 0.0025839279405772686, + 0.002716983202844858, + 0.004479921888560057, + 0.00515484856441617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.1799718141555786 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558416426181793 + ], + "opposite_logprob_tokens": [ + -1.6005841493606567 + ], + "logprob_reward": -0.22558416426181793, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558416426181793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07761148363351822, + 0.17357847094535828, + 0.06591495126485825, + 0.3899296820163727, + 0.14078080654144287, + 0.049106840044260025, + 0.035025112330913544, + 0.012595799751579762, + 0.01031839195638895, + 0.045138370245695114 + ], + "credit_map": [ + 0.011301415041089058, + 0.008196664042770863, + 0.011128238402307034, + 0.009694158099591732, + 0.007512510754168034, + 0.010459773242473602, + 0.004744882695376873, + 0.014573845081031322, + 0.030918648466467857, + 0.017881328240036964, + 0.016705641523003578, + 0.01668049953877926, + 0.01714726723730564, + 0.02635251358151436, + 0.015672678127884865, + 0.032219886779785156, + 0.009608549065887928, + 0.00502464734017849, + 0.00759545061737299, + 0.007205422502011061, + 0.006238006986677647, + 0.009990300051867962, + 0.009930367581546307, + 0.01032220758497715, + 0.09823992103338242, + 0.03695826604962349, + 0.02618085779249668, + 0.021160423755645752, + 0.023904835805296898, + 0.09003589302301407, + 0.020979590713977814, + 0.07246989756822586, + 0.02084270678460598, + 0.011925570666790009, + 0.011857819743454456, + 0.01958371326327324, + 0.01169628743082285, + 0.022707628086209297, + 0.012209043838083744, + 0.029958048835396767, + 0.005050011444836855, + 0.0048567685298621655, + 0.005410198587924242, + 0.0051210299134254456, + 0.006707991473376751, + 0.00864044763147831, + 0.005495589226484299, + 0.007824800908565521, + 0.0035884296521544456, + 0.0025872180704027414, + 0.0072251115925610065, + 0.0032207246404141188, + 0.0031584978569298983, + 0.0052190679125487804, + 0.00512612983584404, + 0.004899934399873018, + 0.0014279320603236556, + 0.0009030303335748613, + 0.0019519071793183684, + 0.0019366414053365588, + 0.0015979721210896969, + 0.0010438802419230342, + 0.0017099688993766904, + 0.0020244678016752005, + 0.0013685304438695312, + 0.001055464381352067, + 0.0012119162129238248, + 0.001516772317700088, + 0.0016182559775188565, + 0.0005225744680501521, + 0.0014132206561043859, + 0.001611657440662384, + 0.005391152109950781, + 0.00595538504421711, + 0.008941898122429848, + 0.005179734900593758, + 0.004538615234196186, + 0.00437168451026082, + 0.004543896298855543, + 0.006216002628207207 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.22558416426181793 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371815383434296 + ], + "opposite_logprob_tokens": [ + -3.4087181091308594 + ], + "logprob_reward": -0.03371815383434296, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371815383434296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1886487603187561, + 0.09991459548473358, + 0.05022357404232025, + 0.2810516357421875, + 0.17430460453033447, + 0.09306291490793228, + 0.06919483095407486, + 0.017122907564044, + 0.008230275474488735, + 0.018245920538902283 + ], + "credit_map": [ + 0.023827308788895607, + 0.022483764216303825, + 0.025714918971061707, + 0.02253727801144123, + 0.024237336590886116, + 0.024215107783675194, + 0.012926265597343445, + 0.03270678594708443, + 0.015982314944267273, + 0.009614718146622181, + 0.012091628275811672, + 0.012770638801157475, + 0.0088910311460495, + 0.014354619197547436, + 0.013667470775544643, + 0.012542173266410828, + 0.0074056037701666355, + 0.004386163782328367, + 0.0073555149137973785, + 0.007642589975148439, + 0.0039579845033586025, + 0.005482041742652655, + 0.005952716339379549, + 0.0080409599468112, + 0.0852685272693634, + 0.033529847860336304, + 0.01804295741021633, + 0.012943740002810955, + 0.01410389319062233, + 0.04361240938305855, + 0.012406564317643642, + 0.06114370748400688, + 0.019075127318501472, + 0.025057567283511162, + 0.014031009748578072, + 0.024130109697580338, + 0.019165506586432457, + 0.032108988612890244, + 0.024262597784399986, + 0.01647369936108589, + 0.008555958978831768, + 0.010686762630939484, + 0.009455920197069645, + 0.009232139214873314, + 0.009267513640224934, + 0.014820204116404057, + 0.013453767634928226, + 0.01759064570069313, + 0.00671731773763895, + 0.006192303262650967, + 0.012261496856808662, + 0.00904920045286417, + 0.010774554684758186, + 0.0068225865252316, + 0.008222319185733795, + 0.009155052714049816, + 0.00167439179494977, + 0.0014656848507001996, + 0.0020121363922953606, + 0.0035600338596850634, + 0.002672759350389242, + 0.0008303801878355443, + 0.0022884258069097996, + 0.002619096077978611, + 0.0011008806759491563, + 0.0007013271097093821, + 0.00107289117295295, + 0.0014369701966643333, + 0.0010279096895828843, + 0.0006334526115097106, + 0.0009383080760017037, + 0.0013185362331569195, + 0.002264596987515688, + 0.0022706640884280205, + 0.002516512293368578, + 0.001482857158407569, + 0.0030641264747828245, + 0.0015746774151921272, + 0.0021734442561864853, + 0.0028990423306822777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.03371815383434296 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a high-precision e-reader with an ultra-thin body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the e-reader's body made of matte, dark graphite magnesium alloy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the e-reader perform complex, deliberate multi-axis flips in a physically precise gyroscope motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the E-Ink screen display a solitary cover design of a withered tree on a barren hill under a pale sky in grayscale tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a slow, smooth orbiting motion around the e-reader?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633899636566639 + ], + "opposite_logprob_tokens": [ + -3.651339054107666 + ], + "logprob_reward": -0.02633899636566639, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633899636566639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07352548092603683, + 0.06522729992866516, + 0.055309828370809555, + 0.48017361760139465, + 0.14376123249530792, + 0.05159507319331169, + 0.041559163480997086, + 0.02110421471297741, + 0.015028937719762325, + 0.05271511524915695 + ], + "credit_map": [ + 0.010629041120409966, + 0.0077791656367480755, + 0.011684969998896122, + 0.005940111819654703, + 0.007264365907758474, + 0.014181388542056084, + 0.00378168816678226, + 0.01226475927978754, + 0.013735353946685791, + 0.0067765903659164906, + 0.007266864646226168, + 0.005546442698687315, + 0.005867707543075085, + 0.008252841420471668, + 0.007940597832202911, + 0.009840909391641617, + 0.008304480463266373, + 0.005854066461324692, + 0.00611955625936389, + 0.006088308524340391, + 0.004568552598357201, + 0.00751653965562582, + 0.0073605249635875225, + 0.009497812017798424, + 0.15987347066402435, + 0.0443682037293911, + 0.025831373408436775, + 0.02597998082637787, + 0.020906567573547363, + 0.06829525530338287, + 0.025706525892019272, + 0.10921230912208557, + 0.018102241680026054, + 0.015381019562482834, + 0.013889534398913383, + 0.021442247554659843, + 0.01418394036591053, + 0.019650552421808243, + 0.018333373591303825, + 0.022778334096074104, + 0.004975988529622555, + 0.006022323854267597, + 0.006362460553646088, + 0.0038359041791409254, + 0.007486314047127962, + 0.007851097732782364, + 0.005670660175383091, + 0.009390328079462051, + 0.005053891334682703, + 0.002975062932819128, + 0.007843387313187122, + 0.004380714613944292, + 0.004943497478961945, + 0.005850205197930336, + 0.004897140432149172, + 0.005615265108644962, + 0.0031156656332314014, + 0.0015725604025647044, + 0.002066713059321046, + 0.003692125203087926, + 0.0032004103995859623, + 0.0012008436024188995, + 0.0028354476671665907, + 0.0034204518888145685, + 0.003451482392847538, + 0.0016403234330937266, + 0.0021898967679589987, + 0.0021750314626842737, + 0.0011148490011692047, + 0.000830632692668587, + 0.001530881505459547, + 0.002095842966809869, + 0.006699769292026758, + 0.004891005344688892, + 0.011177083477377892, + 0.00617740536108613, + 0.005822819657623768, + 0.00395309878513217, + 0.005551313050091267, + 0.008442625403404236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.02633899636566639 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_3/index_tensor([8575, 8575])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_3/index_tensor([8575, 8575])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..94871a62ad6d9cf582fe11ee883a215fe55c0722 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_3/index_tensor([8575, 8575])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level close-up depicts a beautiful woman with long, honey-blonde hair falling in soft waves over her shoulders. She wears a vibrant royal blue jacket crafted from soft suede, the rich color contrasting beautifully with her fair complexion. Delicate, dangling silver earrings dangle from her ears, gently swaying and catching the light. The scene is bathed in the warm, diffuse glow of the golden hour, casting soft highlights on her high cheekbones and illuminating her clear blue eyes. She slowly turns her head to look directly into the camera, a gentle breeze ruffling a lock of her hair, a subtle, serene smile playing on her lips. In the background, a softly blurred cafe terrace, the blurred shapes of other patrons and the flicker of warm string lights create an atmosphere of dreamy, intimate, and effortlessly elegant elegance.", + "seed": 30404, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_3/index_8575_seed_30404_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.028378412406891585, + "vlm_logprob": -0.028378412406891585, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.028378412406891585, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009573973715305328 + ], + "opposite_logprob_tokens": [ + -5.009573936462402 + ], + "logprob_reward": -0.009573973715305328, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009573973715305328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17087505757808685, + 0.09559445828199387, + 0.12921766936779022, + 0.06597071886062622, + 0.09201399981975555, + 0.1153869479894638, + 0.04273630306124687, + 0.045775651931762695, + 0.11139649152755737, + 0.13103267550468445 + ], + "credit_map": [ + 0.013403035700321198, + 0.043138548731803894, + 0.018763232976198196, + 0.019626548513770103, + 0.02639862149953842, + 0.014507057145237923, + 0.018576612696051598, + 0.016461409628391266, + 0.009272539056837559, + 0.013651557266712189, + 0.008875015191733837, + 0.012618720531463623, + 0.014416136778891087, + 0.010804626159369946, + 0.015073545277118683, + 0.010882318019866943, + 0.01657819375395775, + 0.010287193581461906, + 0.013914607465267181, + 0.022881725803017616, + 0.01595648191869259, + 0.027236517518758774, + 0.008060168474912643, + 0.01430278830230236, + 0.01012189220637083, + 0.006947551853954792, + 0.004565667826682329, + 0.007946247234940529, + 0.008205159567296505, + 0.008493877947330475, + 0.006737106014043093, + 0.012953216210007668, + 0.013182423077523708, + 0.012105308473110199, + 0.00976700522005558, + 0.009977336041629314, + 0.008836226537823677, + 0.018163815140724182, + 0.007259397767484188, + 0.012722485698759556, + 0.012486515566706657, + 0.01892094500362873, + 0.008426967076957226, + 0.006426695268601179, + 0.015550871379673481, + 0.027335742488503456, + 0.01892993412911892, + 0.007309277076274157, + 0.005705771967768669, + 0.004193674772977829, + 0.007007762789726257, + 0.00394095666706562, + 0.005877078510820866, + 0.006019284017384052, + 0.005302760750055313, + 0.004689014051109552, + 0.005759861785918474, + 0.004141346085816622, + 0.0061570643447339535, + 0.0057518635876476765, + 0.0072484491392970085, + 0.0064300852827727795, + 0.0054879747331142426, + 0.0047990065068006516, + 0.01382475532591343, + 0.010198806412518024, + 0.0134740574285388, + 0.02918952703475952, + 0.013864179141819477, + 0.013762430287897587, + 0.010758931748569012, + 0.006323813460767269, + 0.012168804183602333, + 0.012537228874862194, + 0.017399871721863747, + 0.032758235931396484, + 0.012382605113089085, + 0.01620505563914776, + 0.012899373658001423, + 0.014681493863463402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.009573973715305328 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038102809339761734 + ], + "opposite_logprob_tokens": [ + -3.288102865219116 + ], + "logprob_reward": -0.038102809339761734, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038102809339761734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17395001649856567, + 0.09840637445449829, + 0.16825315356254578, + 0.0355486124753952, + 0.06397878378629684, + 0.21993103623390198, + 0.05682973936200142, + 0.10836496949195862, + 0.02375282533466816, + 0.05098450183868408 + ], + "credit_map": [ + 0.011627552099525928, + 0.030666084960103035, + 0.014029969461262226, + 0.020028796046972275, + 0.024066563695669174, + 0.01808328367769718, + 0.030798373743891716, + 0.024649396538734436, + 0.011003654450178146, + 0.012532565742731094, + 0.00629916088655591, + 0.01305011659860611, + 0.01612669602036476, + 0.009893037378787994, + 0.019680052995681763, + 0.009821082465350628, + 0.02190634422004223, + 0.013623073697090149, + 0.013889848254621029, + 0.03000917285680771, + 0.02062923274934292, + 0.02330874651670456, + 0.019862772896885872, + 0.025023959577083588, + 0.0043408432975411415, + 0.003656747518107295, + 0.002887546317651868, + 0.004168052691966295, + 0.00546953035518527, + 0.004099680110812187, + 0.0045174723491072655, + 0.006408737972378731, + 0.008745231665670872, + 0.007972664199769497, + 0.004796539433300495, + 0.004670171998441219, + 0.006420803721994162, + 0.01863081566989422, + 0.003311174688860774, + 0.009431383572518826, + 0.026356490328907967, + 0.048408154398202896, + 0.01289109606295824, + 0.014357442036271095, + 0.026883020997047424, + 0.04573235288262367, + 0.03349188342690468, + 0.011810602620244026, + 0.005820121616125107, + 0.004935725592076778, + 0.008843870833516121, + 0.0052337185479700565, + 0.009542908519506454, + 0.008797944523394108, + 0.007898098789155483, + 0.005757350940257311, + 0.012673350051045418, + 0.009757578372955322, + 0.011715900152921677, + 0.020112520083785057, + 0.016090530902147293, + 0.01719369925558567, + 0.01226466428488493, + 0.008556722663342953, + 0.003924576565623283, + 0.002111891284584999, + 0.0027540388982743025, + 0.003617973765358329, + 0.0029694894328713417, + 0.0027320506051182747, + 0.0035884971730411053, + 0.0020543073769658804, + 0.00470409682020545, + 0.004226616583764553, + 0.006985152140259743, + 0.011332899332046509, + 0.0056175366044044495, + 0.005821416154503822, + 0.009525954723358154, + 0.002770827617496252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.038102809339761734 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002181559801102 + ], + "opposite_logprob_tokens": [ + -2.695021867752075 + ], + "logprob_reward": -0.07002181559801102, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002181559801102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10059932619333267, + 0.15391141176223755, + 0.1456957459449768, + 0.05905371159315109, + 0.13387826085090637, + 0.20037473738193512, + 0.03281203657388687, + 0.06256508082151413, + 0.021544823423027992, + 0.08956491202116013 + ], + "credit_map": [ + 0.007844137027859688, + 0.018362466245889664, + 0.007022131234407425, + 0.011931129731237888, + 0.014506321400403976, + 0.010977855883538723, + 0.01613682135939598, + 0.01381844561547041, + 0.014299778267741203, + 0.027180764824151993, + 0.012620587833225727, + 0.01676933467388153, + 0.02344628982245922, + 0.018830616027116776, + 0.023545490577816963, + 0.017218535766005516, + 0.01915603131055832, + 0.011416623368859291, + 0.016993073746562004, + 0.023172978311777115, + 0.01910603791475296, + 0.021420558914542198, + 0.01316468883305788, + 0.021265747025609016, + 0.006893682759255171, + 0.0063173118978738785, + 0.005642548203468323, + 0.008254358544945717, + 0.009534908458590508, + 0.008332357741892338, + 0.007235631346702576, + 0.006842908915132284, + 0.01861714944243431, + 0.01524538267403841, + 0.0090558510273695, + 0.009425575844943523, + 0.011914512142539024, + 0.039547454565763474, + 0.009997999295592308, + 0.02007431536912918, + 0.02296828106045723, + 0.04178868234157562, + 0.011759007349610329, + 0.012105860747396946, + 0.023007294163107872, + 0.04340379312634468, + 0.03467154502868652, + 0.010670260526239872, + 0.004345526918768883, + 0.004790890030562878, + 0.0035038418136537075, + 0.0028988071717321873, + 0.004560604691505432, + 0.005968650337308645, + 0.004141210578382015, + 0.0026025022380053997, + 0.007946535013616085, + 0.005930765066295862, + 0.005533158779144287, + 0.012146116234362125, + 0.01074600126594305, + 0.007601766847074032, + 0.00804833322763443, + 0.004612398333847523, + 0.003527245717123151, + 0.0020382313523441553, + 0.0029535230714827776, + 0.002545481314882636, + 0.0031427275389432907, + 0.0032904541585594416, + 0.002328630303964019, + 0.0017185276374220848, + 0.009415845386683941, + 0.00886988453567028, + 0.012518704868853092, + 0.017663395032286644, + 0.0127276461571455, + 0.013392004184424877, + 0.008922924287617207, + 0.006054500583559275 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.07002181559801102 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036134920082986355 + ], + "opposite_logprob_tokens": [ + -5.628613471984863 + ], + "logprob_reward": -0.0036134920082986355, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036134920082986355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2596289813518524, + 0.12511573731899261, + 0.11917635798454285, + 0.05405736342072487, + 0.10585355758666992, + 0.12486878037452698, + 0.03882298991084099, + 0.11452801525592804, + 0.027436941862106323, + 0.030511263757944107 + ], + "credit_map": [ + 0.02206137776374817, + 0.08408169448375702, + 0.020306792110204697, + 0.026210997253656387, + 0.0422351248562336, + 0.020043419674038887, + 0.02558794803917408, + 0.01910163089632988, + 0.013540549203753471, + 0.022619659081101418, + 0.009039119817316532, + 0.014419546350836754, + 0.014132349751889706, + 0.019131440669298172, + 0.01929255574941635, + 0.012940527871251106, + 0.01475963369011879, + 0.011507069692015648, + 0.013004571199417114, + 0.013884879648685455, + 0.017838187515735626, + 0.02277294360101223, + 0.013293802738189697, + 0.012115276418626308, + 0.0058110859245061874, + 0.006237294990569353, + 0.004865597002208233, + 0.008478008210659027, + 0.005848558619618416, + 0.009790736250579357, + 0.005514186341315508, + 0.007511900272220373, + 0.015860117971897125, + 0.013217496685683727, + 0.011824002489447594, + 0.009702404960989952, + 0.017004724591970444, + 0.014619077555835247, + 0.014278091490268707, + 0.00934764463454485, + 0.014447968453168869, + 0.02821226604282856, + 0.007477134466171265, + 0.007465821225196123, + 0.013078492134809494, + 0.02773599699139595, + 0.02073793113231659, + 0.005713180173188448, + 0.006025309208780527, + 0.006581705529242754, + 0.0056235166266560555, + 0.0038423468358814716, + 0.005173281766474247, + 0.005757123231887817, + 0.0036153518594801426, + 0.002204355550929904, + 0.012141469866037369, + 0.009364807978272438, + 0.021765928715467453, + 0.026966247707605362, + 0.011269636452198029, + 0.015706051141023636, + 0.010732535272836685, + 0.006581340916454792, + 0.005096580367535353, + 0.002158682094886899, + 0.0048225452192127705, + 0.005741931963711977, + 0.002902537351474166, + 0.0035034155007451773, + 0.001956794410943985, + 0.0012544573983177543, + 0.0031185736879706383, + 0.002644288120791316, + 0.0050722574815154076, + 0.0055079287849366665, + 0.0034336140379309654, + 0.003862423589453101, + 0.005273055750876665, + 0.0015991254476830363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0036134920082986355 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579971373081207 + ], + "opposite_logprob_tokens": [ + -3.8955800533294678 + ], + "logprob_reward": -0.020579971373081207, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579971373081207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17456993460655212, + 0.05478646978735924, + 0.1082296296954155, + 0.04894324392080307, + 0.1129877045750618, + 0.41553038358688354, + 0.019939832389354706, + 0.019063500687479973, + 0.01344755943864584, + 0.03250173479318619 + ], + "credit_map": [ + 0.015146465040743351, + 0.04266832768917084, + 0.021652067080140114, + 0.020810460671782494, + 0.029014335945248604, + 0.020182378590106964, + 0.01293069776147604, + 0.012165195308625698, + 0.006534933112561703, + 0.010850135236978531, + 0.007835841737687588, + 0.005618540570139885, + 0.009389706887304783, + 0.006586936768144369, + 0.003952373284846544, + 0.004018007777631283, + 0.01482448447495699, + 0.01021458487957716, + 0.008898556232452393, + 0.019384808838367462, + 0.01463173609226942, + 0.016935529187321663, + 0.00907838810235262, + 0.014261538162827492, + 0.005473202560096979, + 0.005357119254767895, + 0.006178709678351879, + 0.007627047598361969, + 0.008260738104581833, + 0.008492509834468365, + 0.00374888745136559, + 0.003805031068623066, + 0.016782246530056, + 0.013789358548820019, + 0.007415871601551771, + 0.00825630221515894, + 0.00940023921430111, + 0.03313431516289711, + 0.007469098549336195, + 0.01674027182161808, + 0.052764132618904114, + 0.10792253166437149, + 0.0339001789689064, + 0.02783193811774254, + 0.04298444464802742, + 0.07631144672632217, + 0.05459997057914734, + 0.01921570859849453, + 0.00220234552398324, + 0.002681838348507881, + 0.0031530007254332304, + 0.001894535031169653, + 0.003227296518161893, + 0.0025764587335288525, + 0.0021890050265938044, + 0.002015351550653577, + 0.0028018823359161615, + 0.0015112559776753187, + 0.002069065347313881, + 0.004950169939547777, + 0.0021630763076245785, + 0.003121851710602641, + 0.0013328210916370153, + 0.0011133775115013123, + 0.0012326266150921583, + 0.0010609771125018597, + 0.001944263349287212, + 0.00266222539357841, + 0.0021385778672993183, + 0.001676044543273747, + 0.0014176093973219395, + 0.001315235043875873, + 0.0029541871044784784, + 0.0027013784274458885, + 0.0026141370180994272, + 0.005203795153647661, + 0.004015390761196613, + 0.0031080942135304213, + 0.008620248176157475, + 0.0032845032401382923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.020579971373081207 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up depicts a beautiful woman with long, honey-blonde hair falling in soft waves over her shoulders. She wears a vibrant royal blue jacket crafted from soft suede, the rich color contrasting beautifully with her fair complexion. Delicate, dangling silver earrings dangle from her ears, gently swaying and catching the light. The scene is bathed in the warm, diffuse glow of the golden hour, casting soft highlights on her high cheekbones and illuminating her clear blue eyes. She slowly turns her head to look directly into the camera, a gentle breeze ruffling a lock of her hair, a subtle, serene smile playing on her lips. In the background, a softly blurred cafe terrace, the blurred shapes of other patrons and the flicker of warm string lights create an atmosphere of dreamy, intimate, and effortlessly elegant elegance.", + "seed": 30405, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_3/index_8575_seed_30405_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.023775766883045436, + "vlm_logprob": -0.023775766883045436, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.023775766883045436, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010465974919497967 + ], + "opposite_logprob_tokens": [ + -4.8854660987854 + ], + "logprob_reward": -0.010465974919497967, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010465974919497967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3903154730796814, + 0.043205276131629944, + 0.08157120645046234, + 0.07128045707941055, + 0.04192693904042244, + 0.09753373265266418, + 0.03054730035364628, + 0.06564364582300186, + 0.15207135677337646, + 0.025904593989253044 + ], + "credit_map": [ + 0.0368032269179821, + 0.07238653302192688, + 0.03767561912536621, + 0.02793506719172001, + 0.08670985698699951, + 0.028504038229584694, + 0.048528317362070084, + 0.0517728216946125, + 0.003784825326874852, + 0.008170558139681816, + 0.0047699506394565105, + 0.005055466666817665, + 0.005630724132061005, + 0.006507642567157745, + 0.004178696312010288, + 0.005107407458126545, + 0.009945560246706009, + 0.025826310738921165, + 0.006303595844656229, + 0.005295029375702143, + 0.007435156498104334, + 0.005146797280758619, + 0.010062712244689465, + 0.011556049808859825, + 0.006596660241484642, + 0.01241876557469368, + 0.009860631078481674, + 0.007699944078922272, + 0.007121834438294172, + 0.007947450503706932, + 0.009663570672273636, + 0.00997160468250513, + 0.0036378747317939997, + 0.007829452864825726, + 0.005184723995625973, + 0.0037575990427285433, + 0.004378628917038441, + 0.004532287362962961, + 0.0040684533305466175, + 0.008537919260561466, + 0.008177842013537884, + 0.023690788075327873, + 0.007561574224382639, + 0.012383341789245605, + 0.007721360307186842, + 0.005633877590298653, + 0.011737052351236343, + 0.020627889782190323, + 0.0029617866966873407, + 0.0033945331815630198, + 0.00511344987899065, + 0.003144796472042799, + 0.0035456253681331873, + 0.002332835691049695, + 0.0043487888760864735, + 0.0057054846547544, + 0.0076040709391236305, + 0.01127910241484642, + 0.0073927934281528, + 0.005662241019308567, + 0.007496245205402374, + 0.008894446305930614, + 0.006769905798137188, + 0.01054484024643898, + 0.010907979682087898, + 0.02806178294122219, + 0.017567118629813194, + 0.02001533843576908, + 0.019014837220311165, + 0.007200876250863075, + 0.027073945850133896, + 0.022229477763175964, + 0.0019379031145945191, + 0.003166136797517538, + 0.003222721628844738, + 0.0036489120684564114, + 0.0049152495339512825, + 0.0018051870865747333, + 0.004957709927111864, + 0.002250773599371314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.010465974919497967 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635873481631279 + ], + "opposite_logprob_tokens": [ + -3.6513588428497314 + ], + "logprob_reward": -0.02635873481631279, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635873481631279, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12305872142314911, + 0.0684291422367096, + 0.45639681816101074, + 0.05538143962621689, + 0.02793748863041401, + 0.02844817005097866, + 0.0898318737745285, + 0.05035846680402756, + 0.07358185946941376, + 0.026575980708003044 + ], + "credit_map": [ + 0.007342812605202198, + 0.01285287644714117, + 0.009384822100400925, + 0.016992174088954926, + 0.030925264582037926, + 0.009548363275825977, + 0.01960575580596924, + 0.016406649723649025, + 0.005409299395978451, + 0.01221874076873064, + 0.005252273753285408, + 0.007270044181495905, + 0.017512435093522072, + 0.0072966986335814, + 0.005912661552429199, + 0.007556984666734934, + 0.045273225754499435, + 0.17114406824111938, + 0.03038611263036728, + 0.031100858002901077, + 0.03673207014799118, + 0.02243061549961567, + 0.026872707530856133, + 0.09245713800191879, + 0.0054138475097715855, + 0.00855861697345972, + 0.0071070087142288685, + 0.005412728525698185, + 0.006469774059951305, + 0.005592414643615484, + 0.00680357264354825, + 0.010023476555943489, + 0.0038411214482039213, + 0.0056272889487445354, + 0.0025150475557893515, + 0.0037023855838924646, + 0.0028849001973867416, + 0.003254522103816271, + 0.003293256741017103, + 0.0028189667500555515, + 0.00276850163936615, + 0.00436973012983799, + 0.0023196672555059195, + 0.0035295002162456512, + 0.0044104973785579205, + 0.002845539478585124, + 0.004056231584399939, + 0.004148503765463829, + 0.00825914740562439, + 0.007255994714796543, + 0.03071575053036213, + 0.006452040746808052, + 0.007163969799876213, + 0.006360561586916447, + 0.0059804790653288364, + 0.017643926665186882, + 0.006937792524695396, + 0.00652295770123601, + 0.005686299875378609, + 0.006014682352542877, + 0.005726264324039221, + 0.007950304076075554, + 0.005640710238367319, + 0.005879453383386135, + 0.005105988122522831, + 0.014991646632552147, + 0.009707077406346798, + 0.009258569218218327, + 0.007207636255770922, + 0.003329063765704632, + 0.009840806946158409, + 0.014141077175736427, + 0.0021536981221288443, + 0.003677915781736374, + 0.0033071294892579317, + 0.003549152286723256, + 0.005671216640621424, + 0.0023843273520469666, + 0.0036826226860284805, + 0.00214991788379848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.02635873481631279 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029804570600390434 + ], + "opposite_logprob_tokens": [ + -3.5298044681549072 + ], + "logprob_reward": -0.029804570600390434, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029804570600390434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1121104508638382, + 0.09260423481464386, + 0.514186680316925, + 0.03813411667943001, + 0.03991840034723282, + 0.03626193851232529, + 0.034309376031160355, + 0.031934238970279694, + 0.07008321583271027, + 0.030457371845841408 + ], + "credit_map": [ + 0.010734145529568195, + 0.016691340133547783, + 0.014277680777013302, + 0.01517721451818943, + 0.017148587852716446, + 0.009164909832179546, + 0.014006529934704304, + 0.014910046011209488, + 0.009922030381858349, + 0.01230592466890812, + 0.0070621236227452755, + 0.005872156471014023, + 0.02920716255903244, + 0.011378677561879158, + 0.007125212345272303, + 0.009730945341289043, + 0.05893096327781677, + 0.17939329147338867, + 0.03934862092137337, + 0.04074108600616455, + 0.054932259023189545, + 0.023197539150714874, + 0.05311046168208122, + 0.06453240662813187, + 0.004077936988323927, + 0.0067025660537183285, + 0.0037547401152551174, + 0.004233640618622303, + 0.0034820891451090574, + 0.00425902521237731, + 0.005224446300417185, + 0.0063996692188084126, + 0.005903674755245447, + 0.005648181773722172, + 0.005065214820206165, + 0.003051670268177986, + 0.0033027837052941322, + 0.005235546268522739, + 0.003972217906266451, + 0.007739109452813864, + 0.0029976097866892815, + 0.008954965509474277, + 0.0023204260505735874, + 0.004943557549268007, + 0.004180667456239462, + 0.0025784671306610107, + 0.004030800424516201, + 0.006255444139242172, + 0.0033651019912213087, + 0.003415337996557355, + 0.008542786352336407, + 0.003125267568975687, + 0.004445102531462908, + 0.0035129955504089594, + 0.003537055803462863, + 0.004365724511444569, + 0.004562073387205601, + 0.0034902282059192657, + 0.003770935581997037, + 0.0034725877922028303, + 0.0034616540651768446, + 0.005467918701469898, + 0.003556488547474146, + 0.0041523524560034275, + 0.0055056484416127205, + 0.014768498949706554, + 0.008549416437745094, + 0.010284175164997578, + 0.006230889353901148, + 0.003391752252355218, + 0.008833995088934898, + 0.012518844567239285, + 0.003335454035550356, + 0.0035782435443252325, + 0.004171251319348812, + 0.004142737947404385, + 0.006171175744384527, + 0.002440460491925478, + 0.003685565199702978, + 0.0029324847273528576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.029804570600390434 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048635587096214294 + ], + "opposite_logprob_tokens": [ + -3.048635482788086 + ], + "logprob_reward": -0.048635587096214294, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048635587096214294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29173004627227783, + 0.06432052701711655, + 0.10777763277292252, + 0.07703091204166412, + 0.033459682017564774, + 0.05931578204035759, + 0.2346770465373993, + 0.021499738097190857, + 0.043793778866529465, + 0.0663948804140091 + ], + "credit_map": [ + 0.024052415043115616, + 0.031026486307382584, + 0.028018662706017494, + 0.03483317047357559, + 0.03754367679357529, + 0.037079378962516785, + 0.04829034581780434, + 0.050885893404483795, + 0.004931435454636812, + 0.012043111957609653, + 0.008500673808157444, + 0.007804562337696552, + 0.007345006335526705, + 0.006562305148690939, + 0.008326695300638676, + 0.008806740865111351, + 0.009423988871276379, + 0.019081154838204384, + 0.011103792116045952, + 0.007439386565238237, + 0.00849555991590023, + 0.010877542197704315, + 0.011534090153872967, + 0.02982211858034134, + 0.006437445525079966, + 0.011290399357676506, + 0.008715256117284298, + 0.008884393610060215, + 0.006699758116155863, + 0.007090351078659296, + 0.007974622771143913, + 0.019938690587878227, + 0.0030440171249210835, + 0.006394779775291681, + 0.005107651464641094, + 0.004008743911981583, + 0.003547853557392955, + 0.0023793319705873728, + 0.003903444390743971, + 0.005073860753327608, + 0.005693864077329636, + 0.012154710479080677, + 0.004419713281095028, + 0.007265869528055191, + 0.005537677090615034, + 0.004118948709219694, + 0.008239394053816795, + 0.011885602958500385, + 0.01898909918963909, + 0.012387877330183983, + 0.0934552550315857, + 0.01377886813133955, + 0.006925874389708042, + 0.012433133088052273, + 0.011823388747870922, + 0.0648835301399231, + 0.002368363318964839, + 0.0036729732528328896, + 0.0019493295112624764, + 0.0024285344406962395, + 0.0024911260697990656, + 0.0021807358134537935, + 0.0024829404428601265, + 0.003925734665244818, + 0.004549057688564062, + 0.0078299380838871, + 0.005216628313064575, + 0.006210798397660255, + 0.004292109981179237, + 0.003163197310641408, + 0.005600695498287678, + 0.006931350566446781, + 0.007359776645898819, + 0.008795557543635368, + 0.007914874702692032, + 0.010117387399077415, + 0.009066320955753326, + 0.004484780132770538, + 0.012715503573417664, + 0.005940678995102644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.048635587096214294 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036139669828116894 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036139669828116894, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036139669828116894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11274507641792297, + 0.12719878554344177, + 0.11486077308654785, + 0.1259213089942932, + 0.03321275860071182, + 0.04183289408683777, + 0.2686106264591217, + 0.08670879900455475, + 0.02879512496292591, + 0.06011388078331947 + ], + "credit_map": [ + 0.010978960432112217, + 0.013062356039881706, + 0.012302103452384472, + 0.01577838510274887, + 0.01748226210474968, + 0.011888016946613789, + 0.016091804951429367, + 0.015161173418164253, + 0.011792090721428394, + 0.012571493163704872, + 0.014216836541891098, + 0.009334779344499111, + 0.023141391575336456, + 0.024043196812272072, + 0.010664134286344051, + 0.021434856578707695, + 0.013099932111799717, + 0.02366401068866253, + 0.009427325800061226, + 0.008426805026829243, + 0.009775201790034771, + 0.007106588687747717, + 0.015763111412525177, + 0.027597786858677864, + 0.015264171175658703, + 0.018281297758221626, + 0.018341895192861557, + 0.01253575086593628, + 0.012558862566947937, + 0.017352836206555367, + 0.016675425693392754, + 0.01491105742752552, + 0.004836099687963724, + 0.003956188447773457, + 0.004389727022498846, + 0.0035754763521254063, + 0.0030929113272577524, + 0.005122391972690821, + 0.0029664505273103714, + 0.005273510701954365, + 0.00495643075555563, + 0.004074687138199806, + 0.003517289413139224, + 0.0037825615145266056, + 0.002845063107088208, + 0.0052210185676813126, + 0.006232904735952616, + 0.011202932335436344, + 0.02168692834675312, + 0.01304200105369091, + 0.10583898425102234, + 0.014996754005551338, + 0.007536528632044792, + 0.013995506800711155, + 0.013934902846813202, + 0.07757899910211563, + 0.013499638997018337, + 0.00929821003228426, + 0.00919228233397007, + 0.009495350532233715, + 0.009341743774712086, + 0.014666477218270302, + 0.01001801434904337, + 0.011197077110409737, + 0.002428621519356966, + 0.005102827213704586, + 0.004244809038937092, + 0.003960600588470697, + 0.003176737343892455, + 0.001666581374593079, + 0.004618061240762472, + 0.003596881404519081, + 0.006765393540263176, + 0.007211924996227026, + 0.009178339503705502, + 0.0056358808651566505, + 0.00943280104547739, + 0.006955648306757212, + 0.008750289678573608, + 0.0061835963279008865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 3 + ], + "selected_reward": -0.0036139669828116894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up depicts a beautiful woman with long, honey-blonde hair falling in soft waves over her shoulders. She wears a vibrant royal blue jacket crafted from soft suede, the rich color contrasting beautifully with her fair complexion. Delicate, dangling silver earrings dangle from her ears, gently swaying and catching the light. The scene is bathed in the warm, diffuse glow of the golden hour, casting soft highlights on her high cheekbones and illuminating her clear blue eyes. She slowly turns her head to look directly into the camera, a gentle breeze ruffling a lock of her hair, a subtle, serene smile playing on her lips. In the background, a softly blurred cafe terrace, the blurred shapes of other patrons and the flicker of warm string lights create an atmosphere of dreamy, intimate, and effortlessly elegant elegance.", + "seed": 30406, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_3/index_8575_seed_30406_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.020852826163172723, + "vlm_logprob": -0.020852826163172723, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.020852826163172723, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0049200416542589664 + ], + "opposite_logprob_tokens": [ + -6.37992000579834 + ], + "logprob_reward": -0.0049200416542589664, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0049200416542589664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11045295745134354, + 0.0684821680188179, + 0.13247078657150269, + 0.09356944262981415, + 0.04455089941620827, + 0.09878630936145782, + 0.06799818575382233, + 0.07957109808921814, + 0.0730157196521759, + 0.23110246658325195 + ], + "credit_map": [ + 0.015908444300293922, + 0.010973174124956131, + 0.015199380926787853, + 0.009504399262368679, + 0.012208541855216026, + 0.010482151061296463, + 0.01886480115354061, + 0.01731206104159355, + 0.009747267700731754, + 0.00632754759863019, + 0.010027027688920498, + 0.008509814739227295, + 0.007879108190536499, + 0.007757679559290409, + 0.007704839576035738, + 0.01052887924015522, + 0.008880321867763996, + 0.03902942314743996, + 0.007185352500528097, + 0.017649928107857704, + 0.018255334347486496, + 0.008975065313279629, + 0.009095927700400352, + 0.02339942939579487, + 0.007762324996292591, + 0.010600380599498749, + 0.007654711604118347, + 0.012396852485835552, + 0.01743141934275627, + 0.011878526769578457, + 0.011749206110835075, + 0.014096023514866829, + 0.0029336027801036835, + 0.0049400669522583485, + 0.005926607642322779, + 0.005561302416026592, + 0.005697572138160467, + 0.004912009462714195, + 0.008524131961166859, + 0.006055603269487619, + 0.010877657681703568, + 0.01091212872415781, + 0.008376584388315678, + 0.010906010866165161, + 0.008235953748226166, + 0.015106314793229103, + 0.01882178895175457, + 0.015549862757325172, + 0.006719059776514769, + 0.006645792629569769, + 0.007788782007992268, + 0.00713327806442976, + 0.0059078168123960495, + 0.009882695972919464, + 0.01403067447245121, + 0.00989008229225874, + 0.008972609415650368, + 0.00831898208707571, + 0.009899199940264225, + 0.012249669060111046, + 0.009368318133056164, + 0.007610889617353678, + 0.010809030383825302, + 0.012342398054897785, + 0.0074393367394804955, + 0.01111816056072712, + 0.008157556876540184, + 0.007953140884637833, + 0.0065963976085186005, + 0.015412452630698681, + 0.0064657763577997684, + 0.009872895665466785, + 0.020876163616776466, + 0.025393806397914886, + 0.09102016687393188, + 0.012735928408801556, + 0.012462803162634373, + 0.009732386097311974, + 0.007624424993991852, + 0.051256779581308365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.0049200416542589664 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330239862203598 + ], + "opposite_logprob_tokens": [ + -3.7733023166656494 + ], + "logprob_reward": -0.02330239862203598, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330239862203598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12634950876235962, + 0.10819932818412781, + 0.12698471546173096, + 0.08808573335409164, + 0.051511432975530624, + 0.04781712591648102, + 0.11016774922609329, + 0.14202715456485748, + 0.09002082794904709, + 0.10883643478155136 + ], + "credit_map": [ + 0.020748458802700043, + 0.024098791182041168, + 0.019092878326773643, + 0.012235946953296661, + 0.01423635333776474, + 0.010969328694045544, + 0.010639689862728119, + 0.01432806346565485, + 0.012339618057012558, + 0.011244164779782295, + 0.012082928791642189, + 0.026908576488494873, + 0.013761765323579311, + 0.01093343086540699, + 0.01020098477602005, + 0.010727860033512115, + 0.009754734113812447, + 0.020480403676629066, + 0.01634112000465393, + 0.02203298546373844, + 0.014805588871240616, + 0.012555002234876156, + 0.018895380198955536, + 0.012119502760469913, + 0.006037687882781029, + 0.011079400777816772, + 0.008828367106616497, + 0.01843271031975746, + 0.009530678391456604, + 0.007058555260300636, + 0.013895392417907715, + 0.013222942128777504, + 0.004646957851946354, + 0.00583407049998641, + 0.005842028651386499, + 0.006354845594614744, + 0.005664229858666658, + 0.006981787737458944, + 0.008264502510428429, + 0.007923007942736149, + 0.005626706872135401, + 0.004104418680071831, + 0.005001452285796404, + 0.0043931808322668076, + 0.0036162922624498606, + 0.006477758288383484, + 0.010290221311151981, + 0.008307094685733318, + 0.008290006779134274, + 0.009985249489545822, + 0.008764452300965786, + 0.014343553222715855, + 0.009625954553484917, + 0.009857778437435627, + 0.028514660894870758, + 0.020786093547940254, + 0.008809284307062626, + 0.012871332466602325, + 0.014302724972367287, + 0.023364398628473282, + 0.014402279630303383, + 0.01929902844130993, + 0.02777905948460102, + 0.021199043840169907, + 0.007357683498412371, + 0.010552844032645226, + 0.009754316881299019, + 0.013016480952501297, + 0.008967108093202114, + 0.009448275901377201, + 0.018343539908528328, + 0.012580586597323418, + 0.011197970248758793, + 0.015100676566362381, + 0.036741193383932114, + 0.009142409078776836, + 0.008356214500963688, + 0.008540751412510872, + 0.009288966655731201, + 0.010468248277902603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.02330239862203598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029806768521666527 + ], + "opposite_logprob_tokens": [ + -3.5298068523406982 + ], + "logprob_reward": -0.029806768521666527, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029806768521666527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17422327399253845, + 0.11170395463705063, + 0.2668129503726959, + 0.048571228981018066, + 0.05136210098862648, + 0.055088091641664505, + 0.04045211151242256, + 0.09755264967679977, + 0.03141067922115326, + 0.12282294780015945 + ], + "credit_map": [ + 0.020822471007704735, + 0.0334516279399395, + 0.02788681350648403, + 0.01647031120955944, + 0.019750116392970085, + 0.019058579578995705, + 0.019786052405834198, + 0.01699729636311531, + 0.010958070866763592, + 0.011711817234754562, + 0.00952078029513359, + 0.021262353286147118, + 0.017024151980876923, + 0.010083393193781376, + 0.015776872634887695, + 0.015366500243544579, + 0.023048022761940956, + 0.06089868023991585, + 0.020574159920215607, + 0.02849012054502964, + 0.03150518238544464, + 0.028421485796570778, + 0.02245817333459854, + 0.05141713097691536, + 0.004086271394044161, + 0.005070898216217756, + 0.004126166924834251, + 0.004728792235255241, + 0.006516204681247473, + 0.005381960421800613, + 0.01132988277822733, + 0.007331048604100943, + 0.004846304189413786, + 0.004694093018770218, + 0.003922143019735813, + 0.008123152889311314, + 0.006599439773708582, + 0.004257336258888245, + 0.009174306876957417, + 0.009745321236550808, + 0.005920237395912409, + 0.006457951385527849, + 0.0048407395370304585, + 0.007356690708547831, + 0.005768296308815479, + 0.007436190731823444, + 0.00891362875699997, + 0.008394348435103893, + 0.003217139979824424, + 0.005907749757170677, + 0.0037777419202029705, + 0.006432950962334871, + 0.0038159748073667288, + 0.005152115132659674, + 0.006057607010006905, + 0.006090829614549875, + 0.011942876502871513, + 0.01320625189691782, + 0.011182530783116817, + 0.014598165638744831, + 0.013076895847916603, + 0.008392924442887306, + 0.009978254325687885, + 0.015174749307334423, + 0.0032200736459344625, + 0.004437004681676626, + 0.0037272644694894552, + 0.005076056346297264, + 0.0038405980449169874, + 0.0036304257810115814, + 0.003228717250749469, + 0.004250538069754839, + 0.01075239758938551, + 0.012180403806269169, + 0.04675859212875366, + 0.007126708514988422, + 0.006848142482340336, + 0.005662680137902498, + 0.005276246927678585, + 0.028217770159244537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.029806768521666527 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304330423474312 + ], + "opposite_logprob_tokens": [ + -3.168043375015259 + ], + "logprob_reward": -0.04304330423474312, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304330423474312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1993805468082428, + 0.08441876620054245, + 0.12327568978071213, + 0.07498298585414886, + 0.04760162904858589, + 0.05604887381196022, + 0.10233377665281296, + 0.051695261150598526, + 0.06223451346158981, + 0.19802792370319366 + ], + "credit_map": [ + 0.03403006121516228, + 0.01991528458893299, + 0.03375978395342827, + 0.016271518543362617, + 0.03340025991201401, + 0.019293319433927536, + 0.02166125923395157, + 0.02104906737804413, + 0.010799427516758442, + 0.012163815088570118, + 0.008251340128481388, + 0.011491160839796066, + 0.01632680371403694, + 0.008848928846418858, + 0.009214699268341064, + 0.007322587072849274, + 0.009701577946543694, + 0.028838003054261208, + 0.015853362157940865, + 0.0162834320217371, + 0.01601775363087654, + 0.008950863033533096, + 0.008952596224844456, + 0.01867811381816864, + 0.006908303126692772, + 0.010864553973078728, + 0.009629934094846249, + 0.011129696853458881, + 0.012643939815461636, + 0.005174947902560234, + 0.007180708460509777, + 0.011450905352830887, + 0.004655658267438412, + 0.00899174902588129, + 0.005210931412875652, + 0.006055840756744146, + 0.005666465498507023, + 0.006247607991099358, + 0.004629865754395723, + 0.00614351499825716, + 0.005871577654033899, + 0.0043977173045277596, + 0.005079174414277077, + 0.006680530495941639, + 0.005529958289116621, + 0.010469398461282253, + 0.009063435718417168, + 0.008957082405686378, + 0.008736604824662209, + 0.009884096682071686, + 0.012630760669708252, + 0.015234329737722874, + 0.009780634194612503, + 0.016916602849960327, + 0.014269703067839146, + 0.014881040900945663, + 0.006617844570428133, + 0.006720830220729113, + 0.005067090969532728, + 0.0075429463759064674, + 0.007505930028855801, + 0.005069343373179436, + 0.005158493760973215, + 0.008012783713638783, + 0.006931568030267954, + 0.010360917076468468, + 0.006515787448734045, + 0.00972823891788721, + 0.006723925471305847, + 0.0073332092724740505, + 0.005698330234736204, + 0.008942541666328907, + 0.024236837401986122, + 0.03319086879491806, + 0.01919112540781498, + 0.024897608906030655, + 0.02435634844005108, + 0.028080865740776062, + 0.012087495066225529, + 0.03198679909110069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.04304330423474312 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031916177831590176 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031916177831590176, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031916177831590176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11751142144203186, + 0.2201647311449051, + 0.2569938898086548, + 0.0446946881711483, + 0.06990018486976624, + 0.1253882646560669, + 0.06405304372310638, + 0.031689923256635666, + 0.03262163698673248, + 0.03698214516043663 + ], + "credit_map": [ + 0.022113941609859467, + 0.022782685235142708, + 0.020293107256293297, + 0.013647183775901794, + 0.012003635056316853, + 0.008030704222619534, + 0.009240439161658287, + 0.009399725124239922, + 0.03305194899439812, + 0.015800613909959793, + 0.02231941744685173, + 0.05396994575858116, + 0.021747928112745285, + 0.039719101041555405, + 0.01192671712487936, + 0.021629059687256813, + 0.02235386334359646, + 0.0637931302189827, + 0.01763368584215641, + 0.024256469681859016, + 0.028864001855254173, + 0.028536060824990273, + 0.019227825105190277, + 0.052328843623399734, + 0.003902911674231291, + 0.0045654368586838245, + 0.005203793291002512, + 0.004927685018628836, + 0.004444199148565531, + 0.006535807158797979, + 0.008364420384168625, + 0.00675043324008584, + 0.00730241509154439, + 0.005456802435219288, + 0.008854481391608715, + 0.007118777371942997, + 0.007569341454654932, + 0.014624892733991146, + 0.008179744705557823, + 0.010793727822601795, + 0.013531629927456379, + 0.009972200728952885, + 0.0090753473341465, + 0.014972403645515442, + 0.011750195175409317, + 0.02343527227640152, + 0.024024473503232002, + 0.018626749515533447, + 0.005298455245792866, + 0.004738961812108755, + 0.004576863255351782, + 0.007734354585409164, + 0.004819129593670368, + 0.010993392206728458, + 0.014994404278695583, + 0.010897486470639706, + 0.004092094022780657, + 0.0033059834968298674, + 0.00246605952270329, + 0.005185911897569895, + 0.003796130884438753, + 0.004171306733042002, + 0.004283608868718147, + 0.004388829227536917, + 0.0032972851768136024, + 0.003621347714215517, + 0.0034023080952465534, + 0.00384184205904603, + 0.002891028532758355, + 0.00602394575253129, + 0.004270828794687986, + 0.005273050162941217, + 0.003398645669221878, + 0.0039796349592506886, + 0.013870489783585072, + 0.00235344422981143, + 0.0020271039102226496, + 0.002001467626541853, + 0.0010781630408018827, + 0.008273198269307613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.0031916177831590176 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up depicts a beautiful woman with long, honey-blonde hair falling in soft waves over her shoulders. She wears a vibrant royal blue jacket crafted from soft suede, the rich color contrasting beautifully with her fair complexion. Delicate, dangling silver earrings dangle from her ears, gently swaying and catching the light. The scene is bathed in the warm, diffuse glow of the golden hour, casting soft highlights on her high cheekbones and illuminating her clear blue eyes. She slowly turns her head to look directly into the camera, a gentle breeze ruffling a lock of her hair, a subtle, serene smile playing on her lips. In the background, a softly blurred cafe terrace, the blurred shapes of other patrons and the flicker of warm string lights create an atmosphere of dreamy, intimate, and effortlessly elegant elegance.", + "seed": 30407, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_3/index_8575_seed_30407_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01618682760745287, + "vlm_logprob": -0.01618682760745287, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.01618682760745287, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005021816119551659 + ], + "opposite_logprob_tokens": [ + -7.005022048950195 + ], + "logprob_reward": -0.005021816119551659, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.005021816119551659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09543950855731964, + 0.2181665003299713, + 0.04259217530488968, + 0.025748705491423607, + 0.044120173901319504, + 0.04759129509329796, + 0.06293462961912155, + 0.11963259428739548, + 0.315658837556839, + 0.02811560407280922 + ], + "credit_map": [ + 0.018656250089406967, + 0.011192776262760162, + 0.018004920333623886, + 0.007889199070632458, + 0.01113860122859478, + 0.013818781822919846, + 0.00791684165596962, + 0.0068221441470086575, + 0.07254184782505035, + 0.01323117408901453, + 0.05022382736206055, + 0.016302308067679405, + 0.013819565996527672, + 0.03082616627216339, + 0.009228630922734737, + 0.0119929825887084, + 0.005855189170688391, + 0.003809097222983837, + 0.004073216579854488, + 0.008648277260363102, + 0.003988702781498432, + 0.0050122542306780815, + 0.004569708369672298, + 0.006635731551796198, + 0.003199736587703228, + 0.0023679451551288366, + 0.003336593508720398, + 0.0036847777664661407, + 0.003959923982620239, + 0.0031165755353868008, + 0.0023618126288056374, + 0.003721341025084257, + 0.006283922120928764, + 0.004609335213899612, + 0.005563477054238319, + 0.011793923564255238, + 0.0043716332875192165, + 0.003132793353870511, + 0.004395953845232725, + 0.003969137091189623, + 0.007439947687089443, + 0.006343539338558912, + 0.00827733427286148, + 0.007897821255028248, + 0.00470117200165987, + 0.0032248019706457853, + 0.0037898679729551077, + 0.005916808266192675, + 0.008121681399643421, + 0.00845713634043932, + 0.006643746979534626, + 0.008694625459611416, + 0.00731092132627964, + 0.008328454568982124, + 0.008322225883603096, + 0.007055839989334345, + 0.02043580822646618, + 0.009995974600315094, + 0.025349225848913193, + 0.013291051611304283, + 0.01562187448143959, + 0.00955224223434925, + 0.010148424655199051, + 0.015237988904118538, + 0.06929653137922287, + 0.031938955187797546, + 0.10501708090305328, + 0.027625903487205505, + 0.03419604152441025, + 0.016353517770767212, + 0.015209365636110306, + 0.016021475195884705, + 0.003958250395953655, + 0.002263027476146817, + 0.004223885014653206, + 0.0030275569297373295, + 0.0039680334739387035, + 0.006042587570846081, + 0.0017756870947778225, + 0.0028565777465701103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.005021816119551659 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018199361860752106 + ], + "opposite_logprob_tokens": [ + -4.018199443817139 + ], + "logprob_reward": -0.018199361860752106, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018199361860752106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1374579519033432, + 0.24281616508960724, + 0.08162133395671844, + 0.06072160601615906, + 0.07630223035812378, + 0.0943276435136795, + 0.09984225034713745, + 0.06688790023326874, + 0.08683118969202042, + 0.05319175496697426 + ], + "credit_map": [ + 0.022841665893793106, + 0.014230474829673767, + 0.030379479750990868, + 0.012630070559680462, + 0.012897486798465252, + 0.02086668275296688, + 0.00967935100197792, + 0.013932726345956326, + 0.07032670825719833, + 0.01707204431295395, + 0.0486784391105175, + 0.02945011481642723, + 0.019391769543290138, + 0.018006324768066406, + 0.01470054779201746, + 0.025190196931362152, + 0.008592342026531696, + 0.009176532737910748, + 0.007372897583991289, + 0.01943173073232174, + 0.008093249052762985, + 0.010495498776435852, + 0.009273440577089787, + 0.009185639210045338, + 0.009934034198522568, + 0.006958669517189264, + 0.007166864350438118, + 0.006876898929476738, + 0.009060384705662727, + 0.005263811443001032, + 0.006157069467008114, + 0.009303871542215347, + 0.012334533035755157, + 0.00964362919330597, + 0.01003436278551817, + 0.014748737215995789, + 0.006868934724479914, + 0.008107651956379414, + 0.0064983186312019825, + 0.008066062815487385, + 0.012945515103638172, + 0.010567179881036282, + 0.012645474635064602, + 0.01185299176722765, + 0.008425313048064709, + 0.010982375591993332, + 0.013630497269332409, + 0.013278293423354626, + 0.01397627778351307, + 0.010666522197425365, + 0.013088037259876728, + 0.010911398567259312, + 0.012719342485070229, + 0.010106822475790977, + 0.014543152414262295, + 0.013830689713358879, + 0.007706229574978352, + 0.0066128503531217575, + 0.009559019468724728, + 0.006771315820515156, + 0.01065007783472538, + 0.008361825719475746, + 0.008503079414367676, + 0.008723502047359943, + 0.01703542470932007, + 0.007080163806676865, + 0.025270646438002586, + 0.00980405230075121, + 0.010381351225078106, + 0.007633538451045752, + 0.0041221980936825275, + 0.005503802094608545, + 0.008032682351768017, + 0.0040033492259681225, + 0.007777428720146418, + 0.00559168541803956, + 0.008363683708012104, + 0.00627152482047677, + 0.007411539554595947, + 0.005739855580031872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.018199361860752106 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290520533919334 + ], + "opposite_logprob_tokens": [ + -3.7732906341552734 + ], + "logprob_reward": -0.023290520533919334, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290520533919334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07399171590805054, + 0.441658616065979, + 0.057017721235752106, + 0.036068279296159744, + 0.24366708099842072, + 0.03126318007707596, + 0.032800279557704926, + 0.031822409480810165, + 0.01954587735235691, + 0.03216481953859329 + ], + "credit_map": [ + 0.00836926233023405, + 0.007531390059739351, + 0.0106138801202178, + 0.008375406265258789, + 0.01172992680221796, + 0.013036723248660564, + 0.007295073941349983, + 0.0070400540716946125, + 0.12150591611862183, + 0.03101302497088909, + 0.0754234790802002, + 0.044187966734170914, + 0.03822530433535576, + 0.048585399985313416, + 0.02614551968872547, + 0.05657198652625084, + 0.006326963659375906, + 0.005595621187239885, + 0.004747010767459869, + 0.009237458929419518, + 0.005676630884408951, + 0.007736171595752239, + 0.00394913274794817, + 0.01374872587621212, + 0.0055705164559185505, + 0.0037303008139133453, + 0.004173495341092348, + 0.004993905313313007, + 0.0029255172703415155, + 0.003940576687455177, + 0.0027341549284756184, + 0.007999812252819538, + 0.03801732882857323, + 0.020387986674904823, + 0.01869596354663372, + 0.05602194741368294, + 0.03624679520726204, + 0.023086316883563995, + 0.03414962813258171, + 0.017061099410057068, + 0.004141984041780233, + 0.005395975429564714, + 0.005100566893815994, + 0.005667936988174915, + 0.0038757952861487865, + 0.001882406766526401, + 0.0028979077469557524, + 0.002300604712218046, + 0.004902651067823172, + 0.003827771870419383, + 0.006211527623236179, + 0.005235053598880768, + 0.0037907131481915712, + 0.0021197530440986156, + 0.0038931763265281916, + 0.0028196312487125397, + 0.003908043727278709, + 0.003105832962319255, + 0.0040893699042499065, + 0.006362390238791704, + 0.0032174810767173767, + 0.00405810447409749, + 0.0026827030815184116, + 0.004398481920361519, + 0.0033331066370010376, + 0.001852024463005364, + 0.003956872038543224, + 0.0034280105028301477, + 0.002427531173452735, + 0.0015094808768481016, + 0.0011554228840395808, + 0.0018834267975762486, + 0.005216597579419613, + 0.002648641588166356, + 0.004700465593487024, + 0.004865861963480711, + 0.003494108794257045, + 0.005982397124171257, + 0.0021785572171211243, + 0.003078186186030507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.023290520533919334 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978559397161007 + ], + "opposite_logprob_tokens": [ + -3.529785633087158 + ], + "logprob_reward": -0.02978559397161007, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978559397161007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19446824491024017, + 0.20200717449188232, + 0.16557592153549194, + 0.049814533442258835, + 0.036279525607824326, + 0.07385900616645813, + 0.05333757400512695, + 0.04858453944325447, + 0.10930059850215912, + 0.0667729452252388 + ], + "credit_map": [ + 0.02425461634993553, + 0.013257169164717197, + 0.04203564673662186, + 0.0273740254342556, + 0.02252095751464367, + 0.015556840226054192, + 0.02322438918054104, + 0.02624458260834217, + 0.03823205828666687, + 0.015362415462732315, + 0.0313420295715332, + 0.032037876546382904, + 0.02155894972383976, + 0.017533190548419952, + 0.018536120653152466, + 0.02740451693534851, + 0.017638439312577248, + 0.02003943920135498, + 0.01763986237347126, + 0.025855734944343567, + 0.016646871343255043, + 0.01584920473396778, + 0.020403288304805756, + 0.03150307014584541, + 0.005413530860096216, + 0.0073308139108121395, + 0.005986775737255812, + 0.008165767416357994, + 0.006529733072966337, + 0.004902146756649017, + 0.004241642076522112, + 0.007244120817631483, + 0.005269071087241173, + 0.0034521990455687046, + 0.003259081859141588, + 0.004996160045266151, + 0.004967355635017157, + 0.0053885397501289845, + 0.0033606227952986956, + 0.005586492829024792, + 0.01617257483303547, + 0.00796220637857914, + 0.02224068157374859, + 0.008232067339122295, + 0.005641802679747343, + 0.0036726542748510838, + 0.0039756749756634235, + 0.005961335729807615, + 0.005098548252135515, + 0.006834241095930338, + 0.007539601530879736, + 0.012048362754285336, + 0.0059428769163787365, + 0.004321723710745573, + 0.005382482893764973, + 0.0061697340570390224, + 0.008934621699154377, + 0.004228770267218351, + 0.008338374085724354, + 0.004980168770998716, + 0.007062215358018875, + 0.003848582273349166, + 0.005504663567990065, + 0.005687138997018337, + 0.024971455335617065, + 0.00804234016686678, + 0.03818957507610321, + 0.011413680389523506, + 0.011293025687336922, + 0.00573021499440074, + 0.004214882850646973, + 0.00544541422277689, + 0.007756481412798166, + 0.005045032594352961, + 0.008768185041844845, + 0.007320926524698734, + 0.010265546850860119, + 0.012729066424071789, + 0.007994950748980045, + 0.00689275236800313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.02978559397161007 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the subject a woman with long honey-blonde hair styled in soft waves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a vibrant royal blue jacket made of soft suede?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are delicate dangling silver earrings visible on the woman's ears?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman slowly turn her head to look directly into the camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a softly blurred cafe terrace with warm string lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636845551431179 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636845551431179, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636845551431179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2037828117609024, + 0.22135457396507263, + 0.08301173150539398, + 0.06504914909601212, + 0.07475020736455917, + 0.020355619490146637, + 0.026683570817112923, + 0.025239702314138412, + 0.049722589552402496, + 0.23005004227161407 + ], + "credit_map": [ + 0.036978185176849365, + 0.029673490673303604, + 0.031235065311193466, + 0.01995936594903469, + 0.021640267223119736, + 0.03683073818683624, + 0.01373404823243618, + 0.013731667771935463, + 0.061808232218027115, + 0.016253789886832237, + 0.04457232356071472, + 0.020893579348921776, + 0.017570341005921364, + 0.035255394876003265, + 0.013537850230932236, + 0.01146306935697794, + 0.009496290236711502, + 0.007656247355043888, + 0.009598040021955967, + 0.0119441207498312, + 0.010272610932588577, + 0.014852364547550678, + 0.007039819378405809, + 0.012152235954999924, + 0.00941909197717905, + 0.004934747237712145, + 0.007302910089492798, + 0.007580803241580725, + 0.005909723229706287, + 0.007849405519664288, + 0.006601523142307997, + 0.015450948849320412, + 0.011815926991403103, + 0.006590007804334164, + 0.006407248787581921, + 0.01978999748826027, + 0.008608974516391754, + 0.00673610670492053, + 0.008876672014594078, + 0.005925281904637814, + 0.0029952307231724262, + 0.0019208119483664632, + 0.0039000113029032946, + 0.002856253879144788, + 0.0018210156122222543, + 0.0022254129871726036, + 0.0018415023805573583, + 0.0027953796088695526, + 0.0035154419019818306, + 0.001826259889639914, + 0.0030474946834146976, + 0.003063665935769677, + 0.0029011189471930265, + 0.003565652295947075, + 0.004773743916302919, + 0.003990192897617817, + 0.0030132634565234184, + 0.003042965428903699, + 0.003230182221159339, + 0.004600939340889454, + 0.0033879645634442568, + 0.0032119189854711294, + 0.002745195524767041, + 0.0020072704646736383, + 0.009175533428788185, + 0.004317461978644133, + 0.013616192154586315, + 0.0055357967503368855, + 0.0070261345244944096, + 0.0049224477261304855, + 0.002206474542617798, + 0.0029225472826510668, + 0.03322234004735947, + 0.018292609602212906, + 0.0332450270652771, + 0.022595012560486794, + 0.02267397940158844, + 0.054353978484869, + 0.014517737552523613, + 0.03114934265613556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.004636845551431179 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_4/index_tensor([5726, 5726])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_4/index_tensor([5726, 5726])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9d6db4e831a67363bb68bfcaee0fd8812f4621bb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_4/index_tensor([5726, 5726])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a dramatic medium shot, set in a vast, dimly lit tournament arena, with the blurred silhouettes of a huge, cheering crowd in the background. The camera's focus is first pinpointed on a massive totem in the foreground, symbolizing defeat. Composed of rough, dark gray, almost black rock, its surface is riddled with cracks. A faint, flickering flame adorns its summit, and dark, solidified wax drips clutter its shaft. Suddenly, with a muffled, powerful crack, the totem's shaft violently shatters, exploding into countless sharp fragments and dust. At that moment, a blinding white flash instantly engulfs the frame, and the camera shifts focus with remarkable speed, moving from the dark, crumbling debris in the foreground to a smooth, decisive pull toward another totem in the background, symbolizing victory. This totem is crafted from a brilliant golden metal, its surface smooth and reflective, and its tip blazes with a powerful, blazing golden-white flame. The camera finally stops on this victory totem, its golden light illuminates the flying dust around it, forming a static picture full of power and glory. The whole scene, under the high-contrast light and shadow, creates a glorious atmosphere of tension, climaxes and ultimately the joy of victory.", + "seed": 3316, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_4/index_5726_seed_3316_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03428250560536981, + "vlm_logprob": -0.03428250560536981, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03428250560536981, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07131973654031754 + ], + "opposite_logprob_tokens": [ + -2.696319818496704 + ], + "logprob_reward": -0.07131973654031754, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07131973654031754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22683875262737274, + 0.1385868489742279, + 0.15797516703605652, + 0.07697130739688873, + 0.10639959573745728, + 0.028072400018572807, + 0.037031546235084534, + 0.04255950450897217, + 0.10344114154577255, + 0.08212366700172424 + ], + "credit_map": [ + 0.020125694572925568, + 0.016161123290657997, + 0.016694672405719757, + 0.06550469249486923, + 0.018553195521235466, + 0.04759816825389862, + 0.02157272957265377, + 0.020628491416573524, + 0.01379342656582594, + 0.012219295836985111, + 0.012451030313968658, + 0.0373850055038929, + 0.015955545008182526, + 0.012250369414687157, + 0.021752381697297096, + 0.012779793702065945, + 0.016967803239822388, + 0.020031116902828217, + 0.020377300679683685, + 0.028038937598466873, + 0.01960836909711361, + 0.014553189277648926, + 0.01879371702671051, + 0.019604742527008057, + 0.006987446919083595, + 0.012541957199573517, + 0.007685614284127951, + 0.01197522971779108, + 0.013494783081114292, + 0.008734418079257011, + 0.008376848883926868, + 0.007175012491643429, + 0.019295308738946915, + 0.011751611717045307, + 0.01231555175036192, + 0.013565300032496452, + 0.008963372558355331, + 0.020549261942505836, + 0.010080861859023571, + 0.009878333657979965, + 0.004144452977925539, + 0.0027191448025405407, + 0.0030789808370172977, + 0.003390407422557473, + 0.0021499418653547764, + 0.0033638423774391413, + 0.0051140738651156425, + 0.004111555404961109, + 0.004185847006738186, + 0.00318796094506979, + 0.009242775849997997, + 0.004459465388208628, + 0.0027504265308380127, + 0.003955933265388012, + 0.002523373579606414, + 0.0067257629707455635, + 0.005200579296797514, + 0.004104206338524818, + 0.008653867058455944, + 0.00494412612169981, + 0.003187761642038822, + 0.005620887503027916, + 0.003971962258219719, + 0.006876114755868912, + 0.011665190570056438, + 0.00942318420857191, + 0.01850237138569355, + 0.013453920371830463, + 0.008511335588991642, + 0.017290642485022545, + 0.011844469234347343, + 0.012750034220516682, + 0.015096130780875683, + 0.0046605453826487064, + 0.017192643135786057, + 0.007128069642931223, + 0.005488272290676832, + 0.012749651446938515, + 0.005981131922453642, + 0.013827229849994183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.07131973654031754 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088181287050247 + ], + "opposite_logprob_tokens": [ + -4.5110883712768555 + ], + "logprob_reward": -0.011088181287050247, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088181287050247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14879336953163147, + 0.09472688287496567, + 0.11396801471710205, + 0.3035306930541992, + 0.09130806475877762, + 0.053994469344615936, + 0.03515767306089401, + 0.05342785641551018, + 0.06231215223670006, + 0.04278083145618439 + ], + "credit_map": [ + 0.01984500139951706, + 0.017297690734267235, + 0.01557321660220623, + 0.024312419816851616, + 0.01522036548703909, + 0.025086721405386925, + 0.015676813200116158, + 0.015781134366989136, + 0.011831963434815407, + 0.012873616069555283, + 0.010750385001301765, + 0.011794758029282093, + 0.014851262792944908, + 0.00914054736495018, + 0.013623526319861412, + 0.009860831312835217, + 0.01015537977218628, + 0.014634430408477783, + 0.013085332699120045, + 0.019706683233380318, + 0.016911929473280907, + 0.013111202046275139, + 0.014171930029988289, + 0.01219114102423191, + 0.03563474118709564, + 0.03558586537837982, + 0.046929389238357544, + 0.03627631068229675, + 0.04922813922166824, + 0.03954382240772247, + 0.02607521414756775, + 0.0342571958899498, + 0.014914494939148426, + 0.01257379725575447, + 0.012328960001468658, + 0.013433084823191166, + 0.008529112674295902, + 0.012782013043761253, + 0.007437401916831732, + 0.009309201501309872, + 0.009539567865431309, + 0.005565539468079805, + 0.00427826726809144, + 0.004832072649151087, + 0.004520799499005079, + 0.012214254587888718, + 0.008305003866553307, + 0.004738967400044203, + 0.0036985778715461493, + 0.0051143234595656395, + 0.006324430461972952, + 0.004617336206138134, + 0.003997719846665859, + 0.004336378071457148, + 0.0032874438911676407, + 0.003781461389735341, + 0.008458009921014309, + 0.00479091377928853, + 0.009557831101119518, + 0.004631498828530312, + 0.004033055622130632, + 0.008151615038514137, + 0.004087685141712427, + 0.009717247448861599, + 0.006584486924111843, + 0.0064317332580685616, + 0.009630872868001461, + 0.00831384677439928, + 0.006376716308295727, + 0.009373863227665424, + 0.008290504105389118, + 0.007310128305107355, + 0.0065337251871824265, + 0.00361404987052083, + 0.007228014059364796, + 0.00477599585428834, + 0.004252202808856964, + 0.004871689714491367, + 0.004385335836559534, + 0.0071198176592588425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.011088181287050247 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948459263890982 + ], + "opposite_logprob_tokens": [ + -5.130948543548584 + ], + "logprob_reward": -0.005948459263890982, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948459263890982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08796893805265427, + 0.10721834003925323, + 0.13951165974140167, + 0.22380360960960388, + 0.19442598521709442, + 0.04831784963607788, + 0.0666482001543045, + 0.04334168881177902, + 0.042394157499074936, + 0.04636954516172409 + ], + "credit_map": [ + 0.009783458895981312, + 0.011353793554008007, + 0.007446078583598137, + 0.012354076839983463, + 0.014681891538202763, + 0.011012393981218338, + 0.012092618271708488, + 0.009244624525308609, + 0.008689165115356445, + 0.00972730852663517, + 0.010851344093680382, + 0.02910093031823635, + 0.011426884680986404, + 0.010209753178060055, + 0.018941424787044525, + 0.008271531201899052, + 0.0128566212952137, + 0.016265908256173134, + 0.01778518781065941, + 0.02403443492949009, + 0.02154800482094288, + 0.014028496108949184, + 0.01620922051370144, + 0.01678377576172352, + 0.021713105961680412, + 0.027086937800049782, + 0.0290395338088274, + 0.03638075664639473, + 0.0406489223241806, + 0.024724755436182022, + 0.023218343034386635, + 0.020991239696741104, + 0.029618067666888237, + 0.02421419508755207, + 0.032025232911109924, + 0.03434086963534355, + 0.01424629520624876, + 0.026752877980470657, + 0.01559955533593893, + 0.017628900706768036, + 0.007341850083321333, + 0.0056536756455898285, + 0.005960495211184025, + 0.004709294065833092, + 0.006017609499394894, + 0.008141765370965004, + 0.005336291156709194, + 0.005156870000064373, + 0.006021758075803518, + 0.00804932788014412, + 0.01551150158047676, + 0.0075734141282737255, + 0.007487939670681953, + 0.007255019154399633, + 0.0072656963020563126, + 0.007483539637178183, + 0.005720531102269888, + 0.005050834733992815, + 0.0077059632167220116, + 0.0044799321331083775, + 0.00470620347186923, + 0.005353898275643587, + 0.004182425793260336, + 0.00614189961925149, + 0.005136766470968723, + 0.0040198154747486115, + 0.00701872119680047, + 0.00482321297749877, + 0.003921499941498041, + 0.004975523799657822, + 0.005337869748473167, + 0.0071607488207519054, + 0.005704041570425034, + 0.004149725195020437, + 0.006763477344065905, + 0.006893794517964125, + 0.00581838795915246, + 0.005620935000479221, + 0.004905537236481905, + 0.0065136500634253025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.005948459263890982 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091820679605007 + ], + "opposite_logprob_tokens": [ + -5.504091739654541 + ], + "logprob_reward": -0.004091820679605007, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091820679605007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19699285924434662, + 0.06678147614002228, + 0.13455836474895477, + 0.09029950946569443, + 0.07784371078014374, + 0.140835240483284, + 0.0667024701833725, + 0.09466057270765305, + 0.06700315326452255, + 0.0643225759267807 + ], + "credit_map": [ + 0.018898775801062584, + 0.015626078471541405, + 0.017315765842795372, + 0.05389736220240593, + 0.01227424293756485, + 0.03670041263103485, + 0.029097001999616623, + 0.013183251023292542, + 0.006373728159815073, + 0.009736782871186733, + 0.007591762114316225, + 0.01073447521775961, + 0.008648736402392387, + 0.005965114571154118, + 0.01092449389398098, + 0.006806399207562208, + 0.012414705939590931, + 0.013286291621625423, + 0.018723178654909134, + 0.022215262055397034, + 0.01432210486382246, + 0.016391299664974213, + 0.02151712030172348, + 0.015688421204686165, + 0.009725418873131275, + 0.01750069111585617, + 0.011032489128410816, + 0.01232942659407854, + 0.01248079352080822, + 0.009775210171937943, + 0.008537031710147858, + 0.008918458595871925, + 0.01568465866148472, + 0.008411288261413574, + 0.009287464432418346, + 0.007877162657678127, + 0.006930366158485413, + 0.016210047528147697, + 0.006822429597377777, + 0.006620305590331554, + 0.022633029147982597, + 0.015823936089873314, + 0.012231134809553623, + 0.014515328221023083, + 0.011304644867777824, + 0.02794523537158966, + 0.024954577907919884, + 0.01142736803740263, + 0.008291050791740417, + 0.010787270963191986, + 0.011652490124106407, + 0.007281485479325056, + 0.006882722023874521, + 0.008064988069236279, + 0.005400164984166622, + 0.008342298679053783, + 0.010214769281446934, + 0.009550824761390686, + 0.011193890124559402, + 0.00978868268430233, + 0.008943988010287285, + 0.012072929181158543, + 0.010406900197267532, + 0.02248859405517578, + 0.0066835940815508366, + 0.00896438304334879, + 0.009269139729440212, + 0.00924173928797245, + 0.007979222573339939, + 0.007604032289236784, + 0.008047294802963734, + 0.009213750250637531, + 0.008424478583037853, + 0.00600724620744586, + 0.010236045345664024, + 0.00728369876742363, + 0.0070611098781228065, + 0.007698953151702881, + 0.00762544758617878, + 0.009985603392124176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.004091820679605007 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896433025598526 + ], + "opposite_logprob_tokens": [ + -2.5789642333984375 + ], + "logprob_reward": -0.07896433025598526, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896433025598526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10816249996423721, + 0.10141393542289734, + 0.10529644042253494, + 0.13287679851055145, + 0.1394507735967636, + 0.05323675274848938, + 0.12537162005901337, + 0.08730848133563995, + 0.08062463253736496, + 0.06625806540250778 + ], + "credit_map": [ + 0.013418536633253098, + 0.008407349698245525, + 0.010271520353853703, + 0.016833987087011337, + 0.00898722093552351, + 0.020260239019989967, + 0.01838650368154049, + 0.01159715186804533, + 0.009065601974725723, + 0.013164984062314034, + 0.008216769434511662, + 0.02205747365951538, + 0.012581270188093185, + 0.008289678953588009, + 0.019681448116898537, + 0.008356720209121704, + 0.010362341068685055, + 0.012501976452767849, + 0.013026487082242966, + 0.015664931386709213, + 0.013988789170980453, + 0.010391214862465858, + 0.014689689502120018, + 0.014671014621853828, + 0.014918791130185127, + 0.021506784483790398, + 0.013784308917820454, + 0.017396457493305206, + 0.024128347635269165, + 0.015797743573784828, + 0.01293139811605215, + 0.012412968091666698, + 0.01636388525366783, + 0.021681491285562515, + 0.015529186464846134, + 0.02235342562198639, + 0.016681408509612083, + 0.0170171819627285, + 0.01607964187860489, + 0.013744543306529522, + 0.007358570583164692, + 0.008160569705069065, + 0.00533430278301239, + 0.005422394257038832, + 0.006479182746261358, + 0.008003980852663517, + 0.006299001630395651, + 0.006178749725222588, + 0.012643801048398018, + 0.01650162972509861, + 0.02881319634616375, + 0.016830850392580032, + 0.010911145247519016, + 0.013172321021556854, + 0.007779147941619158, + 0.01871953345835209, + 0.016787013038992882, + 0.007384748198091984, + 0.016215629875659943, + 0.007198093459010124, + 0.003969789948314428, + 0.015751924365758896, + 0.0043805064633488655, + 0.015620779246091843, + 0.009343486279249191, + 0.0066930754110217094, + 0.015106264501810074, + 0.010333220474421978, + 0.006621986161917448, + 0.011835791170597076, + 0.007920490577816963, + 0.012770316563546658, + 0.008790455758571625, + 0.0061059254221618176, + 0.008766552433371544, + 0.009948919527232647, + 0.0073242876678705215, + 0.007129895966500044, + 0.00962620135396719, + 0.008565821684896946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 6 + ], + "selected_reward": -0.07896433025598526 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a dramatic medium shot, set in a vast, dimly lit tournament arena, with the blurred silhouettes of a huge, cheering crowd in the background. The camera's focus is first pinpointed on a massive totem in the foreground, symbolizing defeat. Composed of rough, dark gray, almost black rock, its surface is riddled with cracks. A faint, flickering flame adorns its summit, and dark, solidified wax drips clutter its shaft. Suddenly, with a muffled, powerful crack, the totem's shaft violently shatters, exploding into countless sharp fragments and dust. At that moment, a blinding white flash instantly engulfs the frame, and the camera shifts focus with remarkable speed, moving from the dark, crumbling debris in the foreground to a smooth, decisive pull toward another totem in the background, symbolizing victory. This totem is crafted from a brilliant golden metal, its surface smooth and reflective, and its tip blazes with a powerful, blazing golden-white flame. The camera finally stops on this victory totem, its golden light illuminates the flying dust around it, forming a static picture full of power and glory. The whole scene, under the high-contrast light and shadow, creates a glorious atmosphere of tension, climaxes and ultimately the joy of victory.", + "seed": 3317, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_4/index_5726_seed_3317_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016465831967070698, + "vlm_logprob": -0.016465831967070698, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.016465831967070698, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017248153686523438 + ], + "opposite_logprob_tokens": [ + -4.142248153686523 + ], + "logprob_reward": -0.017248153686523438, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.017248153686523438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.200687974691391, + 0.103360116481781, + 0.16874735057353973, + 0.10161343961954117, + 0.2172592729330063, + 0.03244154527783394, + 0.07886238396167755, + 0.03871508687734604, + 0.037763483822345734, + 0.02054929919540882 + ], + "credit_map": [ + 0.018039334565401077, + 0.02772359363734722, + 0.03080032393336296, + 0.028612518683075905, + 0.02541554719209671, + 0.02721020206809044, + 0.02365758642554283, + 0.019228888675570488, + 0.010621128603816032, + 0.013217202387750149, + 0.01652620919048786, + 0.010504563339054585, + 0.00836178008466959, + 0.01442217268049717, + 0.01160338893532753, + 0.018103674054145813, + 0.04559855908155441, + 0.008273403160274029, + 0.028684573248028755, + 0.008900340646505356, + 0.010901727713644505, + 0.033937353640794754, + 0.009460045024752617, + 0.022991366684436798, + 0.010777881368994713, + 0.011567198671400547, + 0.01802779920399189, + 0.009838860481977463, + 0.011637933552265167, + 0.013467323035001755, + 0.013019031845033169, + 0.013277418911457062, + 0.0214184932410717, + 0.026089662685990334, + 0.0260359738022089, + 0.03521864488720894, + 0.02022043988108635, + 0.02885204553604126, + 0.02616582065820694, + 0.03325822204351425, + 0.004164571408182383, + 0.004769506398588419, + 0.0040849121287465096, + 0.00372429471462965, + 0.004309480544179678, + 0.0042822896502912045, + 0.0039276438765227795, + 0.0031788493506610394, + 0.008769059553742409, + 0.022884901612997055, + 0.007996329106390476, + 0.006449516862630844, + 0.006056156475096941, + 0.012519788928329945, + 0.006381218787282705, + 0.007805417291820049, + 0.005364683456718922, + 0.006477745715528727, + 0.0036760298535227776, + 0.002386766951531172, + 0.0032610734924674034, + 0.009326590225100517, + 0.003742429194971919, + 0.0044797733426094055, + 0.004882758017629385, + 0.006281787063926458, + 0.004168802872300148, + 0.0025653631892055273, + 0.0031418742146342993, + 0.009019767865538597, + 0.0029700929298996925, + 0.004733039531856775, + 0.003330405568704009, + 0.0031809136271476746, + 0.0031093270517885685, + 0.0015492478851228952, + 0.001989587675780058, + 0.0044011385180056095, + 0.0015539235901087523, + 0.0014347569085657597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.017248153686523438 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789316914975643 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.009789316914975643, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789316914975643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41081905364990234, + 0.022261345759034157, + 0.4130207300186157, + 0.039289847016334534, + 0.027575718238949776, + 0.018569936975836754, + 0.017492253333330154, + 0.020545657724142075, + 0.014601562172174454, + 0.01582391932606697 + ], + "credit_map": [ + 0.03905785456299782, + 0.055252280086278915, + 0.06853204220533371, + 0.06949859857559204, + 0.04184406250715256, + 0.05456431955099106, + 0.04720010980963707, + 0.03486976772546768, + 0.002085733227431774, + 0.0022263636346906424, + 0.0023422888480126858, + 0.005140583962202072, + 0.0018021868309006095, + 0.002621309133246541, + 0.002978775417432189, + 0.00306410132907331, + 0.09656399488449097, + 0.01765446551144123, + 0.059620440006256104, + 0.02121688611805439, + 0.02660665102303028, + 0.09579763561487198, + 0.023588204756379128, + 0.07197243720293045, + 0.008053110912442207, + 0.003559716511517763, + 0.005159287713468075, + 0.005541803315281868, + 0.0036818438675254583, + 0.005111796781420708, + 0.004401728976517916, + 0.0037805556785315275, + 0.0031499327160418034, + 0.004394973162561655, + 0.0024636785965412855, + 0.005823799408972263, + 0.0024228254333138466, + 0.003942977637052536, + 0.0033756259363144636, + 0.0020019018556922674, + 0.0021230552811175585, + 0.0027862342540174723, + 0.002139428397640586, + 0.0017175087705254555, + 0.0021602241322398186, + 0.0026785896625369787, + 0.002019947860389948, + 0.002944947686046362, + 0.001738114980980754, + 0.004011556971818209, + 0.0019279539119452238, + 0.0017106852028518915, + 0.00161275127902627, + 0.0024148854427039623, + 0.0026726231444627047, + 0.0014036833308637142, + 0.0022436771541833878, + 0.0032275000121444464, + 0.0022834567353129387, + 0.002667280612513423, + 0.0021781730465590954, + 0.0025098260957747698, + 0.0037873403634876013, + 0.001648402656428516, + 0.001791036338545382, + 0.0030283976811915636, + 0.0015736743807792664, + 0.001161327469162643, + 0.0013914561131969094, + 0.0023985840380191803, + 0.0017613348318263888, + 0.001495750853791833, + 0.0021009626798331738, + 0.0021403066348284483, + 0.0023083502892404795, + 0.0015793730271980166, + 0.001226752414368093, + 0.0036528129130601883, + 0.0015912116505205631, + 0.0012241500662639737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.009789316914975643 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634849451482296 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008634849451482296, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634849451482296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4880463480949402, + 0.03497255966067314, + 0.1377190351486206, + 0.05357322841882706, + 0.06612011045217514, + 0.07481691241264343, + 0.03806101903319359, + 0.04795508086681366, + 0.0235741026699543, + 0.03516153618693352 + ], + "credit_map": [ + 0.047891806811094284, + 0.06352926045656204, + 0.079666368663311, + 0.08481891453266144, + 0.048403870314359665, + 0.06540357321500778, + 0.055936574935913086, + 0.04239602014422417, + 0.003795590717345476, + 0.0036466496530920267, + 0.004050371237099171, + 0.006361667066812515, + 0.0036399592645466328, + 0.004253693390637636, + 0.004533185623586178, + 0.0046914503909647465, + 0.03533431887626648, + 0.0064130183309316635, + 0.024555902928113937, + 0.007024715188890696, + 0.008755436167120934, + 0.027656108140945435, + 0.007870122790336609, + 0.02010943926870823, + 0.0055512115359306335, + 0.006324299145489931, + 0.009008445776998997, + 0.006695899181067944, + 0.005822699051350355, + 0.007118492852896452, + 0.006091758143156767, + 0.006960431579500437, + 0.006362957879900932, + 0.010845688171684742, + 0.006336349993944168, + 0.01551520824432373, + 0.005584482103586197, + 0.008704994805157185, + 0.007557532750070095, + 0.005212907679378986, + 0.009331614710390568, + 0.007765903603285551, + 0.00942204613238573, + 0.005301888100802898, + 0.007629387080669403, + 0.011260420083999634, + 0.007820066064596176, + 0.016285598278045654, + 0.0036390572786331177, + 0.006829000543802977, + 0.004114098381251097, + 0.0032770258840173483, + 0.00567682646214962, + 0.00411471351981163, + 0.007332659792155027, + 0.0030776411294937134, + 0.004624798893928528, + 0.010456166230142117, + 0.005008026957511902, + 0.006069941911846399, + 0.005928321275860071, + 0.004696592688560486, + 0.008031956851482391, + 0.0031392842065542936, + 0.002613685792312026, + 0.004454435780644417, + 0.002376825548708439, + 0.002227125223726034, + 0.003357945941388607, + 0.003470627823844552, + 0.003128496464341879, + 0.0019449640531092882, + 0.004710400942713022, + 0.007919100113213062, + 0.004530506208539009, + 0.0036277479957789183, + 0.0041458485648036, + 0.004086086992174387, + 0.0038131861947476864, + 0.002328665228560567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.008634849451482296 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036117101553827524 + ], + "opposite_logprob_tokens": [ + -5.6286115646362305 + ], + "logprob_reward": -0.0036117101553827524, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036117101553827524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23312203586101532, + 0.053060196340084076, + 0.2534993886947632, + 0.1157391145825386, + 0.08126550167798996, + 0.06977858394384384, + 0.04803472012281418, + 0.05192572623491287, + 0.03683188185095787, + 0.056742891669273376 + ], + "credit_map": [ + 0.02211509272456169, + 0.026755066588521004, + 0.03164130076766014, + 0.03689376264810562, + 0.02532942034304142, + 0.03090658038854599, + 0.02633596584200859, + 0.03314482793211937, + 0.004808444529771805, + 0.008070670999586582, + 0.005390200763940811, + 0.00943648163229227, + 0.007563373073935509, + 0.005996727850288153, + 0.00672073382884264, + 0.005073560867458582, + 0.06480958312749863, + 0.011779166758060455, + 0.043299105018377304, + 0.012604573741555214, + 0.017982138320803642, + 0.05117611587047577, + 0.015309171751141548, + 0.036539528518915176, + 0.011898946017026901, + 0.014609911479055882, + 0.018913457170128822, + 0.013417846523225307, + 0.01346362940967083, + 0.01526628341525793, + 0.014223466627299786, + 0.013945563696324825, + 0.008279349654912949, + 0.013738679699599743, + 0.0071706813760101795, + 0.015754716470837593, + 0.007762243505567312, + 0.011213215999305248, + 0.012802815064787865, + 0.004543796181678772, + 0.009661856107413769, + 0.007344023324549198, + 0.008841321803629398, + 0.004929041489958763, + 0.0070720273070037365, + 0.010548562742769718, + 0.007016854826360941, + 0.014364899136126041, + 0.00470332708209753, + 0.0078009190037846565, + 0.005917548667639494, + 0.004257683642208576, + 0.006811570376157761, + 0.005477022845298052, + 0.008712922222912312, + 0.004353723023086786, + 0.005471618380397558, + 0.008760605938732624, + 0.0046523818746209145, + 0.005563764367252588, + 0.006890947464853525, + 0.006348748225718737, + 0.009684666991233826, + 0.004552991129457951, + 0.004542579874396324, + 0.004690632224082947, + 0.0035746535286307335, + 0.004290624056011438, + 0.004176322836428881, + 0.004835699684917927, + 0.007521426305174828, + 0.003199940314516425, + 0.005843441467732191, + 0.009546389803290367, + 0.005955683067440987, + 0.005787509027868509, + 0.004217274021357298, + 0.013841204345226288, + 0.007078577298671007, + 0.004472811706364155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.0036117101553827524 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043045129626989365 + ], + "opposite_logprob_tokens": [ + -3.1680450439453125 + ], + "logprob_reward": -0.043045129626989365, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043045129626989365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47396597266197205, + 0.05722157657146454, + 0.14715790748596191, + 0.06892816722393036, + 0.0395614430308342, + 0.042926497757434845, + 0.03444284945726395, + 0.03743967413902283, + 0.03307633101940155, + 0.06527945399284363 + ], + "credit_map": [ + 0.04711860790848732, + 0.06609014421701431, + 0.06492621451616287, + 0.07860008627176285, + 0.05207112431526184, + 0.060897018760442734, + 0.05647636950016022, + 0.04778646305203438, + 0.004853885620832443, + 0.010068175382912159, + 0.006113952491432428, + 0.00576594565063715, + 0.0076384167186915874, + 0.007028980180621147, + 0.008556189946830273, + 0.007196031976491213, + 0.03485868126153946, + 0.0070211878046393394, + 0.02534961886703968, + 0.008090554736554623, + 0.00906915869563818, + 0.030533619225025177, + 0.008106323890388012, + 0.024128787219524384, + 0.006994597148150206, + 0.00956475269049406, + 0.009672466665506363, + 0.009122390300035477, + 0.007747930008918047, + 0.008974138647317886, + 0.009198030456900597, + 0.007653865497559309, + 0.004185853060334921, + 0.005544278305023909, + 0.0036903047002851963, + 0.006463740486651659, + 0.0042643253691494465, + 0.005503857508301735, + 0.006772501859813929, + 0.0031365850009024143, + 0.006963110528886318, + 0.005534576717764139, + 0.003715930273756385, + 0.004917581100016832, + 0.005245306063443422, + 0.006707977969199419, + 0.0042351423762738705, + 0.005606880411505699, + 0.00345486868172884, + 0.005448481533676386, + 0.004134221002459526, + 0.0028585537802428007, + 0.0040815104730427265, + 0.004769919440150261, + 0.0048484476283192635, + 0.0048468527384102345, + 0.003804422216489911, + 0.008140752092003822, + 0.004850376863032579, + 0.003489948809146881, + 0.003978610038757324, + 0.004720374010503292, + 0.005565081723034382, + 0.0028901132754981518, + 0.004311875440180302, + 0.005965531338006258, + 0.0038720027077943087, + 0.0033665376249700785, + 0.0038399111945182085, + 0.005352355539798737, + 0.0036013147328048944, + 0.002766803838312626, + 0.009554637596011162, + 0.010787195526063442, + 0.006315083708614111, + 0.005667015910148621, + 0.006331045646220446, + 0.016147589311003685, + 0.006247606128454208, + 0.004229281563311815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.043045129626989365 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a dramatic medium shot, set in a vast, dimly lit tournament arena, with the blurred silhouettes of a huge, cheering crowd in the background. The camera's focus is first pinpointed on a massive totem in the foreground, symbolizing defeat. Composed of rough, dark gray, almost black rock, its surface is riddled with cracks. A faint, flickering flame adorns its summit, and dark, solidified wax drips clutter its shaft. Suddenly, with a muffled, powerful crack, the totem's shaft violently shatters, exploding into countless sharp fragments and dust. At that moment, a blinding white flash instantly engulfs the frame, and the camera shifts focus with remarkable speed, moving from the dark, crumbling debris in the foreground to a smooth, decisive pull toward another totem in the background, symbolizing victory. This totem is crafted from a brilliant golden metal, its surface smooth and reflective, and its tip blazes with a powerful, blazing golden-white flame. The camera finally stops on this victory totem, its golden light illuminates the flying dust around it, forming a static picture full of power and glory. The whole scene, under the high-contrast light and shadow, creates a glorious atmosphere of tension, climaxes and ultimately the joy of victory.", + "seed": 3318, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_4/index_5726_seed_3318_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019661614578217267, + "vlm_logprob": -0.019661614578217267, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.019661614578217267, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030698830261826515 + ], + "opposite_logprob_tokens": [ + -3.530698776245117 + ], + "logprob_reward": -0.030698830261826515, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.030698830261826515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2901124954223633, + 0.17003250122070312, + 0.1559929996728897, + 0.0657411739230156, + 0.061561111360788345, + 0.08512190729379654, + 0.036450497806072235, + 0.01684480533003807, + 0.08711598068475723, + 0.031026534736156464 + ], + "credit_map": [ + 0.061148181557655334, + 0.018595539033412933, + 0.0507298968732357, + 0.02936537377536297, + 0.016963498666882515, + 0.03109552338719368, + 0.028313633054494858, + 0.05390084907412529, + 0.03881407156586647, + 0.006464895326644182, + 0.020417528226971626, + 0.010643543675541878, + 0.006308295298367739, + 0.056474488228559494, + 0.009288248606026173, + 0.021621443331241608, + 0.020278753712773323, + 0.017975958064198494, + 0.031573738902807236, + 0.01782708615064621, + 0.012780876830220222, + 0.02164839580655098, + 0.016309574246406555, + 0.01759861595928669, + 0.008403262123465538, + 0.005386936012655497, + 0.007066285703331232, + 0.005800931248813868, + 0.00829008873552084, + 0.011787914671003819, + 0.010842491872608662, + 0.008163266815245152, + 0.01263298187404871, + 0.008118292316794395, + 0.0036496322136372328, + 0.004491154104471207, + 0.007643342949450016, + 0.014013128355145454, + 0.006426082458347082, + 0.004586496856063604, + 0.007982906885445118, + 0.01527173537760973, + 0.008716996759176254, + 0.005411575082689524, + 0.016387583687901497, + 0.015388340689241886, + 0.007447391748428345, + 0.008515382185578346, + 0.00463897455483675, + 0.005205445922911167, + 0.0027944836765527725, + 0.0025689678732305765, + 0.0065430207177996635, + 0.0054148901253938675, + 0.0036034039221704006, + 0.005681311246007681, + 0.0017280227039009333, + 0.0022054254077374935, + 0.002100827172398567, + 0.0023946117144078016, + 0.002573327859863639, + 0.0018402030691504478, + 0.0017926477594301105, + 0.002209739526733756, + 0.006468311417847872, + 0.027988700196146965, + 0.005627941340208054, + 0.006622642744332552, + 0.012509635649621487, + 0.013827198185026646, + 0.007376945577561855, + 0.006694607436656952, + 0.0037779954727739096, + 0.0069600315764546394, + 0.002165276790037751, + 0.0032008024863898754, + 0.003986390307545662, + 0.004482949152588844, + 0.0035631374921649694, + 0.002889948431402445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.030698830261826515 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018193509429693222 + ], + "opposite_logprob_tokens": [ + -4.01819372177124 + ], + "logprob_reward": -0.018193509429693222, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018193509429693222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15820032358169556, + 0.16256731748580933, + 0.08777476102113724, + 0.1778523027896881, + 0.11074331402778625, + 0.11433523148298264, + 0.034532465040683746, + 0.028023473918437958, + 0.07663416117429733, + 0.049336668103933334 + ], + "credit_map": [ + 0.02133023738861084, + 0.012973102740943432, + 0.027393987402319908, + 0.02130274474620819, + 0.017380032688379288, + 0.015736788511276245, + 0.02185838297009468, + 0.020225049927830696, + 0.019878242164850235, + 0.011573579162359238, + 0.019696174189448357, + 0.021045755594968796, + 0.013502250425517559, + 0.030416931957006454, + 0.019795242697000504, + 0.02665914036333561, + 0.010048027150332928, + 0.012526025995612144, + 0.013115350157022476, + 0.010306969285011292, + 0.009756326675415039, + 0.012242334894835949, + 0.010852302424609661, + 0.008927418850362301, + 0.02488619089126587, + 0.02447354607284069, + 0.021662473678588867, + 0.023888317868113518, + 0.01667877286672592, + 0.024801721796393394, + 0.0272541381418705, + 0.014207148924469948, + 0.02633826620876789, + 0.016597017645835876, + 0.004711146000772715, + 0.006877227686345577, + 0.012108254246413708, + 0.030828624963760376, + 0.0068012019619345665, + 0.006481574382632971, + 0.011522010900080204, + 0.025401413440704346, + 0.006776306312531233, + 0.00822987500578165, + 0.021895579993724823, + 0.023920370265841484, + 0.009692596271634102, + 0.0068970732390880585, + 0.0033368414733558893, + 0.003881427925080061, + 0.0035347503144294024, + 0.003745727939531207, + 0.0050224424339830875, + 0.004523468669503927, + 0.004025543574243784, + 0.006462266203016043, + 0.0032369995024055243, + 0.003670560661703348, + 0.0026641206350177526, + 0.0031639805529266596, + 0.003949841950088739, + 0.004598595667630434, + 0.0034472511615604162, + 0.003292124019935727, + 0.011419217102229595, + 0.01572919636964798, + 0.006618308834731579, + 0.005477528087794781, + 0.012354138307273388, + 0.012760559096932411, + 0.004971643444150686, + 0.007303571794182062, + 0.006066416390240192, + 0.008495950140058994, + 0.0034142297226935625, + 0.0036698253825306892, + 0.009398166090250015, + 0.006918311584740877, + 0.0041782790794968605, + 0.007195490412414074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.018193509429693222 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537825852632523 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012537825852632523, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537825852632523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09117712080478668, + 0.13095830380916595, + 0.2437950223684311, + 0.16614405810832977, + 0.13073159754276276, + 0.0682564303278923, + 0.046935632824897766, + 0.027776595205068588, + 0.0786937028169632, + 0.015531487762928009 + ], + "credit_map": [ + 0.019617324694991112, + 0.006201562937349081, + 0.015234422869980335, + 0.008811208419501781, + 0.006914806552231312, + 0.011570079252123833, + 0.00727272592484951, + 0.015554995276033878, + 0.031658437103033066, + 0.006790300831198692, + 0.009542594663798809, + 0.007803359068930149, + 0.006824580021202564, + 0.04369301721453667, + 0.01067810133099556, + 0.013967919163405895, + 0.029375754296779633, + 0.028228404000401497, + 0.04588310047984123, + 0.025104131549596786, + 0.024558503180742264, + 0.0337551012635231, + 0.0290081724524498, + 0.027881871908903122, + 0.01840442232787609, + 0.016461748629808426, + 0.015896571800112724, + 0.024456005543470383, + 0.029818028211593628, + 0.021205835044384003, + 0.024094171822071075, + 0.01580728217959404, + 0.03451155871152878, + 0.016085660085082054, + 0.006904997397214174, + 0.007208798546344042, + 0.01250557228922844, + 0.04024415835738182, + 0.006706437095999718, + 0.006564426701515913, + 0.007044541649520397, + 0.012875582091510296, + 0.004262498579919338, + 0.004986142739653587, + 0.013785995543003082, + 0.012416956014931202, + 0.007501507643610239, + 0.00538320979103446, + 0.006563166156411171, + 0.008087052032351494, + 0.0033226197119802237, + 0.0029889403376728296, + 0.01157392282038927, + 0.007408672012388706, + 0.0032157276291400194, + 0.0037755323573946953, + 0.0029330344405025244, + 0.003468636190518737, + 0.0027873266953974962, + 0.004254911094903946, + 0.004454550798982382, + 0.0026659751310944557, + 0.00372690730728209, + 0.003485254943370819, + 0.00666147843003273, + 0.01838841661810875, + 0.008454647846519947, + 0.007727794349193573, + 0.012719959020614624, + 0.009948891587555408, + 0.007965991273522377, + 0.006826527416706085, + 0.002060376340523362, + 0.002238699235022068, + 0.0018526705680415034, + 0.001771019073203206, + 0.002217880915850401, + 0.0018084847833961248, + 0.0019795959815382957, + 0.0016027613310143352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.012537825852632523 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031898352317512035 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.0031898352317512035, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031898352317512035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16529293358325958, + 0.12743905186653137, + 0.06853223592042923, + 0.04219628497958183, + 0.07778303325176239, + 0.17057223618030548, + 0.09473080188035965, + 0.032556790858507156, + 0.13625745475292206, + 0.08463913947343826 + ], + "credit_map": [ + 0.04183680936694145, + 0.01136801764369011, + 0.020687144249677658, + 0.0140029676258564, + 0.009706673212349415, + 0.014712168835103512, + 0.017139960080385208, + 0.03583921119570732, + 0.022700922563672066, + 0.007405622862279415, + 0.01087197009474039, + 0.007099525537341833, + 0.0062817903235554695, + 0.03918640315532684, + 0.009800937958061695, + 0.02409188449382782, + 0.009262977167963982, + 0.008142254315316677, + 0.013422748073935509, + 0.005769301671534777, + 0.006566944532096386, + 0.010147570632398129, + 0.006450050976127386, + 0.008770395070314407, + 0.004721842706203461, + 0.0041221496649086475, + 0.005893238354474306, + 0.004033361095935106, + 0.0045971740037202835, + 0.005159894470125437, + 0.00792927946895361, + 0.005739349871873856, + 0.011216647922992706, + 0.01191047765314579, + 0.006335255224257708, + 0.006378553342074156, + 0.01038340013474226, + 0.014322580769658089, + 0.0070485398173332214, + 0.010187584906816483, + 0.018359383568167686, + 0.040495865046978, + 0.009021931327879429, + 0.009325943887233734, + 0.028996868059039116, + 0.038189541548490524, + 0.013584084808826447, + 0.012598630972206593, + 0.011281577870249748, + 0.015566147863864899, + 0.007771518547087908, + 0.007366914767771959, + 0.021373458206653595, + 0.014807806350290775, + 0.006862927693873644, + 0.00970045942813158, + 0.0037271766923367977, + 0.0040940046310424805, + 0.0030859080143272877, + 0.004392731003463268, + 0.006747662089765072, + 0.003979815635830164, + 0.0030552230309695005, + 0.0034742732532322407, + 0.01457189954817295, + 0.04291200637817383, + 0.00630028173327446, + 0.00789063610136509, + 0.016907060518860817, + 0.028120974078774452, + 0.008732457645237446, + 0.010822140611708164, + 0.009817002341151237, + 0.013006389141082764, + 0.010426010005176067, + 0.012212865054607391, + 0.008310530334711075, + 0.008679792284965515, + 0.012384690344333649, + 0.009801870211958885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.0031898352317512035 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368807211518288 + ], + "opposite_logprob_tokens": [ + -3.4086880683898926 + ], + "logprob_reward": -0.03368807211518288, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368807211518288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10809078067541122, + 0.08981616795063019, + 0.03741947188973427, + 0.03816385567188263, + 0.08876368403434753, + 0.1985853761434555, + 0.07508011162281036, + 0.043096087872982025, + 0.2667238712310791, + 0.05426052212715149 + ], + "credit_map": [ + 0.020365094766020775, + 0.006890188902616501, + 0.018876465037465096, + 0.009732730686664581, + 0.010053779929876328, + 0.012759379111230373, + 0.008893040008842945, + 0.020520100370049477, + 0.01569228246808052, + 0.003965286538004875, + 0.007393412292003632, + 0.006884397938847542, + 0.003777055535465479, + 0.026692884042859077, + 0.009165358729660511, + 0.016245488077402115, + 0.004826318472623825, + 0.003790042130276561, + 0.0055413395166397095, + 0.005067924968898296, + 0.003721896791830659, + 0.005076603032648563, + 0.004899859428405762, + 0.004495490808039904, + 0.003732925746589899, + 0.0032273810356855392, + 0.006562081631273031, + 0.004167985636740923, + 0.00480907317250967, + 0.006782879121601582, + 0.004372555762529373, + 0.0045089744962751865, + 0.014522096142172813, + 0.013569248840212822, + 0.006868303753435612, + 0.006697054952383041, + 0.01208645012229681, + 0.017778033390641212, + 0.009436196647584438, + 0.007806296925991774, + 0.024360880255699158, + 0.03932620957493782, + 0.00966043584048748, + 0.01373480074107647, + 0.033679354935884476, + 0.04116209223866463, + 0.01867987960577011, + 0.017981726676225662, + 0.008333882316946983, + 0.010408823378384113, + 0.006407488603144884, + 0.006359392777085304, + 0.01230344083160162, + 0.009346073493361473, + 0.009337961673736572, + 0.012583049945533276, + 0.004399609286338091, + 0.00617377320304513, + 0.0046802484430372715, + 0.004865280818194151, + 0.0080895209684968, + 0.004779403563588858, + 0.0058105746284127235, + 0.0042976755648851395, + 0.019596507772803307, + 0.09374675899744034, + 0.017596617341041565, + 0.0177670419216156, + 0.033772531896829605, + 0.04752179980278015, + 0.019551793113350868, + 0.01717081293463707, + 0.005097110290080309, + 0.006260831840336323, + 0.006678884383291006, + 0.009256737306714058, + 0.006664611864835024, + 0.004433899186551571, + 0.006348010618239641, + 0.009520437568426132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.03368807211518288 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a dramatic medium shot, set in a vast, dimly lit tournament arena, with the blurred silhouettes of a huge, cheering crowd in the background. The camera's focus is first pinpointed on a massive totem in the foreground, symbolizing defeat. Composed of rough, dark gray, almost black rock, its surface is riddled with cracks. A faint, flickering flame adorns its summit, and dark, solidified wax drips clutter its shaft. Suddenly, with a muffled, powerful crack, the totem's shaft violently shatters, exploding into countless sharp fragments and dust. At that moment, a blinding white flash instantly engulfs the frame, and the camera shifts focus with remarkable speed, moving from the dark, crumbling debris in the foreground to a smooth, decisive pull toward another totem in the background, symbolizing victory. This totem is crafted from a brilliant golden metal, its surface smooth and reflective, and its tip blazes with a powerful, blazing golden-white flame. The camera finally stops on this victory totem, its golden light illuminates the flying dust around it, forming a static picture full of power and glory. The whole scene, under the high-contrast light and shadow, creates a glorious atmosphere of tension, climaxes and ultimately the joy of victory.", + "seed": 3319, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_4/index_5726_seed_3319_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04128145389258862, + "vlm_logprob": -0.04128145389258862, + "vlm_yes_no_margin": 3.800000023841858, + "vlm_token_credit": -0.04128145389258862, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007291372865438461 + ], + "opposite_logprob_tokens": [ + -5.007291316986084 + ], + "logprob_reward": -0.007291372865438461, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.007291372865438461, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40008053183555603, + 0.07193747907876968, + 0.07808415591716766, + 0.11127501726150513, + 0.10634590685367584, + 0.08152133971452713, + 0.02454659342765808, + 0.02458023838698864, + 0.0728505477309227, + 0.028778141364455223 + ], + "credit_map": [ + 0.05777839943766594, + 0.039219312369823456, + 0.05202757567167282, + 0.08464158326387405, + 0.033083442598581314, + 0.031909819692373276, + 0.0781094953417778, + 0.02331092208623886, + 0.0082220658659935, + 0.01098056510090828, + 0.011800024658441544, + 0.009454114362597466, + 0.006358991842716932, + 0.008511548861861229, + 0.008171644993126392, + 0.008438528515398502, + 0.010456172749400139, + 0.007709276396781206, + 0.01782682165503502, + 0.006573061924427748, + 0.006469733081758022, + 0.011151673272252083, + 0.005198113154619932, + 0.012699311599135399, + 0.010183551348745823, + 0.012827415019273758, + 0.016588643193244934, + 0.01434269454330206, + 0.018762823194265366, + 0.012991727329790592, + 0.013303052634000778, + 0.012275123968720436, + 0.010128643363714218, + 0.025119813159108162, + 0.009829930029809475, + 0.006420362275093794, + 0.022907763719558716, + 0.0177061278373003, + 0.005787178408354521, + 0.008446098305284977, + 0.006704365368932486, + 0.013752021826803684, + 0.012036378495395184, + 0.005763523746281862, + 0.013560890220105648, + 0.014274151995778084, + 0.0065463148057460785, + 0.008883703500032425, + 0.002696704352274537, + 0.00405859062448144, + 0.0029045133851468563, + 0.003355309832841158, + 0.0036929198540747166, + 0.0029238376300781965, + 0.0024054674431681633, + 0.0025092519354075193, + 0.003653786377981305, + 0.002385190688073635, + 0.003497466677799821, + 0.0025957159232348204, + 0.00223317788913846, + 0.0025043191853910685, + 0.0036048367619514465, + 0.004105745814740658, + 0.00958386342972517, + 0.007687503471970558, + 0.013298987410962582, + 0.006668669171631336, + 0.008755449205636978, + 0.010073045268654823, + 0.006786842830479145, + 0.009996199049055576, + 0.003951168619096279, + 0.003426196053624153, + 0.005195945501327515, + 0.002246279502287507, + 0.0026483868714421988, + 0.004075516946613789, + 0.0016608622390776873, + 0.005573789589107037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.007291372865438461 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288074880838394 + ], + "opposite_logprob_tokens": [ + -3.7732880115509033 + ], + "logprob_reward": -0.023288074880838394, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288074880838394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4677830934524536, + 0.14195355772972107, + 0.04989148676395416, + 0.02970312163233757, + 0.06345037370920181, + 0.06992068886756897, + 0.035653918981552124, + 0.02790827676653862, + 0.07422615587711334, + 0.03950938209891319 + ], + "credit_map": [ + 0.06810783594846725, + 0.05118761211633682, + 0.07590940594673157, + 0.08887393027544022, + 0.04647986218333244, + 0.04188312217593193, + 0.06974661350250244, + 0.02559468522667885, + 0.0145768066868186, + 0.018934471532702446, + 0.013939689844846725, + 0.020519807934761047, + 0.0165814571082592, + 0.02079148218035698, + 0.02783072553575039, + 0.008779121562838554, + 0.005411822814494371, + 0.004612114746123552, + 0.008277582935988903, + 0.005777325481176376, + 0.008960260078310966, + 0.006886402145028114, + 0.004337953869253397, + 0.0056280214339494705, + 0.003386185271665454, + 0.004322072025388479, + 0.004276821855455637, + 0.0023362720385193825, + 0.004728366155177355, + 0.005460216198116541, + 0.0027162733022123575, + 0.0024769145529717207, + 0.005324393045157194, + 0.010034783743321896, + 0.008013090118765831, + 0.0053939176723361015, + 0.010066046379506588, + 0.01203868631273508, + 0.005164456553757191, + 0.007415000814944506, + 0.0057088034227490425, + 0.010852186009287834, + 0.010288090445101261, + 0.0046302746050059795, + 0.014417757280170918, + 0.011794731952250004, + 0.004759849514812231, + 0.007468992844223976, + 0.004464045166969299, + 0.005652558524161577, + 0.004359798040241003, + 0.00329581624828279, + 0.005393192172050476, + 0.0041474634781479836, + 0.0039026429876685143, + 0.004438399337232113, + 0.0029348088428378105, + 0.003441674867644906, + 0.0033893887884914875, + 0.0032253649551421404, + 0.004143170081079006, + 0.0036042206920683384, + 0.003657389199361205, + 0.003512260038405657, + 0.009184232912957668, + 0.007566045969724655, + 0.016577256843447685, + 0.0072840009815990925, + 0.006597227416932583, + 0.008572723716497421, + 0.006154530216008425, + 0.01229013130068779, + 0.0065054078586399555, + 0.003941051661968231, + 0.008652294985949993, + 0.003014974296092987, + 0.0024099803995341063, + 0.0059486874379217625, + 0.0027539064176380634, + 0.0062830811366438866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.023288074880838394 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023278523236513138 + ], + "opposite_logprob_tokens": [ + -3.7732784748077393 + ], + "logprob_reward": -0.023278523236513138, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023278523236513138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4783377945423126, + 0.0959775298833847, + 0.04070984944701195, + 0.03796800225973129, + 0.05051625147461891, + 0.053743764758110046, + 0.05974016338586807, + 0.038811419159173965, + 0.0995699092745781, + 0.04462533816695213 + ], + "credit_map": [ + 0.06953401118516922, + 0.04798229783773422, + 0.06230347976088524, + 0.09882491827011108, + 0.038744907826185226, + 0.03791818767786026, + 0.09764991700649261, + 0.02538003772497177, + 0.01032387837767601, + 0.01261733565479517, + 0.007341479416936636, + 0.013619919307529926, + 0.010440236888825893, + 0.01390402764081955, + 0.02084897831082344, + 0.006881680339574814, + 0.005098319612443447, + 0.004553663544356823, + 0.006214506458491087, + 0.005512846168130636, + 0.004639653023332357, + 0.006050323136150837, + 0.003809403395280242, + 0.004831131547689438, + 0.0032371049746870995, + 0.006051394157111645, + 0.004387921188026667, + 0.0034157708287239075, + 0.007855185307562351, + 0.00561335776001215, + 0.0037394247483462095, + 0.003667842596769333, + 0.003932933788746595, + 0.00950763188302517, + 0.0038136299699544907, + 0.004364450462162495, + 0.008636143058538437, + 0.012309303507208824, + 0.003383835544809699, + 0.004568324889987707, + 0.005170266609638929, + 0.010849512182176113, + 0.0034048694651573896, + 0.0032711487729102373, + 0.01461127307265997, + 0.009665842168033123, + 0.003259937046095729, + 0.0035109163727611303, + 0.005492966156452894, + 0.008346655406057835, + 0.007418319117277861, + 0.005531011614948511, + 0.011131301522254944, + 0.00850551575422287, + 0.005769409704953432, + 0.007544983644038439, + 0.004019965883344412, + 0.007228860631585121, + 0.0036842504050582647, + 0.005022899247705936, + 0.004580891691148281, + 0.004400314763188362, + 0.005214829463511705, + 0.004659406840801239, + 0.01672840490937233, + 0.00850089080631733, + 0.012835175730288029, + 0.010314300656318665, + 0.010950802825391293, + 0.019408779218792915, + 0.00911031849682331, + 0.011721235699951649, + 0.0058401585556566715, + 0.007737976033240557, + 0.005863567348569632, + 0.004984096623957157, + 0.0038944680709391832, + 0.005705438554286957, + 0.004240937530994415, + 0.006358696613460779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.023278523236513138 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009773852303624153 + ], + "opposite_logprob_tokens": [ + -4.6347737312316895 + ], + "logprob_reward": -0.009773852303624153, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009773852303624153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46744465827941895, + 0.07987169921398163, + 0.05518155172467232, + 0.04690517112612724, + 0.05432804301381111, + 0.06526663899421692, + 0.05054587125778198, + 0.06077433004975319, + 0.09164322912693024, + 0.02803870663046837 + ], + "credit_map": [ + 0.0679381936788559, + 0.04572984576225281, + 0.060871124267578125, + 0.10229359567165375, + 0.03867131471633911, + 0.03601895272731781, + 0.08928927779197693, + 0.0266324020922184, + 0.007932277396321297, + 0.010471051558852196, + 0.01167750358581543, + 0.008816642686724663, + 0.009166191332042217, + 0.010333959013223648, + 0.012644111178815365, + 0.008829976432025433, + 0.006996603216975927, + 0.0060566384345293045, + 0.009980497881770134, + 0.006013920065015554, + 0.006733452435582876, + 0.007413904182612896, + 0.004322974011301994, + 0.007663565222173929, + 0.005001721903681755, + 0.008370053954422474, + 0.004438675940036774, + 0.00489663053303957, + 0.007675506174564362, + 0.007286102510988712, + 0.004654993768781424, + 0.0045814914628863335, + 0.004156908951699734, + 0.008026625961065292, + 0.005683172959834337, + 0.004574561957269907, + 0.008824390359222889, + 0.010864974930882454, + 0.00522508192807436, + 0.006972332019358873, + 0.005546333268284798, + 0.00995080079883337, + 0.007548089604824781, + 0.0038620212581008673, + 0.017261257395148277, + 0.01100927870720625, + 0.0036270590499043465, + 0.00646180659532547, + 0.005946163088083267, + 0.008362606167793274, + 0.005742364097386599, + 0.00481946486979723, + 0.006914662662893534, + 0.006509141530841589, + 0.005439753644168377, + 0.006811716593801975, + 0.006208560895174742, + 0.00948404986411333, + 0.0062877521850168705, + 0.006195761729031801, + 0.01079364400357008, + 0.00831840094178915, + 0.007286352105438709, + 0.006199815310537815, + 0.012427435256540775, + 0.010431835427880287, + 0.015597413294017315, + 0.008152145892381668, + 0.012019602581858635, + 0.013368570245802402, + 0.008956686593592167, + 0.010689552873373032, + 0.003501519560813904, + 0.002866020193323493, + 0.005920727271586657, + 0.0029102719854563475, + 0.0028843898326158524, + 0.0030638198368251324, + 0.003041811054572463, + 0.0038501506205648184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.009773852303624153 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the initial scene set in a dimly lit tournament arena with blurred silhouettes of a cheering crowd in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the first totem composed of rough, dark gray rock riddled with cracks and adorned with a faint flickering flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the first totem violently shatter into sharp fragments and dust, followed by a blinding white flash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the second totem crafted from smooth, reflective brilliant golden metal with a blazing golden-white flame at its tip?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera shift focus from the crumbling debris to the golden totem and stop on a static shot illuminated by its golden light?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277544617652893 + ], + "opposite_logprob_tokens": [ + -2.017775535583496 + ], + "logprob_reward": -0.14277544617652893, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14277544617652893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11482729762792587, + 0.09733685106039047, + 0.06541077792644501, + 0.153390571475029, + 0.1251368224620819, + 0.07608865201473236, + 0.1439099758863449, + 0.0778530016541481, + 0.107911616563797, + 0.0381343849003315 + ], + "credit_map": [ + 0.016701633110642433, + 0.011084415949881077, + 0.013471981510519981, + 0.021818555891513824, + 0.012197619304060936, + 0.008930517360568047, + 0.02218211069703102, + 0.00844047125428915, + 0.010120846331119537, + 0.013682544231414795, + 0.014104065485298634, + 0.012330683879554272, + 0.011783777736127377, + 0.01167162973433733, + 0.013645823113620281, + 0.009997480548918247, + 0.00784322153776884, + 0.010399759747087955, + 0.010170596651732922, + 0.0069723837077617645, + 0.007707071490585804, + 0.00883151963353157, + 0.006629663519561291, + 0.006856563501060009, + 0.015951350331306458, + 0.027945319190621376, + 0.014920615591108799, + 0.011988500133156776, + 0.03237895667552948, + 0.024557529017329216, + 0.013234364800155163, + 0.012413936667144299, + 0.008845566771924496, + 0.02408910170197487, + 0.017369931563735008, + 0.00864565372467041, + 0.02735772170126438, + 0.020008081570267677, + 0.007449331693351269, + 0.011371436528861523, + 0.006763806566596031, + 0.015364700928330421, + 0.009737006388604641, + 0.004541235510259867, + 0.015962831676006317, + 0.011725641787052155, + 0.004391063936054707, + 0.007602368947118521, + 0.014331876300275326, + 0.021154755726456642, + 0.013557404279708862, + 0.01554456353187561, + 0.026490775868296623, + 0.020087208598852158, + 0.015744740143418312, + 0.016998641192913055, + 0.008596803992986679, + 0.010696307756006718, + 0.00841063354164362, + 0.008417624048888683, + 0.012024365365505219, + 0.009069972671568394, + 0.013118271715939045, + 0.007519016973674297, + 0.013447711244225502, + 0.01240992546081543, + 0.016904162243008614, + 0.008612140081822872, + 0.014161045663058758, + 0.015040243044495583, + 0.013078081421554089, + 0.0142583092674613, + 0.005132078193128109, + 0.0050535742193460464, + 0.006635850761085749, + 0.004164732061326504, + 0.003554724855348468, + 0.004970385227352381, + 0.004830069374293089, + 0.003792969509959221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 4 + ], + "selected_reward": -0.14277544617652893 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_5/index_tensor([6634, 6634])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_5/index_tensor([6634, 6634])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ff73a88225b5445d4051242861ba9199ee54e755 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_5/index_tensor([6634, 6634])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static medium shot, positioned at a slightly elevated angle, observes a pristine, smooth white crystal kitchen counter. Soft, bright, indirect light streams in from an unseen window, casting gentle shadows and creating a clean, tranquil atmosphere. At the center of the frame lies a folded, round, bright cyan silicone food container, its concentric rings forming a flat disc. A subtle, almost imperceptible vibration creates ripples across the polished counter surface, causing the container's small, translucent latch to momentarily loosen before decisively withdrawing from its slot. The container immediately springs to life in a swift, satisfying motion, unfolding upward, its accordion-like walls rapidly expanding to their full height, transforming from a flat disc into a solid rectangular bowl in a fraction of a second. The now fully formed container stands rigidly, its vibrant cyan surface catching the diffused light, appearing brighter and more saturated against the minimalist backdrop, capturing the moment of a sudden, elegant mechanical transformation.", + "seed": 33414, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_5/index_6634_seed_33414_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07657618774101138, + "vlm_logprob": -0.07657618774101138, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.07657618774101138, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06887222081422806 + ], + "opposite_logprob_tokens": [ + -3.0688722133636475 + ], + "logprob_reward": -0.06887222081422806, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06887222081422806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31293344497680664, + 0.09629169851541519, + 0.03302019089460373, + 0.10412506014108658, + 0.09942372143268585, + 0.11118483543395996, + 0.06889436393976212, + 0.05201656371355057, + 0.07624800503253937, + 0.045862119644880295 + ], + "credit_map": [ + 0.036804426461458206, + 0.02798122726380825, + 0.031028712168335915, + 0.0369347408413887, + 0.03854735568165779, + 0.048299163579940796, + 0.02268362045288086, + 0.07065421342849731, + 0.014677613042294979, + 0.010986030101776123, + 0.010967586189508438, + 0.008621840737760067, + 0.010382020846009254, + 0.0194973424077034, + 0.008311447687447071, + 0.01284781564027071, + 0.003949839621782303, + 0.003665377851575613, + 0.0035358467139303684, + 0.0038372401613742113, + 0.0041830032132565975, + 0.004281512461602688, + 0.0039626723155379295, + 0.00560469925403595, + 0.0126339765265584, + 0.012341401539742947, + 0.0099040437489748, + 0.017813656479120255, + 0.009876813739538193, + 0.017714010551571846, + 0.010798810981214046, + 0.013042348437011242, + 0.012995582073926926, + 0.011786265298724174, + 0.010620078071951866, + 0.008346941322088242, + 0.009500485844910145, + 0.022138221189379692, + 0.007173128891736269, + 0.016863012686371803, + 0.013060483150184155, + 0.011166187934577465, + 0.01309364102780819, + 0.011863759718835354, + 0.014272584579885006, + 0.008887052536010742, + 0.011765445582568645, + 0.02707567997276783, + 0.00954176764935255, + 0.011423910968005657, + 0.00801707524806261, + 0.007765292655676603, + 0.009824084118008614, + 0.008844267576932907, + 0.005571715533733368, + 0.007906246930360794, + 0.005441809538751841, + 0.004888520110398531, + 0.006022825837135315, + 0.0067636179737746716, + 0.006346928421407938, + 0.004153999965637922, + 0.006280101835727692, + 0.012118758633732796, + 0.008464018814265728, + 0.0069358027540147305, + 0.011070395819842815, + 0.019595211371779442, + 0.0047852881252765656, + 0.009381559677422047, + 0.007246045395731926, + 0.008769682608544827, + 0.005922834854573011, + 0.006910743657499552, + 0.005605853628367186, + 0.008295747451484203, + 0.003984313458204269, + 0.006118322256952524, + 0.004081541206687689, + 0.004942764528095722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.06887222081422806 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01612784154713154 + ], + "opposite_logprob_tokens": [ + -4.141128063201904 + ], + "logprob_reward": -0.01612784154713154, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01612784154713154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2581101357936859, + 0.07143978029489517, + 0.04947555810213089, + 0.13774095475673676, + 0.0806814581155777, + 0.08805454522371292, + 0.11016372591257095, + 0.11155188828706741, + 0.03394648805260658, + 0.05883548781275749 + ], + "credit_map": [ + 0.03506770357489586, + 0.03594985976815224, + 0.026401886716485023, + 0.02334575355052948, + 0.03253916651010513, + 0.037462957203388214, + 0.01824537105858326, + 0.04909740388393402, + 0.008519314229488373, + 0.008983954787254333, + 0.008649921044707298, + 0.006306472700089216, + 0.01324135810136795, + 0.008968275971710682, + 0.004586216993629932, + 0.012184257619082928, + 0.005442939233034849, + 0.005413438659161329, + 0.006301775574684143, + 0.004095235373824835, + 0.0067349812015891075, + 0.00816349033266306, + 0.006031294818967581, + 0.007292401511222124, + 0.014167005196213722, + 0.015023062005639076, + 0.015950962901115417, + 0.019684765487909317, + 0.010553538799285889, + 0.01375112310051918, + 0.013852119445800781, + 0.03475836291909218, + 0.009771659970283508, + 0.008795798756182194, + 0.00794792827218771, + 0.007443083915859461, + 0.008671150542795658, + 0.012266864068806171, + 0.010402967222034931, + 0.01538199931383133, + 0.012716107070446014, + 0.009615733288228512, + 0.007975874468684196, + 0.009707724675536156, + 0.010148024186491966, + 0.008372901938855648, + 0.008070264011621475, + 0.021447908133268356, + 0.01726972870528698, + 0.018992280587553978, + 0.012127331458032131, + 0.01019479800015688, + 0.014066300354897976, + 0.015603845939040184, + 0.009783931076526642, + 0.012125503271818161, + 0.015541872009634972, + 0.010263095609843731, + 0.010945603251457214, + 0.01202249713242054, + 0.010297629982233047, + 0.015539545565843582, + 0.012923243455588818, + 0.02401839755475521, + 0.005649232771247625, + 0.003901939606294036, + 0.003912445157766342, + 0.0026225324254482985, + 0.0029137730598449707, + 0.008199798874557018, + 0.0025222923140972853, + 0.004224472679197788, + 0.010571594350039959, + 0.009530200622975826, + 0.004864188376814127, + 0.004211801569908857, + 0.006211781408637762, + 0.011854872107505798, + 0.005276109091937542, + 0.006314933765679598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.01612784154713154 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002211864572018385 + ], + "opposite_logprob_tokens": [ + -6.127212047576904 + ], + "logprob_reward": -0.002211864572018385, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002211864572018385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18352268636226654, + 0.12828083336353302, + 0.07795125246047974, + 0.10559746623039246, + 0.12362588942050934, + 0.15326328575611115, + 0.15260732173919678, + 0.033521078526973724, + 0.022268768399953842, + 0.019361386075615883 + ], + "credit_map": [ + 0.020940613001585007, + 0.020785164088010788, + 0.02071322873234749, + 0.016459977254271507, + 0.018707722425460815, + 0.0301811546087265, + 0.027183864265680313, + 0.028550975024700165, + 0.017731787636876106, + 0.013978011906147003, + 0.014774004928767681, + 0.010632611811161041, + 0.012289796024560928, + 0.021371295675635338, + 0.016811396926641464, + 0.02069193869829178, + 0.007651649881154299, + 0.006782121490687132, + 0.011641236953437328, + 0.007497058715671301, + 0.010439220815896988, + 0.008845861069858074, + 0.006705307401716709, + 0.01838880032300949, + 0.012988220900297165, + 0.012470600195229053, + 0.011311621405184269, + 0.017788441851735115, + 0.009896979667246342, + 0.01103622280061245, + 0.011497553437948227, + 0.01860782504081726, + 0.017906324937939644, + 0.014162644743919373, + 0.013750209473073483, + 0.010444452054798603, + 0.016285941004753113, + 0.01761621981859207, + 0.012687508948147297, + 0.02077259123325348, + 0.02526373602449894, + 0.01928432658314705, + 0.015559868887066841, + 0.01236046478152275, + 0.01394012663513422, + 0.027538247406482697, + 0.013423584401607513, + 0.025892941281199455, + 0.02126445062458515, + 0.020254027098417282, + 0.01553185936063528, + 0.014894613064825535, + 0.019511818885803223, + 0.021008001640439034, + 0.01729096844792366, + 0.022851577028632164, + 0.004029138013720512, + 0.004323748406022787, + 0.0032549072057008743, + 0.003365939250215888, + 0.004003542009741068, + 0.004217988345772028, + 0.00404083076864481, + 0.006284988485276699, + 0.003470767755061388, + 0.00257591693662107, + 0.0022703749127686024, + 0.0035112856421619654, + 0.002058441750705242, + 0.003194926306605339, + 0.002210981911048293, + 0.0029760743491351604, + 0.0028304012957960367, + 0.002642316510900855, + 0.0018698713975027204, + 0.0024963021278381348, + 0.001972687430679798, + 0.0033057951368391514, + 0.001945315976627171, + 0.0022986973635852337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.002211864572018385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002348452806473 + ], + "opposite_logprob_tokens": [ + -2.695023536682129 + ], + "logprob_reward": -0.07002348452806473, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002348452806473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18839894235134125, + 0.10633872449398041, + 0.03640628233551979, + 0.23227591812610626, + 0.09138309955596924, + 0.1123342216014862, + 0.06631722301244736, + 0.07408406585454941, + 0.04516226425766945, + 0.04729928448796272 + ], + "credit_map": [ + 0.02686847560107708, + 0.023515723645687103, + 0.023051762953400612, + 0.01488508004695177, + 0.01862436719238758, + 0.02857021987438202, + 0.029247073456645012, + 0.0236362237483263, + 0.015206214040517807, + 0.011258182115852833, + 0.011059459298849106, + 0.008184944279491901, + 0.010647210292518139, + 0.021137919276952744, + 0.014905625954270363, + 0.013939173892140388, + 0.005456516053527594, + 0.004300759639590979, + 0.004206612706184387, + 0.0028230170719325542, + 0.0039050092454999685, + 0.006549661513417959, + 0.005543791688978672, + 0.003620915114879608, + 0.022957779467105865, + 0.02148962765932083, + 0.02868952788412571, + 0.02960633486509323, + 0.014883449301123619, + 0.023561328649520874, + 0.023733114823698997, + 0.0673547312617302, + 0.010014046914875507, + 0.008624430745840073, + 0.008497387170791626, + 0.009262381121516228, + 0.00907113216817379, + 0.010472799651324749, + 0.013789799064397812, + 0.021651122719049454, + 0.014224008657038212, + 0.01378554105758667, + 0.010818330571055412, + 0.013172997161746025, + 0.014456161297857761, + 0.008246801793575287, + 0.013978804461658001, + 0.023651570081710815, + 0.009034352377057076, + 0.008369909599423409, + 0.0065593840554356575, + 0.005968014243990183, + 0.011880338191986084, + 0.009094719775021076, + 0.008805621415376663, + 0.006604882422834635, + 0.008783506229519844, + 0.010270098224282265, + 0.006922969128936529, + 0.007748514413833618, + 0.011290345340967178, + 0.010523972101509571, + 0.008987730368971825, + 0.009556925855576992, + 0.00573954451829195, + 0.003992700949311256, + 0.006415152456611395, + 0.00626805005595088, + 0.007138436660170555, + 0.00507873622700572, + 0.004523609299212694, + 0.0060060336254537106, + 0.005291676614433527, + 0.00850677490234375, + 0.004159391857683659, + 0.004358105827122927, + 0.008113767020404339, + 0.005442475434392691, + 0.006884636823087931, + 0.004542455542832613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.07002348452806473 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256455272436142 + ], + "opposite_logprob_tokens": [ + -1.6006455421447754 + ], + "logprob_reward": -0.2256455272436142, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256455272436142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15936830639839172, + 0.0728791207075119, + 0.10931076854467392, + 0.15715061128139496, + 0.16358314454555511, + 0.129092738032341, + 0.06726265698671341, + 0.06641090661287308, + 0.04430434852838516, + 0.0306374691426754 + ], + "credit_map": [ + 0.01568068191409111, + 0.015644770115613937, + 0.01936078444123268, + 0.018792806193232536, + 0.019033191725611687, + 0.014455873519182205, + 0.01728549599647522, + 0.039114683866500854, + 0.009355179034173489, + 0.009114756248891354, + 0.008566807024180889, + 0.00607518432661891, + 0.010759655386209488, + 0.011423621326684952, + 0.00951385498046875, + 0.008070061914622784, + 0.013266857713460922, + 0.01162875909358263, + 0.012012703344225883, + 0.01081549096852541, + 0.014482539147138596, + 0.023478563874959946, + 0.011567404493689537, + 0.012058434076607227, + 0.017955774441361427, + 0.01901937648653984, + 0.01676357537508011, + 0.017399590462446213, + 0.014395414851605892, + 0.016046255826950073, + 0.01987062580883503, + 0.035699985921382904, + 0.01740942895412445, + 0.015157334506511688, + 0.018330112099647522, + 0.012992369942367077, + 0.024683108553290367, + 0.022959237918257713, + 0.02020048163831234, + 0.03185105696320534, + 0.020159557461738586, + 0.015349066816270351, + 0.014685047790408134, + 0.012571060098707676, + 0.014667367562651634, + 0.010830510407686234, + 0.011165161617100239, + 0.02966495417058468, + 0.010470010340213776, + 0.008715329691767693, + 0.006880114320665598, + 0.008391104638576508, + 0.008863546885550022, + 0.006925108842551708, + 0.005616362672299147, + 0.011401071213185787, + 0.009139631874859333, + 0.007775936741381884, + 0.007213289849460125, + 0.0076563055627048016, + 0.0072396788746118546, + 0.005852774251252413, + 0.006656464654952288, + 0.01487682294100523, + 0.005191899370402098, + 0.004960949067026377, + 0.005735504440963268, + 0.008415624499320984, + 0.006202662363648415, + 0.004402472637593746, + 0.005028416868299246, + 0.004366815090179443, + 0.004126973450183868, + 0.002631888957694173, + 0.0040874481201171875, + 0.0047042942605912685, + 0.0019951933063566685, + 0.007407642435282469, + 0.0023508823942393064, + 0.0033331462182104588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.2256455272436142 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, positioned at a slightly elevated angle, observes a pristine, smooth white crystal kitchen counter. Soft, bright, indirect light streams in from an unseen window, casting gentle shadows and creating a clean, tranquil atmosphere. At the center of the frame lies a folded, round, bright cyan silicone food container, its concentric rings forming a flat disc. A subtle, almost imperceptible vibration creates ripples across the polished counter surface, causing the container's small, translucent latch to momentarily loosen before decisively withdrawing from its slot. The container immediately springs to life in a swift, satisfying motion, unfolding upward, its accordion-like walls rapidly expanding to their full height, transforming from a flat disc into a solid rectangular bowl in a fraction of a second. The now fully formed container stands rigidly, its vibrant cyan surface catching the diffused light, appearing brighter and more saturated against the minimalist backdrop, capturing the moment of a sudden, elegant mechanical transformation.", + "seed": 33415, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_5/index_6634_seed_33415_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07116203531622886, + "vlm_logprob": -0.07116203531622886, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.07116203531622886, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019270699471235275 + ], + "opposite_logprob_tokens": [ + -4.394270896911621 + ], + "logprob_reward": -0.019270699471235275, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.019270699471235275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1023780107498169, + 0.2089838981628418, + 0.15642286837100983, + 0.06798538565635681, + 0.09521529078483582, + 0.07194268703460693, + 0.13821399211883545, + 0.054188407957553864, + 0.06734977662563324, + 0.03731970489025116 + ], + "credit_map": [ + 0.015806201845407486, + 0.01069008931517601, + 0.016941966488957405, + 0.012827236205339432, + 0.007971012964844704, + 0.008687343448400497, + 0.017723323777318, + 0.011730828322470188, + 0.0205644890666008, + 0.032840095460414886, + 0.03260546177625656, + 0.026059161871671677, + 0.013602829538285732, + 0.022112581878900528, + 0.03973139449954033, + 0.02146787941455841, + 0.01896592415869236, + 0.025438537821173668, + 0.027300981804728508, + 0.014089866541326046, + 0.01397558581084013, + 0.02135918103158474, + 0.01615193858742714, + 0.019140830263495445, + 0.007462755776941776, + 0.011047075502574444, + 0.012047531083226204, + 0.006917315535247326, + 0.00630197674036026, + 0.006885450333356857, + 0.00859021581709385, + 0.008733061142265797, + 0.015954041853547096, + 0.010094515979290009, + 0.01600838638842106, + 0.00965350866317749, + 0.01108191441744566, + 0.00936292763799429, + 0.008410916663706303, + 0.014649070799350739, + 0.008830365724861622, + 0.015948623418807983, + 0.011756516993045807, + 0.007102591916918755, + 0.0056534637697041035, + 0.004899912513792515, + 0.006413708906620741, + 0.011337497271597385, + 0.01213917974382639, + 0.022218089550733566, + 0.017908863723278046, + 0.02111770212650299, + 0.014141897670924664, + 0.008738772943615913, + 0.025882262736558914, + 0.01606721617281437, + 0.008778664283454418, + 0.0048662275075912476, + 0.008939554914832115, + 0.006055071018636227, + 0.005681213457137346, + 0.005913775879889727, + 0.006774371489882469, + 0.007179525680840015, + 0.012036736123263836, + 0.0053382813930511475, + 0.011793693527579308, + 0.006946412846446037, + 0.006591334939002991, + 0.009336564689874649, + 0.006736467592418194, + 0.00857027992606163, + 0.004965933505445719, + 0.004527908284217119, + 0.00594334676861763, + 0.005342037416994572, + 0.004458992276340723, + 0.002513373037800193, + 0.004677322693169117, + 0.004890786949545145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 6 + ], + "selected_reward": -0.019270699471235275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012568432837724686 + ], + "opposite_logprob_tokens": [ + -4.387568473815918 + ], + "logprob_reward": -0.012568432837724686, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012568432837724686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08751428127288818, + 0.06671541184186935, + 0.11419747024774551, + 0.1445375680923462, + 0.08817671984434128, + 0.08815141022205353, + 0.1251363605260849, + 0.10787247866392136, + 0.14434762299060822, + 0.033350732177495956 + ], + "credit_map": [ + 0.013403866440057755, + 0.009476367384195328, + 0.013313417322933674, + 0.0101954136043787, + 0.006968025583773851, + 0.009935902431607246, + 0.014966909773647785, + 0.009254369884729385, + 0.008066222071647644, + 0.008991774171590805, + 0.01094616949558258, + 0.006338491570204496, + 0.005711117293685675, + 0.009629612788558006, + 0.008697603829205036, + 0.008334418758749962, + 0.011394371278584003, + 0.015934333205223083, + 0.019181277602910995, + 0.010230140760540962, + 0.011000254191458225, + 0.010970246978104115, + 0.01872229389846325, + 0.01676454208791256, + 0.018942993134260178, + 0.021329214796423912, + 0.02514749951660633, + 0.011590264737606049, + 0.010085862129926682, + 0.021092094480991364, + 0.009985147044062614, + 0.02636447735130787, + 0.012897293083369732, + 0.011689627543091774, + 0.01348823495209217, + 0.009291027672588825, + 0.012618908658623695, + 0.0065040127374231815, + 0.012632360681891441, + 0.0090552419424057, + 0.009994513355195522, + 0.014958469197154045, + 0.014289052225649357, + 0.009617245756089687, + 0.008528823032975197, + 0.008663864806294441, + 0.00872423779219389, + 0.013375197537243366, + 0.012803827412426472, + 0.01778644509613514, + 0.019262149930000305, + 0.016437707468867302, + 0.012832297943532467, + 0.009432122111320496, + 0.02017124742269516, + 0.016410542652010918, + 0.01330067403614521, + 0.010991851799190044, + 0.01583205908536911, + 0.016247939318418503, + 0.013488061726093292, + 0.009817343205213547, + 0.016047824174165726, + 0.012146711349487305, + 0.014546716585755348, + 0.013731096871197224, + 0.023208672180771828, + 0.016817288473248482, + 0.019198918715119362, + 0.011311818845570087, + 0.021221237257122993, + 0.024311836808919907, + 0.0054327393881976604, + 0.004384736530482769, + 0.005973484832793474, + 0.003247916931286454, + 0.0036810056772083044, + 0.0028246406000107527, + 0.0028686642181128263, + 0.004937541671097279 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.012568432837724686 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032059960067272186 + ], + "opposite_logprob_tokens": [ + -5.7532057762146 + ], + "logprob_reward": -0.0032059960067272186, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032059960067272186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14836689829826355, + 0.11582301557064056, + 0.26044970750808716, + 0.03235570713877678, + 0.029228845611214638, + 0.06279139965772629, + 0.131678506731987, + 0.0677829384803772, + 0.13363873958587646, + 0.017884287983179092 + ], + "credit_map": [ + 0.018877029418945312, + 0.02007937617599964, + 0.022958509624004364, + 0.02005629800260067, + 0.010346777737140656, + 0.013469461351633072, + 0.02294010855257511, + 0.01963934302330017, + 0.013655910268425941, + 0.018094735220074654, + 0.018069162964820862, + 0.015362930484116077, + 0.006867722142487764, + 0.012277135625481606, + 0.01913370005786419, + 0.012361719273030758, + 0.0320354588329792, + 0.045644454658031464, + 0.04741403087973595, + 0.021717291325330734, + 0.02149362303316593, + 0.03465147316455841, + 0.021641047671437263, + 0.03585235774517059, + 0.002779127564281225, + 0.0026405577082186937, + 0.0053148940205574036, + 0.0040265279822051525, + 0.004608101211488247, + 0.0024028995539993048, + 0.005988959223031998, + 0.004594642203301191, + 0.0028349936474114656, + 0.0027863644063472748, + 0.004831624682992697, + 0.004407815635204315, + 0.0018161532934755087, + 0.001994131598621607, + 0.005794385448098183, + 0.0047633773647248745, + 0.008068257011473179, + 0.006034822668880224, + 0.01034678053110838, + 0.0092162499204278, + 0.007357419468462467, + 0.004791271407157183, + 0.007929014973342419, + 0.009047584608197212, + 0.012840678915381432, + 0.017888959497213364, + 0.018644727766513824, + 0.018875639885663986, + 0.013890859670937061, + 0.008826116099953651, + 0.02361992746591568, + 0.01709159091114998, + 0.006316077895462513, + 0.007616665214300156, + 0.009823603555560112, + 0.011725747026503086, + 0.007951470091938972, + 0.005442237947136164, + 0.010589294135570526, + 0.008317841216921806, + 0.012442230246961117, + 0.01283413078635931, + 0.022523146122694016, + 0.016568804159760475, + 0.018174007534980774, + 0.01066298596560955, + 0.018739458173513412, + 0.021693969145417213, + 0.0022072182036936283, + 0.0016323174349963665, + 0.0026667837519198656, + 0.0023245650809258223, + 0.0023343684151768684, + 0.0018472884548828006, + 0.0023927006404846907, + 0.0024790451861917973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.0032059960067272186 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035906970500946 + ], + "opposite_logprob_tokens": [ + -1.9103590250015259 + ], + "logprob_reward": -0.16035906970500946, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035906970500946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13379907608032227, + 0.11894445866346359, + 0.43466904759407043, + 0.024578845128417015, + 0.019852695986628532, + 0.05181852728128433, + 0.11839558184146881, + 0.03549904376268387, + 0.04851868003606796, + 0.013924134895205498 + ], + "credit_map": [ + 0.01749332994222641, + 0.017392436042428017, + 0.020575597882270813, + 0.017609361559152603, + 0.0093990508466959, + 0.012529573403298855, + 0.021455757319927216, + 0.01734396256506443, + 0.013204391114413738, + 0.01902008429169655, + 0.018919214606285095, + 0.01604052446782589, + 0.007570559624582529, + 0.012357719242572784, + 0.017307274043560028, + 0.01452468428760767, + 0.05607086420059204, + 0.0770828053355217, + 0.07948900014162064, + 0.03413575515151024, + 0.03498820215463638, + 0.05822155624628067, + 0.0344712994992733, + 0.06020951271057129, + 0.0027198633179068565, + 0.002582186833024025, + 0.003722826251760125, + 0.0027832910418510437, + 0.0022351699881255627, + 0.0022543638478964567, + 0.0044390130788087845, + 0.0038421275094151497, + 0.0024958669673651457, + 0.0021493444219231606, + 0.003094395389780402, + 0.0030422911513596773, + 0.0017531337216496468, + 0.0012022280134260654, + 0.003020800184458494, + 0.0030946347396820784, + 0.006303095258772373, + 0.005193213000893593, + 0.008186268620193005, + 0.0076467725448310375, + 0.005507550202310085, + 0.004406408406794071, + 0.007001874502748251, + 0.007573341950774193, + 0.01240910217165947, + 0.01604706421494484, + 0.016924070194363594, + 0.016732845455408096, + 0.012294539250433445, + 0.0076346187852323055, + 0.020115770399570465, + 0.01623755879700184, + 0.0025077241007238626, + 0.003959471359848976, + 0.00497114472091198, + 0.007542496547102928, + 0.003821323625743389, + 0.002187068108469248, + 0.0061184111982584, + 0.004391399212181568, + 0.004655178170651197, + 0.00465267663821578, + 0.008360770530998707, + 0.005797747988253832, + 0.006414526607841253, + 0.0038836828898638487, + 0.006695582997053862, + 0.008058510720729828, + 0.0009817894315347075, + 0.0018142685294151306, + 0.0017778045730665326, + 0.0022759814746677876, + 0.00177717674523592, + 0.0011393929598852992, + 0.0022199428640305996, + 0.0019377778517082334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.16035906970500946 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604059785604477 + ], + "opposite_logprob_tokens": [ + -1.9104059934616089 + ], + "logprob_reward": -0.1604059785604477, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604059785604477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15471310913562775, + 0.16646230220794678, + 0.356008380651474, + 0.06057557836174965, + 0.04959150776267052, + 0.052907370030879974, + 0.03934618830680847, + 0.03349106386303902, + 0.07394969463348389, + 0.012954838573932648 + ], + "credit_map": [ + 0.019873781129717827, + 0.020445210859179497, + 0.024059612303972244, + 0.020108943805098534, + 0.010184667073190212, + 0.017062656581401825, + 0.023207390680909157, + 0.019770869985222816, + 0.01774439588189125, + 0.024097144603729248, + 0.024857565760612488, + 0.022526053711771965, + 0.010017697699368, + 0.016485601663589478, + 0.03284478187561035, + 0.01788906753063202, + 0.04514237493276596, + 0.06184590235352516, + 0.06556874513626099, + 0.028533076867461205, + 0.02867693267762661, + 0.04903823509812355, + 0.026124510914087296, + 0.05107858031988144, + 0.007209520321339369, + 0.005307626444846392, + 0.00986863486468792, + 0.0062650577165186405, + 0.007926120422780514, + 0.0037677159998565912, + 0.010943234898149967, + 0.009287673979997635, + 0.004272581543773413, + 0.005313668865710497, + 0.007308220025151968, + 0.00722232973203063, + 0.007238504011183977, + 0.0032566539011895657, + 0.00860789604485035, + 0.006371655035763979, + 0.00656709261238575, + 0.005222916137427092, + 0.008556750603020191, + 0.006908298935741186, + 0.004410552326589823, + 0.004543924238532782, + 0.007419702596962452, + 0.009278139099478722, + 0.005358906928449869, + 0.004133887123316526, + 0.0058985077776014805, + 0.005648244172334671, + 0.0034363465383648872, + 0.003654774511232972, + 0.005144914612174034, + 0.006070610135793686, + 0.0028483939822763205, + 0.003729474963620305, + 0.004885536152869463, + 0.00604473240673542, + 0.004202177748084068, + 0.002713005756959319, + 0.005093713290989399, + 0.003974032122641802, + 0.007917224429547787, + 0.006676231976598501, + 0.012763232924044132, + 0.009528611786663532, + 0.00883827917277813, + 0.006558955181390047, + 0.009623449295759201, + 0.012043707072734833, + 0.001563173602335155, + 0.001241473713889718, + 0.0019248700700700283, + 0.0016418695449829102, + 0.001455566962249577, + 0.0011612721718847752, + 0.0017237954307347536, + 0.002242817310616374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.1604059785604477 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, positioned at a slightly elevated angle, observes a pristine, smooth white crystal kitchen counter. Soft, bright, indirect light streams in from an unseen window, casting gentle shadows and creating a clean, tranquil atmosphere. At the center of the frame lies a folded, round, bright cyan silicone food container, its concentric rings forming a flat disc. A subtle, almost imperceptible vibration creates ripples across the polished counter surface, causing the container's small, translucent latch to momentarily loosen before decisively withdrawing from its slot. The container immediately springs to life in a swift, satisfying motion, unfolding upward, its accordion-like walls rapidly expanding to their full height, transforming from a flat disc into a solid rectangular bowl in a fraction of a second. The now fully formed container stands rigidly, its vibrant cyan surface catching the diffused light, appearing brighter and more saturated against the minimalist backdrop, capturing the moment of a sudden, elegant mechanical transformation.", + "seed": 33416, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_5/index_6634_seed_33416_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14816164104267954, + "vlm_logprob": -0.14816164104267954, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.14816164104267954, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.29131782054901123 + ], + "opposite_logprob_tokens": [ + -1.4163178205490112 + ], + "logprob_reward": -0.29131782054901123, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.29131782054901123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044489532709121704, + 0.04269390180706978, + 0.025170868262648582, + 0.01239657960832119, + 0.7646273970603943, + 0.03039250336587429, + 0.013223904184997082, + 0.009180168621242046, + 0.011215482838451862, + 0.046609677374362946 + ], + "credit_map": [ + 0.010352689772844315, + 0.0035390700213611126, + 0.007532632444053888, + 0.0035950131714344025, + 0.002524916548281908, + 0.0063262395560741425, + 0.0025729788467288017, + 0.008045990020036697, + 0.005613215733319521, + 0.0035807788372039795, + 0.006398412864655256, + 0.002966358559206128, + 0.002031859941780567, + 0.006950246170163155, + 0.0020527574233710766, + 0.013100272975862026, + 0.00586708215996623, + 0.0019862856715917587, + 0.004068007692694664, + 0.0017482740804553032, + 0.001678040949627757, + 0.00394685473293066, + 0.001267897547222674, + 0.0046084229834377766, + 0.001672144397161901, + 0.0009683517273515463, + 0.0022611108142882586, + 0.0017450593877583742, + 0.0012271445011720061, + 0.001524455612525344, + 0.0013344262260943651, + 0.0016638863598927855, + 0.06509041041135788, + 0.04458176717162132, + 0.05230279639363289, + 0.1953631192445755, + 0.054930344223976135, + 0.06516237556934357, + 0.1787325143814087, + 0.1084640696644783, + 0.007666402030736208, + 0.0025876315776258707, + 0.0064628287218511105, + 0.0023891383316367865, + 0.0017288010567426682, + 0.003044040873646736, + 0.0024807399604469538, + 0.0040329210460186005, + 0.0018130281241610646, + 0.001177096040919423, + 0.0018398147076368332, + 0.0013888163957744837, + 0.0027529727667570114, + 0.0015702616656199098, + 0.0013315740507096052, + 0.001350340317003429, + 0.0010286851320415735, + 0.0011529377661645412, + 0.0009590299450792372, + 0.0012969927629455924, + 0.001401567948050797, + 0.0009974954882636666, + 0.0015753301559016109, + 0.0007681290153414011, + 0.0015929552027955651, + 0.0011432096362113953, + 0.0016350263031199574, + 0.0012568620732054114, + 0.0012369912583380938, + 0.0019768758211284876, + 0.0011645257472991943, + 0.0012090366799384356, + 0.002943460363894701, + 0.005777087528258562, + 0.0038869406562298536, + 0.00685723265632987, + 0.010020713321864605, + 0.004621730651706457, + 0.0085858553647995, + 0.003916656598448753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.29131782054901123 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014254650101065636 + ], + "opposite_logprob_tokens": [ + -4.264254570007324 + ], + "logprob_reward": -0.014254650101065636, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014254650101065636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1379205882549286, + 0.10584672540426254, + 0.031400468200445175, + 0.01754356361925602, + 0.5456375479698181, + 0.06356970965862274, + 0.034560658037662506, + 0.016046669334173203, + 0.010632513090968132, + 0.036841534078121185 + ], + "credit_map": [ + 0.028439702466130257, + 0.011207787320017815, + 0.020367885008454323, + 0.016178438439965248, + 0.007694211322814226, + 0.018281470984220505, + 0.01087209302932024, + 0.02487899176776409, + 0.01419873721897602, + 0.007980700582265854, + 0.014908869750797749, + 0.006912760902196169, + 0.004969148896634579, + 0.01800466515123844, + 0.005113964434713125, + 0.033757876604795456, + 0.007578956428915262, + 0.002209837781265378, + 0.004796516615897417, + 0.0021623976062983274, + 0.0018408913165330887, + 0.005016612820327282, + 0.0014020316302776337, + 0.006393223535269499, + 0.004245263058692217, + 0.0018359256209805608, + 0.0020117356907576323, + 0.0015776501968502998, + 0.0016338553978130221, + 0.002542972331866622, + 0.001442247536033392, + 0.0022539126221090555, + 0.046432074159383774, + 0.03160972148180008, + 0.037519995123147964, + 0.13885468244552612, + 0.03978866711258888, + 0.04689729958772659, + 0.1268797069787979, + 0.07765543460845947, + 0.016749078407883644, + 0.005804113578051329, + 0.01301224622875452, + 0.0041864546947181225, + 0.004372172523289919, + 0.005994919687509537, + 0.004979490302503109, + 0.008471234701573849, + 0.0038912661839276552, + 0.0031069032847881317, + 0.002964850515127182, + 0.00595454964786768, + 0.00920142699033022, + 0.0033526895567774773, + 0.0034079693723469973, + 0.00268100225366652, + 0.0023919830564409494, + 0.0017975916853174567, + 0.002459008479490876, + 0.001605940400622785, + 0.0020873777102679014, + 0.002295261947438121, + 0.0015262680826708674, + 0.0018832386704161763, + 0.001628179452382028, + 0.0011622277088463306, + 0.0018465886823832989, + 0.0013571117306128144, + 0.001374450046569109, + 0.0008979660924524069, + 0.0013796468265354633, + 0.0009863418526947498, + 0.002802410861477256, + 0.0046637291088700294, + 0.0032674791291356087, + 0.005107426084578037, + 0.008393112570047379, + 0.003299730597063899, + 0.006445574574172497, + 0.0028620725497603416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.014254650101065636 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003206114750355482 + ], + "opposite_logprob_tokens": [ + -5.753206253051758 + ], + "logprob_reward": -0.003206114750355482, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003206114750355482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12920762598514557, + 0.027304399758577347, + 0.03816946968436241, + 0.023092670366168022, + 0.6386092305183411, + 0.03486046567559242, + 0.031712606549263, + 0.0208413265645504, + 0.009896751493215561, + 0.04630545154213905 + ], + "credit_map": [ + 0.027960266917943954, + 0.009526602923870087, + 0.01959434524178505, + 0.015228520147502422, + 0.006601440254598856, + 0.017390459775924683, + 0.009787404909729958, + 0.023118574172258377, + 0.006384185515344143, + 0.002795054344460368, + 0.003171731485053897, + 0.0037522066850215197, + 0.0016757299890741706, + 0.0038539995439350605, + 0.0021981773898005486, + 0.0034733160864561796, + 0.006587308831512928, + 0.002890468342229724, + 0.004564268048852682, + 0.0030826996080577374, + 0.002355944598093629, + 0.007792528253048658, + 0.0020456525962799788, + 0.008850596845149994, + 0.0034379763528704643, + 0.0015523219481110573, + 0.00450911046937108, + 0.003694734303280711, + 0.0017894472694024444, + 0.002816529246047139, + 0.002456369809806347, + 0.0028361789882183075, + 0.054458923637866974, + 0.03735072910785675, + 0.044095467776060104, + 0.16266366839408875, + 0.046997278928756714, + 0.05431792140007019, + 0.1486968696117401, + 0.09002836048603058, + 0.009593542665243149, + 0.0034113789442926645, + 0.007520706858485937, + 0.0022078915499150753, + 0.001866556005552411, + 0.003062994685024023, + 0.0024463406298309565, + 0.0047510541044175625, + 0.0036553998943418264, + 0.0028458726592361927, + 0.004125080071389675, + 0.004188749939203262, + 0.008536859415471554, + 0.002829762641340494, + 0.002771653002128005, + 0.002759226830676198, + 0.0021644949447363615, + 0.0026718229055404663, + 0.0020228619687259197, + 0.0029363268986344337, + 0.0020767871756106615, + 0.003293336369097233, + 0.0034729826729744673, + 0.002202713629230857, + 0.0012507245410233736, + 0.0010960277868434787, + 0.0014475269708782434, + 0.0011032448383048177, + 0.001298827468417585, + 0.0011358300689607859, + 0.0016910412814468145, + 0.0008735280716791749, + 0.0037914796266704798, + 0.004516903776675463, + 0.005734977312386036, + 0.006217547692358494, + 0.010025441646575928, + 0.0043159425258636475, + 0.008164566941559315, + 0.0035385892260819674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.003206114750355482 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25208619236946106 + ], + "opposite_logprob_tokens": [ + -1.5020861625671387 + ], + "logprob_reward": -0.25208619236946106, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25208619236946106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07701410353183746, + 0.024729212746024132, + 0.07122886925935745, + 0.02927507646381855, + 0.2244676649570465, + 0.08488047122955322, + 0.11624232679605484, + 0.11780817806720734, + 0.0655863881111145, + 0.1887677162885666 + ], + "credit_map": [ + 0.016368890181183815, + 0.005628910847008228, + 0.009960235096514225, + 0.008483652025461197, + 0.004119416698813438, + 0.011300533078610897, + 0.005143607500940561, + 0.016008850187063217, + 0.003927912097424269, + 0.001986738760024309, + 0.004798910114914179, + 0.0027561315800994635, + 0.0022295971866697073, + 0.0035593716893345118, + 0.001755803357809782, + 0.0037147491239011288, + 0.018429996445775032, + 0.005843361373990774, + 0.008413632400333881, + 0.004840298555791378, + 0.0040444983169436455, + 0.012515274807810783, + 0.003868382191285491, + 0.013273426331579685, + 0.003726809984073043, + 0.00301493052393198, + 0.004645042587071657, + 0.00445874547585845, + 0.00292816199362278, + 0.0032578047830611467, + 0.003942401148378849, + 0.0033011792693287134, + 0.019628388807177544, + 0.013281052000820637, + 0.016948705539107323, + 0.05633419379591942, + 0.015009850263595581, + 0.01995105855166912, + 0.05052609369158745, + 0.032788317650556564, + 0.018105702474713326, + 0.010431419126689434, + 0.015237187966704369, + 0.00630829855799675, + 0.007942216470837593, + 0.008991730399429798, + 0.005079545546323061, + 0.012784370221197605, + 0.010511118918657303, + 0.018727906048297882, + 0.013380191288888454, + 0.017625389620661736, + 0.012231838889420033, + 0.01703292317688465, + 0.015140184201300144, + 0.01159277930855751, + 0.013040955178439617, + 0.013129469938576221, + 0.010133441537618637, + 0.014795137569308281, + 0.021751338616013527, + 0.014827921986579895, + 0.01968894898891449, + 0.010440964251756668, + 0.007191994693130255, + 0.007201068568974733, + 0.00861653033643961, + 0.007385286968201399, + 0.00991733931005001, + 0.012316963635385036, + 0.006720441393554211, + 0.006236759014427662, + 0.011282967403531075, + 0.02251524105668068, + 0.01739964820444584, + 0.028036896139383316, + 0.041313108056783676, + 0.01772666908800602, + 0.03552631661295891, + 0.014966879040002823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.25208619236946106 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994342744350433 + ], + "opposite_logprob_tokens": [ + -1.8049434423446655 + ], + "logprob_reward": -0.17994342744350433, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994342744350433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027593230828642845, + 0.2602667510509491, + 0.04733456298708916, + 0.049365729093551636, + 0.21614064276218414, + 0.05481119826436043, + 0.09094563126564026, + 0.030703941360116005, + 0.052994877099990845, + 0.16984346508979797 + ], + "credit_map": [ + 0.004052482545375824, + 0.0023457431234419346, + 0.0039136591367423534, + 0.003183528780937195, + 0.0020084609277546406, + 0.004037017002701759, + 0.002808307996019721, + 0.005244031082838774, + 0.034511785954236984, + 0.018224749714136124, + 0.03901684284210205, + 0.018570149317383766, + 0.011734399944543839, + 0.04388606920838356, + 0.013098827563226223, + 0.08122391998767853, + 0.009910103864967823, + 0.002863615984097123, + 0.006549646146595478, + 0.004495853558182716, + 0.0031997424084693193, + 0.008202783763408661, + 0.003609166946262121, + 0.008503645658493042, + 0.009809490293264389, + 0.004869313910603523, + 0.006066469009965658, + 0.007026663515716791, + 0.004254324361681938, + 0.006451912224292755, + 0.00508801219984889, + 0.005799542646855116, + 0.018355289474129677, + 0.012874029576778412, + 0.015923574566841125, + 0.05492072552442551, + 0.015883993357419968, + 0.018331779167056084, + 0.04951242730021477, + 0.030338827520608902, + 0.011387803591787815, + 0.006431496236473322, + 0.009566383436322212, + 0.004483708180487156, + 0.003932178486138582, + 0.006576162297278643, + 0.004802848678082228, + 0.0076306164264678955, + 0.012792270630598068, + 0.012168247252702713, + 0.012950249016284943, + 0.01062024012207985, + 0.013456851243972778, + 0.01034458912909031, + 0.009035894647240639, + 0.009577286429703236, + 0.0032724577467888594, + 0.003982235677540302, + 0.004209134262055159, + 0.0038635602686554193, + 0.0060282000340521336, + 0.0029079339001327753, + 0.004113534465432167, + 0.0023268857039511204, + 0.004165247082710266, + 0.006736776325851679, + 0.004657738376408815, + 0.0063042715191841125, + 0.012380212545394897, + 0.005922622047364712, + 0.008342742919921875, + 0.0044852690771222115, + 0.010490522719919682, + 0.0192719679325819, + 0.015103526413440704, + 0.025706524029374123, + 0.037042759358882904, + 0.014912394806742668, + 0.03370543569326401, + 0.01361031737178564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.17994342744350433 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static medium shot, positioned at a slightly elevated angle, observes a pristine, smooth white crystal kitchen counter. Soft, bright, indirect light streams in from an unseen window, casting gentle shadows and creating a clean, tranquil atmosphere. At the center of the frame lies a folded, round, bright cyan silicone food container, its concentric rings forming a flat disc. A subtle, almost imperceptible vibration creates ripples across the polished counter surface, causing the container's small, translucent latch to momentarily loosen before decisively withdrawing from its slot. The container immediately springs to life in a swift, satisfying motion, unfolding upward, its accordion-like walls rapidly expanding to their full height, transforming from a flat disc into a solid rectangular bowl in a fraction of a second. The now fully formed container stands rigidly, its vibrant cyan surface catching the diffused light, appearing brighter and more saturated against the minimalist backdrop, capturing the moment of a sudden, elegant mechanical transformation.", + "seed": 33417, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_5/index_6634_seed_33417_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10359588665887713, + "vlm_logprob": -0.10359588665887713, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.10359588665887713, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04225464537739754 + ], + "opposite_logprob_tokens": [ + -3.417254686355591 + ], + "logprob_reward": -0.04225464537739754, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04225464537739754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1408824771642685, + 0.13523083925247192, + 0.334584504365921, + 0.06678736209869385, + 0.032631561160087585, + 0.04496448487043381, + 0.044533949345350266, + 0.048296477645635605, + 0.08229795843362808, + 0.06979037821292877 + ], + "credit_map": [ + 0.019684648141264915, + 0.010252957232296467, + 0.01926615461707115, + 0.008258908055722713, + 0.013099015690386295, + 0.026111118495464325, + 0.02508731372654438, + 0.01912236213684082, + 0.015261735767126083, + 0.010250216349959373, + 0.0291072279214859, + 0.009418627247214317, + 0.010302809067070484, + 0.03038778528571129, + 0.010476037859916687, + 0.020026395097374916, + 0.03816104680299759, + 0.020979231223464012, + 0.07502435147762299, + 0.020345449447631836, + 0.022610904648900032, + 0.0833643227815628, + 0.022081702947616577, + 0.052017491310834885, + 0.008535590954124928, + 0.006825596559792757, + 0.010099107399582863, + 0.0054176184348762035, + 0.005990720819681883, + 0.012350854463875294, + 0.009482672438025475, + 0.008085197769105434, + 0.00433118874207139, + 0.005974791944026947, + 0.0051450300961732864, + 0.0026291944086551666, + 0.0027245571836829185, + 0.0038529790472239256, + 0.003746317233890295, + 0.004227501805871725, + 0.005628196522593498, + 0.003098276676610112, + 0.009059222415089607, + 0.0032257428392767906, + 0.0035523097030818462, + 0.00552653381600976, + 0.008606522344052792, + 0.006267680786550045, + 0.0069562941789627075, + 0.003354543587192893, + 0.01040264405310154, + 0.0025715003721415997, + 0.0026190588250756264, + 0.008671224117279053, + 0.00286569120362401, + 0.0070929937064647675, + 0.006272120401263237, + 0.002626920584589243, + 0.0073435381054878235, + 0.006246423348784447, + 0.003388413693755865, + 0.010418622754514217, + 0.0047327387146651745, + 0.0072676981799304485, + 0.017568305134773254, + 0.007505625952035189, + 0.014273681677877903, + 0.006504826247692108, + 0.00440637581050396, + 0.011891193687915802, + 0.007300265599042177, + 0.012847683392465115, + 0.010352795012295246, + 0.005697854328900576, + 0.01237998902797699, + 0.007326684892177582, + 0.005476799327880144, + 0.011789466254413128, + 0.007557994220405817, + 0.009208795614540577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.04225464537739754 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009814579039812088 + ], + "opposite_logprob_tokens": [ + -4.634814739227295 + ], + "logprob_reward": -0.009814579039812088, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009814579039812088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15921221673488617, + 0.08379336446523666, + 0.34677422046661377, + 0.05953192338347435, + 0.13425692915916443, + 0.04201492667198181, + 0.07097245007753372, + 0.03140709921717644, + 0.03859388828277588, + 0.03344302996993065 + ], + "credit_map": [ + 0.011749770492315292, + 0.017430046573281288, + 0.022261736914515495, + 0.01067604124546051, + 0.016465187072753906, + 0.030145209282636642, + 0.02865591272711754, + 0.021828316152095795, + 0.007155633065849543, + 0.009271234273910522, + 0.014915037900209427, + 0.007529028691351414, + 0.007144605740904808, + 0.012811004184186459, + 0.013630538247525692, + 0.011336281895637512, + 0.03948981314897537, + 0.02179018221795559, + 0.07728543877601624, + 0.022090217098593712, + 0.022308744490146637, + 0.08483778685331345, + 0.025320952758193016, + 0.05365104600787163, + 0.005176140461117029, + 0.004089283291250467, + 0.007819038815796375, + 0.006693774368613958, + 0.004501678515225649, + 0.012325095944106579, + 0.011216756887733936, + 0.007710153236985207, + 0.014889980666339397, + 0.014041482470929623, + 0.02275778353214264, + 0.006380059290677309, + 0.009614729322493076, + 0.03126566484570503, + 0.015486200340092182, + 0.019821016117930412, + 0.003896405454725027, + 0.002185649471357465, + 0.005141489673405886, + 0.0023568274918943644, + 0.003303934819996357, + 0.009875819087028503, + 0.009141607210040092, + 0.006113188341259956, + 0.006510111968964338, + 0.004776170011609793, + 0.011800743639469147, + 0.0044500264339149, + 0.005208848975598812, + 0.020999623462557793, + 0.0056835985742509365, + 0.011543326079845428, + 0.00367638748139143, + 0.0032942232210189104, + 0.00445004366338253, + 0.0035960348322987556, + 0.0029754680581390858, + 0.006168290041387081, + 0.0033916684333235025, + 0.003854979993775487, + 0.004298617132008076, + 0.0028086656238883734, + 0.006585323251783848, + 0.004729889333248138, + 0.002857930026948452, + 0.0076007964089512825, + 0.0038400785997509956, + 0.005872586276382208, + 0.005619272589683533, + 0.0026136087253689766, + 0.005572466179728508, + 0.002056790515780449, + 0.0027249876875430346, + 0.007181343622505665, + 0.002252007368952036, + 0.005422552116215229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.009814579039812088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004651559051126242 + ], + "opposite_logprob_tokens": [ + -5.3796515464782715 + ], + "logprob_reward": -0.004651559051126242, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004651559051126242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09284679591655731, + 0.263593465089798, + 0.2696847915649414, + 0.1639583557844162, + 0.044499799609184265, + 0.05262940376996994, + 0.03597657009959221, + 0.021100640296936035, + 0.03681311383843422, + 0.01889706775546074 + ], + "credit_map": [ + 0.008317189291119576, + 0.009718293324112892, + 0.012849612161517143, + 0.006864276248961687, + 0.009214256890118122, + 0.01420724205672741, + 0.019375400617718697, + 0.012300522066652775, + 0.02640209160745144, + 0.022243764251470566, + 0.05432567372918129, + 0.02070804312825203, + 0.01842980459332466, + 0.05259370803833008, + 0.03068273328244686, + 0.038207679986953735, + 0.027798205614089966, + 0.016776081174612045, + 0.05584552884101868, + 0.021166592836380005, + 0.0183015838265419, + 0.05439203232526779, + 0.0377248078584671, + 0.037679947912693024, + 0.013304432854056358, + 0.014026809483766556, + 0.025141632184386253, + 0.014467517845332623, + 0.011689545586705208, + 0.039649996906518936, + 0.021859142929315567, + 0.023819271475076675, + 0.0043670907616615295, + 0.006547045893967152, + 0.005889812018722296, + 0.003270632354542613, + 0.004726503975689411, + 0.006462509278208017, + 0.0073226443491876125, + 0.0059135607443749905, + 0.005811003968119621, + 0.003990896977484226, + 0.007436854764819145, + 0.0025611096061766148, + 0.0038892729207873344, + 0.013933039270341396, + 0.007031391840428114, + 0.007975834421813488, + 0.0030223862268030643, + 0.0030864349100738764, + 0.005321613047271967, + 0.0034142853692173958, + 0.002992156893014908, + 0.008640335872769356, + 0.004179652314633131, + 0.005319709423929453, + 0.002087961183860898, + 0.002574786078184843, + 0.003165989648550749, + 0.0022996612824499607, + 0.0021105357445776463, + 0.0040061757899820805, + 0.0020863045938313007, + 0.002769225975498557, + 0.0031183736864477396, + 0.0030884495936334133, + 0.0047981454990804195, + 0.003990948665887117, + 0.0031455003190785646, + 0.00914029125124216, + 0.003760510589927435, + 0.005770892836153507, + 0.00284828944131732, + 0.0015761383110657334, + 0.004163131583482027, + 0.0015374599024653435, + 0.001347638200968504, + 0.003228813176974654, + 0.0012807250022888184, + 0.0029148708563297987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.004651559051126242 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813163101673126 + ], + "opposite_logprob_tokens": [ + -1.4063162803649902 + ], + "logprob_reward": -0.2813163101673126, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813163101673126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1369592398405075, + 0.12282112240791321, + 0.3628770112991333, + 0.057511016726493835, + 0.07259128242731094, + 0.02626853808760643, + 0.09280714392662048, + 0.03172217309474945, + 0.0739409551024437, + 0.022501550614833832 + ], + "credit_map": [ + 0.014201239682734013, + 0.011077869683504105, + 0.026725301519036293, + 0.008268098346889019, + 0.010546436533331871, + 0.02802785113453865, + 0.018819548189640045, + 0.019292892888188362, + 0.011944308876991272, + 0.01110096089541912, + 0.02462616376578808, + 0.009573603980243206, + 0.009174547158181667, + 0.023588400334119797, + 0.014905611984431744, + 0.017907526344060898, + 0.041398875415325165, + 0.023205962032079697, + 0.08127299696207047, + 0.022782685235142708, + 0.024193909019231796, + 0.09010091423988342, + 0.023484351113438606, + 0.05643733590841293, + 0.004707276821136475, + 0.005320303607732058, + 0.009779157117009163, + 0.004907696973532438, + 0.005383508745580912, + 0.010761454701423645, + 0.008435462601482868, + 0.00821615755558014, + 0.006824973504990339, + 0.004299384541809559, + 0.012228712439537048, + 0.0037682887632399797, + 0.004993509501218796, + 0.01935296319425106, + 0.009307865984737873, + 0.011815580539405346, + 0.0023735282011330128, + 0.00192710489500314, + 0.003388676093891263, + 0.001096792402677238, + 0.001965629169717431, + 0.007519523147493601, + 0.0037145924288779497, + 0.004282691515982151, + 0.00805097445845604, + 0.005768526811152697, + 0.014204874634742737, + 0.0067331260070204735, + 0.007198019418865442, + 0.026879463344812393, + 0.008957347832620144, + 0.015014809556305408, + 0.0055834464728832245, + 0.002628495916724205, + 0.0040259468369185925, + 0.002520274370908737, + 0.0023531930055469275, + 0.007513612043112516, + 0.0024595032446086407, + 0.004637700971215963, + 0.009521572850644588, + 0.005052164662629366, + 0.011004481464624405, + 0.004830422345548868, + 0.0045580314472317696, + 0.020162155851721764, + 0.006011371035128832, + 0.012800752185285091, + 0.0023418916389346123, + 0.0022403919138014317, + 0.003513965755701065, + 0.002713580848649144, + 0.0023061134852468967, + 0.0034847247879952192, + 0.00278307287953794, + 0.0031178079079836607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.2813163101673126 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a static medium shot of a pristine, smooth white crystal kitchen counter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a folded, round, bright cyan silicone food container visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the container unfold rapidly from a flat disc into a solid rectangular bowl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the container's latch loosen and withdraw before the container springs open?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17994233965873718 + ], + "opposite_logprob_tokens": [ + -1.8049423694610596 + ], + "logprob_reward": -0.17994233965873718, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17994233965873718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08648296445608139, + 0.17438143491744995, + 0.2856317460536957, + 0.04193160682916641, + 0.12256032228469849, + 0.08575203269720078, + 0.06737890839576721, + 0.03492448478937149, + 0.07575376331806183, + 0.025202663615345955 + ], + "credit_map": [ + 0.007978228852152824, + 0.0070572164840996265, + 0.014261720702052116, + 0.006078986916691065, + 0.008689118549227715, + 0.016245853155851364, + 0.014789842069149017, + 0.011381999589502811, + 0.017369678243994713, + 0.01465004775673151, + 0.03610747307538986, + 0.013594472780823708, + 0.012765072286128998, + 0.03520172834396362, + 0.019418619573116302, + 0.025274351239204407, + 0.031679313629865646, + 0.01842426508665085, + 0.06366242468357086, + 0.018366223201155663, + 0.01950579136610031, + 0.06796255707740784, + 0.0223237331956625, + 0.043707460165023804, + 0.003951425198465586, + 0.0041554877534508705, + 0.005101167131215334, + 0.0037302253767848015, + 0.00476253405213356, + 0.006503937300294638, + 0.008764373138546944, + 0.0049624573439359665, + 0.012214773334562778, + 0.011611456982791424, + 0.020471807569265366, + 0.005648888181895018, + 0.008647236041724682, + 0.030757931992411613, + 0.014242135919630527, + 0.018966078758239746, + 0.008791918866336346, + 0.005047197919338942, + 0.013829415664076805, + 0.00349251925945282, + 0.0061203246004879475, + 0.024014826864004135, + 0.010294652543962002, + 0.014161171391606331, + 0.005833155941218138, + 0.0032428184058517218, + 0.011587543413043022, + 0.005000095348805189, + 0.004538176581263542, + 0.020810168236494064, + 0.00456021586433053, + 0.011806733906269073, + 0.0030174388084560633, + 0.0036038889084011316, + 0.004022387787699699, + 0.003547183470800519, + 0.0030618126038461924, + 0.008512131869792938, + 0.004027065355330706, + 0.0051325722597539425, + 0.0076494100503623486, + 0.008625333197414875, + 0.00966600514948368, + 0.01050036121159792, + 0.007118594367057085, + 0.011744524352252483, + 0.00970360729843378, + 0.010745927691459656, + 0.003294231602922082, + 0.0028093543369323015, + 0.0036762706004083157, + 0.0034070033580064774, + 0.002675033640116453, + 0.0033887000754475594, + 0.0032409706618636847, + 0.002711097477003932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.17994233965873718 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_6/index_tensor([3510, 3510])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_6/index_tensor([3510, 3510])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..97ccac142dd949a3ac0ca4de11613ae6f2a5cde7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_6/index_tensor([3510, 3510])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera begins with a low-angle view of a forgotten wooden toy car on the floor, then slowly and steadily moves forward, focusing on an off-white plush cat climbing structure. From a circular cavity in the middle of the structure, a sleepy, cream-colored kitten cautiously pokes its furry head out. Its fur is soft and fluffy, streaked with light ginger, and shimmers softly in the warm afternoon sunlight slanting in from the window. Its disproportionately large, obsidian-like eyes, moist and bright, blink in a hazy, half-awake state, filled with innocent curiosity about this new world. It tentatively places a fleshy pink paw on the edge of the cavity, then clumsily yet calmly wriggles its entire body completely out of the darkness. Finally, it opens its mouth in a silent yawn, revealing its tiny pink tongue. The entire picture is immersed in a warm and delicate Ghibli animation style, with sparkling dust particles floating in the air, the background is a soft-focus wooden floor and a touch of green potted plants, and the main colors are composed of soft cream, warm wood color and golden sunlight, creating a tranquil, healing and fairy-tale atmosphere.", + "seed": 30294, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_6/index_3510_seed_30294_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.4390942264348269, + "vlm_logprob": -0.4390942264348269, + "vlm_yes_no_margin": 2.950000023841858, + "vlm_token_credit": -0.4390942264348269, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.129260540008545 + ], + "opposite_logprob_tokens": [ + -0.1292606145143509 + ], + "logprob_reward": -2.129260540008545, + "yes_no_margin_reward": -1.9999998807907104, + "token_credit_reward": -2.129260540008545, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09023527055978775, + 0.04961508512496948, + 0.19080480933189392, + 0.1308688223361969, + 0.1360146701335907, + 0.060352981090545654, + 0.13697658479213715, + 0.15531280636787415, + 0.037889569997787476, + 0.011929440312087536 + ], + "credit_map": [ + 0.01002467144280672, + 0.007211874704807997, + 0.012251134030520916, + 0.008628672920167446, + 0.0162031427025795, + 0.009441867470741272, + 0.007494660094380379, + 0.018979232758283615, + 0.006450257729738951, + 0.0032877647317945957, + 0.004924478475004435, + 0.012170068919658661, + 0.005858227144926786, + 0.0052609872072935104, + 0.004069535061717033, + 0.007593757938593626, + 0.020857425406575203, + 0.008845146745443344, + 0.009750258177518845, + 0.006970708258450031, + 0.011616235598921776, + 0.011876583099365234, + 0.06325919926166534, + 0.05762922391295433, + 0.008330472745001316, + 0.01792345568537712, + 0.016009831801056862, + 0.008143888786435127, + 0.009487374685704708, + 0.00896988995373249, + 0.03733358159661293, + 0.024670317769050598, + 0.01116681843996048, + 0.006773444823920727, + 0.048461828380823135, + 0.012848485261201859, + 0.007856443524360657, + 0.009025425650179386, + 0.006405356340110302, + 0.03347685560584068, + 0.0045387279242277145, + 0.005075232125818729, + 0.016392134130001068, + 0.004899622406810522, + 0.004797934554517269, + 0.006185811012983322, + 0.00220574252307415, + 0.016257768496870995, + 0.013348681852221489, + 0.015961607918143272, + 0.03592948615550995, + 0.01772402413189411, + 0.018361151218414307, + 0.010379876010119915, + 0.0076817478984594345, + 0.017589999362826347, + 0.013070564717054367, + 0.024197624996304512, + 0.021666269749403, + 0.01959121786057949, + 0.025287427008152008, + 0.018104733899235725, + 0.009575937874615192, + 0.023819010704755783, + 0.0033732173033058643, + 0.007367451209574938, + 0.0033000255934894085, + 0.005330115091055632, + 0.008148273453116417, + 0.003840359626337886, + 0.002528812037780881, + 0.004001312889158726, + 0.0011444388655945659, + 0.0020698062144219875, + 0.0013768337666988373, + 0.0019510933198034763, + 0.002395001472905278, + 0.00098631891887635, + 0.0008905866416171193, + 0.0011153592495247722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 6 + ], + "selected_reward": -2.129260540008545 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016137342900037766 + ], + "opposite_logprob_tokens": [ + -4.14113712310791 + ], + "logprob_reward": -0.016137342900037766, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016137342900037766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08692679554224014, + 0.04734507203102112, + 0.17604871094226837, + 0.10447525233030319, + 0.1546063870191574, + 0.053941115736961365, + 0.1892954707145691, + 0.12727293372154236, + 0.03825334832072258, + 0.021834934130311012 + ], + "credit_map": [ + 0.008790231309831142, + 0.006226703990250826, + 0.010798498056828976, + 0.006022633984684944, + 0.024177061393857002, + 0.007228704169392586, + 0.005544465035200119, + 0.018138496205210686, + 0.006157249212265015, + 0.0034412196837365627, + 0.0053847599774599075, + 0.010642213746905327, + 0.0068001835606992245, + 0.004809140227735043, + 0.0038946671411395073, + 0.006215638015419245, + 0.018705688416957855, + 0.00820985808968544, + 0.008649676106870174, + 0.006581711117178202, + 0.011022207327187061, + 0.010983943939208984, + 0.05834067240357399, + 0.053554970771074295, + 0.006665146443992853, + 0.015109874308109283, + 0.013005809858441353, + 0.006833408493548632, + 0.007640734780579805, + 0.008752738125622272, + 0.027531282976269722, + 0.018936261534690857, + 0.012370000593364239, + 0.0067493063397705555, + 0.05491233617067337, + 0.013093872927129269, + 0.008946246467530727, + 0.010012996383011341, + 0.006614712532609701, + 0.04190691187977791, + 0.003979000262916088, + 0.004661302547901869, + 0.015098563395440578, + 0.005295775830745697, + 0.005323685240000486, + 0.004748110193759203, + 0.002182808006182313, + 0.012651869095861912, + 0.01800871267914772, + 0.02073408104479313, + 0.05256963521242142, + 0.023463690653443336, + 0.02353757806122303, + 0.012584742158651352, + 0.00978803914040327, + 0.02860899828374386, + 0.01093768049031496, + 0.019925398752093315, + 0.01821085438132286, + 0.016219578683376312, + 0.02010255679488182, + 0.014803548343479633, + 0.00788776483386755, + 0.019185543060302734, + 0.0032180179841816425, + 0.004869155120104551, + 0.00427338108420372, + 0.003761764382943511, + 0.005421024281531572, + 0.005337278824299574, + 0.002021694788709283, + 0.009351029060781002, + 0.002103660488501191, + 0.002133389934897423, + 0.005499592516571283, + 0.002099963603541255, + 0.003007356310263276, + 0.0016463695792481303, + 0.001019109389744699, + 0.00432549137622118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.016137342900037766 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00763675756752491 + ], + "opposite_logprob_tokens": [ + -4.882636547088623 + ], + "logprob_reward": -0.00763675756752491, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00763675756752491, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07847610116004944, + 0.05905260890722275, + 0.2237129956483841, + 0.11073955148458481, + 0.16301371157169342, + 0.05559013783931732, + 0.0927751287817955, + 0.15941309928894043, + 0.04656419903039932, + 0.010662402957677841 + ], + "credit_map": [ + 0.00829567201435566, + 0.007023910526186228, + 0.011143642477691174, + 0.008810478262603283, + 0.013098808005452156, + 0.008908848278224468, + 0.006779794115573168, + 0.014414958655834198, + 0.005354592576622963, + 0.004105053376406431, + 0.0051003191620111465, + 0.010523997247219086, + 0.013203420676290989, + 0.006906593218445778, + 0.004627360496670008, + 0.009231271222233772, + 0.0234089195728302, + 0.014296313747763634, + 0.011157459579408169, + 0.00811737310141325, + 0.013139124028384686, + 0.01634763553738594, + 0.07088825106620789, + 0.06635793298482895, + 0.007960518822073936, + 0.011194797232747078, + 0.014008811675012112, + 0.006651498842984438, + 0.007004713639616966, + 0.006177640985697508, + 0.035858310759067535, + 0.021883277222514153, + 0.013815533369779587, + 0.005459693260490894, + 0.06156240776181221, + 0.01133513543754816, + 0.010151945985853672, + 0.009608421474695206, + 0.005534854251891375, + 0.04554573446512222, + 0.004330819007009268, + 0.00839413795620203, + 0.010003187693655491, + 0.006368569098412991, + 0.005556792486459017, + 0.005434144288301468, + 0.002807122189551592, + 0.012695367448031902, + 0.006707893684506416, + 0.015629010275006294, + 0.011497198604047298, + 0.017823902890086174, + 0.019465217366814613, + 0.006880634464323521, + 0.006966439075767994, + 0.007804845459759235, + 0.01352941244840622, + 0.023789821192622185, + 0.022130463272333145, + 0.019346855580806732, + 0.024661902338266373, + 0.01967674493789673, + 0.009235799312591553, + 0.02704210765659809, + 0.004270034376531839, + 0.008926684968173504, + 0.00451625743880868, + 0.005863096099346876, + 0.00958935171365738, + 0.005232675466686487, + 0.0029904150869697332, + 0.005175688304007053, + 0.0010982806561514735, + 0.0018652050057426095, + 0.001096199150197208, + 0.0023469380103051662, + 0.0014799803029745817, + 0.0008724742801859975, + 0.0006660516955889761, + 0.0012372747296467423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 7 + ], + "selected_reward": -0.00763675756752491 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016078105196356773 + ], + "opposite_logprob_tokens": [ + -4.141077995300293 + ], + "logprob_reward": -0.016078105196356773, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016078105196356773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06660353392362595, + 0.050660230219364166, + 0.16868668794631958, + 0.11540006101131439, + 0.08859468251466751, + 0.028897453099489212, + 0.07720357924699783, + 0.13785287737846375, + 0.06138048321008682, + 0.2047203928232193 + ], + "credit_map": [ + 0.006528925150632858, + 0.0051894411444664, + 0.010561798699200153, + 0.008635512553155422, + 0.010902256704866886, + 0.006641983985900879, + 0.004569280426949263, + 0.01357434131205082, + 0.00495204096660018, + 0.003411811077967286, + 0.004558760207146406, + 0.01141044870018959, + 0.008414501324295998, + 0.005644916091114283, + 0.003755000187084079, + 0.00851275585591793, + 0.01049242913722992, + 0.014356447383761406, + 0.009502537548542023, + 0.00596202164888382, + 0.011148383840918541, + 0.00872533768415451, + 0.05984345078468323, + 0.048656076192855835, + 0.008142114616930485, + 0.014794709160923958, + 0.01600721664726734, + 0.008984561078250408, + 0.007806723937392235, + 0.006738480646163225, + 0.03521428629755974, + 0.017711972817778587, + 0.006626651622354984, + 0.009246102534234524, + 0.01913805678486824, + 0.01223682425916195, + 0.007476200349628925, + 0.017056632786989212, + 0.006206038873642683, + 0.01060817576944828, + 0.0039969454519450665, + 0.003953270614147186, + 0.0027825185097754, + 0.004071143455803394, + 0.0041191792115569115, + 0.0033888863399624825, + 0.002739336108788848, + 0.0038461757358163595, + 0.006853183265775442, + 0.012219606898725033, + 0.008992251008749008, + 0.011504678055644035, + 0.014840533025562763, + 0.006916472688317299, + 0.008512288331985474, + 0.007364562712609768, + 0.013148314319550991, + 0.017652468755841255, + 0.020668644458055496, + 0.01608922891318798, + 0.01458275318145752, + 0.018118126317858696, + 0.007310665212571621, + 0.030282672494649887, + 0.0059113530442118645, + 0.009821899235248566, + 0.00920642726123333, + 0.011720728129148483, + 0.007652695290744305, + 0.0057059689424932, + 0.004819139372557402, + 0.0065422686748206615, + 0.018462594598531723, + 0.009550102986395359, + 0.06468138098716736, + 0.016168849542737007, + 0.010957256890833378, + 0.018023334443569183, + 0.004819453228265047, + 0.06205743923783302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.016078105196356773 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358386501669884 + ], + "opposite_logprob_tokens": [ + -3.6513583660125732 + ], + "logprob_reward": -0.026358386501669884, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358386501669884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10142616927623749, + 0.02993091009557247, + 0.11074723303318024, + 0.08389714360237122, + 0.15345260500907898, + 0.06828168779611588, + 0.10584288835525513, + 0.21333183348178864, + 0.03041556291282177, + 0.1026739552617073 + ], + "credit_map": [ + 0.012030057609081268, + 0.009722656570374966, + 0.016448182985186577, + 0.011388847604393959, + 0.017517412081360817, + 0.013530555181205273, + 0.011130436323583126, + 0.00965801440179348, + 0.004209139849990606, + 0.002343075582757592, + 0.003258438315242529, + 0.005943473894149065, + 0.003691453719511628, + 0.00338638829998672, + 0.002864647191017866, + 0.004234294407069683, + 0.008926155976951122, + 0.00637952471151948, + 0.00580168329179287, + 0.004687892738729715, + 0.007631615735590458, + 0.00541338324546814, + 0.03938233479857445, + 0.03252463787794113, + 0.005450100637972355, + 0.014640831388533115, + 0.01017090491950512, + 0.00537622207775712, + 0.006758471485227346, + 0.006762264762073755, + 0.020733866840600967, + 0.014004490338265896, + 0.013300980441272259, + 0.004683237057179213, + 0.05760788545012474, + 0.009865905158221722, + 0.00942605547606945, + 0.009138980880379677, + 0.0054818675853312016, + 0.043947700411081314, + 0.004888200666755438, + 0.0064154695719480515, + 0.013978381641209126, + 0.006842735223472118, + 0.006079126615077257, + 0.008777872659265995, + 0.00349301565438509, + 0.0178068820387125, + 0.010035236366093159, + 0.011987278237938881, + 0.027291197329759598, + 0.0117977699264884, + 0.013470953330397606, + 0.006817562505602837, + 0.006401692051440477, + 0.01804119534790516, + 0.017886565998196602, + 0.032726120203733444, + 0.02996942400932312, + 0.026854941621422768, + 0.03412434831261635, + 0.024603793397545815, + 0.012755296193063259, + 0.03441134840250015, + 0.0026838972698897123, + 0.005386758595705032, + 0.0037045092321932316, + 0.003616935573518276, + 0.005130698438733816, + 0.003651375649496913, + 0.0021694425959140062, + 0.0040719457902014256, + 0.009523053653538227, + 0.005476681981235743, + 0.033259931951761246, + 0.006033351179212332, + 0.006020818371325731, + 0.009100519120693207, + 0.0021650302223861217, + 0.031094571575522423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.026358386501669884 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a low-angle view of a forgotten wooden toy car on the floor, then slowly and steadily moves forward, focusing on an off-white plush cat climbing structure. From a circular cavity in the middle of the structure, a sleepy, cream-colored kitten cautiously pokes its furry head out. Its fur is soft and fluffy, streaked with light ginger, and shimmers softly in the warm afternoon sunlight slanting in from the window. Its disproportionately large, obsidian-like eyes, moist and bright, blink in a hazy, half-awake state, filled with innocent curiosity about this new world. It tentatively places a fleshy pink paw on the edge of the cavity, then clumsily yet calmly wriggles its entire body completely out of the darkness. Finally, it opens its mouth in a silent yawn, revealing its tiny pink tongue. The entire picture is immersed in a warm and delicate Ghibli animation style, with sparkling dust particles floating in the air, the background is a soft-focus wooden floor and a touch of green potted plants, and the main colors are composed of soft cream, warm wood color and golden sunlight, creating a tranquil, healing and fairy-tale atmosphere.", + "seed": 30295, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_6/index_3510_seed_30295_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02403093557804823, + "vlm_logprob": -0.02403093557804823, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.02403093557804823, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.052263401448726654 + ], + "opposite_logprob_tokens": [ + -3.0522634983062744 + ], + "logprob_reward": -0.052263401448726654, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.052263401448726654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.014983393251895905, + 0.017485657706856728, + 0.011390726082026958, + 0.008888987824320793, + 0.297087162733078, + 0.0069680181331932545, + 0.004759663715958595, + 0.016067584976553917, + 0.04384135454893112, + 0.5785274505615234 + ], + "credit_map": [ + 0.0012220676289871335, + 0.0013198120286688209, + 0.0026790243573486805, + 0.0034533371217548847, + 0.0008073186036199331, + 0.0012091167736798525, + 0.0018857921240851283, + 0.0024069256614893675, + 0.00275639770552516, + 0.0013351843226701021, + 0.001911412924528122, + 0.004853217396885157, + 0.0009077595314010978, + 0.0013119260547682643, + 0.0014151420909911394, + 0.002994620008394122, + 0.001452983939088881, + 0.0010862791677936912, + 0.0017949598841369152, + 0.0016621312825009227, + 0.0008998503908514977, + 0.0014894746709614992, + 0.0014122894499450922, + 0.0015927593922242522, + 0.00122944510076195, + 0.0008368081180378795, + 0.0011174998944625258, + 0.0012530065141618252, + 0.0007214288925752044, + 0.001326314639300108, + 0.0011544297449290752, + 0.0012500555021688342, + 0.0233630184084177, + 0.015295869670808315, + 0.01974368654191494, + 0.00565178832039237, + 0.011702251620590687, + 0.1302364468574524, + 0.010207243263721466, + 0.08088692277669907, + 0.0005548254121094942, + 0.0014458027435466647, + 0.000996425049379468, + 0.0006939255981706083, + 0.0007863279897719622, + 0.0010796255664899945, + 0.0005900647374801338, + 0.0008210218511521816, + 0.0004388038069009781, + 0.0006892699166201055, + 0.0006582575151696801, + 0.00042293494334444404, + 0.000714731402695179, + 0.0008452069014310837, + 0.00042295901221223176, + 0.0005675008869729936, + 0.0016849924577400088, + 0.0016499990597367287, + 0.0031641332898288965, + 0.001253395457752049, + 0.0013597141951322556, + 0.0029791344422847033, + 0.002366736764088273, + 0.0016094836173579097, + 0.004101587925106287, + 0.004408663138747215, + 0.006179505493491888, + 0.004433403257280588, + 0.004923390224575996, + 0.004773884546011686, + 0.00894936639815569, + 0.006071557756513357, + 0.03495773300528526, + 0.007038291078060865, + 0.22307579219341278, + 0.03382204845547676, + 0.03599794954061508, + 0.1899503767490387, + 0.022888630628585815, + 0.030796656385064125 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.052263401448726654 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004118295386433601 + ], + "opposite_logprob_tokens": [ + -5.5041184425354 + ], + "logprob_reward": -0.004118295386433601, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004118295386433601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01020323671400547, + 0.007340546231716871, + 0.02768748067319393, + 0.00738157844170928, + 0.5343491435050964, + 0.006536827422678471, + 0.007604644633829594, + 0.027300504967570305, + 0.07142812758684158, + 0.30016788840293884 + ], + "credit_map": [ + 0.0010937770130112767, + 0.0009710881276987493, + 0.0018759112572297454, + 0.0015100761083886027, + 0.0007208303431980312, + 0.0013825433561578393, + 0.001422050059773028, + 0.0012269606813788414, + 0.000699016556609422, + 0.0007034757290966809, + 0.0007395612774416804, + 0.001820319565013051, + 0.00035176202072761953, + 0.001039551687426865, + 0.0008319653570652008, + 0.0011548936599865556, + 0.003356217872351408, + 0.0027038471307605505, + 0.00402027228847146, + 0.0047053680755198, + 0.002359195612370968, + 0.004239299800246954, + 0.0033140273299068213, + 0.0029892525635659695, + 0.0010783072793856263, + 0.0008017606451176107, + 0.0010910846758633852, + 0.0009586918749846518, + 0.00046715009375475347, + 0.0011697980808094144, + 0.0008489248575642705, + 0.0009658609633333981, + 0.042060814797878265, + 0.027590027078986168, + 0.035300362855196, + 0.010367357172071934, + 0.021370546892285347, + 0.233668714761734, + 0.01843578740954399, + 0.1455555409193039, + 0.0005879176314920187, + 0.0014558227267116308, + 0.0007980447844602168, + 0.0006522212061099708, + 0.0008409954607486725, + 0.0007793726399540901, + 0.0006578254397027194, + 0.0007646277663297951, + 0.000693267909809947, + 0.001030962448567152, + 0.0011863698018714786, + 0.0007049711421132088, + 0.0010058755287900567, + 0.0014729048125445843, + 0.0007070022984407842, + 0.0008032908081077039, + 0.0027192397974431515, + 0.002858612919226289, + 0.005102493334561586, + 0.0025865938514471054, + 0.002530002500861883, + 0.004763533361256123, + 0.003993669990450144, + 0.0027463610749691725, + 0.006803540512919426, + 0.007077761925756931, + 0.010081599466502666, + 0.007169816177338362, + 0.007949844002723694, + 0.007833871059119701, + 0.014650693163275719, + 0.009861000813543797, + 0.01820489950478077, + 0.0036344430409371853, + 0.11577887833118439, + 0.01763119362294674, + 0.01868417300283909, + 0.0983966588973999, + 0.011938872747123241, + 0.015898769721388817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.004118295386433601 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077570728957653 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.011077570728957653, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077570728957653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.004711170680820942, + 0.0068789576180279255, + 0.01316624041646719, + 0.005992154125124216, + 0.3626459240913391, + 0.012632426805794239, + 0.004730506334453821, + 0.009647376835346222, + 0.029118966311216354, + 0.5504763126373291 + ], + "credit_map": [ + 0.0006459876894950867, + 0.0004143835394643247, + 0.0006654779426753521, + 0.0008612457313574851, + 0.0003609070845413953, + 0.0005670239916071296, + 0.0004625451983883977, + 0.0007335993577726185, + 0.0009158355533145368, + 0.0009766072034835815, + 0.0005950202466920018, + 0.0018028501654043794, + 0.0005063803400844336, + 0.0006354611250571907, + 0.0006053049582988024, + 0.0008414981421083212, + 0.0016011210391297936, + 0.0014595474349334836, + 0.0017293578712269664, + 0.002212935360148549, + 0.0013387149665504694, + 0.0022353993263095617, + 0.0011941581033170223, + 0.001395006780512631, + 0.0007832564297132194, + 0.0006088053341954947, + 0.0008858254295773804, + 0.0008296933956444263, + 0.000650567642878741, + 0.0007813467527739704, + 0.0007379814051091671, + 0.0007146772113628685, + 0.029026886448264122, + 0.019334500655531883, + 0.02373865246772766, + 0.007028997875750065, + 0.014136658981442451, + 0.15794327855110168, + 0.012384291738271713, + 0.09905262291431427, + 0.0010870707919821143, + 0.002785312244668603, + 0.0015900229336693883, + 0.0013490415876731277, + 0.0018373633502051234, + 0.0012762481346726418, + 0.0011309640249237418, + 0.0015764044364914298, + 0.00035974144702777267, + 0.0008936096564866602, + 0.00048313027946278453, + 0.0005185614572837949, + 0.0009283755789510906, + 0.0005118305562064052, + 0.0005407675635069609, + 0.0004944896791130304, + 0.000892720534466207, + 0.0010969284921884537, + 0.0019014616264030337, + 0.0008277176530100405, + 0.0009389616316184402, + 0.0016827459912747145, + 0.0012865896569564939, + 0.0010202504927292466, + 0.0028218876104801893, + 0.0029246557969599962, + 0.004011391196399927, + 0.002924228087067604, + 0.0033374065533280373, + 0.0031016834545880556, + 0.005917055066674948, + 0.00408066064119339, + 0.03339913859963417, + 0.006579019594937563, + 0.21285800635814667, + 0.032297369092702866, + 0.034341759979724884, + 0.18004274368286133, + 0.0218261256814003, + 0.029132133349776268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.011077570728957653 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026349911466240883 + ], + "opposite_logprob_tokens": [ + -3.6513500213623047 + ], + "logprob_reward": -0.026349911466240883, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026349911466240883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.004688045009970665, + 0.0070117162540555, + 0.01746169850230217, + 0.00859129335731268, + 0.43975746631622314, + 0.005699068773537874, + 0.005477519240230322, + 0.011152115650475025, + 0.022271322086453438, + 0.477889746427536 + ], + "credit_map": [ + 0.000650732486974448, + 0.0005440311506390572, + 0.0007967503624968231, + 0.0006961809122003615, + 0.0003681017260532826, + 0.0005941817071288824, + 0.00036037209792993963, + 0.0006776950322091579, + 0.0009522222098894417, + 0.0009581497288309038, + 0.0006954859127290547, + 0.0015169641701504588, + 0.0005566072068177164, + 0.0007828501984477043, + 0.0006311953184194863, + 0.0009182419162243605, + 0.002139196265488863, + 0.0018929749494418502, + 0.002386255655437708, + 0.002762134186923504, + 0.0017524739960208535, + 0.0029476715717464685, + 0.0018716957420110703, + 0.0017092970665544271, + 0.001147367525845766, + 0.0011751304846256971, + 0.0017402528319507837, + 0.0009552264818921685, + 0.0008527506142854691, + 0.0010146177373826504, + 0.0008791551226750016, + 0.0008267927332781255, + 0.03478531911969185, + 0.02273775264620781, + 0.028919922187924385, + 0.008340684697031975, + 0.017072467133402824, + 0.19227275252342224, + 0.015026532113552094, + 0.1206020712852478, + 0.0005605339538305998, + 0.0010116007179021835, + 0.0007283916347660124, + 0.0004149639862589538, + 0.0006936697172932327, + 0.0009709773003123701, + 0.0005039267707616091, + 0.0008150055073201656, + 0.0004376988217700273, + 0.0008511096239089966, + 0.000734928878955543, + 0.0004879900661762804, + 0.0007827130611985922, + 0.0010945509420707822, + 0.0005317563773132861, + 0.0005567716434597969, + 0.0011822934029623866, + 0.0011466835858300328, + 0.0025315184611827135, + 0.0009263971587643027, + 0.0008337511098943651, + 0.0017730135004967451, + 0.001572952140122652, + 0.0011855072807520628, + 0.0020867001730948687, + 0.0022758636623620987, + 0.0030726129189133644, + 0.002346639521420002, + 0.0025350789073854685, + 0.0023949644528329372, + 0.004553001839667559, + 0.0030064620077610016, + 0.028974542394280434, + 0.005886816885322332, + 0.18459594249725342, + 0.02798916958272457, + 0.029918454587459564, + 0.15623970329761505, + 0.0189626794308424, + 0.025322461500763893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.026349911466240883 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026345498859882355 + ], + "opposite_logprob_tokens": [ + -3.6513454914093018 + ], + "logprob_reward": -0.026345498859882355, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026345498859882355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02071780525147915, + 0.013144339434802532, + 0.016434939578175545, + 0.010561277158558369, + 0.15160751342773438, + 0.0056218840181827545, + 0.004307800903916359, + 0.01649201102554798, + 0.034607142210006714, + 0.7265052795410156 + ], + "credit_map": [ + 0.0022468091920018196, + 0.0022335918620228767, + 0.0030307448469102383, + 0.003904904704540968, + 0.001509970286861062, + 0.0023744998034089804, + 0.0027615393046289682, + 0.0026557431556284428, + 0.001913460437208414, + 0.0007953240419737995, + 0.0014182556187734008, + 0.003639424918219447, + 0.0007541993982158601, + 0.0011417646892368793, + 0.001186621026135981, + 0.0022952877916395664, + 0.002496410859748721, + 0.001976531697437167, + 0.001639451482333243, + 0.002532881684601307, + 0.0014827377162873745, + 0.002204360207542777, + 0.0021920008584856987, + 0.0019105643732473254, + 0.0010992001043632627, + 0.0012395852245390415, + 0.0011954831425100565, + 0.0015943959588184953, + 0.0015150868566706777, + 0.0012384834699332714, + 0.0014038679655641317, + 0.0012751736212521791, + 0.011901157908141613, + 0.008198387920856476, + 0.009830348193645477, + 0.002938513644039631, + 0.00587885407730937, + 0.0659368485212326, + 0.005215954501181841, + 0.041707444936037064, + 0.0005468761664815247, + 0.0012230374850332737, + 0.0006787774618715048, + 0.0006123460480012, + 0.0007818014128133655, + 0.0005683209747076035, + 0.0004689386405516416, + 0.0007417849847115576, + 0.0005552651127800345, + 0.0006142149213701487, + 0.0005446373834274709, + 0.00031887704972177744, + 0.0005618461291305721, + 0.0004997808136977255, + 0.0006444955361075699, + 0.000568683841265738, + 0.0013877387391403317, + 0.0018501291051506996, + 0.0029662714805454016, + 0.0019544302485883236, + 0.0017085896106436849, + 0.0026859131176024675, + 0.0022149833384901285, + 0.001723953988403082, + 0.0033122829627245665, + 0.003216287586838007, + 0.005144031252712011, + 0.00322116375900805, + 0.0036811758764088154, + 0.00372892152518034, + 0.007353791035711765, + 0.004949485417455435, + 0.04419844225049019, + 0.008535794913768768, + 0.28136223554611206, + 0.042712919414043427, + 0.04538973048329353, + 0.2373598963022232, + 0.02871588058769703, + 0.03823035955429077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.026345498859882355 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a low-angle view of a forgotten wooden toy car on the floor, then slowly and steadily moves forward, focusing on an off-white plush cat climbing structure. From a circular cavity in the middle of the structure, a sleepy, cream-colored kitten cautiously pokes its furry head out. Its fur is soft and fluffy, streaked with light ginger, and shimmers softly in the warm afternoon sunlight slanting in from the window. Its disproportionately large, obsidian-like eyes, moist and bright, blink in a hazy, half-awake state, filled with innocent curiosity about this new world. It tentatively places a fleshy pink paw on the edge of the cavity, then clumsily yet calmly wriggles its entire body completely out of the darkness. Finally, it opens its mouth in a silent yawn, revealing its tiny pink tongue. The entire picture is immersed in a warm and delicate Ghibli animation style, with sparkling dust particles floating in the air, the background is a soft-focus wooden floor and a touch of green potted plants, and the main colors are composed of soft cream, warm wood color and golden sunlight, creating a tranquil, healing and fairy-tale atmosphere.", + "seed": 30296, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_6/index_3510_seed_30296_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3164029087871313, + "vlm_logprob": -0.3164029087871313, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.3164029087871313, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5045576095581055 + ], + "opposite_logprob_tokens": [ + -0.25455763936042786 + ], + "logprob_reward": -1.5045576095581055, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5045576095581055, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.21638533473014832, + 0.19135408103466034, + 0.13818241655826569, + 0.10224034637212753, + 0.08487645536661148, + 0.07212091982364655, + 0.07106036692857742, + 0.023451844230294228, + 0.05377032607793808, + 0.04655790328979492 + ], + "credit_map": [ + 0.02758498676121235, + 0.016723064705729485, + 0.028909580782055855, + 0.025113068521022797, + 0.01934824138879776, + 0.023162968456745148, + 0.020626628771424294, + 0.054916780441999435, + 0.016523664817214012, + 0.021173978224396706, + 0.01479415688663721, + 0.03429770469665527, + 0.03489864245057106, + 0.02086574211716652, + 0.03257672116160393, + 0.016223471611738205, + 0.022632436826825142, + 0.00745764933526516, + 0.010465343482792377, + 0.027837222442030907, + 0.011443814262747765, + 0.013472873717546463, + 0.018642544746398926, + 0.02623054012656212, + 0.011336391791701317, + 0.00932347308844328, + 0.011949948035180569, + 0.009349471889436245, + 0.007205410860478878, + 0.010645123198628426, + 0.03380840644240379, + 0.008622121065855026, + 0.02262730896472931, + 0.006251435726881027, + 0.009122931398451328, + 0.007280908990651369, + 0.003454945283010602, + 0.009082505479454994, + 0.004192770924419165, + 0.02286365255713463, + 0.007769581396132708, + 0.0071082282811403275, + 0.012721627950668335, + 0.006778458133339882, + 0.006956545170396566, + 0.013516437262296677, + 0.006364641711115837, + 0.010905398055911064, + 0.008795159868896008, + 0.00886042695492506, + 0.008343585766851902, + 0.0066868700087070465, + 0.006419082172214985, + 0.011862601153552532, + 0.007587152533233166, + 0.012505490332841873, + 0.0018187843961641192, + 0.0025658239610493183, + 0.002947588451206684, + 0.004187419544905424, + 0.00367868784815073, + 0.0030965562909841537, + 0.002418817486613989, + 0.0027381686959415674, + 0.005694495979696512, + 0.00432170182466507, + 0.0068172886967659, + 0.006270820274949074, + 0.0038736346177756786, + 0.008570297621190548, + 0.008078687824308872, + 0.010143399238586426, + 0.0042416853830218315, + 0.003980899229645729, + 0.00508458586409688, + 0.004103389102965593, + 0.0035886873956769705, + 0.00979859009385109, + 0.0035039084032177925, + 0.012256155721843243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -1.5045576095581055 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006771119311451912 + ], + "opposite_logprob_tokens": [ + -5.006771087646484 + ], + "logprob_reward": -0.006771119311451912, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006771119311451912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22491593658924103, + 0.19346484541893005, + 0.12449713796377182, + 0.0769028291106224, + 0.06829307973384857, + 0.05463792756199837, + 0.13771872222423553, + 0.02917320467531681, + 0.06784103810787201, + 0.022555245086550713 + ], + "credit_map": [ + 0.025867074728012085, + 0.01876584067940712, + 0.03202693164348602, + 0.025519581511616707, + 0.02407265268266201, + 0.0242994986474514, + 0.02115076780319214, + 0.05321360379457474, + 0.016489900648593903, + 0.02047087624669075, + 0.014732657931745052, + 0.033612243831157684, + 0.03845120593905449, + 0.02063712663948536, + 0.0320085845887661, + 0.017062244936823845, + 0.020464537665247917, + 0.006923159584403038, + 0.009576212614774704, + 0.0236252099275589, + 0.010181684046983719, + 0.01319644320756197, + 0.0162892434746027, + 0.024240642786026, + 0.006948981434106827, + 0.004095159471035004, + 0.010374276898801327, + 0.006260001566261053, + 0.005585983395576477, + 0.00938051100820303, + 0.026518559083342552, + 0.007739350199699402, + 0.016914069652557373, + 0.005274462513625622, + 0.006850164849311113, + 0.0069464812986552715, + 0.003080270253121853, + 0.00788111425936222, + 0.0040366207249462605, + 0.017309904098510742, + 0.005872214678674936, + 0.008355164900422096, + 0.008120435290038586, + 0.007384670898318291, + 0.005684790667146444, + 0.005939732771366835, + 0.004419669974595308, + 0.00886125210672617, + 0.02540956251323223, + 0.00990329496562481, + 0.014744390733540058, + 0.009721975773572922, + 0.006128909531980753, + 0.025213150307536125, + 0.009212191216647625, + 0.037385243922472, + 0.002799996640533209, + 0.003111861879006028, + 0.003284298814833164, + 0.005112082697451115, + 0.0047330125235021114, + 0.00391347985714674, + 0.002758866874501109, + 0.003459605388343334, + 0.007959423586726189, + 0.0056900447234511375, + 0.008520836010575294, + 0.00800026673823595, + 0.005221644416451454, + 0.010694862343370914, + 0.010343563742935658, + 0.011410394683480263, + 0.00228497083298862, + 0.002777895424515009, + 0.0024036099202930927, + 0.002380049554631114, + 0.0028190326411277056, + 0.003443891881033778, + 0.0022861408069729805, + 0.004159656353294849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.006771119311451912 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952132865786552 + ], + "opposite_logprob_tokens": [ + -5.13095235824585 + ], + "logprob_reward": -0.005952132865786552, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952132865786552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17782478034496307, + 0.12758953869342804, + 0.11764400452375412, + 0.11498144268989563, + 0.07500249147415161, + 0.09238846600055695, + 0.09073843061923981, + 0.06194485351443291, + 0.07175150513648987, + 0.07013452053070068 + ], + "credit_map": [ + 0.02513551525771618, + 0.01324651949107647, + 0.02327180653810501, + 0.02131268009543419, + 0.0156682301312685, + 0.018946640193462372, + 0.019022932276129723, + 0.041220441460609436, + 0.010795582085847855, + 0.014288650825619698, + 0.010521638207137585, + 0.020205451175570488, + 0.0244312584400177, + 0.013279014267027378, + 0.02085845358669758, + 0.013209488242864609, + 0.015716802328824997, + 0.007533875294029713, + 0.008980577811598778, + 0.02672949805855751, + 0.009431933984160423, + 0.01151188462972641, + 0.01804530806839466, + 0.019694119691848755, + 0.011093487963080406, + 0.008746587671339512, + 0.010125821456313133, + 0.01046818122267723, + 0.008977659046649933, + 0.011240826919674873, + 0.0431809201836586, + 0.011147952638566494, + 0.019975589588284492, + 0.005941974930465221, + 0.008351946249604225, + 0.006572849117219448, + 0.0031116544269025326, + 0.007490959018468857, + 0.003776404308155179, + 0.01978110522031784, + 0.008204481564462185, + 0.014627494849264622, + 0.010980647057294846, + 0.01236451230943203, + 0.011661931872367859, + 0.015458103269338608, + 0.00812110211700201, + 0.01097018737345934, + 0.013100696727633476, + 0.008903181180357933, + 0.00938077736645937, + 0.0078018223866820335, + 0.007422247901558876, + 0.016763955354690552, + 0.007376046851277351, + 0.01998969539999962, + 0.0061945682391524315, + 0.006178098730742931, + 0.005601189564913511, + 0.013674180954694748, + 0.011056054383516312, + 0.00679023377597332, + 0.005951872561126947, + 0.006498653907328844, + 0.007576769217848778, + 0.005955943837761879, + 0.009164071641862392, + 0.008703186176717281, + 0.0057352459989488125, + 0.011139515787363052, + 0.01013359148055315, + 0.01334317959845066, + 0.006629707291722298, + 0.005257755052298307, + 0.007265276741236448, + 0.005550520960241556, + 0.00458819093182683, + 0.01627163030207157, + 0.004724618513137102, + 0.01984681934118271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005952132865786552 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866249859333038 + ], + "opposite_logprob_tokens": [ + -3.0486624240875244 + ], + "logprob_reward": -0.04866249859333038, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866249859333038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057586342096328735, + 0.0454113744199276, + 0.04518566653132439, + 0.09199993312358856, + 0.11217574030160904, + 0.18426531553268433, + 0.22435779869556427, + 0.06635122746229172, + 0.1357622742652893, + 0.03690434619784355 + ], + "credit_map": [ + 0.009423717856407166, + 0.004263405688107014, + 0.0059420205652713776, + 0.007274925708770752, + 0.00449135759845376, + 0.006048404611647129, + 0.009333858266472816, + 0.010808651335537434, + 0.0033184310887008905, + 0.0033535375259816647, + 0.0023843906819820404, + 0.003599711460992694, + 0.01403143908828497, + 0.0052143665961921215, + 0.004091400653123856, + 0.009418097324669361, + 0.004902783781290054, + 0.0029654784593731165, + 0.003543826751410961, + 0.009500078856945038, + 0.0032179963309317827, + 0.009194851852953434, + 0.006499033421278, + 0.005361616145819426, + 0.011625471524894238, + 0.01667501963675022, + 0.009135110303759575, + 0.013590915128588676, + 0.008321464993059635, + 0.00863638799637556, + 0.01653156988322735, + 0.007484001107513905, + 0.028864331543445587, + 0.006707875058054924, + 0.012633579783141613, + 0.01103195920586586, + 0.004524938762187958, + 0.010653811506927013, + 0.007954037748277187, + 0.02980520948767662, + 0.021640228107571602, + 0.025276506319642067, + 0.027516407892107964, + 0.024562370032072067, + 0.019699154421687126, + 0.022058967500925064, + 0.030576374381780624, + 0.012935304082930088, + 0.0432015135884285, + 0.0127604641020298, + 0.025380147621035576, + 0.014403405599296093, + 0.008813180960714817, + 0.04476897791028023, + 0.010575625114142895, + 0.06445448100566864, + 0.006332778837531805, + 0.007120469119399786, + 0.006699635647237301, + 0.012339659966528416, + 0.01052580401301384, + 0.007570684887468815, + 0.0062138852663338184, + 0.009548311121761799, + 0.014522895216941833, + 0.011609780602157116, + 0.018275577574968338, + 0.017329474911093712, + 0.011923529207706451, + 0.018919043242931366, + 0.01904357224702835, + 0.024138404056429863, + 0.0034831638913601637, + 0.002573869889602065, + 0.003864570055156946, + 0.0026091430336236954, + 0.0019653593190014362, + 0.00934903509914875, + 0.0022749691270291805, + 0.010784236714243889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.04866249859333038 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607118360698223 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.01607118360698223, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607118360698223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19891878962516785, + 0.14317096769809723, + 0.13008955121040344, + 0.06902214884757996, + 0.025626569986343384, + 0.19259467720985413, + 0.08962873369455338, + 0.029698800295591354, + 0.06662949174642563, + 0.05462031811475754 + ], + "credit_map": [ + 0.02518729493021965, + 0.01245824247598648, + 0.025722123682498932, + 0.02159999869763851, + 0.013378174975514412, + 0.024373890832066536, + 0.01866050250828266, + 0.05753854289650917, + 0.011250920593738556, + 0.016653385013341904, + 0.011771400459110737, + 0.024102067574858665, + 0.022904641926288605, + 0.016203654929995537, + 0.02410324290394783, + 0.016181640326976776, + 0.019820841029286385, + 0.007188154850155115, + 0.009709243662655354, + 0.026651248335838318, + 0.01061225775629282, + 0.01369817741215229, + 0.018146855756640434, + 0.024262767285108566, + 0.00472235819324851, + 0.004504601005464792, + 0.007701895199716091, + 0.006383880041539669, + 0.004038998857140541, + 0.007211437914520502, + 0.024475734680891037, + 0.009983236901462078, + 0.0026483756955713034, + 0.0032517651561647654, + 0.0022318747360259295, + 0.00450354116037488, + 0.002054141601547599, + 0.0045318822376430035, + 0.003305673599243164, + 0.003099313471466303, + 0.019283859059214592, + 0.024262163788080215, + 0.0248749740421772, + 0.02135072834789753, + 0.02254551090300083, + 0.04140350967645645, + 0.015138048678636551, + 0.02373587153851986, + 0.01510210894048214, + 0.010303022339940071, + 0.008101441897451878, + 0.008422648534178734, + 0.0054478817619383335, + 0.014938929118216038, + 0.007324874866753817, + 0.019987815991044044, + 0.002964288229122758, + 0.0031418136786669493, + 0.004411518573760986, + 0.0033736154437065125, + 0.003049100050702691, + 0.0037414939142763615, + 0.0037913834676146507, + 0.005225582513958216, + 0.007068697828799486, + 0.004987341817468405, + 0.008903300389647484, + 0.0069220587611198425, + 0.004449788480997086, + 0.010571926832199097, + 0.008270434103906155, + 0.01545593049377203, + 0.004916980396956205, + 0.005122504197061062, + 0.005550617352128029, + 0.005169453099370003, + 0.005274658557027578, + 0.010931977070868015, + 0.004362784791737795, + 0.013291336596012115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.01607118360698223 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a low-angle view of a forgotten wooden toy car on the floor, then slowly and steadily moves forward, focusing on an off-white plush cat climbing structure. From a circular cavity in the middle of the structure, a sleepy, cream-colored kitten cautiously pokes its furry head out. Its fur is soft and fluffy, streaked with light ginger, and shimmers softly in the warm afternoon sunlight slanting in from the window. Its disproportionately large, obsidian-like eyes, moist and bright, blink in a hazy, half-awake state, filled with innocent curiosity about this new world. It tentatively places a fleshy pink paw on the edge of the cavity, then clumsily yet calmly wriggles its entire body completely out of the darkness. Finally, it opens its mouth in a silent yawn, revealing its tiny pink tongue. The entire picture is immersed in a warm and delicate Ghibli animation style, with sparkling dust particles floating in the air, the background is a soft-focus wooden floor and a touch of green potted plants, and the main colors are composed of soft cream, warm wood color and golden sunlight, creating a tranquil, healing and fairy-tale atmosphere.", + "seed": 30297, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_6/index_3510_seed_30297_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12322805095463991, + "vlm_logprob": -0.12322805095463991, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.12322805095463991, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5780239701271057 + ], + "opposite_logprob_tokens": [ + -0.8280239701271057 + ], + "logprob_reward": -0.5780239701271057, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5780239701271057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1943483054637909, + 0.12144986540079117, + 0.06657380610704422, + 0.06828901171684265, + 0.07038535922765732, + 0.10244373232126236, + 0.2215673178434372, + 0.05425998941063881, + 0.07301637530326843, + 0.027666231617331505 + ], + "credit_map": [ + 0.014079217799007893, + 0.009144337847828865, + 0.017830880358815193, + 0.012053204700350761, + 0.012845546007156372, + 0.05906970798969269, + 0.009178943932056427, + 0.06014646589756012, + 0.01060115173459053, + 0.00978077296167612, + 0.020709536969661713, + 0.026920516043901443, + 0.009813348762691021, + 0.011782540008425713, + 0.017597662284970284, + 0.01424433384090662, + 0.005825670436024666, + 0.00405834149569273, + 0.009179978631436825, + 0.005150364711880684, + 0.006862585432827473, + 0.014389121904969215, + 0.011322707869112492, + 0.009785044007003307, + 0.007734533399343491, + 0.008055347017943859, + 0.007728309370577335, + 0.007672215346246958, + 0.007943973876535892, + 0.009211095049977303, + 0.008358766324818134, + 0.011584775522351265, + 0.005381349939852953, + 0.003881066106259823, + 0.006852343212813139, + 0.006377484183758497, + 0.0041001662611961365, + 0.014698878861963749, + 0.00888121034950018, + 0.020212866365909576, + 0.010777225717902184, + 0.011905292049050331, + 0.01205358374863863, + 0.011512923054397106, + 0.01034946832805872, + 0.017237568274140358, + 0.013532212935388088, + 0.015075452625751495, + 0.022598443552851677, + 0.013746319338679314, + 0.07751774787902832, + 0.013550328090786934, + 0.01622779481112957, + 0.008735784329473972, + 0.06003227457404137, + 0.009158628061413765, + 0.005631235893815756, + 0.005933873355388641, + 0.008735775016248226, + 0.004732160363346338, + 0.012349002063274384, + 0.004751270171254873, + 0.008933479897677898, + 0.003193190321326256, + 0.009682364761829376, + 0.006321909371763468, + 0.01710537075996399, + 0.005480210762470961, + 0.011155513115227222, + 0.006884041707962751, + 0.012027561664581299, + 0.004359400365501642, + 0.004505114164203405, + 0.0026713949628174305, + 0.004981633275747299, + 0.0030272076837718487, + 0.0044900765642523766, + 0.00315282354131341, + 0.003438131883740425, + 0.0013998495414853096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.5780239701271057 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032087289728224277 + ], + "opposite_logprob_tokens": [ + -5.753208637237549 + ], + "logprob_reward": -0.0032087289728224277, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032087289728224277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08513463288545609, + 0.08999998867511749, + 0.05018002167344093, + 0.048831548541784286, + 0.09293774515390396, + 0.13803839683532715, + 0.30506637692451477, + 0.08883535116910934, + 0.08025789260864258, + 0.020718056708574295 + ], + "credit_map": [ + 0.005380863789469004, + 0.004088906571269035, + 0.006169207394123077, + 0.004548114258795977, + 0.006045161746442318, + 0.026996759697794914, + 0.005044121760874987, + 0.026861490681767464, + 0.007861804217100143, + 0.00828569382429123, + 0.01032837200909853, + 0.022021690383553505, + 0.007410891354084015, + 0.010472286492586136, + 0.012557504698634148, + 0.011061732657253742, + 0.004705414175987244, + 0.005963286384940147, + 0.006217809394001961, + 0.006297693122178316, + 0.0058420985005795956, + 0.007143024820834398, + 0.0078031946904957294, + 0.006207495927810669, + 0.005604350473731756, + 0.005134038161486387, + 0.008085787296295166, + 0.0053413729183375835, + 0.004818668123334646, + 0.007734025828540325, + 0.006732907611876726, + 0.00538039393723011, + 0.009142330847680569, + 0.00699883047491312, + 0.007303380407392979, + 0.008799566887319088, + 0.005257680546492338, + 0.028311289846897125, + 0.009406805969774723, + 0.01771785505115986, + 0.035218629986047745, + 0.006586240604519844, + 0.03258305415511131, + 0.006633914541453123, + 0.007873834110796452, + 0.026587996631860733, + 0.012577950954437256, + 0.009976773522794247, + 0.0543142594397068, + 0.017029544338583946, + 0.11462368816137314, + 0.020138727501034737, + 0.01621338166296482, + 0.017745835706591606, + 0.05210448428988457, + 0.012896423228085041, + 0.00879636313766241, + 0.009648224338889122, + 0.014554088935256004, + 0.0071608698926866055, + 0.021043043583631516, + 0.007783024571835995, + 0.01464766077697277, + 0.005202069878578186, + 0.00965278409421444, + 0.00662342319265008, + 0.018452344462275505, + 0.00664419075474143, + 0.010298890061676502, + 0.007624380756169558, + 0.016405310481786728, + 0.004556557163596153, + 0.003294390859082341, + 0.0017611737130209804, + 0.0027379821985960007, + 0.002225952921435237, + 0.0031768805347383022, + 0.002740449272096157, + 0.0034245180431753397, + 0.0013567074202001095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.0032087289728224277 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046369642950594425 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.0046369642950594425, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046369642950594425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05725026875734329, + 0.040933556854724884, + 0.041949737817049026, + 0.07118389010429382, + 0.1907964050769806, + 0.05801260098814964, + 0.29698270559310913, + 0.07347513735294342, + 0.0498679094016552, + 0.11954773962497711 + ], + "credit_map": [ + 0.004643662832677364, + 0.005384971387684345, + 0.0061936089769005775, + 0.004502025432884693, + 0.006151560693979263, + 0.01203927118331194, + 0.004786877892911434, + 0.013548288494348526, + 0.005254438612610102, + 0.005108779761940241, + 0.006316486280411482, + 0.006258158944547176, + 0.003387385280802846, + 0.005835646763443947, + 0.004234482999891043, + 0.004538176115602255, + 0.0038569762837141752, + 0.004541680682450533, + 0.00378528144210577, + 0.0042100939899683, + 0.007061799988150597, + 0.007981253787875175, + 0.005877441260963678, + 0.004635210614651442, + 0.011414929293096066, + 0.0082522789016366, + 0.008908197283744812, + 0.00930033903568983, + 0.006144513376057148, + 0.011901638470590115, + 0.007120138965547085, + 0.008141848258674145, + 0.021120378747582436, + 0.0065865362994372845, + 0.007360705640166998, + 0.008685573004186153, + 0.006229311693459749, + 0.07564853876829147, + 0.009602155536413193, + 0.05556320771574974, + 0.0072069973684847355, + 0.005292156245559454, + 0.010797171853482723, + 0.006090168841183186, + 0.006010320968925953, + 0.008631459437310696, + 0.007373315282166004, + 0.006611016113311052, + 0.05147700756788254, + 0.016487406566739082, + 0.12450549751520157, + 0.01732618361711502, + 0.01414601318538189, + 0.011863759718835354, + 0.047391973435878754, + 0.013784864917397499, + 0.008652647957205772, + 0.007550118491053581, + 0.01064356043934822, + 0.006345769856125116, + 0.01826183870434761, + 0.005962487310171127, + 0.011114942841231823, + 0.004943773150444031, + 0.006972366012632847, + 0.004062873777002096, + 0.0065612392500042915, + 0.004827172961086035, + 0.007841474376618862, + 0.007276118732988834, + 0.009359538555145264, + 0.0029671250376850367, + 0.01885673776268959, + 0.012404573149979115, + 0.026605548337101936, + 0.005949761252850294, + 0.02538284845650196, + 0.010936105623841286, + 0.012827136553823948, + 0.006585028022527695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.0046369642950594425 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197062700986862 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.014197062700986862, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197062700986862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039407990872859955, + 0.028189169242978096, + 0.03521500527858734, + 0.028231628239154816, + 0.26151058077812195, + 0.06089600920677185, + 0.35877296328544617, + 0.08649707585573196, + 0.052967824041843414, + 0.04831170663237572 + ], + "credit_map": [ + 0.001970853190869093, + 0.002034100703895092, + 0.002574086654931307, + 0.002367613837122917, + 0.0028571453876793385, + 0.01279095746576786, + 0.0023215869441628456, + 0.012491646222770214, + 0.002243706723675132, + 0.0023037465289235115, + 0.003488838905468583, + 0.007315695285797119, + 0.002467677928507328, + 0.004311340395361185, + 0.0035510179586708546, + 0.002507145283743739, + 0.002818153938278556, + 0.0027441391721367836, + 0.0028489644173532724, + 0.004244337789714336, + 0.004161880351603031, + 0.008184896782040596, + 0.005500781815499067, + 0.004711850080639124, + 0.005334747023880482, + 0.00352082634344697, + 0.0027040515560656786, + 0.003736612619832158, + 0.0025740666314959526, + 0.003917814698070288, + 0.003209430258721113, + 0.003234078176319599, + 0.027901822701096535, + 0.010302333161234856, + 0.011698703281581402, + 0.013126078993082047, + 0.007460777182132006, + 0.09981939941644669, + 0.00960494764149189, + 0.08159653097391129, + 0.012058685533702374, + 0.004021143540740013, + 0.0033612074330449104, + 0.0033330516889691353, + 0.003472101641818881, + 0.022623196244239807, + 0.006069852039217949, + 0.00595677038654685, + 0.06664441525936127, + 0.018151575699448586, + 0.14387445151805878, + 0.022277910262346268, + 0.01595422811806202, + 0.018140679225325584, + 0.057777684181928635, + 0.01595202088356018, + 0.008882125839591026, + 0.009902323596179485, + 0.016394371166825294, + 0.009250780567526817, + 0.016560252755880356, + 0.007067242171615362, + 0.013506793417036533, + 0.004933193791657686, + 0.006646980065852404, + 0.004282587207853794, + 0.012793020345270634, + 0.004574623424559832, + 0.005222992971539497, + 0.005120251327753067, + 0.010734195820987225, + 0.0035931726451963186, + 0.007238104473799467, + 0.006158158648759127, + 0.00726936012506485, + 0.00525346165522933, + 0.007806305307894945, + 0.0054663741029798985, + 0.006307276431471109, + 0.0028126658871769905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.014197062700986862 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera initially positioned at a low angle showing a forgotten wooden toy car on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move forward steadily to focus on an off-white plush cat climbing structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cream-colored kitten with soft, fluffy fur streaked with light ginger visible poking its head out of a circular cavity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the kitten wriggle its body out of the cavity and open its mouth in a silent yawn revealing a tiny pink tongue?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a warm Ghibli animation style with sparkling dust particles, soft-focus background, and golden sunlight?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073528677225113 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073528677225113, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073528677225113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07125749439001083, + 0.05495226010680199, + 0.059265799820423126, + 0.08949098736047745, + 0.3469160497188568, + 0.06328348070383072, + 0.06898702681064606, + 0.10760790854692459, + 0.09348590672016144, + 0.044753190129995346 + ], + "credit_map": [ + 0.008053954690694809, + 0.005897799506783485, + 0.00800754688680172, + 0.005437525920569897, + 0.004719515331089497, + 0.015608900226652622, + 0.004530881065875292, + 0.019001375883817673, + 0.004745075479149818, + 0.005746654234826565, + 0.005439389031380415, + 0.014518970623612404, + 0.004744898993521929, + 0.008347025141119957, + 0.005932353436946869, + 0.005477900616824627, + 0.004568684846162796, + 0.0067602465860545635, + 0.007755146827548742, + 0.007463744841516018, + 0.006623073946684599, + 0.008931045420467854, + 0.010329917073249817, + 0.006833935156464577, + 0.009248469956219196, + 0.011826788075268269, + 0.010911217890679836, + 0.009759685024619102, + 0.011103118769824505, + 0.015301317907869816, + 0.011148189194500446, + 0.010192196816205978, + 0.03272119536995888, + 0.014761123806238174, + 0.01851322315633297, + 0.020434606820344925, + 0.012197828851640224, + 0.11948294937610626, + 0.01981278508901596, + 0.1089923158288002, + 0.0058211167342960835, + 0.003390647703781724, + 0.016748502850532532, + 0.005641557741910219, + 0.0059209768660366535, + 0.014942608773708344, + 0.006792477332055569, + 0.004025592934340239, + 0.010061733424663544, + 0.008474776521325111, + 0.014789548702538013, + 0.006078548263758421, + 0.0068540251813828945, + 0.007377755828201771, + 0.010866384021937847, + 0.004484251607209444, + 0.01397103350609541, + 0.011510627344250679, + 0.009825644083321095, + 0.011298595927655697, + 0.0204305537045002, + 0.01363326609134674, + 0.01927793212234974, + 0.007660259492695332, + 0.011705883778631687, + 0.009149326011538506, + 0.013571128249168396, + 0.0084923654794693, + 0.015307089313864708, + 0.01153490785509348, + 0.01766216568648815, + 0.006063032429665327, + 0.007259573321789503, + 0.005241925362497568, + 0.004819421097636223, + 0.003002450568601489, + 0.01063122320920229, + 0.004954332020133734, + 0.006352834403514862, + 0.002491430379450321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.016073528677225113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_7/index_tensor([5502, 5502])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_7/index_tensor([5502, 5502])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9bb340831baf2da135a7383d46535248c6b21f15 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_7/index_tensor([5502, 5502])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly low-angle close-up captures a thin, transparent piece of plastic film, held perfectly still by an invisible static charge against the vast, smooth surface of a large office window. An imperfect rectangle with a slightly curled corner, the film catches the intense sunlight, revealing a subtle, iridescent glow across its surface and highlighting a few tiny dust particles clinging to it. Bright, golden afternoon sunlight streams directly through the clean glass, bringing the plastic film into focus with a bright, warm lens flare. Through the window, the deep blue of a clear afternoon sky and the blurred architectural forms of distant city buildings are visible, all slightly distorted by the glass. Blurred, ghostly reflections of the empty office corridor—the edges of empty desks and the geometric pattern of the ceiling—superimposed on the view outside enhance the quiet loneliness. The film appears to float weightlessly, its edge casting a thin, sharp shadow on the glass, testament to the unseen force holding it in place. The entire scene is bathed in warm, high-contrast light, creating a tranquil and contemplative atmosphere that transforms mundane details into moments of serene, minimalist beauty.", + "seed": 3130, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_7/index_5502_seed_3130_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.22358395755290986, + "vlm_logprob": -0.22358395755290986, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.22358395755290986, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01830412447452545 + ], + "opposite_logprob_tokens": [ + -4.143304347991943 + ], + "logprob_reward": -0.01830412447452545, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01830412447452545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15316082537174225, + 0.0445435456931591, + 0.10331020504236221, + 0.2342505007982254, + 0.2034696787595749, + 0.053216736763715744, + 0.047133203595876694, + 0.08276532590389252, + 0.036428410559892654, + 0.04172160103917122 + ], + "credit_map": [ + 0.02227696403861046, + 0.041657522320747375, + 0.006827947683632374, + 0.00745655270293355, + 0.017894485965371132, + 0.02478916570544243, + 0.016720624640583992, + 0.0155375637114048, + 0.0037354116793721914, + 0.005243504419922829, + 0.004314752761274576, + 0.007141506299376488, + 0.0055239093489944935, + 0.004740158561617136, + 0.007848960347473621, + 0.005995340179651976, + 0.007366212084889412, + 0.012912612408399582, + 0.007596820127218962, + 0.009942447766661644, + 0.018540242686867714, + 0.01494609471410513, + 0.020320719107985497, + 0.01168504636734724, + 0.02725585550069809, + 0.02776632457971573, + 0.029378879815340042, + 0.03745635971426964, + 0.03684624284505844, + 0.028049707412719727, + 0.026959126815199852, + 0.020538005977869034, + 0.011026070453226566, + 0.027505282312631607, + 0.011413507163524628, + 0.019866060465574265, + 0.05970492586493492, + 0.025373701006174088, + 0.033014047890901566, + 0.015566103160381317, + 0.0035794959403574467, + 0.006715031806379557, + 0.0038686010520905256, + 0.00565378600731492, + 0.01298475544899702, + 0.006623453460633755, + 0.007092208135873079, + 0.0066994065418839455, + 0.006565365940332413, + 0.007771212141960859, + 0.006990660913288593, + 0.0045366850681602955, + 0.006341050378978252, + 0.006446120794862509, + 0.00398189527913928, + 0.004500214010477066, + 0.009121970273554325, + 0.011253759264945984, + 0.008085181005299091, + 0.007357809226959944, + 0.012724827975034714, + 0.017870990559458733, + 0.00728650251403451, + 0.009064284153282642, + 0.0031607227865606546, + 0.0036986672785133123, + 0.007945384830236435, + 0.0034348585177212954, + 0.0030144283082336187, + 0.009629558771848679, + 0.002519455039873719, + 0.0030253359582275152, + 0.005774927791208029, + 0.004540513269603252, + 0.006237397901713848, + 0.005322650540620089, + 0.004272853955626488, + 0.007281162776052952, + 0.003802548162639141, + 0.004489545710384846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.01830412447452545 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743155241012573 + ], + "opposite_logprob_tokens": [ + -0.4743155241012573 + ], + "logprob_reward": -0.9743155241012573, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743155241012573, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.146549791097641, + 0.036002811044454575, + 0.037482794374227524, + 0.09957197308540344, + 0.15643252432346344, + 0.08163034915924072, + 0.08216254413127899, + 0.07745632529258728, + 0.1704515665769577, + 0.11225928366184235 + ], + "credit_map": [ + 0.028652042150497437, + 0.051341667771339417, + 0.004796911031007767, + 0.009154604747891426, + 0.02160823345184326, + 0.014080023393034935, + 0.008745521306991577, + 0.008170798420906067, + 0.0036724500823765993, + 0.004051187541335821, + 0.004215073771774769, + 0.0059815519489347935, + 0.004033156670629978, + 0.0039804959669709206, + 0.004652639385312796, + 0.005416260566562414, + 0.003182021901011467, + 0.004437228664755821, + 0.003578677074983716, + 0.004100640770047903, + 0.0062979431822896, + 0.004176887217909098, + 0.008795053698122501, + 0.0029143430292606354, + 0.010748532600700855, + 0.010916344821453094, + 0.011572441086173058, + 0.014879310503602028, + 0.015240564942359924, + 0.01159935537725687, + 0.013103769160807133, + 0.011511663906276226, + 0.01486531924456358, + 0.017393581569194794, + 0.012319643050432205, + 0.02452751062810421, + 0.027486955747008324, + 0.0197166595607996, + 0.02528366819024086, + 0.014839205890893936, + 0.007073399145156145, + 0.00938394758850336, + 0.007154111284762621, + 0.009753714315593243, + 0.018338529393076897, + 0.012047533877193928, + 0.010070542804896832, + 0.007808574475347996, + 0.008463225327432156, + 0.014215567149221897, + 0.010243270546197891, + 0.006849952973425388, + 0.01954161375761032, + 0.012493614107370377, + 0.005032092798501253, + 0.005323214456439018, + 0.0086250901222229, + 0.009371061809360981, + 0.007500411942601204, + 0.0074723828583955765, + 0.009408446960151196, + 0.016562631353735924, + 0.006019058637320995, + 0.012497244402766228, + 0.01716783083975315, + 0.016647210344672203, + 0.0389225035905838, + 0.013164735399186611, + 0.014055246487259865, + 0.040563080459833145, + 0.010217267088592052, + 0.019713707268238068, + 0.00861139502376318, + 0.005305496975779533, + 0.03474981710314751, + 0.007382798474282026, + 0.009543953463435173, + 0.03293464332818985, + 0.006145688239485025, + 0.007585497107356787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.9743155241012573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10037458688020706 + ], + "opposite_logprob_tokens": [ + -2.350374698638916 + ], + "logprob_reward": -0.10037458688020706, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10037458688020706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11666382849216461, + 0.035077597945928574, + 0.13314148783683777, + 0.031215306371450424, + 0.14520975947380066, + 0.08703024685382843, + 0.14881183207035065, + 0.05558064952492714, + 0.03979739174246788, + 0.20747195184230804 + ], + "credit_map": [ + 0.014337068423628807, + 0.028311176225543022, + 0.006102990359067917, + 0.01150453183799982, + 0.013626775704324245, + 0.015471592545509338, + 0.01110071875154972, + 0.016208969056606293, + 0.0036087282933294773, + 0.005695453844964504, + 0.002876273123547435, + 0.00425321701914072, + 0.004234473221004009, + 0.0038453645538538694, + 0.0044773416593670845, + 0.006086744833737612, + 0.014018279500305653, + 0.014704976230859756, + 0.015648499131202698, + 0.015465858392417431, + 0.02199583873152733, + 0.016679314896464348, + 0.022341914474964142, + 0.012286797165870667, + 0.003817776683717966, + 0.003329173196107149, + 0.003986978903412819, + 0.004123510327190161, + 0.003732123877853155, + 0.004468909464776516, + 0.004819659516215324, + 0.0029371713753789663, + 0.015530095435678959, + 0.02233494631946087, + 0.010091602802276611, + 0.014070873148739338, + 0.030506925657391548, + 0.014749156311154366, + 0.023257959634065628, + 0.014668176881968975, + 0.006999571342021227, + 0.009367930702865124, + 0.006501700263470411, + 0.009614264592528343, + 0.01847650483250618, + 0.013636923395097256, + 0.013341986574232578, + 0.009091357700526714, + 0.016631020233035088, + 0.026261208578944206, + 0.014720330946147442, + 0.01691998541355133, + 0.027388306334614754, + 0.02336576022207737, + 0.010498378425836563, + 0.01302682887762785, + 0.00698597077280283, + 0.008268211036920547, + 0.00461276201531291, + 0.006275115069001913, + 0.006929262541234493, + 0.010721895843744278, + 0.00558787165209651, + 0.0061995526775717735, + 0.006367489695549011, + 0.004057764075696468, + 0.005638887174427509, + 0.004070638213306665, + 0.0037942659109830856, + 0.006602663081139326, + 0.003012373810634017, + 0.006253306288272142, + 0.02963080070912838, + 0.020160090178251266, + 0.03668420389294624, + 0.022493697702884674, + 0.020848756656050682, + 0.029905099421739578, + 0.02625073306262493, + 0.021498577669262886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.10037458688020706 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067422278225421906 + ], + "opposite_logprob_tokens": [ + -5.006742000579834 + ], + "logprob_reward": -0.0067422278225421906, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067422278225421906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6118022203445435, + 0.023721303790807724, + 0.04222344234585762, + 0.04891204833984375, + 0.0773872658610344, + 0.03750021383166313, + 0.04215015843510628, + 0.04071713238954544, + 0.031069546937942505, + 0.044516611844301224 + ], + "credit_map": [ + 0.11753363907337189, + 0.24301327764987946, + 0.01692124456167221, + 0.0399450957775116, + 0.05567321926355362, + 0.08360593020915985, + 0.026147278025746346, + 0.02896258607506752, + 0.002396496944129467, + 0.002567509887740016, + 0.002123278332874179, + 0.0043164948001503944, + 0.00368289346806705, + 0.0026642358861863613, + 0.0034196628257632256, + 0.00255073350854218, + 0.003112997394055128, + 0.004871452692896128, + 0.004439900163561106, + 0.006807183846831322, + 0.006466824561357498, + 0.0049902102909982204, + 0.007320322096347809, + 0.004214548040181398, + 0.005794951692223549, + 0.006100526545196772, + 0.005989255383610725, + 0.006974207237362862, + 0.008520599454641342, + 0.005347002297639847, + 0.0060132346116006374, + 0.004172275308519602, + 0.0066825877875089645, + 0.008468457497656345, + 0.005490703042596579, + 0.01096686813980341, + 0.018328875303268433, + 0.007738606538623571, + 0.01310607697814703, + 0.006605093367397785, + 0.0024480316787958145, + 0.005602654069662094, + 0.002952604554593563, + 0.0034923863131552935, + 0.011179632507264614, + 0.004601794295012951, + 0.004426827188581228, + 0.002796283457428217, + 0.005352448206394911, + 0.007529091089963913, + 0.005983929615467787, + 0.004898990970104933, + 0.004791073966771364, + 0.006841693539172411, + 0.0029900444205850363, + 0.0037628880236297846, + 0.0050033265724778175, + 0.0045404392294585705, + 0.0051905191503465176, + 0.003011515596881509, + 0.0048725539818406105, + 0.009904607199132442, + 0.002914352109655738, + 0.005279819015413523, + 0.0031602641101926565, + 0.003256726311519742, + 0.006269197445362806, + 0.0029277075082063675, + 0.002640043618157506, + 0.006763210985809565, + 0.0021041184663772583, + 0.003948278725147247, + 0.0030442040879279375, + 0.003511846764013171, + 0.013198635540902615, + 0.003326818812638521, + 0.0027040389832109213, + 0.013122264295816422, + 0.0021443755831569433, + 0.0034644263796508312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0067422278225421906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183324486017227 + ], + "opposite_logprob_tokens": [ + -4.01818323135376 + ], + "logprob_reward": -0.018183324486017227, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183324486017227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3781701624393463, + 0.029518647119402885, + 0.07573734223842621, + 0.05182298645377159, + 0.14993694424629211, + 0.04515102133154869, + 0.10315126180648804, + 0.04972098022699356, + 0.0653715580701828, + 0.05141912028193474 + ], + "credit_map": [ + 0.06441743671894073, + 0.1434866487979889, + 0.010445937514305115, + 0.018016081303358078, + 0.036177415400743484, + 0.05922657251358032, + 0.02021360583603382, + 0.026186462491750717, + 0.0029089655727148056, + 0.003038856666535139, + 0.003005112987011671, + 0.0029732459224760532, + 0.00454372214153409, + 0.0036856243386864662, + 0.005854198709130287, + 0.0035089198499917984, + 0.00524511793628335, + 0.008587631396949291, + 0.0077051278203725815, + 0.009880833327770233, + 0.011324696242809296, + 0.009538866579532623, + 0.014280523173511028, + 0.009174542501568794, + 0.004858752246946096, + 0.005922206211835146, + 0.00589335011318326, + 0.007084991317242384, + 0.006680075544863939, + 0.006348572671413422, + 0.009879548102617264, + 0.005155488848686218, + 0.011966231279075146, + 0.017370885238051414, + 0.013075144030153751, + 0.019424909725785255, + 0.038946475833654404, + 0.017625896260142326, + 0.021109841763973236, + 0.010417561046779156, + 0.0031703440472483635, + 0.006826893426477909, + 0.0028983585070818663, + 0.004328375682234764, + 0.011191828176379204, + 0.006256830412894487, + 0.007461998146027327, + 0.003016393631696701, + 0.009726542979478836, + 0.01711447536945343, + 0.0056844535283744335, + 0.010329069569706917, + 0.026114467531442642, + 0.017914390191435814, + 0.009113109670579433, + 0.00715474970638752, + 0.008155508898198605, + 0.006540877278894186, + 0.004270981531590223, + 0.006195388734340668, + 0.005176091566681862, + 0.008034441620111465, + 0.00444659823551774, + 0.00690109608694911, + 0.006954457610845566, + 0.005502544343471527, + 0.015126342885196209, + 0.005909601226449013, + 0.005138535983860493, + 0.016019903123378754, + 0.004092419985681772, + 0.006627758499234915, + 0.0040425509214401245, + 0.003688874887302518, + 0.010303018614649773, + 0.004943488631397486, + 0.00568848941475153, + 0.01405003760010004, + 0.0036328169517219067, + 0.005069838836789131 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.018183324486017227 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle close-up captures a thin, transparent piece of plastic film, held perfectly still by an invisible static charge against the vast, smooth surface of a large office window. An imperfect rectangle with a slightly curled corner, the film catches the intense sunlight, revealing a subtle, iridescent glow across its surface and highlighting a few tiny dust particles clinging to it. Bright, golden afternoon sunlight streams directly through the clean glass, bringing the plastic film into focus with a bright, warm lens flare. Through the window, the deep blue of a clear afternoon sky and the blurred architectural forms of distant city buildings are visible, all slightly distorted by the glass. Blurred, ghostly reflections of the empty office corridor—the edges of empty desks and the geometric pattern of the ceiling—superimposed on the view outside enhance the quiet loneliness. The film appears to float weightlessly, its edge casting a thin, sharp shadow on the glass, testament to the unseen force holding it in place. The entire scene is bathed in warm, high-contrast light, creating a tranquil and contemplative atmosphere that transforms mundane details into moments of serene, minimalist beauty.", + "seed": 3131, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_7/index_5502_seed_3131_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05418462222442031, + "vlm_logprob": -0.05418462222442031, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.05418462222442031, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00514943478628993 + ], + "opposite_logprob_tokens": [ + -6.2551493644714355 + ], + "logprob_reward": -0.00514943478628993, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00514943478628993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3700099289417267, + 0.12967798113822937, + 0.05590500682592392, + 0.04674220830202103, + 0.05914418026804924, + 0.06571497023105621, + 0.05262262746691704, + 0.05675171688199043, + 0.045323800295591354, + 0.11810764670372009 + ], + "credit_map": [ + 0.03827729821205139, + 0.05328790470957756, + 0.05598478391766548, + 0.03636666387319565, + 0.05707348510622978, + 0.04145178943872452, + 0.04843306913971901, + 0.039134904742240906, + 0.01485368050634861, + 0.017668766900897026, + 0.0194668211042881, + 0.012234734371304512, + 0.01822466403245926, + 0.014831730164587498, + 0.017245152965188026, + 0.015152428299188614, + 0.0063438923098146915, + 0.006721802055835724, + 0.005800201091915369, + 0.0067199235782027245, + 0.010948041453957558, + 0.006835209205746651, + 0.008456419222056866, + 0.004079517442733049, + 0.005094091407954693, + 0.0064880335703492165, + 0.006070820149034262, + 0.004692845046520233, + 0.00908611435443163, + 0.006127229426056147, + 0.0054364013485610485, + 0.0037466755602508783, + 0.008659308776259422, + 0.005089562851935625, + 0.012071614153683186, + 0.005341857206076384, + 0.007542798295617104, + 0.006114506162703037, + 0.008242717944085598, + 0.006081809289753437, + 0.006327629089355469, + 0.005697838496416807, + 0.01132266502827406, + 0.007455980405211449, + 0.012749927118420601, + 0.007672849576920271, + 0.006129124667495489, + 0.008358960971236229, + 0.007055468857288361, + 0.004936960060149431, + 0.008657158352434635, + 0.005688582081347704, + 0.009117933921515942, + 0.0054737720638513565, + 0.006015470717102289, + 0.0056772781535983086, + 0.006068705581128597, + 0.007676385343074799, + 0.010013961233198643, + 0.006417693104594946, + 0.008697939105331898, + 0.006215214729309082, + 0.006356424652040005, + 0.00530539033934474, + 0.005598675925284624, + 0.005365179385989904, + 0.0063711535185575485, + 0.005824259016662836, + 0.005794231779873371, + 0.004925575107336044, + 0.006467034574598074, + 0.00497769471257925, + 0.020165668800473213, + 0.015097650699317455, + 0.0201385747641325, + 0.012387129478156567, + 0.012219672091305256, + 0.013869100250303745, + 0.011098059825599194, + 0.013131787069141865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.00514943478628993 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161204040050507 + ], + "opposite_logprob_tokens": [ + -1.7016119956970215 + ], + "logprob_reward": -0.20161204040050507, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161204040050507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32017821073532104, + 0.1741565465927124, + 0.06537314504384995, + 0.05035397782921791, + 0.07643876224756241, + 0.05678509920835495, + 0.1283755600452423, + 0.032853830605745316, + 0.05193386599421501, + 0.043550971895456314 + ], + "credit_map": [ + 0.03047564998269081, + 0.03890916332602501, + 0.05112988501787186, + 0.04371625930070877, + 0.06102755665779114, + 0.04159824922680855, + 0.03145420178771019, + 0.021867241710424423, + 0.0169716477394104, + 0.013840449042618275, + 0.03489603474736214, + 0.020089391618967056, + 0.021696878597140312, + 0.01718379184603691, + 0.03175611421465874, + 0.0177222341299057, + 0.009924164973199368, + 0.008205157704651356, + 0.008662506937980652, + 0.007186373230069876, + 0.008843909949064255, + 0.007803266402333975, + 0.006208492908626795, + 0.008539274334907532, + 0.005176305305212736, + 0.006007956340909004, + 0.0059250774793326855, + 0.008579584769904613, + 0.008177259936928749, + 0.005485241301357746, + 0.006238766014575958, + 0.0047637843526899815, + 0.008293516002595425, + 0.007108623161911964, + 0.011436018161475658, + 0.010605898685753345, + 0.00909214373677969, + 0.007853192277252674, + 0.015185032039880753, + 0.006864339113235474, + 0.006172503344714642, + 0.0076910508796572685, + 0.00827818177640438, + 0.005201298743486404, + 0.0072779580950737, + 0.006147736683487892, + 0.007588075939565897, + 0.008428297936916351, + 0.01504445169121027, + 0.011200913228094578, + 0.031526051461696625, + 0.014297877438366413, + 0.013424163684248924, + 0.012496059760451317, + 0.010878704488277435, + 0.019507335498929024, + 0.0035492898896336555, + 0.004009956959635019, + 0.004878472536802292, + 0.0034994683228433132, + 0.005640117451548576, + 0.0041429428383708, + 0.003911919891834259, + 0.0032216638792306185, + 0.0042112842202186584, + 0.0049110352993011475, + 0.00826440192759037, + 0.004913276061415672, + 0.007946915924549103, + 0.005725867114961147, + 0.006532050669193268, + 0.009429038502275944, + 0.005841544363647699, + 0.004390384536236525, + 0.007100414019078016, + 0.005166471470147371, + 0.005066452547907829, + 0.004844256676733494, + 0.007913989014923573, + 0.0032274583354592323 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.20161204040050507 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029809776693582535 + ], + "opposite_logprob_tokens": [ + -3.5298097133636475 + ], + "logprob_reward": -0.029809776693582535, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029809776693582535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3280351161956787, + 0.20373666286468506, + 0.09072260558605194, + 0.0339355543255806, + 0.03875599056482315, + 0.051146671175956726, + 0.16069909930229187, + 0.025305232033133507, + 0.02610824443399906, + 0.04155479371547699 + ], + "credit_map": [ + 0.03942594677209854, + 0.03421223163604736, + 0.05763647332787514, + 0.05248849093914032, + 0.03709384426474571, + 0.031233767047524452, + 0.04126235470175743, + 0.0346820093691349, + 0.023353079333901405, + 0.015992337837815285, + 0.03383541852235794, + 0.022515365853905678, + 0.021868661046028137, + 0.019623521715402603, + 0.03090554103255272, + 0.035642728209495544, + 0.008830416947603226, + 0.007325842976570129, + 0.011396326124668121, + 0.009804915636777878, + 0.010786197148263454, + 0.010151482187211514, + 0.013489849865436554, + 0.018937580287456512, + 0.0030170013196766376, + 0.003912830725312233, + 0.004911401774734259, + 0.0046760085970163345, + 0.004860054235905409, + 0.0036708435509353876, + 0.004622998647391796, + 0.0042644161731004715, + 0.004486867692321539, + 0.004141354467719793, + 0.006280060857534409, + 0.005335008259862661, + 0.005304175894707441, + 0.003731674514710903, + 0.005743259098380804, + 0.0037335928063839674, + 0.005250804126262665, + 0.00364737119525671, + 0.011345483362674713, + 0.00870660413056612, + 0.004926855210214853, + 0.004664415493607521, + 0.0076797762885689735, + 0.0049253590404987335, + 0.016081849113106728, + 0.014068111777305603, + 0.03102746419608593, + 0.018106279894709587, + 0.024332374334335327, + 0.01669575832784176, + 0.01650916412472725, + 0.02387809194624424, + 0.002990684937685728, + 0.002438606694340706, + 0.004559084307402372, + 0.00445239432156086, + 0.002681904239580035, + 0.002363273175433278, + 0.003542517079040408, + 0.0022767670452594757, + 0.00343648181296885, + 0.002788909012451768, + 0.004145251587033272, + 0.0034452832769602537, + 0.003018629737198353, + 0.002906774403527379, + 0.0022152827586978674, + 0.004151631612330675, + 0.00539387296885252, + 0.004267379641532898, + 0.006435326766222715, + 0.005481659900397062, + 0.004776910878717899, + 0.004545080941170454, + 0.004282088950276375, + 0.006372472736984491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.029809776693582535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074034497141838 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011074034497141838, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074034497141838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26211026310920715, + 0.15064987540245056, + 0.12929600477218628, + 0.04045850411057472, + 0.09424144774675369, + 0.06250444054603577, + 0.05002463236451149, + 0.07972879707813263, + 0.035631924867630005, + 0.0953541174530983 + ], + "credit_map": [ + 0.03266267478466034, + 0.029499631375074387, + 0.0438852421939373, + 0.026927988976240158, + 0.03993334248661995, + 0.02832910232245922, + 0.035083141177892685, + 0.025789126753807068, + 0.015606004744768143, + 0.012100400403141975, + 0.03249196335673332, + 0.014237560331821442, + 0.018959393724799156, + 0.017003722488880157, + 0.015566970221698284, + 0.02468384988605976, + 0.02124553918838501, + 0.012891797348856926, + 0.025885120034217834, + 0.0131162628531456, + 0.016186485067009926, + 0.01355106569826603, + 0.01474068034440279, + 0.011679055169224739, + 0.0035230759531259537, + 0.004225342068821192, + 0.005918610841035843, + 0.004564461298286915, + 0.008054769597947598, + 0.005021187476813793, + 0.004848513752222061, + 0.004302540794014931, + 0.017259500920772552, + 0.009970276616513729, + 0.013034479692578316, + 0.0102460952475667, + 0.01425580121576786, + 0.010424304753541946, + 0.009729787707328796, + 0.009321204386651516, + 0.009896752424538136, + 0.0073439739644527435, + 0.00865983311086893, + 0.006630957592278719, + 0.009753569960594177, + 0.00702096289023757, + 0.006833996623754501, + 0.006364399567246437, + 0.004715423099696636, + 0.004218114074319601, + 0.008874115534126759, + 0.0055691637098789215, + 0.007178901229053736, + 0.00583761977031827, + 0.0052343388088047504, + 0.008396957069635391, + 0.009447073563933372, + 0.0071845813654363155, + 0.015570994466543198, + 0.009770630858838558, + 0.009430835954844952, + 0.008318066596984863, + 0.00823896937072277, + 0.011767645366489887, + 0.0035959782544523478, + 0.0033467744942754507, + 0.0047156112268567085, + 0.006994477938860655, + 0.004457505419850349, + 0.003523118793964386, + 0.006121218670159578, + 0.002877241699025035, + 0.017587559297680855, + 0.011802274733781815, + 0.01585390418767929, + 0.011556410230696201, + 0.010991966351866722, + 0.01054562721401453, + 0.009176705963909626, + 0.007839665748178959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011074034497141838 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023277824744582176 + ], + "opposite_logprob_tokens": [ + -3.773277759552002 + ], + "logprob_reward": -0.023277824744582176, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023277824744582176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37755951285362244, + 0.0838928297162056, + 0.1040654182434082, + 0.053785596042871475, + 0.06502200663089752, + 0.060479260981082916, + 0.09404140710830688, + 0.022552160546183586, + 0.0383814200758934, + 0.10022037476301193 + ], + "credit_map": [ + 0.055398665368556976, + 0.047512300312519073, + 0.05058345943689346, + 0.04121799394488335, + 0.0558161661028862, + 0.04537974298000336, + 0.044170182198286057, + 0.03748099133372307, + 0.013620221056044102, + 0.009842962957918644, + 0.014701764099299908, + 0.00901411846280098, + 0.011377258226275444, + 0.008705907501280308, + 0.008130816742777824, + 0.008499772287905216, + 0.018717436119914055, + 0.013675670139491558, + 0.016824107617139816, + 0.009531151503324509, + 0.010621470399200916, + 0.011624276638031006, + 0.009501015767455101, + 0.013570280745625496, + 0.005127495154738426, + 0.005366843659430742, + 0.01160787045955658, + 0.006535425316542387, + 0.005425597075372934, + 0.005557307507842779, + 0.00512458011507988, + 0.009040473960340023, + 0.011466806754469872, + 0.007752124685794115, + 0.008279277943074703, + 0.00713915703818202, + 0.007763889152556658, + 0.006392031908035278, + 0.009820467792451382, + 0.006408242043107748, + 0.008016664534807205, + 0.00848166923969984, + 0.008147083222866058, + 0.005830592010170221, + 0.007787475362420082, + 0.006494282744824886, + 0.009225667454302311, + 0.006495823618024588, + 0.009324843063950539, + 0.008355471305549145, + 0.01733328215777874, + 0.010053071193397045, + 0.01519079227000475, + 0.010049723088741302, + 0.00909329205751419, + 0.014640928246080875, + 0.0028713031206279993, + 0.003178786253556609, + 0.002584272064268589, + 0.0031554100569337606, + 0.00400985311716795, + 0.0023383728694170713, + 0.002760171191766858, + 0.0016539886128157377, + 0.004053991287946701, + 0.004311691969633102, + 0.005150319542735815, + 0.005513523705303669, + 0.005683308467268944, + 0.003910314291715622, + 0.006613058503717184, + 0.003145206021144986, + 0.014025972224771976, + 0.012974400073289871, + 0.01590721122920513, + 0.009845063090324402, + 0.013035242445766926, + 0.011404396034777164, + 0.011098206043243408, + 0.01192986499518156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.023277824744582176 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle close-up captures a thin, transparent piece of plastic film, held perfectly still by an invisible static charge against the vast, smooth surface of a large office window. An imperfect rectangle with a slightly curled corner, the film catches the intense sunlight, revealing a subtle, iridescent glow across its surface and highlighting a few tiny dust particles clinging to it. Bright, golden afternoon sunlight streams directly through the clean glass, bringing the plastic film into focus with a bright, warm lens flare. Through the window, the deep blue of a clear afternoon sky and the blurred architectural forms of distant city buildings are visible, all slightly distorted by the glass. Blurred, ghostly reflections of the empty office corridor—the edges of empty desks and the geometric pattern of the ceiling—superimposed on the view outside enhance the quiet loneliness. The film appears to float weightlessly, its edge casting a thin, sharp shadow on the glass, testament to the unseen force holding it in place. The entire scene is bathed in warm, high-contrast light, creating a tranquil and contemplative atmosphere that transforms mundane details into moments of serene, minimalist beauty.", + "seed": 3132, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_7/index_5502_seed_3132_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08356214053928852, + "vlm_logprob": -0.08356214053928852, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.08356214053928852, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009178586304187775 + ], + "opposite_logprob_tokens": [ + -5.134178638458252 + ], + "logprob_reward": -0.009178586304187775, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.009178586304187775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10263711214065552, + 0.16599631309509277, + 0.10165716707706451, + 0.07969211041927338, + 0.11048998683691025, + 0.07233479619026184, + 0.0920679122209549, + 0.10740470141172409, + 0.09226232022047043, + 0.07545758783817291 + ], + "credit_map": [ + 0.013571584597229958, + 0.024179497733712196, + 0.00928884744644165, + 0.010662850923836231, + 0.014538079500198364, + 0.009346513077616692, + 0.013422674499452114, + 0.007627064362168312, + 0.020210599526762962, + 0.018201449885964394, + 0.028442202135920525, + 0.019132954999804497, + 0.017891988158226013, + 0.02199367992579937, + 0.0260981023311615, + 0.014025325886905193, + 0.010247836820781231, + 0.01991390436887741, + 0.008617536164820194, + 0.008861822076141834, + 0.01876991242170334, + 0.01608104258775711, + 0.012482739984989166, + 0.0066823698580265045, + 0.013085323385894299, + 0.013585014268755913, + 0.010102658532559872, + 0.00846258644014597, + 0.01093603391200304, + 0.011492897756397724, + 0.0074953618459403515, + 0.004532230086624622, + 0.015318203717470169, + 0.015069122426211834, + 0.013315846212208271, + 0.012907623313367367, + 0.018768878653645515, + 0.01513402909040451, + 0.011431044898927212, + 0.008545243181288242, + 0.010308204218745232, + 0.011612164787948132, + 0.008415676653385162, + 0.006065754685550928, + 0.011234577745199203, + 0.01009950041770935, + 0.008976261131465435, + 0.005622657015919685, + 0.010544417425990105, + 0.012472056783735752, + 0.01702887937426567, + 0.010860953480005264, + 0.009715324267745018, + 0.012829326093196869, + 0.008302283473312855, + 0.010314670391380787, + 0.014964125119149685, + 0.01854180358350277, + 0.012035917490720749, + 0.01123439148068428, + 0.012980829924345016, + 0.014204194769263268, + 0.008413000032305717, + 0.015030433423817158, + 0.012504592537879944, + 0.012839709408581257, + 0.01668436825275421, + 0.009589721448719501, + 0.013728204183280468, + 0.012734269723296165, + 0.008234410546720028, + 0.005947041790932417, + 0.006100313272327185, + 0.012197324074804783, + 0.011059521697461605, + 0.008968990296125412, + 0.012757998891174793, + 0.010928727686405182, + 0.0071543920785188675, + 0.006290321238338947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.009178586304187775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.313495010137558 + ], + "opposite_logprob_tokens": [ + -1.3134950399398804 + ], + "logprob_reward": -0.313495010137558, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.313495010137558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14473696053028107, + 0.06506463885307312, + 0.08636876940727234, + 0.09874621033668518, + 0.0516115166246891, + 0.06118617579340935, + 0.09908730536699295, + 0.2070743441581726, + 0.11027825623750687, + 0.07584581524133682 + ], + "credit_map": [ + 0.017069775611162186, + 0.01285236980766058, + 0.02632661536335945, + 0.015621659345924854, + 0.01620372012257576, + 0.014101165346801281, + 0.0267488993704319, + 0.01581275276839733, + 0.007365072146058083, + 0.006722995080053806, + 0.008894819766283035, + 0.009398632682859898, + 0.008508001454174519, + 0.009121393784880638, + 0.008525685407221317, + 0.0065280417911708355, + 0.0061309803277254105, + 0.0119656166061759, + 0.010182566940784454, + 0.015593256801366806, + 0.010369690135121346, + 0.013042224571108818, + 0.008639606647193432, + 0.010444821789860725, + 0.009426448494195938, + 0.016277369111776352, + 0.008080749772489071, + 0.008743587881326675, + 0.017118630930781364, + 0.01568589173257351, + 0.012503410689532757, + 0.01091012079268694, + 0.005977456457912922, + 0.005876102950423956, + 0.007939118891954422, + 0.007498330436646938, + 0.005631196312606335, + 0.00719393091276288, + 0.007317846175283194, + 0.004177531227469444, + 0.00483484985306859, + 0.010221187025308609, + 0.005559252109378576, + 0.005105405114591122, + 0.007231491152197123, + 0.009751373901963234, + 0.005744808353483677, + 0.012737804092466831, + 0.010648229159414768, + 0.02213062159717083, + 0.014525054022669792, + 0.009232653304934502, + 0.013469372875988483, + 0.0146378418430686, + 0.008056706748902798, + 0.0063868314027786255, + 0.030087420716881752, + 0.03511232137680054, + 0.03052644617855549, + 0.01905723474919796, + 0.025146305561065674, + 0.030390402302145958, + 0.013359779492020607, + 0.02339443378150463, + 0.01602509245276451, + 0.017108896747231483, + 0.014057107269763947, + 0.01104732695966959, + 0.01985158585011959, + 0.015208487398922443, + 0.007679985370486975, + 0.009299778379499912, + 0.0060873860493302345, + 0.010748098604381084, + 0.010366338305175304, + 0.007439451292157173, + 0.014090054668486118, + 0.01104395929723978, + 0.007330271881073713, + 0.008740256540477276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.313495010137558 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206968426704407 + ], + "opposite_logprob_tokens": [ + -2.8120696544647217 + ], + "logprob_reward": -0.06206968426704407, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206968426704407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1604166328907013, + 0.051621969789266586, + 0.06890605390071869, + 0.1363798826932907, + 0.05961637943983078, + 0.30508098006248474, + 0.05110358074307442, + 0.05672668293118477, + 0.051059968769550323, + 0.0590878501534462 + ], + "credit_map": [ + 0.020984450355172157, + 0.03239383548498154, + 0.01815257966518402, + 0.01452688593417406, + 0.02461421675980091, + 0.016433240845799446, + 0.018654754385352135, + 0.014656678773462772, + 0.005599223077297211, + 0.005762928631156683, + 0.005915752612054348, + 0.0063077970407903194, + 0.008544579148292542, + 0.0059081451036036015, + 0.007141672074794769, + 0.006441871635615826, + 0.011375727131962776, + 0.01177034992724657, + 0.006386774592101574, + 0.006806868128478527, + 0.008499112911522388, + 0.009796530939638615, + 0.007436373271048069, + 0.006834312342107296, + 0.01737074926495552, + 0.015839407220482826, + 0.016379669308662415, + 0.018244855105876923, + 0.013056190684437752, + 0.0179336816072464, + 0.018767233937978745, + 0.01878809928894043, + 0.008681762963533401, + 0.007568991743028164, + 0.007150663994252682, + 0.006928509101271629, + 0.009823854081332684, + 0.007929312065243721, + 0.007413328625261784, + 0.004119954537600279, + 0.04333433881402016, + 0.04290540888905525, + 0.0413767509162426, + 0.0302809476852417, + 0.042170945554971695, + 0.03917311877012253, + 0.0403028205037117, + 0.025536667555570602, + 0.006236253306269646, + 0.006379580125212669, + 0.011867799796164036, + 0.004614402074366808, + 0.00477948784828186, + 0.007192203775048256, + 0.005291165318340063, + 0.004742689896374941, + 0.007675644941627979, + 0.006572897080332041, + 0.009673717431724072, + 0.00522646913304925, + 0.007039370946586132, + 0.00709338765591383, + 0.005786587484180927, + 0.007658607792109251, + 0.007200939580798149, + 0.008434015326201916, + 0.005577308125793934, + 0.004404623992741108, + 0.008006746880710125, + 0.006968220695853233, + 0.004670795053243637, + 0.005797316785901785, + 0.004469734616577625, + 0.011344467289745808, + 0.0074003636837005615, + 0.006693211384117603, + 0.007602724712342024, + 0.008217930793762207, + 0.007947529666125774, + 0.0054118880070745945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.06206968426704407 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784122928977013 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.009784122928977013, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784122928977013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19361509382724762, + 0.0848216563463211, + 0.09587215632200241, + 0.0558459609746933, + 0.1695336550474167, + 0.11173852533102036, + 0.07062867283821106, + 0.08076522499322891, + 0.07584203034639359, + 0.061337001621723175 + ], + "credit_map": [ + 0.030863329768180847, + 0.021893789991736412, + 0.03407851234078407, + 0.016646886244416237, + 0.02457936480641365, + 0.018659699708223343, + 0.029649395495653152, + 0.017244137823581696, + 0.011742164380848408, + 0.015381510369479656, + 0.008619382046163082, + 0.008586368523538113, + 0.012908952310681343, + 0.011723252944648266, + 0.010434441268444061, + 0.005425583571195602, + 0.009006903506815434, + 0.015571463853120804, + 0.006841377355158329, + 0.01335367001593113, + 0.01526101678609848, + 0.014921347610652447, + 0.011948800645768642, + 0.008967582136392593, + 0.006980313919484615, + 0.010043053887784481, + 0.005723113659769297, + 0.005789536517113447, + 0.008116928860545158, + 0.007821322418749332, + 0.007204537279903889, + 0.004167152103036642, + 0.02407224476337433, + 0.021264828741550446, + 0.02101915515959263, + 0.019080601632595062, + 0.0290253683924675, + 0.02234007604420185, + 0.016350893303751945, + 0.016380473971366882, + 0.0175028033554554, + 0.014675452373921871, + 0.01596107892692089, + 0.007818380370736122, + 0.01491500437259674, + 0.015221931040287018, + 0.010736743919551373, + 0.01490713283419609, + 0.007884709164500237, + 0.010035484097898006, + 0.014422008767724037, + 0.006166841834783554, + 0.008193881250917912, + 0.00991959311068058, + 0.0080182533711195, + 0.005987904965877533, + 0.00910892616957426, + 0.01203763298690319, + 0.010808147490024567, + 0.007283678278326988, + 0.013284544460475445, + 0.01099475845694542, + 0.010849922895431519, + 0.006397610530257225, + 0.0073106549680233, + 0.013617119751870632, + 0.007118106819689274, + 0.006344971247017384, + 0.014532068744301796, + 0.01101191807538271, + 0.007822218351066113, + 0.00808497704565525, + 0.005196562502533197, + 0.0093074357137084, + 0.0091858534142375, + 0.007252213545143604, + 0.008949953131377697, + 0.008820478804409504, + 0.007023962680250406, + 0.005600543227046728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.009784122928977013 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283299058675766 + ], + "opposite_logprob_tokens": [ + -3.7732832431793213 + ], + "logprob_reward": -0.023283299058675766, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283299058675766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17574730515480042, + 0.03489585965871811, + 0.03076016902923584, + 0.10025406628847122, + 0.11318459361791611, + 0.06191264092922211, + 0.17509792745113373, + 0.15541133284568787, + 0.06050761416554451, + 0.09222851693630219 + ], + "credit_map": [ + 0.025497574359178543, + 0.0208748746663332, + 0.0296992938965559, + 0.014038052409887314, + 0.025306906551122665, + 0.017620431259274483, + 0.026559878140687943, + 0.016150284558534622, + 0.005822313949465752, + 0.002882446860894561, + 0.004452300723642111, + 0.003419064451009035, + 0.004974358715116978, + 0.003869342617690563, + 0.00716900359839201, + 0.0023070252500474453, + 0.005485289730131626, + 0.0047677066177129745, + 0.0038802458439022303, + 0.0026011012960225344, + 0.004179686773568392, + 0.004133081994950771, + 0.002708907937631011, + 0.0030041455756872892, + 0.013684552162885666, + 0.019014758989214897, + 0.013486179523169994, + 0.00952932145446539, + 0.01415703259408474, + 0.015212916769087315, + 0.009888527914881706, + 0.005280767567455769, + 0.01870747283101082, + 0.012515868991613388, + 0.01860373467206955, + 0.014456121250987053, + 0.011937241069972515, + 0.014014482498168945, + 0.013414589688181877, + 0.009535077959299088, + 0.006357905920594931, + 0.010227401740849018, + 0.005747383926063776, + 0.005343250930309296, + 0.010118137113749981, + 0.008894526399672031, + 0.008845692500472069, + 0.006378339137881994, + 0.01593114249408245, + 0.030902722850441933, + 0.029011448845267296, + 0.015159941278398037, + 0.02781939134001732, + 0.024104131385684013, + 0.021052146330475807, + 0.011116999201476574, + 0.021344680339097977, + 0.019041715189814568, + 0.03121512196958065, + 0.013676107861101627, + 0.020504118874669075, + 0.02022285759449005, + 0.016285663470625877, + 0.01312105730175972, + 0.006890500895678997, + 0.008744234219193459, + 0.008917721919715405, + 0.0068026334047317505, + 0.008558528497815132, + 0.007824527099728584, + 0.006322205066680908, + 0.0064472598023712635, + 0.008128205314278603, + 0.015047336928546429, + 0.014673423022031784, + 0.011157435365021229, + 0.013511675409972668, + 0.01269784476608038, + 0.007963652722537518, + 0.009048939682543278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.023283299058675766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angle close-up captures a thin, transparent piece of plastic film, held perfectly still by an invisible static charge against the vast, smooth surface of a large office window. An imperfect rectangle with a slightly curled corner, the film catches the intense sunlight, revealing a subtle, iridescent glow across its surface and highlighting a few tiny dust particles clinging to it. Bright, golden afternoon sunlight streams directly through the clean glass, bringing the plastic film into focus with a bright, warm lens flare. Through the window, the deep blue of a clear afternoon sky and the blurred architectural forms of distant city buildings are visible, all slightly distorted by the glass. Blurred, ghostly reflections of the empty office corridor—the edges of empty desks and the geometric pattern of the ceiling—superimposed on the view outside enhance the quiet loneliness. The film appears to float weightlessly, its edge casting a thin, sharp shadow on the glass, testament to the unseen force holding it in place. The entire scene is bathed in warm, high-contrast light, creating a tranquil and contemplative atmosphere that transforms mundane details into moments of serene, minimalist beauty.", + "seed": 3133, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_7/index_5502_seed_3133_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08079187348484992, + "vlm_logprob": -0.08079187348484992, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.08079187348484992, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03349441662430763 + ], + "opposite_logprob_tokens": [ + -3.533494472503662 + ], + "logprob_reward": -0.03349441662430763, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03349441662430763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13124413788318634, + 0.2077319622039795, + 0.05397450551390648, + 0.06458181887865067, + 0.1289675086736679, + 0.2787524461746216, + 0.048658691346645355, + 0.020062224939465523, + 0.019010843709111214, + 0.047015830874443054 + ], + "credit_map": [ + 0.013291175477206707, + 0.022342614829540253, + 0.012565044686198235, + 0.016997013241052628, + 0.017365790903568268, + 0.01923331245779991, + 0.016730161383748055, + 0.012719023041427135, + 0.026833461597561836, + 0.0410483293235302, + 0.017155544832348824, + 0.02453606389462948, + 0.03025742992758751, + 0.032789211720228195, + 0.019262641668319702, + 0.015849271789193153, + 0.006655944976955652, + 0.010038163512945175, + 0.004992777016013861, + 0.006152022629976273, + 0.010364288464188576, + 0.006702923681586981, + 0.005645326804369688, + 0.003423057496547699, + 0.00797145627439022, + 0.008554776199162006, + 0.005997404456138611, + 0.01265597902238369, + 0.011017999611794949, + 0.004887356422841549, + 0.007406136021018028, + 0.006090710870921612, + 0.017824087291955948, + 0.021814478561282158, + 0.010282274335622787, + 0.018069332465529442, + 0.024145547300577164, + 0.020969850942492485, + 0.009349372237920761, + 0.006512559950351715, + 0.0439835786819458, + 0.04718152806162834, + 0.019852545112371445, + 0.04541270434856415, + 0.04773472994565964, + 0.034358832985162735, + 0.02616121433675289, + 0.014067329466342926, + 0.007869753986597061, + 0.008143018931150436, + 0.0034144544042646885, + 0.007096005138009787, + 0.005634800996631384, + 0.007547625806182623, + 0.005290655419230461, + 0.0036623787600547075, + 0.003390972502529621, + 0.0034427207428961992, + 0.0012691024458035827, + 0.003394629340618849, + 0.0029623666778206825, + 0.0017242649337276816, + 0.00249978294596076, + 0.0013783854665234685, + 0.0032148994505405426, + 0.003146602539345622, + 0.0012405974557623267, + 0.002477027475833893, + 0.003132730955258012, + 0.002459946321323514, + 0.0016082156216725707, + 0.001730823190882802, + 0.003960927948355675, + 0.009753533639013767, + 0.0023752145934849977, + 0.003488511312752962, + 0.01799697056412697, + 0.0038687046617269516, + 0.004094006959348917, + 0.001477963523939252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.03349441662430763 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31350719928741455 + ], + "opposite_logprob_tokens": [ + -1.3135071992874146 + ], + "logprob_reward": -0.31350719928741455, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31350719928741455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14690174162387848, + 0.15564283728599548, + 0.23669342696666718, + 0.039762694388628006, + 0.06850837916135788, + 0.10016713291406631, + 0.06648547202348709, + 0.0227341391146183, + 0.028140155598521233, + 0.13496403396129608 + ], + "credit_map": [ + 0.015857351943850517, + 0.025318745523691177, + 0.012726008892059326, + 0.013716417364776134, + 0.02880353108048439, + 0.021896813064813614, + 0.017557954415678978, + 0.011024932377040386, + 0.019711367785930634, + 0.03107113018631935, + 0.013128506019711494, + 0.018407564610242844, + 0.021388983353972435, + 0.0242678914219141, + 0.015863114967942238, + 0.011804274283349514, + 0.020836297422647476, + 0.05504535138607025, + 0.02213672548532486, + 0.033986691385507584, + 0.05049816891551018, + 0.022757254540920258, + 0.020171629264950752, + 0.011261316016316414, + 0.005166562739759684, + 0.006336180027574301, + 0.003948705270886421, + 0.005062185227870941, + 0.007334650494158268, + 0.004390256945043802, + 0.0043603237718343735, + 0.003163827583193779, + 0.011588000692427158, + 0.010968738235533237, + 0.006736651994287968, + 0.011727772653102875, + 0.01158200390636921, + 0.006069423630833626, + 0.005837619304656982, + 0.003998170606791973, + 0.014524459838867188, + 0.01633063703775406, + 0.007190220057964325, + 0.01520566362887621, + 0.018497347831726074, + 0.012978509068489075, + 0.01087974850088358, + 0.004560547415167093, + 0.009127032943069935, + 0.011433873325586319, + 0.006579223554581404, + 0.011824854649603367, + 0.008263755589723587, + 0.008495456539094448, + 0.005950456950813532, + 0.004810818936675787, + 0.003509125905111432, + 0.0037408024072647095, + 0.0017186518525704741, + 0.0028906851075589657, + 0.003353669075295329, + 0.0029736035503447056, + 0.002496404806151986, + 0.002051197923719883, + 0.004571146331727505, + 0.004483406897634268, + 0.0020430819131433964, + 0.0034224940463900566, + 0.0036823609843850136, + 0.003008609637618065, + 0.00391545332968235, + 0.0030136010609567165, + 0.011767611838877201, + 0.028999799862504005, + 0.006005807779729366, + 0.009003414772450924, + 0.0526617094874382, + 0.010041547007858753, + 0.012114332988858223, + 0.004369819536805153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.31350719928741455 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029813595116138458 + ], + "opposite_logprob_tokens": [ + -3.529813528060913 + ], + "logprob_reward": -0.029813595116138458, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029813595116138458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08614850044250488, + 0.26289117336273193, + 0.440133273601532, + 0.04048280045390129, + 0.03971388563513756, + 0.024668343365192413, + 0.04175672307610512, + 0.019073788076639175, + 0.00871312990784645, + 0.03641839325428009 + ], + "credit_map": [ + 0.00964928325265646, + 0.014987282454967499, + 0.005740814842283726, + 0.009535280987620354, + 0.013011055998504162, + 0.012199461460113525, + 0.011261697858572006, + 0.00976362731307745, + 0.03672896325588226, + 0.06720320135354996, + 0.01875305362045765, + 0.02642299048602581, + 0.03868751972913742, + 0.03397957608103752, + 0.02247493527829647, + 0.018640950322151184, + 0.03814687952399254, + 0.11287336051464081, + 0.03292989730834961, + 0.05119716003537178, + 0.11296219378709793, + 0.03787017986178398, + 0.03635820746421814, + 0.017795417457818985, + 0.005768941715359688, + 0.006002485286444426, + 0.0032034898176789284, + 0.0037161810323596, + 0.008389478549361229, + 0.0044295936822891235, + 0.0037407332565635443, + 0.005231898743659258, + 0.004675952717661858, + 0.006905796006321907, + 0.00276801735162735, + 0.0038858267944306135, + 0.008476072922348976, + 0.006119105964899063, + 0.003917649853974581, + 0.00296546658501029, + 0.00364531553350389, + 0.003830984700471163, + 0.0023722227197140455, + 0.0036319338250905275, + 0.003884719219058752, + 0.0032892797607928514, + 0.002340221544727683, + 0.0016736661782488227, + 0.005335202906280756, + 0.006413220893591642, + 0.004404205828905106, + 0.00656493566930294, + 0.004946572240442038, + 0.005951361730694771, + 0.004840106703341007, + 0.0033011185005307198, + 0.001968540484085679, + 0.0034119829069823027, + 0.0015971206594258547, + 0.0032334583811461926, + 0.004313324578106403, + 0.0012604780495166779, + 0.0021549661178141832, + 0.0011339174816384912, + 0.0011768161784857512, + 0.0013426193036139011, + 0.0006925821071490645, + 0.001244723447598517, + 0.0011454133782535791, + 0.0012326077558100224, + 0.0009248384740203619, + 0.0009535298449918628, + 0.0032044588588178158, + 0.007453870493918657, + 0.0016816799761727452, + 0.002189388731494546, + 0.013176383450627327, + 0.0030762769747525454, + 0.003758253064006567, + 0.0018780820537358522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.029813595116138458 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071300953626633 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.016071300953626633, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071300953626633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15915261209011078, + 0.11794187873601913, + 0.19062857329845428, + 0.09716776013374329, + 0.15060082077980042, + 0.06456514447927475, + 0.06191195547580719, + 0.038211628794670105, + 0.028933128342032433, + 0.09088652580976486 + ], + "credit_map": [ + 0.020122500136494637, + 0.03289042040705681, + 0.010601172223687172, + 0.014963895082473755, + 0.02918054349720478, + 0.020269570872187614, + 0.01846468262374401, + 0.012659802101552486, + 0.014007880352437496, + 0.016878262162208557, + 0.009927233681082726, + 0.013364517129957676, + 0.01980266347527504, + 0.020886652171611786, + 0.012144827283918858, + 0.010929829441010952, + 0.0156859140843153, + 0.04855974763631821, + 0.017238367348909378, + 0.025999227538704872, + 0.04123840108513832, + 0.017271410673856735, + 0.015323063358664513, + 0.009312435053288937, + 0.010366484522819519, + 0.01598254404962063, + 0.00718076853081584, + 0.010582289658486843, + 0.027796851471066475, + 0.010194659233093262, + 0.008102779276669025, + 0.006961380131542683, + 0.016472311690449715, + 0.022594237700104713, + 0.012284345924854279, + 0.01970169134438038, + 0.027625882998108864, + 0.028142478317022324, + 0.014577913098037243, + 0.009201937355101109, + 0.0079287588596344, + 0.011201700195670128, + 0.005481213331222534, + 0.012958272360265255, + 0.012915771454572678, + 0.0052932933904230595, + 0.005108503624796867, + 0.0036776307970285416, + 0.009594082832336426, + 0.009940065443515778, + 0.005344875156879425, + 0.00947327259927988, + 0.007126106880605221, + 0.00911602471023798, + 0.006496279966086149, + 0.004821244161576033, + 0.006121423561125994, + 0.0065630641765892506, + 0.002689980436116457, + 0.005630137398838997, + 0.00594516284763813, + 0.0037328284233808517, + 0.004226953722536564, + 0.003302072873339057, + 0.005269263871014118, + 0.005205024033784866, + 0.002247001975774765, + 0.004841889254748821, + 0.0029818713665008545, + 0.002422567456960678, + 0.003616622881963849, + 0.002348885405808687, + 0.011222449131309986, + 0.014202678576111794, + 0.0062996926717460155, + 0.010183140635490417, + 0.01933898590505123, + 0.014255878515541553, + 0.006388894747942686, + 0.008994805626571178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.016071300953626633 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a thin, transparent piece of plastic film visible in the foreground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the plastic film held perfectly still against a large office window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature bright, golden afternoon sunlight streaming through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred architectural forms of distant city buildings visible through the window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are ghostly reflections of an empty office corridor superimposed on the view outside?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072855442762375 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072855442762375, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072855442762375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1582009196281433, + 0.2289111465215683, + 0.11618701368570328, + 0.04183420538902283, + 0.16416311264038086, + 0.051114585250616074, + 0.04900706559419632, + 0.036315031349658966, + 0.04444653168320656, + 0.10982044041156769 + ], + "credit_map": [ + 0.017620334401726723, + 0.03733798861503601, + 0.010724686086177826, + 0.012304877862334251, + 0.033080413937568665, + 0.01798076555132866, + 0.016077762469649315, + 0.013074084185063839, + 0.02901639975607395, + 0.04595831409096718, + 0.01860026828944683, + 0.027572454884648323, + 0.03304291144013405, + 0.03554731234908104, + 0.02160843461751938, + 0.017565041780471802, + 0.011195146478712559, + 0.02490491420030594, + 0.011967732571065426, + 0.015530874021351337, + 0.024280523881316185, + 0.01310261245816946, + 0.008987574838101864, + 0.006217632908374071, + 0.004496308974921703, + 0.005341269541531801, + 0.004330739378929138, + 0.005295388866215944, + 0.007090180646628141, + 0.006204108241945505, + 0.004928628448396921, + 0.004147583618760109, + 0.01947650872170925, + 0.029912350699305534, + 0.011369558051228523, + 0.020122850313782692, + 0.033206719905138016, + 0.03067406639456749, + 0.011333856731653214, + 0.008067206479609013, + 0.00676209619268775, + 0.008269386366009712, + 0.004151108209043741, + 0.00777916144579649, + 0.008439221419394016, + 0.006484014913439751, + 0.006021990440785885, + 0.0032076032366603613, + 0.007382386364042759, + 0.007896832190454006, + 0.003911430481821299, + 0.005912161432206631, + 0.006725040730088949, + 0.007590902969241142, + 0.006020301952958107, + 0.003568013198673725, + 0.004460691940039396, + 0.005591388326138258, + 0.002508551813662052, + 0.005267005413770676, + 0.004157012794166803, + 0.005013264715671539, + 0.005367945414036512, + 0.003949170466512442, + 0.006159194279462099, + 0.006404767278581858, + 0.0029050379525870085, + 0.0053423806093633175, + 0.007175120525062084, + 0.006333978846669197, + 0.005554993636906147, + 0.004571058321744204, + 0.010245890356600285, + 0.02166781947016716, + 0.0051857344806194305, + 0.008398359641432762, + 0.039066869765520096, + 0.009707032702863216, + 0.010609283111989498, + 0.004939446225762367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.011072855442762375 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_8/index_tensor([3724, 3724])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_8/index_tensor([3724, 3724])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..165b50c11921aa29ac6373657d80ef4bf6a9a479 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_8/index_tensor([3724, 3724])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot with a shaky handheld perspective, this wide-angle shot emulates the unsteady gaze of a breathless onlooker observing a magnificent, multi-story mansion violently engulfed in flames beneath a turbulent night sky. The entire scene is rendered in the dramatic, swirling style of a Van Gogh painting, with thick, impassive brushstrokes visible on every surface. Blazing flames, composed of vibrant smears of cadmium yellow, fiery orange, and crimson, burst from arched windows and greedily lick at the stone walls, their light casting a frantic, flickering glow. Above, the turbulent night sky churns deep, anxious blues and violets, while clouds and billowing smoke swirl together in a thick, textured vortex of paint, reflecting the chaos below. The mansion's silhouette begins to bend and twist under the intense heat, its architectural details fused into the fiery spectacle, sparks and embers rendered as bright, fleeting brushstrokes that drift on the wind. The camera's constant, shaky movement heightens the overwhelming sense of panic and awe, completely immersing the viewer in the terrifying yet beautiful spectacle of destruction.", + "seed": 29831, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_8/index_3724_seed_29831_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3057701658457518, + "vlm_logprob": -0.3057701658457518, + "vlm_yes_no_margin": 2.2749999880790712, + "vlm_token_credit": -0.3057701658457518, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027895521372556686 + ], + "opposite_logprob_tokens": [ + -3.777895450592041 + ], + "logprob_reward": -0.027895521372556686, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027895521372556686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5031113028526306, + 0.0412411130964756, + 0.09967184066772461, + 0.05820072814822197, + 0.12891879677772522, + 0.08605298399925232, + 0.023805953562259674, + 0.007351134438067675, + 0.04046018421649933, + 0.011185928247869015 + ], + "credit_map": [ + 0.07436191290616989, + 0.0872853547334671, + 0.04383336752653122, + 0.04299897328019142, + 0.062403514981269836, + 0.10869264602661133, + 0.04310857504606247, + 0.040426962077617645, + 0.004458218347281218, + 0.004835003986954689, + 0.004610816948115826, + 0.005304323043674231, + 0.005445774644613266, + 0.003773650387302041, + 0.004803421441465616, + 0.008009904064238071, + 0.008091696538031101, + 0.01146028097718954, + 0.009372033178806305, + 0.018318861722946167, + 0.010968947783112526, + 0.009810577146708965, + 0.016241561621427536, + 0.015407883562147617, + 0.008455791510641575, + 0.007555094081908464, + 0.0075379833579063416, + 0.005450785160064697, + 0.0064746318385005, + 0.013227353803813457, + 0.003988000564277172, + 0.005511084571480751, + 0.017202425748109818, + 0.0436428003013134, + 0.00548706715926528, + 0.005307476036250591, + 0.012761574238538742, + 0.03602590039372444, + 0.0037812404334545135, + 0.004710308741778135, + 0.019578831270337105, + 0.0270107239484787, + 0.008807012811303139, + 0.005111037753522396, + 0.010795188136398792, + 0.007114459294825792, + 0.0031851036474108696, + 0.004450627602636814, + 0.002574169309809804, + 0.003969147335737944, + 0.002835614373907447, + 0.0032817968167364597, + 0.002677486976608634, + 0.002919838996604085, + 0.002721476601436734, + 0.0028264212887734175, + 0.0010661710985004902, + 0.000795943196862936, + 0.000991982757113874, + 0.0013101431541144848, + 0.0008918203529901803, + 0.0006328227464109659, + 0.000890687748324126, + 0.0007715633255429566, + 0.004547532182186842, + 0.00496867997571826, + 0.004618923179805279, + 0.004374969284981489, + 0.00853062141686678, + 0.0048148357309401035, + 0.004020571243017912, + 0.0045840502716600895, + 0.002697800053283572, + 0.001595215522684157, + 0.0012273330939933658, + 0.0013949130661785603, + 0.0009739276138134301, + 0.0010367375798523426, + 0.0011234261328354478, + 0.0011365751270204782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.027895521372556686 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983187697827816 + ], + "opposite_logprob_tokens": [ + -3.529831886291504 + ], + "logprob_reward": -0.02983187697827816, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983187697827816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6196039915084839, + 0.06509318202733994, + 0.02934149093925953, + 0.008318730629980564, + 0.1386805772781372, + 0.0835564136505127, + 0.011254286393523216, + 0.009220143780112267, + 0.020467832684516907, + 0.014463347382843494 + ], + "credit_map": [ + 0.09133785218000412, + 0.11621274799108505, + 0.0655360221862793, + 0.04809070751070976, + 0.06969540566205978, + 0.13387316465377808, + 0.04811859503388405, + 0.046739500015974045, + 0.01245423499494791, + 0.010862988419830799, + 0.010156246833503246, + 0.0071686566807329655, + 0.006698367651551962, + 0.00844988040626049, + 0.0038207401521503925, + 0.005482071544975042, + 0.0036450931802392006, + 0.0029693772085011005, + 0.0028166589327156544, + 0.0050894226878881454, + 0.0029380249325186014, + 0.0029586961027234793, + 0.004502867814153433, + 0.004421351011842489, + 0.0008986840257421136, + 0.0011307536624372005, + 0.0011047455482184887, + 0.0008771581924520433, + 0.0012771657202392817, + 0.0011089053004980087, + 0.0008139941492117941, + 0.0011073248460888863, + 0.025845132768154144, + 0.03460260108113289, + 0.008496987633407116, + 0.005167531780898571, + 0.009914308786392212, + 0.04552069678902626, + 0.003517372766509652, + 0.0056159538216888905, + 0.011222231201827526, + 0.02700057066977024, + 0.01037012878805399, + 0.005162539426237345, + 0.012331949546933174, + 0.009911803528666496, + 0.0027064743917435408, + 0.004850724712014198, + 0.0018130744574591517, + 0.0017097315285354853, + 0.0011557387188076973, + 0.0010942643275484443, + 0.0013526164693757892, + 0.0014819662319496274, + 0.0010105848778039217, + 0.0016363104805350304, + 0.0011028668377548456, + 0.0012489092769101262, + 0.0016180264065042138, + 0.0008468955638818443, + 0.0015428407350555062, + 0.0009677051566541195, + 0.0010224872967228293, + 0.0008704132633283734, + 0.0018427208997309208, + 0.002585216425359249, + 0.0023152488283813, + 0.0018946296768262982, + 0.00464034965261817, + 0.0036770692095160484, + 0.0016687045572325587, + 0.0018438944825902581, + 0.0021143376361578703, + 0.0015335078351199627, + 0.0018890004139393568, + 0.0016488626133650541, + 0.0022686703596264124, + 0.0015456159599125385, + 0.0019068026449531317, + 0.001556551898829639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.02983187697827816 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31345805525779724 + ], + "opposite_logprob_tokens": [ + -1.3134580850601196 + ], + "logprob_reward": -0.31345805525779724, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31345805525779724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3872316777706146, + 0.07205589860677719, + 0.03712010383605957, + 0.04094954580068588, + 0.3198007345199585, + 0.07749482244253159, + 0.014429798349738121, + 0.016158772632479668, + 0.017295047640800476, + 0.017463553696870804 + ], + "credit_map": [ + 0.057063616812229156, + 0.07458342611789703, + 0.04656556248664856, + 0.02945016697049141, + 0.04106202721595764, + 0.08346941322088242, + 0.027410266920924187, + 0.02762722410261631, + 0.0117785120382905, + 0.011134574189782143, + 0.012744245119392872, + 0.0076752351596951485, + 0.008181331679224968, + 0.008656922727823257, + 0.0055548567324876785, + 0.006330218631774187, + 0.0037754406221210957, + 0.0045683858916163445, + 0.003738881554454565, + 0.0047583500854671, + 0.004702837206423283, + 0.0034046913497149944, + 0.004732735455036163, + 0.007438784930855036, + 0.005295596085488796, + 0.005722121801227331, + 0.003483224892988801, + 0.003934819716960192, + 0.005390682723373175, + 0.008144320920109749, + 0.0027363516855984926, + 0.006242427974939346, + 0.05196842551231384, + 0.09483763575553894, + 0.011543353088200092, + 0.01074673980474472, + 0.025129832327365875, + 0.10526492446660995, + 0.008513444103300571, + 0.011796358041465282, + 0.013973934575915337, + 0.01124703697860241, + 0.017520515248179436, + 0.005410078447312117, + 0.009962350130081177, + 0.009589266031980515, + 0.003395745297893882, + 0.006395895499736071, + 0.0024439815897494555, + 0.0017750344704836607, + 0.0018328165169805288, + 0.001679863897152245, + 0.0017191683873534203, + 0.002133964328095317, + 0.0012223956873640418, + 0.0016225731233134866, + 0.0022755200043320656, + 0.0018777173245325685, + 0.0023939309176057577, + 0.001845354912802577, + 0.0022518171463161707, + 0.0014699377352371812, + 0.0017263964982703328, + 0.0023180977441370487, + 0.002660707337781787, + 0.0019494622247293591, + 0.0024667498655617237, + 0.001803821767680347, + 0.0028140658978372812, + 0.0022699199616909027, + 0.0016764228930696845, + 0.0016538994386792183, + 0.004961776547133923, + 0.0024322057142853737, + 0.0021294099278748035, + 0.0017462376272305846, + 0.0022617997601628304, + 0.0014939634129405022, + 0.0013714722590520978, + 0.0010666892630979419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.31345805525779724 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575884729623795 + ], + "opposite_logprob_tokens": [ + -3.895576000213623 + ], + "logprob_reward": -0.020575884729623795, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575884729623795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1053132638335228, + 0.22048179805278778, + 0.04166487604379654, + 0.0332811176776886, + 0.21887119114398956, + 0.27095353603363037, + 0.01851298287510872, + 0.013019909150898457, + 0.025385698303580284, + 0.052515577524900436 + ], + "credit_map": [ + 0.01319111604243517, + 0.027441255748271942, + 0.014715881086885929, + 0.008147470653057098, + 0.017343850806355476, + 0.009984656237065792, + 0.007190709933638573, + 0.007298333570361137, + 0.04422297328710556, + 0.03131586313247681, + 0.030517514795064926, + 0.023539777845144272, + 0.022976143285632133, + 0.029170623049139977, + 0.013475712388753891, + 0.025263190269470215, + 0.006459279451519251, + 0.004201656207442284, + 0.004175346344709396, + 0.006642588879913092, + 0.004387489520013332, + 0.00453758891671896, + 0.0054633161053061485, + 0.005797609221190214, + 0.005829783622175455, + 0.004862090572714806, + 0.003925170283764601, + 0.003257329808548093, + 0.0035343794152140617, + 0.006513306871056557, + 0.0024409727193415165, + 0.0029180869460105896, + 0.04963381215929985, + 0.039914511144161224, + 0.016552003100514412, + 0.008922056294977665, + 0.01398790068924427, + 0.0737120732665062, + 0.005825106054544449, + 0.010323727503418922, + 0.04108425974845886, + 0.0978967621922493, + 0.028030866757035255, + 0.015351518988609314, + 0.03944896534085274, + 0.028505055233836174, + 0.0075205545872449875, + 0.013115540146827698, + 0.0028505930677056313, + 0.002438229974359274, + 0.0026074042543768883, + 0.002134874230250716, + 0.001954598119482398, + 0.0020890142768621445, + 0.001644108910113573, + 0.0027941600419580936, + 0.0015591428382322192, + 0.0018081831512972713, + 0.002699990523979068, + 0.0015459428541362286, + 0.0014667115174233913, + 0.001282054465264082, + 0.001257981057278812, + 0.0013999014627188444, + 0.0017560824053362012, + 0.0035903588868677616, + 0.003040942130610347, + 0.002071390626952052, + 0.006502654403448105, + 0.0033919180277734995, + 0.0027676306199282408, + 0.0022647185251116753, + 0.005285375751554966, + 0.005961618386209011, + 0.00749093247577548, + 0.007486602757126093, + 0.007168599404394627, + 0.0044807628728449345, + 0.00696696899831295, + 0.007674719672650099 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 4 + ], + "selected_reward": -0.020575884729623795 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.137089490890503 + ], + "opposite_logprob_tokens": [ + -0.38708943128585815 + ], + "logprob_reward": -1.137089490890503, + "yes_no_margin_reward": -0.7500000596046448, + "token_credit_reward": -1.137089490890503, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3892097771167755, + 0.05328326299786568, + 0.02180839329957962, + 0.06906185299158096, + 0.280631959438324, + 0.12213553488254547, + 0.037452373653650284, + 0.0076474701054394245, + 0.011657909490168095, + 0.007111448794603348 + ], + "credit_map": [ + 0.0552714504301548, + 0.07916320860385895, + 0.04068928584456444, + 0.03296162560582161, + 0.04724099487066269, + 0.07324472069740295, + 0.03303004801273346, + 0.0276084803044796, + 0.008921647444367409, + 0.00738401897251606, + 0.007100616581737995, + 0.006063740700483322, + 0.006602991372346878, + 0.0063505591824650764, + 0.005154413171112537, + 0.005705272778868675, + 0.003985289484262466, + 0.0029159358236938715, + 0.003207682864740491, + 0.0025017219595611095, + 0.0023418248165398836, + 0.0021649464033544064, + 0.0019403852056711912, + 0.0027506074402481318, + 0.008642242290079594, + 0.011345826089382172, + 0.009990470483899117, + 0.006627949420362711, + 0.010013028047978878, + 0.00747303944081068, + 0.00451288977637887, + 0.010456412099301815, + 0.040464848279953, + 0.09385007619857788, + 0.03466510772705078, + 0.012425785884261131, + 0.03051280416548252, + 0.046758756041526794, + 0.01151295006275177, + 0.01044165063649416, + 0.02705536037683487, + 0.030835559591650963, + 0.01572946086525917, + 0.008323203772306442, + 0.013196956366300583, + 0.01291385293006897, + 0.00575844245031476, + 0.0083227027207613, + 0.0035318885929882526, + 0.005121780559420586, + 0.004388067405670881, + 0.005785094108432531, + 0.004579178523272276, + 0.004901558626443148, + 0.004365867469459772, + 0.004778937436640263, + 0.0007870274712331593, + 0.0009872694499790668, + 0.0009810782503336668, + 0.0010791962267830968, + 0.0013237087987363338, + 0.0007521772640757263, + 0.0008562798611819744, + 0.0008807331905700266, + 0.0015275402693077922, + 0.001390647841617465, + 0.0015990849351510406, + 0.0010640202090144157, + 0.0021524836774915457, + 0.0015035913093015552, + 0.0013143907999619842, + 0.0011061497498303652, + 0.0009235405013896525, + 0.00099277519620955, + 0.0008798126364126801, + 0.000931252259761095, + 0.00106428109575063, + 0.0006426466861739755, + 0.0008452192414551973, + 0.0008319210610352457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -1.137089490890503 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "Shot with a shaky handheld perspective, this wide-angle shot emulates the unsteady gaze of a breathless onlooker observing a magnificent, multi-story mansion violently engulfed in flames beneath a turbulent night sky. The entire scene is rendered in the dramatic, swirling style of a Van Gogh painting, with thick, impassive brushstrokes visible on every surface. Blazing flames, composed of vibrant smears of cadmium yellow, fiery orange, and crimson, burst from arched windows and greedily lick at the stone walls, their light casting a frantic, flickering glow. Above, the turbulent night sky churns deep, anxious blues and violets, while clouds and billowing smoke swirl together in a thick, textured vortex of paint, reflecting the chaos below. The mansion's silhouette begins to bend and twist under the intense heat, its architectural details fused into the fiery spectacle, sparks and embers rendered as bright, fleeting brushstrokes that drift on the wind. The camera's constant, shaky movement heightens the overwhelming sense of panic and awe, completely immersing the viewer in the terrifying yet beautiful spectacle of destruction.", + "seed": 29832, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_8/index_3724_seed_29832_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2861453376710415, + "vlm_logprob": -0.2861453376710415, + "vlm_yes_no_margin": 2.7749999880790712, + "vlm_token_credit": -0.2861453376710415, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025768499821424484 + ], + "opposite_logprob_tokens": [ + -3.775768518447876 + ], + "logprob_reward": -0.025768499821424484, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025768499821424484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12058958411216736, + 0.09532865881919861, + 0.1870894432067871, + 0.041764382272958755, + 0.05573948472738266, + 0.052376460283994675, + 0.28636273741722107, + 0.021165015175938606, + 0.09975935518741608, + 0.03982485458254814 + ], + "credit_map": [ + 0.010097426362335682, + 0.009414054453372955, + 0.010778393596410751, + 0.0103862090036273, + 0.03441011533141136, + 0.024937961250543594, + 0.009457830339670181, + 0.011107598431408405, + 0.006224682088941336, + 0.005956041626632214, + 0.007243547588586807, + 0.0077586364932358265, + 0.02953033335506916, + 0.0233937855809927, + 0.007818621583282948, + 0.007403009571135044, + 0.02155391126871109, + 0.018504228442907333, + 0.01684003509581089, + 0.020887745544314384, + 0.030678007751703262, + 0.020910946652293205, + 0.03259323909878731, + 0.025121323764324188, + 0.005095663480460644, + 0.007041573990136385, + 0.00435000890865922, + 0.004276270978152752, + 0.008213107474148273, + 0.0043425546027719975, + 0.004180234391242266, + 0.004264967981725931, + 0.004344089888036251, + 0.008184334263205528, + 0.008043606765568256, + 0.009952823631465435, + 0.0073042819276452065, + 0.004766685888171196, + 0.007179807871580124, + 0.0059638540260493755, + 0.005875808652490377, + 0.004936563782393932, + 0.009978537447750568, + 0.007067461963742971, + 0.006043401081115007, + 0.0054412586614489555, + 0.007107284385710955, + 0.005926142446696758, + 0.03294841945171356, + 0.04871862381696701, + 0.03297687694430351, + 0.03489332273602486, + 0.04758452996611595, + 0.029955433681607246, + 0.02766943909227848, + 0.031616099178791046, + 0.002783017000183463, + 0.002545777475461364, + 0.0026783638168126345, + 0.0018078493885695934, + 0.004121881444007158, + 0.0025433884002268314, + 0.001840845332480967, + 0.002843893365934491, + 0.008952916599810123, + 0.01379996445029974, + 0.007565378211438656, + 0.011083856225013733, + 0.02074982412159443, + 0.01247121300548315, + 0.01327779795974493, + 0.011858407407999039, + 0.004696691408753395, + 0.005287855863571167, + 0.004445902071893215, + 0.0035733149852603674, + 0.00857486855238676, + 0.004061865154653788, + 0.0037476816214621067, + 0.005436676554381847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.025768499821424484 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421798299998045 + ], + "opposite_logprob_tokens": [ + -4.264217853546143 + ], + "logprob_reward": -0.01421798299998045, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421798299998045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08355110138654709, + 0.04626989737153053, + 0.4192951023578644, + 0.04095608368515968, + 0.05819220840930939, + 0.03763577714562416, + 0.0712517499923706, + 0.036376453936100006, + 0.1172478199005127, + 0.08922378718852997 + ], + "credit_map": [ + 0.008496781811118126, + 0.007540940772742033, + 0.010386389680206776, + 0.009921804070472717, + 0.017140891402959824, + 0.010876556858420372, + 0.009287713095545769, + 0.00990003440529108, + 0.003464721841737628, + 0.005169112700968981, + 0.005767946597188711, + 0.005821715109050274, + 0.007857832126319408, + 0.006426400970667601, + 0.006328171119093895, + 0.005434001330286264, + 0.04688703641295433, + 0.04297998547554016, + 0.036719828844070435, + 0.0347960963845253, + 0.07628855109214783, + 0.05164800211787224, + 0.07298558205366135, + 0.05699004977941513, + 0.00485055148601532, + 0.0037994992453604937, + 0.005915642715990543, + 0.0047466023825109005, + 0.006019615102559328, + 0.003264554776251316, + 0.004680073820054531, + 0.0076795537024736404, + 0.006786382291465998, + 0.009278702549636364, + 0.008631209842860699, + 0.00942178163677454, + 0.0061797997914254665, + 0.004814412910491228, + 0.007382968440651894, + 0.005696962587535381, + 0.004541328642517328, + 0.00439474917948246, + 0.0049300300888717175, + 0.005235247779637575, + 0.006475441623479128, + 0.0033749036956578493, + 0.004438393283635378, + 0.004245684947818518, + 0.007548321038484573, + 0.010394272394478321, + 0.008748834952712059, + 0.008738108910620213, + 0.012604976072907448, + 0.00787063967436552, + 0.006878201384097338, + 0.008468400686979294, + 0.003158118575811386, + 0.002888489281758666, + 0.00507714506238699, + 0.004832098726183176, + 0.007743846159428358, + 0.005081717390567064, + 0.003549803514033556, + 0.0040452382527291775, + 0.011824877932667732, + 0.013324768282473087, + 0.012746197171509266, + 0.01135367900133133, + 0.02343798242509365, + 0.013250518590211868, + 0.012921348214149475, + 0.01838845945894718, + 0.011829384602606297, + 0.007201836444437504, + 0.006946623791009188, + 0.006000759080052376, + 0.024521155282855034, + 0.010098283179104328, + 0.011334242299199104, + 0.011291508562862873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.01421798299998045 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035936772823334 + ], + "opposite_logprob_tokens": [ + -1.9103593826293945 + ], + "logprob_reward": -0.16035936772823334, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035936772823334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12766024470329285, + 0.037858542054891586, + 0.41466906666755676, + 0.082148477435112, + 0.08656483888626099, + 0.025124821811914444, + 0.05919056013226509, + 0.04946407303214073, + 0.06755600869655609, + 0.04976346716284752 + ], + "credit_map": [ + 0.010717776603996754, + 0.007726628798991442, + 0.010240798816084862, + 0.010269277729094028, + 0.04085589945316315, + 0.031149420887231827, + 0.008514533750712872, + 0.008185883983969688, + 0.0042578415013849735, + 0.0039201099425554276, + 0.005421004258096218, + 0.0039635165594518185, + 0.005612161010503769, + 0.006303049623966217, + 0.003370266407728195, + 0.005010589491575956, + 0.04514777660369873, + 0.03862587735056877, + 0.035915400832891464, + 0.03566693142056465, + 0.07097144424915314, + 0.049610819667577744, + 0.07611317932605743, + 0.06261761486530304, + 0.008517851121723652, + 0.008857011795043945, + 0.0073211160488426685, + 0.007743009831756353, + 0.01508221123367548, + 0.007187479175627232, + 0.01133860182017088, + 0.016101183369755745, + 0.010938555933535099, + 0.01054102648049593, + 0.013932141475379467, + 0.013411271385848522, + 0.009690509177744389, + 0.0074315788224339485, + 0.011687936261296272, + 0.00893181562423706, + 0.002600600942969322, + 0.00287534948438406, + 0.0032687357161194086, + 0.002970548812299967, + 0.004646903369575739, + 0.002481789095327258, + 0.002872997196391225, + 0.003407893469557166, + 0.007021164521574974, + 0.011148049496114254, + 0.006254198960959911, + 0.008307515643537045, + 0.007406159769743681, + 0.005164595786482096, + 0.006831848993897438, + 0.00705702044069767, + 0.004960578866302967, + 0.004968988709151745, + 0.005997367203235626, + 0.006865568459033966, + 0.008331907913088799, + 0.005185456480830908, + 0.0052667162381112576, + 0.007887482643127441, + 0.006467548198997974, + 0.006385218817740679, + 0.008592999540269375, + 0.008124369196593761, + 0.013891780748963356, + 0.007908042520284653, + 0.007090932689607143, + 0.009095113724470139, + 0.006508032791316509, + 0.0049276151694357395, + 0.006627551279962063, + 0.008130748756229877, + 0.00839630700647831, + 0.00505180424079299, + 0.0051130312494933605, + 0.005008370615541935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.16035936772823334 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731097586452961 + ], + "opposite_logprob_tokens": [ + -5.006731033325195 + ], + "logprob_reward": -0.006731097586452961, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731097586452961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10255195200443268, + 0.0537564717233181, + 0.19022701680660248, + 0.14797982573509216, + 0.08095389604568481, + 0.040546271950006485, + 0.04605714976787567, + 0.04761284589767456, + 0.24364207684993744, + 0.04667249321937561 + ], + "credit_map": [ + 0.007345480844378471, + 0.006914210505783558, + 0.00919811800122261, + 0.008603808470070362, + 0.02859368920326233, + 0.02553035132586956, + 0.008082304149866104, + 0.008284000679850578, + 0.003949970472604036, + 0.004296604543924332, + 0.0041121081449091434, + 0.005017601884901524, + 0.013999858871102333, + 0.011845207773149014, + 0.004597754217684269, + 0.00593737093731761, + 0.01848345622420311, + 0.019673414528369904, + 0.015963127836585045, + 0.014615214429795742, + 0.035792525857686996, + 0.02859794907271862, + 0.03080715984106064, + 0.026294179260730743, + 0.013596942648291588, + 0.017437858507037163, + 0.015098103322088718, + 0.01458123978227377, + 0.023068388924002647, + 0.013069245964288712, + 0.018072916194796562, + 0.0330551452934742, + 0.012091321870684624, + 0.009970147162675858, + 0.008375468663871288, + 0.00844596792012453, + 0.01290991622954607, + 0.007544985041022301, + 0.010018112137913704, + 0.011597982607781887, + 0.004505694843828678, + 0.00566988717764616, + 0.005590879358351231, + 0.005859835539013147, + 0.005493041593581438, + 0.0035758265294134617, + 0.00516968360170722, + 0.0046814302913844585, + 0.0054383473470807076, + 0.006589088588953018, + 0.00577751686796546, + 0.005712023004889488, + 0.007137840148061514, + 0.004459988791495562, + 0.004789119586348534, + 0.006153229624032974, + 0.004832028411328793, + 0.005170505028218031, + 0.005010455846786499, + 0.0039960239082574844, + 0.01167959813028574, + 0.006521778646856546, + 0.004076569806784391, + 0.006325889378786087, + 0.023769140243530273, + 0.02275889739394188, + 0.018496911972761154, + 0.01864119991660118, + 0.06801344454288483, + 0.03336423635482788, + 0.026880227029323578, + 0.031718041747808456, + 0.005885099992156029, + 0.0042191543616354465, + 0.0036118680145591497, + 0.0027889886405318975, + 0.011692974716424942, + 0.0049561928026378155, + 0.006769417319446802, + 0.006748798303306103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.006731097586452961 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2236497402191162 + ], + "opposite_logprob_tokens": [ + -0.34864968061447144 + ], + "logprob_reward": -1.2236497402191162, + "yes_no_margin_reward": -0.8750000596046448, + "token_credit_reward": -1.2236497402191162, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14916673302650452, + 0.1150200217962265, + 0.3001018762588501, + 0.026501454412937164, + 0.032337162643671036, + 0.06491200625896454, + 0.05093732103705406, + 0.02792501263320446, + 0.13909316062927246, + 0.09400521963834763 + ], + "credit_map": [ + 0.013305198401212692, + 0.012167840264737606, + 0.012358758598566055, + 0.01381667796522379, + 0.0392819344997406, + 0.03186516836285591, + 0.012094622477889061, + 0.014276538044214249, + 0.007678697817027569, + 0.008633112534880638, + 0.005472846329212189, + 0.005253973882645369, + 0.043186649680137634, + 0.03162417560815811, + 0.0062003484927117825, + 0.006970216985791922, + 0.035774488002061844, + 0.03091561608016491, + 0.02688058279454708, + 0.028883909806609154, + 0.05070405825972557, + 0.03396974131464958, + 0.05175510048866272, + 0.04121837392449379, + 0.0026964263524860144, + 0.0024992080871015787, + 0.0034195638727396727, + 0.0028942886274307966, + 0.004405226558446884, + 0.00235349521972239, + 0.003736600512638688, + 0.004496644716709852, + 0.0037543971557170153, + 0.0031669405288994312, + 0.006116621661931276, + 0.004473505541682243, + 0.0038777925074100494, + 0.0027477352414280176, + 0.0047177341766655445, + 0.0034824348986148834, + 0.00712870666757226, + 0.008304521441459656, + 0.00964272115379572, + 0.008848803117871284, + 0.008219340816140175, + 0.005232219584286213, + 0.009094877168536186, + 0.008440817706286907, + 0.005055774934589863, + 0.006747021339833736, + 0.006303467787802219, + 0.0061473725363612175, + 0.009064159356057644, + 0.005238157231360674, + 0.00466514565050602, + 0.007716221269220114, + 0.002929103560745716, + 0.0031521848868578672, + 0.0035868387203663588, + 0.0050010355189442635, + 0.00400251429527998, + 0.002824989380314946, + 0.0033680424094200134, + 0.0030603036284446716, + 0.012804307043552399, + 0.015171695500612259, + 0.014760429039597511, + 0.01198912039399147, + 0.03681682050228119, + 0.01707782782614231, + 0.010898694396018982, + 0.019574277102947235, + 0.011182655580341816, + 0.0075834644958376884, + 0.008741226978600025, + 0.0064284480176866055, + 0.02823290042579174, + 0.011536465026438236, + 0.009531337767839432, + 0.010768716223537922 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -1.2236497402191162 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "Shot with a shaky handheld perspective, this wide-angle shot emulates the unsteady gaze of a breathless onlooker observing a magnificent, multi-story mansion violently engulfed in flames beneath a turbulent night sky. The entire scene is rendered in the dramatic, swirling style of a Van Gogh painting, with thick, impassive brushstrokes visible on every surface. Blazing flames, composed of vibrant smears of cadmium yellow, fiery orange, and crimson, burst from arched windows and greedily lick at the stone walls, their light casting a frantic, flickering glow. Above, the turbulent night sky churns deep, anxious blues and violets, while clouds and billowing smoke swirl together in a thick, textured vortex of paint, reflecting the chaos below. The mansion's silhouette begins to bend and twist under the intense heat, its architectural details fused into the fiery spectacle, sparks and embers rendered as bright, fleeting brushstrokes that drift on the wind. The camera's constant, shaky movement heightens the overwhelming sense of panic and awe, completely immersing the viewer in the terrifying yet beautiful spectacle of destruction.", + "seed": 29833, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_8/index_3724_seed_29833_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.434418919775635, + "vlm_logprob": -0.434418919775635, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.434418919775635, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011974239721894264 + ], + "opposite_logprob_tokens": [ + -4.636974334716797 + ], + "logprob_reward": -0.011974239721894264, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011974239721894264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3365557789802551, + 0.15544164180755615, + 0.05189085006713867, + 0.044109780341386795, + 0.08323642611503601, + 0.054645147174596786, + 0.05259599909186363, + 0.07426140457391739, + 0.07994116097688675, + 0.06732182949781418 + ], + "credit_map": [ + 0.02532503753900528, + 0.018907593563199043, + 0.020230557769536972, + 0.02951657958328724, + 0.018880492076277733, + 0.021402660757303238, + 0.11007780581712723, + 0.09221503138542175, + 0.009643550962209702, + 0.015756625682115555, + 0.012786198407411575, + 0.014296569861471653, + 0.014247771352529526, + 0.01213100180029869, + 0.03569371998310089, + 0.04088620841503143, + 0.00728886853903532, + 0.003479480044916272, + 0.0034712706692516804, + 0.008346698246896267, + 0.0030584584455937147, + 0.005134974606335163, + 0.005919467192143202, + 0.015191633254289627, + 0.0055821542628109455, + 0.005379708483815193, + 0.007093623746186495, + 0.0065956260077655315, + 0.003922104369848967, + 0.003722052788361907, + 0.006773841567337513, + 0.005040673539042473, + 0.0053960285149514675, + 0.01101579051464796, + 0.011023959144949913, + 0.01482191402465105, + 0.007644247729331255, + 0.006030376069247723, + 0.015373372472822666, + 0.011930735781788826, + 0.003914212808012962, + 0.008402242325246334, + 0.007803852669894695, + 0.009380373172461987, + 0.004172959830611944, + 0.004596743267029524, + 0.008059807121753693, + 0.008314953185617924, + 0.006656065583229065, + 0.00704124616459012, + 0.007076642941683531, + 0.007120536174625158, + 0.004931054078042507, + 0.00604248559102416, + 0.006806172896176577, + 0.006921796593815088, + 0.008224033750593662, + 0.010103429667651653, + 0.008648618124425411, + 0.007716495543718338, + 0.00835749413818121, + 0.009879764169454575, + 0.009909183718264103, + 0.011422382667660713, + 0.011194802820682526, + 0.007134140469133854, + 0.014757394790649414, + 0.010451680049300194, + 0.007119199726730585, + 0.010440285317599773, + 0.010745326988399029, + 0.008098328486084938, + 0.004969227593392134, + 0.007159694563597441, + 0.008634654805064201, + 0.014584321528673172, + 0.007728151511400938, + 0.005283117759972811, + 0.012875061482191086, + 0.0060876039788126945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.011974239721894264 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764515670016408 + ], + "opposite_logprob_tokens": [ + -4.882645130157471 + ], + "logprob_reward": -0.00764515670016408, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764515670016408, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19407005608081818, + 0.2540973722934723, + 0.06717086583375931, + 0.0637524276971817, + 0.04596426337957382, + 0.041315436363220215, + 0.14362061023712158, + 0.06169472634792328, + 0.06570170819759369, + 0.0626126378774643 + ], + "credit_map": [ + 0.01396651566028595, + 0.014600192196667194, + 0.01598777063190937, + 0.02409985102713108, + 0.01571330428123474, + 0.012191152200102806, + 0.045236386358737946, + 0.05227486044168472, + 0.016387520357966423, + 0.023407991975545883, + 0.022848395630717278, + 0.030892131850123405, + 0.021881308406591415, + 0.020301630720496178, + 0.05091461166739464, + 0.06746375560760498, + 0.0077560655772686005, + 0.007077959831804037, + 0.00552220456302166, + 0.010675286874175072, + 0.005063796415925026, + 0.007066207006573677, + 0.007505199871957302, + 0.016504142433404922, + 0.004368762020021677, + 0.005915868561714888, + 0.008130093105137348, + 0.01366568636149168, + 0.006125597283244133, + 0.005196265876293182, + 0.013483187183737755, + 0.006866959854960442, + 0.005149975884705782, + 0.004750590305775404, + 0.005261706188321114, + 0.006869189906865358, + 0.005075025372207165, + 0.004264988005161285, + 0.007632349617779255, + 0.006960433442145586, + 0.003834731411188841, + 0.004286648239940405, + 0.00815819576382637, + 0.005574540700763464, + 0.0029792997520416975, + 0.004372222814708948, + 0.005374168045818806, + 0.006735626608133316, + 0.017487745732069016, + 0.012721491977572441, + 0.03412775695323944, + 0.016222599893808365, + 0.01073779072612524, + 0.01624573953449726, + 0.014412962831556797, + 0.02166450023651123, + 0.010874735191464424, + 0.007929155603051186, + 0.00952916033565998, + 0.006978146266192198, + 0.004917137790471315, + 0.007837804034352303, + 0.006810650695115328, + 0.006817931775003672, + 0.00804657582193613, + 0.006086048204451799, + 0.00935043953359127, + 0.008704589679837227, + 0.006571230478584766, + 0.007341872900724411, + 0.009114365093410015, + 0.01048657950013876, + 0.008666414767503738, + 0.004911558236926794, + 0.012802815064787865, + 0.009176240302622318, + 0.005352416075766087, + 0.006383839063346386, + 0.007155361585319042, + 0.008163990452885628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.00764515670016408 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270517259836197 + ], + "opposite_logprob_tokens": [ + -2.127051830291748 + ], + "logprob_reward": -0.1270517259836197, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270517259836197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14900176227092743, + 0.2443280965089798, + 0.14391165971755981, + 0.06993330270051956, + 0.11178553849458694, + 0.10226711630821228, + 0.04718524217605591, + 0.07026882469654083, + 0.045655153691768646, + 0.015663299709558487 + ], + "credit_map": [ + 0.017129946500062943, + 0.014401569962501526, + 0.014927727170288563, + 0.020822148770093918, + 0.01408569049090147, + 0.009750326164066792, + 0.033516135066747665, + 0.02436821535229683, + 0.01930375210940838, + 0.028045669198036194, + 0.021122481673955917, + 0.030723795294761658, + 0.019480686634778976, + 0.021220430731773376, + 0.0424061194062233, + 0.06202516704797745, + 0.021159103140234947, + 0.009576166979968548, + 0.009913191199302673, + 0.02355247735977173, + 0.005075938068330288, + 0.011907154694199562, + 0.013873938471078873, + 0.048853687942028046, + 0.00902129989117384, + 0.009280663914978504, + 0.007172796875238419, + 0.012072371318936348, + 0.006493208464235067, + 0.007195563055574894, + 0.010032552294433117, + 0.008664844557642937, + 0.007164014969021082, + 0.010026383213698864, + 0.01610902138054371, + 0.01683695986866951, + 0.01063455082476139, + 0.00890287384390831, + 0.020679645240306854, + 0.02143208310008049, + 0.01036041509360075, + 0.010707231238484383, + 0.021785560995340347, + 0.012583163566887379, + 0.006592332385480404, + 0.010935219004750252, + 0.012315650470554829, + 0.016987545415759087, + 0.006288258824497461, + 0.005417156033217907, + 0.009186082519590855, + 0.006323486566543579, + 0.0036684665828943253, + 0.006039245519787073, + 0.005161948036402464, + 0.005100596696138382, + 0.007187291514128447, + 0.00987283419817686, + 0.01125334668904543, + 0.009088394232094288, + 0.006757101509720087, + 0.0071381572633981705, + 0.00817821454256773, + 0.010793483816087246, + 0.0050918180495500565, + 0.003992567770183086, + 0.006771945860236883, + 0.006712865084409714, + 0.0049951523542404175, + 0.004205980338156223, + 0.0081621203571558, + 0.005722700152546167, + 0.002311223652213812, + 0.0014401075895875692, + 0.0020412523299455643, + 0.002666481537744403, + 0.0016925460658967495, + 0.0015567379305139184, + 0.0024700395297259092, + 0.0014849113067612052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.1270517259836197 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624099031090736 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.007624099031090736, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624099031090736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2563993036746979, + 0.17424029111862183, + 0.30667147040367126, + 0.06331807374954224, + 0.054462503641843796, + 0.050570953637361526, + 0.023145049810409546, + 0.02370079606771469, + 0.0335305891931057, + 0.013960919342935085 + ], + "credit_map": [ + 0.03319327533245087, + 0.013426193967461586, + 0.02020598016679287, + 0.03790506348013878, + 0.017650898545980453, + 0.02425302378833294, + 0.06783059984445572, + 0.04193427413702011, + 0.013169807381927967, + 0.01873147115111351, + 0.01663171499967575, + 0.014965522103011608, + 0.019015371799468994, + 0.01609141193330288, + 0.03145359456539154, + 0.04418139159679413, + 0.04036390781402588, + 0.020974870771169662, + 0.023082919418811798, + 0.0524400994181633, + 0.01367131993174553, + 0.025648964568972588, + 0.030045853927731514, + 0.10044354945421219, + 0.008822301402688026, + 0.0062714931555092335, + 0.00871098693460226, + 0.006469284649938345, + 0.005475095938891172, + 0.007668651174753904, + 0.008484480902552605, + 0.011415776796638966, + 0.006770362611860037, + 0.0052375392988324165, + 0.011275284923613071, + 0.005278301890939474, + 0.004425087943673134, + 0.007057793438434601, + 0.0058536832220852375, + 0.008564449846744537, + 0.0046379477716982365, + 0.005615130066871643, + 0.010764721781015396, + 0.006237434688955545, + 0.004067507106810808, + 0.006335928570479155, + 0.005793693475425243, + 0.007118591573089361, + 0.0022026165388524532, + 0.0025692270137369633, + 0.003915705252438784, + 0.0033169875387102365, + 0.001970169134438038, + 0.002917147008702159, + 0.003216253826394677, + 0.0030369432643055916, + 0.0035144048742949963, + 0.0029776825103908777, + 0.004048673436045647, + 0.002083367668092251, + 0.0016448668902739882, + 0.002252283738926053, + 0.0028975142631679773, + 0.004282001871615648, + 0.0033752836752682924, + 0.0034929560497403145, + 0.003953014966100454, + 0.005089494399726391, + 0.003907531499862671, + 0.0032482859678566456, + 0.006512833759188652, + 0.003951188642531633, + 0.0016283515142276883, + 0.0013262671418488026, + 0.0023835627362132072, + 0.0023042119573801756, + 0.0015637109754607081, + 0.0014027849538251758, + 0.0018474680837243795, + 0.0015045625623315573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.007624099031090736 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.0177993774414062 + ], + "opposite_logprob_tokens": [ + -0.14279942214488983 + ], + "logprob_reward": -2.0177993774414062, + "yes_no_margin_reward": -1.875, + "token_credit_reward": -2.0177993774414062, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4874504804611206, + 0.08877682685852051, + 0.08854014426469803, + 0.052712008357048035, + 0.03491023927927017, + 0.08326809853315353, + 0.03572720289230347, + 0.04355701059103012, + 0.041923608630895615, + 0.04313434660434723 + ], + "credit_map": [ + 0.0223051980137825, + 0.015843916684389114, + 0.020800957456231117, + 0.03806281089782715, + 0.02009350061416626, + 0.01918802410364151, + 0.17188489437103271, + 0.17927120625972748, + 0.007546747103333473, + 0.012339497916400433, + 0.008037443272769451, + 0.009711862541735172, + 0.009928769432008266, + 0.008138889446854591, + 0.01670081913471222, + 0.016372807323932648, + 0.01152000017464161, + 0.006950774230062962, + 0.006119822151958942, + 0.014515756629407406, + 0.003968001808971167, + 0.007305890321731567, + 0.009129798039793968, + 0.029030097648501396, + 0.004429166205227375, + 0.006559351459145546, + 0.006782307289540768, + 0.009508165530860424, + 0.005127736367285252, + 0.003751592244952917, + 0.009049040265381336, + 0.007504646200686693, + 0.0023456718772649765, + 0.0037134981248527765, + 0.006369188893586397, + 0.004701508209109306, + 0.003160239430144429, + 0.003246838692575693, + 0.0049802870489656925, + 0.006393007468432188, + 0.008195427246391773, + 0.009598334319889545, + 0.013537093065679073, + 0.014283278025686741, + 0.005850516259670258, + 0.008442065678536892, + 0.011335975490510464, + 0.012025407515466213, + 0.0039367834106087685, + 0.003991834353655577, + 0.00617488706484437, + 0.005246005952358246, + 0.002760726260021329, + 0.004905529785901308, + 0.0035847118124365807, + 0.005126724950969219, + 0.004627414513379335, + 0.006275756750255823, + 0.006720246281474829, + 0.003904821118339896, + 0.003277568845078349, + 0.005244812462478876, + 0.004631511401385069, + 0.008874879218637943, + 0.005807699169963598, + 0.0049455114640295506, + 0.00456340191885829, + 0.005044290330260992, + 0.004371835384517908, + 0.004861054942011833, + 0.0077020470052957535, + 0.004627768881618977, + 0.004834835883229971, + 0.004013226833194494, + 0.004784771706908941, + 0.007447671610862017, + 0.005217887926846743, + 0.006548313423991203, + 0.006223941687494516, + 0.004063697066158056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -2.0177993774414062 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "Shot with a shaky handheld perspective, this wide-angle shot emulates the unsteady gaze of a breathless onlooker observing a magnificent, multi-story mansion violently engulfed in flames beneath a turbulent night sky. The entire scene is rendered in the dramatic, swirling style of a Van Gogh painting, with thick, impassive brushstrokes visible on every surface. Blazing flames, composed of vibrant smears of cadmium yellow, fiery orange, and crimson, burst from arched windows and greedily lick at the stone walls, their light casting a frantic, flickering glow. Above, the turbulent night sky churns deep, anxious blues and violets, while clouds and billowing smoke swirl together in a thick, textured vortex of paint, reflecting the chaos below. The mansion's silhouette begins to bend and twist under the intense heat, its architectural details fused into the fiery spectacle, sparks and embers rendered as bright, fleeting brushstrokes that drift on the wind. The camera's constant, shaky movement heightens the overwhelming sense of panic and awe, completely immersing the viewer in the terrifying yet beautiful spectacle of destruction.", + "seed": 29834, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_8/index_3724_seed_29834_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3099532128311694, + "vlm_logprob": -0.3099532128311694, + "vlm_yes_no_margin": 2.575000023841858, + "vlm_token_credit": -0.3099532128311694, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06608590483665466 + ], + "opposite_logprob_tokens": [ + -2.8160858154296875 + ], + "logprob_reward": -0.06608590483665466, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06608590483665466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41665276885032654, + 0.08368412405252457, + 0.038310706615448, + 0.0476938858628273, + 0.18084187805652618, + 0.044791050255298615, + 0.03589970991015434, + 0.0387098528444767, + 0.03181370347738266, + 0.08160227537155151 + ], + "credit_map": [ + 0.05337171629071236, + 0.06118464097380638, + 0.03365461900830269, + 0.029483646154403687, + 0.07141096144914627, + 0.10739420354366302, + 0.028516868129372597, + 0.03163615241646767, + 0.00632495665922761, + 0.01560811698436737, + 0.011172992177307606, + 0.006684690713882446, + 0.015248721465468407, + 0.01267315074801445, + 0.009172549471259117, + 0.006798954214900732, + 0.007302555255591869, + 0.0036685350351035595, + 0.00274188001640141, + 0.003077344736084342, + 0.0041832998394966125, + 0.006093613337725401, + 0.004570741672068834, + 0.0066727385856211185, + 0.003985777497291565, + 0.009219938889145851, + 0.0032835546880960464, + 0.004796951077878475, + 0.009614461101591587, + 0.005223321262747049, + 0.006180587690323591, + 0.005389297381043434, + 0.01764431782066822, + 0.016652360558509827, + 0.012784358114004135, + 0.019256487488746643, + 0.02445336990058422, + 0.026307320222258568, + 0.027032341808080673, + 0.036711327731609344, + 0.004217773675918579, + 0.0035530782770365477, + 0.0037075786385685205, + 0.004253332503139973, + 0.011473499238491058, + 0.005374138709157705, + 0.006220882758498192, + 0.0059907701797783375, + 0.0011818396160379052, + 0.0026469891890883446, + 0.0014582969015464187, + 0.0019439745228737593, + 0.014537637121975422, + 0.010407515801489353, + 0.001892788801342249, + 0.0018306659767404199, + 0.005949382204562426, + 0.003415956860408187, + 0.0031842945609241724, + 0.002722673350945115, + 0.004451161716133356, + 0.006409783847630024, + 0.003580746240913868, + 0.00899585522711277, + 0.001981074921786785, + 0.0090592410415411, + 0.0020916620269417763, + 0.0023193652741611004, + 0.004518745467066765, + 0.0061019351705908775, + 0.002936012577265501, + 0.0028056695591658354, + 0.004020157735794783, + 0.005480258259922266, + 0.002330588409677148, + 0.0021300362423062325, + 0.03513801097869873, + 0.024016594514250755, + 0.003360115922987461, + 0.00512652425095439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.06608590483665466 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023314977064728737 + ], + "opposite_logprob_tokens": [ + -3.773314952850342 + ], + "logprob_reward": -0.023314977064728737, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023314977064728737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32785242795944214, + 0.017040539532899857, + 0.18941812217235565, + 0.035804495215415955, + 0.04555821418762207, + 0.021098047494888306, + 0.07215717434883118, + 0.03331027925014496, + 0.24531589448451996, + 0.012444836087524891 + ], + "credit_map": [ + 0.04303727298974991, + 0.04406438022851944, + 0.02727692760527134, + 0.02495552971959114, + 0.054240062832832336, + 0.08486882597208023, + 0.023490237072110176, + 0.025919172912836075, + 0.0038107612635940313, + 0.0012943183537572622, + 0.004236007574945688, + 0.0010119950165972114, + 0.0010983763495460153, + 0.0017955534858629107, + 0.0015657453332096338, + 0.002227781806141138, + 0.039128635078668594, + 0.013523981906473637, + 0.010654146783053875, + 0.01901254802942276, + 0.016615895554423332, + 0.04310907423496246, + 0.013921116478741169, + 0.03345271199941635, + 0.003703903639689088, + 0.007849115878343582, + 0.0022385399788618088, + 0.004046284127980471, + 0.005736137740314007, + 0.003707223804667592, + 0.003999375272542238, + 0.004523913376033306, + 0.0042777168564498425, + 0.004129767883569002, + 0.003944778814911842, + 0.004618048667907715, + 0.006197143346071243, + 0.0063118101097643375, + 0.0060311355628073215, + 0.010047812014818192, + 0.002105131046846509, + 0.001997422194108367, + 0.0012219318887218833, + 0.001122604007832706, + 0.006220476236194372, + 0.004461632110178471, + 0.0016827487852424383, + 0.0022861005272716284, + 0.0018431675853207707, + 0.005751862656325102, + 0.0015689324354752898, + 0.0015652523143216968, + 0.0328671857714653, + 0.023117776960134506, + 0.0024979233276098967, + 0.002945059910416603, + 0.0037280386313796043, + 0.002774564316496253, + 0.002369225723668933, + 0.0026551333721727133, + 0.0051492443308234215, + 0.006516746245324612, + 0.00407474534586072, + 0.00604257732629776, + 0.010619457811117172, + 0.0981103926897049, + 0.004345282446593046, + 0.00770987756550312, + 0.020732935518026352, + 0.0724315345287323, + 0.014064944349229336, + 0.017301447689533234, + 0.0012610525591298938, + 0.0012917412677779794, + 0.0020079752430319786, + 0.0011075426591560245, + 0.0016596127534285188, + 0.0011580521240830421, + 0.002275867387652397, + 0.0016829901142045856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.023314977064728737 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280955493450165 + ], + "opposite_logprob_tokens": [ + -2.0178096294403076 + ], + "logprob_reward": -0.14280955493450165, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14280955493450165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42909231781959534, + 0.030320726335048676, + 0.021645262837409973, + 0.09505724161863327, + 0.016381481662392616, + 0.011475484818220139, + 0.05680948495864868, + 0.013543152250349522, + 0.2966809868812561, + 0.028993820771574974 + ], + "credit_map": [ + 0.052363328635692596, + 0.0666261538863182, + 0.03343033418059349, + 0.03365931287407875, + 0.07736531645059586, + 0.10242140293121338, + 0.03124089352786541, + 0.03198559209704399, + 0.005157249979674816, + 0.0026559815742075443, + 0.0036104689352214336, + 0.0021621505729854107, + 0.002121062483638525, + 0.006571809761226177, + 0.003552317153662443, + 0.004489685408771038, + 0.00451295729726553, + 0.00150499667506665, + 0.0013739443384110928, + 0.002155621536076069, + 0.001585763180628419, + 0.0047858767211437225, + 0.0014712326228618622, + 0.0042548696510493755, + 0.00919894129037857, + 0.01759703829884529, + 0.007210709620267153, + 0.011690249666571617, + 0.009265420958399773, + 0.010619229637086391, + 0.012941150926053524, + 0.01653449796140194, + 0.0022776478435844183, + 0.001270212815143168, + 0.0022020083852112293, + 0.001470673829317093, + 0.0018358876695856452, + 0.0033972228411585093, + 0.0017969527980312705, + 0.002130876062437892, + 0.0012113080592826009, + 0.001007239567115903, + 0.0011183383176103234, + 0.00142491958104074, + 0.0026644268073141575, + 0.0013810015516355634, + 0.0014291070401668549, + 0.001239143661223352, + 0.0014942088164389133, + 0.00448496313765645, + 0.0013673852663487196, + 0.00152150122448802, + 0.024299316108226776, + 0.018786676228046417, + 0.002112289657816291, + 0.0027431461494416, + 0.0013314747484400868, + 0.0015524751506745815, + 0.0010604152921587229, + 0.0008992077200673521, + 0.0035451280418783426, + 0.0022779146675020456, + 0.001475503551773727, + 0.0014010329032316804, + 0.01303458958864212, + 0.11704245954751968, + 0.00596048915758729, + 0.010175157338380814, + 0.024756476283073425, + 0.08812636882066727, + 0.01730124093592167, + 0.020284190773963928, + 0.0017543187132105231, + 0.001836177660152316, + 0.0011505559086799622, + 0.0010084762470796704, + 0.011804772540926933, + 0.00768519239500165, + 0.0016475635347887874, + 0.0021067627239972353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.14280955493450165 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091939423233271 + ], + "opposite_logprob_tokens": [ + -5.504091739654541 + ], + "logprob_reward": -0.004091939423233271, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091939423233271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.123785100877285, + 0.05504883453249931, + 0.09783496707677841, + 0.3381342887878418, + 0.021917611360549927, + 0.04840552806854248, + 0.12192299962043762, + 0.0373542457818985, + 0.13441674411296844, + 0.021179761737585068 + ], + "credit_map": [ + 0.015183942392468452, + 0.00800809171050787, + 0.016308538615703583, + 0.015383673831820488, + 0.01500400435179472, + 0.02410764805972576, + 0.012697035446763039, + 0.017092155292630196, + 0.012162749655544758, + 0.005854770075529814, + 0.008180184289813042, + 0.0023766974918544292, + 0.005230375565588474, + 0.011542871594429016, + 0.0029788368847221136, + 0.006722338031977415, + 0.01989630050957203, + 0.006250809878110886, + 0.006016907747834921, + 0.007735850755125284, + 0.008519227616488934, + 0.022985178977251053, + 0.00720225740224123, + 0.01922842115163803, + 0.03349524363875389, + 0.0606081597507, + 0.025315307080745697, + 0.04075225442647934, + 0.03190018609166145, + 0.03918489068746567, + 0.045513227581977844, + 0.061364974826574326, + 0.0023672410752624273, + 0.0023654575925320387, + 0.002309066941961646, + 0.0019768807105720043, + 0.002988367574289441, + 0.004941206891089678, + 0.001918070251122117, + 0.003051317995414138, + 0.0052224332466721535, + 0.0038519911468029022, + 0.005572157446295023, + 0.0033349136356264353, + 0.006224615499377251, + 0.009131471626460552, + 0.0041709099896252155, + 0.01089702732861042, + 0.003125239396467805, + 0.00980929285287857, + 0.002591079333797097, + 0.0024087768979370594, + 0.054902199655771255, + 0.04020573943853378, + 0.004363417159765959, + 0.004517246037721634, + 0.004270851146429777, + 0.0031698073726147413, + 0.0027541599702090025, + 0.0030253506265580654, + 0.0046266308054327965, + 0.007015358190983534, + 0.0048492481000721455, + 0.007642835378646851, + 0.009503147564828396, + 0.042898379266262054, + 0.006334290839731693, + 0.0072807613760232925, + 0.014597855508327484, + 0.0331449955701828, + 0.008692728355526924, + 0.01196457352489233, + 0.0012021790025755763, + 0.0013335677795112133, + 0.001717650331556797, + 0.001568356528878212, + 0.006112281233072281, + 0.004330776631832123, + 0.0017452924512326717, + 0.0031696553342044353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.004091939423233271 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a multi-story mansion engulfed in flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the flames rendered with vibrant smears of cadmium yellow, fiery orange, and crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the visual style that of a Van Gogh painting with thick, impassive brushstrokes visible on every surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the sky turbulent with deep blues and violets, clouds, and billowing smoke swirling in a thick vortex?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot shaky and handheld, moving constantly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134636878967285 + ], + "opposite_logprob_tokens": [ + -0.3134637176990509 + ], + "logprob_reward": -1.3134636878967285, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134636878967285, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09518707543611526, + 0.03860015794634819, + 0.14231811463832855, + 0.029722455888986588, + 0.034712549299001694, + 0.015216256491839886, + 0.019106077030301094, + 0.010353956371545792, + 0.6054171323776245, + 0.0093662329018116 + ], + "credit_map": [ + 0.00967733096331358, + 0.016521235927939415, + 0.008369683288037777, + 0.009190685115754604, + 0.018576567992568016, + 0.01666094735264778, + 0.008380909450352192, + 0.007809716742485762, + 0.004158237017691135, + 0.0019520880887284875, + 0.010017091408371925, + 0.0018287583952769637, + 0.001601425581611693, + 0.01360020600259304, + 0.0016959572676569223, + 0.003746395232155919, + 0.027099894359707832, + 0.010467412881553173, + 0.008040497079491615, + 0.012939413078129292, + 0.012805814854800701, + 0.031299032270908356, + 0.013060035184025764, + 0.026606008410453796, + 0.0026274756528437138, + 0.0037036279682070017, + 0.002844268223270774, + 0.0031969284173101187, + 0.00423169881105423, + 0.004015286918729544, + 0.0037490311078727245, + 0.005354137625545263, + 0.004358137026429176, + 0.0024120877496898174, + 0.003691741032525897, + 0.0028006145730614662, + 0.0043288529850542545, + 0.008604713715612888, + 0.0033043816220015287, + 0.005212021991610527, + 0.0018098526634275913, + 0.0013717852998524904, + 0.0010103307431563735, + 0.0012453163508325815, + 0.003008926287293434, + 0.0032806082163006067, + 0.0014285692013800144, + 0.002060866914689541, + 0.0008004906121641397, + 0.0014397205086424947, + 0.0009781699627637863, + 0.0006852204678580165, + 0.007538023870438337, + 0.005971163976937532, + 0.0007022524368949234, + 0.0009910346707329154, + 0.0012958998559042811, + 0.0009496146230958402, + 0.0007062768563628197, + 0.0008517599198967218, + 0.002663220511749387, + 0.0014553142245858908, + 0.001201426493935287, + 0.0012304437113925815, + 0.02669811248779297, + 0.23994658887386322, + 0.011746424250304699, + 0.0207842830568552, + 0.05131757631897926, + 0.17865754663944244, + 0.03554503619670868, + 0.04072153568267822, + 0.0008822757517918944, + 0.000745550321880728, + 0.0011653134133666754, + 0.0008906683651730418, + 0.0015299320220947266, + 0.0009061642340384424, + 0.0015477330889552832, + 0.0016985967522487044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -1.3134636878967285 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_9/index_tensor([2003, 2003])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_9/index_tensor([2003, 2003])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..80136ef18cdbda90de8a48d44f21a12c12fd5c82 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_9/index_tensor([2003, 2003])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, taken from above and tilted slightly to the right, traces a string of loose pearls across a deep black velvet surface. The pearls are creamy white and almost perfectly spherical, radiating a soft, milky sheen as a single, gentle light source from outside the camera kisses their surface, creating delicate, mirror-like highlights. The string of beads rolls with a deliberate, yet unhurried grace, each pearl gently pressing into the plush, non-reflective fabric, leaving faint, temporary marks in the soft pile that perfectly embodies the sensation of sliding across a rich texture. The striking contrast between the luminous pearls and the light-absorbing velvet, combined with the hypnotic, steady movement, creates an atmosphere of profound tranquility, elegance, and serene luxury.", + "seed": 10742, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_9/index_2003_seed_10742_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04434101236984134, + "vlm_logprob": -0.04434101236984134, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.04434101236984134, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02421482466161251 + ], + "opposite_logprob_tokens": [ + -3.899214744567871 + ], + "logprob_reward": -0.02421482466161251, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02421482466161251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02369746007025242, + 0.014278877526521683, + 0.04512875899672508, + 0.03572338819503784, + 0.10730916261672974, + 0.018525751307606697, + 0.5985333919525146, + 0.007937716320157051, + 0.1392425298690796, + 0.009622864425182343 + ], + "credit_map": [ + 0.0025837623979896307, + 0.002062307670712471, + 0.0036443874705582857, + 0.002574090613052249, + 0.002315724501386285, + 0.0028972637373954058, + 0.003366925986483693, + 0.004253001883625984, + 0.0010056061437353492, + 0.0011823817621916533, + 0.002834657207131386, + 0.003282696707174182, + 0.0015730352606624365, + 0.0013751569204032421, + 0.0012270723236724734, + 0.0017982723657041788, + 0.0038783305790275335, + 0.004692488815635443, + 0.01044336799532175, + 0.007260797079652548, + 0.0049209315329790115, + 0.004026691894978285, + 0.0063851140439510345, + 0.0035210412461310625, + 0.0032948569860309362, + 0.003397980472072959, + 0.006041877903044224, + 0.005851535126566887, + 0.0029572430066764355, + 0.005235616583377123, + 0.0062498124316334724, + 0.002694466384127736, + 0.011459912173449993, + 0.010018008761107922, + 0.009911257773637772, + 0.023622827604413033, + 0.004897432867437601, + 0.005247214343398809, + 0.01878710463643074, + 0.02336541749536991, + 0.002319145714864135, + 0.0019268043106421828, + 0.0018896376714110374, + 0.002800795715302229, + 0.001178072765469551, + 0.0012294872431084514, + 0.0036009049508720636, + 0.003580905729904771, + 0.03111349605023861, + 0.05053464323282242, + 0.21317684650421143, + 0.045596733689308167, + 0.014607184566557407, + 0.015259898267686367, + 0.04083201289176941, + 0.18741266429424286, + 0.0007548948633484542, + 0.0013744968455284834, + 0.0006626328686252236, + 0.0013592269970104098, + 0.0007717270636931062, + 0.000683874124661088, + 0.0009275837219320238, + 0.0014032807666808367, + 0.009598973207175732, + 0.023047292605042458, + 0.025932317599654198, + 0.02216469682753086, + 0.020909227430820465, + 0.009154819883406162, + 0.013128111138939857, + 0.015307104215025902, + 0.0008104746812023222, + 0.0016319940332323313, + 0.0008827147539705038, + 0.0008693278068676591, + 0.0017245498020201921, + 0.0005619744770228863, + 0.0021074439864605665, + 0.0010343854082748294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.02421482466161251 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076707093976438046 + ], + "opposite_logprob_tokens": [ + -4.882670879364014 + ], + "logprob_reward": -0.0076707093976438046, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076707093976438046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07400266081094742, + 0.04461849108338356, + 0.030266284942626953, + 0.04174025356769562, + 0.05049184709787369, + 0.027419643476605415, + 0.41539886593818665, + 0.023254409432411194, + 0.2157273292541504, + 0.07708020508289337 + ], + "credit_map": [ + 0.00672468775883317, + 0.010844477452337742, + 0.009885624051094055, + 0.007150900084525347, + 0.009430785663425922, + 0.009146236814558506, + 0.00936362985521555, + 0.01145632378757, + 0.0039896718226373196, + 0.004893903154879808, + 0.007494383957237005, + 0.008326144888997078, + 0.004152332898229361, + 0.0037506367079913616, + 0.005048873834311962, + 0.006962548475712538, + 0.002115304348990321, + 0.0026086336001753807, + 0.007156446576118469, + 0.007818960584700108, + 0.002670244313776493, + 0.0026480972301214933, + 0.003480971558019519, + 0.0017676251009106636, + 0.003489204915240407, + 0.004191452171653509, + 0.006619696971029043, + 0.007462233770638704, + 0.0025468801613897085, + 0.004149593412876129, + 0.009645085781812668, + 0.0036361082457005978, + 0.004162712953984737, + 0.0036379953380674124, + 0.004982711747288704, + 0.006455628201365471, + 0.002336417557671666, + 0.0017626521876081824, + 0.014542199671268463, + 0.012611527927219868, + 0.0038570170290768147, + 0.00336994300596416, + 0.0027780381496995687, + 0.003627487923949957, + 0.0018623793730512261, + 0.002025087596848607, + 0.0045320638455450535, + 0.005367627367377281, + 0.02288847230374813, + 0.043623726814985275, + 0.1676931530237198, + 0.015168779529631138, + 0.00808736588805914, + 0.007268112618476152, + 0.02071547508239746, + 0.12995381653308868, + 0.002003792440518737, + 0.004145362414419651, + 0.0020075105130672455, + 0.004854240454733372, + 0.0015732228057458997, + 0.0016657582018524408, + 0.0021248722914606333, + 0.004879648331552744, + 0.015493965707719326, + 0.03393525257706642, + 0.042604126036167145, + 0.03294852748513222, + 0.030553380027413368, + 0.014467036351561546, + 0.020847810432314873, + 0.024877233430743217, + 0.005121123511344194, + 0.013718287460505962, + 0.005792644340544939, + 0.0077525596134364605, + 0.0141750518232584, + 0.0036938367411494255, + 0.018917061388492584, + 0.007909640669822693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.0076707093976438046 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026381492614746094 + ], + "opposite_logprob_tokens": [ + -3.651381492614746 + ], + "logprob_reward": -0.026381492614746094, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026381492614746094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09070783853530884, + 0.04335559159517288, + 0.03708234801888466, + 0.03701018914580345, + 0.04685056954622269, + 0.05044589191675186, + 0.22155648469924927, + 0.023299969732761383, + 0.1630290299654007, + 0.2866620719432831 + ], + "credit_map": [ + 0.007366429548710585, + 0.006769272498786449, + 0.012960424646735191, + 0.008688338100910187, + 0.008324285969138145, + 0.011490835808217525, + 0.013660967350006104, + 0.021447278559207916, + 0.00303150643594563, + 0.002895742654800415, + 0.007868755608797073, + 0.010507871396839619, + 0.004543926101177931, + 0.005158409476280212, + 0.003227044129744172, + 0.00612233579158783, + 0.0016515530878677964, + 0.0021600541658699512, + 0.013841346837580204, + 0.012950586155056953, + 0.0014696021098643541, + 0.0012072742683812976, + 0.0023865930270403624, + 0.0014153365045785904, + 0.003434686688706279, + 0.0037837582640349865, + 0.005788232199847698, + 0.006529617123305798, + 0.00259784166701138, + 0.004076092969626188, + 0.0071245720610022545, + 0.0036753856111317873, + 0.004553173203021288, + 0.003642717842012644, + 0.0051355864852666855, + 0.006653728894889355, + 0.0026816069148480892, + 0.0022823854815214872, + 0.010735249146819115, + 0.011166120879352093, + 0.006673253607004881, + 0.006707322783768177, + 0.005647165700793266, + 0.007386246230453253, + 0.0028488028328865767, + 0.002298758365213871, + 0.008945081382989883, + 0.009939259849488735, + 0.01257227547466755, + 0.021549846976995468, + 0.06879629194736481, + 0.019749917089939117, + 0.006738577038049698, + 0.006459694821387529, + 0.010629275813698769, + 0.07506059110164642, + 0.0011568850604817271, + 0.004155682399868965, + 0.0026435754261910915, + 0.004820343106985092, + 0.0016326055629178882, + 0.0015319590456783772, + 0.002126560779288411, + 0.005232356954365969, + 0.011053819209337234, + 0.026716548949480057, + 0.029856525361537933, + 0.027282891795039177, + 0.02343493141233921, + 0.010641837492585182, + 0.016420455649495125, + 0.017622027546167374, + 0.024840552359819412, + 0.06067810207605362, + 0.024601811543107033, + 0.0231604166328907, + 0.05341138690710068, + 0.010953398421406746, + 0.061752982437610626, + 0.027263421565294266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.026381492614746094 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020601926371455193 + ], + "opposite_logprob_tokens": [ + -3.895601987838745 + ], + "logprob_reward": -0.020601926371455193, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020601926371455193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17536897957324982, + 0.07166868448257446, + 0.09325090795755386, + 0.09255924820899963, + 0.05533168464899063, + 0.08207293599843979, + 0.18192745745182037, + 0.068886898458004, + 0.05456918105483055, + 0.12436401844024658 + ], + "credit_map": [ + 0.012585535645484924, + 0.018246669322252274, + 0.027140336111187935, + 0.028318271040916443, + 0.02282237447798252, + 0.017118124291300774, + 0.02731253020465374, + 0.02182513102889061, + 0.006042237859219313, + 0.006623874884098768, + 0.012449033558368683, + 0.012976064346730709, + 0.008304420858621597, + 0.0065652052871882915, + 0.008060203865170479, + 0.010647648014128208, + 0.003595333779230714, + 0.004230552352964878, + 0.033869046717882156, + 0.033179428428411484, + 0.005848840344697237, + 0.0021671252325177193, + 0.007890342734754086, + 0.002470236038789153, + 0.009062821045517921, + 0.009019458666443825, + 0.01719108410179615, + 0.017070457339286804, + 0.005772633012384176, + 0.009701802395284176, + 0.017597584053874016, + 0.007143412716686726, + 0.0050718956626951694, + 0.007204256486147642, + 0.006412899121642113, + 0.007552274037152529, + 0.004891814664006233, + 0.0034730369225144386, + 0.010781354270875454, + 0.009944154880940914, + 0.01047474890947342, + 0.008366678841412067, + 0.00831445399671793, + 0.010156331583857536, + 0.005001754034310579, + 0.005051415413618088, + 0.013709311373531818, + 0.020998245105147362, + 0.014953016303479671, + 0.03256683796644211, + 0.05721397325396538, + 0.02588341198861599, + 0.00721949664875865, + 0.008778072893619537, + 0.012959088198840618, + 0.02235357277095318, + 0.006433365400880575, + 0.012988746166229248, + 0.007350258529186249, + 0.010042602196335793, + 0.0048556141555309296, + 0.006505976896733046, + 0.005941162817180157, + 0.0147691760212183, + 0.0029039212968200445, + 0.009564572013914585, + 0.01007335539907217, + 0.01190333440899849, + 0.005716738291084766, + 0.002789563499391079, + 0.0051741464994847775, + 0.006443552207201719, + 0.009841009974479675, + 0.025386052206158638, + 0.010672020725905895, + 0.008936976082623005, + 0.026439310982823372, + 0.005362395662814379, + 0.02695588208734989, + 0.010770372115075588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.020601926371455193 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283610880374908 + ], + "opposite_logprob_tokens": [ + -2.017836093902588 + ], + "logprob_reward": -0.14283610880374908, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14283610880374908, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05769007280468941, + 0.029203621670603752, + 0.09807731211185455, + 0.10053304582834244, + 0.14768797159194946, + 0.16657710075378418, + 0.138792023062706, + 0.0187982264906168, + 0.17626315355300903, + 0.06637750566005707 + ], + "credit_map": [ + 0.004555127117782831, + 0.006825129501521587, + 0.007148211356252432, + 0.0062045929953455925, + 0.006204204633831978, + 0.007383767049759626, + 0.009691755287349224, + 0.009677277877926826, + 0.0023988985922187567, + 0.002264218870550394, + 0.005150088109076023, + 0.005966796074062586, + 0.003652864834293723, + 0.0026804201770573854, + 0.003286264603957534, + 0.0038040694780647755, + 0.003434961661696434, + 0.005023373756557703, + 0.03915244713425636, + 0.03644382208585739, + 0.0028392872773110867, + 0.0038069202564656734, + 0.00500689959153533, + 0.0023695877753198147, + 0.008619214408099651, + 0.012482157908380032, + 0.011793610639870167, + 0.01603451743721962, + 0.007079531438648701, + 0.01932651177048683, + 0.016656339168548584, + 0.008541163057088852, + 0.015625251457095146, + 0.013859882950782776, + 0.012590775266289711, + 0.039746399968862534, + 0.005671514198184013, + 0.007216534577310085, + 0.026815930381417274, + 0.02616168186068535, + 0.01980718784034252, + 0.02023576758801937, + 0.018192654475569725, + 0.02418520301580429, + 0.009198244661092758, + 0.009634810499846935, + 0.029378483071923256, + 0.03594474494457245, + 0.007012621499598026, + 0.010733581148087978, + 0.04855766519904137, + 0.010352701880037785, + 0.003731874516233802, + 0.004311205353587866, + 0.009352429769933224, + 0.04473992809653282, + 0.0013364686165004969, + 0.003895512316375971, + 0.0014381746295839548, + 0.004144774284213781, + 0.001206883229315281, + 0.001227493048645556, + 0.0011608994100242853, + 0.0043880208395421505, + 0.012678353115916252, + 0.027896521613001823, + 0.034310877323150635, + 0.027355078607797623, + 0.02471277117729187, + 0.011425543576478958, + 0.017486324533820152, + 0.02039767988026142, + 0.005056403111666441, + 0.012926481664180756, + 0.004984315950423479, + 0.0056999363005161285, + 0.012224402278661728, + 0.002352288691326976, + 0.016567735001444817, + 0.0065659345127642155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.14283610880374908 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, taken from above and tilted slightly to the right, traces a string of loose pearls across a deep black velvet surface. The pearls are creamy white and almost perfectly spherical, radiating a soft, milky sheen as a single, gentle light source from outside the camera kisses their surface, creating delicate, mirror-like highlights. The string of beads rolls with a deliberate, yet unhurried grace, each pearl gently pressing into the plush, non-reflective fabric, leaving faint, temporary marks in the soft pile that perfectly embodies the sensation of sliding across a rich texture. The striking contrast between the luminous pearls and the light-absorbing velvet, combined with the hypnotic, steady movement, creates an atmosphere of profound tranquility, elegance, and serene luxury.", + "seed": 10743, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_9/index_2003_seed_10743_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03733988255262375, + "vlm_logprob": -0.03733988255262375, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03733988255262375, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04902113229036331 + ], + "opposite_logprob_tokens": [ + -3.1740212440490723 + ], + "logprob_reward": -0.04902113229036331, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04902113229036331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12396107614040375, + 0.08863359689712524, + 0.04236689209938049, + 0.13428495824337006, + 0.0840127170085907, + 0.07499615848064423, + 0.2314659059047699, + 0.06200791895389557, + 0.10358534008264542, + 0.054685454815626144 + ], + "credit_map": [ + 0.012488821521401405, + 0.016381477937102318, + 0.015703819692134857, + 0.011921152472496033, + 0.013008690439164639, + 0.009410019032657146, + 0.028815384954214096, + 0.0162317156791687, + 0.006573187652975321, + 0.010913671925663948, + 0.019770752638578415, + 0.008309852331876755, + 0.015904225409030914, + 0.010467869229614735, + 0.010479222983121872, + 0.00621481379494071, + 0.0034954207949340343, + 0.0045110248029232025, + 0.0071331895887851715, + 0.010448438115417957, + 0.004190725274384022, + 0.006451463792473078, + 0.0032553148921579123, + 0.002881318097934127, + 0.009476917795836926, + 0.015446072444319725, + 0.03440825268626213, + 0.031638335436582565, + 0.012472053989768028, + 0.006651136092841625, + 0.013867689296603203, + 0.010324490256607533, + 0.009504454210400581, + 0.017823100090026855, + 0.013637776486575603, + 0.014251095242798328, + 0.007347108330577612, + 0.003834955859929323, + 0.011724691838026047, + 0.005889534950256348, + 0.004661296959966421, + 0.009382768534123898, + 0.014452245086431503, + 0.0177063699811697, + 0.005676541943103075, + 0.003954628482460976, + 0.014184998348355293, + 0.0049773030914366245, + 0.025966765359044075, + 0.018343165516853333, + 0.05277678370475769, + 0.05454845726490021, + 0.017008520662784576, + 0.0178337711840868, + 0.02656036615371704, + 0.018428068608045578, + 0.004427568055689335, + 0.004357344936579466, + 0.013356024399399757, + 0.01861930824816227, + 0.004865691531449556, + 0.003063535550609231, + 0.00856588501483202, + 0.0047525581903755665, + 0.006937208119779825, + 0.0056885164231061935, + 0.034580934792757034, + 0.03364871069788933, + 0.004083956126123667, + 0.004889054223895073, + 0.007621018681675196, + 0.006135944742709398, + 0.004152604844421148, + 0.004298904445022345, + 0.009745282121002674, + 0.010777230374515057, + 0.00442715547978878, + 0.004909275099635124, + 0.008684543892741203, + 0.007690455298870802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.04902113229036331 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005990882404148579 + ], + "opposite_logprob_tokens": [ + -5.130990982055664 + ], + "logprob_reward": -0.005990882404148579, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005990882404148579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1095045730471611, + 0.14391285181045532, + 0.12949183583259583, + 0.13621412217617035, + 0.06896725296974182, + 0.0847526416182518, + 0.09152302891016006, + 0.0576583631336689, + 0.07777116447687149, + 0.10020417720079422 + ], + "credit_map": [ + 0.012306919321417809, + 0.014466348104178905, + 0.011050435714423656, + 0.014826602302491665, + 0.009760351851582527, + 0.009850909002125263, + 0.01932073011994362, + 0.01792227290570736, + 0.00806812010705471, + 0.021497411653399467, + 0.026398640125989914, + 0.018457774072885513, + 0.031192675232887268, + 0.012884505093097687, + 0.01470405887812376, + 0.010709667578339577, + 0.009172717109322548, + 0.012778548523783684, + 0.02341867983341217, + 0.03835387900471687, + 0.016484860330820084, + 0.007373347878456116, + 0.011599897406995296, + 0.010309889912605286, + 0.008627811446785927, + 0.026506029069423676, + 0.03748451545834541, + 0.028460612520575523, + 0.014590826816856861, + 0.0056533208116889, + 0.00833852868527174, + 0.006552482023835182, + 0.006549017503857613, + 0.008130185306072235, + 0.011374841444194317, + 0.017576953396201134, + 0.004151315428316593, + 0.004218753892928362, + 0.010910817421972752, + 0.006055367179214954, + 0.006674149539321661, + 0.009567239321768284, + 0.013534676283597946, + 0.016754325479269028, + 0.006030591204762459, + 0.005037085153162479, + 0.02047445811331272, + 0.006680116523057222, + 0.01076434925198555, + 0.0075192200019955635, + 0.020857902243733406, + 0.02426554262638092, + 0.005423745606094599, + 0.0052350847981870174, + 0.010490786284208298, + 0.006966398097574711, + 0.005742652807384729, + 0.008654526434838772, + 0.011438964866101742, + 0.011403712444007397, + 0.006143901031464338, + 0.0034053155686706305, + 0.006994791328907013, + 0.0038744972553104162, + 0.005183599889278412, + 0.0052433316595852375, + 0.015980031341314316, + 0.01665058359503746, + 0.005955190863460302, + 0.006211020052433014, + 0.015277018770575523, + 0.00727038225159049, + 0.0071467687375843525, + 0.007129267323762178, + 0.02051267772912979, + 0.022384680807590485, + 0.007417888380587101, + 0.00889409240335226, + 0.01761130429804325, + 0.00910751149058342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.005990882404148579 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011090539395809174 + ], + "opposite_logprob_tokens": [ + -4.5110907554626465 + ], + "logprob_reward": -0.011090539395809174, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011090539395809174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040510132908821106, + 0.1323028951883316, + 0.17621174454689026, + 0.09937792271375656, + 0.20730328559875488, + 0.05216840282082558, + 0.11093760281801224, + 0.07227898389101028, + 0.06812328845262527, + 0.04078568518161774 + ], + "credit_map": [ + 0.005157358478754759, + 0.005450930446386337, + 0.0049333469942212105, + 0.0056083109229803085, + 0.0044875419698655605, + 0.004107921849936247, + 0.0064896815456449986, + 0.004275039304047823, + 0.014314982108771801, + 0.021237904205918312, + 0.028693420812487602, + 0.014203313738107681, + 0.01796797476708889, + 0.013265765272080898, + 0.014578486792743206, + 0.008041049353778362, + 0.011510450392961502, + 0.022549714893102646, + 0.039335884153842926, + 0.05535494163632393, + 0.019025003537535667, + 0.012815359979867935, + 0.008437425829470158, + 0.007182960864156485, + 0.0053923302330076694, + 0.013258764520287514, + 0.030258234590291977, + 0.02541607804596424, + 0.006708176340907812, + 0.004281457047909498, + 0.006910698022693396, + 0.007152180653065443, + 0.021825222298502922, + 0.03267785906791687, + 0.036024533212184906, + 0.04292648658156395, + 0.020868582651019096, + 0.01186734065413475, + 0.026035118848085403, + 0.015078145079314709, + 0.003175433725118637, + 0.005384577438235283, + 0.011910699307918549, + 0.013605323620140553, + 0.0032058516517281532, + 0.0032996253576129675, + 0.007266680244356394, + 0.004320213105529547, + 0.01187269389629364, + 0.009701660834252834, + 0.025713983923196793, + 0.028109034523367882, + 0.00892676506191492, + 0.00985775701701641, + 0.009137466549873352, + 0.0076182386837899685, + 0.006402766332030296, + 0.004286708310246468, + 0.01676543615758419, + 0.015638934448361397, + 0.0045699309557676315, + 0.003827808191999793, + 0.010617155581712723, + 0.01017024740576744, + 0.005399979185312986, + 0.003707005875185132, + 0.02050820365548134, + 0.02020120434463024, + 0.0029569154139608145, + 0.00494607537984848, + 0.006328807678073645, + 0.004075094126164913, + 0.002368703717365861, + 0.0027079207357019186, + 0.007680509239435196, + 0.010091068223118782, + 0.002631539711728692, + 0.0029709867667406797, + 0.00787283107638359, + 0.004462129902094603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 1 + ], + "selected_reward": -0.011090539395809174 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076385317370295525 + ], + "opposite_logprob_tokens": [ + -4.882638454437256 + ], + "logprob_reward": -0.0076385317370295525, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076385317370295525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06983818113803864, + 0.07148758322000504, + 0.0613766685128212, + 0.17391212284564972, + 0.086019366979599, + 0.06884528696537018, + 0.05778414011001587, + 0.07717287540435791, + 0.22329391539096832, + 0.11026988923549652 + ], + "credit_map": [ + 0.008769584819674492, + 0.009492486715316772, + 0.01098703034222126, + 0.007828251458704472, + 0.0065783243626356125, + 0.0065477387979626656, + 0.008993353694677353, + 0.010641410946846008, + 0.007792394608259201, + 0.011409660801291466, + 0.011841781437397003, + 0.007684668991714716, + 0.00865071453154087, + 0.009014079347252846, + 0.010423681698739529, + 0.00467060599476099, + 0.007977025583386421, + 0.008314114063978195, + 0.0089731989428401, + 0.009807354770600796, + 0.007099410984665155, + 0.005970271769911051, + 0.00568896159529686, + 0.007546333596110344, + 0.012300055474042892, + 0.03237177059054375, + 0.03675418719649315, + 0.03442048281431198, + 0.020120838657021523, + 0.009834137745201588, + 0.015407600440084934, + 0.012703055515885353, + 0.010283119045197964, + 0.013082318007946014, + 0.013330520130693913, + 0.015177702531218529, + 0.006877377163618803, + 0.00472521036863327, + 0.015711450949311256, + 0.006831674370914698, + 0.005979952402412891, + 0.007152152713388205, + 0.01317260880023241, + 0.015031421557068825, + 0.00383005035109818, + 0.0027650175616145134, + 0.01534755527973175, + 0.005566528532654047, + 0.006229328457266092, + 0.0041436487808823586, + 0.012786702252924442, + 0.01473347470164299, + 0.003424997441470623, + 0.0030002272687852383, + 0.009301586076617241, + 0.0041641732677817345, + 0.004732925910502672, + 0.0035809038672596216, + 0.018807169049978256, + 0.021249720826745033, + 0.004918726161122322, + 0.002868467476218939, + 0.013941061682999134, + 0.0070738946087658405, + 0.01823371648788452, + 0.013202929869294167, + 0.06362316757440567, + 0.06129255145788193, + 0.010342498309910297, + 0.009590154513716698, + 0.03128233551979065, + 0.01572655513882637, + 0.013753595761954784, + 0.009535400196909904, + 0.02327801287174225, + 0.019004130735993385, + 0.006698185577988625, + 0.009777208790183067, + 0.014779015444219112, + 0.013444339856505394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 9 + ], + "selected_reward": -0.0076385317370295525 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295832693576813 + ], + "opposite_logprob_tokens": [ + -2.2379584312438965 + ], + "logprob_reward": -0.11295832693576813, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295832693576813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10365565866231918, + 0.14246253669261932, + 0.09996280074119568, + 0.14016516506671906, + 0.06583213061094284, + 0.10199707001447678, + 0.10424882918596268, + 0.08605272322893143, + 0.11532214283943176, + 0.040300846099853516 + ], + "credit_map": [ + 0.010303805582225323, + 0.014958292245864868, + 0.016897032037377357, + 0.012845789082348347, + 0.01446077786386013, + 0.00856515672057867, + 0.013949253596365452, + 0.011675567366182804, + 0.015260717831552029, + 0.02165764942765236, + 0.025649359449744225, + 0.02454797364771366, + 0.01848568767309189, + 0.01436783280223608, + 0.012709815986454487, + 0.009783516637980938, + 0.007821358740329742, + 0.01272392924875021, + 0.01734275184571743, + 0.029146932065486908, + 0.011708301492035389, + 0.010604823008179665, + 0.005285926163196564, + 0.00532878702506423, + 0.012576209381222725, + 0.02337280660867691, + 0.02579803392291069, + 0.0313291922211647, + 0.014678443782031536, + 0.00882627908140421, + 0.01167948916554451, + 0.011904720216989517, + 0.004334095865488052, + 0.00856101419776678, + 0.012464787811040878, + 0.018954236060380936, + 0.004255913197994232, + 0.0032038686331361532, + 0.008692455478012562, + 0.005365762859582901, + 0.007669475395232439, + 0.010713201947510242, + 0.02233939617872238, + 0.025642050430178642, + 0.007283183746039867, + 0.006105904467403889, + 0.01210042741149664, + 0.010143444873392582, + 0.00943513959646225, + 0.008051207289099693, + 0.02277388609945774, + 0.025412363931536674, + 0.006549271289259195, + 0.008931227959692478, + 0.011577464640140533, + 0.011518272571265697, + 0.00693244906142354, + 0.004588690586388111, + 0.02012711577117443, + 0.021457182243466377, + 0.005831672810018063, + 0.0043661585077643394, + 0.01300375908613205, + 0.009745710529386997, + 0.010439768433570862, + 0.009762774221599102, + 0.023975662887096405, + 0.024782581254839897, + 0.01229132805019617, + 0.007940678857266903, + 0.013815900310873985, + 0.01231345720589161, + 0.004244646523147821, + 0.0042001912370324135, + 0.006462856661528349, + 0.005520131438970566, + 0.004081698600202799, + 0.0052728550508618355, + 0.005213871598243713, + 0.005304599646478891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.11295832693576813 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, taken from above and tilted slightly to the right, traces a string of loose pearls across a deep black velvet surface. The pearls are creamy white and almost perfectly spherical, radiating a soft, milky sheen as a single, gentle light source from outside the camera kisses their surface, creating delicate, mirror-like highlights. The string of beads rolls with a deliberate, yet unhurried grace, each pearl gently pressing into the plush, non-reflective fabric, leaving faint, temporary marks in the soft pile that perfectly embodies the sensation of sliding across a rich texture. The striking contrast between the luminous pearls and the light-absorbing velvet, combined with the hypnotic, steady movement, creates an atmosphere of profound tranquility, elegance, and serene luxury.", + "seed": 10744, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_9/index_2003_seed_10744_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.049410976376384495, + "vlm_logprob": -0.049410976376384495, + "vlm_yes_no_margin": 3.724999952316284, + "vlm_token_credit": -0.049410976376384495, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027963802218437195 + ], + "opposite_logprob_tokens": [ + -3.777963876724243 + ], + "logprob_reward": -0.027963802218437195, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.027963802218437195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0969427153468132, + 0.033196598291397095, + 0.04773952439427376, + 0.04101147502660751, + 0.12527476251125336, + 0.04393519461154938, + 0.05647037550806999, + 0.1693318784236908, + 0.25688713788986206, + 0.12921029329299927 + ], + "credit_map": [ + 0.010367150418460369, + 0.006354103796184063, + 0.011722882278263569, + 0.011269045993685722, + 0.006955942139029503, + 0.012205800041556358, + 0.02151750586926937, + 0.016550295054912567, + 0.004164228215813637, + 0.0024240536149591208, + 0.0044651152566075325, + 0.005090807564556599, + 0.0024681277573108673, + 0.0035629877820611, + 0.005859468597918749, + 0.005161813925951719, + 0.0044882772490382195, + 0.003002396784722805, + 0.010137345641851425, + 0.006321865599602461, + 0.00408811355009675, + 0.00535240676254034, + 0.0074042659252882, + 0.006944857072085142, + 0.005912881810218096, + 0.0030721768271178007, + 0.004433886613696814, + 0.005443749483674765, + 0.0030671751592308283, + 0.0063097355887293816, + 0.005860340315848589, + 0.006911532022058964, + 0.03878694027662277, + 0.0057597593404352665, + 0.013916362076997757, + 0.013031478971242905, + 0.0057057105004787445, + 0.025015119463205338, + 0.007816893979907036, + 0.0152424992993474, + 0.00814900640398264, + 0.0029518476221710443, + 0.007105440832674503, + 0.006673565600067377, + 0.004332748707383871, + 0.0036592993419617414, + 0.006337971426546574, + 0.004725319799035788, + 0.008862810209393501, + 0.007020887453109026, + 0.008012947626411915, + 0.00784219242632389, + 0.004968378692865372, + 0.005055825691670179, + 0.00935483817011118, + 0.005352497100830078, + 0.02027881145477295, + 0.01663920097053051, + 0.028917202726006508, + 0.03434233367443085, + 0.01849590241909027, + 0.013212410733103752, + 0.01891320012509823, + 0.018532821908593178, + 0.03684568777680397, + 0.013614426366984844, + 0.041241541504859924, + 0.0484326109290123, + 0.01339869573712349, + 0.013408464379608631, + 0.044728558510541916, + 0.045217160135507584, + 0.00543178291991353, + 0.003889624495059252, + 0.02179715782403946, + 0.036438748240470886, + 0.004552655387669802, + 0.0048882667906582355, + 0.03829285502433777, + 0.013919222168624401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.027963802218437195 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006773605477064848 + ], + "opposite_logprob_tokens": [ + -5.006773471832275 + ], + "logprob_reward": -0.006773605477064848, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006773605477064848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0947888121008873, + 0.13843128085136414, + 0.042269475758075714, + 0.12484731525182724, + 0.121523916721344, + 0.04769406467676163, + 0.0404156856238842, + 0.06413812190294266, + 0.2081070840358734, + 0.1177842766046524 + ], + "credit_map": [ + 0.009765361435711384, + 0.007527137640863657, + 0.014122817665338516, + 0.013543833047151566, + 0.008665506727993488, + 0.014131319709122181, + 0.012883776798844337, + 0.014149059541523457, + 0.011806625872850418, + 0.007174317725002766, + 0.0296962670981884, + 0.024710435420274734, + 0.010175623930990696, + 0.018448734655976295, + 0.01722644455730915, + 0.019192824140191078, + 0.004080873914062977, + 0.002790171420201659, + 0.007883562706410885, + 0.006276407279074192, + 0.0028159336652606726, + 0.007384652737528086, + 0.005256022792309523, + 0.005781853571534157, + 0.01567029394209385, + 0.011609331704676151, + 0.020286574959754944, + 0.017375364899635315, + 0.010431325994431973, + 0.013066803105175495, + 0.018593959510326385, + 0.017813656479120255, + 0.04436995089054108, + 0.003904668614268303, + 0.015579459257423878, + 0.005199816077947617, + 0.004538205452263355, + 0.03067045286297798, + 0.006008652504533529, + 0.011252707801759243, + 0.006709799636155367, + 0.0029691432137042284, + 0.006054944824427366, + 0.007107164710760117, + 0.005122868809849024, + 0.005464132875204086, + 0.007043182384222746, + 0.0072228251956403255, + 0.0061043621972203255, + 0.004119059536606073, + 0.005727745592594147, + 0.006990951485931873, + 0.002915098564699292, + 0.003631453262642026, + 0.007011756766587496, + 0.003915258217602968, + 0.009816832840442657, + 0.003453297773376107, + 0.009415728971362114, + 0.009622788056731224, + 0.00470196595415473, + 0.005154065787792206, + 0.012456966564059258, + 0.009516475722193718, + 0.032664015889167786, + 0.00946372002363205, + 0.030710412189364433, + 0.04914165660738945, + 0.009951271116733551, + 0.010399851948022842, + 0.03608132153749466, + 0.029694829136133194, + 0.006725273560732603, + 0.005771603435277939, + 0.020653190091252327, + 0.03214326500892639, + 0.00562645448371768, + 0.007234957069158554, + 0.022883689031004906, + 0.01674584113061428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.006773605477064848 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211518689990044 + ], + "opposite_logprob_tokens": [ + -4.264211654663086 + ], + "logprob_reward": -0.014211518689990044, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211518689990044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1749696284532547, + 0.16382041573524475, + 0.11474583297967911, + 0.04326779022812843, + 0.10009434819221497, + 0.04677552357316017, + 0.042216405272483826, + 0.12974770367145538, + 0.10243426263332367, + 0.0819280669093132 + ], + "credit_map": [ + 0.014985702000558376, + 0.008014358580112457, + 0.017143351957201958, + 0.01820203848183155, + 0.011784462258219719, + 0.019232844933867455, + 0.049008775502443314, + 0.03659810125827789, + 0.018716653808951378, + 0.00948494765907526, + 0.013941854238510132, + 0.037720054388046265, + 0.009660868905484676, + 0.011671843938529491, + 0.023746268823742867, + 0.03887791559100151, + 0.012251957319676876, + 0.0092642642557621, + 0.016097471117973328, + 0.012937633320689201, + 0.011130775325000286, + 0.010022859089076519, + 0.02554509788751602, + 0.017495762556791306, + 0.006930659059435129, + 0.002846831688657403, + 0.0055724745616316795, + 0.004944379907101393, + 0.0022062158677726984, + 0.005081353709101677, + 0.0077804336324334145, + 0.00790544506162405, + 0.023276429623365402, + 0.006645504385232925, + 0.018813177943229675, + 0.00942491926252842, + 0.004789629485458136, + 0.013246542774140835, + 0.012379363179206848, + 0.011518783867359161, + 0.008744231425225735, + 0.00286887027323246, + 0.00798973347991705, + 0.00735916942358017, + 0.004978302866220474, + 0.0034932822454720736, + 0.006497398018836975, + 0.004844534210860729, + 0.009187804535031319, + 0.0031414239201694727, + 0.006272949744015932, + 0.006998623721301556, + 0.00264742411673069, + 0.002598992781713605, + 0.00672989571467042, + 0.004639293532818556, + 0.017221635207533836, + 0.0051503959111869335, + 0.027551935985684395, + 0.0348413921892643, + 0.005744850728660822, + 0.006021550390869379, + 0.020009314641356468, + 0.013206631876528263, + 0.015778185799717903, + 0.004452642984688282, + 0.015843309462070465, + 0.024939466267824173, + 0.0043585700914263725, + 0.0054570515640079975, + 0.016988791525363922, + 0.014616242609918118, + 0.004991481080651283, + 0.0040583256632089615, + 0.016521591693162918, + 0.02028374932706356, + 0.003898230381309986, + 0.004400079138576984, + 0.018005384132266045, + 0.009769225493073463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.014211518689990044 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018200064077973366 + ], + "opposite_logprob_tokens": [ + -4.018199920654297 + ], + "logprob_reward": -0.018200064077973366, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018200064077973366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13807159662246704, + 0.2068520486354828, + 0.13886798918247223, + 0.13148237764835358, + 0.07678873091936111, + 0.04542737826704979, + 0.03808238357305527, + 0.02586062252521515, + 0.06924185156822205, + 0.12932509183883667 + ], + "credit_map": [ + 0.013525568880140781, + 0.007163120899349451, + 0.01533767394721508, + 0.01884765364229679, + 0.011521756649017334, + 0.016934597864747047, + 0.02752707339823246, + 0.027214154601097107, + 0.015396546572446823, + 0.014291598461568356, + 0.02620735950767994, + 0.050583116710186005, + 0.01331387273967266, + 0.021799881011247635, + 0.02424684166908264, + 0.04101283475756645, + 0.013796217739582062, + 0.00949444156140089, + 0.03297317400574684, + 0.013327959924936295, + 0.009415952488780022, + 0.01383231207728386, + 0.025682387873530388, + 0.020345538854599, + 0.020627280697226524, + 0.011312637478113174, + 0.016232647001743317, + 0.01751840114593506, + 0.010990163311362267, + 0.014506000094115734, + 0.01790054515004158, + 0.02239469811320305, + 0.01508188247680664, + 0.004590515047311783, + 0.0076470267958939075, + 0.013502825051546097, + 0.005165204405784607, + 0.011313673108816147, + 0.009401788003742695, + 0.010085820220410824, + 0.009754241444170475, + 0.0031585602555423975, + 0.0077404240146279335, + 0.006290166638791561, + 0.004923006985336542, + 0.0037133097648620605, + 0.005504415836185217, + 0.00434325123205781, + 0.005075139459222555, + 0.0030356072820723057, + 0.004739042371511459, + 0.005147283431142569, + 0.002847410039976239, + 0.0030167337972670794, + 0.009941278956830502, + 0.004279888700693846, + 0.003101803595200181, + 0.0014207469066604972, + 0.005049590952694416, + 0.0053484817035496235, + 0.0017385396640747786, + 0.0019161754753440619, + 0.0037928796373307705, + 0.003492405405268073, + 0.005941803567111492, + 0.005590281914919615, + 0.010747894644737244, + 0.010534849017858505, + 0.006962690036743879, + 0.0058323536068201065, + 0.011655986309051514, + 0.011975985951721668, + 0.006829221732914448, + 0.00653889449313283, + 0.0229173693805933, + 0.03629285842180252, + 0.005672947503626347, + 0.005849521141499281, + 0.03184366598725319, + 0.013380607590079308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.018200064077973366 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990589141845703 + ], + "opposite_logprob_tokens": [ + -1.804905891418457 + ], + "logprob_reward": -0.17990589141845703, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990589141845703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1027844026684761, + 0.06027785688638687, + 0.1015939861536026, + 0.08702104538679123, + 0.17501898109912872, + 0.09459106624126434, + 0.06039859727025032, + 0.08664515614509583, + 0.10652708262205124, + 0.12514182925224304 + ], + "credit_map": [ + 0.00965162180364132, + 0.005535542964935303, + 0.020001521334052086, + 0.016598789021372795, + 0.005789840128272772, + 0.012185333296656609, + 0.01613997109234333, + 0.016881784424185753, + 0.009305723942816257, + 0.003816254436969757, + 0.008324693888425827, + 0.007409415207803249, + 0.004194188397377729, + 0.007439610082656145, + 0.010719755664467812, + 0.009068216197192669, + 0.008174722082912922, + 0.007528981193900108, + 0.01811937801539898, + 0.015365819446742535, + 0.010041071102023125, + 0.01017953734844923, + 0.01769465021789074, + 0.014489826746284962, + 0.020891450345516205, + 0.005994819104671478, + 0.012413116171956062, + 0.009449193254113197, + 0.005342400167137384, + 0.008085970766842365, + 0.009414032101631165, + 0.015430061146616936, + 0.050282374024391174, + 0.008877887390553951, + 0.021722210571169853, + 0.0202712994068861, + 0.00728441821411252, + 0.031561776995658875, + 0.014911239966750145, + 0.02010776475071907, + 0.018536582589149475, + 0.00581066869199276, + 0.016437001526355743, + 0.012540935538709164, + 0.011049101129174232, + 0.006708301138132811, + 0.013514727354049683, + 0.009993746876716614, + 0.008558016270399094, + 0.00633199093863368, + 0.006760329008102417, + 0.007708496879786253, + 0.004128650296479464, + 0.005156327039003372, + 0.015488703735172749, + 0.00626608869060874, + 0.0068372804671525955, + 0.005555533803999424, + 0.016408728435635567, + 0.018622487783432007, + 0.0071212369948625565, + 0.007344558835029602, + 0.012928814627230167, + 0.011826510541141033, + 0.008622409775853157, + 0.004939239472150803, + 0.01922028698027134, + 0.02078886330127716, + 0.0074798655696213245, + 0.007434946019202471, + 0.016656920313835144, + 0.021384555846452713, + 0.009871424175798893, + 0.005397729575634003, + 0.0213947594165802, + 0.02752053737640381, + 0.0060390145517885685, + 0.00837706308811903, + 0.02968287095427513, + 0.016858424991369247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.17990589141845703 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, taken from above and tilted slightly to the right, traces a string of loose pearls across a deep black velvet surface. The pearls are creamy white and almost perfectly spherical, radiating a soft, milky sheen as a single, gentle light source from outside the camera kisses their surface, creating delicate, mirror-like highlights. The string of beads rolls with a deliberate, yet unhurried grace, each pearl gently pressing into the plush, non-reflective fabric, leaving faint, temporary marks in the soft pile that perfectly embodies the sensation of sliding across a rich texture. The striking contrast between the luminous pearls and the light-absorbing velvet, combined with the hypnotic, steady movement, creates an atmosphere of profound tranquility, elegance, and serene luxury.", + "seed": 10745, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000074/group_9/index_2003_seed_10745_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.046908419020473956, + "vlm_logprob": -0.046908419020473956, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.046908419020473956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02416653372347355 + ], + "opposite_logprob_tokens": [ + -4.024166584014893 + ], + "logprob_reward": -0.02416653372347355, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02416653372347355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0924881175160408, + 0.34401991963386536, + 0.09928391873836517, + 0.10705775767564774, + 0.04770796000957489, + 0.06547502428293228, + 0.04392186924815178, + 0.1172017902135849, + 0.036139532923698425, + 0.04670408368110657 + ], + "credit_map": [ + 0.014840492978692055, + 0.012993922457098961, + 0.011044586077332497, + 0.009143995121121407, + 0.010193069465458393, + 0.009445000439882278, + 0.011099936440587044, + 0.01372712291777134, + 0.02057167887687683, + 0.026827428489923477, + 0.04009854421019554, + 0.08377031981945038, + 0.03351246193051338, + 0.03251652792096138, + 0.054270509630441666, + 0.05245247110724449, + 0.019153853878378868, + 0.008207199163734913, + 0.009056996554136276, + 0.0084706274792552, + 0.008969997987151146, + 0.011320664547383785, + 0.012020798400044441, + 0.02208378165960312, + 0.012179581448435783, + 0.008475051261484623, + 0.010773790068924427, + 0.007628235500305891, + 0.009070159867405891, + 0.014206788502633572, + 0.020051460713148117, + 0.024672703817486763, + 0.007194082718342543, + 0.004100379999727011, + 0.006852531805634499, + 0.004872473422437906, + 0.003799986094236374, + 0.007453544065356255, + 0.006158299744129181, + 0.007276664022356272, + 0.008640068583190441, + 0.004723592661321163, + 0.013497663661837578, + 0.00786032434552908, + 0.00457069231197238, + 0.007280384190380573, + 0.007461377419531345, + 0.01144092995673418, + 0.0108413090929389, + 0.0027886710595339537, + 0.007010664325207472, + 0.0048555959947407246, + 0.0029714913107454777, + 0.0036163402255624533, + 0.003964542411267757, + 0.00787325855344534, + 0.01582067273557186, + 0.011439262889325619, + 0.015742622315883636, + 0.014640315435826778, + 0.020425882190465927, + 0.01536327414214611, + 0.01279425248503685, + 0.010975513607263565, + 0.005315810441970825, + 0.004193390253931284, + 0.0050688995979726315, + 0.008400339633226395, + 0.002433213870972395, + 0.002439949195832014, + 0.0037396808620542288, + 0.004548248834908009, + 0.004644994158297777, + 0.007046862505376339, + 0.007198499981313944, + 0.00804267730563879, + 0.002975668990984559, + 0.002805646974593401, + 0.007827349938452244, + 0.006162391044199467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.02416653372347355 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007678990252315998 + ], + "opposite_logprob_tokens": [ + -4.882678985595703 + ], + "logprob_reward": -0.007678990252315998, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007678990252315998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32616502046585083, + 0.13540661334991455, + 0.09307029843330383, + 0.04294808208942413, + 0.047560639679431915, + 0.03694570064544678, + 0.10997768491506577, + 0.09005186706781387, + 0.03740496188402176, + 0.08046915382146835 + ], + "credit_map": [ + 0.018973292782902718, + 0.027758140116930008, + 0.02350720390677452, + 0.02458975836634636, + 0.024388786405324936, + 0.02246350422501564, + 0.10178876668214798, + 0.08269554376602173, + 0.007850742898881435, + 0.012016062624752522, + 0.019601009786128998, + 0.029835321009159088, + 0.01477836538106203, + 0.013585987500846386, + 0.018277769908308983, + 0.01946134865283966, + 0.010235334746539593, + 0.008603825233876705, + 0.007767647039145231, + 0.012646764516830444, + 0.012295927852392197, + 0.009713469073176384, + 0.013444286771118641, + 0.018363039940595627, + 0.004672217182815075, + 0.007522690109908581, + 0.004920283332467079, + 0.004627041518688202, + 0.0034741878043860197, + 0.005844860803335905, + 0.0048709409311413765, + 0.007015862502157688, + 0.006228186655789614, + 0.0038745286874473095, + 0.007220031693577766, + 0.006980582140386105, + 0.0033583762124180794, + 0.006274616811424494, + 0.006095595192164183, + 0.007528720423579216, + 0.004643282853066921, + 0.0028602387756109238, + 0.0045300694182515144, + 0.004787903744727373, + 0.0029591659549623728, + 0.0049825310707092285, + 0.007171629462391138, + 0.005010878201574087, + 0.01860376074910164, + 0.008099615573883057, + 0.017962250858545303, + 0.015345476567745209, + 0.00840152706950903, + 0.011129912920296192, + 0.014934643171727657, + 0.015500489622354507, + 0.01345271710306406, + 0.010691441595554352, + 0.012161986902356148, + 0.013287700712680817, + 0.007179358974099159, + 0.009231600910425186, + 0.01282312162220478, + 0.011223937384784222, + 0.0038044038228690624, + 0.004121730569750071, + 0.005423966329544783, + 0.008992152288556099, + 0.00307273305952549, + 0.003572325222194195, + 0.0045570628717541695, + 0.0038605905137956142, + 0.00753183476626873, + 0.006496242247521877, + 0.014764906838536263, + 0.017463866621255875, + 0.005482949316501617, + 0.004890558775514364, + 0.01500905305147171, + 0.00882973987609148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.007678990252315998 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05501466244459152 + ], + "opposite_logprob_tokens": [ + -2.9300146102905273 + ], + "logprob_reward": -0.05501466244459152, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05501466244459152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43958544731140137, + 0.08543241024017334, + 0.023802928626537323, + 0.07067791372537613, + 0.02867269329726696, + 0.0589929036796093, + 0.03575098514556885, + 0.18176941573619843, + 0.044085901230573654, + 0.03122936375439167 + ], + "credit_map": [ + 0.025998150929808617, + 0.031203804537653923, + 0.03477019816637039, + 0.036956626921892166, + 0.03137194737792015, + 0.030598146840929985, + 0.13947747647762299, + 0.10920914262533188, + 0.005481001455336809, + 0.008149468339979649, + 0.010192034766077995, + 0.02109646610915661, + 0.008956382051110268, + 0.007745092734694481, + 0.011227360926568508, + 0.012584613636136055, + 0.0028242203406989574, + 0.0034957437310367823, + 0.0021477516274899244, + 0.00251591089181602, + 0.002198841655626893, + 0.0026890102308243513, + 0.002523165661841631, + 0.005408288445323706, + 0.00891164131462574, + 0.008800352923572063, + 0.007847839966416359, + 0.006177294533699751, + 0.005949473474174738, + 0.00967353954911232, + 0.010586768388748169, + 0.012731009162962437, + 0.0022859089076519012, + 0.0033332097809761763, + 0.0031977654434740543, + 0.004089280962944031, + 0.002063505118712783, + 0.0056356461718678474, + 0.0036778494250029325, + 0.004389530513435602, + 0.008828360587358475, + 0.004135803319513798, + 0.007389469537883997, + 0.008219534531235695, + 0.00429764436557889, + 0.008390489034354687, + 0.008263343945145607, + 0.009468266740441322, + 0.005637737922370434, + 0.0030788348522037268, + 0.006702365819364786, + 0.004087198060005903, + 0.00338307349011302, + 0.003262402955442667, + 0.003983813337981701, + 0.005615563597530127, + 0.03430652990937233, + 0.014012770727276802, + 0.034609850496053696, + 0.02898222953081131, + 0.008738513104617596, + 0.011885259300470352, + 0.013383191078901291, + 0.035851094871759415, + 0.00509277218952775, + 0.004140686709433794, + 0.006487291306257248, + 0.008858725428581238, + 0.0035817213356494904, + 0.004377078264951706, + 0.006708977743983269, + 0.004838655237108469, + 0.0039604343473911285, + 0.0030004759319126606, + 0.004964736290276051, + 0.0065426575019955635, + 0.0016002198681235313, + 0.0016351784579455853, + 0.005268900655210018, + 0.004256764892488718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.05501466244459152 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598890259861946 + ], + "opposite_logprob_tokens": [ + -3.895598888397217 + ], + "logprob_reward": -0.020598890259861946, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598890259861946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4164723753929138, + 0.1498367041349411, + 0.036166638135910034, + 0.030804168432950974, + 0.06898575276136398, + 0.07536949217319489, + 0.08686935901641846, + 0.07254637777805328, + 0.014484711922705173, + 0.04846443235874176 + ], + "credit_map": [ + 0.026251573115587234, + 0.030612686648964882, + 0.034049633890390396, + 0.037540141493082047, + 0.03087194822728634, + 0.030302859842777252, + 0.12535832822322845, + 0.10148517787456512, + 0.01078711822628975, + 0.013901004567742348, + 0.020129019394516945, + 0.03970005735754967, + 0.015738822519779205, + 0.01370246522128582, + 0.015946093946695328, + 0.01993212290108204, + 0.003829610301181674, + 0.0033879897091537714, + 0.0031014017295092344, + 0.005315971095114946, + 0.0047812447883188725, + 0.004856050945818424, + 0.00577959232032299, + 0.0051147788763046265, + 0.0034401335287839174, + 0.002130018314346671, + 0.005430831573903561, + 0.0032373748254030943, + 0.002496512606739998, + 0.0032749006059020758, + 0.004671222995966673, + 0.006123174913227558, + 0.005750162526965141, + 0.009835815988481045, + 0.006737576797604561, + 0.009636623784899712, + 0.005177288316190243, + 0.013122597709298134, + 0.0090296296402812, + 0.009696057997643948, + 0.006272470578551292, + 0.0051116677932441235, + 0.009621947072446346, + 0.011596285738050938, + 0.006846691947430372, + 0.012529969215393066, + 0.013178524561226368, + 0.010211936198174953, + 0.01380227692425251, + 0.006476710084825754, + 0.014232317917048931, + 0.012493133544921875, + 0.006772591732442379, + 0.007928971201181412, + 0.012682086788117886, + 0.012481272220611572, + 0.015485808253288269, + 0.0073782773688435555, + 0.01159175205975771, + 0.010133442468941212, + 0.0060008070431649685, + 0.005862055346369743, + 0.006920631974935532, + 0.009173603728413582, + 0.0017455690540373325, + 0.001980756176635623, + 0.0022901915945112705, + 0.0036102773156017065, + 0.0008023654809221625, + 0.001059431699104607, + 0.0013860887847840786, + 0.0016100320499390364, + 0.00436990475282073, + 0.007508860435336828, + 0.007206509355455637, + 0.007959100417792797, + 0.0038846759125590324, + 0.004197928588837385, + 0.007836470380425453, + 0.005500981584191322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.020598890259861946 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a string of loose pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the pearls creamy white, almost perfectly spherical, and radiating a soft milky sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a deep black velvet surface that is plush and non-reflective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a single gentle light source from outside the camera create mirror-like highlights on the pearls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera angle taken from above and tilted slightly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708301842212677 + ], + "opposite_logprob_tokens": [ + -2.1270830631256104 + ], + "logprob_reward": -0.12708301842212677, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12708301842212677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28786560893058777, + 0.14681418240070343, + 0.025991834700107574, + 0.12194278836250305, + 0.12869133055210114, + 0.07022643834352493, + 0.06284572184085846, + 0.05138011276721954, + 0.05148439109325409, + 0.05275752767920494 + ], + "credit_map": [ + 0.033519234508275986, + 0.021776879206299782, + 0.027138905599713326, + 0.029997024685144424, + 0.022554298862814903, + 0.01697375625371933, + 0.07762283086776733, + 0.05828268826007843, + 0.014300108887255192, + 0.010195350274443626, + 0.015489297918975353, + 0.02337402105331421, + 0.013626361265778542, + 0.016254739835858345, + 0.031261760741472244, + 0.02231253869831562, + 0.0026450289878994226, + 0.002180312527343631, + 0.0026213445235043764, + 0.004932986572384834, + 0.0028122561052441597, + 0.0025847472716122866, + 0.004546918906271458, + 0.003668242134153843, + 0.012940742075443268, + 0.014725781045854092, + 0.013160860165953636, + 0.011729532852768898, + 0.009641911834478378, + 0.015846965834498405, + 0.02332606539130211, + 0.02057093009352684, + 0.021180039271712303, + 0.015162413939833641, + 0.020654788240790367, + 0.015971148386597633, + 0.00755509827286005, + 0.01284920796751976, + 0.016246184706687927, + 0.019072452560067177, + 0.012204292230308056, + 0.005917478818446398, + 0.010402585379779339, + 0.009161064401268959, + 0.006874194834381342, + 0.009026650339365005, + 0.00655942689627409, + 0.010080741718411446, + 0.011770256794989109, + 0.0047628311440348625, + 0.00994977168738842, + 0.008851077407598495, + 0.004205912351608276, + 0.004812227562069893, + 0.009941985830664635, + 0.008551656268537045, + 0.010870813392102718, + 0.004515133798122406, + 0.00899980403482914, + 0.005871254950761795, + 0.0049842349253594875, + 0.004497114103287458, + 0.004748122300952673, + 0.006893638055771589, + 0.005952125880867243, + 0.005577402655035257, + 0.008079302497208118, + 0.013192659243941307, + 0.0031592363957315683, + 0.0037914311978965998, + 0.005740334279835224, + 0.005991899874061346, + 0.005258987192064524, + 0.004889272153377533, + 0.008716171607375145, + 0.010766023769974709, + 0.0035127787850797176, + 0.00284014199860394, + 0.010230026207864285, + 0.006544128991663456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.12708301842212677 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_0/index_tensor([9416, 9416])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_0/index_tensor([9416, 9416])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..fa6fe00d936cbe20c65a23f2c4322c9f765d2e21 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_0/index_tensor([9416, 9416])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slightly low-angle, static medium shot captures a solitary, compact shrub nestled in the heart of a tranquil, rolling meadow. The entire scene is bathed in an ethereal, warm, and soft glow of golden light. The low-hanging sun casts long, soft shadows, creating a brilliant golden rim of light around the subject. In a breathtaking, timeless movement, the shrub springs to life; its woody, gnarled branches rapidly and organically stretch outward and upward, unfurling like awakening limbs. Simultaneously, from these extended branches, countless new shoots sprout, their tightly curled leaf buds explosively unfurling. The new leaves, a vibrant, almost translucent spring green, transform from a crinkled state to fully textured forms in a fluid, continuous wave of growth. Warm light filters through the newly formed crown, causing the leaves to glow from within and highlighting the delicate, downy texture on the young stems. The background remains a soft-focus tapestry of golden grass, creating an atmosphere of pure, tranquil, and powerful natural magic.", + "seed": 10361, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_0/index_9416_seed_10361_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03330610101111233, + "vlm_logprob": -0.03330610101111233, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.03330610101111233, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007729265373200178 + ], + "opposite_logprob_tokens": [ + -5.5077290534973145 + ], + "logprob_reward": -0.007729265373200178, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.007729265373200178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5138151049613953, + 0.03883414342999458, + 0.0860573947429657, + 0.07740695774555206, + 0.06812405586242676, + 0.030416345223784447, + 0.03695640340447426, + 0.04742467403411865, + 0.040423158556222916, + 0.06054176390171051 + ], + "credit_map": [ + 0.03755435347557068, + 0.1836693286895752, + 0.04828013479709625, + 0.04018788039684296, + 0.07823900878429413, + 0.03170263394713402, + 0.03782515972852707, + 0.05635662004351616, + 0.0029395977035164833, + 0.004736431874334812, + 0.005381368100643158, + 0.0059434399008750916, + 0.0033457621466368437, + 0.0044674351811409, + 0.0077796271070837975, + 0.004240480717271566, + 0.006826409604400396, + 0.028665464371442795, + 0.00836339220404625, + 0.006185057573020458, + 0.00932951457798481, + 0.006147092208266258, + 0.006169522181153297, + 0.014370941556990147, + 0.006533375941216946, + 0.02465176396071911, + 0.007281207479536533, + 0.004458229057490826, + 0.0066077811643481255, + 0.004952289164066315, + 0.004743779078125954, + 0.018178537487983704, + 0.003986733499914408, + 0.004047687631100416, + 0.02121114358305931, + 0.007039587013423443, + 0.005787879694253206, + 0.009788990020751953, + 0.0057263728231191635, + 0.010535656474530697, + 0.0029876825865358114, + 0.0028227518778294325, + 0.00399908609688282, + 0.004749780520796776, + 0.0040970551781356335, + 0.0033805235289037228, + 0.004700098652392626, + 0.0036793684121221304, + 0.004428106360137463, + 0.003554926486685872, + 0.006094153504818678, + 0.004404440056532621, + 0.003664839081466198, + 0.004673823714256287, + 0.0038548691663891077, + 0.006281247362494469, + 0.009474404156208038, + 0.003937634639441967, + 0.014615198597311974, + 0.004155399743467569, + 0.002346767345443368, + 0.004501970484852791, + 0.0022564632818102837, + 0.006136836484074593, + 0.008190348744392395, + 0.004011168610304594, + 0.010469856671988964, + 0.004897433798760176, + 0.0020779764745384455, + 0.002981751225888729, + 0.0035020404029637575, + 0.004292584024369717, + 0.013254749588668346, + 0.0031828759238123894, + 0.024516532197594643, + 0.003124418668448925, + 0.002156798029318452, + 0.002908625639975071, + 0.002396690659224987, + 0.00900107342749834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.007729265373200178 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270732581615448 + ], + "opposite_logprob_tokens": [ + -2.127073287963867 + ], + "logprob_reward": -0.1270732581615448, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270732581615448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7462327480316162, + 0.022391339763998985, + 0.027016781270503998, + 0.07034285366535187, + 0.06425028294324875, + 0.019393308088183403, + 0.008207692764699459, + 0.0038286985363811255, + 0.010614940896630287, + 0.027721382677555084 + ], + "credit_map": [ + 0.05419212207198143, + 0.26935988664627075, + 0.06178213655948639, + 0.057987239211797714, + 0.11677111685276031, + 0.04375018551945686, + 0.053503748029470444, + 0.08888627588748932, + 0.0017460405360907316, + 0.0032349745742976665, + 0.002288500079885125, + 0.0026516022626310587, + 0.0027858498506247997, + 0.0023601318243891, + 0.005332403816282749, + 0.0019918347243219614, + 0.0025258564855903387, + 0.006235501263290644, + 0.004385814536362886, + 0.002703014062717557, + 0.00291678449138999, + 0.0020312678534537554, + 0.0029376093298196793, + 0.0032809318508952856, + 0.004967768210917711, + 0.013239114545285702, + 0.015245432034134865, + 0.004065333399921656, + 0.004303113557398319, + 0.0032797842286527157, + 0.005307699553668499, + 0.019934605807065964, + 0.0041369167156517506, + 0.006080920808017254, + 0.011802897788584232, + 0.0092932665720582, + 0.008480167016386986, + 0.008528050035238266, + 0.007259821519255638, + 0.00866824109107256, + 0.0017845788970589638, + 0.0018959371373057365, + 0.0027370615862309933, + 0.0026159470435231924, + 0.002811357146129012, + 0.0022890635300427675, + 0.00220349058508873, + 0.0030558700673282146, + 0.0007349615916609764, + 0.0010418236488476396, + 0.001866508973762393, + 0.0011542182182893157, + 0.0007419225294142962, + 0.0006598286563530564, + 0.0007725327159278095, + 0.001235896022990346, + 0.0005986791220493615, + 0.00039427902083843946, + 0.0006651111762039363, + 0.0005622479948215187, + 0.00035651898360811174, + 0.0004766398051287979, + 0.00036521992296911776, + 0.0004100020742043853, + 0.0025909682735800743, + 0.000600297236815095, + 0.0035556063521653414, + 0.0008109956979751587, + 0.0003696990024764091, + 0.0010148915462195873, + 0.0004403145285323262, + 0.001232167356647551, + 0.005750247742980719, + 0.001999026397243142, + 0.011122218333184719, + 0.0015839310362935066, + 0.0008894212078303099, + 0.0012096442515030503, + 0.0010726131731644273, + 0.004094278905540705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.1270732581615448 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031948259565979242 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031948259565979242, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031948259565979242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4770996570587158, + 0.034450046718120575, + 0.06989076733589172, + 0.1398286521434784, + 0.12582767009735107, + 0.04612872377038002, + 0.026528101414442062, + 0.011083313263952732, + 0.01578998565673828, + 0.05337308347225189 + ], + "credit_map": [ + 0.03557795658707619, + 0.16931046545505524, + 0.04345094412565231, + 0.03791120648384094, + 0.06900157034397125, + 0.030432499945163727, + 0.03810689225792885, + 0.053308144211769104, + 0.0028165820986032486, + 0.0047906567342579365, + 0.003209112212061882, + 0.004099409561604261, + 0.0037616288755089045, + 0.0030778830405324697, + 0.009130910970270634, + 0.0035638611298054457, + 0.00623986404389143, + 0.02119286172091961, + 0.008497446775436401, + 0.006071742158383131, + 0.008259248919785023, + 0.005328606814146042, + 0.006781264673918486, + 0.00751973083242774, + 0.010407308116555214, + 0.03149417042732239, + 0.025449393317103386, + 0.00809770729392767, + 0.009309046901762486, + 0.00718333525583148, + 0.01053022313863039, + 0.037357479333877563, + 0.007975630462169647, + 0.009290671907365322, + 0.029736220836639404, + 0.015209547244012356, + 0.01270537544041872, + 0.022621892392635345, + 0.011801013723015785, + 0.01648731343448162, + 0.004214609507471323, + 0.004266887903213501, + 0.008368909358978271, + 0.005740273278206587, + 0.006475793197751045, + 0.005168779753148556, + 0.005142585374414921, + 0.006750886328518391, + 0.0026038081850856543, + 0.0025919820182025433, + 0.00497034564614296, + 0.0031716085504740477, + 0.002688041189685464, + 0.0031421210151165724, + 0.002693647751584649, + 0.004666547756642103, + 0.0018132434925064445, + 0.0007742468733340502, + 0.002937347162514925, + 0.001312716631218791, + 0.0007767811184749007, + 0.001196826109662652, + 0.0007959428476169705, + 0.0014762089122086763, + 0.003616764908656478, + 0.0011885567801073194, + 0.004676350392401218, + 0.0014863588148728013, + 0.0010058064945042133, + 0.0013844874920323491, + 0.0009418490226380527, + 0.0014898108784109354, + 0.01031177956610918, + 0.004422144498676062, + 0.020878495648503304, + 0.0030961991287767887, + 0.0021659766789525747, + 0.002306237816810608, + 0.0023895734921097755, + 0.007802673149853945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0031948259565979242 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254980642348528 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.005254980642348528, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254980642348528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6848195791244507, + 0.014135465957224369, + 0.021805716678500175, + 0.04607728496193886, + 0.054199956357479095, + 0.012994982302188873, + 0.014941480942070484, + 0.014121190644800663, + 0.06736055761575699, + 0.06954379379749298 + ], + "credit_map": [ + 0.049394939094781876, + 0.24763621389865875, + 0.05593010410666466, + 0.05347380414605141, + 0.10986362397670746, + 0.039445873349905014, + 0.04608754813671112, + 0.08298745006322861, + 0.001297213020734489, + 0.0017387158004567027, + 0.002911901567131281, + 0.001745274174027145, + 0.0011970307677984238, + 0.001438930630683899, + 0.0022994154132902622, + 0.0015069840010255575, + 0.0020820368081331253, + 0.0063204835169017315, + 0.0026869261637330055, + 0.001917547546327114, + 0.002804249757900834, + 0.0014567658072337508, + 0.002543109469115734, + 0.0019945972599089146, + 0.004296035040169954, + 0.015762152150273323, + 0.003126564435660839, + 0.0026745463255792856, + 0.003946500830352306, + 0.0037300733383744955, + 0.00295977178029716, + 0.009581643156707287, + 0.0037130285054445267, + 0.003019675612449646, + 0.015640635043382645, + 0.003969034180045128, + 0.003688005730509758, + 0.016149545088410378, + 0.003025318030267954, + 0.004994713235646486, + 0.0009836496319621801, + 0.00153740041423589, + 0.0020069757010787725, + 0.0016942964866757393, + 0.00224138586781919, + 0.0014553137589246035, + 0.0013906077947467566, + 0.0016853520646691322, + 0.0017311882693320513, + 0.0012236259644851089, + 0.0027321460656821728, + 0.001576526090502739, + 0.001199144870042801, + 0.002126061823219061, + 0.0014717023586854339, + 0.0028810857329517603, + 0.0021313601173460484, + 0.0024155059363693, + 0.002261959481984377, + 0.0018989290110766888, + 0.0008888913434930146, + 0.0014683250337839127, + 0.0012109801173210144, + 0.0018452390795573592, + 0.01435316726565361, + 0.0027145221829414368, + 0.02847045660018921, + 0.005515506956726313, + 0.0020421799272298813, + 0.003566123079508543, + 0.0021396237425506115, + 0.008558976463973522, + 0.013150578364729881, + 0.005927796475589275, + 0.02648843266069889, + 0.004418959375470877, + 0.0032807434909045696, + 0.0032712994143366814, + 0.0033665229566395283, + 0.009639464318752289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.005254980642348528 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327817492187023 + ], + "opposite_logprob_tokens": [ + -3.77327823638916 + ], + "logprob_reward": -0.02327817492187023, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327817492187023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6192311644554138, + 0.045028187334537506, + 0.01989429071545601, + 0.009712942875921726, + 0.09982822090387344, + 0.03216879814863205, + 0.0113958939909935, + 0.031029926612973213, + 0.11942438036203384, + 0.012286223471164703 + ], + "credit_map": [ + 0.04456573352217674, + 0.22855593264102936, + 0.051890239119529724, + 0.047416478395462036, + 0.10020684450864792, + 0.03604090213775635, + 0.04027486592531204, + 0.07028012722730637, + 0.004704127553850412, + 0.01123624574393034, + 0.007172019220888615, + 0.0033809144515544176, + 0.004949313122779131, + 0.0035705515183508396, + 0.004759098403155804, + 0.005255911499261856, + 0.002034007105976343, + 0.004210605286061764, + 0.003412725869566202, + 0.0024352187756448984, + 0.0016370315570384264, + 0.0015995537396520376, + 0.002564294496551156, + 0.002000850159674883, + 0.0012505445629358292, + 0.0014810683205723763, + 0.0016187640139833093, + 0.001058763824403286, + 0.0006078721489757299, + 0.0010725955944508314, + 0.0009673897293396294, + 0.001655943924561143, + 0.009848948568105698, + 0.007833349518477917, + 0.026061946526169777, + 0.004817708395421505, + 0.005127939861267805, + 0.03497748076915741, + 0.003673662431538105, + 0.007487166672945023, + 0.0036391683388501406, + 0.002642922569066286, + 0.008832667954266071, + 0.0037019061855971813, + 0.0032032974995672703, + 0.0026007143314927816, + 0.001761905150488019, + 0.005786212161183357, + 0.0010084904497489333, + 0.001300548785366118, + 0.0017526294104754925, + 0.0016333609819412231, + 0.0010481075150892138, + 0.0012106894282624125, + 0.0013637340161949396, + 0.002078332006931305, + 0.003700156696140766, + 0.004322877619415522, + 0.003781060455366969, + 0.0037388689815998077, + 0.002355464966967702, + 0.004656343255192041, + 0.0025826103519648314, + 0.005892540793865919, + 0.01860218122601509, + 0.0055558886379003525, + 0.048762209713459015, + 0.009597163647413254, + 0.004484502132982016, + 0.009953614324331284, + 0.004188537131994963, + 0.01828026957809925, + 0.0010870099067687988, + 0.0018116974970325828, + 0.0014991137431934476, + 0.0014596483670175076, + 0.0016195341013371944, + 0.0012146196095272899, + 0.0014818321214988828, + 0.0021127674262970686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.02327817492187023 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle, static medium shot captures a solitary, compact shrub nestled in the heart of a tranquil, rolling meadow. The entire scene is bathed in an ethereal, warm, and soft glow of golden light. The low-hanging sun casts long, soft shadows, creating a brilliant golden rim of light around the subject. In a breathtaking, timeless movement, the shrub springs to life; its woody, gnarled branches rapidly and organically stretch outward and upward, unfurling like awakening limbs. Simultaneously, from these extended branches, countless new shoots sprout, their tightly curled leaf buds explosively unfurling. The new leaves, a vibrant, almost translucent spring green, transform from a crinkled state to fully textured forms in a fluid, continuous wave of growth. Warm light filters through the newly formed crown, causing the leaves to glow from within and highlighting the delicate, downy texture on the young stems. The background remains a soft-focus tapestry of golden grass, creating an atmosphere of pure, tranquil, and powerful natural magic.", + "seed": 10362, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_0/index_9416_seed_10362_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07004564041271806, + "vlm_logprob": -0.07004564041271806, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.07004564041271806, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009246266447007656 + ], + "opposite_logprob_tokens": [ + -5.009246349334717 + ], + "logprob_reward": -0.009246266447007656, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009246266447007656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24009300768375397, + 0.1600697934627533, + 0.042855799198150635, + 0.05219965800642967, + 0.13542361557483673, + 0.04751474782824516, + 0.07304076105356216, + 0.07896696031093597, + 0.08848114311695099, + 0.08135458081960678 + ], + "credit_map": [ + 0.015345294959843159, + 0.02393626607954502, + 0.02442028746008873, + 0.04816211760044098, + 0.029001174494624138, + 0.018428146839141846, + 0.03795656934380531, + 0.04284314811229706, + 0.006729912478476763, + 0.013124844990670681, + 0.033400628715753555, + 0.021257223561406136, + 0.019322168081998825, + 0.03127995878458023, + 0.019762514159083366, + 0.015192554332315922, + 0.003419547574594617, + 0.0051598455756902695, + 0.0077110780403018, + 0.005500265397131443, + 0.0046388753689825535, + 0.004459742456674576, + 0.004913613200187683, + 0.007052835077047348, + 0.004903695546090603, + 0.0047462438233196735, + 0.00667366711422801, + 0.007097285240888596, + 0.008075794205069542, + 0.006709001958370209, + 0.008392691612243652, + 0.00560128316283226, + 0.013391021639108658, + 0.0190264955163002, + 0.015021173283457756, + 0.01637278124690056, + 0.02183021418750286, + 0.01839650794863701, + 0.019475936889648438, + 0.011909480206668377, + 0.005419957917183638, + 0.005471760872751474, + 0.005613055545836687, + 0.005490593146532774, + 0.006691585294902325, + 0.004354695789515972, + 0.009893950074911118, + 0.004579148720949888, + 0.007996918633580208, + 0.008623619563877583, + 0.007589044980704784, + 0.008279136382043362, + 0.00838435348123312, + 0.007423142436891794, + 0.010015251114964485, + 0.01472929771989584, + 0.013113021850585938, + 0.011945392936468124, + 0.009301590733230114, + 0.007989241741597652, + 0.008182642981410027, + 0.0069274017587304115, + 0.005957707762718201, + 0.01554996520280838, + 0.008811983279883862, + 0.011703570373356342, + 0.010472699999809265, + 0.0077059268951416016, + 0.016209430992603302, + 0.008092978969216347, + 0.00878366082906723, + 0.01670088805258274, + 0.0065141841769218445, + 0.010056349448859692, + 0.008852433413267136, + 0.009998694062232971, + 0.012710275128483772, + 0.010400880128145218, + 0.011822417378425598, + 0.010999344289302826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.009246266447007656 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134462237358093 + ], + "opposite_logprob_tokens": [ + -1.313446283340454 + ], + "logprob_reward": -0.3134462237358093, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134462237358093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.061376895755529404, + 0.229094997048378, + 0.10024895519018173, + 0.10039392858743668, + 0.044105712324380875, + 0.08609151840209961, + 0.028313666582107544, + 0.2627289295196533, + 0.04605097696185112, + 0.041594382375478745 + ], + "credit_map": [ + 0.005253245588392019, + 0.007097610272467136, + 0.013211716897785664, + 0.006923353299498558, + 0.006469955202192068, + 0.00638140132650733, + 0.008497301489114761, + 0.007542319595813751, + 0.007852312177419662, + 0.009806654416024685, + 0.07292770594358444, + 0.020739153027534485, + 0.022414157167077065, + 0.06439653784036636, + 0.012037986889481544, + 0.018920497968792915, + 0.006290351506322622, + 0.013296360149979591, + 0.015050502493977547, + 0.016482848674058914, + 0.008363572880625725, + 0.007229622919112444, + 0.012824990786612034, + 0.020710717886686325, + 0.00743479048833251, + 0.007382986601442099, + 0.017076078802347183, + 0.015031897462904453, + 0.015930557623505592, + 0.010117030702531338, + 0.01600559800863266, + 0.01141499262303114, + 0.003113530343398452, + 0.004858237225562334, + 0.004644354805350304, + 0.006302847992628813, + 0.007312345318496227, + 0.004011751152575016, + 0.007757118437439203, + 0.006105531007051468, + 0.008946198970079422, + 0.01656392030417919, + 0.008862581104040146, + 0.012690491043031216, + 0.011238943785429, + 0.006331657525151968, + 0.009821516461670399, + 0.01163622085005045, + 0.002173604676499963, + 0.003815094009041786, + 0.0030072680674493313, + 0.0032506969291716814, + 0.004186119418591261, + 0.003134530270472169, + 0.004302540794014931, + 0.004443816840648651, + 0.03161926195025444, + 0.03455108031630516, + 0.031022915616631508, + 0.025571361184120178, + 0.02906162478029728, + 0.026333317160606384, + 0.017920706421136856, + 0.06664867699146271, + 0.005411916878074408, + 0.00521685229614377, + 0.005063476972281933, + 0.004679484758526087, + 0.007472219876945019, + 0.0052396804094314575, + 0.004300167318433523, + 0.008667181245982647, + 0.0035819781478494406, + 0.005315656308084726, + 0.0041082194074988365, + 0.005110093858093023, + 0.00683900760486722, + 0.007236754521727562, + 0.004628568422049284, + 0.004774107597768307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.3134462237358093 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256047938019037 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256047938019037, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256047938019037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07679888606071472, + 0.20320983231067657, + 0.14830325543880463, + 0.11329993605613708, + 0.028378624469041824, + 0.06714143604040146, + 0.028612034395337105, + 0.24777856469154358, + 0.051567405462265015, + 0.034909967333078384 + ], + "credit_map": [ + 0.006152798887342215, + 0.009373248554766178, + 0.017491010949015617, + 0.00750390300527215, + 0.0072107878513634205, + 0.009617730975151062, + 0.01030154898762703, + 0.009147848933935165, + 0.006123142316937447, + 0.0074514709413051605, + 0.06659132987260818, + 0.017707202583551407, + 0.01809241622686386, + 0.060376062989234924, + 0.011108146980404854, + 0.01576005294919014, + 0.00903195608407259, + 0.02006516605615616, + 0.02124299295246601, + 0.02469143643975258, + 0.013136227615177631, + 0.010850679129362106, + 0.01865471340715885, + 0.030630087479948997, + 0.008055412210524082, + 0.008313736878335476, + 0.01866300404071808, + 0.01733105070888996, + 0.01796630024909973, + 0.011729883961379528, + 0.018093887716531754, + 0.013146664947271347, + 0.002662540413439274, + 0.004135543014854193, + 0.0030536858830600977, + 0.003293039742857218, + 0.0036611328832805157, + 0.002829977311193943, + 0.004977898672223091, + 0.003764807479456067, + 0.008885416202247143, + 0.014365780167281628, + 0.006668774876743555, + 0.009084584191441536, + 0.00868682935833931, + 0.004849047400057316, + 0.007240363862365484, + 0.007360643241554499, + 0.0025595580227673054, + 0.0042564221657812595, + 0.004228418227285147, + 0.003430702956393361, + 0.005900757387280464, + 0.0033812737092375755, + 0.0024230859708040953, + 0.00243181805126369, + 0.029111044481396675, + 0.03243226930499077, + 0.02942669950425625, + 0.023929210379719734, + 0.027473870664834976, + 0.02514078840613365, + 0.017259959131479263, + 0.06300473213195801, + 0.007141551002860069, + 0.005592214409261942, + 0.005375159438699484, + 0.0052374075166881084, + 0.008282681927084923, + 0.006038731895387173, + 0.004789239261299372, + 0.009110423736274242, + 0.0033684750087559223, + 0.0043658483773469925, + 0.0033802951220422983, + 0.004348199814558029, + 0.004965223837643862, + 0.005085581913590431, + 0.004405742511153221, + 0.004990601446479559 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 2 + ], + "selected_reward": -0.005256047938019037 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0040968069806694984 + ], + "opposite_logprob_tokens": [ + -5.504096984863281 + ], + "logprob_reward": -0.0040968069806694984, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0040968069806694984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05917655676603317, + 0.12330227345228195, + 0.02189188450574875, + 0.046517275273799896, + 0.042625464498996735, + 0.026762472465634346, + 0.04253153130412102, + 0.5099031329154968, + 0.0858437567949295, + 0.041445598006248474 + ], + "credit_map": [ + 0.004835665691643953, + 0.006143629550933838, + 0.007057785987854004, + 0.00933875422924757, + 0.006800688803195953, + 0.00828890223056078, + 0.009468874894082546, + 0.007242261432111263, + 0.0046518356539309025, + 0.007234722841531038, + 0.039634380489587784, + 0.009839929640293121, + 0.01114717312157154, + 0.03974588215351105, + 0.004967607092112303, + 0.006080755032598972, + 0.0022059609182178974, + 0.003551783971488476, + 0.002018958330154419, + 0.0021461693104356527, + 0.003242264036089182, + 0.0021135832648724318, + 0.002667725086212158, + 0.003945442382246256, + 0.003277442418038845, + 0.006599947344511747, + 0.005254540126770735, + 0.0050560771487653255, + 0.011313478462398052, + 0.004569550044834614, + 0.006861069705337286, + 0.003585175611078739, + 0.0039172121323645115, + 0.005930124316364527, + 0.004478342365473509, + 0.005489646457135677, + 0.006446440704166889, + 0.005347339902073145, + 0.0058782463893294334, + 0.005138115491718054, + 0.0022169474977999926, + 0.004291040822863579, + 0.0038649134803563356, + 0.004017963539808989, + 0.003315690206363797, + 0.002383994637057185, + 0.0029209302738308907, + 0.0037509927060455084, + 0.005604088772088289, + 0.005046130623668432, + 0.007459735497832298, + 0.004776997957378626, + 0.007364162243902683, + 0.004364505875855684, + 0.0036626493092626333, + 0.00425326731055975, + 0.05971239507198334, + 0.06476598232984543, + 0.06120672821998596, + 0.050179604440927505, + 0.05398646369576454, + 0.04979638382792473, + 0.03458080068230629, + 0.13567481935024261, + 0.018344761803746223, + 0.011694508604705334, + 0.010154234245419502, + 0.00908757746219635, + 0.009258553385734558, + 0.011239728890359402, + 0.006311470177024603, + 0.009752933867275715, + 0.00746419932693243, + 0.005808490794152021, + 0.004295650869607925, + 0.005037633236497641, + 0.005425723269581795, + 0.00436934269964695, + 0.0036391348112374544, + 0.005405424162745476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.0040968069806694984 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818285696208477 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.01818285696208477, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818285696208477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12738414108753204, + 0.3562805950641632, + 0.07048667222261429, + 0.08843611180782318, + 0.09646888077259064, + 0.04342827945947647, + 0.045031361281871796, + 0.07708507776260376, + 0.05739519000053406, + 0.038003675639629364 + ], + "credit_map": [ + 0.009248633868992329, + 0.010660774074494839, + 0.015751035884022713, + 0.021156679838895798, + 0.014258953742682934, + 0.01020137034356594, + 0.025461511686444283, + 0.020645182579755783, + 0.012427588924765587, + 0.01766812987625599, + 0.11777753382921219, + 0.027073489502072334, + 0.03144393861293793, + 0.11537633091211319, + 0.013959201984107494, + 0.02055438794195652, + 0.005053206346929073, + 0.009509378112852573, + 0.010265199467539787, + 0.011220769956707954, + 0.006261180620640516, + 0.005806589033454657, + 0.006832791492342949, + 0.015537561848759651, + 0.005954473279416561, + 0.006032555364072323, + 0.011011937633156776, + 0.015472040511667728, + 0.012872208841145039, + 0.00839230790734291, + 0.016070421785116196, + 0.012630166485905647, + 0.0058653089217841625, + 0.0134358499199152, + 0.009805218316614628, + 0.013372156769037247, + 0.019656194373965263, + 0.008415388874709606, + 0.013009564019739628, + 0.012909202836453915, + 0.007353311404585838, + 0.009894304908812046, + 0.005078558344393969, + 0.00510371383279562, + 0.005231287330389023, + 0.0025976859033107758, + 0.0036363364197313786, + 0.004533080384135246, + 0.004113953094929457, + 0.00496372114866972, + 0.004665672779083252, + 0.0056103612296283245, + 0.006937636062502861, + 0.004617676604539156, + 0.007349695079028606, + 0.006772646214812994, + 0.011085067875683308, + 0.009335648268461227, + 0.008176400326192379, + 0.007850785739719868, + 0.009859082289040089, + 0.008331849239766598, + 0.005870338995009661, + 0.016575908288359642, + 0.007007807493209839, + 0.00810894463211298, + 0.00631949445232749, + 0.005445267539471388, + 0.010633895173668861, + 0.0051599666476249695, + 0.0065909866243600845, + 0.00812882836908102, + 0.005325512494891882, + 0.005104111507534981, + 0.005067693069577217, + 0.004125547595322132, + 0.007364781107753515, + 0.004797611851245165, + 0.003006233600899577, + 0.0032121825497597456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.01818285696208477 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle, static medium shot captures a solitary, compact shrub nestled in the heart of a tranquil, rolling meadow. The entire scene is bathed in an ethereal, warm, and soft glow of golden light. The low-hanging sun casts long, soft shadows, creating a brilliant golden rim of light around the subject. In a breathtaking, timeless movement, the shrub springs to life; its woody, gnarled branches rapidly and organically stretch outward and upward, unfurling like awakening limbs. Simultaneously, from these extended branches, countless new shoots sprout, their tightly curled leaf buds explosively unfurling. The new leaves, a vibrant, almost translucent spring green, transform from a crinkled state to fully textured forms in a fluid, continuous wave of growth. Warm light filters through the newly formed crown, causing the leaves to glow from within and highlighting the delicate, downy texture on the young stems. The background remains a soft-focus tapestry of golden grass, creating an atmosphere of pure, tranquil, and powerful natural magic.", + "seed": 10363, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_0/index_9416_seed_10363_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.016800403641536833, + "vlm_logprob": -0.016800403641536833, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.016800403641536833, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008456375449895859 + ], + "opposite_logprob_tokens": [ + -5.133456230163574 + ], + "logprob_reward": -0.008456375449895859, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.008456375449895859, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11144021898508072, + 0.4097062349319458, + 0.15502765774726868, + 0.05274598300457001, + 0.04040553420782089, + 0.042645759880542755, + 0.041743308305740356, + 0.02740021049976349, + 0.019038081169128418, + 0.09984695911407471 + ], + "credit_map": [ + 0.01084351260215044, + 0.01363063883036375, + 0.012434926815330982, + 0.01339871808886528, + 0.0186687670648098, + 0.017328474670648575, + 0.010753383859992027, + 0.01438179425895214, + 0.06845330446958542, + 0.16642233729362488, + 0.010785553604364395, + 0.021884290501475334, + 0.07518154382705688, + 0.02601652592420578, + 0.02666996419429779, + 0.014292711392045021, + 0.0172472782433033, + 0.03846533223986626, + 0.010502012446522713, + 0.011717253364622593, + 0.04166843742132187, + 0.01234670914709568, + 0.011708837933838367, + 0.011371797882020473, + 0.0055275787599384785, + 0.0064435480162501335, + 0.007262866944074631, + 0.007309554144740105, + 0.005689568351954222, + 0.010365202091634274, + 0.005290327128022909, + 0.004857341758906841, + 0.006209908053278923, + 0.0040877568535506725, + 0.009956142865121365, + 0.005058944225311279, + 0.003422878449782729, + 0.0035117247607558966, + 0.0038191855419427156, + 0.00433899462223053, + 0.006594433449208736, + 0.006296908017247915, + 0.005026911851018667, + 0.007516325451433659, + 0.004373208619654179, + 0.004032605793327093, + 0.005053849425166845, + 0.00375151913613081, + 0.004916743375360966, + 0.006063397973775864, + 0.007859473116695881, + 0.004828210920095444, + 0.004862969741225243, + 0.0040721287950873375, + 0.004855939652770758, + 0.004284446127712727, + 0.004316928330808878, + 0.002418555086478591, + 0.004172959830611944, + 0.0025987024419009686, + 0.0027987917419523, + 0.0030730378348380327, + 0.0032078318763524294, + 0.004813402891159058, + 0.0022443600464612246, + 0.0022113865707069635, + 0.0038051374722272158, + 0.0017029130831360817, + 0.002313076751306653, + 0.0019672736525535583, + 0.0018542491598054767, + 0.002939684083685279, + 0.025543007999658585, + 0.018126443028450012, + 0.014496286399662495, + 0.012699015438556671, + 0.008316257037222385, + 0.006528831087052822, + 0.008408105932176113, + 0.005729012191295624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.008456375449895859 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549958199262619 + ], + "opposite_logprob_tokens": [ + -2.9299957752227783 + ], + "logprob_reward": -0.0549958199262619, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549958199262619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06175258010625839, + 0.6113805770874023, + 0.1718989759683609, + 0.057667866349220276, + 0.043517593294382095, + 0.01364195253700018, + 0.010487417690455914, + 0.007800225168466568, + 0.009841828607022762, + 0.012011094950139523 + ], + "credit_map": [ + 0.004427041858434677, + 0.005158782005310059, + 0.006865261122584343, + 0.014033031649887562, + 0.007833143696188927, + 0.006350801791995764, + 0.010834425687789917, + 0.006250086706131697, + 0.0998615175485611, + 0.24219557642936707, + 0.019469385966658592, + 0.03488878160715103, + 0.10679955780506134, + 0.04045301675796509, + 0.037242449820041656, + 0.030470268800854683, + 0.03256632015109062, + 0.02727046236395836, + 0.012079759500920773, + 0.011068258434534073, + 0.03161619231104851, + 0.03806518390774727, + 0.0098771583288908, + 0.009355630725622177, + 0.01317223347723484, + 0.00438957754522562, + 0.0034574021119624376, + 0.005182613153010607, + 0.005011847708374262, + 0.01949176751077175, + 0.003441444132477045, + 0.003520979778841138, + 0.005818271543830633, + 0.005717730149626732, + 0.0067735705524683, + 0.008814726956188679, + 0.004102832637727261, + 0.0034154353197664022, + 0.005127353593707085, + 0.0037476751022040844, + 0.0024385990109294653, + 0.0016754369717091322, + 0.0020457289647310972, + 0.0016161759849637747, + 0.001307871425524354, + 0.0014617980923503637, + 0.0013698128750547767, + 0.0017265294445678592, + 0.0012571178376674652, + 0.0018795095384120941, + 0.001768198562785983, + 0.001399754430167377, + 0.0014031610917299986, + 0.0011410049628466368, + 0.0008166615152731538, + 0.0008220094023272395, + 0.001331005827523768, + 0.0013670785119757056, + 0.0012046530609950423, + 0.0007919956115074456, + 0.0008545544696971774, + 0.0006504320190288126, + 0.0007008218090049922, + 0.0008996835676953197, + 0.0013364845653995872, + 0.0016492052236571908, + 0.0017621746519580483, + 0.0008533962536603212, + 0.0013447999954223633, + 0.0009033425594680011, + 0.0008948732865974307, + 0.001097551779821515, + 0.0011708326637744904, + 0.0009523642947897315, + 0.0022498960606753826, + 0.002493469975888729, + 0.0014177093980833888, + 0.0008228959050029516, + 0.0015084970509633422, + 0.0013954296009615064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0549958199262619 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001718117156997323 + ], + "opposite_logprob_tokens": [ + -6.376718044281006 + ], + "logprob_reward": -0.001718117156997323, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001718117156997323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10413112491369247, + 0.5030572414398193, + 0.16935990750789642, + 0.11088145524263382, + 0.04763743653893471, + 0.011920712888240814, + 0.012949264608323574, + 0.007681386079639196, + 0.009096418507397175, + 0.02328512631356716 + ], + "credit_map": [ + 0.009105279110372066, + 0.011336412280797958, + 0.010560350492596626, + 0.02174224704504013, + 0.018612554296851158, + 0.0077325874008238316, + 0.01471288874745369, + 0.01032879389822483, + 0.08152342587709427, + 0.19656121730804443, + 0.01575310528278351, + 0.029736943542957306, + 0.08746495097875595, + 0.0329558365046978, + 0.03319622203707695, + 0.02586548775434494, + 0.03540150076150894, + 0.02181985042989254, + 0.012699509970843792, + 0.011638405732810497, + 0.025592871010303497, + 0.04231791943311691, + 0.00975146796554327, + 0.0101383738219738, + 0.026629524305462837, + 0.00801596324890852, + 0.0059992787428200245, + 0.010642298497259617, + 0.008596445433795452, + 0.03653627634048462, + 0.007412747945636511, + 0.007048917934298515, + 0.006272392347455025, + 0.006116410717368126, + 0.007654187269508839, + 0.009285244159400463, + 0.004596553277224302, + 0.003788694040849805, + 0.00555173447355628, + 0.004372218623757362, + 0.0015411254717037082, + 0.0017948643071576953, + 0.0016696968814358115, + 0.0020293144043534994, + 0.0014538108371198177, + 0.0011671200627461076, + 0.001201531384140253, + 0.0010632490739226341, + 0.0016509321285411716, + 0.002435419475659728, + 0.002170019783079624, + 0.0014849472790956497, + 0.0016531021101400256, + 0.001445530797354877, + 0.0009672780870459974, + 0.0011420345399528742, + 0.0011412955354899168, + 0.0008910904871299863, + 0.0013764804461970925, + 0.0006881908630020916, + 0.0007988663855940104, + 0.0008710786933079362, + 0.0005804667598567903, + 0.0013339163269847631, + 0.0010649556061252952, + 0.0019852404948323965, + 0.001076759654097259, + 0.0007710130303166807, + 0.001261823228560388, + 0.0011128217447549105, + 0.0008444407139904797, + 0.0009793636854737997, + 0.007482906803488731, + 0.004278984386473894, + 0.003428388386964798, + 0.001856412272900343, + 0.0012295276392251253, + 0.00192639313172549, + 0.0010475001763552427, + 0.0020350127015262842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.001718117156997323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640049301087856 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.004640049301087856, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640049301087856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09677162766456604, + 0.6021865606307983, + 0.14501354098320007, + 0.04110226407647133, + 0.04403563216328621, + 0.017152445390820503, + 0.014156991615891457, + 0.011888392269611359, + 0.012066109105944633, + 0.015626447275280952 + ], + "credit_map": [ + 0.007544711697846651, + 0.011704663746058941, + 0.01480735745280981, + 0.013242490589618683, + 0.011268764734268188, + 0.01154311653226614, + 0.01135634072124958, + 0.015304182656109333, + 0.09709589183330536, + 0.23416845500469208, + 0.022144101560115814, + 0.034539591521024704, + 0.10229262709617615, + 0.04008212685585022, + 0.03683008253574371, + 0.0350336953997612, + 0.03166899457573891, + 0.02217557467520237, + 0.008133694529533386, + 0.00856915395706892, + 0.025690270587801933, + 0.03513467684388161, + 0.007776081096380949, + 0.005865097977221012, + 0.0075471471063792706, + 0.0033914661034941673, + 0.004820023197680712, + 0.002635770244523883, + 0.0034061770420521498, + 0.013979597948491573, + 0.0018209486734122038, + 0.003501135390251875, + 0.004992518108338118, + 0.004416295327246189, + 0.00769314169883728, + 0.008994740433990955, + 0.0038522756658494473, + 0.004408894572407007, + 0.005919347982853651, + 0.0037584155797958374, + 0.0027147987857460976, + 0.002288719406351447, + 0.0029764326754957438, + 0.0015748636797070503, + 0.0018718937644734979, + 0.0015377126401290298, + 0.0015198702458292246, + 0.002668153727427125, + 0.0019157335627824068, + 0.0011618908029049635, + 0.0027976834680885077, + 0.0016905685188248754, + 0.0014878128422424197, + 0.0016242486890405416, + 0.0013267815811559558, + 0.0021522725000977516, + 0.0018440056592226028, + 0.0009361208649352193, + 0.0015403684228658676, + 0.0009675931069068611, + 0.0015466333134099841, + 0.002591115888208151, + 0.0008441141690127552, + 0.0016184408450499177, + 0.0015336323995143175, + 0.0021464156452566385, + 0.0015548929804936051, + 0.0014092614874243736, + 0.0015260193031281233, + 0.0013572841417044401, + 0.0011825455585494637, + 0.0013560570077970624, + 0.0019070882117375731, + 0.0017528991447761655, + 0.0022878979798406363, + 0.0019501190399751067, + 0.0017265441128984094, + 0.002027682727202773, + 0.0021882392466068268, + 0.0017859757645055652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.004640049301087856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419165637344122 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.01419165637344122, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419165637344122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06391556560993195, + 0.676827073097229, + 0.0410773791372776, + 0.10830050706863403, + 0.016428690403699875, + 0.011705519631505013, + 0.0072122784331440926, + 0.008145898580551147, + 0.022992542013525963, + 0.04339461773633957 + ], + "credit_map": [ + 0.004788595251739025, + 0.00621230760589242, + 0.0070447660982608795, + 0.012869935482740402, + 0.010806036181747913, + 0.008246962912380695, + 0.006814612541347742, + 0.007132344413548708, + 0.10672859102487564, + 0.25690174102783203, + 0.03080984763801098, + 0.041539739817380905, + 0.11051509529352188, + 0.04553989693522453, + 0.04031427204608917, + 0.04447781294584274, + 0.004979991819709539, + 0.006217545364052057, + 0.004959991667419672, + 0.0041654412634670734, + 0.006782678887248039, + 0.004776055924594402, + 0.003189018229022622, + 0.006006652023643255, + 0.025627994909882545, + 0.008016069419682026, + 0.008628284558653831, + 0.004140299744904041, + 0.009026959538459778, + 0.04271985962986946, + 0.0036110514774918556, + 0.006529978476464748, + 0.002133969683200121, + 0.0016308274352923036, + 0.0032385308295488358, + 0.0026335064321756363, + 0.0013512149453163147, + 0.0015216211322695017, + 0.002137115690857172, + 0.0017819023923948407, + 0.0017994111403822899, + 0.0013746751938015223, + 0.0022165141999721527, + 0.0016695036320015788, + 0.0011992508079856634, + 0.0011288399109616876, + 0.0010374329285696149, + 0.0012798903044313192, + 0.0009352806373499334, + 0.0008282935596071184, + 0.0012387688038870692, + 0.0008641497115604579, + 0.0008860906818881631, + 0.0008438091026619077, + 0.000648476358037442, + 0.0009674089378677309, + 0.0012047915952280164, + 0.0006279503577388823, + 0.0013129013823345304, + 0.0007180059910751879, + 0.000983571633696556, + 0.0013883033534511924, + 0.000578469829633832, + 0.0013319030404090881, + 0.0048590125516057014, + 0.001986660063266754, + 0.004490514285862446, + 0.0023075502831488848, + 0.0019618943333625793, + 0.0031308268662542105, + 0.001954940613359213, + 0.0023011411540210247, + 0.010813027620315552, + 0.006347613409161568, + 0.008495194837450981, + 0.004787756595760584, + 0.0034153321757912636, + 0.004304373636841774, + 0.0026396233588457108, + 0.002591690979897976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.01419165637344122 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slightly low-angle, static medium shot captures a solitary, compact shrub nestled in the heart of a tranquil, rolling meadow. The entire scene is bathed in an ethereal, warm, and soft glow of golden light. The low-hanging sun casts long, soft shadows, creating a brilliant golden rim of light around the subject. In a breathtaking, timeless movement, the shrub springs to life; its woody, gnarled branches rapidly and organically stretch outward and upward, unfurling like awakening limbs. Simultaneously, from these extended branches, countless new shoots sprout, their tightly curled leaf buds explosively unfurling. The new leaves, a vibrant, almost translucent spring green, transform from a crinkled state to fully textured forms in a fluid, continuous wave of growth. Warm light filters through the newly formed crown, causing the leaves to glow from within and highlighting the delicate, downy texture on the young stems. The background remains a soft-focus tapestry of golden grass, creating an atmosphere of pure, tranquil, and powerful natural magic.", + "seed": 10364, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_0/index_9416_seed_10364_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.025502901896834372, + "vlm_logprob": -0.025502901896834372, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.025502901896834372, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008948815986514091 + ], + "opposite_logprob_tokens": [ + -5.008948802947998 + ], + "logprob_reward": -0.008948815986514091, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008948815986514091, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05508136376738548, + 0.051164183765649796, + 0.04188060387969017, + 0.10072489827871323, + 0.031713224947452545, + 0.028599951416254044, + 0.07347110658884048, + 0.31147444248199463, + 0.2558537721633911, + 0.050036437809467316 + ], + "credit_map": [ + 0.004829412326216698, + 0.008519443683326244, + 0.007155194412916899, + 0.005943036638200283, + 0.007211638148874044, + 0.009978271089494228, + 0.0049935453571379185, + 0.006450818385928869, + 0.004444679245352745, + 0.003925632685422897, + 0.008460303768515587, + 0.006773805245757103, + 0.005727941635996103, + 0.008022353053092957, + 0.007741080597043037, + 0.006068387534469366, + 0.002883558627218008, + 0.012289286591112614, + 0.004350228235125542, + 0.0027374601922929287, + 0.0035287351347506046, + 0.010220901109278202, + 0.003314690198749304, + 0.0025557412300258875, + 0.004730751737952232, + 0.04147585853934288, + 0.0037787356413900852, + 0.0044850935228168964, + 0.008189328014850616, + 0.03065187856554985, + 0.003512422554194927, + 0.0039008271414786577, + 0.0027580875903367996, + 0.005475161597132683, + 0.0045669605024158955, + 0.002492576139047742, + 0.0030721372459083796, + 0.0073473891243338585, + 0.0019499188056215644, + 0.004050991963595152, + 0.0029015426989644766, + 0.0033693870063871145, + 0.0027851799968630075, + 0.0038060543593019247, + 0.003297387855127454, + 0.004494985565543175, + 0.002714460249990225, + 0.005230950657278299, + 0.008088630624115467, + 0.0076364814303815365, + 0.011995507404208183, + 0.005330212414264679, + 0.006300599779933691, + 0.00867445208132267, + 0.00448165088891983, + 0.020963573828339577, + 0.04173062741756439, + 0.03077881596982479, + 0.04689820483326912, + 0.025280790403485298, + 0.020959803834557533, + 0.03472297266125679, + 0.018882445991039276, + 0.09222079813480377, + 0.021930508315563202, + 0.09464951604604721, + 0.07371077686548233, + 0.010840605944395065, + 0.013483881950378418, + 0.01276722177863121, + 0.005429060664027929, + 0.023042215034365654, + 0.005760828964412212, + 0.009165022522211075, + 0.003850366221740842, + 0.0061621530912816525, + 0.005162947811186314, + 0.009525572881102562, + 0.004380603786557913, + 0.006028944626450539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.008948815986514091 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905375003814697 + ], + "opposite_logprob_tokens": [ + -2.4640536308288574 + ], + "logprob_reward": -0.08905375003814697, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905375003814697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.033069759607315063, + 0.13188722729682922, + 0.06740260124206543, + 0.24354137480258942, + 0.029269374907016754, + 0.030029354616999626, + 0.0394584983587265, + 0.06864945590496063, + 0.3398020565509796, + 0.01689027063548565 + ], + "credit_map": [ + 0.002536698477342725, + 0.003906899131834507, + 0.003911109641194344, + 0.003591683693230152, + 0.006010163575410843, + 0.00478286063298583, + 0.0026211084332317114, + 0.0057092406786978245, + 0.005420435220003128, + 0.029928436502814293, + 0.006113238632678986, + 0.010015523992478848, + 0.04615218937397003, + 0.015037233941257, + 0.011499174870550632, + 0.00772100780159235, + 0.004534281324595213, + 0.012979122810065746, + 0.010422280989587307, + 0.005166599992662668, + 0.006070317234843969, + 0.015540377236902714, + 0.00627801101654768, + 0.00641161622479558, + 0.012262708507478237, + 0.10272058099508286, + 0.010200039483606815, + 0.008105066604912281, + 0.01658092997968197, + 0.08229507505893707, + 0.00412833783775568, + 0.007248660083860159, + 0.003133614780381322, + 0.003652733750641346, + 0.00429402245208621, + 0.001797012402676046, + 0.0023802719078958035, + 0.007212270982563496, + 0.0021791665349155664, + 0.004620287101715803, + 0.0018060451839119196, + 0.007668131496757269, + 0.0025293261278420687, + 0.00245417351834476, + 0.003142185742035508, + 0.003665587631985545, + 0.002321957377716899, + 0.006441951263695955, + 0.002555253216996789, + 0.014076386578381062, + 0.0029166920576244593, + 0.0020824121311306953, + 0.0028585500549525023, + 0.003959826659411192, + 0.0015006106114014983, + 0.009508772753179073, + 0.00844805222004652, + 0.008482896722853184, + 0.0105737941339612, + 0.00592573219910264, + 0.006043849512934685, + 0.007579127326607704, + 0.004040672909468412, + 0.017555341124534607, + 0.024899624288082123, + 0.15097707509994507, + 0.06958582252264023, + 0.013321622274816036, + 0.023310167714953423, + 0.01845185086131096, + 0.0063781957142055035, + 0.03287774324417114, + 0.0020810626447200775, + 0.0034342692233622074, + 0.0020464931149035692, + 0.001169852213934064, + 0.0016787780914455652, + 0.0030314151663333178, + 0.0012166162487119436, + 0.0022317860275506973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.08905375003814697 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021990183740854263 + ], + "opposite_logprob_tokens": [ + -6.127199172973633 + ], + "logprob_reward": -0.0021990183740854263, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021990183740854263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04869295656681061, + 0.17067435383796692, + 0.1273631751537323, + 0.27890995144844055, + 0.021664917469024658, + 0.025580434128642082, + 0.062211912125349045, + 0.09554295241832733, + 0.15378990769386292, + 0.015569452196359634 + ], + "credit_map": [ + 0.0034828726202249527, + 0.00638636713847518, + 0.005131440702825785, + 0.0050993612967431545, + 0.011913764290511608, + 0.006676611490547657, + 0.004316800739616156, + 0.005685737822204828, + 0.007087752688676119, + 0.03895749896764755, + 0.005967813543975353, + 0.011816872283816338, + 0.0626690462231636, + 0.02335035242140293, + 0.012811032123863697, + 0.00801398977637291, + 0.008702275343239307, + 0.030247380957007408, + 0.01657726988196373, + 0.009049603715538979, + 0.011091840453445911, + 0.03161078691482544, + 0.010405062697827816, + 0.009678961709141731, + 0.014292259700596333, + 0.11971014738082886, + 0.010159051977097988, + 0.009110619314014912, + 0.020218076184391975, + 0.09331222623586655, + 0.004550836514681578, + 0.007556746248155832, + 0.002227692399173975, + 0.003227901179343462, + 0.0031806870829313993, + 0.0017840922810137272, + 0.0017318809404969215, + 0.005037248134613037, + 0.00177437299862504, + 0.0027010408230125904, + 0.0014213052345439792, + 0.00657394016161561, + 0.0024244326632469893, + 0.0017957148374989629, + 0.002542670350521803, + 0.0021202131174504757, + 0.001775058452039957, + 0.006927099544554949, + 0.00431316951289773, + 0.018894819542765617, + 0.005754388403147459, + 0.0027077484410256147, + 0.003479602048173547, + 0.006692961789667606, + 0.0024464691523462534, + 0.017922749742865562, + 0.012368845753371716, + 0.009625869803130627, + 0.014922535046935081, + 0.007725135888904333, + 0.007422838360071182, + 0.010324940085411072, + 0.005896023474633694, + 0.02725677564740181, + 0.008361466228961945, + 0.0657333955168724, + 0.003951302729547024, + 0.004837097600102425, + 0.012493785470724106, + 0.00876808911561966, + 0.0037612973246723413, + 0.04588346183300018, + 0.0022516779135912657, + 0.001980452099815011, + 0.002263729926198721, + 0.0017413548193871975, + 0.0017619358841329813, + 0.002436555689200759, + 0.0010456267045810819, + 0.0020881188102066517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.0021990183740854263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738675758242607 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.006738675758242607, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738675758242607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09912947565317154, + 0.02971651405096054, + 0.08849677443504333, + 0.4079209864139557, + 0.013902977108955383, + 0.035385020077228546, + 0.088304303586483, + 0.03290868550539017, + 0.18080812692642212, + 0.023427117615938187 + ], + "credit_map": [ + 0.006972667761147022, + 0.015268470160663128, + 0.0071867951191961765, + 0.006367201451212168, + 0.024683380499482155, + 0.024661971256136894, + 0.00462687574326992, + 0.009362109936773777, + 0.0025614378973841667, + 0.0027240945491939783, + 0.004699042998254299, + 0.003833820577710867, + 0.0033496792893856764, + 0.005808536428958178, + 0.0032230862416327, + 0.003516815137118101, + 0.006591592449694872, + 0.021347973495721817, + 0.01159397792071104, + 0.005739475600421429, + 0.008405392058193684, + 0.022115135565400124, + 0.006202474236488342, + 0.006500756833702326, + 0.021001454442739487, + 0.17943324148654938, + 0.012818059884011745, + 0.013147328980267048, + 0.032847121357917786, + 0.1335120052099228, + 0.0048825922422111034, + 0.010279187001287937, + 0.001406228868290782, + 0.0024916089605540037, + 0.0016177614452317357, + 0.0013469092082232237, + 0.0014012797037139535, + 0.0028934022411704063, + 0.001143523259088397, + 0.0016022627241909504, + 0.0022053932771086693, + 0.005203103646636009, + 0.0027580326423048973, + 0.002434658585116267, + 0.003792365314438939, + 0.006292342208325863, + 0.002328225877135992, + 0.010370897129178047, + 0.007565981708467007, + 0.020391421392560005, + 0.0104749146848917, + 0.004412088077515364, + 0.005528108682483435, + 0.009896612726151943, + 0.0043752738274633884, + 0.02565990388393402, + 0.00278846244327724, + 0.002313659293577075, + 0.010089614428579807, + 0.003233881201595068, + 0.001983846537768841, + 0.002912118099629879, + 0.002078043296933174, + 0.007509058341383934, + 0.014929136261343956, + 0.06205255165696144, + 0.05466333404183388, + 0.007685627322643995, + 0.00823970977216959, + 0.008726135827600956, + 0.0039049494080245495, + 0.02060667611658573, + 0.0030781221576035023, + 0.0036153187975287437, + 0.0026708270888775587, + 0.002275527687743306, + 0.0026139835827052593, + 0.004273827653378248, + 0.002009177580475807, + 0.00289033236913383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.006738675758242607 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary, compact shrub with woody, gnarled branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the shrub rapidly stretch outward and upward while new shoots and leaf buds sprout from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the new leaves transform from crinkled buds to fully textured forms in a fluid, continuous wave of growth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the new leaves a vibrant, translucent spring green that glow from within due to warm light filtering through?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene captured with a slightly low-angle, static medium shot in a soft-focus meadow with long, soft shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057424932718277 + ], + "opposite_logprob_tokens": [ + -3.8955743312835693 + ], + "logprob_reward": -0.02057424932718277, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057424932718277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10325011610984802, + 0.0491354800760746, + 0.03002339042723179, + 0.0895242691040039, + 0.03235182166099548, + 0.03640512004494667, + 0.018516041338443756, + 0.1463346779346466, + 0.45707967877388, + 0.0373794361948967 + ], + "credit_map": [ + 0.009897839277982712, + 0.012300309725105762, + 0.009643607772886753, + 0.009689887054264545, + 0.018677327781915665, + 0.022077100351452827, + 0.007894068956375122, + 0.013069966807961464, + 0.002766588469967246, + 0.010538052767515182, + 0.004215564578771591, + 0.0033534604590386152, + 0.013202005065977573, + 0.0075723957270383835, + 0.0030871189665049314, + 0.004400286823511124, + 0.002253839047625661, + 0.008255518041551113, + 0.003380671376362443, + 0.0020470842719078064, + 0.003491933923214674, + 0.00631605414673686, + 0.0020848552230745554, + 0.0021934325341135263, + 0.005112518090754747, + 0.03905583545565605, + 0.0025553812738507986, + 0.0029652672819793224, + 0.008572308346629143, + 0.026667892932891846, + 0.0016681557754054666, + 0.002926906803622842, + 0.00447466317564249, + 0.005115301813930273, + 0.00422306451946497, + 0.0021818913519382477, + 0.0023948370944708586, + 0.007060674950480461, + 0.001710345153696835, + 0.0051910411566495895, + 0.002057274803519249, + 0.004117853008210659, + 0.005630146246403456, + 0.0016167891444638371, + 0.002281309338286519, + 0.0066077266819775105, + 0.001667080563493073, + 0.012426936998963356, + 0.0015974066918715835, + 0.0028429424855858088, + 0.003111055353656411, + 0.00164239841978997, + 0.002593401586636901, + 0.0024798517115414143, + 0.001350605278275907, + 0.002898378297686577, + 0.019641056656837463, + 0.018508825451135635, + 0.015118982642889023, + 0.01393427886068821, + 0.01386813260614872, + 0.0224355086684227, + 0.00960707850754261, + 0.0332207977771759, + 0.03098249062895775, + 0.2036379873752594, + 0.08176625519990921, + 0.01720554567873478, + 0.031169714406132698, + 0.02472250908613205, + 0.008695501834154129, + 0.05889961123466492, + 0.0036288639530539513, + 0.006679872516542673, + 0.004269843455404043, + 0.004028251860290766, + 0.0040480163879692554, + 0.005930726882070303, + 0.0037673301994800568, + 0.005026529543101788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.02057424932718277 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_1/index_tensor([4617, 4617])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_1/index_tensor([4617, 4617])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7166604b011ac64c06b83847df5c39e94947e04a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_1/index_tensor([4617, 4617])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long, static shot, shot from the ground, captures a strangely shaped spacecraft landing on a rugged, rocky landscape shrouded in a thick, milky fog. The craft, an asymmetrical, biomechanical design, is constructed from a dull, carapace-like black metal, its surface covered in intricate etchings that reflect no light in the humid air. With a low mechanical hum and the hiss of steam, a heavy hydraulic ramp on the ship's belly slowly deploys downward, finally making contact with the ground. A strange, cold, cyan light emanates from the depths of the gloomy hatch, dispersing into a hazy halo in the dense fog. A vague, slender figure emerges at the top of the ramp. It struggles to crawl outward, one slender limb tentatively reaching out to grasp the damp ramp's edge. The rest of its body, struggling and laborious, inched its way out of the abysmal darkness within the craft, its unsteady movements suggesting fatigue or pain. The entire picture is dominated by an almost silent silence, full of suspense and awe of unknown creatures, creating a tense and mysterious sci-fi atmosphere.", + "seed": 32330, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_1/index_4617_seed_32330_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09827385265380144, + "vlm_logprob": -0.09827385265380144, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.09827385265380144, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19400453567504883 + ], + "opposite_logprob_tokens": [ + -2.069004535675049 + ], + "logprob_reward": -0.19400453567504883, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.19400453567504883, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19261908531188965, + 0.13524213433265686, + 0.06650878489017487, + 0.04124758392572403, + 0.07005120068788528, + 0.0898892804980278, + 0.10540498048067093, + 0.13681164383888245, + 0.06294972449541092, + 0.09927557408809662 + ], + "credit_map": [ + 0.024545056745409966, + 0.01983572170138359, + 0.029381783679127693, + 0.04251222312450409, + 0.017755722627043724, + 0.01768624037504196, + 0.02282216027379036, + 0.01808016560971737, + 0.031198780983686447, + 0.008527275174856186, + 0.04870051518082619, + 0.013817290775477886, + 0.00936145056039095, + 0.007901805453002453, + 0.007666192948818207, + 0.008068829774856567, + 0.007751354016363621, + 0.00549800880253315, + 0.008336398750543594, + 0.014293743297457695, + 0.008293879218399525, + 0.006933247204869986, + 0.006009048782289028, + 0.00939310435205698, + 0.005502875428646803, + 0.003976015839725733, + 0.004390707705169916, + 0.005610656924545765, + 0.007124718278646469, + 0.005380465183407068, + 0.004545022267848253, + 0.0047171227633953094, + 0.007652128115296364, + 0.006995304487645626, + 0.008119574747979641, + 0.007827273570001125, + 0.013084751553833485, + 0.009310265071690083, + 0.008705594576895237, + 0.008356310427188873, + 0.007892351597547531, + 0.00896524079144001, + 0.014653674326837063, + 0.009806385263800621, + 0.011513965204358101, + 0.013567992486059666, + 0.008904791437089443, + 0.014584878459572792, + 0.008813477121293545, + 0.010706834495067596, + 0.021961022168397903, + 0.005935788154602051, + 0.030745606869459152, + 0.010293754749000072, + 0.012172183021903038, + 0.00477631576359272, + 0.010102789849042892, + 0.010470009408891201, + 0.027168776839971542, + 0.010843993164598942, + 0.041091546416282654, + 0.010442806407809258, + 0.015341196209192276, + 0.011350535787642002, + 0.009488396346569061, + 0.007660359609872103, + 0.009302953258156776, + 0.007293353788554668, + 0.010564889758825302, + 0.006712003145366907, + 0.0062269968912005424, + 0.005700772162526846, + 0.013074051588773727, + 0.01144493743777275, + 0.011585350148379803, + 0.014708351343870163, + 0.014704922214150429, + 0.008603713475167751, + 0.0143582196906209, + 0.01079602912068367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.19400453567504883 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033458471298218 + ], + "opposite_logprob_tokens": [ + -2.350334644317627 + ], + "logprob_reward": -0.10033458471298218, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033458471298218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2149776816368103, + 0.1292751282453537, + 0.09222215414047241, + 0.09859412163496017, + 0.04420342668890953, + 0.038142964243888855, + 0.041037268936634064, + 0.117585688829422, + 0.05225619673728943, + 0.17170535027980804 + ], + "credit_map": [ + 0.026173794642090797, + 0.03061843477189541, + 0.02830454520881176, + 0.026630325242877007, + 0.025697991251945496, + 0.023153889924287796, + 0.03133140131831169, + 0.023067297413945198, + 0.024170635268092155, + 0.010362147353589535, + 0.03803981840610504, + 0.0150457127019763, + 0.013535033911466599, + 0.009325091727077961, + 0.010532441549003124, + 0.008264257572591305, + 0.009917320683598518, + 0.00929925125092268, + 0.01340132113546133, + 0.020874816924333572, + 0.01126713678240776, + 0.006854753475636244, + 0.01143572200089693, + 0.009171837009489536, + 0.013723119162023067, + 0.008785830810666084, + 0.01646328717470169, + 0.011228254996240139, + 0.015769070014357567, + 0.010695855133235455, + 0.010477613657712936, + 0.011451094411313534, + 0.00382818141952157, + 0.0038621847052127123, + 0.006162471603602171, + 0.0032619615085422993, + 0.007449239958077669, + 0.006087823770940304, + 0.0071816821582615376, + 0.006369882728904486, + 0.0036782880779355764, + 0.004762973636388779, + 0.004041125066578388, + 0.0025792510714381933, + 0.0054674046114087105, + 0.009342621080577374, + 0.004337703809142113, + 0.00393359549343586, + 0.0032549328170716763, + 0.0040782238356769085, + 0.007596262264996767, + 0.002710113301873207, + 0.010787916369736195, + 0.004989003762602806, + 0.0051381830126047134, + 0.0024826356675475836, + 0.011621774174273014, + 0.015281789936125278, + 0.0149355698376894, + 0.01747097261250019, + 0.02121930755674839, + 0.012808619067072868, + 0.012779458425939083, + 0.011468196287751198, + 0.005762523040175438, + 0.006749712862074375, + 0.005931757856160402, + 0.0066390084102749825, + 0.009544549509882927, + 0.005147785414010286, + 0.00817042775452137, + 0.0043104300275444984, + 0.014730561524629593, + 0.01813974417746067, + 0.022173315286636353, + 0.021076472476124763, + 0.031178193166851997, + 0.021284114569425583, + 0.024910692125558853, + 0.01821225881576538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.10033458471298218 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905331045389175 + ], + "opposite_logprob_tokens": [ + -2.4640533924102783 + ], + "logprob_reward": -0.08905331045389175, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905331045389175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1361912190914154, + 0.17938795685768127, + 0.1662943959236145, + 0.04993622004985809, + 0.17454852163791656, + 0.05577736720442772, + 0.04231216385960579, + 0.11713393032550812, + 0.04559588432312012, + 0.032822392880916595 + ], + "credit_map": [ + 0.017686257138848305, + 0.01771485060453415, + 0.02129860408604145, + 0.02068639174103737, + 0.018980015069246292, + 0.011719753965735435, + 0.015011593699455261, + 0.013093750923871994, + 0.04651081562042236, + 0.009551463648676872, + 0.06466981023550034, + 0.014141789637506008, + 0.014466281980276108, + 0.010652611963450909, + 0.010950889438390732, + 0.008444295264780521, + 0.014883278869092464, + 0.01530467253178358, + 0.026661671698093414, + 0.031985003501176834, + 0.023501455783843994, + 0.013912346214056015, + 0.023890098556876183, + 0.01615585759282112, + 0.006533521227538586, + 0.004043225664645433, + 0.007174257654696703, + 0.006227081175893545, + 0.007551707327365875, + 0.006576618179678917, + 0.004413573071360588, + 0.007416235748678446, + 0.0162204522639513, + 0.01754746213555336, + 0.01825626753270626, + 0.012405574321746826, + 0.033041637390851974, + 0.038216136395931244, + 0.02192004956305027, + 0.016940932720899582, + 0.007908542640507221, + 0.006639911327511072, + 0.007367332465946674, + 0.006042130757123232, + 0.008203591220080853, + 0.007018431089818478, + 0.007374417502433062, + 0.0052230097353458405, + 0.0038646915927529335, + 0.005443761125206947, + 0.005959274712949991, + 0.005149036645889282, + 0.00646550627425313, + 0.00555777782574296, + 0.005223325919359922, + 0.004648791626095772, + 0.008822575211524963, + 0.01705409772694111, + 0.015283969230949879, + 0.009617521427571774, + 0.022805651649832726, + 0.019692974165081978, + 0.013165448792278767, + 0.010691690258681774, + 0.005153825972229242, + 0.004439088050276041, + 0.006460830569267273, + 0.006423737853765488, + 0.007689819671213627, + 0.004130612127482891, + 0.007250006776303053, + 0.004047960974276066, + 0.003462844528257847, + 0.004290045239031315, + 0.004704231396317482, + 0.0037864050827920437, + 0.005644495598971844, + 0.0040620178915560246, + 0.0037083798088133335, + 0.00316397356800735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.08905331045389175 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034871101379395 + ], + "opposite_logprob_tokens": [ + -2.350348711013794 + ], + "logprob_reward": -0.10034871101379395, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034871101379395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2025095671415329, + 0.1448623538017273, + 0.09616335481405258, + 0.040821436792612076, + 0.2012348622083664, + 0.05367230996489525, + 0.042904920876026154, + 0.14979887008666992, + 0.03797229751944542, + 0.030059952288866043 + ], + "credit_map": [ + 0.02358657494187355, + 0.029590412974357605, + 0.033033587038517, + 0.032401908189058304, + 0.02094591222703457, + 0.020826326683163643, + 0.02533053606748581, + 0.016794314607977867, + 0.03381827473640442, + 0.007721828296780586, + 0.04958886653184891, + 0.013592803850769997, + 0.012818961404263973, + 0.010287825018167496, + 0.008443656377494335, + 0.008590136654675007, + 0.008692719973623753, + 0.0102730393409729, + 0.014556570909917355, + 0.01716499589383602, + 0.013405866920948029, + 0.007753781042993069, + 0.01443647500127554, + 0.009879902936518192, + 0.005535405594855547, + 0.0032890820875763893, + 0.004864982794970274, + 0.005336213856935501, + 0.0059380847960710526, + 0.004920425359159708, + 0.004561066161841154, + 0.0063761756755411625, + 0.018138380721211433, + 0.019320426508784294, + 0.021899204701185226, + 0.013757814653217793, + 0.040363967418670654, + 0.04181265830993652, + 0.025088338181376457, + 0.02085406333208084, + 0.0069931079633533955, + 0.006932619493454695, + 0.008547578006982803, + 0.004586708731949329, + 0.007487973663955927, + 0.006431792862713337, + 0.006498699076473713, + 0.006193829234689474, + 0.004107756540179253, + 0.004504058044403791, + 0.00830354355275631, + 0.00299186073243618, + 0.010070032440125942, + 0.005591259337961674, + 0.004016928840428591, + 0.003319482784718275, + 0.011343419551849365, + 0.01821017451584339, + 0.032275304198265076, + 0.00979855190962553, + 0.03580057621002197, + 0.016770483925938606, + 0.016980420798063278, + 0.008619929663836956, + 0.0038507163990288973, + 0.004987852647900581, + 0.004686814732849598, + 0.005593375768512487, + 0.005987642798572779, + 0.004828108008950949, + 0.005000461358577013, + 0.0030373241752386093, + 0.0043178885243833065, + 0.0035036855842918158, + 0.003529452718794346, + 0.0033887638710439205, + 0.00496602663770318, + 0.0035460249055176973, + 0.003785580163821578, + 0.003022529883310199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.10034871101379395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628121413290501 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007628121413290501, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628121413290501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1240077093243599, + 0.36897751688957214, + 0.08465838432312012, + 0.09238582104444504, + 0.09073376655578613, + 0.10240573436021805, + 0.0317053347826004, + 0.051328469067811966, + 0.021328724920749664, + 0.0324685238301754 + ], + "credit_map": [ + 0.014826108701527119, + 0.012763715349137783, + 0.020646320655941963, + 0.023176101967692375, + 0.01579640991985798, + 0.010950589552521706, + 0.013782448135316372, + 0.012066018767654896, + 0.11595749855041504, + 0.011647054925560951, + 0.1610921025276184, + 0.02768530324101448, + 0.015207754448056221, + 0.008068205788731575, + 0.014937970787286758, + 0.014381621032953262, + 0.009341280907392502, + 0.008334334939718246, + 0.013751170597970486, + 0.019683385267853737, + 0.00898088701069355, + 0.007225862238556147, + 0.008117357268929482, + 0.009224102832376957, + 0.014844496734440327, + 0.008735778741538525, + 0.012448656372725964, + 0.01408461481332779, + 0.013072855770587921, + 0.01207368541508913, + 0.007724277209490538, + 0.009401459246873856, + 0.0092745590955019, + 0.010019229725003242, + 0.011700794100761414, + 0.00692966440692544, + 0.017782114446163177, + 0.014390048570930958, + 0.012145426124334335, + 0.008491931483149529, + 0.014577693305909634, + 0.01185921672731638, + 0.018427930772304535, + 0.0077902874909341335, + 0.013066246174275875, + 0.011319706216454506, + 0.012850559316575527, + 0.012514091096818447, + 0.0026956896763294935, + 0.004185270983725786, + 0.005228632595390081, + 0.0024626560043543577, + 0.006571798585355282, + 0.0044531021267175674, + 0.0029147362802177668, + 0.003193450393155217, + 0.005310714710503817, + 0.00499196071177721, + 0.009654607623815536, + 0.0047078849747776985, + 0.010565211996436119, + 0.006253145635128021, + 0.005591987632215023, + 0.00425295252352953, + 0.002889753319323063, + 0.002269395859912038, + 0.0025049294345080853, + 0.0029521610122174025, + 0.003708476200699806, + 0.0023883895482867956, + 0.002718128962442279, + 0.0018974911654368043, + 0.003561273915693164, + 0.004141324665397406, + 0.0046815718524158, + 0.003529783571138978, + 0.006871913094073534, + 0.0029282269533723593, + 0.00411396287381649, + 0.0026404689997434616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.007628121413290501 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, static shot, shot from the ground, captures a strangely shaped spacecraft landing on a rugged, rocky landscape shrouded in a thick, milky fog. The craft, an asymmetrical, biomechanical design, is constructed from a dull, carapace-like black metal, its surface covered in intricate etchings that reflect no light in the humid air. With a low mechanical hum and the hiss of steam, a heavy hydraulic ramp on the ship's belly slowly deploys downward, finally making contact with the ground. A strange, cold, cyan light emanates from the depths of the gloomy hatch, dispersing into a hazy halo in the dense fog. A vague, slender figure emerges at the top of the ramp. It struggles to crawl outward, one slender limb tentatively reaching out to grasp the damp ramp's edge. The rest of its body, struggling and laborious, inched its way out of the abysmal darkness within the craft, its unsteady movements suggesting fatigue or pain. The entire picture is dominated by an almost silent silence, full of suspense and awe of unknown creatures, creating a tense and mysterious sci-fi atmosphere.", + "seed": 32331, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_1/index_4617_seed_32331_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09107699133455753, + "vlm_logprob": -0.09107699133455753, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.09107699133455753, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.26466599106788635 + ], + "opposite_logprob_tokens": [ + -1.639665961265564 + ], + "logprob_reward": -0.26466599106788635, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.26466599106788635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12619230151176453, + 0.15638203918933868, + 0.14202634990215302, + 0.1623087078332901, + 0.22469541430473328, + 0.06254826486110687, + 0.032523684203624725, + 0.025354444980621338, + 0.02053118497133255, + 0.0474376305937767 + ], + "credit_map": [ + 0.014924373477697372, + 0.020006917417049408, + 0.016633301973342896, + 0.01536976546049118, + 0.013096342794597149, + 0.020329445600509644, + 0.013571210205554962, + 0.012260939925909042, + 0.01962033286690712, + 0.01762019656598568, + 0.021772822365164757, + 0.02073168009519577, + 0.014281508512794971, + 0.024785814806818962, + 0.018735211342573166, + 0.018834460526704788, + 0.018382834270596504, + 0.016840314492583275, + 0.0257143322378397, + 0.022398794069886208, + 0.013777275569736958, + 0.011752200312912464, + 0.018538428470492363, + 0.014622144401073456, + 0.014909563586115837, + 0.017590217292308807, + 0.025587471202015877, + 0.024597594514489174, + 0.019303934648633003, + 0.015395093709230423, + 0.021157246083021164, + 0.023767586797475815, + 0.02913297526538372, + 0.02020983397960663, + 0.028198760002851486, + 0.027881501242518425, + 0.01972338743507862, + 0.01957867294549942, + 0.026271741837263107, + 0.053698524832725525, + 0.008076420053839684, + 0.006703819148242474, + 0.013248027302324772, + 0.006461949087679386, + 0.01066138967871666, + 0.0040552979335188866, + 0.005635399371385574, + 0.00770596181973815, + 0.005337570793926716, + 0.003280034987255931, + 0.008376396261155605, + 0.003864060854539275, + 0.002341424347832799, + 0.002177532995119691, + 0.003145810216665268, + 0.004000849090516567, + 0.004104047082364559, + 0.0024142898619174957, + 0.0052750990726053715, + 0.003198426915332675, + 0.0020700059831142426, + 0.0021024569869041443, + 0.0030536882113665342, + 0.0031364294700324535, + 0.0024632217828184366, + 0.0026386957615613937, + 0.002466889563947916, + 0.0020183264277875423, + 0.0024957717396318913, + 0.002057744422927499, + 0.004516162443906069, + 0.0018743707332760096, + 0.007463205140084028, + 0.006666603498160839, + 0.00628192164003849, + 0.005109483376145363, + 0.005845706909894943, + 0.003922380972653627, + 0.006490017287433147, + 0.005658310372382402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.26466599106788635 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.112964928150177 + ], + "opposite_logprob_tokens": [ + -2.2379648685455322 + ], + "logprob_reward": -0.112964928150177, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.112964928150177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2528606057167053, + 0.19489359855651855, + 0.06568390876054764, + 0.0683816596865654, + 0.05420037731528282, + 0.07446873188018799, + 0.0400739349424839, + 0.14721156656742096, + 0.040823500603437424, + 0.06140202283859253 + ], + "credit_map": [ + 0.012330434285104275, + 0.009988892823457718, + 0.09350758790969849, + 0.014020424336194992, + 0.008773691952228546, + 0.014793487265706062, + 0.014965623617172241, + 0.08448047935962677, + 0.02425548806786537, + 0.025704074651002884, + 0.029940802603960037, + 0.023641014471650124, + 0.01771511137485504, + 0.023913389071822166, + 0.019571270793676376, + 0.030152449384331703, + 0.005771805997937918, + 0.007622530683875084, + 0.011590583249926567, + 0.014642421156167984, + 0.005815125070512295, + 0.0047301980666816235, + 0.010794535279273987, + 0.004716711118817329, + 0.005464137531816959, + 0.011450378224253654, + 0.009054174646735191, + 0.009700615890324116, + 0.009496092796325684, + 0.005834932904690504, + 0.010118653066456318, + 0.007262673228979111, + 0.005084943026304245, + 0.004952185787260532, + 0.007721899077296257, + 0.007763198111206293, + 0.007474882062524557, + 0.004262084141373634, + 0.0105388592928648, + 0.006402326747775078, + 0.009433564729988575, + 0.009090902283787727, + 0.011368022300302982, + 0.012631586752831936, + 0.01000609528273344, + 0.00574566749855876, + 0.007474652025848627, + 0.008718241937458515, + 0.004264298360794783, + 0.004463226068764925, + 0.0068407743237912655, + 0.005057982634752989, + 0.004076807294040918, + 0.0037676794454455376, + 0.0068798670545220375, + 0.004723300226032734, + 0.016728529706597328, + 0.01679874025285244, + 0.024976611137390137, + 0.016610942780971527, + 0.019909746944904327, + 0.012951602227985859, + 0.02167116478085518, + 0.017564229667186737, + 0.004542929586023092, + 0.004315254744142294, + 0.007005693390965462, + 0.004626081790775061, + 0.005217850673943758, + 0.0035114570055156946, + 0.006205876357853413, + 0.005398354958742857, + 0.00796592514961958, + 0.011010772548615932, + 0.007429470773786306, + 0.00706366216763854, + 0.006135549396276474, + 0.006579885724931955, + 0.008447127416729927, + 0.006769627798348665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.112964928150177 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810751438140869 + ], + "opposite_logprob_tokens": [ + -3.288107395172119 + ], + "logprob_reward": -0.03810751438140869, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810751438140869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18156367540359497, + 0.344575971364975, + 0.07565214484930038, + 0.034964669495821, + 0.22509722411632538, + 0.037406668066978455, + 0.012792427092790604, + 0.037166111171245575, + 0.024265777319669724, + 0.026515336707234383 + ], + "credit_map": [ + 0.009671434760093689, + 0.014470919035375118, + 0.054131560027599335, + 0.011248141527175903, + 0.00795784406363964, + 0.01286417618393898, + 0.01326700672507286, + 0.05795259773731232, + 0.035441454499959946, + 0.0450957715511322, + 0.067104272544384, + 0.041117072105407715, + 0.03065907396376133, + 0.04902089387178421, + 0.031545113772153854, + 0.044592320919036865, + 0.005697327200323343, + 0.008570431731641293, + 0.01566198468208313, + 0.015594635158777237, + 0.0077374884858727455, + 0.0066755968146026134, + 0.010738307610154152, + 0.004976377822458744, + 0.0034390317741781473, + 0.0036789888981729746, + 0.005506285931915045, + 0.00463583180680871, + 0.0037550083361566067, + 0.0028400346636772156, + 0.004247152712196112, + 0.0068623339757323265, + 0.029428912326693535, + 0.018892256543040276, + 0.028479263186454773, + 0.02647394686937332, + 0.020266495645046234, + 0.019406229257583618, + 0.02902347780764103, + 0.05312665179371834, + 0.004083135165274143, + 0.005729481112211943, + 0.004514620639383793, + 0.004482390359044075, + 0.00798382330685854, + 0.0028951559215784073, + 0.004453079774975777, + 0.0032649829518049955, + 0.0017687100917100906, + 0.0013983125099912286, + 0.0036058968398720026, + 0.0014258096925914288, + 0.0011622686870396137, + 0.0009566130465827882, + 0.0011587007902562618, + 0.0013161150272935629, + 0.004136586561799049, + 0.00413885060697794, + 0.005520477890968323, + 0.003925666678696871, + 0.007120783906430006, + 0.0031351789366453886, + 0.005705317482352257, + 0.00348324584774673, + 0.0026369458064436913, + 0.0032921843230724335, + 0.0036818431690335274, + 0.002727033104747534, + 0.0034710958134382963, + 0.0023269152734428644, + 0.0037477349396795034, + 0.002382023725658655, + 0.0037700252141803503, + 0.0035054816398769617, + 0.003321277443319559, + 0.0026535976212471724, + 0.004427637439221144, + 0.0024728085845708847, + 0.0034851760137826204, + 0.002879333682358265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.03810751438140869 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369936719536781 + ], + "opposite_logprob_tokens": [ + -3.4086992740631104 + ], + "logprob_reward": -0.03369936719536781, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369936719536781, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09647195786237717, + 0.1520007997751236, + 0.056541625410318375, + 0.05409689620137215, + 0.21522554755210876, + 0.044377878308296204, + 0.04182916879653931, + 0.19919686019420624, + 0.09331255406141281, + 0.04694676399230957 + ], + "credit_map": [ + 0.005706841126084328, + 0.005815283861011267, + 0.02685142494738102, + 0.008033807389438152, + 0.0049853273667395115, + 0.013658017851412296, + 0.008416298776865005, + 0.023004954680800438, + 0.01638670824468136, + 0.01925053261220455, + 0.02839776501059532, + 0.017421426251530647, + 0.013728881254792213, + 0.021846957504749298, + 0.014149291440844536, + 0.02081923745572567, + 0.004143393132835627, + 0.007421633694320917, + 0.011666898615658283, + 0.007575397379696369, + 0.006465551909059286, + 0.005921193398535252, + 0.008837818168103695, + 0.004509741440415382, + 0.004642342682927847, + 0.00555744394659996, + 0.007669040001928806, + 0.007390825543552637, + 0.0053658438846468925, + 0.004156327340751886, + 0.00928118173032999, + 0.01003388874232769, + 0.028545040637254715, + 0.019382299855351448, + 0.027325402945280075, + 0.026471374556422234, + 0.020402520895004272, + 0.018817517906427383, + 0.02396976575255394, + 0.050311632454395294, + 0.0044185081496834755, + 0.006147566717118025, + 0.008764451369643211, + 0.005849859677255154, + 0.005775813478976488, + 0.0029843354132026434, + 0.006878162268549204, + 0.003559179836884141, + 0.004350245464593172, + 0.004351997282356024, + 0.006938368547707796, + 0.0038271434605121613, + 0.0075782667845487595, + 0.003928057383745909, + 0.006764132995158434, + 0.0040909587405622005, + 0.02335127256810665, + 0.02325945720076561, + 0.02311531826853752, + 0.020375650376081467, + 0.029465680941939354, + 0.016434606164693832, + 0.04137185961008072, + 0.021823005750775337, + 0.009791538119316101, + 0.008766560815274715, + 0.009162820875644684, + 0.011584452353417873, + 0.011280350387096405, + 0.0070952726528048515, + 0.02563069388270378, + 0.010000860318541527, + 0.006746799685060978, + 0.005869835615158081, + 0.005203288979828358, + 0.005762383341789246, + 0.006063539069145918, + 0.0035278545692563057, + 0.009229941293597221, + 0.0045431237667799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 1 + ], + "selected_reward": -0.03369936719536781 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947155877947807 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947155877947807, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947155877947807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1182335913181305, + 0.3385578989982605, + 0.022125566378235817, + 0.04902837052941322, + 0.3049633800983429, + 0.0216391459107399, + 0.011561094783246517, + 0.014745775610208511, + 0.02342257834970951, + 0.09572263807058334 + ], + "credit_map": [ + 0.00650055194273591, + 0.006050687283277512, + 0.03933035582304001, + 0.008465582504868507, + 0.004432135261595249, + 0.008798656053841114, + 0.007603790611028671, + 0.037051837891340256, + 0.03241482377052307, + 0.04318451136350632, + 0.06445842981338501, + 0.038033757358789444, + 0.03083840198814869, + 0.048765815794467926, + 0.03198479488492012, + 0.04887736961245537, + 0.0021591049153357744, + 0.0024679440539330244, + 0.0035461413208395243, + 0.00396424625068903, + 0.0026193868834525347, + 0.002040108200162649, + 0.002776063745841384, + 0.002552570542320609, + 0.005065233912318945, + 0.005462247878313065, + 0.008107995614409447, + 0.008146542124450207, + 0.004868935327976942, + 0.0038617122918367386, + 0.006806558929383755, + 0.006709148641675711, + 0.041323814541101456, + 0.02741830423474312, + 0.03838449344038963, + 0.037072427570819855, + 0.026583118364214897, + 0.027150362730026245, + 0.03295983374118805, + 0.07407104223966599, + 0.0021836271043866873, + 0.0026561932172626257, + 0.00304663204587996, + 0.0028044653590768576, + 0.004670844879001379, + 0.0017685185885056853, + 0.0025538422632962465, + 0.0019550221040844917, + 0.0013542782980948687, + 0.0013161824317649007, + 0.0020051144529134035, + 0.0013794387923553586, + 0.0014853313332423568, + 0.0012015728279948235, + 0.0016669081524014473, + 0.0011522684944793582, + 0.001559714088216424, + 0.001592801185324788, + 0.003237995319068432, + 0.0016908443067222834, + 0.001549102715216577, + 0.0012064942857250571, + 0.002374935895204544, + 0.0015338872326537967, + 0.002685368061065674, + 0.002213880652561784, + 0.003354590153321624, + 0.0027797967195510864, + 0.0027571097016334534, + 0.001710572512820363, + 0.005478898528963327, + 0.002442362252622843, + 0.016904883086681366, + 0.013885624706745148, + 0.0136201661080122, + 0.010456840507686138, + 0.010513423942029476, + 0.00828923936933279, + 0.010764631442725658, + 0.011287827976047993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.005947155877947807 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, static shot, shot from the ground, captures a strangely shaped spacecraft landing on a rugged, rocky landscape shrouded in a thick, milky fog. The craft, an asymmetrical, biomechanical design, is constructed from a dull, carapace-like black metal, its surface covered in intricate etchings that reflect no light in the humid air. With a low mechanical hum and the hiss of steam, a heavy hydraulic ramp on the ship's belly slowly deploys downward, finally making contact with the ground. A strange, cold, cyan light emanates from the depths of the gloomy hatch, dispersing into a hazy halo in the dense fog. A vague, slender figure emerges at the top of the ramp. It struggles to crawl outward, one slender limb tentatively reaching out to grasp the damp ramp's edge. The rest of its body, struggling and laborious, inched its way out of the abysmal darkness within the craft, its unsteady movements suggesting fatigue or pain. The entire picture is dominated by an almost silent silence, full of suspense and awe of unknown creatures, creating a tense and mysterious sci-fi atmosphere.", + "seed": 32332, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_1/index_4617_seed_32332_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04822018137201667, + "vlm_logprob": -0.04822018137201667, + "vlm_yes_no_margin": 3.6499999523162843, + "vlm_token_credit": -0.04822018137201667, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07260244339704514 + ], + "opposite_logprob_tokens": [ + -3.1976025104522705 + ], + "logprob_reward": -0.07260244339704514, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.07260244339704514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11110004037618637, + 0.2666797935962677, + 0.07268428057432175, + 0.08665166050195694, + 0.05278449505567551, + 0.029842115938663483, + 0.05726950615644455, + 0.060277123004198074, + 0.05163010209798813, + 0.2110809087753296 + ], + "credit_map": [ + 0.013894151896238327, + 0.014074565842747688, + 0.016135714948177338, + 0.01817290298640728, + 0.013768663629889488, + 0.00865933671593666, + 0.011914307251572609, + 0.01448039524257183, + 0.013385523110628128, + 0.021979600191116333, + 0.08002543449401855, + 0.02917615696787834, + 0.013724417425692081, + 0.008323497138917446, + 0.0846293568611145, + 0.015435825102031231, + 0.004492344334721565, + 0.011721826158463955, + 0.013997047208249569, + 0.011581871658563614, + 0.008256458677351475, + 0.005953076761215925, + 0.010865041986107826, + 0.005816618446260691, + 0.0037596041802316904, + 0.00699329050257802, + 0.019566159695386887, + 0.036029715090990067, + 0.0031895721331238747, + 0.004259332548826933, + 0.00892319530248642, + 0.003930789418518543, + 0.0036512231454253197, + 0.007283654995262623, + 0.00735511165112257, + 0.013793614692986012, + 0.0063512735068798065, + 0.0046982550993561745, + 0.005724506452679634, + 0.003926856443285942, + 0.002477213740348816, + 0.0041829426772892475, + 0.003758523380383849, + 0.007441249210387468, + 0.0026470436714589596, + 0.002191811567172408, + 0.0044196592643857, + 0.002723671728745103, + 0.003231797367334366, + 0.005702396854758263, + 0.003745388239622116, + 0.02189400978386402, + 0.0035985929425805807, + 0.0028426586650311947, + 0.009734384715557098, + 0.006520277354866266, + 0.003099653869867325, + 0.006351142190396786, + 0.004966383799910545, + 0.01898900792002678, + 0.005690368823707104, + 0.004362252075225115, + 0.011321491561830044, + 0.005496818572282791, + 0.00811867043375969, + 0.0057005686685442924, + 0.005796563345938921, + 0.006515772081911564, + 0.0034499159082770348, + 0.004909077193588018, + 0.007742596324533224, + 0.009396940469741821, + 0.013499030843377113, + 0.011414116248488426, + 0.08832130581140518, + 0.06579500436782837, + 0.003798293648287654, + 0.011820406652987003, + 0.006547163240611553, + 0.009885597974061966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.07260244339704514 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298207193613052 + ], + "opposite_logprob_tokens": [ + -2.2379820346832275 + ], + "logprob_reward": -0.11298207193613052, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298207193613052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036817748099565506, + 0.20232203602790833, + 0.031576842069625854, + 0.03297978267073631, + 0.03030070848762989, + 0.044329043477773666, + 0.08145922422409058, + 0.027007225900888443, + 0.10087351500988007, + 0.41233381628990173 + ], + "credit_map": [ + 0.003646749071776867, + 0.006475504487752914, + 0.003037380753085017, + 0.0071920473128557205, + 0.004524255637079477, + 0.004456184804439545, + 0.0037857305724173784, + 0.0036998989526182413, + 0.00881513673812151, + 0.01435364130884409, + 0.06485913693904877, + 0.01697620563209057, + 0.006836244370788336, + 0.004359528888016939, + 0.07653651386499405, + 0.009585649706423283, + 0.002053258242085576, + 0.0056185368448495865, + 0.0035841097123920918, + 0.007615359965711832, + 0.003565282793715596, + 0.0023032636381685734, + 0.0046770889312028885, + 0.0021599445026367903, + 0.0016209331806749105, + 0.0031736043747514486, + 0.005055759102106094, + 0.012523174285888672, + 0.0017040529055520892, + 0.0016107301926240325, + 0.004898764658719301, + 0.0023927681613713503, + 0.003393680090084672, + 0.003913456574082375, + 0.0029665075708180666, + 0.004261417780071497, + 0.004243011586368084, + 0.0029908481519669294, + 0.005165887530893087, + 0.0033659019973129034, + 0.002970866858959198, + 0.006920716259628534, + 0.004535863175988197, + 0.009985576383769512, + 0.0058379340916872025, + 0.002824368653818965, + 0.007520736660808325, + 0.0037329853512346745, + 0.00480805104598403, + 0.006358779966831207, + 0.004628376103937626, + 0.03354118764400482, + 0.00310093117877841, + 0.004381104838103056, + 0.01181054674088955, + 0.01283025648444891, + 0.001679867971688509, + 0.0037600118666887283, + 0.004041600972414017, + 0.0038618820253759623, + 0.0020105948206037283, + 0.0018662072252482176, + 0.004252155311405659, + 0.00553490687161684, + 0.011523978784680367, + 0.02243799716234207, + 0.009393761865794659, + 0.012548746541142464, + 0.009964428842067719, + 0.012688452377915382, + 0.013756475411355495, + 0.008559685200452805, + 0.026019517332315445, + 0.02218576893210411, + 0.17207391560077667, + 0.13278727233409882, + 0.0074073816649615765, + 0.022833218798041344, + 0.013588139787316322, + 0.015438640490174294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.11298207193613052 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819479651749134 + ], + "opposite_logprob_tokens": [ + -4.018194675445557 + ], + "logprob_reward": -0.01819479651749134, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819479651749134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05787745118141174, + 0.09759477525949478, + 0.04440626502037048, + 0.0940215215086937, + 0.049847107380628586, + 0.022877518087625504, + 0.1359180510044098, + 0.028697416186332703, + 0.04569276049733162, + 0.423067182302475 + ], + "credit_map": [ + 0.006752624176442623, + 0.008726853877305984, + 0.005742251873016357, + 0.009226826019585133, + 0.005367356352508068, + 0.0037432515528053045, + 0.008647722192108631, + 0.009670563973486423, + 0.007710730191320181, + 0.014407938346266747, + 0.02001016214489937, + 0.01627068780362606, + 0.009260778315365314, + 0.006132878828793764, + 0.015101264230906963, + 0.00870033260434866, + 0.002530735917389393, + 0.007232996169477701, + 0.00623001903295517, + 0.00992506556212902, + 0.004999191965907812, + 0.0031616261694580317, + 0.007100216578692198, + 0.00322641315869987, + 0.004220128990709782, + 0.009241420775651932, + 0.012110086157917976, + 0.038728486746549606, + 0.004749800078570843, + 0.003941366448998451, + 0.013940298929810524, + 0.007089938037097454, + 0.0031058606691658497, + 0.007869120687246323, + 0.005745599512010813, + 0.006957460660487413, + 0.010489419102668762, + 0.004550201818346977, + 0.0062978556379675865, + 0.0048315864987671375, + 0.0017625958425924182, + 0.0031884005293250084, + 0.003081075381487608, + 0.004711000714451075, + 0.0018101121531799436, + 0.0015755051281303167, + 0.004263611510396004, + 0.002485217759385705, + 0.007952905260026455, + 0.009196251630783081, + 0.00786464475095272, + 0.053893428295850754, + 0.007049032021313906, + 0.0069416887126863, + 0.023678693920373917, + 0.019341401755809784, + 0.0016906720120459795, + 0.00325735192745924, + 0.0021731702145189047, + 0.008496769703924656, + 0.001703632646240294, + 0.002005110029131174, + 0.0037608081474900246, + 0.00560990022495389, + 0.0036125504411756992, + 0.011107291094958782, + 0.004176088608801365, + 0.006116800010204315, + 0.006539517547935247, + 0.004710190463811159, + 0.005635201930999756, + 0.003795122029259801, + 0.027042575180530548, + 0.022380832582712173, + 0.17694327235221863, + 0.13561125099658966, + 0.007026943378150463, + 0.023476362228393555, + 0.01355359423905611, + 0.01703231781721115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.01819479651749134 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370501473546028 + ], + "opposite_logprob_tokens": [ + -3.408704996109009 + ], + "logprob_reward": -0.03370501473546028, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370501473546028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037357885390520096, + 0.017585119232535362, + 0.049668360501527786, + 0.14612220227718353, + 0.02724943310022354, + 0.017665648832917213, + 0.027243267744779587, + 0.02555406093597412, + 0.016859477385878563, + 0.634694516658783 + ], + "credit_map": [ + 0.00434707710519433, + 0.0074739763513207436, + 0.003581214463338256, + 0.006752525921911001, + 0.003814269322901964, + 0.002631404437124729, + 0.003631364554166794, + 0.005126055795699358, + 0.0020007947459816933, + 0.0024862210266292095, + 0.002590238582342863, + 0.0026214742101728916, + 0.0018270990112796426, + 0.001479555619880557, + 0.0026113619096577168, + 0.0019683775026351213, + 0.003415136830881238, + 0.00968195777386427, + 0.006944911554455757, + 0.006601706612855196, + 0.008322370238602161, + 0.004211049526929855, + 0.006301780231297016, + 0.0041894493624567986, + 0.006197416689246893, + 0.013737766072154045, + 0.029147552326321602, + 0.06290105730295181, + 0.003992222715169191, + 0.005131927318871021, + 0.01646839827299118, + 0.008545877411961555, + 0.0020925309509038925, + 0.003837053431198001, + 0.0023581543937325478, + 0.006454740185290575, + 0.004590020049363375, + 0.002669554203748703, + 0.00272920960560441, + 0.002518170978873968, + 0.0012602565111592412, + 0.0022222776897251606, + 0.002097325399518013, + 0.002724628895521164, + 0.001303254859521985, + 0.0012124874629080296, + 0.004902595188468695, + 0.0019428242230787873, + 0.0014477197546511889, + 0.002792108105495572, + 0.0019588000141084194, + 0.010102598927915096, + 0.001885277801193297, + 0.001419289386831224, + 0.004453377332538366, + 0.0031840966548770666, + 0.001463141874410212, + 0.0035518882796168327, + 0.002056782366707921, + 0.004614340607076883, + 0.004157501272857189, + 0.001683322829194367, + 0.006171211134642363, + 0.0018558738520368934, + 0.001424851012416184, + 0.0033297149930149317, + 0.0012676309561356902, + 0.0023342894855886698, + 0.0020691922400146723, + 0.0015253395540639758, + 0.003301000688225031, + 0.0016074592713266611, + 0.04278741404414177, + 0.03362688794732094, + 0.2617056369781494, + 0.19590872526168823, + 0.011427712626755238, + 0.035498373210430145, + 0.019809991121292114, + 0.03392982855439186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.03370501473546028 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036165802739560604 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.0036165802739560604, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036165802739560604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0758313462138176, + 0.06051158532500267, + 0.0867866799235344, + 0.17757566273212433, + 0.1280304342508316, + 0.07973343133926392, + 0.14776398241519928, + 0.142488032579422, + 0.0431734137237072, + 0.058105453848838806 + ], + "credit_map": [ + 0.007963048294186592, + 0.010828088968992233, + 0.0100771589204669, + 0.01233051996678114, + 0.007603902835398912, + 0.007239715661853552, + 0.009830271825194359, + 0.009958629496395588, + 0.0052033355459570885, + 0.009858530014753342, + 0.00944086629897356, + 0.010242638178169727, + 0.005762305110692978, + 0.004599538166075945, + 0.01004237961024046, + 0.005361990537494421, + 0.004667895380407572, + 0.012377413921058178, + 0.014913062565028667, + 0.01780639961361885, + 0.009945638477802277, + 0.006832036655396223, + 0.014270440675318241, + 0.005973787512630224, + 0.0080798314884305, + 0.014695994555950165, + 0.03609978035092354, + 0.07431832700967789, + 0.0046057021245360374, + 0.0066357627511024475, + 0.02367873676121235, + 0.009461534209549427, + 0.012922470457851887, + 0.014285782352089882, + 0.02402079664170742, + 0.016190845519304276, + 0.012194821611046791, + 0.01164609007537365, + 0.019411621615290642, + 0.017357995733618736, + 0.005385307129472494, + 0.010344278067350388, + 0.009431661106646061, + 0.010204145684838295, + 0.009090371429920197, + 0.006720160599797964, + 0.020597638562321663, + 0.007959866896271706, + 0.00876292772591114, + 0.009452765807509422, + 0.009222018532454967, + 0.06277748942375183, + 0.003865040373057127, + 0.007588110398501158, + 0.023810623213648796, + 0.022284995764493942, + 0.006663719657808542, + 0.013515998609364033, + 0.012937849387526512, + 0.04972943291068077, + 0.01165817491710186, + 0.008780855685472488, + 0.026314126327633858, + 0.012887870892882347, + 0.006021199282258749, + 0.005460819229483604, + 0.004772062413394451, + 0.005702880676835775, + 0.0037445819471031427, + 0.00388269848190248, + 0.006826503202319145, + 0.00676266523078084, + 0.0030370312742888927, + 0.005284676793962717, + 0.016211606562137604, + 0.014225800521671772, + 0.0031044555362313986, + 0.0049023861065506935, + 0.00776100717484951, + 0.0035784824285656214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.0036165802739560604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long, static shot, shot from the ground, captures a strangely shaped spacecraft landing on a rugged, rocky landscape shrouded in a thick, milky fog. The craft, an asymmetrical, biomechanical design, is constructed from a dull, carapace-like black metal, its surface covered in intricate etchings that reflect no light in the humid air. With a low mechanical hum and the hiss of steam, a heavy hydraulic ramp on the ship's belly slowly deploys downward, finally making contact with the ground. A strange, cold, cyan light emanates from the depths of the gloomy hatch, dispersing into a hazy halo in the dense fog. A vague, slender figure emerges at the top of the ramp. It struggles to crawl outward, one slender limb tentatively reaching out to grasp the damp ramp's edge. The rest of its body, struggling and laborious, inched its way out of the abysmal darkness within the craft, its unsteady movements suggesting fatigue or pain. The entire picture is dominated by an almost silent silence, full of suspense and awe of unknown creatures, creating a tense and mysterious sci-fi atmosphere.", + "seed": 32333, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_1/index_4617_seed_32333_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04819727251306176, + "vlm_logprob": -0.04819727251306176, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.04819727251306176, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09513364732265472 + ], + "opposite_logprob_tokens": [ + -2.8451335430145264 + ], + "logprob_reward": -0.09513364732265472, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.09513364732265472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28739964962005615, + 0.08955792337656021, + 0.10404480248689651, + 0.06949301064014435, + 0.056573256850242615, + 0.05024361237883568, + 0.10952213406562805, + 0.10827486962080002, + 0.07148296386003494, + 0.05340779200196266 + ], + "credit_map": [ + 0.02729218266904354, + 0.031752295792102814, + 0.04216090589761734, + 0.04170679301023483, + 0.04049660265445709, + 0.02666913904249668, + 0.038131408393383026, + 0.03919031843543053, + 0.009355292655527592, + 0.007419813424348831, + 0.015520702116191387, + 0.013843879103660583, + 0.016604214906692505, + 0.006092341151088476, + 0.012921626679599285, + 0.0078000472858548164, + 0.006796760950237513, + 0.009296505711972713, + 0.007991775870323181, + 0.011021901853382587, + 0.030298307538032532, + 0.0065925451926887035, + 0.019716918468475342, + 0.012330080382525921, + 0.006041791755706072, + 0.007197156548500061, + 0.011611582711338997, + 0.011614106595516205, + 0.007212378550320864, + 0.004839122295379639, + 0.009833709336817265, + 0.011143156327307224, + 0.003503701649606228, + 0.006544019561260939, + 0.008059593848884106, + 0.009529595263302326, + 0.00929525587707758, + 0.005044768564403057, + 0.007417219690978527, + 0.007179103326052427, + 0.004548360127955675, + 0.005864886101335287, + 0.008601714856922626, + 0.009177935309708118, + 0.004827337805181742, + 0.0032655270770192146, + 0.007419242057949305, + 0.006538609974086285, + 0.01312379539012909, + 0.013552357442677021, + 0.014747818931937218, + 0.013738865964114666, + 0.01551648136228323, + 0.009853734634816647, + 0.013103236444294453, + 0.015885839238762856, + 0.010187527164816856, + 0.015076180920004845, + 0.012994133867323399, + 0.013832262717187405, + 0.016622668132185936, + 0.007798634935170412, + 0.015948794782161713, + 0.015814661979675293, + 0.010796629823744297, + 0.008090109564363956, + 0.008873007260262966, + 0.010298789478838444, + 0.009360584430396557, + 0.005673118866980076, + 0.008775316178798676, + 0.009615401737391949, + 0.005378609523177147, + 0.00685515720397234, + 0.01130423229187727, + 0.006852650549262762, + 0.008087780326604843, + 0.003499357495456934, + 0.005160931963473558, + 0.006269070319831371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.09513364732265472 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901301980018616 + ], + "opposite_logprob_tokens": [ + -2.5790131092071533 + ], + "logprob_reward": -0.07901301980018616, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901301980018616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10734966397285461, + 0.12929528951644897, + 0.10067696124315262, + 0.15477147698402405, + 0.06640097498893738, + 0.0824924111366272, + 0.0984041690826416, + 0.0987444594502449, + 0.09815871715545654, + 0.06370585411787033 + ], + "credit_map": [ + 0.007841105572879314, + 0.00997570063918829, + 0.01385777723044157, + 0.013527775183320045, + 0.01585465855896473, + 0.009244181215763092, + 0.023739870637655258, + 0.013308599591255188, + 0.007924795150756836, + 0.01430295966565609, + 0.012287091463804245, + 0.015006288886070251, + 0.04073207080364227, + 0.007838113233447075, + 0.019966622814536095, + 0.01123736146837473, + 0.005940137431025505, + 0.009671979583799839, + 0.007704105693846941, + 0.01019980013370514, + 0.030207274481654167, + 0.008093329146504402, + 0.02002924494445324, + 0.008831098675727844, + 0.008862416259944439, + 0.021567482501268387, + 0.0198096614331007, + 0.018809068948030472, + 0.019617047160863876, + 0.013652011752128601, + 0.027378490194678307, + 0.02507530152797699, + 0.0046571847051382065, + 0.007030857726931572, + 0.007703172974288464, + 0.01015862450003624, + 0.00888131558895111, + 0.00528288958594203, + 0.013559958897531033, + 0.009126980789005756, + 0.007403420750051737, + 0.008361445739865303, + 0.014470790512859821, + 0.012420861050486565, + 0.009375480934977531, + 0.004968262277543545, + 0.014253467321395874, + 0.011238683946430683, + 0.008468450978398323, + 0.013645933009684086, + 0.010705753229558468, + 0.012921278364956379, + 0.018697455525398254, + 0.010774698108434677, + 0.013148191384971142, + 0.010042412206530571, + 0.008131233975291252, + 0.012054054997861385, + 0.01919776201248169, + 0.012856462970376015, + 0.016267141327261925, + 0.00836316030472517, + 0.012424790300428867, + 0.009449861943721771, + 0.007726945448666811, + 0.010421589016914368, + 0.016178106889128685, + 0.015344479121267796, + 0.018589863553643227, + 0.007920175790786743, + 0.012976855039596558, + 0.009000711143016815, + 0.006784196477383375, + 0.00782257504761219, + 0.009202160872519016, + 0.006532937753945589, + 0.010766597464680672, + 0.004045028239488602, + 0.007715319748967886, + 0.010837039910256863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.07901301980018616 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058674581348896 + ], + "opposite_logprob_tokens": [ + -3.8955867290496826 + ], + "logprob_reward": -0.02058674581348896, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058674581348896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11046187579631805, + 0.16210898756980896, + 0.06223233789205551, + 0.08922573924064636, + 0.33060339093208313, + 0.04211237281560898, + 0.03554306551814079, + 0.05537787824869156, + 0.060121238231658936, + 0.05221321061253548 + ], + "credit_map": [ + 0.010358838364481926, + 0.011534169316291809, + 0.014753674156963825, + 0.017068129032850266, + 0.015453197993338108, + 0.00768403010442853, + 0.018673721700906754, + 0.014936097897589207, + 0.017212968319654465, + 0.013717453926801682, + 0.026380371302366257, + 0.030307702720165253, + 0.022060204297304153, + 0.009438423439860344, + 0.024266058579087257, + 0.018725784495472908, + 0.003622341202571988, + 0.004345332272350788, + 0.01036074012517929, + 0.011370854452252388, + 0.012744336389005184, + 0.0037303061690181494, + 0.010005485266447067, + 0.006052939221262932, + 0.009197496809065342, + 0.005875397939234972, + 0.02081473171710968, + 0.012527056969702244, + 0.008424931205809116, + 0.0062443651258945465, + 0.012649817392230034, + 0.013491926714777946, + 0.021225912496447563, + 0.04195701703429222, + 0.044062238186597824, + 0.035760123282670975, + 0.05354838818311691, + 0.03552047163248062, + 0.05833042412996292, + 0.040198806673288345, + 0.003089936450123787, + 0.006147394888103008, + 0.004955546464771032, + 0.00616700341925025, + 0.005627529229968786, + 0.0033067078329622746, + 0.00825384072959423, + 0.004564410075545311, + 0.003622534452006221, + 0.003816528245806694, + 0.0052829310297966, + 0.004401633981615305, + 0.004988097120076418, + 0.003061410505324602, + 0.004709874279797077, + 0.005660051014274359, + 0.005738177336752415, + 0.006776202470064163, + 0.006529298610985279, + 0.00706657487899065, + 0.009112624451518059, + 0.003624972887337208, + 0.0080728093162179, + 0.008457210846245289, + 0.0048732031136751175, + 0.007812906987965107, + 0.00786049384623766, + 0.009505103342235088, + 0.009006081148982048, + 0.005067776422947645, + 0.007573907729238272, + 0.008421758189797401, + 0.004946989938616753, + 0.0062234001234173775, + 0.011889009736478329, + 0.006539157126098871, + 0.007397029083222151, + 0.003851990681141615, + 0.0053841732442379, + 0.005981454625725746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.02058674581348896 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306168854236603 + ], + "opposite_logprob_tokens": [ + -3.1680617332458496 + ], + "logprob_reward": -0.04306168854236603, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306168854236603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1418968141078949, + 0.09836027771234512, + 0.15195728838443756, + 0.09280191361904144, + 0.17886199057102203, + 0.11310464143753052, + 0.06285935640335083, + 0.05878835171461105, + 0.06612520664930344, + 0.03524414822459221 + ], + "credit_map": [ + 0.009586489759385586, + 0.012219085358083248, + 0.018486453220248222, + 0.02414274960756302, + 0.022673474624753, + 0.013230114243924618, + 0.024167120456695557, + 0.017391327768564224, + 0.005760452710092068, + 0.010157167911529541, + 0.013050907291471958, + 0.015003648586571217, + 0.021606577560305595, + 0.006266074255108833, + 0.013827119953930378, + 0.012688323855400085, + 0.01073676161468029, + 0.011872567236423492, + 0.017451675608754158, + 0.01827474683523178, + 0.03622736036777496, + 0.009325395338237286, + 0.02960732765495777, + 0.018461458384990692, + 0.008227614685893059, + 0.007710831239819527, + 0.024256635457277298, + 0.013555859215557575, + 0.009842722676694393, + 0.0057340627536177635, + 0.011866341345012188, + 0.011607841588556767, + 0.012153931893408298, + 0.021306689828634262, + 0.023161420598626137, + 0.02339022234082222, + 0.03156285732984543, + 0.012549242004752159, + 0.03183691203594208, + 0.022900719195604324, + 0.00896273273974657, + 0.013154962100088596, + 0.015433622524142265, + 0.01582060195505619, + 0.015561612322926521, + 0.009064876474440098, + 0.023734813556075096, + 0.011371420696377754, + 0.006024884060025215, + 0.006589790340512991, + 0.011671643704175949, + 0.007142261136323214, + 0.010844971984624863, + 0.005388092715293169, + 0.0061877076514065266, + 0.009010004810988903, + 0.0056802695617079735, + 0.008768158964812756, + 0.008718790486454964, + 0.0067492821253836155, + 0.009141082875430584, + 0.004195555113255978, + 0.007680891547352076, + 0.007854320108890533, + 0.005785077344626188, + 0.008734120987355709, + 0.010832750238478184, + 0.009527241811156273, + 0.00991768203675747, + 0.004805881064385176, + 0.008160625584423542, + 0.008361820131540298, + 0.0035917358472943306, + 0.0034251061733812094, + 0.01106808427721262, + 0.004924159962683916, + 0.0034939853940159082, + 0.0020016850903630257, + 0.0027001737616956234, + 0.004039216786623001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.04306168854236603 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low ground-level perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the spacecraft designed with an asymmetrical, biomechanical shape made of dull black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a heavy hydraulic ramp deploy from the ship's belly and make contact with the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a vague, slender figure emerge from the hatch and crawl outward onto the ramp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a rugged, rocky landscape shrouded in thick, milky fog with a cyan light source?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319126108661294 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.00319126108661294, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319126108661294, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18757829070091248, + 0.10122186690568924, + 0.033619459718465805, + 0.08832897990942001, + 0.15119032561779022, + 0.0689680352807045, + 0.07989922165870667, + 0.08460192382335663, + 0.03077232651412487, + 0.1738196164369583 + ], + "credit_map": [ + 0.017824212089180946, + 0.017822876572608948, + 0.032552290707826614, + 0.02866605669260025, + 0.024336498230695724, + 0.01751997508108616, + 0.023580173030495644, + 0.02527621015906334, + 0.010935761034488678, + 0.006530407350510359, + 0.0241300780326128, + 0.013215157203376293, + 0.011689052917063236, + 0.005028779152780771, + 0.01679152622818947, + 0.012901113368570805, + 0.0032893624156713486, + 0.002384162973612547, + 0.00448202807456255, + 0.00554682407528162, + 0.004848710261285305, + 0.0026444189716130495, + 0.0049678124487400055, + 0.005456137470901012, + 0.008400795049965382, + 0.008411330170929432, + 0.018425432965159416, + 0.013925560750067234, + 0.00866567250341177, + 0.006680374965071678, + 0.012172323651611805, + 0.011647490784525871, + 0.010954903438687325, + 0.019581133499741554, + 0.021202778443694115, + 0.017387913540005684, + 0.022784670814871788, + 0.015587749890983105, + 0.024921946227550507, + 0.01876923069357872, + 0.0053926557302474976, + 0.0077206892892718315, + 0.008702686056494713, + 0.008776835165917873, + 0.01046215370297432, + 0.005169488023966551, + 0.013138401322066784, + 0.009605132043361664, + 0.010766169987618923, + 0.00876627303659916, + 0.010562676936388016, + 0.009843029081821442, + 0.01294525433331728, + 0.0063903494738042355, + 0.009378579445183277, + 0.011246883310377598, + 0.009061220102012157, + 0.011603521183133125, + 0.014603213407099247, + 0.01049744337797165, + 0.011494893580675125, + 0.007095856126397848, + 0.009663361124694347, + 0.010582414455711842, + 0.002164953388273716, + 0.004266038071364164, + 0.004199916031211615, + 0.004623680841177702, + 0.004469836596399546, + 0.0027323164977133274, + 0.004107429645955563, + 0.0042081549763679504, + 0.018027042970061302, + 0.02411222830414772, + 0.03387294337153435, + 0.022663088515400887, + 0.026676082983613014, + 0.012391867116093636, + 0.016738886013627052, + 0.019337479025125504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.00319126108661294 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_10/index_tensor([4521, 4521])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_10/index_tensor([4521, 4521])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c296e9e9647c871d5e7249c68301da7e5010e114 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_10/index_tensor([4521, 4521])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, upward, low-angle shot captures a majestic peacock taking flight from a desolate, rocky clearing, rendered in the distinctive style of a dark fairytale illustration. The creature's body is a mass of deep, velvety charcoal feathers, but its tail is a stunning and terrifying fan of seemingly sharp obsidian shards, each side catching the dim ambient light with a cold, glassy gleam. With a single, powerful, and deliberate beat, the peacock's massive wings launch vertically into the gloomy air. The entire scene is bathed in a heavy, desaturated color gradient, reducing the world to a palette of muted grays, deep blacks, and faded earth tones. The only exception to this monochromatic scheme is a faint, unsettling crimson glow within the peacock's eyes, casting a subtle, sinister aura. As its wings continued to beat vigorously, they spread a cloud of black, shimmering feathers that looked like tiny, sharp fragments of the night, which slowly drifted and spiraled downward through the silent, hazy atmosphere, creating a scene of haunting, malevolent beauty.", + "seed": 8844, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_10/index_4521_seed_8844_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10859797783195972, + "vlm_logprob": -0.10859797783195972, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.10859797783195972, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03235512226819992 + ], + "opposite_logprob_tokens": [ + -4.032355308532715 + ], + "logprob_reward": -0.03235512226819992, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.03235512226819992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11337326467037201, + 0.05306664854288101, + 0.11500603705644608, + 0.093784399330616, + 0.10271210968494415, + 0.07690716534852982, + 0.04685952514410019, + 0.16092641651630402, + 0.1990804523229599, + 0.038284000009298325 + ], + "credit_map": [ + 0.021523941308259964, + 0.010347823612391949, + 0.009622056037187576, + 0.013121465221047401, + 0.009719342924654484, + 0.017621519044041634, + 0.011528601869940758, + 0.019888505339622498, + 0.007157283369451761, + 0.0067789894528687, + 0.009091988205909729, + 0.005895782727748156, + 0.006324326153844595, + 0.005415801424533129, + 0.004636813420802355, + 0.007765667047351599, + 0.011239342391490936, + 0.009759089909493923, + 0.034634266048669815, + 0.008830579929053783, + 0.013978734612464905, + 0.008180651813745499, + 0.008697215467691422, + 0.01968616247177124, + 0.006637977436184883, + 0.006129502318799496, + 0.02338293567299843, + 0.0034134401939809322, + 0.007539781276136637, + 0.011502113193273544, + 0.003853687783703208, + 0.03132496029138565, + 0.023428743705153465, + 0.011696448549628258, + 0.01203299593180418, + 0.004148037172853947, + 0.010265005752444267, + 0.018849173560738564, + 0.00580730801448226, + 0.016484396532177925, + 0.019744673743844032, + 0.007013927213847637, + 0.007663584779947996, + 0.007795142009854317, + 0.0077372388914227486, + 0.0075492518953979015, + 0.0060050166212022305, + 0.013398329727351665, + 0.007200588937848806, + 0.004786927253007889, + 0.007018058095127344, + 0.0035073005128651857, + 0.002744575496762991, + 0.004662178456783295, + 0.00491372961550951, + 0.0120261674746871, + 0.020298125222325325, + 0.014828888699412346, + 0.01973670721054077, + 0.017483264207839966, + 0.013510588556528091, + 0.01582242175936699, + 0.02728721871972084, + 0.03195919841527939, + 0.015963422134518623, + 0.008322685956954956, + 0.07026386260986328, + 0.005351589526981115, + 0.00850070919841528, + 0.056130338460206985, + 0.02721906453371048, + 0.007328756153583527, + 0.0036622295156121254, + 0.008299728855490685, + 0.005224338732659817, + 0.006004365626722574, + 0.004582738038152456, + 0.002925590146332979, + 0.004039217717945576, + 0.0035457918420434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.03235512226819992 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811784088611603 + ], + "opposite_logprob_tokens": [ + -3.2881178855895996 + ], + "logprob_reward": -0.03811784088611603, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811784088611603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10695577412843704, + 0.052966248244047165, + 0.03281545266509056, + 0.07651825249195099, + 0.02776763029396534, + 0.04021046310663223, + 0.09326298534870148, + 0.10060562193393707, + 0.457468718290329, + 0.01142884325236082 + ], + "credit_map": [ + 0.024912044405937195, + 0.009786801412701607, + 0.009498866274952888, + 0.0074400813318789005, + 0.009919779375195503, + 0.02493579499423504, + 0.007854397408664227, + 0.012608004733920097, + 0.008994460105895996, + 0.004570131190121174, + 0.012258511036634445, + 0.005187653936445713, + 0.005110178142786026, + 0.005888930521905422, + 0.00463093351572752, + 0.00632544606924057, + 0.0035866284742951393, + 0.001852797926403582, + 0.009569268673658371, + 0.001610978739336133, + 0.0027817199006676674, + 0.00240225694142282, + 0.0021841919515281916, + 0.008827609941363335, + 0.0063049932941794395, + 0.005125665571540594, + 0.021442947909235954, + 0.0031993649899959564, + 0.004436195828020573, + 0.0074238707311451435, + 0.0028314292430877686, + 0.02575378306210041, + 0.004312139004468918, + 0.003308650804683566, + 0.0032488759607076645, + 0.0020124632865190506, + 0.003994733560830355, + 0.004915275145322084, + 0.0024674341548234224, + 0.003508059773594141, + 0.006518541369587183, + 0.004846527241170406, + 0.004903067369014025, + 0.004766521975398064, + 0.0040987092070281506, + 0.0051975795067846775, + 0.003271768568083644, + 0.0066077494993805885, + 0.01582537591457367, + 0.00794039387255907, + 0.014852117747068405, + 0.0068540871143341064, + 0.005149899981915951, + 0.007460565306246281, + 0.009093957021832466, + 0.026086585596203804, + 0.010014362633228302, + 0.00956142507493496, + 0.011702611111104488, + 0.010984525084495544, + 0.011215816251933575, + 0.011094653978943825, + 0.0126073332503438, + 0.023424895480275154, + 0.03487905114889145, + 0.015233444049954414, + 0.1725366860628128, + 0.011534733697772026, + 0.017115995287895203, + 0.12137263268232346, + 0.07078000903129578, + 0.014016148634254932, + 0.001093974569812417, + 0.0016110297292470932, + 0.0018007878679782152, + 0.0017515209037810564, + 0.0009414901142008603, + 0.0014933205675333738, + 0.0015314402990043163, + 0.0012052793754264712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.03811784088611603 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209007278084755 + ], + "opposite_logprob_tokens": [ + -2.8120901584625244 + ], + "logprob_reward": -0.06209007278084755, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209007278084755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1230701357126236, + 0.07585461437702179, + 0.04956042766571045, + 0.08297038823366165, + 0.11091913282871246, + 0.05419733002781868, + 0.16512157022953033, + 0.06725353002548218, + 0.14263658225536346, + 0.12841631472110748 + ], + "credit_map": [ + 0.021965594962239265, + 0.012216908857226372, + 0.011161035858094692, + 0.009773964993655682, + 0.01269281655550003, + 0.018740583211183548, + 0.012592407874763012, + 0.02392682246863842, + 0.009400533512234688, + 0.0074267820455133915, + 0.014078058302402496, + 0.007186551112681627, + 0.010473203845322132, + 0.009643306024372578, + 0.0073586925864219666, + 0.01028748694807291, + 0.005975518375635147, + 0.004884439054876566, + 0.009118274785578251, + 0.0029788322281092405, + 0.0062327333725988865, + 0.004617211874574423, + 0.0032694865949451923, + 0.0124839311465621, + 0.009394745342433453, + 0.009692332707345486, + 0.015647750347852707, + 0.0037821936421096325, + 0.008933430537581444, + 0.014939439482986927, + 0.004163390025496483, + 0.01641710288822651, + 0.01702590100467205, + 0.014554071240127087, + 0.011388844810426235, + 0.008634041994810104, + 0.013857915997505188, + 0.01837659440934658, + 0.012277699075639248, + 0.014804057776927948, + 0.010645666159689426, + 0.005964614450931549, + 0.006691715680062771, + 0.0054972488433122635, + 0.005713830702006817, + 0.00481444364413619, + 0.004079642239958048, + 0.010790168307721615, + 0.017401061952114105, + 0.015600515529513359, + 0.02247420884668827, + 0.017161492258310318, + 0.0356362946331501, + 0.017843762412667274, + 0.0184802133589983, + 0.02052401192486286, + 0.007399687543511391, + 0.006917132530361414, + 0.008939595893025398, + 0.007012565620243549, + 0.006376293022185564, + 0.006151954643428326, + 0.012015302665531635, + 0.012441005557775497, + 0.015377828851342201, + 0.018354196101427078, + 0.03467880189418793, + 0.00978273618966341, + 0.012352358549833298, + 0.012686744332313538, + 0.019990650936961174, + 0.01941327564418316, + 0.010848741047084332, + 0.01737748086452484, + 0.016454193741083145, + 0.023307327181100845, + 0.01831326261162758, + 0.014304272830486298, + 0.01329793594777584, + 0.014513093046844006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.06209007278084755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023311017081141472 + ], + "opposite_logprob_tokens": [ + -3.773310899734497 + ], + "logprob_reward": -0.023311017081141472, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023311017081141472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05108160898089409, + 0.04423448443412781, + 0.052214644849300385, + 0.06603473424911499, + 0.13395479321479797, + 0.03473992273211479, + 0.1432807296514511, + 0.03133586421608925, + 0.40789857506752014, + 0.03522466495633125 + ], + "credit_map": [ + 0.008539720438420773, + 0.004255876410752535, + 0.005528280511498451, + 0.006317811086773872, + 0.004027987830340862, + 0.006375944707542658, + 0.0062728263437747955, + 0.009763160720467567, + 0.0040049925446510315, + 0.0046157436445355415, + 0.007518863305449486, + 0.004963577724993229, + 0.0062342677265405655, + 0.006326856557279825, + 0.004961811937391758, + 0.005608373787254095, + 0.004507850855588913, + 0.004531431011855602, + 0.01109176967293024, + 0.004887420684099197, + 0.005547329317778349, + 0.005193023942410946, + 0.0038420159835368395, + 0.012613806873559952, + 0.007234076038002968, + 0.00778054166585207, + 0.012380658648908138, + 0.003010123036801815, + 0.006715114694088697, + 0.011058701202273369, + 0.0032582946587353945, + 0.014597220346331596, + 0.019501231610774994, + 0.016392502933740616, + 0.013764647766947746, + 0.011156010441482067, + 0.018816262483596802, + 0.02112450636923313, + 0.017024841159582138, + 0.016174783930182457, + 0.006101371720433235, + 0.0038207846228033304, + 0.00394896836951375, + 0.0034494290594011545, + 0.0038422083016484976, + 0.003353347070515156, + 0.0027970210649073124, + 0.007426795549690723, + 0.020301127806305885, + 0.012627353891730309, + 0.01844051480293274, + 0.014098228886723518, + 0.019302090629935265, + 0.011582765728235245, + 0.013835501857101917, + 0.03309316188097, + 0.0038457168266177177, + 0.0029553500935435295, + 0.0036346076522022486, + 0.0032489027362316847, + 0.0037151125725358725, + 0.0032196934334933758, + 0.003101934678852558, + 0.00761454738676548, + 0.032130591571331024, + 0.017520003020763397, + 0.15047521889209747, + 0.012301560491323471, + 0.016451627016067505, + 0.10242840647697449, + 0.059722911566495895, + 0.016868237406015396, + 0.003379819681867957, + 0.005613397341221571, + 0.006243285723030567, + 0.003198940074071288, + 0.004038217943161726, + 0.004304352682083845, + 0.0047361901961266994, + 0.0037104585207998753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.023311017081141472 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38711583614349365 + ], + "opposite_logprob_tokens": [ + -1.1371158361434937 + ], + "logprob_reward": -0.38711583614349365, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38711583614349365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04900185018777847, + 0.03359918296337128, + 0.0740344226360321, + 0.02501225471496582, + 0.06104784831404686, + 0.06609443575143814, + 0.1539885699748993, + 0.0665181577205658, + 0.4377339780330658, + 0.0329694002866745 + ], + "credit_map": [ + 0.006013492587953806, + 0.004955945070832968, + 0.00574212521314621, + 0.00758441723883152, + 0.004902062006294727, + 0.007050361949950457, + 0.005113741382956505, + 0.007639703340828419, + 0.0030318775679916143, + 0.0029337950982153416, + 0.005032793618738651, + 0.0042485399171710014, + 0.004228934179991484, + 0.005295904818922281, + 0.003976490814238787, + 0.004850844386965036, + 0.006497958209365606, + 0.007013064809143543, + 0.0187478419393301, + 0.007093031890690327, + 0.007679486647248268, + 0.006512534338980913, + 0.006618442013859749, + 0.013872054405510426, + 0.0035505308769643307, + 0.0024631135165691376, + 0.004473106004297733, + 0.0012421129504218698, + 0.002744539175182581, + 0.0035324920900166035, + 0.002068126806989312, + 0.004938232712447643, + 0.009311962872743607, + 0.007487109862267971, + 0.00631985766813159, + 0.003341758856549859, + 0.007312392815947533, + 0.013541930355131626, + 0.006906827911734581, + 0.0068259998224675655, + 0.011771862395107746, + 0.006421639118343592, + 0.008245037868618965, + 0.005240580998361111, + 0.006166933104395866, + 0.007119752932339907, + 0.005558652803301811, + 0.015569977462291718, + 0.0217464417219162, + 0.01199325267225504, + 0.01972801610827446, + 0.01411726139485836, + 0.022217297926545143, + 0.0134518351405859, + 0.016422489657998085, + 0.034311965107917786, + 0.006485617719590664, + 0.0064938925206661224, + 0.011688695289194584, + 0.007379515096545219, + 0.004780012182891369, + 0.006181019823998213, + 0.012821434997022152, + 0.01068795844912529, + 0.03425147384405136, + 0.014949196018278599, + 0.16134482622146606, + 0.012643890455365181, + 0.017093224450945854, + 0.10706479847431183, + 0.07475242018699646, + 0.015634076669812202, + 0.0028035033028572798, + 0.003391929203644395, + 0.003935222048312426, + 0.007524056825786829, + 0.004256102256476879, + 0.0038880002684891224, + 0.004202661104500294, + 0.0029679222498089075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.38711583614349365 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward, low-angle shot captures a majestic peacock taking flight from a desolate, rocky clearing, rendered in the distinctive style of a dark fairytale illustration. The creature's body is a mass of deep, velvety charcoal feathers, but its tail is a stunning and terrifying fan of seemingly sharp obsidian shards, each side catching the dim ambient light with a cold, glassy gleam. With a single, powerful, and deliberate beat, the peacock's massive wings launch vertically into the gloomy air. The entire scene is bathed in a heavy, desaturated color gradient, reducing the world to a palette of muted grays, deep blacks, and faded earth tones. The only exception to this monochromatic scheme is a faint, unsettling crimson glow within the peacock's eyes, casting a subtle, sinister aura. As its wings continued to beat vigorously, they spread a cloud of black, shimmering feathers that looked like tiny, sharp fragments of the night, which slowly drifted and spiraled downward through the silent, hazy atmosphere, creating a scene of haunting, malevolent beauty.", + "seed": 8845, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_10/index_4521_seed_8845_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18552803471684456, + "vlm_logprob": -0.18552803471684456, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.18552803471684456, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054049063473939896 + ], + "opposite_logprob_tokens": [ + -3.429049015045166 + ], + "logprob_reward": -0.054049063473939896, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.054049063473939896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09123227000236511, + 0.046482425183057785, + 0.09013044089078903, + 0.09722545742988586, + 0.2747623324394226, + 0.023023832589387894, + 0.023079467937350273, + 0.052823975682258606, + 0.1685349941253662, + 0.13270482420921326 + ], + "credit_map": [ + 0.015446276403963566, + 0.01350991427898407, + 0.016066133975982666, + 0.009547058492898941, + 0.009425941854715347, + 0.010522895492613316, + 0.009240618906915188, + 0.007473435252904892, + 0.005829715635627508, + 0.0066663650795817375, + 0.007726529147475958, + 0.00821433961391449, + 0.0045895990915596485, + 0.004701692145317793, + 0.00403095968067646, + 0.004723225720226765, + 0.01087463553994894, + 0.007124492898583412, + 0.01175545621663332, + 0.027066202834248543, + 0.005589189473539591, + 0.008853874169290066, + 0.008383998647332191, + 0.01048259250819683, + 0.007912652567029, + 0.007593201473355293, + 0.021253958344459534, + 0.020240401849150658, + 0.00843525305390358, + 0.010149084031581879, + 0.011335078626871109, + 0.010305826552212238, + 0.028415633365511894, + 0.03524693846702576, + 0.05293930321931839, + 0.05816662684082985, + 0.017924988642334938, + 0.021910028532147408, + 0.027254173532128334, + 0.032904643565416336, + 0.0027662001084536314, + 0.0025946972891688347, + 0.0035930818412452936, + 0.004011316224932671, + 0.002131745219230652, + 0.0019966280087828636, + 0.0037083846982568502, + 0.002221778966486454, + 0.002445711987093091, + 0.0022995942272245884, + 0.004395297262817621, + 0.006703283172100782, + 0.001492027658969164, + 0.0016669458709657192, + 0.0024263018276542425, + 0.0016503073275089264, + 0.006181091070175171, + 0.005352954380214214, + 0.008998476900160313, + 0.008988620713353157, + 0.003912751097232103, + 0.005955947563052177, + 0.009285617619752884, + 0.0041485135443508625, + 0.036380548030138016, + 0.013366593047976494, + 0.03590577095746994, + 0.039751533418893814, + 0.014371369034051895, + 0.012052535079419613, + 0.00976293534040451, + 0.00694370036944747, + 0.021400587633252144, + 0.018937483429908752, + 0.01629476808011532, + 0.01821293495595455, + 0.015621055848896503, + 0.015007207170128822, + 0.016645077615976334, + 0.010585718788206577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.054049063473939896 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297525465488434 + ], + "opposite_logprob_tokens": [ + -2.2379753589630127 + ], + "logprob_reward": -0.11297525465488434, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297525465488434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32754620909690857, + 0.05239994078874588, + 0.05078066140413284, + 0.04586326703429222, + 0.15590138733386993, + 0.06937193125486374, + 0.048075392842292786, + 0.04958874359726906, + 0.12278511375188828, + 0.0776873454451561 + ], + "credit_map": [ + 0.043701108545064926, + 0.05549747496843338, + 0.046514689922332764, + 0.036845725029706955, + 0.039633844047784805, + 0.06243240833282471, + 0.02570926770567894, + 0.017211711034178734, + 0.006117673125118017, + 0.009003028273582458, + 0.0066334703005850315, + 0.006307105999439955, + 0.0045914058573544025, + 0.005926365032792091, + 0.006194749381393194, + 0.007626140955835581, + 0.006466402672231197, + 0.004035239107906818, + 0.00698434142395854, + 0.013375714421272278, + 0.003764545079320669, + 0.004392975941300392, + 0.005461851134896278, + 0.006299586966633797, + 0.0040219672955572605, + 0.004462935496121645, + 0.00792373064905405, + 0.011058238334953785, + 0.004275625571608543, + 0.004502197727560997, + 0.0045886049047112465, + 0.005029967054724693, + 0.015002047643065453, + 0.015656176954507828, + 0.032595328986644745, + 0.03847865015268326, + 0.009319937787950039, + 0.01286078616976738, + 0.01549493707716465, + 0.01649351418018341, + 0.008261415176093578, + 0.00587849784642458, + 0.011882924474775791, + 0.015623440034687519, + 0.005054397042840719, + 0.004637339152395725, + 0.0104154571890831, + 0.0076184626668691635, + 0.00475934986025095, + 0.0032954979687929153, + 0.00817651953548193, + 0.013887795619666576, + 0.0025035222060978413, + 0.003012315835803747, + 0.005875104572623968, + 0.006565289106220007, + 0.0056658959947526455, + 0.004148281179368496, + 0.009513042867183685, + 0.011975240893661976, + 0.0029656810220330954, + 0.002857985207810998, + 0.006927506998181343, + 0.00553510757163167, + 0.021575553342700005, + 0.010068762116134167, + 0.02856195718050003, + 0.031271930783987045, + 0.01086838636547327, + 0.008970174938440323, + 0.005798109341412783, + 0.005670248530805111, + 0.008991550654172897, + 0.010193342342972755, + 0.009098397567868233, + 0.012118611484766006, + 0.010662567801773548, + 0.011011443100869656, + 0.008561914786696434, + 0.007049519568681717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.11297525465488434 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07904330641031265 + ], + "opposite_logprob_tokens": [ + -2.579043388366699 + ], + "logprob_reward": -0.07904330641031265, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07904330641031265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1682967096567154, + 0.11992712318897247, + 0.062057603150606155, + 0.08223749697208405, + 0.0823281854391098, + 0.03617275133728981, + 0.045730870217084885, + 0.07309456169605255, + 0.14987929165363312, + 0.18027544021606445 + ], + "credit_map": [ + 0.025720825418829918, + 0.02759183570742607, + 0.025127757340669632, + 0.01744314469397068, + 0.025330429896712303, + 0.020927799865603447, + 0.014125100336968899, + 0.012029816396534443, + 0.020595183596014977, + 0.008756056427955627, + 0.038352809846401215, + 0.010499531403183937, + 0.005542352329939604, + 0.011349959298968315, + 0.008914063684642315, + 0.01591716893017292, + 0.010253621265292168, + 0.00539900129660964, + 0.009218172170221806, + 0.01598595455288887, + 0.004110192880034447, + 0.004775685723870993, + 0.0053946711122989655, + 0.006920306012034416, + 0.0069578541442751884, + 0.005457557737827301, + 0.01255675870925188, + 0.01678496040403843, + 0.005109878256917, + 0.008574476465582848, + 0.011662743985652924, + 0.015133267268538475, + 0.00841307919472456, + 0.013306119479238987, + 0.011487065814435482, + 0.018365295603871346, + 0.005438866093754768, + 0.006632985081523657, + 0.010042097419500351, + 0.008642677217721939, + 0.005612039472907782, + 0.003204708220437169, + 0.005168560426682234, + 0.008784904144704342, + 0.002249225042760372, + 0.0017990795895457268, + 0.0050944313406944275, + 0.004259800538420677, + 0.005545554216951132, + 0.0036963277962058783, + 0.007205743808299303, + 0.0094552431255579, + 0.0032528794836252928, + 0.004194291774183512, + 0.004689985420554876, + 0.007690844126045704, + 0.008519399911165237, + 0.006423032842576504, + 0.011592280119657516, + 0.013605131767690182, + 0.00713225407525897, + 0.008453182876110077, + 0.011956315487623215, + 0.005412970203906298, + 0.024784095585346222, + 0.011569825932383537, + 0.035466648638248444, + 0.039917051792144775, + 0.014172788709402084, + 0.010334369726479053, + 0.00709021370857954, + 0.006544293370097876, + 0.024241337552666664, + 0.024844540283083916, + 0.02055382914841175, + 0.02656676061451435, + 0.02412690781056881, + 0.02381235547363758, + 0.021685229614377022, + 0.01444448996335268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.07904330641031265 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869406670331955 + ], + "opposite_logprob_tokens": [ + -3.048694133758545 + ], + "logprob_reward": -0.04869406670331955, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869406670331955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12449756264686584, + 0.06674647331237793, + 0.06092708185315132, + 0.17494738101959229, + 0.156483456492424, + 0.020155441015958786, + 0.02628488466143608, + 0.036420416086912155, + 0.12545453011989594, + 0.20808282494544983 + ], + "credit_map": [ + 0.0161571241915226, + 0.0176982544362545, + 0.021360969170928, + 0.014300317503511906, + 0.020533405244350433, + 0.014869303442537785, + 0.01102222129702568, + 0.008555966429412365, + 0.007110219448804855, + 0.008517762646079063, + 0.009452760219573975, + 0.010003912262618542, + 0.005610404070466757, + 0.006393842399120331, + 0.00680948281660676, + 0.012848085723817348, + 0.010052461177110672, + 0.006591455545276403, + 0.00905478186905384, + 0.012246654368937016, + 0.004601146560162306, + 0.004213857464492321, + 0.006200248841196299, + 0.007966474629938602, + 0.01183188334107399, + 0.015264449641108513, + 0.027752229943871498, + 0.03147395700216293, + 0.010574414394795895, + 0.02227349393069744, + 0.024846814572811127, + 0.030930133536458015, + 0.01022359449416399, + 0.010608606040477753, + 0.054628241807222366, + 0.042772937566041946, + 0.008211293257772923, + 0.00797365978360176, + 0.01108896266669035, + 0.010976158082485199, + 0.0021898779086768627, + 0.0021978672593832016, + 0.003084359457716346, + 0.004395151045173407, + 0.001745895016938448, + 0.0016250843182206154, + 0.0022848115768283606, + 0.0026323951315134764, + 0.0033883184660226107, + 0.002725288039073348, + 0.003667691024020314, + 0.004658099263906479, + 0.0022763689048588276, + 0.0030811289325356483, + 0.002975862706080079, + 0.003512126626446843, + 0.0037793770898133516, + 0.003907348029315472, + 0.005460600834339857, + 0.006746300961822271, + 0.004321238957345486, + 0.003577189054340124, + 0.00478470791131258, + 0.0038436513859778643, + 0.02033577486872673, + 0.009965857490897179, + 0.029250934720039368, + 0.032526981085538864, + 0.012602070346474648, + 0.00863044336438179, + 0.005814669653773308, + 0.006327793002128601, + 0.027157636359333992, + 0.028301961719989777, + 0.023517809808254242, + 0.029934437945485115, + 0.02679787389934063, + 0.029117949306964874, + 0.026339402422308922, + 0.016915759071707726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.04869406670331955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328784823417664 + ], + "opposite_logprob_tokens": [ + -0.7578784823417664 + ], + "logprob_reward": -0.6328784823417664, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328784823417664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2050715684890747, + 0.11416826397180557, + 0.07908961921930313, + 0.07581181824207306, + 0.09690765291452408, + 0.05062331259250641, + 0.04053319990634918, + 0.03524398431181908, + 0.07738947123289108, + 0.22516117990016937 + ], + "credit_map": [ + 0.026177655905485153, + 0.030626526102423668, + 0.03223458677530289, + 0.017854677513241768, + 0.04139290004968643, + 0.02460470050573349, + 0.015414133667945862, + 0.01676638238132, + 0.017619837075471878, + 0.011467519216239452, + 0.03533909097313881, + 0.015428337268531322, + 0.007841886952519417, + 0.010733026079833508, + 0.005953317508101463, + 0.009785250760614872, + 0.01024254597723484, + 0.0061788600869476795, + 0.010690152645111084, + 0.024731475859880447, + 0.005966757424175739, + 0.006866110023111105, + 0.005975244101136923, + 0.008438470773398876, + 0.008498581126332283, + 0.009840152226388454, + 0.011000790633261204, + 0.014798790216445923, + 0.005291770212352276, + 0.009563314728438854, + 0.009747607633471489, + 0.007070811465382576, + 0.007596032693982124, + 0.01088625006377697, + 0.01726166345179081, + 0.019106576219201088, + 0.006160228047519922, + 0.010015184059739113, + 0.013805299997329712, + 0.01207641139626503, + 0.005669999402016401, + 0.008991733193397522, + 0.008046302944421768, + 0.007560339756309986, + 0.00489033292979002, + 0.004133022855967283, + 0.005338235758244991, + 0.00599335040897131, + 0.005149134434759617, + 0.004743045661598444, + 0.005538754630833864, + 0.006697649136185646, + 0.00358121027238667, + 0.005140725057572126, + 0.004431082401424646, + 0.005251599010080099, + 0.005254787392914295, + 0.0029523377306759357, + 0.006309229880571365, + 0.0075761801563203335, + 0.0028291973285377026, + 0.002783911069855094, + 0.0047050234861671925, + 0.0028333179652690887, + 0.00985401775687933, + 0.007178450468927622, + 0.01714862324297428, + 0.01882154867053032, + 0.008591923862695694, + 0.006908230949193239, + 0.0031984797678887844, + 0.005688194185495377, + 0.029995450749993324, + 0.02674407884478569, + 0.02720271795988083, + 0.03564717248082161, + 0.030192870646715164, + 0.03028855472803116, + 0.023437393829226494, + 0.021652935072779655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.6328784823417664 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward, low-angle shot captures a majestic peacock taking flight from a desolate, rocky clearing, rendered in the distinctive style of a dark fairytale illustration. The creature's body is a mass of deep, velvety charcoal feathers, but its tail is a stunning and terrifying fan of seemingly sharp obsidian shards, each side catching the dim ambient light with a cold, glassy gleam. With a single, powerful, and deliberate beat, the peacock's massive wings launch vertically into the gloomy air. The entire scene is bathed in a heavy, desaturated color gradient, reducing the world to a palette of muted grays, deep blacks, and faded earth tones. The only exception to this monochromatic scheme is a faint, unsettling crimson glow within the peacock's eyes, casting a subtle, sinister aura. As its wings continued to beat vigorously, they spread a cloud of black, shimmering feathers that looked like tiny, sharp fragments of the night, which slowly drifted and spiraled downward through the silent, hazy atmosphere, creating a scene of haunting, malevolent beauty.", + "seed": 8846, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_10/index_4521_seed_8846_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07221742235124111, + "vlm_logprob": -0.07221742235124111, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.07221742235124111, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06446614861488342 + ], + "opposite_logprob_tokens": [ + -3.0644662380218506 + ], + "logprob_reward": -0.06446614861488342, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.06446614861488342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0789719820022583, + 0.0485537126660347, + 0.02933160960674286, + 0.36758431792259216, + 0.04650956392288208, + 0.10596012324094772, + 0.07483989745378494, + 0.07324757426977158, + 0.10705503821372986, + 0.0679461807012558 + ], + "credit_map": [ + 0.008967384696006775, + 0.01057417318224907, + 0.006156241986900568, + 0.009998713620007038, + 0.019208481535315514, + 0.00855807401239872, + 0.00737646734341979, + 0.008132446557283401, + 0.008989525958895683, + 0.0057594021782279015, + 0.006569027900695801, + 0.004774652421474457, + 0.005695146042853594, + 0.005508473142981529, + 0.0055677276104688644, + 0.0056897602044045925, + 0.002370029455050826, + 0.0038105486892163754, + 0.004497136455029249, + 0.00363104953430593, + 0.0035584864672273397, + 0.003960106987506151, + 0.00400862842798233, + 0.003495623357594013, + 0.056019656360149384, + 0.03963368758559227, + 0.04402927681803703, + 0.031777214258909225, + 0.02161170355975628, + 0.04167035594582558, + 0.04753551632165909, + 0.08530692011117935, + 0.005429006181657314, + 0.006472499575465918, + 0.006554261781275272, + 0.004987719468772411, + 0.004990051034837961, + 0.006829582620412111, + 0.005515883676707745, + 0.0057305581867694855, + 0.008230924606323242, + 0.00944221206009388, + 0.025511043146252632, + 0.010590781457722187, + 0.008454486727714539, + 0.010223392397165298, + 0.009999576024711132, + 0.023507708683609962, + 0.008465093560516834, + 0.009952339343726635, + 0.008993728086352348, + 0.008005528710782528, + 0.010380466468632221, + 0.01019273791462183, + 0.00860372930765152, + 0.010246279649436474, + 0.008235104382038116, + 0.00996145699173212, + 0.011097278445959091, + 0.00923135969787836, + 0.011910760775208473, + 0.009863999672234058, + 0.006622812710702419, + 0.006324795074760914, + 0.013870210386812687, + 0.010577717795968056, + 0.015298127196729183, + 0.010428107343614101, + 0.01281727384775877, + 0.019413091242313385, + 0.011012323200702667, + 0.013638181611895561, + 0.006078509613871574, + 0.009092986583709717, + 0.009357518516480923, + 0.0052039735019207, + 0.0055876667611300945, + 0.02051234059035778, + 0.006716648116707802, + 0.005396540276706219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 5 + ], + "selected_reward": -0.06446614861488342 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371607884764671 + ], + "opposite_logprob_tokens": [ + -3.4087159633636475 + ], + "logprob_reward": -0.03371607884764671, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371607884764671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3346134424209595, + 0.07568918168544769, + 0.0524674728512764, + 0.05113600194454193, + 0.12689939141273499, + 0.11519353091716766, + 0.029007351025938988, + 0.06523500382900238, + 0.1109912246465683, + 0.038767412304878235 + ], + "credit_map": [ + 0.03763937950134277, + 0.05063860863447189, + 0.02178337238729, + 0.04264373704791069, + 0.06742352247238159, + 0.04376724362373352, + 0.03727179393172264, + 0.033445775508880615, + 0.007141204085201025, + 0.012175360694527626, + 0.012013343162834644, + 0.006503227632492781, + 0.011952555738389492, + 0.00812394730746746, + 0.009129181504249573, + 0.008650359697639942, + 0.005850156769156456, + 0.0033179097808897495, + 0.012916314415633678, + 0.006712905131280422, + 0.004521224182099104, + 0.005888817831873894, + 0.006406309548765421, + 0.00685383565723896, + 0.008281049318611622, + 0.006365321110934019, + 0.007610670290887356, + 0.007605386897921562, + 0.003930495586246252, + 0.004702241625636816, + 0.006411321461200714, + 0.0062295179814100266, + 0.013125981204211712, + 0.009664959274232388, + 0.013093437068164349, + 0.008463239297270775, + 0.01183395367115736, + 0.03239864483475685, + 0.011026282794773579, + 0.027292894199490547, + 0.011131270788609982, + 0.01047584693878889, + 0.02677774243056774, + 0.0111296521499753, + 0.008868859149515629, + 0.01358797401189804, + 0.012957033701241016, + 0.02026514708995819, + 0.0033588518854230642, + 0.0024782323744148016, + 0.003675583517178893, + 0.0026110243052244186, + 0.0032536955550312996, + 0.005152454134076834, + 0.003663021605461836, + 0.004814483691006899, + 0.009578341618180275, + 0.008601696230471134, + 0.010472716763615608, + 0.006600390188395977, + 0.009070243686437607, + 0.008572559803724289, + 0.0058023626916110516, + 0.006536688655614853, + 0.01124570518732071, + 0.011231639422476292, + 0.01892169378697872, + 0.01472311932593584, + 0.009526905603706837, + 0.02064085565507412, + 0.013223632238805294, + 0.011477667838335037, + 0.0027764129918068647, + 0.00487478356808424, + 0.0055843377485871315, + 0.005189583171159029, + 0.0035058222711086273, + 0.009610513225197792, + 0.0037879387382417917, + 0.0034380198922008276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.03371607884764671 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04311284050345421 + ], + "opposite_logprob_tokens": [ + -3.1681127548217773 + ], + "logprob_reward": -0.04311284050345421, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311284050345421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1655069887638092, + 0.09901119768619537, + 0.057995229959487915, + 0.24607744812965393, + 0.025060061365365982, + 0.12101050466299057, + 0.02058834210038185, + 0.0241779163479805, + 0.09167042374610901, + 0.1489018350839615 + ], + "credit_map": [ + 0.020347880199551582, + 0.026115531101822853, + 0.010214729234576225, + 0.021008295938372612, + 0.03198525309562683, + 0.02130059525370598, + 0.01824895665049553, + 0.01628577709197998, + 0.010263201780617237, + 0.008537392131984234, + 0.019701343029737473, + 0.008413503877818584, + 0.01292993500828743, + 0.01198135782033205, + 0.014577023684978485, + 0.012607453390955925, + 0.0036777013447135687, + 0.006819394417107105, + 0.008033230900764465, + 0.008983133360743523, + 0.006969771347939968, + 0.0037875929847359657, + 0.009400833398103714, + 0.010323579423129559, + 0.02792169898748398, + 0.029965544119477272, + 0.022017212584614754, + 0.019838009029626846, + 0.012078151106834412, + 0.02408798784017563, + 0.03441851958632469, + 0.07575035095214844, + 0.003922195639461279, + 0.0029337892774492502, + 0.003067410085350275, + 0.00274289702065289, + 0.002833061618730426, + 0.0034563886001706123, + 0.0031379153952002525, + 0.002966407686471939, + 0.011737894266843796, + 0.014736276119947433, + 0.019308816641569138, + 0.010664846748113632, + 0.013976944610476494, + 0.015757950022816658, + 0.015423769131302834, + 0.019404025748372078, + 0.0019213295308873057, + 0.003935107029974461, + 0.002414243295788765, + 0.002185731427744031, + 0.002755148336291313, + 0.0032251733355224133, + 0.0020719885360449553, + 0.0020796230528503656, + 0.0015430786879733205, + 0.002799007808789611, + 0.0035773722920566797, + 0.0019699071999639273, + 0.004296005703508854, + 0.003842091653496027, + 0.002671618014574051, + 0.0034788381308317184, + 0.010200001299381256, + 0.009274032898247242, + 0.015904100611805916, + 0.00997442752122879, + 0.01028474885970354, + 0.014098752290010452, + 0.011813213117420673, + 0.010121157392859459, + 0.010203693062067032, + 0.020537640899419785, + 0.025663774460554123, + 0.010509062558412552, + 0.008265632204711437, + 0.054735396057367325, + 0.008918838575482368, + 0.01006781030446291 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.04311284050345421 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018210014328360558 + ], + "opposite_logprob_tokens": [ + -4.018209934234619 + ], + "logprob_reward": -0.018210014328360558, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018210014328360558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06124745309352875, + 0.0910005271434784, + 0.09595804661512375, + 0.3630081117153168, + 0.022410230711102486, + 0.12997697293758392, + 0.04112910479307175, + 0.020273013040423393, + 0.08709249645471573, + 0.08790401369333267 + ], + "credit_map": [ + 0.0059503247030079365, + 0.009693634696304798, + 0.00513177365064621, + 0.00793374888598919, + 0.010992606170475483, + 0.007573914714157581, + 0.007431244943290949, + 0.006540205329656601, + 0.010284161195158958, + 0.008038570173084736, + 0.017033154144883156, + 0.007952934131026268, + 0.011027459986507893, + 0.011912647634744644, + 0.013084611855447292, + 0.011666989885270596, + 0.005535707343369722, + 0.009548604488372803, + 0.014921773225069046, + 0.014561590738594532, + 0.009613163769245148, + 0.006845883559435606, + 0.015519951470196247, + 0.019411375746130943, + 0.052645377814769745, + 0.040059611201286316, + 0.03777679055929184, + 0.030990762636065483, + 0.02014150843024254, + 0.040361929684877396, + 0.04603945091366768, + 0.09499267488718033, + 0.0030411824118345976, + 0.0023440713994205, + 0.002686353400349617, + 0.0028422828763723373, + 0.001834488706663251, + 0.003609538311138749, + 0.0028801453299820423, + 0.00317216943949461, + 0.013467847369611263, + 0.015608404763042927, + 0.021842090412974358, + 0.011664028279483318, + 0.013854573480784893, + 0.016581453382968903, + 0.017168642953038216, + 0.01978992111980915, + 0.005760538391768932, + 0.0062671834602952, + 0.004808741621673107, + 0.004196994472295046, + 0.004820255097001791, + 0.006633902434259653, + 0.00480694230645895, + 0.0038345486391335726, + 0.001525582280009985, + 0.0031516237650066614, + 0.0026714240666478872, + 0.0026230320800095797, + 0.0025463909842073917, + 0.0023236488923430443, + 0.0031578855123370886, + 0.0022734261583536863, + 0.007870448753237724, + 0.009001788683235645, + 0.016066083684563637, + 0.009805618785321712, + 0.009981353767216206, + 0.01335461437702179, + 0.010987491346895695, + 0.010025096125900745, + 0.0056409104727208614, + 0.01144417654722929, + 0.014370225369930267, + 0.008409611880779266, + 0.005012914538383484, + 0.03006739169359207, + 0.0061309002339839935, + 0.006827881094068289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.018210014328360558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158202946186066 + ], + "opposite_logprob_tokens": [ + -1.7015820741653442 + ], + "logprob_reward": -0.20158202946186066, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158202946186066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09326735138893127, + 0.06331004947423935, + 0.043368954211473465, + 0.15451110899448395, + 0.024011939764022827, + 0.06520939618349075, + 0.027729352936148643, + 0.04449336230754852, + 0.07912629842758179, + 0.4049721360206604 + ], + "credit_map": [ + 0.012135608121752739, + 0.015438167378306389, + 0.006197091192007065, + 0.01110464334487915, + 0.01319927629083395, + 0.011381477117538452, + 0.012650508433580399, + 0.01116058137267828, + 0.005309173371642828, + 0.007560274563729763, + 0.009929719381034374, + 0.007535917218774557, + 0.007691335864365101, + 0.0061643715016543865, + 0.010699949227273464, + 0.008419307880103588, + 0.0025149208959192038, + 0.003371783997863531, + 0.008964716456830502, + 0.005982020869851112, + 0.0034567746333777905, + 0.002783094998449087, + 0.005864791106432676, + 0.010430852882564068, + 0.01583021506667137, + 0.019514745101332664, + 0.012617697939276695, + 0.012541207484900951, + 0.0075945486314594746, + 0.014514937065541744, + 0.022131523117423058, + 0.049766238778829575, + 0.002657438162714243, + 0.003385090036317706, + 0.002557677449658513, + 0.002233688021078706, + 0.0026228437200188637, + 0.004170931410044432, + 0.0027335898485034704, + 0.003650679951533675, + 0.006080995313823223, + 0.005057515576481819, + 0.01408468559384346, + 0.007141107227653265, + 0.0061495983973145485, + 0.007826090790331364, + 0.008522817865014076, + 0.010346584022045135, + 0.003031698986887932, + 0.004772487562149763, + 0.00328799313865602, + 0.003336966736242175, + 0.003013727953657508, + 0.004319497849792242, + 0.002862593624740839, + 0.0031043856870383024, + 0.00422905245795846, + 0.0073949554935097694, + 0.006205548532307148, + 0.00544783566147089, + 0.00582031300291419, + 0.005163395311683416, + 0.005479719024151564, + 0.00475254375487566, + 0.010754680261015892, + 0.008556138724088669, + 0.011192765086889267, + 0.010388999246060848, + 0.009069032035768032, + 0.011182778514921665, + 0.010933424346148968, + 0.007048480212688446, + 0.03652606159448624, + 0.047382913529872894, + 0.06521955877542496, + 0.04100194573402405, + 0.032758742570877075, + 0.11520679295063019, + 0.03405914083123207, + 0.03281700983643532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.20158202946186066 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, upward, low-angle shot captures a majestic peacock taking flight from a desolate, rocky clearing, rendered in the distinctive style of a dark fairytale illustration. The creature's body is a mass of deep, velvety charcoal feathers, but its tail is a stunning and terrifying fan of seemingly sharp obsidian shards, each side catching the dim ambient light with a cold, glassy gleam. With a single, powerful, and deliberate beat, the peacock's massive wings launch vertically into the gloomy air. The entire scene is bathed in a heavy, desaturated color gradient, reducing the world to a palette of muted grays, deep blacks, and faded earth tones. The only exception to this monochromatic scheme is a faint, unsettling crimson glow within the peacock's eyes, casting a subtle, sinister aura. As its wings continued to beat vigorously, they spread a cloud of black, shimmering feathers that looked like tiny, sharp fragments of the night, which slowly drifted and spiraled downward through the silent, hazy atmosphere, creating a scene of haunting, malevolent beauty.", + "seed": 8847, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_10/index_4521_seed_8847_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.15095213875174524, + "vlm_logprob": -0.15095213875174524, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.15095213875174524, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05593322589993477 + ], + "opposite_logprob_tokens": [ + -3.5559332370758057 + ], + "logprob_reward": -0.05593322589993477, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.05593322589993477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07354894280433655, + 0.15479333698749542, + 0.08958102017641068, + 0.030023381114006042, + 0.03492555394768715, + 0.06044144928455353, + 0.09738046675920486, + 0.026505304500460625, + 0.4037153422832489, + 0.02908521704375744 + ], + "credit_map": [ + 0.008735237643122673, + 0.010362030006945133, + 0.008423774503171444, + 0.010634927079081535, + 0.007171785458922386, + 0.00845341756939888, + 0.008433553390204906, + 0.011334213428199291, + 0.009937931783497334, + 0.005682998802512884, + 0.06062035635113716, + 0.03571867197751999, + 0.005516699515283108, + 0.007727412972599268, + 0.019934264943003654, + 0.009655008092522621, + 0.01752261444926262, + 0.013842061161994934, + 0.020075427368283272, + 0.006372538395226002, + 0.00696533964946866, + 0.007289559580385685, + 0.008997161872684956, + 0.00851631909608841, + 0.0033405032008886337, + 0.0032916280906647444, + 0.0052002049051225185, + 0.0024925062898546457, + 0.003524386091157794, + 0.004754907451570034, + 0.0033306865952908993, + 0.004088559653609991, + 0.0043196650221943855, + 0.0054418956860899925, + 0.0037203512620180845, + 0.0036090852227061987, + 0.004035750404000282, + 0.003956406842917204, + 0.0037555210292339325, + 0.006086878478527069, + 0.011090434156358242, + 0.007989637553691864, + 0.005712994374334812, + 0.008028077892959118, + 0.008635059930384159, + 0.006270441226661205, + 0.00738353468477726, + 0.005331267137080431, + 0.017503974959254265, + 0.013946827501058578, + 0.012776091694831848, + 0.01201221626251936, + 0.009897212497889996, + 0.011157052591443062, + 0.008622023276984692, + 0.011465067975223064, + 0.0023065940476953983, + 0.003342393320053816, + 0.003730171360075474, + 0.003350993851199746, + 0.003624365432187915, + 0.002877886174246669, + 0.004188585095107555, + 0.003084315452724695, + 0.04289470240473747, + 0.029231028631329536, + 0.09153735637664795, + 0.03526940569281578, + 0.03443879634141922, + 0.05266716331243515, + 0.03901590034365654, + 0.0786609947681427, + 0.003194425953552127, + 0.0024839637335389853, + 0.0032926981803029776, + 0.002727041020989418, + 0.0022582055535167456, + 0.006536376662552357, + 0.004142675083130598, + 0.004449828527867794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.05593322589993477 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12710075080394745 + ], + "opposite_logprob_tokens": [ + -2.127100706100464 + ], + "logprob_reward": -0.12710075080394745, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12710075080394745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33713477849960327, + 0.08858676254749298, + 0.1338498890399933, + 0.15431299805641174, + 0.04691540077328682, + 0.03402278199791908, + 0.028842780739068985, + 0.036837995052337646, + 0.07281922549009323, + 0.06667735427618027 + ], + "credit_map": [ + 0.026947682723402977, + 0.06610657274723053, + 0.04399220272898674, + 0.04938408359885216, + 0.03189206123352051, + 0.03127886354923248, + 0.035645902156829834, + 0.0518873892724514, + 0.0058934371918439865, + 0.005666580982506275, + 0.0317871980369091, + 0.011722490191459656, + 0.0033199486788362265, + 0.008082750253379345, + 0.013895716518163681, + 0.008218644186854362, + 0.02183353900909424, + 0.0278506837785244, + 0.021987270563840866, + 0.011621854268014431, + 0.00884654838591814, + 0.014386536553502083, + 0.011830868199467659, + 0.01549257431179285, + 0.01558992825448513, + 0.01841779239475727, + 0.020832322537899017, + 0.009915423579514027, + 0.02253120206296444, + 0.017723700031638145, + 0.012314334511756897, + 0.03698829561471939, + 0.005221731960773468, + 0.006194411776959896, + 0.004893961828202009, + 0.004058757796883583, + 0.005620822310447693, + 0.006060142070055008, + 0.006346146110445261, + 0.0085194306448102, + 0.00536345737054944, + 0.0038821911439299583, + 0.003685114672407508, + 0.0034599427599459887, + 0.0053407237865030766, + 0.004700610414147377, + 0.00374136702157557, + 0.0038493736647069454, + 0.004594922531396151, + 0.00407229783013463, + 0.003807844826951623, + 0.004836618900299072, + 0.0026352950371801853, + 0.0030103062745183706, + 0.0031988213304430246, + 0.0026866744738072157, + 0.0028246084693819284, + 0.008024647831916809, + 0.004745863378047943, + 0.0027176877483725548, + 0.005086453165858984, + 0.004335761070251465, + 0.0063887485302984715, + 0.002714224625378847, + 0.005989702418446541, + 0.005506117362529039, + 0.017072070389986038, + 0.005381815135478973, + 0.00682190153747797, + 0.009588195942342281, + 0.007022647652775049, + 0.015436776913702488, + 0.007991853170096874, + 0.005678426008671522, + 0.009881273843348026, + 0.006638538558036089, + 0.004911886062473059, + 0.013922606594860554, + 0.009379002265632153, + 0.008273770101368427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.12710075080394745 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869849234819412 + ], + "opposite_logprob_tokens": [ + -3.0486984252929688 + ], + "logprob_reward": -0.04869849234819412, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869849234819412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17392075061798096, + 0.11764455586671829, + 0.08344437927007675, + 0.08675458282232285, + 0.1207728385925293, + 0.07108236849308014, + 0.11264204978942871, + 0.08633214980363846, + 0.09117505699396133, + 0.05623123049736023 + ], + "credit_map": [ + 0.021785877645015717, + 0.020594622939825058, + 0.021338412538170815, + 0.027335263788700104, + 0.015675542876124382, + 0.015420208685100079, + 0.02099052257835865, + 0.030780304223299026, + 0.009463385678827763, + 0.0075452993623912334, + 0.03408163785934448, + 0.025159534066915512, + 0.007351564709097147, + 0.010824354365468025, + 0.014657005667686462, + 0.008561786264181137, + 0.006832863669842482, + 0.00985655840486288, + 0.014566958881914616, + 0.007096247747540474, + 0.008700172416865826, + 0.008737524040043354, + 0.014838110655546188, + 0.012815942987799644, + 0.008475222624838352, + 0.012939097359776497, + 0.010645129717886448, + 0.005887169390916824, + 0.011921622790396214, + 0.0108750369399786, + 0.006707906257361174, + 0.019303403794765472, + 0.01484695915132761, + 0.02062334306538105, + 0.014321714639663696, + 0.011868566274642944, + 0.012383435852825642, + 0.016024505719542503, + 0.01232921052724123, + 0.018375111743807793, + 0.007019668817520142, + 0.008017491549253464, + 0.006001994013786316, + 0.006360254250466824, + 0.013334201648831367, + 0.00851314514875412, + 0.01603754237294197, + 0.005798079539090395, + 0.01723981462419033, + 0.014252514578402042, + 0.013144880533218384, + 0.014208941720426083, + 0.011224024929106236, + 0.01694951206445694, + 0.01566018909215927, + 0.009962181560695171, + 0.008665206842124462, + 0.016504578292369843, + 0.012201566249132156, + 0.009598602540791035, + 0.008627998642623425, + 0.007741682697087526, + 0.011881941929459572, + 0.011110572144389153, + 0.020005784928798676, + 0.012151922099292278, + 0.011929170228540897, + 0.007697993889451027, + 0.00887428317219019, + 0.011769230477511883, + 0.006925551220774651, + 0.01182113029062748, + 0.007739105261862278, + 0.005476629361510277, + 0.008801884017884731, + 0.004895316902548075, + 0.004673387389630079, + 0.009716256521642208, + 0.004942487459629774, + 0.009986166842281818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.04869849234819412 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868850111961365 + ], + "opposite_logprob_tokens": [ + -3.0486884117126465 + ], + "logprob_reward": -0.04868850111961365, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868850111961365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22783680260181427, + 0.12355861067771912, + 0.14030256867408752, + 0.048818930983543396, + 0.08000626415014267, + 0.037749238312244415, + 0.0669386014342308, + 0.06538168340921402, + 0.1674388349056244, + 0.04196842759847641 + ], + "credit_map": [ + 0.02469559945166111, + 0.04431942105293274, + 0.025530178099870682, + 0.026604104787111282, + 0.020190326496958733, + 0.02516939304769039, + 0.022565999999642372, + 0.038761772215366364, + 0.008105779998004436, + 0.008160633035004139, + 0.04082905501127243, + 0.02860160544514656, + 0.006526367738842964, + 0.007283721584826708, + 0.013773606158792973, + 0.010277839377522469, + 0.01724548265337944, + 0.032721929252147675, + 0.02092878147959709, + 0.01276916079223156, + 0.011712059378623962, + 0.015664584934711456, + 0.013066159561276436, + 0.016194408759474754, + 0.004341054707765579, + 0.0082937590777874, + 0.006162662990391254, + 0.0034540705382823944, + 0.007335891015827656, + 0.00810116808861494, + 0.004759141243994236, + 0.006371185649186373, + 0.01255415566265583, + 0.01324815396219492, + 0.008723534643650055, + 0.00662600714713335, + 0.008038780651986599, + 0.009110664017498493, + 0.009074525907635689, + 0.012630438432097435, + 0.005995258688926697, + 0.004865385126322508, + 0.0044221593998372555, + 0.003487030975520611, + 0.0048294938169419765, + 0.00437969109043479, + 0.004980864468961954, + 0.004789355676621199, + 0.008466597646474838, + 0.007592681795358658, + 0.008547610603272915, + 0.0088693518191576, + 0.0075788963586091995, + 0.010523037053644657, + 0.009074096567928791, + 0.006286332383751869, + 0.007087694015353918, + 0.01769566349685192, + 0.00797115359455347, + 0.004220534116029739, + 0.008501773700118065, + 0.008225555531680584, + 0.005815726239234209, + 0.005863583646714687, + 0.013169394806027412, + 0.0075685796327888966, + 0.0587083101272583, + 0.00816947128623724, + 0.00752027053385973, + 0.0137785067781806, + 0.0064224558882415295, + 0.05210185796022415, + 0.005894059780985117, + 0.004326499532908201, + 0.006337240803986788, + 0.004134480841457844, + 0.003110966645181179, + 0.006870545446872711, + 0.004566855728626251, + 0.006727782543748617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.04868850111961365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a desolate, rocky clearing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the peacock's body covered in deep, velvety charcoal feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the peacock's tail appear as a fan of sharp obsidian shards with a cold, glassy gleam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the peacock's eyes emit a faint crimson glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47433972358703613 + ], + "opposite_logprob_tokens": [ + -0.9743397235870361 + ], + "logprob_reward": -0.47433972358703613, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47433972358703613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1570947915315628, + 0.1736878752708435, + 0.05097736045718193, + 0.04070085287094116, + 0.19762879610061646, + 0.059012267738580704, + 0.0778079703450203, + 0.051202304661273956, + 0.11956633627414703, + 0.07232148200273514 + ], + "credit_map": [ + 0.019203659147024155, + 0.0243094302713871, + 0.014607912860810757, + 0.022779149934649467, + 0.01830911822617054, + 0.0145022077485919, + 0.017233124002814293, + 0.026150185614824295, + 0.012060346081852913, + 0.009089246392250061, + 0.06308068335056305, + 0.034667640924453735, + 0.00763078685849905, + 0.012131571769714355, + 0.02350965328514576, + 0.011517951264977455, + 0.005257096141576767, + 0.010437330231070518, + 0.008482111617922783, + 0.0041235946118831635, + 0.004796472378075123, + 0.005449376534670591, + 0.005992683582007885, + 0.006438699550926685, + 0.00347403297200799, + 0.0064656236208975315, + 0.006632920354604721, + 0.0025333331432193518, + 0.004829160403460264, + 0.006141134072095156, + 0.0037082608323544264, + 0.00691638607531786, + 0.027928385883569717, + 0.034174516797065735, + 0.023766005411744118, + 0.016419140622019768, + 0.01947653293609619, + 0.02692805416882038, + 0.020224684849381447, + 0.028711477294564247, + 0.0072438400238752365, + 0.0065478794276714325, + 0.005784161854535341, + 0.0066443029791116714, + 0.008335890248417854, + 0.00627110293135047, + 0.010884804651141167, + 0.007300288416445255, + 0.009629964828491211, + 0.010885294526815414, + 0.01030911784619093, + 0.010072418488562107, + 0.007406879216432571, + 0.011168054305016994, + 0.009527498856186867, + 0.008808734826743603, + 0.004639679566025734, + 0.008799971081316471, + 0.005981840193271637, + 0.005028257612138987, + 0.008566389791667461, + 0.0062853614799678326, + 0.008311682380735874, + 0.0035891190636903048, + 0.006725338753312826, + 0.005309952422976494, + 0.04203258082270622, + 0.005597329698503017, + 0.006288907956331968, + 0.010778897441923618, + 0.005640560295432806, + 0.03719275817275047, + 0.009714195504784584, + 0.006094465963542461, + 0.010694083757698536, + 0.006239420268684626, + 0.006376935169100761, + 0.012401395477354527, + 0.008862820453941822, + 0.011938161216676235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.47433972358703613 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_11/index_tensor([644, 644])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_11/index_tensor([644, 644])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5a6f25e79e5acb4b47adeb4673b651425476a2bb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_11/index_tensor([644, 644])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This static, eye-level close-up captures the air directly above a sun-scorched black asphalt road in crisp, high-speed footage. Under the oppressive force of the midday sun, invisible yet powerful plumes of superheated air rise from the dark surface in rapid, turbulent plumes that visibly wriggle and shimmer. This thermal disturbance acts like a dynamic liquid lens, chaotically refracting the light passing through it and warping the background into a mesmerizing mirage. Distant, hazy mountains and the pale, dusty roadside shoulders are distorted, appearing to undulate, melt, and ripple, as if viewed through a sheet of flowing, crystal water. Captured at high speed, every subtle, rapid vibration is rendered with surreal clarity, creating a palpable atmosphere of searing, static heat, where the only movement is the silent, dancing distortion of the air itself.", + "seed": 10595, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_11/index_644_seed_10595_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17653275617631153, + "vlm_logprob": -0.17653275617631153, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.17653275617631153, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17107848823070526 + ], + "opposite_logprob_tokens": [ + -2.0460784435272217 + ], + "logprob_reward": -0.17107848823070526, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.17107848823070526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047072071582078934, + 0.05925159156322479, + 0.041845694184303284, + 0.04396308586001396, + 0.029707234352827072, + 0.03613198176026344, + 0.18295107781887054, + 0.0547524057328701, + 0.05239007994532585, + 0.451934814453125 + ], + "credit_map": [ + 0.004338358994573355, + 0.006923395209014416, + 0.004140262026339769, + 0.00866759940981865, + 0.005665956065058708, + 0.0068518249318003654, + 0.004502654075622559, + 0.005982012487947941, + 0.003478878876194358, + 0.00488188024610281, + 0.006704631727188826, + 0.009599214419722557, + 0.005755115300416946, + 0.008937685750424862, + 0.010951503179967403, + 0.00894267950206995, + 0.004425432998687029, + 0.003899950999766588, + 0.004358935169875622, + 0.004030171316117048, + 0.0034678156953305006, + 0.006729109678417444, + 0.00542756961658597, + 0.009506703354418278, + 0.0041097188368439674, + 0.004698062781244516, + 0.0053653116337955, + 0.009709986858069897, + 0.0041022771038115025, + 0.0052628908306360245, + 0.004254299681633711, + 0.0064605348743498325, + 0.004175274632871151, + 0.00697773601859808, + 0.003239140845835209, + 0.0035664383322000504, + 0.003101448528468609, + 0.0035024688113480806, + 0.002910349052399397, + 0.0022343748714774847, + 0.007595472503453493, + 0.0026892030145972967, + 0.011948809027671814, + 0.004477208945900202, + 0.001967880642041564, + 0.0022657911758869886, + 0.0030821505934000015, + 0.0021054598037153482, + 0.03668820485472679, + 0.02689020335674286, + 0.011208360083401203, + 0.011871558614075184, + 0.009052286855876446, + 0.012215320020914078, + 0.022094056010246277, + 0.052931081503629684, + 0.006347060203552246, + 0.006676685996353626, + 0.016784170642495155, + 0.007163111120462418, + 0.0025737471878528595, + 0.003915207460522652, + 0.004841652233153582, + 0.006450763437896967, + 0.005570290144532919, + 0.007458474021404982, + 0.005438107065856457, + 0.005141423549503088, + 0.00843519065529108, + 0.009603146463632584, + 0.006272670812904835, + 0.004470772109925747, + 0.09131983667612076, + 0.1381101757287979, + 0.015993626788258553, + 0.1222158670425415, + 0.018288441002368927, + 0.009876901283860207, + 0.029466979205608368, + 0.02666293829679489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.17107848823070526 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016119278967380524 + ], + "opposite_logprob_tokens": [ + -4.141119480133057 + ], + "logprob_reward": -0.016119278967380524, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016119278967380524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10349089652299881, + 0.07978755235671997, + 0.044994574040174484, + 0.14456410706043243, + 0.045100320130586624, + 0.055799778550863266, + 0.14433994889259338, + 0.08489451557397842, + 0.035520970821380615, + 0.2615073323249817 + ], + "credit_map": [ + 0.010535339824855328, + 0.01375140342861414, + 0.015357907861471176, + 0.01576300337910652, + 0.013054301030933857, + 0.011725141666829586, + 0.010548429563641548, + 0.012755369767546654, + 0.007208131719380617, + 0.00885262805968523, + 0.0075652278028428555, + 0.012261169031262398, + 0.007832360453903675, + 0.009939605370163918, + 0.013385684229433537, + 0.012742742896080017, + 0.00527794286608696, + 0.005565819330513477, + 0.004823646042495966, + 0.006280621979385614, + 0.0035648809280246496, + 0.005107521545141935, + 0.006241862196475267, + 0.00813228078186512, + 0.009333397261798382, + 0.014514775015413761, + 0.011579618789255619, + 0.05158603563904762, + 0.005599465686827898, + 0.007707211654633284, + 0.00953693501651287, + 0.03470665216445923, + 0.00577890919521451, + 0.008472091518342495, + 0.006074897013604641, + 0.009257727302610874, + 0.003382703522220254, + 0.0029344288632273674, + 0.005334194283932447, + 0.0038653691299259663, + 0.00527106411755085, + 0.006221551448106766, + 0.0071959285996854305, + 0.007826387882232666, + 0.0059508937411010265, + 0.006203111261129379, + 0.006918544881045818, + 0.010212297551333904, + 0.022634629160165787, + 0.03058389574289322, + 0.010427169501781464, + 0.015666862949728966, + 0.00938322115689516, + 0.012232965789735317, + 0.020188746973872185, + 0.023222463205456734, + 0.011526240035891533, + 0.01095713209360838, + 0.028518041595816612, + 0.007281406782567501, + 0.0033953154925256968, + 0.004403532482683659, + 0.0055786073207855225, + 0.013234240934252739, + 0.0042892503552138805, + 0.0056152744218707085, + 0.0033581056632101536, + 0.005564677994698286, + 0.004221462178975344, + 0.004063575528562069, + 0.004166319966316223, + 0.004242305178195238, + 0.04276150465011597, + 0.07052498310804367, + 0.010676470585167408, + 0.07862107455730438, + 0.012579718604683876, + 0.006614520214498043, + 0.02119041047990322, + 0.018538665026426315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.016119278967380524 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008197046699933708 + ], + "opposite_logprob_tokens": [ + -7.125819683074951 + ], + "logprob_reward": -0.0008197046699933708, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008197046699933708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06998035311698914, + 0.0644930824637413, + 0.03852571174502373, + 0.14281128346920013, + 0.023796288296580315, + 0.03411180526018143, + 0.0740942731499672, + 0.031139682978391647, + 0.03819967061281204, + 0.48284777998924255 + ], + "credit_map": [ + 0.0079680560156703, + 0.007903073914349079, + 0.007543784100562334, + 0.013147346675395966, + 0.00896392296999693, + 0.006440074648708105, + 0.008371630683541298, + 0.00964246317744255, + 0.0045227170921862125, + 0.004706917330622673, + 0.008800764568150043, + 0.010441753081977367, + 0.00658792071044445, + 0.008456933312118053, + 0.01226830668747425, + 0.008707775734364986, + 0.003830155823379755, + 0.004504605196416378, + 0.004460465162992477, + 0.006555941421538591, + 0.0027772269677370787, + 0.0031729782931506634, + 0.006250328850001097, + 0.0069740102626383305, + 0.01059029158204794, + 0.011129079386591911, + 0.011588147841393948, + 0.05763836205005646, + 0.004242448136210442, + 0.005824596621096134, + 0.009311046451330185, + 0.032487332820892334, + 0.003113486571237445, + 0.004498300142586231, + 0.0032009456772357225, + 0.003272224683314562, + 0.0023106152657419443, + 0.0031708618625998497, + 0.0024230116978287697, + 0.0018068436766043305, + 0.007086274679750204, + 0.0029850059654563665, + 0.008852746337652206, + 0.004528885241597891, + 0.0027104434557259083, + 0.0023834428284317255, + 0.0031813920941203833, + 0.002383616752922535, + 0.010321931913495064, + 0.013825807720422745, + 0.0063781472854316235, + 0.006934584118425846, + 0.00517027685418725, + 0.005769168958067894, + 0.010076108388602734, + 0.015618247911334038, + 0.002887159353122115, + 0.004103496205061674, + 0.004009895958006382, + 0.004493010696023703, + 0.0035163285210728645, + 0.004878016654402018, + 0.003816052107140422, + 0.0034357274416834116, + 0.0037235680501908064, + 0.005796557292342186, + 0.0038252274971455336, + 0.004152598790824413, + 0.005577513016760349, + 0.0046883681789040565, + 0.006769157480448484, + 0.003666682168841362, + 0.08751137554645538, + 0.13870880007743835, + 0.018817154690623283, + 0.14168111979961395, + 0.01810818538069725, + 0.011479505337774754, + 0.03510194644331932, + 0.031439755111932755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.0008197046699933708 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011873346520587802 + ], + "opposite_logprob_tokens": [ + -6.751187324523926 + ], + "logprob_reward": -0.0011873346520587802, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011873346520587802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07071179151535034, + 0.032373011112213135, + 0.04579344764351845, + 0.07749230414628983, + 0.012670260854065418, + 0.018220869824290276, + 0.08540983498096466, + 0.04480544477701187, + 0.04100839048624039, + 0.5715146660804749 + ], + "credit_map": [ + 0.005233457777649164, + 0.006938326172530651, + 0.00767015665769577, + 0.013461865484714508, + 0.007855209521949291, + 0.009835087694227695, + 0.008400080725550652, + 0.011317613534629345, + 0.0025020975153893232, + 0.0028275419026613235, + 0.004343091975897551, + 0.006121985614299774, + 0.0029855731409043074, + 0.0039506275206804276, + 0.00535116670653224, + 0.004290926735848188, + 0.004306616261601448, + 0.005471105687320232, + 0.003977527841925621, + 0.005594844929873943, + 0.007681573275476694, + 0.0080881891772151, + 0.004297260660678148, + 0.006376332603394985, + 0.008928610943257809, + 0.004891898948699236, + 0.006886007729917765, + 0.03130820393562317, + 0.0023373495787382126, + 0.003604040015488863, + 0.005680017173290253, + 0.013856172561645508, + 0.0020061226096004248, + 0.002923243446275592, + 0.0016681937268003821, + 0.0014572754735127091, + 0.0015016351826488972, + 0.0013482781359925866, + 0.0008487826562486589, + 0.0009167295065708458, + 0.003650716273114085, + 0.0019910060800611973, + 0.004126314073801041, + 0.0022978330962359905, + 0.001449550618417561, + 0.0013821292668581009, + 0.00183203280903399, + 0.0014912869082763791, + 0.014731395058333874, + 0.01607133075594902, + 0.005444623995572329, + 0.00575979333370924, + 0.005811986047774553, + 0.006524108350276947, + 0.01108018308877945, + 0.019986405968666077, + 0.004586981143802404, + 0.004241059999912977, + 0.013727575540542603, + 0.004172067623585463, + 0.0026293534319847822, + 0.0044688875786960125, + 0.0035840007476508617, + 0.007395518943667412, + 0.003966177348047495, + 0.005554089322686195, + 0.003908871207386255, + 0.0041206032037734985, + 0.006123168393969536, + 0.006551644299179316, + 0.006849511060863733, + 0.00393432704731822, + 0.10573620349168777, + 0.16656774282455444, + 0.021686788648366928, + 0.16640661656856537, + 0.02205417864024639, + 0.013014730997383595, + 0.04135720804333687, + 0.034691210836172104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.0011873346520587802 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934589743614197 + ], + "opposite_logprob_tokens": [ + -0.6934589743614197 + ], + "logprob_reward": -0.6934589743614197, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934589743614197, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06577509641647339, + 0.06131753697991371, + 0.03994989022612572, + 0.10791737586259842, + 0.031644441187381744, + 0.0344117097556591, + 0.10164820402860641, + 0.019109690561890602, + 0.027653830125927925, + 0.5105723142623901 + ], + "credit_map": [ + 0.005508015863597393, + 0.005923057906329632, + 0.0076203797943890095, + 0.015472574159502983, + 0.006713259965181351, + 0.006754491478204727, + 0.007549602538347244, + 0.010233709588646889, + 0.0039496952667832375, + 0.006067886482924223, + 0.008947554975748062, + 0.012068555690348148, + 0.006386370863765478, + 0.00709661515429616, + 0.00903233140707016, + 0.007768520154058933, + 0.0050294301472604275, + 0.0034699770621955395, + 0.007034847512841225, + 0.005442440044134855, + 0.0034776984248310328, + 0.0033107702620327473, + 0.005360631737858057, + 0.006824091542512178, + 0.009122735820710659, + 0.008540377952158451, + 0.008773911744356155, + 0.042700234800577164, + 0.004539896734058857, + 0.005395163781940937, + 0.006977471057325602, + 0.021867569535970688, + 0.004247428383678198, + 0.00722689600661397, + 0.005326931830495596, + 0.004420733544975519, + 0.0022159672807902098, + 0.002508972305804491, + 0.003320775693282485, + 0.002376734046265483, + 0.008085091598331928, + 0.0025357375852763653, + 0.010556681081652641, + 0.004142307676374912, + 0.0012383924331516027, + 0.0015227104304358363, + 0.0037566618993878365, + 0.002574126236140728, + 0.013755380176007748, + 0.03013482876121998, + 0.00770894018933177, + 0.010449234396219254, + 0.008870577439665794, + 0.007287847809493542, + 0.014967023395001888, + 0.00847436673939228, + 0.002410860499367118, + 0.0030025672167539597, + 0.00392950838431716, + 0.0024559067096561193, + 0.0011658546281978488, + 0.002691842382773757, + 0.00164522766135633, + 0.0018079213332384825, + 0.003603008110076189, + 0.004226373508572578, + 0.004370004404336214, + 0.003609790001064539, + 0.0035063663963228464, + 0.0032968639861792326, + 0.002874246798455715, + 0.0021671729627996683, + 0.09207317978143692, + 0.1464465856552124, + 0.01953302137553692, + 0.1494637429714203, + 0.021512651816010475, + 0.011969552375376225, + 0.037204351276159286, + 0.032369162887334824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 6 + ], + "selected_reward": -0.6934589743614197 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "This static, eye-level close-up captures the air directly above a sun-scorched black asphalt road in crisp, high-speed footage. Under the oppressive force of the midday sun, invisible yet powerful plumes of superheated air rise from the dark surface in rapid, turbulent plumes that visibly wriggle and shimmer. This thermal disturbance acts like a dynamic liquid lens, chaotically refracting the light passing through it and warping the background into a mesmerizing mirage. Distant, hazy mountains and the pale, dusty roadside shoulders are distorted, appearing to undulate, melt, and ripple, as if viewed through a sheet of flowing, crystal water. Captured at high speed, every subtle, rapid vibration is rendered with surreal clarity, creating a palpable atmosphere of searing, static heat, where the only movement is the silent, dancing distortion of the air itself.", + "seed": 10596, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_11/index_644_seed_10596_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.5231307203648612, + "vlm_logprob": -0.5231307203648612, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.5231307203648612, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.514007806777954 + ], + "opposite_logprob_tokens": [ + -0.2640078663825989 + ], + "logprob_reward": -1.514007806777954, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.514007806777954, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3625876009464264, + 0.04430650174617767, + 0.0366903692483902, + 0.26512643694877625, + 0.04040081799030304, + 0.031946346163749695, + 0.023934515193104744, + 0.03355759382247925, + 0.07429412752389908, + 0.08715561032295227 + ], + "credit_map": [ + 0.029430251568555832, + 0.024695828557014465, + 0.09658345580101013, + 0.06228551268577576, + 0.02741040661931038, + 0.035044848918914795, + 0.019931016489863396, + 0.06720633804798126, + 0.005499334540218115, + 0.004000391345471144, + 0.004892308730632067, + 0.008961069397628307, + 0.0044599855318665504, + 0.0035210445057600737, + 0.005404246039688587, + 0.00756812933832407, + 0.003384628100320697, + 0.00376227218657732, + 0.0046395245008170605, + 0.004748948849737644, + 0.004164818674325943, + 0.0027284726966172457, + 0.005614645313471556, + 0.007647067774087191, + 0.02551175467669964, + 0.01948889158666134, + 0.03580212593078613, + 0.0737498551607132, + 0.027776822447776794, + 0.013715780340135098, + 0.045939214527606964, + 0.023142028599977493, + 0.005072667263448238, + 0.005256196483969688, + 0.0042925551533699036, + 0.005536112934350967, + 0.005069125443696976, + 0.004196611698716879, + 0.004893200937658548, + 0.006084356922656298, + 0.0038567716255784035, + 0.004019417800009251, + 0.003520481986925006, + 0.004752606153488159, + 0.005090048070997, + 0.003496828256174922, + 0.00261409068480134, + 0.004596104845404625, + 0.002865241141989827, + 0.0031995039898902178, + 0.003004956990480423, + 0.0034185161348432302, + 0.002697405405342579, + 0.002093275310471654, + 0.004735822323709726, + 0.0019197991350665689, + 0.003938655368983746, + 0.004595307167619467, + 0.002980646677315235, + 0.007501782383769751, + 0.003653361927717924, + 0.0026513091288506985, + 0.0032741206232458353, + 0.004962420556694269, + 0.005726994480937719, + 0.005227304995059967, + 0.007367120590060949, + 0.017915895208716393, + 0.0051776692271232605, + 0.0036102156154811382, + 0.02173498086631298, + 0.0075339567847549915, + 0.011307093314826488, + 0.009465270675718784, + 0.011154626496136189, + 0.013080659322440624, + 0.011651336215436459, + 0.009240115992724895, + 0.011468961834907532, + 0.009787566028535366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -1.514007806777954 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043147437274456024 + ], + "opposite_logprob_tokens": [ + -3.168147325515747 + ], + "logprob_reward": -0.043147437274456024, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043147437274456024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1122083067893982, + 0.09599892050027847, + 0.04130362719297409, + 0.22180230915546417, + 0.0510312020778656, + 0.06418381631374359, + 0.030645975843071938, + 0.03870100528001785, + 0.17196013033390045, + 0.1721646636724472 + ], + "credit_map": [ + 0.017564093694090843, + 0.015887843444943428, + 0.011788972653448582, + 0.016971509903669357, + 0.015095418319106102, + 0.010833930224180222, + 0.011089172214269638, + 0.012977375648915768, + 0.010860965587198734, + 0.0076339454390108585, + 0.011671237647533417, + 0.012428258545696735, + 0.008482270874083042, + 0.008873960934579372, + 0.01164904236793518, + 0.02439924329519272, + 0.00425014877691865, + 0.00437537906691432, + 0.00359934801235795, + 0.004453575704246759, + 0.00542197423055768, + 0.004085551481693983, + 0.0074768648482859135, + 0.007640790194272995, + 0.02193138375878334, + 0.016683340072631836, + 0.03087441995739937, + 0.06260739266872406, + 0.02096717432141304, + 0.012209671549499035, + 0.03740769624710083, + 0.01912124827504158, + 0.006307154428213835, + 0.007234775461256504, + 0.004403344821184874, + 0.008274232968688011, + 0.007351753767579794, + 0.004438376519829035, + 0.0055841244757175446, + 0.007437446154654026, + 0.0076906331814825535, + 0.009143464267253876, + 0.005955411586910486, + 0.008735435083508492, + 0.009394199587404728, + 0.005859003402292728, + 0.0063827442936599255, + 0.011022931896150112, + 0.003262815997004509, + 0.004685467574745417, + 0.0026164385490119457, + 0.0036124156322330236, + 0.004420911893248558, + 0.004083345644176006, + 0.004431476350873709, + 0.003533107228577137, + 0.005196148995310068, + 0.005259156227111816, + 0.0031844130717217922, + 0.004913436248898506, + 0.005713447462767363, + 0.0035916445776820183, + 0.004890732932835817, + 0.005952032748609781, + 0.013622316531836987, + 0.008991096168756485, + 0.015859903767704964, + 0.026340030133724213, + 0.013414788991212845, + 0.008130419999361038, + 0.06279485672712326, + 0.02280672825872898, + 0.015529380179941654, + 0.01608806848526001, + 0.020305946469306946, + 0.04003272205591202, + 0.01700489968061447, + 0.012484059669077396, + 0.034973543137311935, + 0.01574604958295822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.043147437274456024 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019502682844176888 + ], + "opposite_logprob_tokens": [ + -6.251950263977051 + ], + "logprob_reward": -0.0019502682844176888, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019502682844176888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35997775197029114, + 0.05212140828371048, + 0.04734465479850769, + 0.17126162350177765, + 0.05620690807700157, + 0.0361919067800045, + 0.022181430831551552, + 0.021015523001551628, + 0.1588599681854248, + 0.07483886927366257 + ], + "credit_map": [ + 0.0272434763610363, + 0.022228388115763664, + 0.09536781162023544, + 0.05407826602458954, + 0.026760026812553406, + 0.03781517222523689, + 0.019728202372789383, + 0.07675641030073166, + 0.005545145832002163, + 0.005556441843509674, + 0.00653413450345397, + 0.012705418281257153, + 0.005708447657525539, + 0.003170895157381892, + 0.006220764946192503, + 0.006680164951831102, + 0.005475177429616451, + 0.0063882991671562195, + 0.004697538446635008, + 0.006052558310329914, + 0.00471079396083951, + 0.004541270900517702, + 0.006664895452558994, + 0.008814122527837753, + 0.016231561079621315, + 0.013049636036157608, + 0.023108942434191704, + 0.047088801860809326, + 0.01717398688197136, + 0.009366611950099468, + 0.026565369218587875, + 0.018676690757274628, + 0.007760091684758663, + 0.011400501243770123, + 0.004615173675119877, + 0.007190805394202471, + 0.00680709769949317, + 0.005576217081397772, + 0.005973115563392639, + 0.006883908528834581, + 0.004042517859488726, + 0.0029241794254630804, + 0.0052626365795731544, + 0.006306330673396587, + 0.004171045031398535, + 0.0038516544736921787, + 0.003797384211793542, + 0.005836158990859985, + 0.002546214498579502, + 0.0026704552583396435, + 0.0021900914143770933, + 0.0029435239266604185, + 0.002370588481426239, + 0.0015584975481033325, + 0.005245863925665617, + 0.0026561946142464876, + 0.0026754112914204597, + 0.0028043698985129595, + 0.002081051003187895, + 0.0027022617869079113, + 0.0033249245025217533, + 0.001816965639591217, + 0.0029679795261472464, + 0.0026425581891089678, + 0.01264125108718872, + 0.007891830988228321, + 0.014659486711025238, + 0.02600327506661415, + 0.011815935373306274, + 0.006926028989255428, + 0.06106050685048103, + 0.017861660569906235, + 0.0069435457699000835, + 0.006208195351064205, + 0.010054421611130238, + 0.0148466806858778, + 0.006211999338120222, + 0.0057059465907514095, + 0.01558304950594902, + 0.009285028092563152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.0019502682844176888 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00250038574449718 + ], + "opposite_logprob_tokens": [ + -6.002500534057617 + ], + "logprob_reward": -0.00250038574449718, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00250038574449718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2940080165863037, + 0.06146552413702011, + 0.04298205301165581, + 0.1995362788438797, + 0.03795261308550835, + 0.0655527263879776, + 0.04510033130645752, + 0.03386981412768364, + 0.15705908834934235, + 0.062473490834236145 + ], + "credit_map": [ + 0.022936103865504265, + 0.01675226166844368, + 0.0861213356256485, + 0.05396600812673569, + 0.017535466700792313, + 0.024211160838603973, + 0.016323434188961983, + 0.05616224929690361, + 0.006760176736861467, + 0.00803416222333908, + 0.006821945775300264, + 0.01309339702129364, + 0.007614753674715757, + 0.005215459503233433, + 0.007168130483478308, + 0.006757499184459448, + 0.004584162496030331, + 0.004811015911400318, + 0.005947902798652649, + 0.006310696713626385, + 0.004606142640113831, + 0.003839744720607996, + 0.006678657606244087, + 0.0062037319876253605, + 0.01828698068857193, + 0.016006585210561752, + 0.026481106877326965, + 0.05364326760172844, + 0.01901196502149105, + 0.011557615362107754, + 0.030796371400356293, + 0.023752376437187195, + 0.0047516836784780025, + 0.00479540741071105, + 0.003705996321514249, + 0.007065185811370611, + 0.0042496537789702415, + 0.0033574355766177177, + 0.005080836825072765, + 0.004946410655975342, + 0.0068147494457662106, + 0.006269371602684259, + 0.00860682874917984, + 0.011504972353577614, + 0.007900960743427277, + 0.006781347095966339, + 0.007086306344717741, + 0.010588191449642181, + 0.0053020319901406765, + 0.00631282152608037, + 0.003713031066581607, + 0.005895733367651701, + 0.005523767322301865, + 0.005047148559242487, + 0.00785790104418993, + 0.005447894334793091, + 0.0035813122522085905, + 0.003071797313168645, + 0.0027706401888281107, + 0.004738734569400549, + 0.006629404611885548, + 0.003205778542906046, + 0.005471238400787115, + 0.004400905687361956, + 0.01289400178939104, + 0.007408676669001579, + 0.014166304841637611, + 0.0261012502014637, + 0.010924376547336578, + 0.008006548509001732, + 0.057960089296102524, + 0.019597843289375305, + 0.004743809811770916, + 0.0054540191777050495, + 0.007787142880260944, + 0.010911640711128712, + 0.0052603548392653465, + 0.00602644681930542, + 0.014400043524801731, + 0.007890029810369015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.00250038574449718 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.054047703742981 + ], + "opposite_logprob_tokens": [ + -0.42904767394065857 + ], + "logprob_reward": -1.054047703742981, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.054047703742981, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.29502129554748535, + 0.09406058490276337, + 0.02955007739365101, + 0.08033982664346695, + 0.09146513044834137, + 0.02698083594441414, + 0.02448461577296257, + 0.06739157438278198, + 0.17746570706367493, + 0.11324039101600647 + ], + "credit_map": [ + 0.025322481989860535, + 0.01628872938454151, + 0.08746661990880966, + 0.059900738298892975, + 0.019541969522833824, + 0.024303719401359558, + 0.01739644818007946, + 0.04480058327317238, + 0.011143535375595093, + 0.00868732575327158, + 0.011702378280460835, + 0.012878158129751682, + 0.007075316738337278, + 0.008446956053376198, + 0.011475512757897377, + 0.022651398554444313, + 0.0034826805349439383, + 0.004665843676775694, + 0.003621097654104233, + 0.004714773502200842, + 0.002830041106790304, + 0.0024718446657061577, + 0.004457262344658375, + 0.0033065336756408215, + 0.007678842172026634, + 0.006456343457102776, + 0.009687063284218311, + 0.021646110340952873, + 0.0076143089681863785, + 0.005446307361125946, + 0.01380180474370718, + 0.008009051904082298, + 0.011715664528310299, + 0.013059580698609352, + 0.009894415736198425, + 0.016809681430459023, + 0.009793931618332863, + 0.006572548300027847, + 0.014520871452987194, + 0.00909843947738409, + 0.002937858924269676, + 0.003761646570637822, + 0.0025950190611183643, + 0.004069942981004715, + 0.004031264688819647, + 0.0029976419173181057, + 0.002503527794033289, + 0.004083935637027025, + 0.0024608667008578777, + 0.003010832704603672, + 0.002554322360083461, + 0.003444865345954895, + 0.0037869608495384455, + 0.0026974526699632406, + 0.004108005668967962, + 0.002421308308839798, + 0.008553736843168736, + 0.007798732258379459, + 0.005667447112500668, + 0.011213365010917187, + 0.00951689574867487, + 0.007232454605400562, + 0.005919780116528273, + 0.011489166878163815, + 0.01348314993083477, + 0.009885492734611034, + 0.0162447951734066, + 0.029516754671931267, + 0.014238351956009865, + 0.00763127813115716, + 0.06516619026660919, + 0.021299690008163452, + 0.011072766967117786, + 0.012755225412547588, + 0.016499679535627365, + 0.022012608125805855, + 0.012581040151417255, + 0.006841566413640976, + 0.020133743062615395, + 0.011343756690621376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -1.054047703742981 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "This static, eye-level close-up captures the air directly above a sun-scorched black asphalt road in crisp, high-speed footage. Under the oppressive force of the midday sun, invisible yet powerful plumes of superheated air rise from the dark surface in rapid, turbulent plumes that visibly wriggle and shimmer. This thermal disturbance acts like a dynamic liquid lens, chaotically refracting the light passing through it and warping the background into a mesmerizing mirage. Distant, hazy mountains and the pale, dusty roadside shoulders are distorted, appearing to undulate, melt, and ripple, as if viewed through a sheet of flowing, crystal water. Captured at high speed, every subtle, rapid vibration is rendered with surreal clarity, creating a palpable atmosphere of searing, static heat, where the only movement is the silent, dancing distortion of the air itself.", + "seed": 10597, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_11/index_644_seed_10597_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.4085941703990102, + "vlm_logprob": -0.4085941703990102, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.4085941703990102, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3238749504089355 + ], + "opposite_logprob_tokens": [ + -0.32387495040893555 + ], + "logprob_reward": -1.3238749504089355, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3238749504089355, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19373300671577454, + 0.5606701374053955, + 0.020709572359919548, + 0.028498925268650055, + 0.017584336921572685, + 0.01975846290588379, + 0.02672167494893074, + 0.022585788741707802, + 0.0762721598148346, + 0.033465977758169174 + ], + "credit_map": [ + 0.025623824447393417, + 0.04049229249358177, + 0.01714463345706463, + 0.025802675634622574, + 0.016316190361976624, + 0.01607370935380459, + 0.030283821746706963, + 0.021995851770043373, + 0.17655408382415771, + 0.2326311320066452, + 0.023127693682909012, + 0.020367542281746864, + 0.059955574572086334, + 0.011168657802045345, + 0.017602093517780304, + 0.019263358786702156, + 0.002853465499356389, + 0.0026582740247249603, + 0.002997171599417925, + 0.003173442557454109, + 0.002076745266094804, + 0.001546473358757794, + 0.002362546743825078, + 0.003041455987840891, + 0.0036101241130381823, + 0.003004261292517185, + 0.004956046584993601, + 0.004700399003922939, + 0.0033896341919898987, + 0.0020976087544113398, + 0.003011025721207261, + 0.0037298244424164295, + 0.0016785993939265609, + 0.0018103994661942124, + 0.0023351211566478014, + 0.0045154765248298645, + 0.0018113967962563038, + 0.0017676770221441984, + 0.0016738183330744505, + 0.0019918489269912243, + 0.001403866452164948, + 0.0023343171924352646, + 0.002040322171524167, + 0.0031169659923762083, + 0.002529542427510023, + 0.0034168448764830828, + 0.0028664374258369207, + 0.0020501655526459217, + 0.003461783053353429, + 0.0031398129649460316, + 0.0032679301220923662, + 0.0037854774855077267, + 0.00360721698962152, + 0.003835138864815235, + 0.0024728013668209314, + 0.003151512937620282, + 0.002489479724317789, + 0.0019305422902107239, + 0.0035823220387101173, + 0.00413143914192915, + 0.002221952425315976, + 0.0019516213797032833, + 0.0030040135607123375, + 0.0032744170166552067, + 0.012520004995167255, + 0.008903604000806808, + 0.011037279851734638, + 0.009505761787295341, + 0.00810175109654665, + 0.0068932934664189816, + 0.010530771687626839, + 0.008779692463576794, + 0.0030274875462055206, + 0.0030315520707517862, + 0.00483274320140481, + 0.007472548633813858, + 0.0028870999813079834, + 0.0032048451248556376, + 0.004433217458426952, + 0.004576484207063913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -1.3238749504089355 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790589451789856 + ], + "opposite_logprob_tokens": [ + -2.579058885574341 + ], + "logprob_reward": -0.0790589451789856, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790589451789856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1460590362548828, + 0.31821703910827637, + 0.21603934466838837, + 0.03909038379788399, + 0.06018073111772537, + 0.030805902555584908, + 0.06636204570531845, + 0.03959110006690025, + 0.06137309595942497, + 0.0222813468426466 + ], + "credit_map": [ + 0.012876303866505623, + 0.011937745846807957, + 0.02054011821746826, + 0.029850730672478676, + 0.013252212665975094, + 0.012381933629512787, + 0.02493879944086075, + 0.02028118632733822, + 0.09811858087778091, + 0.12138895690441132, + 0.015903262421488762, + 0.010423065163195133, + 0.030073998495936394, + 0.008496081456542015, + 0.014795406721532345, + 0.019017675891518593, + 0.02602764405310154, + 0.041255541145801544, + 0.035476766526699066, + 0.04252390190958977, + 0.01671816036105156, + 0.010340739041566849, + 0.023092778399586678, + 0.020603813230991364, + 0.006189200561493635, + 0.009389958344399929, + 0.005772759206593037, + 0.004539803136140108, + 0.004786208737641573, + 0.0025185621343553066, + 0.00273595261387527, + 0.0031579413916915655, + 0.00843737367540598, + 0.007379444316029549, + 0.009135517291724682, + 0.013909139670431614, + 0.005727555137127638, + 0.004857766907662153, + 0.005385662894695997, + 0.005348270758986473, + 0.0020851041190326214, + 0.002423166064545512, + 0.003990684170275927, + 0.004659175407141447, + 0.003681028028950095, + 0.0046273814514279366, + 0.005437045823782682, + 0.0039023186545819044, + 0.008608642034232616, + 0.00741951260715723, + 0.010858495719730854, + 0.010297203436493874, + 0.007393656764179468, + 0.007885826751589775, + 0.00753153208643198, + 0.006367180496454239, + 0.0032320423051714897, + 0.0043126740492880344, + 0.0054780724458396435, + 0.0061359708197414875, + 0.0045613753609359264, + 0.00456414045765996, + 0.005801967810839415, + 0.00550485635176301, + 0.010514013469219208, + 0.007054782938212156, + 0.00802222453057766, + 0.007302192505449057, + 0.006701190024614334, + 0.005192873999476433, + 0.008734562434256077, + 0.007851256057620049, + 0.0023964170832186937, + 0.0034937935415655375, + 0.0034400669392198324, + 0.002912650816142559, + 0.002390121342614293, + 0.002133391099050641, + 0.003048207610845566, + 0.0024666995741426945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0790589451789856 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002495034597814083 + ], + "opposite_logprob_tokens": [ + -6.002494812011719 + ], + "logprob_reward": -0.002495034597814083, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002495034597814083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11806797981262207, + 0.6114751100540161, + 0.031779881566762924, + 0.03980023041367531, + 0.023074163123965263, + 0.019322844222187996, + 0.029535513371229172, + 0.03928816691040993, + 0.06261887401342392, + 0.025037258863449097 + ], + "credit_map": [ + 0.014005031436681747, + 0.02068702131509781, + 0.01438723411411047, + 0.01613805629312992, + 0.012607606127858162, + 0.011879334226250648, + 0.015586419962346554, + 0.01277727447450161, + 0.1933450996875763, + 0.25556737184524536, + 0.024066003039479256, + 0.019966572523117065, + 0.06540467590093613, + 0.011837066151201725, + 0.018399424850940704, + 0.02288888953626156, + 0.005415384657680988, + 0.007856376469135284, + 0.0027208488900214434, + 0.003986900672316551, + 0.0036347401328384876, + 0.002317390637472272, + 0.0025907799135893583, + 0.003257457632571459, + 0.005038259085267782, + 0.003965363372117281, + 0.00808117538690567, + 0.007107682526111603, + 0.0037998808547854424, + 0.0023000678047537804, + 0.0045255632139742374, + 0.004982237238436937, + 0.0022713658399879932, + 0.0024488577619194984, + 0.004344051703810692, + 0.004634748212993145, + 0.002088115317746997, + 0.0023044177796691656, + 0.0027003404684364796, + 0.0022822634782642126, + 0.001464770408347249, + 0.0019382809987291694, + 0.0019688087049871683, + 0.0025818166323006153, + 0.0028086320962756872, + 0.0031436309218406677, + 0.0033227072563022375, + 0.0020941956900060177, + 0.0041019245982170105, + 0.003441378241404891, + 0.0040333629585802555, + 0.003970550373196602, + 0.003532977541908622, + 0.0036991999950259924, + 0.0027749070432037115, + 0.003981208894401789, + 0.003243895247578621, + 0.003702946240082383, + 0.006196030415594578, + 0.007256071083247662, + 0.003658632980659604, + 0.004944346845149994, + 0.004691003821790218, + 0.005595236551016569, + 0.010187369771301746, + 0.007254261989146471, + 0.008838566020131111, + 0.007547165732830763, + 0.006703090853989124, + 0.005703459959477186, + 0.008713777177035809, + 0.007671174593269825, + 0.002152242697775364, + 0.00375454337336123, + 0.0034181331284344196, + 0.004206310957670212, + 0.0024191448464989662, + 0.0021433138754218817, + 0.003764855209738016, + 0.0031787122134119272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.002495034597814083 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647525027394295 + ], + "opposite_logprob_tokens": [ + -5.379647731781006 + ], + "logprob_reward": -0.004647525027394295, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647525027394295, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1230500265955925, + 0.522402286529541, + 0.05126839131116867, + 0.07750272750854492, + 0.0370207205414772, + 0.022450242191553116, + 0.026711640879511833, + 0.04123443737626076, + 0.04837682098150253, + 0.04998264089226723 + ], + "credit_map": [ + 0.015276286751031876, + 0.023327575996518135, + 0.016703670844435692, + 0.01852407492697239, + 0.011817201040685177, + 0.013508714735507965, + 0.01248864084482193, + 0.011403882876038551, + 0.1474122852087021, + 0.2038000375032425, + 0.02272837609052658, + 0.02331150509417057, + 0.05634254217147827, + 0.01497847493737936, + 0.023682793602347374, + 0.030146336182951927, + 0.007869970053434372, + 0.010874168016016483, + 0.006191105581820011, + 0.007361697964370251, + 0.005626980680972338, + 0.004153862129896879, + 0.004704407881945372, + 0.004486205521970987, + 0.015733810141682625, + 0.02385788783431053, + 0.005282226018607616, + 0.006035301368683577, + 0.011052872985601425, + 0.006040714215487242, + 0.0047240788117051125, + 0.004775842651724815, + 0.004848122131079435, + 0.0030924712773412466, + 0.006602805107831955, + 0.008562019094824791, + 0.003112850012257695, + 0.0028296159580349922, + 0.004035797901451588, + 0.003937042783945799, + 0.001784292166121304, + 0.0020980702247470617, + 0.0022222234401851892, + 0.0035806456580758095, + 0.0028516286984086037, + 0.0033325052354484797, + 0.0037049681413918734, + 0.0028759099077433348, + 0.00362324807792902, + 0.003399625886231661, + 0.0035603379365056753, + 0.0035360294859856367, + 0.0031768789049237967, + 0.0028260464314371347, + 0.002511512255296111, + 0.0040779635310173035, + 0.0031972865108400583, + 0.005530884955078363, + 0.004681685473769903, + 0.004909402225166559, + 0.005719790700823069, + 0.005450184922665358, + 0.005977145396173, + 0.0057680620811879635, + 0.00790630467236042, + 0.005616835784167051, + 0.007330401334911585, + 0.00636661471799016, + 0.004897538106888533, + 0.004748369101434946, + 0.006281283684074879, + 0.00522947870194912, + 0.005339911673218012, + 0.004337707534432411, + 0.00873059406876564, + 0.0073064174503088, + 0.00409372104331851, + 0.007069692946970463, + 0.007152889389544725, + 0.005951711907982826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.004647525027394295 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328943967819214 + ], + "opposite_logprob_tokens": [ + -0.7578943967819214 + ], + "logprob_reward": -0.6328943967819214, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328943967819214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039354532957077026, + 0.38998448848724365, + 0.21784867346286774, + 0.02298862300813198, + 0.21711936593055725, + 0.015594616532325745, + 0.024744318798184395, + 0.04414135217666626, + 0.015202623791992664, + 0.01302135270088911 + ], + "credit_map": [ + 0.0060921404510736465, + 0.00877678208053112, + 0.00526877585798502, + 0.0045369346626102924, + 0.004471426829695702, + 0.0033219873439520597, + 0.0036802066024392843, + 0.003206283086910844, + 0.12228924036026001, + 0.1638210117816925, + 0.012873681262135506, + 0.011990081518888474, + 0.04256296902894974, + 0.006449716631323099, + 0.012352829799056053, + 0.01764494739472866, + 0.03843330219388008, + 0.061172157526016235, + 0.026230860501527786, + 0.03025248646736145, + 0.02129262499511242, + 0.008574353531002998, + 0.016043568029999733, + 0.015849340707063675, + 0.0027178579475730658, + 0.004120498895645142, + 0.00298292376101017, + 0.004058513790369034, + 0.00352198607288301, + 0.0017525425646454096, + 0.001976077212020755, + 0.0018582247430458665, + 0.029938045889139175, + 0.019310995936393738, + 0.03957565128803253, + 0.04592244699597359, + 0.02017337456345558, + 0.017827708274126053, + 0.024001499637961388, + 0.02036965824663639, + 0.0019278734689578414, + 0.0018104459159076214, + 0.001365849282592535, + 0.0020960592664778233, + 0.0021409860346466303, + 0.0022716959938406944, + 0.0017004032852128148, + 0.0022813051473349333, + 0.003018460003659129, + 0.0033431367482990026, + 0.0030956827104091644, + 0.0034971910063177347, + 0.0034428962972015142, + 0.0034731205087155104, + 0.0026499037630856037, + 0.002223932184278965, + 0.003527054563164711, + 0.005630933679640293, + 0.005323936231434345, + 0.005904337391257286, + 0.0061071282252669334, + 0.005243758205324411, + 0.006229685619473457, + 0.006174520123749971, + 0.0019046174129471183, + 0.002031933981925249, + 0.00232279603369534, + 0.002056689001619816, + 0.001993838232010603, + 0.0017127076862379909, + 0.0016046466771513224, + 0.001575396046973765, + 0.0013678210088983178, + 0.0013228458119556308, + 0.0017444256227463484, + 0.0025244848802685738, + 0.0014544643927365541, + 0.001515558804385364, + 0.001569718006066978, + 0.0015220348723232746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.6328943967819214 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, eye-level close-up captures the air directly above a sun-scorched black asphalt road in crisp, high-speed footage. Under the oppressive force of the midday sun, invisible yet powerful plumes of superheated air rise from the dark surface in rapid, turbulent plumes that visibly wriggle and shimmer. This thermal disturbance acts like a dynamic liquid lens, chaotically refracting the light passing through it and warping the background into a mesmerizing mirage. Distant, hazy mountains and the pale, dusty roadside shoulders are distorted, appearing to undulate, melt, and ripple, as if viewed through a sheet of flowing, crystal water. Captured at high speed, every subtle, rapid vibration is rendered with surreal clarity, creating a palpable atmosphere of searing, static heat, where the only movement is the silent, dancing distortion of the air itself.", + "seed": 10598, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_11/index_644_seed_10598_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.38454026347026227, + "vlm_logprob": -0.38454026347026227, + "vlm_yes_no_margin": 3.175000011920929, + "vlm_token_credit": -0.38454026347026227, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0724835395812988 + ], + "opposite_logprob_tokens": [ + -0.4474835991859436 + ], + "logprob_reward": -1.0724835395812988, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.0724835395812988, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.173436239361763, + 0.09112548828125, + 0.15300039947032928, + 0.1695232391357422, + 0.13801001012325287, + 0.06354260444641113, + 0.0662771537899971, + 0.06995457410812378, + 0.0249839685857296, + 0.05014631897211075 + ], + "credit_map": [ + 0.021439604461193085, + 0.031206589192152023, + 0.03254791721701622, + 0.023930037394165993, + 0.014120211824774742, + 0.014518131501972675, + 0.01884845271706581, + 0.01682530716061592, + 0.01652355305850506, + 0.013205088675022125, + 0.019692804664373398, + 0.007226096000522375, + 0.01686934567987919, + 0.007092006504535675, + 0.0061325994320213795, + 0.004384004510939121, + 0.01619049534201622, + 0.015572994016110897, + 0.02579203061759472, + 0.021685566753149033, + 0.01654997654259205, + 0.015734292566776276, + 0.014843987300992012, + 0.02663106843829155, + 0.048569221049547195, + 0.009765343740582466, + 0.06439033150672913, + 0.014990937896072865, + 0.008344440720975399, + 0.00592143926769495, + 0.0075207483023405075, + 0.010020794346928596, + 0.015945693477988243, + 0.02210693061351776, + 0.020141595974564552, + 0.020386360585689545, + 0.022540317848324776, + 0.012188924476504326, + 0.013303123414516449, + 0.01139707863330841, + 0.00581824267283082, + 0.006719381548464298, + 0.008456090465188026, + 0.008479377254843712, + 0.009821639396250248, + 0.004137818701565266, + 0.008275697007775307, + 0.011834361590445042, + 0.008739215321838856, + 0.0064096939750015736, + 0.007327086757868528, + 0.010239429771900177, + 0.006351016461849213, + 0.005973896011710167, + 0.012014678679406643, + 0.009222140535712242, + 0.006155605427920818, + 0.008149897679686546, + 0.009138871915638447, + 0.0162262711673975, + 0.010385135188698769, + 0.0042509776540100574, + 0.009846655651926994, + 0.005801168270409107, + 0.0025921857450157404, + 0.002693560905754566, + 0.004304138943552971, + 0.004519753623753786, + 0.0026653327513486147, + 0.001663530245423317, + 0.004091216251254082, + 0.002454252215102315, + 0.00561031186953187, + 0.007373965345323086, + 0.008181574754416943, + 0.008663848973810673, + 0.006543627940118313, + 0.0033564746845513582, + 0.0041542863473296165, + 0.00626223674044013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -1.0724835395812988 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020660780370235443 + ], + "opposite_logprob_tokens": [ + -3.895660877227783 + ], + "logprob_reward": -0.020660780370235443, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020660780370235443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17367534339427948, + 0.10945134609937668, + 0.14158770442008972, + 0.15788400173187256, + 0.060793496668338776, + 0.06512581557035446, + 0.08411437273025513, + 0.06705018877983093, + 0.06944134086370468, + 0.07087638229131699 + ], + "credit_map": [ + 0.01983271911740303, + 0.021979793906211853, + 0.029848400503396988, + 0.022073375061154366, + 0.02246915176510811, + 0.017901096493005753, + 0.01909930445253849, + 0.02047150395810604, + 0.01655786670744419, + 0.01006048172712326, + 0.023806914687156677, + 0.011557226069271564, + 0.013729157857596874, + 0.0081098061054945, + 0.014894399791955948, + 0.010735489428043365, + 0.015817916020751, + 0.015963239595294, + 0.022986846044659615, + 0.023621750995516777, + 0.016430100426077843, + 0.013676832430064678, + 0.013033419847488403, + 0.020057599991559982, + 0.023702623322606087, + 0.017284611240029335, + 0.028096992522478104, + 0.019029511138796806, + 0.020420854911208153, + 0.007436975371092558, + 0.02326798439025879, + 0.01864444650709629, + 0.007651222869753838, + 0.006809763610363007, + 0.010980076156556606, + 0.012915529310703278, + 0.005183420144021511, + 0.0033559275325387716, + 0.006640423554927111, + 0.007257131393998861, + 0.0067433244548738, + 0.006023559719324112, + 0.009234534576535225, + 0.011984949000179768, + 0.006722614169120789, + 0.00361245172098279, + 0.011493932455778122, + 0.009310447610914707, + 0.01014699973165989, + 0.009184609167277813, + 0.00919236708432436, + 0.01739002950489521, + 0.008350944146513939, + 0.00760525930672884, + 0.01433743815869093, + 0.00790673028677702, + 0.007152137812227011, + 0.007817045785486698, + 0.006948789581656456, + 0.01282182801514864, + 0.008413553237915039, + 0.004678005818277597, + 0.012802492827177048, + 0.006416338495910168, + 0.008407765068113804, + 0.009180882014334202, + 0.013094835914671421, + 0.0072830673307180405, + 0.009328708983957767, + 0.005668483674526215, + 0.0066771903075277805, + 0.009800398722290993, + 0.008406083099544048, + 0.010831858031451702, + 0.009812092408537865, + 0.010844280011951923, + 0.01209932379424572, + 0.004462437238544226, + 0.007013069000095129, + 0.007407249882817268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.020660780370235443 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013441346818581223 + ], + "opposite_logprob_tokens": [ + -6.626344203948975 + ], + "logprob_reward": -0.0013441346818581223, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013441346818581223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12792527675628662, + 0.13279692828655243, + 0.18483184278011322, + 0.09514505416154861, + 0.17699548602104187, + 0.07246869802474976, + 0.06628568470478058, + 0.057906508445739746, + 0.0414523184299469, + 0.04419220983982086 + ], + "credit_map": [ + 0.014250908978283405, + 0.015003879554569721, + 0.020396128296852112, + 0.01694587990641594, + 0.01934840902686119, + 0.013909592293202877, + 0.012842201627790928, + 0.015228274278342724, + 0.013589678332209587, + 0.023029806092381477, + 0.015101236291229725, + 0.014684449881315231, + 0.02941139042377472, + 0.010687178000807762, + 0.017328735440969467, + 0.00896445568650961, + 0.020773757249116898, + 0.022303227335214615, + 0.032607775181531906, + 0.026668934151530266, + 0.020601486787199974, + 0.01625291258096695, + 0.018113115802407265, + 0.027510635554790497, + 0.010736756026744843, + 0.012452241033315659, + 0.010299510322511196, + 0.013234959915280342, + 0.012496721930801868, + 0.006927441339939833, + 0.013454034924507141, + 0.015543385408818722, + 0.01995999924838543, + 0.029879773035645485, + 0.024501649662852287, + 0.02934923954308033, + 0.028422262519598007, + 0.016839785501360893, + 0.013957224786281586, + 0.014085548929870129, + 0.008728327229619026, + 0.00936927180737257, + 0.008339357562363148, + 0.01064448431134224, + 0.011466000229120255, + 0.005679950583726168, + 0.005420691799372435, + 0.012820609845221043, + 0.009903560392558575, + 0.007747572381049395, + 0.007865228690207005, + 0.008344963192939758, + 0.0073394253849983215, + 0.00839606486260891, + 0.009446673095226288, + 0.007242193445563316, + 0.0076126656495034695, + 0.007472996599972248, + 0.008049214258790016, + 0.0058871423825621605, + 0.014428680762648582, + 0.005351581145077944, + 0.0038756572175771, + 0.005228572525084019, + 0.004719278775155544, + 0.003530105110257864, + 0.00832301750779152, + 0.004361055791378021, + 0.005936277098953724, + 0.0030763663817197084, + 0.0052031357772648335, + 0.0063030775636434555, + 0.005295464303344488, + 0.006455016788095236, + 0.007229323964565992, + 0.007703909184783697, + 0.005789034068584442, + 0.0025924197398126125, + 0.004008562304079533, + 0.005118481814861298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.0013441346818581223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019496734021231532 + ], + "opposite_logprob_tokens": [ + -6.251949787139893 + ], + "logprob_reward": -0.0019496734021231532, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019496734021231532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14075951278209686, + 0.1283729374408722, + 0.1666564643383026, + 0.13187885284423828, + 0.1804611086845398, + 0.06335490196943283, + 0.06371515244245529, + 0.04931104928255081, + 0.02185574546456337, + 0.053634267300367355 + ], + "credit_map": [ + 0.01761682890355587, + 0.018330292776226997, + 0.026420271024107933, + 0.01693103089928627, + 0.019853150472044945, + 0.008502406999468803, + 0.01497718971222639, + 0.018128342926502228, + 0.022458260878920555, + 0.01805836707353592, + 0.026302624493837357, + 0.016360653564333916, + 0.014701132662594318, + 0.007731599733233452, + 0.01427085418254137, + 0.008489450439810753, + 0.01836393028497696, + 0.02075023204088211, + 0.030033336952328682, + 0.024594545364379883, + 0.01984289288520813, + 0.014366754330694675, + 0.016051724553108215, + 0.02265304885804653, + 0.01866834983229637, + 0.016700031235814095, + 0.025426607578992844, + 0.016079768538475037, + 0.013931069523096085, + 0.009713871404528618, + 0.01270512118935585, + 0.01865403726696968, + 0.02067941054701805, + 0.02647165209054947, + 0.028660215437412262, + 0.03043694794178009, + 0.028323069214820862, + 0.015028324909508228, + 0.01585979014635086, + 0.015001706779003143, + 0.007390943821519613, + 0.007585326675325632, + 0.008589621633291245, + 0.00849530566483736, + 0.00879800133407116, + 0.004827572498470545, + 0.005398983601480722, + 0.012269144877791405, + 0.009314694441854954, + 0.007380443159490824, + 0.00767172034829855, + 0.00979401171207428, + 0.006052176468074322, + 0.00643501291051507, + 0.010301150381565094, + 0.006765944883227348, + 0.006220079492777586, + 0.006345498841255903, + 0.007006137166172266, + 0.006999034434556961, + 0.009523149579763412, + 0.0040403264574706554, + 0.004565199371427298, + 0.004611623473465443, + 0.0029576881788671017, + 0.0017195029649883509, + 0.006271928548812866, + 0.0020670497324317694, + 0.0028149140998721123, + 0.0016145199770107865, + 0.0018976583378389478, + 0.002512483624741435, + 0.006206711288541555, + 0.008710211142897606, + 0.008123435080051422, + 0.009475874714553356, + 0.007525957189500332, + 0.0032165534794330597, + 0.004892009310424328, + 0.005483515560626984 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.0019496734021231532 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a close-up shot of the air directly above a black asphalt road?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are visible plumes of superheated air rising from the road surface in rapid, turbulent motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the air distortion refract light and warp the background into a mirage effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are distant mountains and roadside shoulders visible and distorted to appear undulating or melting?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the footage captured at high speed to render rapid vibrations with clarity?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8262631893157959 + ], + "opposite_logprob_tokens": [ + -0.5762631893157959 + ], + "logprob_reward": -0.8262631893157959, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8262631893157959, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06978185474872589, + 0.09944318234920502, + 0.0683729350566864, + 0.5322815179824829, + 0.08717387169599533, + 0.047704968601465225, + 0.022285010665655136, + 0.03763961046934128, + 0.018331246450543404, + 0.016985803842544556 + ], + "credit_map": [ + 0.009072419255971909, + 0.011401042342185974, + 0.012160095386207104, + 0.008744151331484318, + 0.008517042733728886, + 0.004571318160742521, + 0.008569498546421528, + 0.006746291182935238, + 0.018620718270540237, + 0.010820265859365463, + 0.022312860935926437, + 0.011076622642576694, + 0.009523244574666023, + 0.006354985758662224, + 0.013305704109370708, + 0.007428786251693964, + 0.008137408643960953, + 0.009289220906794071, + 0.011145031079649925, + 0.008945678360760212, + 0.004847268108278513, + 0.0061981710605323315, + 0.007732493337243795, + 0.012077664025127888, + 0.1615620255470276, + 0.027375856414437294, + 0.21470384299755096, + 0.041984278708696365, + 0.024903081357479095, + 0.010900160297751427, + 0.033704664558172226, + 0.017147613689303398, + 0.009672919288277626, + 0.0078113810159265995, + 0.015231678262352943, + 0.022039681673049927, + 0.00612300680950284, + 0.003558022202923894, + 0.011687079444527626, + 0.011050103232264519, + 0.005464569665491581, + 0.004973875358700752, + 0.010099521838128567, + 0.009480748325586319, + 0.004374133888632059, + 0.0022094787564128637, + 0.0064270687289536, + 0.004675572272390127, + 0.003399844281375408, + 0.002988312626257539, + 0.003155844984576106, + 0.003575369017198682, + 0.0018707511480897665, + 0.0017455725464969873, + 0.002027691574767232, + 0.0035216237884014845, + 0.003761197440326214, + 0.005595202092081308, + 0.004435395821928978, + 0.008374316617846489, + 0.005996185354888439, + 0.0026187740731984377, + 0.0046105170622467995, + 0.002248024335131049, + 0.0025903102941811085, + 0.00199419679120183, + 0.004834818188101053, + 0.0018459237180650234, + 0.002296879654750228, + 0.001386524410918355, + 0.0012821851996704936, + 0.0021004092413932085, + 0.0020260855089873075, + 0.0020740784239023924, + 0.0025371243245899677, + 0.0017810161225497723, + 0.0037833950482308865, + 0.0014473031042143703, + 0.0014576648827642202, + 0.0018791392212733626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.8262631893157959 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_12/index_tensor([1727, 1727])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_12/index_tensor([1727, 1727])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2504e944ef75b2c6a983aeeef5e0eb0814658665 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_12/index_tensor([1727, 1727])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from a slightly upward angle, the scene takes on a unique stop-motion quality. Each frame feels tangible, with tiny, almost imperceptible jumps between movements, imbuing the characters and objects with a handmade, slightly jerky, yet highly expressive quality. The image focuses on an elderly man with a deeply wrinkled face, his thinning white hair plastered against a weathered scalp, his eyes kind yet slightly weary. He stands in a dimly lit, sparsely furnished wooden house, set against a backdrop of rough, dark wood walls. A soft, warm light streams in through a small side window, vividly outlining every crease on his face and illuminating his plain, earth-brown linen shirt. With hands punctuated by age spots and bulging veins, he carefully, almost reverently, holds a simple, unglazed stoneware bowl, its surface rough to the touch, bearing the unmistakable marks of handcraft. Due to the slight but constant trembling of his hands, tiny ripples appeared on the clear water in the bowl, reflecting the faint light from the window. He slowly brought the rim of the bowl to his chapped lips, tilted his head slightly, and took a deliberate sip. The whole movement was imbued with the weight of time. The warm earth tones and strong contrasts of light and dark conveyed a sense of quiet dignity and a moving atmosphere of contemplation.", + "seed": 11840, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_12/index_1727_seed_11840_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17510681897401809, + "vlm_logprob": -0.17510681897401809, + "vlm_yes_no_margin": 2.7249999761581423, + "vlm_token_credit": -0.17510681897401809, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.249123215675354 + ], + "opposite_logprob_tokens": [ + -2.1241230964660645 + ], + "logprob_reward": -0.249123215675354, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.249123215675354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1587415337562561, + 0.09635011106729507, + 0.04504561051726341, + 0.0437329038977623, + 0.08178852498531342, + 0.0642176941037178, + 0.04973878711462021, + 0.32603758573532104, + 0.0991184338927269, + 0.03522881865501404 + ], + "credit_map": [ + 0.008608011528849602, + 0.02219817042350769, + 0.01699938252568245, + 0.02077636867761612, + 0.028748853132128716, + 0.01065871026366949, + 0.029714658856391907, + 0.021037377417087555, + 0.006884522270411253, + 0.011977306567132473, + 0.006024619564414024, + 0.014849799685180187, + 0.020918801426887512, + 0.0063796499744057655, + 0.014574972912669182, + 0.014740442857146263, + 0.0036024884320795536, + 0.003984236624091864, + 0.0036211179103702307, + 0.003117925487458706, + 0.008903752081096172, + 0.0028516899328678846, + 0.01068041380494833, + 0.008283985778689384, + 0.003826092928647995, + 0.0050689224153757095, + 0.0029984323773533106, + 0.0029699131846427917, + 0.006861046887934208, + 0.0032199143897742033, + 0.011177712120115757, + 0.007610870059579611, + 0.0062952423468232155, + 0.007990112528204918, + 0.006486521102488041, + 0.007098776753991842, + 0.014513961970806122, + 0.006989479064941406, + 0.01982012577354908, + 0.01259430032223463, + 0.0063575138337910175, + 0.00663559278473258, + 0.004556544590741396, + 0.005761453881859779, + 0.015686560422182083, + 0.005303565412759781, + 0.010718872770667076, + 0.009197589010000229, + 0.003261126112192869, + 0.004277022089809179, + 0.0032225679606199265, + 0.0050417715683579445, + 0.010334692895412445, + 0.0037061264738440514, + 0.011120233684778214, + 0.008775247260928154, + 0.009325498715043068, + 0.0082175824791193, + 0.008518481627106667, + 0.013991114683449268, + 0.06762179732322693, + 0.009883747436106205, + 0.13545404374599457, + 0.07302531599998474, + 0.006053780671209097, + 0.009635528549551964, + 0.010695572942495346, + 0.022408893331885338, + 0.022761370986700058, + 0.0037620600778609514, + 0.017949147149920464, + 0.0058520776219666, + 0.002460939809679985, + 0.0035798395983874798, + 0.0031364725437015295, + 0.006069948896765709, + 0.007721718866378069, + 0.0037453884724527597, + 0.005255882628262043, + 0.0032586294692009687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.249123215675354 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234285593032837 + ], + "opposite_logprob_tokens": [ + -0.8984285593032837 + ], + "logprob_reward": -0.5234285593032837, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234285593032837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21773402392864227, + 0.08855714648962021, + 0.060480594635009766, + 0.06233840808272362, + 0.029114026576280594, + 0.20787131786346436, + 0.04866814613342285, + 0.22064486145973206, + 0.028831763193011284, + 0.03575977310538292 + ], + "credit_map": [ + 0.0167107954621315, + 0.017275815829634666, + 0.018431328237056732, + 0.029430747032165527, + 0.03752962127327919, + 0.01701260730624199, + 0.050550635904073715, + 0.030792459845542908, + 0.009035726077854633, + 0.008268740028142929, + 0.006531045772135258, + 0.011097763665020466, + 0.015149532817304134, + 0.008248802274465561, + 0.016917914152145386, + 0.013307624496519566, + 0.0050049591809511185, + 0.006319667678326368, + 0.006487127859145403, + 0.006784990895539522, + 0.013173848390579224, + 0.004333462566137314, + 0.009993097744882107, + 0.008383444510400295, + 0.0052468664944171906, + 0.007412086706608534, + 0.0035990567412227392, + 0.007047212217003107, + 0.013377579860389233, + 0.0056885480880737305, + 0.010654267854988575, + 0.009312795475125313, + 0.002030659466981888, + 0.0027976189740002155, + 0.0022109602577984333, + 0.002871230710297823, + 0.00596189871430397, + 0.002378858160227537, + 0.005970397032797337, + 0.004892404656857252, + 0.016944479197263718, + 0.021639114245772362, + 0.011871100403368473, + 0.017852185294032097, + 0.05983296409249306, + 0.014415275305509567, + 0.035464443266391754, + 0.029851749539375305, + 0.0030146113131195307, + 0.005511464085429907, + 0.0025119588244706392, + 0.005429612472653389, + 0.013583815656602383, + 0.0040110074914991856, + 0.006548125296831131, + 0.008057553321123123, + 0.005858805496245623, + 0.003917159046977758, + 0.005381126422435045, + 0.009061815217137337, + 0.046490345150232315, + 0.005979674402624369, + 0.09362993389368057, + 0.05032598227262497, + 0.0013270649360492826, + 0.0041098156943917274, + 0.0023567993193864822, + 0.004336439538747072, + 0.007535222917795181, + 0.002249986631795764, + 0.004796285182237625, + 0.0021201481577008963, + 0.0031671940814703703, + 0.004146267659962177, + 0.003370999824255705, + 0.004484961740672588, + 0.009550289250910282, + 0.003000942524522543, + 0.005091231781989336, + 0.0029478862415999174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.5234285593032837 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009785893373191357 + ], + "opposite_logprob_tokens": [ + -4.634786128997803 + ], + "logprob_reward": -0.009785893373191357, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009785893373191357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31469470262527466, + 0.14425764977931976, + 0.13438834249973297, + 0.030260805040597916, + 0.03808186203241348, + 0.1311463564634323, + 0.05428615212440491, + 0.09141483902931213, + 0.03410571441054344, + 0.02736358717083931 + ], + "credit_map": [ + 0.022815676406025887, + 0.031124265864491463, + 0.027217593044042587, + 0.04814978688955307, + 0.04586154222488403, + 0.022374112159013748, + 0.07152344286441803, + 0.045628294348716736, + 0.010638744570314884, + 0.014024664647877216, + 0.012380155734717846, + 0.03207939863204956, + 0.01973741129040718, + 0.00846785120666027, + 0.0242008026689291, + 0.022728627547621727, + 0.011832193471491337, + 0.016724826768040657, + 0.011541897431015968, + 0.014693389646708965, + 0.022517623379826546, + 0.009527189657092094, + 0.027227936312556267, + 0.020323291420936584, + 0.00279283057898283, + 0.0021651308052241802, + 0.002090228023007512, + 0.0025758990086615086, + 0.006952987983822823, + 0.0018084951443597674, + 0.006794163957238197, + 0.005081072449684143, + 0.002980084391310811, + 0.003547040047124028, + 0.002897399477660656, + 0.004262357018887997, + 0.007845760323107243, + 0.00308110099285841, + 0.0072190240025520325, + 0.006249099969863892, + 0.007081125862896442, + 0.011385509744286537, + 0.012040944769978523, + 0.01292034424841404, + 0.034656964242458344, + 0.007187128532677889, + 0.027463318780064583, + 0.018411025404930115, + 0.004719533957540989, + 0.005644712597131729, + 0.0026829291600733995, + 0.006129451561719179, + 0.011141257360577583, + 0.00412844680249691, + 0.009685195051133633, + 0.010154630057513714, + 0.0032575237564742565, + 0.0032827388495206833, + 0.0027336888015270233, + 0.004119482357054949, + 0.019731465727090836, + 0.0031862410251051188, + 0.03677932545542717, + 0.01832437515258789, + 0.0018061307491734624, + 0.0041746520437300205, + 0.0028330478817224503, + 0.006999319419264793, + 0.008155833929777145, + 0.0023368929978460073, + 0.005573405884206295, + 0.0022264320868998766, + 0.0019739349372684956, + 0.002517035696655512, + 0.0019907266832888126, + 0.004091904032975435, + 0.007723050657659769, + 0.00164754968136549, + 0.005019500385969877, + 0.0023998890537768602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.009785893373191357 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900100946426392 + ], + "opposite_logprob_tokens": [ + -2.579000949859619 + ], + "logprob_reward": -0.07900100946426392, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900100946426392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5029072761535645, + 0.05484785884618759, + 0.11744806915521622, + 0.03082222118973732, + 0.04669840633869171, + 0.05929424986243248, + 0.027067972347140312, + 0.08472567051649094, + 0.022247837856411934, + 0.05394047126173973 + ], + "credit_map": [ + 0.031968921422958374, + 0.07848350703716278, + 0.04276838153600693, + 0.06541292369365692, + 0.10065658390522003, + 0.041989799588918686, + 0.07519220560789108, + 0.06643493473529816, + 0.006148603279143572, + 0.004999347031116486, + 0.004380350466817617, + 0.004990229848772287, + 0.01099464762955904, + 0.004278167616575956, + 0.010895234532654285, + 0.00816127099096775, + 0.007251854985952377, + 0.00798679981380701, + 0.006291539873927832, + 0.005958672147244215, + 0.020939698442816734, + 0.006956863682717085, + 0.039665479212999344, + 0.022397147491574287, + 0.0018069088691845536, + 0.002598338760435581, + 0.0020224011968821287, + 0.0030493643134832382, + 0.0046301353722810745, + 0.001811156515032053, + 0.007733802776783705, + 0.00717011047527194, + 0.00336872274056077, + 0.004816173575818539, + 0.0034798469860106707, + 0.0057282778434455395, + 0.00929086934775114, + 0.004381272941827774, + 0.008627915754914284, + 0.007005324587225914, + 0.003371509024873376, + 0.004767556209117174, + 0.004172501154243946, + 0.005514276213943958, + 0.01689537800848484, + 0.003429901786148548, + 0.012329786084592342, + 0.008813337422907352, + 0.0023673109244555235, + 0.003001304343342781, + 0.0017389869317412376, + 0.0027810214087367058, + 0.00495613319799304, + 0.002456278307363391, + 0.004669031593948603, + 0.005097901914268732, + 0.0028977186884731054, + 0.001847755047492683, + 0.002543276408687234, + 0.003558094846084714, + 0.01853632740676403, + 0.003002839395776391, + 0.034787554293870926, + 0.0175520908087492, + 0.001249076216481626, + 0.0025397816207259893, + 0.0023311059921979904, + 0.005513188894838095, + 0.004137637559324503, + 0.0011196044506505132, + 0.003768068039789796, + 0.001589373336173594, + 0.0040216329507529736, + 0.005549443885684013, + 0.0065564243122935295, + 0.010324940085411072, + 0.010584203526377678, + 0.0036325815599411726, + 0.00848120916634798, + 0.004790032282471657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.07900100946426392 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195417053997517 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195417053997517, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195417053997517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09804942458868027, + 0.04792962595820427, + 0.08312136679887772, + 0.07794079184532166, + 0.09499041736125946, + 0.03704984113574028, + 0.04896407946944237, + 0.2420656383037567, + 0.05921559780836105, + 0.2106732726097107 + ], + "credit_map": [ + 0.00764343561604619, + 0.008236357942223549, + 0.010812176391482353, + 0.012896861881017685, + 0.017746463418006897, + 0.007699923124164343, + 0.02043551206588745, + 0.012578684836626053, + 0.004236343316733837, + 0.005291371140629053, + 0.00321850273758173, + 0.007223299238830805, + 0.008040297776460648, + 0.00462264334782958, + 0.008090775460004807, + 0.007206387352198362, + 0.004847567528486252, + 0.005261550657451153, + 0.0054603624157607555, + 0.006427091546356678, + 0.0142337242141366, + 0.005745209753513336, + 0.02527056820690632, + 0.015875281766057014, + 0.007102686911821365, + 0.009042061865329742, + 0.004129424691200256, + 0.006835137493908405, + 0.018742883577942848, + 0.007732957601547241, + 0.012010394595563412, + 0.012345238588750362, + 0.005561630707234144, + 0.008780494332313538, + 0.00552381994202733, + 0.008547377772629261, + 0.02243277058005333, + 0.009638529270887375, + 0.019386054947972298, + 0.015119732357561588, + 0.0029124568682163954, + 0.003784564323723316, + 0.0023110276088118553, + 0.0032524322159588337, + 0.008994637057185173, + 0.002739689778536558, + 0.007432630751281977, + 0.005622397642582655, + 0.0033888567704707384, + 0.005669100675731897, + 0.0033723386004567146, + 0.004931112285703421, + 0.011079936288297176, + 0.003815262345597148, + 0.008024530485272408, + 0.008682940155267715, + 0.007165343035012484, + 0.005219902843236923, + 0.005877125076949596, + 0.009729467332363129, + 0.05201094225049019, + 0.006835071835666895, + 0.10221255570650101, + 0.05301519110798836, + 0.0033576858695596457, + 0.007728237193077803, + 0.007129283156245947, + 0.015451249666512012, + 0.009374146349728107, + 0.0030682438518851995, + 0.009382876567542553, + 0.0037238686345517635, + 0.012831754051148891, + 0.017681051045656204, + 0.021978087723255157, + 0.04980107396841049, + 0.041154127568006516, + 0.006576919928193092, + 0.04091892018914223, + 0.019731326028704643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.014195417053997517 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly upward angle, the scene takes on a unique stop-motion quality. Each frame feels tangible, with tiny, almost imperceptible jumps between movements, imbuing the characters and objects with a handmade, slightly jerky, yet highly expressive quality. The image focuses on an elderly man with a deeply wrinkled face, his thinning white hair plastered against a weathered scalp, his eyes kind yet slightly weary. He stands in a dimly lit, sparsely furnished wooden house, set against a backdrop of rough, dark wood walls. A soft, warm light streams in through a small side window, vividly outlining every crease on his face and illuminating his plain, earth-brown linen shirt. With hands punctuated by age spots and bulging veins, he carefully, almost reverently, holds a simple, unglazed stoneware bowl, its surface rough to the touch, bearing the unmistakable marks of handcraft. Due to the slight but constant trembling of his hands, tiny ripples appeared on the clear water in the bowl, reflecting the faint light from the window. He slowly brought the rim of the bowl to his chapped lips, tilted his head slightly, and took a deliberate sip. The whole movement was imbued with the weight of time. The warm earth tones and strong contrasts of light and dark conveyed a sense of quiet dignity and a moving atmosphere of contemplation.", + "seed": 11841, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_12/index_1727_seed_11841_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.26257394030690195, + "vlm_logprob": -0.26257394030690195, + "vlm_yes_no_margin": 2.1, + "vlm_token_credit": -0.26257394030690195, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.41243863105773926 + ], + "opposite_logprob_tokens": [ + -1.6624386310577393 + ], + "logprob_reward": -0.41243863105773926, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.41243863105773926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5090910792350769, + 0.06773767620325089, + 0.04917164146900177, + 0.06146005168557167, + 0.04784369841217995, + 0.023637814447283745, + 0.019980721175670624, + 0.060173068195581436, + 0.12856994569301605, + 0.03233432397246361 + ], + "credit_map": [ + 0.16315951943397522, + 0.01680668629705906, + 0.01829114928841591, + 0.045506954193115234, + 0.017004460096359253, + 0.19711564481258392, + 0.023344794288277626, + 0.027861889451742172, + 0.009226324036717415, + 0.00859726034104824, + 0.004899374209344387, + 0.005895643495023251, + 0.007242650259286165, + 0.00976160541176796, + 0.010928741656243801, + 0.011186080984771252, + 0.004432698246091604, + 0.004736314062029123, + 0.004463628400117159, + 0.005654288921505213, + 0.005779155530035496, + 0.007283634506165981, + 0.009302319027483463, + 0.007519600912928581, + 0.00838225707411766, + 0.008121667429804802, + 0.006442604586482048, + 0.007962212897837162, + 0.004833676852285862, + 0.012221037410199642, + 0.005800006445497274, + 0.00769658712670207, + 0.00417296402156353, + 0.0035969517193734646, + 0.0034089633263647556, + 0.005969400051981211, + 0.005821348633617163, + 0.005224388092756271, + 0.01283329352736473, + 0.006816388573497534, + 0.0022945452947169542, + 0.002210699487477541, + 0.0031147149857133627, + 0.0038616133388131857, + 0.002254208317026496, + 0.0031998048070818186, + 0.0035165026783943176, + 0.0031857271678745747, + 0.0021510999649763107, + 0.002103928243741393, + 0.002385925268754363, + 0.002473591361194849, + 0.002920081838965416, + 0.0020321423653513193, + 0.002158073242753744, + 0.0037558795884251595, + 0.0078952731564641, + 0.011086368933320045, + 0.005518388468772173, + 0.0062014032155275345, + 0.007095346227288246, + 0.006486071739345789, + 0.0060463277623057365, + 0.009843884967267513, + 0.033762041479349136, + 0.00991095695644617, + 0.008978703990578651, + 0.00696409959346056, + 0.005378850735723972, + 0.015241206623613834, + 0.00710430508479476, + 0.04122978076338768, + 0.002773026702925563, + 0.006525097414851189, + 0.002339449245482683, + 0.005194883327931166, + 0.0031864072661846876, + 0.004694273229688406, + 0.0032775416038930416, + 0.00434364378452301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.41243863105773926 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934534907341003 + ], + "opposite_logprob_tokens": [ + -0.6934534907341003 + ], + "logprob_reward": -0.6934534907341003, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934534907341003, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20795008540153503, + 0.34226885437965393, + 0.04497302696108818, + 0.09711581468582153, + 0.06621375679969788, + 0.0466512106359005, + 0.024429786950349808, + 0.0502663217484951, + 0.054029449820518494, + 0.06610172986984253 + ], + "credit_map": [ + 0.03500571846961975, + 0.021521804854273796, + 0.016427157446742058, + 0.021315105259418488, + 0.012913745827972889, + 0.03969712182879448, + 0.04488351568579674, + 0.016185909509658813, + 0.05102523788809776, + 0.02243378758430481, + 0.030377786606550217, + 0.03307406231760979, + 0.02904777228832245, + 0.05009010061621666, + 0.06063706427812576, + 0.06558306515216827, + 0.00566762825474143, + 0.003641044721007347, + 0.005232359282672405, + 0.004216653760522604, + 0.004749600775539875, + 0.010216319002211094, + 0.005643702112138271, + 0.0056057218462228775, + 0.012944287620484829, + 0.011379336938261986, + 0.009129172191023827, + 0.009332604706287384, + 0.00877335388213396, + 0.019186213612556458, + 0.007429766934365034, + 0.018941078335046768, + 0.006583758629858494, + 0.0072305635549128056, + 0.005037451628595591, + 0.00890118908137083, + 0.006389810238033533, + 0.00843083020299673, + 0.014340613037347794, + 0.009299539029598236, + 0.00454527186229825, + 0.005370996426790953, + 0.003840584075078368, + 0.006349449511617422, + 0.006180976051837206, + 0.004200462717562914, + 0.00899833720177412, + 0.007165132090449333, + 0.0025372691452503204, + 0.0022079930640757084, + 0.0019880200270563364, + 0.0026769437827169895, + 0.003235669108107686, + 0.003208073088899255, + 0.004055787809193134, + 0.004520030692219734, + 0.0055066924542188644, + 0.005219653248786926, + 0.004046597518026829, + 0.007413686718791723, + 0.00816984474658966, + 0.00737648643553257, + 0.0068968492560088634, + 0.005636512767523527, + 0.011231902986764908, + 0.005450834054499865, + 0.004303479567170143, + 0.003803675062954426, + 0.0048496718518435955, + 0.007100686430931091, + 0.0030702983494848013, + 0.01421890128403902, + 0.005417443346232176, + 0.005841589067131281, + 0.004570588003844023, + 0.014890054240822792, + 0.006838918197900057, + 0.009234395809471607, + 0.007150072604417801, + 0.012158665806055069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.6934534907341003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330414578318596 + ], + "opposite_logprob_tokens": [ + -3.7733042240142822 + ], + "logprob_reward": -0.02330414578318596, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330414578318596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5767908096313477, + 0.06263680011034012, + 0.05639998987317085, + 0.01822025328874588, + 0.15428826212882996, + 0.03468051180243492, + 0.007477218750864267, + 0.03582582622766495, + 0.04431791231036186, + 0.009362369775772095 + ], + "credit_map": [ + 0.1931946724653244, + 0.01712076924741268, + 0.016206176951527596, + 0.049403298646211624, + 0.01542356051504612, + 0.23406797647476196, + 0.018823396414518356, + 0.03255099803209305, + 0.00893484428524971, + 0.005359329283237457, + 0.007717338390648365, + 0.007340417243540287, + 0.0037926877848803997, + 0.008504999801516533, + 0.012080113403499126, + 0.008907073177397251, + 0.00661027617752552, + 0.0038972124457359314, + 0.009149503894150257, + 0.007475752849131823, + 0.006049616262316704, + 0.009265802800655365, + 0.007982155308127403, + 0.005969669204205275, + 0.0017697092844173312, + 0.0030527845956385136, + 0.0019624540582299232, + 0.0018781287362799048, + 0.0026352705899626017, + 0.002753220498561859, + 0.0018617947353050113, + 0.0023068913724273443, + 0.014206226915121078, + 0.013043491169810295, + 0.012045294046401978, + 0.014318006113171577, + 0.024533791467547417, + 0.01950984261929989, + 0.03525366261601448, + 0.02137794904410839, + 0.0034654634073376656, + 0.002859134692698717, + 0.0034838663414120674, + 0.003405441762879491, + 0.0036982193123549223, + 0.0041106385178864, + 0.00848398171365261, + 0.005173766519874334, + 0.0007343958131968975, + 0.0006535243592225015, + 0.000721100892405957, + 0.000892311567440629, + 0.0007395150023512542, + 0.0007167448638938367, + 0.0017496109940111637, + 0.0012700152583420277, + 0.00444862199947238, + 0.006696321070194244, + 0.003277387237176299, + 0.005411198828369379, + 0.003219826379790902, + 0.00426073232665658, + 0.0035950655583292246, + 0.00491667166352272, + 0.012538807466626167, + 0.0030952137894928455, + 0.003052301472052932, + 0.0026228763163089752, + 0.0017340078484266996, + 0.004734060727059841, + 0.002092883922159672, + 0.014447759836912155, + 0.0008330291020683944, + 0.0012399080442264676, + 0.0005940409610047936, + 0.0012966027716174722, + 0.0013882216298952699, + 0.0014319466426968575, + 0.0008875556522980332, + 0.0016910654958337545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.02330414578318596 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038303077220917 + ], + "opposite_logprob_tokens": [ + -1.9103829860687256 + ], + "logprob_reward": -0.16038303077220917, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038303077220917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5175233483314514, + 0.11312786489725113, + 0.07919038087129593, + 0.014660561457276344, + 0.03729613125324249, + 0.03186680004000664, + 0.019986560568213463, + 0.07292318344116211, + 0.09215369820594788, + 0.02127152495086193 + ], + "credit_map": [ + 0.16540710628032684, + 0.017695505172014236, + 0.015545951202511787, + 0.044486433267593384, + 0.014497264288365841, + 0.19935055077075958, + 0.03369338437914848, + 0.0268471110612154, + 0.01766403764486313, + 0.00705846818163991, + 0.010483265854418278, + 0.01066453754901886, + 0.007952634245157242, + 0.015690520405769348, + 0.020511077716946602, + 0.023103313520550728, + 0.008832072839140892, + 0.004845939110964537, + 0.009307313710451126, + 0.01001738104969263, + 0.006843881215900183, + 0.014755617827177048, + 0.015621040016412735, + 0.008967123925685883, + 0.0020238603465259075, + 0.0010302207665517926, + 0.0012580007314682007, + 0.0014617249835282564, + 0.0011166817275807261, + 0.004562671296298504, + 0.0012325551360845566, + 0.0019748436752706766, + 0.003736364422366023, + 0.002567305928096175, + 0.0027372194454073906, + 0.0035057500936090946, + 0.004916081205010414, + 0.007257068529725075, + 0.00856864545494318, + 0.0040076919831335545, + 0.002417362527921796, + 0.003078284440562129, + 0.0027074147947132587, + 0.003184732049703598, + 0.003042846452444792, + 0.0033374291379004717, + 0.009733080863952637, + 0.004365646280348301, + 0.0016097818734124303, + 0.0016773135866969824, + 0.0017119443509727716, + 0.0038983344566076994, + 0.0026060331147164106, + 0.0016099084168672562, + 0.004117315169423819, + 0.0027559262234717607, + 0.007658970542252064, + 0.008090415969491005, + 0.007300252094864845, + 0.008775611408054829, + 0.012717674486339092, + 0.009093856438994408, + 0.00791288074105978, + 0.011373513378202915, + 0.025002464652061462, + 0.005987048149108887, + 0.005550742615014315, + 0.004163043107837439, + 0.003825640305876732, + 0.014081504195928574, + 0.0044607254676520824, + 0.029082514345645905, + 0.002230211393907666, + 0.002885655267164111, + 0.001432126620784402, + 0.0030337905045598745, + 0.0022021979093551636, + 0.003364692907780409, + 0.0030927173793315887, + 0.0030301299411803484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.16038303077220917 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023290403187274933 + ], + "opposite_logprob_tokens": [ + -3.7732903957366943 + ], + "logprob_reward": -0.023290403187274933, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023290403187274933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11336725205183029, + 0.13393180072307587, + 0.07500335574150085, + 0.020826080814003944, + 0.058519396930933, + 0.03539874032139778, + 0.0543648824095726, + 0.3765624165534973, + 0.10137999802827835, + 0.030646100640296936 + ], + "credit_map": [ + 0.013986187987029552, + 0.009787348099052906, + 0.009058000519871712, + 0.011267000809311867, + 0.008142581209540367, + 0.019949447363615036, + 0.0324108749628067, + 0.008765799924731255, + 0.01895037479698658, + 0.009420515038073063, + 0.012981955893337727, + 0.014214164577424526, + 0.010691015049815178, + 0.01801348105072975, + 0.02888374775648117, + 0.020776545628905296, + 0.008500265888869762, + 0.004659510217607021, + 0.0048249089159071445, + 0.010172254405915737, + 0.006648021284490824, + 0.017825372517108917, + 0.01326893363147974, + 0.009104093536734581, + 0.00246389489620924, + 0.0033253871370106936, + 0.0018094420665875077, + 0.0025540438946336508, + 0.0015230829594656825, + 0.0036108612548559904, + 0.0022888570092618465, + 0.003250510897487402, + 0.005133187863975763, + 0.003950192127376795, + 0.005511420778930187, + 0.006287598516792059, + 0.008379662409424782, + 0.005713504273444414, + 0.01519104465842247, + 0.008352779783308506, + 0.0030324477702379227, + 0.002704896032810211, + 0.004288625903427601, + 0.0044214436784386635, + 0.006476312875747681, + 0.004165618680417538, + 0.006049099378287792, + 0.004260294605046511, + 0.0049995849840343, + 0.003950423561036587, + 0.004487441852688789, + 0.005688794422894716, + 0.003975600469857454, + 0.004343722481280565, + 0.01727592758834362, + 0.009643386118113995, + 0.0430251844227314, + 0.040521759539842606, + 0.030409350991249084, + 0.04603986069560051, + 0.07157420367002487, + 0.057528115808963776, + 0.04605366662144661, + 0.04141024500131607, + 0.02569863758981228, + 0.007559077348560095, + 0.006886039860546589, + 0.005135711748152971, + 0.005002839956432581, + 0.013454734347760677, + 0.005442757159471512, + 0.032200202345848083, + 0.0025330709759145975, + 0.004867884796112776, + 0.0037175994366407394, + 0.0036929668858647346, + 0.003132648067548871, + 0.003267175517976284, + 0.005758765619248152, + 0.003675988642498851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.023290403187274933 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly upward angle, the scene takes on a unique stop-motion quality. Each frame feels tangible, with tiny, almost imperceptible jumps between movements, imbuing the characters and objects with a handmade, slightly jerky, yet highly expressive quality. The image focuses on an elderly man with a deeply wrinkled face, his thinning white hair plastered against a weathered scalp, his eyes kind yet slightly weary. He stands in a dimly lit, sparsely furnished wooden house, set against a backdrop of rough, dark wood walls. A soft, warm light streams in through a small side window, vividly outlining every crease on his face and illuminating his plain, earth-brown linen shirt. With hands punctuated by age spots and bulging veins, he carefully, almost reverently, holds a simple, unglazed stoneware bowl, its surface rough to the touch, bearing the unmistakable marks of handcraft. Due to the slight but constant trembling of his hands, tiny ripples appeared on the clear water in the bowl, reflecting the faint light from the window. He slowly brought the rim of the bowl to his chapped lips, tilted his head slightly, and took a deliberate sip. The whole movement was imbued with the weight of time. The warm earth tones and strong contrasts of light and dark conveyed a sense of quiet dignity and a moving atmosphere of contemplation.", + "seed": 11842, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_12/index_1727_seed_11842_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.16516422973945738, + "vlm_logprob": -0.16516422973945738, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.16516422973945738, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20786738395690918 + ], + "opposite_logprob_tokens": [ + -2.332867383956909 + ], + "logprob_reward": -0.20786738395690918, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.20786738395690918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27976465225219727, + 0.11661702394485474, + 0.039443813264369965, + 0.1862582266330719, + 0.1206892728805542, + 0.038691069930791855, + 0.0679141953587532, + 0.044578347355127335, + 0.03643394634127617, + 0.06960941851139069 + ], + "credit_map": [ + 0.035040631890296936, + 0.015161264687776566, + 0.02405100129544735, + 0.042259130626916885, + 0.05146612226963043, + 0.038436029106378555, + 0.04838177561759949, + 0.024968694895505905, + 0.022459516301751137, + 0.005020781885832548, + 0.013810377568006516, + 0.012738213874399662, + 0.017574945464730263, + 0.015973150730133057, + 0.01915815658867359, + 0.00988188199698925, + 0.00450902571901679, + 0.0028169513680040836, + 0.0033249289263039827, + 0.0053713079541921616, + 0.008223245851695538, + 0.004434882663190365, + 0.006630762480199337, + 0.0041327103972435, + 0.014552862383425236, + 0.012700118124485016, + 0.01881919801235199, + 0.024197498336434364, + 0.04402819648385048, + 0.02324248105287552, + 0.031769391149282455, + 0.016948482021689415, + 0.014245073311030865, + 0.005622429773211479, + 0.01323656365275383, + 0.0165705569088459, + 0.023085925728082657, + 0.021630795672535896, + 0.010319962166249752, + 0.015977958217263222, + 0.004993861075490713, + 0.002163692843168974, + 0.003660290502011776, + 0.003578085219487548, + 0.006958694197237492, + 0.007572810165584087, + 0.004632490687072277, + 0.005131145939230919, + 0.009284867905080318, + 0.004025034606456757, + 0.008698119781911373, + 0.007372537162154913, + 0.011380521580576897, + 0.010984271764755249, + 0.00681591359898448, + 0.009352927096188068, + 0.006397504359483719, + 0.0030668694525957108, + 0.007496429607272148, + 0.005391110200434923, + 0.004386625252664089, + 0.005985623225569725, + 0.0062778121791779995, + 0.005576373543590307, + 0.006710302084684372, + 0.0019387133652344346, + 0.006155182141810656, + 0.0032897056080400944, + 0.004418078809976578, + 0.005203865002840757, + 0.003985654097050428, + 0.004732444882392883, + 0.009678170084953308, + 0.002911035204306245, + 0.011004498228430748, + 0.004063035361468792, + 0.007352363783866167, + 0.01671641506254673, + 0.007513292133808136, + 0.010370606556534767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.20786738395690918 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234078168869019 + ], + "opposite_logprob_tokens": [ + -0.8984078168869019 + ], + "logprob_reward": -0.5234078168869019, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234078168869019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3222697675228119, + 0.07917127758264542, + 0.04418887570500374, + 0.04510249197483063, + 0.07765226066112518, + 0.08133473247289658, + 0.06615880876779556, + 0.04593576118350029, + 0.030155781656503677, + 0.20803026854991913 + ], + "credit_map": [ + 0.0334809273481369, + 0.01644456386566162, + 0.0262216217815876, + 0.052917104214429855, + 0.06147145479917526, + 0.04424490034580231, + 0.05923169478774071, + 0.02825748734176159, + 0.014798303134739399, + 0.0038313616532832384, + 0.00991785153746605, + 0.010236261412501335, + 0.012853519059717655, + 0.010567547753453255, + 0.010915794409811497, + 0.006050639320164919, + 0.003726972034201026, + 0.002553891157731414, + 0.004622569307684898, + 0.004631499759852886, + 0.011312260292470455, + 0.005599280819296837, + 0.007386812008917332, + 0.004355586599558592, + 0.004021483473479748, + 0.0024554997216910124, + 0.00442085275426507, + 0.004476191941648722, + 0.005948407109826803, + 0.008091973140835762, + 0.008989105932414532, + 0.006698976270854473, + 0.009533912874758244, + 0.004244724754244089, + 0.009934221394360065, + 0.009932540357112885, + 0.010313044302165508, + 0.01392996683716774, + 0.0079489191994071, + 0.011814919300377369, + 0.009928205981850624, + 0.004464031662791967, + 0.010296815074980259, + 0.011263027787208557, + 0.01732153818011284, + 0.01084381714463234, + 0.009896419011056423, + 0.007320876698940992, + 0.007641166914254427, + 0.004510922823101282, + 0.006384754087775946, + 0.010172262787818909, + 0.013098683208227158, + 0.00800919160246849, + 0.009616177529096603, + 0.006725643761456013, + 0.005869120825082064, + 0.0024375387001782656, + 0.006428683176636696, + 0.005299055483192205, + 0.006107328925281763, + 0.005604431498795748, + 0.008503798395395279, + 0.005685801617801189, + 0.006744214333593845, + 0.0014038702938705683, + 0.003940138500183821, + 0.0037537598982453346, + 0.003312243614345789, + 0.004279889166355133, + 0.0038841748610138893, + 0.0028374872636049986, + 0.03015884943306446, + 0.012111445888876915, + 0.02619660273194313, + 0.01993645168840885, + 0.03296949341893196, + 0.03073662519454956, + 0.03133435919880867, + 0.024586418643593788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.5234078168869019 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002825318370014429 + ], + "opposite_logprob_tokens": [ + -5.8778252601623535 + ], + "logprob_reward": -0.002825318370014429, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002825318370014429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3366893231868744, + 0.16303491592407227, + 0.10723913460969925, + 0.06146537512540817, + 0.10301051288843155, + 0.026325207203626633, + 0.04150905832648277, + 0.02971697598695755, + 0.03375178202986717, + 0.09725777059793472 + ], + "credit_map": [ + 0.03101569041609764, + 0.016144290566444397, + 0.028375839814543724, + 0.057684920728206635, + 0.06954911351203918, + 0.04215821623802185, + 0.06090303882956505, + 0.030858172103762627, + 0.022185077890753746, + 0.00676456606015563, + 0.016157332807779312, + 0.01520781684666872, + 0.027918251231312752, + 0.021527010947465897, + 0.03571893274784088, + 0.017555899918079376, + 0.010341907851397991, + 0.0073349252343177795, + 0.010629326105117798, + 0.01340675912797451, + 0.02783866785466671, + 0.014501864090561867, + 0.01268562488257885, + 0.010500042699277401, + 0.004974705632776022, + 0.004022957291454077, + 0.00634283572435379, + 0.006902549881488085, + 0.013208725489675999, + 0.009212875738739967, + 0.008758684620261192, + 0.008042031899094582, + 0.010461188852787018, + 0.004837049636989832, + 0.011650403030216694, + 0.015059557743370533, + 0.013200218789279461, + 0.02158663421869278, + 0.014026517979800701, + 0.012188931927084923, + 0.004321778658777475, + 0.0011437706416472793, + 0.0028024371713399887, + 0.0024629284162074327, + 0.005018793512135744, + 0.004806853830814362, + 0.002065347507596016, + 0.0037032943218946457, + 0.0049072690308094025, + 0.0026285515632480383, + 0.004833332262933254, + 0.005810210481286049, + 0.006761821918189526, + 0.006012440659105778, + 0.0061679407954216, + 0.0043874881230294704, + 0.003984145354479551, + 0.0019733773078769445, + 0.00431737070903182, + 0.003073073923587799, + 0.00513098482042551, + 0.004179213661700487, + 0.003516520606353879, + 0.0035422868095338345, + 0.008466036058962345, + 0.0015668397536501288, + 0.006003542337566614, + 0.0027107293717563152, + 0.0029075471684336662, + 0.005069355480372906, + 0.0036493707448244095, + 0.003378357272595167, + 0.013508157804608345, + 0.004988277330994606, + 0.013348612003028393, + 0.005930527579039335, + 0.011099115014076233, + 0.02312071993947029, + 0.01032500620931387, + 0.014937344007194042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.002825318370014429 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866613447666168 + ], + "opposite_logprob_tokens": [ + -3.04866623878479 + ], + "logprob_reward": -0.04866613447666168, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866613447666168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3939788341522217, + 0.08753662556409836, + 0.04132111743092537, + 0.03552285581827164, + 0.07661578059196472, + 0.08415607362985611, + 0.06277331709861755, + 0.05654627084732056, + 0.04152727127075195, + 0.12002180516719818 + ], + "credit_map": [ + 0.037415340542793274, + 0.020606407895684242, + 0.03352918475866318, + 0.07191717624664307, + 0.07509606331586838, + 0.05241946130990982, + 0.06758171319961548, + 0.035413503646850586, + 0.015915151685476303, + 0.00453892070800066, + 0.012215545400977135, + 0.008860697969794273, + 0.013727028854191303, + 0.011506740003824234, + 0.01042049191892147, + 0.010352051816880703, + 0.004514762666076422, + 0.004111866932362318, + 0.004878917708992958, + 0.004455753602087498, + 0.009032969363033772, + 0.005056231282651424, + 0.005698358174413443, + 0.003572259796783328, + 0.0028568855486810207, + 0.0018928989302366972, + 0.003585835685953498, + 0.004677816294133663, + 0.005840002093464136, + 0.005170486867427826, + 0.0065499115735292435, + 0.004949020221829414, + 0.007030153647065163, + 0.003296396229416132, + 0.007678685709834099, + 0.010317458771169186, + 0.010637564584612846, + 0.01671871915459633, + 0.010862539522349834, + 0.01007426530122757, + 0.007789270486682653, + 0.0034318258985877037, + 0.01076553389430046, + 0.009015927091240883, + 0.011792979203164577, + 0.016219420358538628, + 0.01052612904459238, + 0.014614986255764961, + 0.00891921203583479, + 0.0033818401861935854, + 0.008849185891449451, + 0.008346844464540482, + 0.01200918573886156, + 0.008230103179812431, + 0.006130974739789963, + 0.0069059706293046474, + 0.007234697230160236, + 0.004224078729748726, + 0.008981353603303432, + 0.0067072962410748005, + 0.008228065446019173, + 0.006731974892318249, + 0.006249892525374889, + 0.008188911713659763, + 0.006008376367390156, + 0.0018248484702780843, + 0.007185273338109255, + 0.005340986419469118, + 0.006318544968962669, + 0.004967596847563982, + 0.005753656383603811, + 0.0041279904544353485, + 0.018739787861704826, + 0.006294103339314461, + 0.016248252242803574, + 0.00865965150296688, + 0.015210517682135105, + 0.025065042078495026, + 0.014180806465446949, + 0.015623643063008785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.04866613447666168 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430544950067997 + ], + "opposite_logprob_tokens": [ + -3.1680545806884766 + ], + "logprob_reward": -0.0430544950067997, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430544950067997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18642307817935944, + 0.08940955251455307, + 0.03343275934457779, + 0.0902639701962471, + 0.1269904226064682, + 0.06663402169942856, + 0.11054372042417526, + 0.07196952402591705, + 0.07181502878665924, + 0.1525178849697113 + ], + "credit_map": [ + 0.020481647923588753, + 0.008447540923953056, + 0.017406225204467773, + 0.025135161355137825, + 0.03289606049656868, + 0.02699807658791542, + 0.0320902019739151, + 0.02296815812587738, + 0.019639918580651283, + 0.004485899582505226, + 0.012166833505034447, + 0.01023185532540083, + 0.013799117878079414, + 0.013063671998679638, + 0.009635846130549908, + 0.00638640858232975, + 0.003265487030148506, + 0.0023439135402441025, + 0.004245379474014044, + 0.003913622349500656, + 0.0075081102550029755, + 0.004820722620934248, + 0.003756975056603551, + 0.0035785501822829247, + 0.009276168420910835, + 0.006404695101082325, + 0.010322129353880882, + 0.01322726346552372, + 0.01864534243941307, + 0.011890827678143978, + 0.011126681230962276, + 0.009370863437652588, + 0.012499894015491009, + 0.004590582102537155, + 0.015766294673085213, + 0.013977304100990295, + 0.017500516027212143, + 0.02963942289352417, + 0.016667086631059647, + 0.016349326819181442, + 0.00728085171431303, + 0.004036313854157925, + 0.007812583819031715, + 0.009936842136085033, + 0.011579707264900208, + 0.009389997459948063, + 0.007494235411286354, + 0.00910349190235138, + 0.0134152602404356, + 0.005748630501329899, + 0.011104271747171879, + 0.019366078078746796, + 0.023764226585626602, + 0.013081866316497326, + 0.014443566091358662, + 0.009619826450943947, + 0.008729943074285984, + 0.004872939083725214, + 0.008992237970232964, + 0.008252120576798916, + 0.01001382153481245, + 0.009519286453723907, + 0.012177180498838425, + 0.009411990642547607, + 0.007702534087002277, + 0.003960841801017523, + 0.010626062750816345, + 0.010105741210281849, + 0.010391400195658207, + 0.01018162164837122, + 0.010460807010531425, + 0.008386022411286831, + 0.016630396246910095, + 0.009107321500778198, + 0.019422616809606552, + 0.014581544324755669, + 0.022204028442502022, + 0.03047400526702404, + 0.017204808071255684, + 0.02289316989481449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.0430544950067997 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly upward angle, the scene takes on a unique stop-motion quality. Each frame feels tangible, with tiny, almost imperceptible jumps between movements, imbuing the characters and objects with a handmade, slightly jerky, yet highly expressive quality. The image focuses on an elderly man with a deeply wrinkled face, his thinning white hair plastered against a weathered scalp, his eyes kind yet slightly weary. He stands in a dimly lit, sparsely furnished wooden house, set against a backdrop of rough, dark wood walls. A soft, warm light streams in through a small side window, vividly outlining every crease on his face and illuminating his plain, earth-brown linen shirt. With hands punctuated by age spots and bulging veins, he carefully, almost reverently, holds a simple, unglazed stoneware bowl, its surface rough to the touch, bearing the unmistakable marks of handcraft. Due to the slight but constant trembling of his hands, tiny ripples appeared on the clear water in the bowl, reflecting the faint light from the window. He slowly brought the rim of the bowl to his chapped lips, tilted his head slightly, and took a deliberate sip. The whole movement was imbued with the weight of time. The warm earth tones and strong contrasts of light and dark conveyed a sense of quiet dignity and a moving atmosphere of contemplation.", + "seed": 11843, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_12/index_1727_seed_11843_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2259257594589144, + "vlm_logprob": -0.2259257594589144, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.2259257594589144, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3719492554664612 + ], + "opposite_logprob_tokens": [ + -1.7469491958618164 + ], + "logprob_reward": -0.3719492554664612, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.3719492554664612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10346345603466034, + 0.0570380873978138, + 0.06104178726673126, + 0.04502032697200775, + 0.07802999019622803, + 0.04457823187112808, + 0.4051654636859894, + 0.06846488267183304, + 0.07047829031944275, + 0.06671948730945587 + ], + "credit_map": [ + 0.015231933444738388, + 0.007559759542346001, + 0.013407377526164055, + 0.009820569306612015, + 0.015849679708480835, + 0.009262706153094769, + 0.013786014169454575, + 0.01854541338980198, + 0.007879061624407768, + 0.002822690177708864, + 0.006825949531048536, + 0.0057495045475661755, + 0.010265938937664032, + 0.005932399537414312, + 0.009669995866715908, + 0.007892546243965626, + 0.009480413980782032, + 0.0032268473878502846, + 0.0053595807403326035, + 0.005424012430012226, + 0.009562444873154163, + 0.007831453345716, + 0.012312263250350952, + 0.007844771258533001, + 0.006682960316538811, + 0.00241782795637846, + 0.004868638701736927, + 0.008222020231187344, + 0.0072289942763745785, + 0.003913348075002432, + 0.0064995684660971165, + 0.005186969414353371, + 0.01249595358967781, + 0.0070738960057497025, + 0.00612519821152091, + 0.008148477412760258, + 0.010563419200479984, + 0.01087730098515749, + 0.008576741442084312, + 0.014169003814458847, + 0.006118706427514553, + 0.003208168549463153, + 0.004560286644846201, + 0.004538506735116243, + 0.010243253782391548, + 0.0031984588131308556, + 0.0063071600161492825, + 0.006403691601008177, + 0.062474265694618225, + 0.02326437085866928, + 0.0591588132083416, + 0.07106030732393265, + 0.06291107833385468, + 0.023488543927669525, + 0.04490463435649872, + 0.057903435081243515, + 0.010769310407340527, + 0.005248743109405041, + 0.00922077614814043, + 0.012573824264109135, + 0.008446168154478073, + 0.003624941688030958, + 0.008396749384701252, + 0.010184367187321186, + 0.011693509295582771, + 0.007460711523890495, + 0.011743689887225628, + 0.008668624795973301, + 0.008848382160067558, + 0.003988761454820633, + 0.009958111681044102, + 0.00811650138348341, + 0.011875143274664879, + 0.00830820482224226, + 0.005610677879303694, + 0.005772542208433151, + 0.01332724280655384, + 0.00528064277023077, + 0.009947553277015686, + 0.006597480271011591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.3719492554664612 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762665271759033 + ], + "opposite_logprob_tokens": [ + -0.8262665271759033 + ], + "logprob_reward": -0.5762665271759033, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762665271759033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09135208278894424, + 0.05805055797100067, + 0.08938053995370865, + 0.1268816888332367, + 0.11786588281393051, + 0.03913077339529991, + 0.13800714910030365, + 0.07907730340957642, + 0.0867529809474945, + 0.17350102961063385 + ], + "credit_map": [ + 0.012740667909383774, + 0.007654718589037657, + 0.007929261773824692, + 0.010109135881066322, + 0.015601271763443947, + 0.015059682540595531, + 0.011405091732740402, + 0.010852253995835781, + 0.008596854284405708, + 0.002698692725971341, + 0.007254682015627623, + 0.005142720881849527, + 0.00676024379208684, + 0.007475677877664566, + 0.009386622346937656, + 0.010735063813626766, + 0.012983758002519608, + 0.004508655052632093, + 0.008533565327525139, + 0.007131603080779314, + 0.01724444329738617, + 0.014039944857358932, + 0.013961692340672016, + 0.010976879857480526, + 0.018573403358459473, + 0.006736533250659704, + 0.017562786117196083, + 0.02889331430196762, + 0.013194466941058636, + 0.008493596687912941, + 0.017485538497567177, + 0.015942037105560303, + 0.017631154507398605, + 0.007400220725685358, + 0.014521385543048382, + 0.023151591420173645, + 0.01361228246241808, + 0.00941572804003954, + 0.017199188470840454, + 0.014934324659407139, + 0.006156666204333305, + 0.0024113729596138, + 0.004669422283768654, + 0.0049985856749117374, + 0.007095739711076021, + 0.002272294834256172, + 0.0057257371954619884, + 0.005800956394523382, + 0.0204026997089386, + 0.0075166476890444756, + 0.01787610910832882, + 0.025926101952791214, + 0.020832370966672897, + 0.008025141432881355, + 0.02002476342022419, + 0.017403310164809227, + 0.013111197389662266, + 0.0045669362880289555, + 0.010028650052845478, + 0.018080273643136024, + 0.009702319279313087, + 0.0038022249937057495, + 0.009190664626657963, + 0.010595042258501053, + 0.01394682377576828, + 0.014562038704752922, + 0.010400715284049511, + 0.009356647729873657, + 0.010899514891207218, + 0.006812307517975569, + 0.010224058292806149, + 0.010550877079367638, + 0.029345907270908356, + 0.018879273906350136, + 0.015522749163210392, + 0.021103622391819954, + 0.03846905007958412, + 0.00941359531134367, + 0.027122613042593002, + 0.013644208200275898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.5762665271759033 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002831618534401059 + ], + "opposite_logprob_tokens": [ + -5.87783145904541 + ], + "logprob_reward": -0.002831618534401059, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002831618534401059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2448144108057022, + 0.1915816217660904, + 0.06194580718874931, + 0.05345642566680908, + 0.06037753447890282, + 0.11266940087080002, + 0.06550101190805435, + 0.047315169125795364, + 0.043323975056409836, + 0.11901462078094482 + ], + "credit_map": [ + 0.03705906495451927, + 0.014034385792911053, + 0.0336189866065979, + 0.026140857487916946, + 0.037281375378370285, + 0.026425790041685104, + 0.03405572474002838, + 0.03619825094938278, + 0.029062479734420776, + 0.010087116621434689, + 0.01843976601958275, + 0.01749391108751297, + 0.02846800722181797, + 0.021303249523043633, + 0.03525915369391441, + 0.03146793693304062, + 0.009606434963643551, + 0.0027786290738731623, + 0.006209359969943762, + 0.006457097362726927, + 0.010913314297795296, + 0.007316051982343197, + 0.011171188205480576, + 0.007493733894079924, + 0.00824406836181879, + 0.002595587633550167, + 0.006806353107094765, + 0.009164568036794662, + 0.008514042012393475, + 0.003282003104686737, + 0.008909909054636955, + 0.00593989435583353, + 0.008315684273838997, + 0.0036845889408141375, + 0.0058436463586986065, + 0.0054323249496519566, + 0.014899441041052341, + 0.005018128547817469, + 0.007755700498819351, + 0.009428022429347038, + 0.022507280111312866, + 0.006495215464383364, + 0.008929131552577019, + 0.013066094368696213, + 0.015356697142124176, + 0.009632785804569721, + 0.016684971749782562, + 0.019997239112854004, + 0.009940952993929386, + 0.003996050916612148, + 0.009239565581083298, + 0.01212841086089611, + 0.011127734556794167, + 0.003857051720842719, + 0.0061574289575219154, + 0.009053818881511688, + 0.008092696778476238, + 0.003758994396775961, + 0.006249208934605122, + 0.00857500359416008, + 0.007613224443048239, + 0.0021620867773890495, + 0.0062957205809652805, + 0.004568235948681831, + 0.008342037908732891, + 0.00636177696287632, + 0.005979820154607296, + 0.006168666295707226, + 0.005400585941970348, + 0.0024441310670226812, + 0.004516595043241978, + 0.0041103670373559, + 0.018312184140086174, + 0.012632563710212708, + 0.015714257955551147, + 0.019807176664471626, + 0.015780653804540634, + 0.010988421738147736, + 0.015523099340498447, + 0.010256268084049225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.002831618534401059 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16039104759693146 + ], + "opposite_logprob_tokens": [ + -1.910391092300415 + ], + "logprob_reward": -0.16039104759693146, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16039104759693146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1619809865951538, + 0.06348815560340881, + 0.07320000976324081, + 0.05757870897650719, + 0.07640518248081207, + 0.1082562804222107, + 0.17126570641994476, + 0.16785408556461334, + 0.03368169814348221, + 0.08628920465707779 + ], + "credit_map": [ + 0.022633908316493034, + 0.009256604127585888, + 0.01696673221886158, + 0.012672937475144863, + 0.025614505633711815, + 0.01038013119250536, + 0.033118750900030136, + 0.031337399035692215, + 0.009163313545286655, + 0.00328553793951869, + 0.007027519401162863, + 0.005949954967945814, + 0.00762611348181963, + 0.009033235721290112, + 0.011224033311009407, + 0.010178438387811184, + 0.011037481017410755, + 0.004247871693223715, + 0.0060384864918887615, + 0.006685775704681873, + 0.009928555227816105, + 0.009916148148477077, + 0.015075315721333027, + 0.010270373895764351, + 0.009326516650617123, + 0.003510301234200597, + 0.004637531936168671, + 0.005949951708316803, + 0.01157644297927618, + 0.005532044917345047, + 0.00975877232849598, + 0.007287135347723961, + 0.012121140956878662, + 0.004359852988272905, + 0.008997118100523949, + 0.01219523511826992, + 0.009817930869758129, + 0.0049494109116494656, + 0.009847614914178848, + 0.014116877689957619, + 0.019009871408343315, + 0.007065833080559969, + 0.006550934631377459, + 0.012134695425629616, + 0.014154401607811451, + 0.012140106409788132, + 0.016248373314738274, + 0.02095205895602703, + 0.02526082657277584, + 0.011006791144609451, + 0.018461186438798904, + 0.020181842148303986, + 0.03852631151676178, + 0.012449168600142002, + 0.01963752880692482, + 0.025742027908563614, + 0.024756135419011116, + 0.014461657963693142, + 0.01919781044125557, + 0.0176044050604105, + 0.017123214900493622, + 0.015183139592409134, + 0.02686699852347374, + 0.032660700380802155, + 0.004989296197891235, + 0.0031035393476486206, + 0.0054869456216692924, + 0.0051631866954267025, + 0.004368260037153959, + 0.0024562119506299496, + 0.003881893353536725, + 0.004232362378388643, + 0.014027107506990433, + 0.014639809727668762, + 0.008629050105810165, + 0.008358852937817574, + 0.01605937071144581, + 0.005629780702292919, + 0.011895792558789253, + 0.0070494310930371284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.16039104759693146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video exhibit a stop-motion quality with tiny, imperceptible jumps between movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the main character an elderly man with a deeply wrinkled face, thinning white hair, and age spots on his hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the man holding a simple, unglazed stoneware bowl containing clear water that shows tiny ripples?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man bring the bowl to his lips, tilt his head slightly, and take a deliberate sip?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018190348520874977 + ], + "opposite_logprob_tokens": [ + -4.018190383911133 + ], + "logprob_reward": -0.018190348520874977, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018190348520874977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07071389257907867, + 0.032049067318439484, + 0.05013060197234154, + 0.04571784660220146, + 0.1545155644416809, + 0.07074484974145889, + 0.09995028376579285, + 0.27401334047317505, + 0.1446262001991272, + 0.05753842368721962 + ], + "credit_map": [ + 0.012010201811790466, + 0.004473963752388954, + 0.009060419164597988, + 0.006022337824106216, + 0.007271486800163984, + 0.011383398436009884, + 0.007066427264362574, + 0.013425651006400585, + 0.0045893313363194466, + 0.001503227511420846, + 0.002405793871730566, + 0.0021299030631780624, + 0.006541583687067032, + 0.004186336882412434, + 0.004476475995033979, + 0.006216415204107761, + 0.007440216839313507, + 0.0018149811076000333, + 0.003981649409979582, + 0.0036555668339133263, + 0.008339975960552692, + 0.008271092548966408, + 0.009341422468423843, + 0.0072856927290558815, + 0.007123117335140705, + 0.002570871263742447, + 0.0052078706212341785, + 0.007250702008605003, + 0.0051500918343663216, + 0.004351550247520208, + 0.0060782707296311855, + 0.007985367439687252, + 0.023378292098641396, + 0.009188618510961533, + 0.01815134845674038, + 0.030700774863362312, + 0.018728770315647125, + 0.011870059184730053, + 0.01883150450885296, + 0.02366618812084198, + 0.010859725065529346, + 0.004558937158435583, + 0.00655898405238986, + 0.010028094053268433, + 0.011369023472070694, + 0.007365509867668152, + 0.009047587402164936, + 0.010956978425383568, + 0.01547588873654604, + 0.0051497602835297585, + 0.01157362386584282, + 0.017651932314038277, + 0.015960775315761566, + 0.006350684445351362, + 0.011508935131132603, + 0.016278671100735664, + 0.04526226967573166, + 0.022950317710638046, + 0.02853057160973549, + 0.035162173211574554, + 0.033450592309236526, + 0.01660211570560932, + 0.035161424428224564, + 0.056893859058618546, + 0.025648312643170357, + 0.01499991212040186, + 0.011176052503287792, + 0.019573276862502098, + 0.02521788515150547, + 0.008909427560865879, + 0.022307923063635826, + 0.016793416813015938, + 0.008646690286695957, + 0.007858620956540108, + 0.0070801894180476665, + 0.008742050267755985, + 0.00836050882935524, + 0.005273509304970503, + 0.006195822264999151, + 0.005381025839596987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.018190348520874977 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_13/index_tensor([9668, 9668])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_13/index_tensor([9668, 9668])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b9a7f02e3ea0ec2511286f506824238c4fa4d653 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_13/index_tensor([9668, 9668])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a level angle, slowly pans left, following a clown struggling forward. He wears a vibrant rainbow-hued afro wig, absurdly topped with an oversized blue bowler hat. His face is thickly smeared with white paint, and a dramatic red smile stretches from ear to ear, contrasting with his brow furrowed in intense concentration and his eyes widened. A painted-on blue teardrop dangles precariously beneath his right eye. He wears an extremely baggy, bright yellow jumpsuit covered in giant red and blue polka dots, and a cascading white ruffled neckpiece wraps around his neck. He is riding a shiny chrome unicycle, his legs laboriously pedaling in an uncoordinated yet determined rhythm, his upper body swaying violently from side to side in an effort to maintain balance, his arms fully extended, waving futilely in the air like wings. The bright afternoon sun shines on the wide asphalt road in the park, casting a vivid luster on his colorful clothes and dragging a clear shadow behind him that distorts and stretches as he sways. The whole scene unfolds in a joyful atmosphere full of comic tension.", + "seed": 22376, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_13/index_9668_seed_22376_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08804112849757076, + "vlm_logprob": -0.08804112849757076, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.08804112849757076, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07992565631866455 + ], + "opposite_logprob_tokens": [ + -2.829925537109375 + ], + "logprob_reward": -0.07992565631866455, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07992565631866455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0839909017086029, + 0.091115303337574, + 0.090727798640728, + 0.07893944531679153, + 0.04871683940291405, + 0.06761813163757324, + 0.19750630855560303, + 0.14134494960308075, + 0.09992747008800507, + 0.10011277347803116 + ], + "credit_map": [ + 0.0061021954752504826, + 0.011180438101291656, + 0.017273593693971634, + 0.012063595466315746, + 0.008558950386941433, + 0.009828723967075348, + 0.009534372016787529, + 0.009449038654565811, + 0.006926408037543297, + 0.01800895482301712, + 0.010190004482865334, + 0.013734191656112671, + 0.012000955641269684, + 0.007149006240069866, + 0.013321926817297935, + 0.00978386215865612, + 0.006517153698951006, + 0.012605772353708744, + 0.014689301140606403, + 0.017243199050426483, + 0.010277924127876759, + 0.009080756455659866, + 0.008178909309208393, + 0.012134796939790249, + 0.014139313250780106, + 0.008918385952711105, + 0.009456485509872437, + 0.013359542936086655, + 0.0065492503345012665, + 0.008039550855755806, + 0.009651979431509972, + 0.008824941702187061, + 0.00481564924120903, + 0.005925734527409077, + 0.007511982694268227, + 0.0069070919416844845, + 0.004773900844156742, + 0.005980559159070253, + 0.006992720998823643, + 0.005809206515550613, + 0.00492709968239069, + 0.008200505748391151, + 0.009790947660803795, + 0.011089161969721317, + 0.005643527954816818, + 0.010019912384450436, + 0.010904195718467236, + 0.007042783312499523, + 0.017273562029004097, + 0.024372102692723274, + 0.042620036751031876, + 0.02763487957417965, + 0.016891175881028175, + 0.0290734451264143, + 0.02145085111260414, + 0.018190274015069008, + 0.01229532714933157, + 0.006275232415646315, + 0.03200278431177139, + 0.013834400102496147, + 0.005814045201987028, + 0.029553020372986794, + 0.012433269992470741, + 0.029136894270777702, + 0.007022756617516279, + 0.007661907467991114, + 0.018747733905911446, + 0.0080263651907444, + 0.005546108819544315, + 0.03617140278220177, + 0.005177969112992287, + 0.011573241092264652, + 0.008030949160456657, + 0.009341733530163765, + 0.012661593034863472, + 0.014212372712790966, + 0.009596629068255424, + 0.015377883799374104, + 0.016862891614437103, + 0.014028736390173435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.07992565631866455 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134581446647644 + ], + "opposite_logprob_tokens": [ + -1.3134582042694092 + ], + "logprob_reward": -0.3134581446647644, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134581446647644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10306072980165482, + 0.052248045802116394, + 0.09456101804971695, + 0.06386261433362961, + 0.06600978970527649, + 0.058803390711545944, + 0.06684692949056625, + 0.10916715115308762, + 0.20745812356472015, + 0.17798224091529846 + ], + "credit_map": [ + 0.011387325823307037, + 0.008324826136231422, + 0.01750832423567772, + 0.009785500355064869, + 0.009954242967069149, + 0.012130112387239933, + 0.011185983195900917, + 0.022784411907196045, + 0.005032809916883707, + 0.00706705916672945, + 0.007407947443425655, + 0.008769091218709946, + 0.00449347123503685, + 0.005674319341778755, + 0.006354690529406071, + 0.007448661606758833, + 0.007771567441523075, + 0.014541083946824074, + 0.014985614456236362, + 0.01820150762796402, + 0.00797189585864544, + 0.008229310624301434, + 0.011795895174145699, + 0.011064153164625168, + 0.016688013449311256, + 0.004698267672210932, + 0.008546428754925728, + 0.005245162174105644, + 0.007129662204533815, + 0.00792276207357645, + 0.005192285403609276, + 0.00844003539532423, + 0.0054124463349580765, + 0.011380151845514774, + 0.007559658493846655, + 0.012147814966738224, + 0.00613014679402113, + 0.006338367238640785, + 0.009261068888008595, + 0.007780139334499836, + 0.004836576990783215, + 0.006909807212650776, + 0.00771883362904191, + 0.01153873372823, + 0.005718845874071121, + 0.005795842036604881, + 0.010042717680335045, + 0.006242034025490284, + 0.008749949745833874, + 0.003641924588009715, + 0.0157919991761446, + 0.006717021111398935, + 0.0031708425376564264, + 0.017340412363409996, + 0.004311004187911749, + 0.007123775780200958, + 0.010775748640298843, + 0.011257080361247063, + 0.02009687013924122, + 0.010515721514821053, + 0.0095293577760458, + 0.01800364814698696, + 0.013440935872495174, + 0.015547786839306355, + 0.012041551992297173, + 0.008389193564653397, + 0.048923660069704056, + 0.013836286962032318, + 0.009537002071738243, + 0.08709099143743515, + 0.008467868901789188, + 0.01917158253490925, + 0.020608829334378242, + 0.01921326480805874, + 0.02167121320962906, + 0.02580750361084938, + 0.01887185499072075, + 0.02486431784927845, + 0.024938788264989853, + 0.02200647071003914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.3134581446647644 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075684800744057 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075684800744057, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075684800744057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10921374708414078, + 0.05946361646056175, + 0.03490881621837616, + 0.07087431848049164, + 0.09077420830726624, + 0.36261186003685, + 0.032726939767599106, + 0.0287358146160841, + 0.029144668951630592, + 0.1815459430217743 + ], + "credit_map": [ + 0.008500959724187851, + 0.014068062417209148, + 0.020736806094646454, + 0.016926761716604233, + 0.00819973275065422, + 0.010463973507285118, + 0.013809981755912304, + 0.01650746539235115, + 0.005754127632826567, + 0.00778178172186017, + 0.009654433466494083, + 0.011870166286826134, + 0.006898333318531513, + 0.00490187294781208, + 0.006756005343049765, + 0.0058469027280807495, + 0.0024268878623843193, + 0.0077461013570427895, + 0.004838481079787016, + 0.0060936445370316505, + 0.003257637144997716, + 0.0029872397426515818, + 0.003605061676353216, + 0.0039537628181278706, + 0.014988763257861137, + 0.00893890019506216, + 0.009614869952201843, + 0.006837815046310425, + 0.007422547787427902, + 0.008138757199048996, + 0.0060687619261443615, + 0.008863908238708973, + 0.009547870606184006, + 0.01816675253212452, + 0.010234130546450615, + 0.01578536443412304, + 0.007598524913191795, + 0.00778172304853797, + 0.010266867466270924, + 0.011392981745302677, + 0.029593730345368385, + 0.07478921860456467, + 0.036098986864089966, + 0.05654720216989517, + 0.04419736936688423, + 0.027429668232798576, + 0.05151064321398735, + 0.042445071041584015, + 0.003515086602419615, + 0.0019576926715672016, + 0.00884223822504282, + 0.0031996064353734255, + 0.001974292565137148, + 0.006184015888720751, + 0.0025802874006330967, + 0.004473723005503416, + 0.0027629605028778315, + 0.003723316825926304, + 0.0031126984395086765, + 0.00331825390458107, + 0.0027909937780350447, + 0.005194483324885368, + 0.0034768935292959213, + 0.004356216173619032, + 0.002384247025474906, + 0.002670618472620845, + 0.005582783371210098, + 0.0031865162309259176, + 0.002105426974594593, + 0.006468997802585363, + 0.002523197792470455, + 0.004222884774208069, + 0.01612163707613945, + 0.011672548949718475, + 0.023410039022564888, + 0.0219577606767416, + 0.01395539939403534, + 0.0262419655919075, + 0.04302321746945381, + 0.02516338787972927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.011075684800744057 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029804570600390434 + ], + "opposite_logprob_tokens": [ + -3.5298044681549072 + ], + "logprob_reward": -0.029804570600390434, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029804570600390434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14945419132709503, + 0.06639247387647629, + 0.06156911700963974, + 0.09533683210611343, + 0.09782874584197998, + 0.05563519895076752, + 0.056788552552461624, + 0.07883413881063461, + 0.04613059014081955, + 0.2920301854610443 + ], + "credit_map": [ + 0.010768277570605278, + 0.012829281389713287, + 0.02366826869547367, + 0.02536896802484989, + 0.012679136358201504, + 0.014009947888553143, + 0.017991729080677032, + 0.03213857486844063, + 0.0067671495489776134, + 0.007689666468650103, + 0.007211985997855663, + 0.008740619756281376, + 0.008976466953754425, + 0.00825489591807127, + 0.010129466652870178, + 0.008622219786047935, + 0.0053510162979364395, + 0.0071080829948186874, + 0.009033477865159512, + 0.014340807683765888, + 0.005017892457544804, + 0.008543494157493114, + 0.005670633167028427, + 0.006503714714199305, + 0.01320838462561369, + 0.01711052469909191, + 0.012858462519943714, + 0.013201273046433926, + 0.00798912812024355, + 0.009957285597920418, + 0.009313608519732952, + 0.011698164977133274, + 0.014764326624572277, + 0.010888674296438694, + 0.010549676604568958, + 0.015349503606557846, + 0.011432960629463196, + 0.010702533647418022, + 0.013706237077713013, + 0.010434839874505997, + 0.003450254211202264, + 0.005948598962277174, + 0.0066603003069758415, + 0.009515175595879555, + 0.004050820134580135, + 0.006067081354558468, + 0.012569914571940899, + 0.007373054511845112, + 0.003845279337838292, + 0.005162681452929974, + 0.013772774487733841, + 0.00695166876539588, + 0.003654893022030592, + 0.011090841144323349, + 0.005326032638549805, + 0.006984376814216375, + 0.007705373223870993, + 0.007469126023352146, + 0.013332152739167213, + 0.008977498859167099, + 0.005893630441278219, + 0.011751534417271614, + 0.00964149460196495, + 0.01406332477927208, + 0.004050163552165031, + 0.004113673232495785, + 0.006437608506530523, + 0.005522008053958416, + 0.004834015388041735, + 0.005712567828595638, + 0.005405702628195286, + 0.010054850950837135, + 0.0293282438069582, + 0.02801385521888733, + 0.0423327311873436, + 0.02789207734167576, + 0.02911144308745861, + 0.045867253094911575, + 0.05728469789028168, + 0.03219985589385033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.029804570600390434 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059415861032903194 + ], + "opposite_logprob_tokens": [ + -5.130941390991211 + ], + "logprob_reward": -0.0059415861032903194, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059415861032903194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21825693547725677, + 0.053076498210430145, + 0.06737800687551498, + 0.08374562859535217, + 0.13462047278881073, + 0.0669734850525856, + 0.07599066942930222, + 0.058983247727155685, + 0.02930258773267269, + 0.21167248487472534 + ], + "credit_map": [ + 0.014326469041407108, + 0.02179703302681446, + 0.03571081534028053, + 0.03746969997882843, + 0.02106797881424427, + 0.02905118092894554, + 0.02741065062582493, + 0.03142309933900833, + 0.005239532329142094, + 0.004658478312194347, + 0.00719241239130497, + 0.006859185639768839, + 0.006794991437345743, + 0.007272873539477587, + 0.008373552933335304, + 0.006685473956167698, + 0.0055112075060606, + 0.00862171035259962, + 0.011548688635230064, + 0.013391237705945969, + 0.005295764189213514, + 0.007699388079345226, + 0.008451431058347225, + 0.006858580280095339, + 0.017982199788093567, + 0.008492197841405869, + 0.012284860946238041, + 0.007775333244353533, + 0.007924654521048069, + 0.010625790804624557, + 0.007728553842753172, + 0.01093203853815794, + 0.013232861645519733, + 0.021071968600153923, + 0.015649843961000443, + 0.024429824203252792, + 0.009822782129049301, + 0.014929840341210365, + 0.01869700476527214, + 0.016786349937319756, + 0.005323001183569431, + 0.007526895496994257, + 0.008783006109297276, + 0.0068534966558218, + 0.005065451841801405, + 0.009765438735485077, + 0.015379455871880054, + 0.008276743814349174, + 0.010569702833890915, + 0.005961088929325342, + 0.022699914872646332, + 0.005640564486384392, + 0.003772323951125145, + 0.011200927197933197, + 0.0036057825200259686, + 0.0125403618440032, + 0.005173585377633572, + 0.0038401177152991295, + 0.008106326684355736, + 0.00788093451410532, + 0.003394643310457468, + 0.008421415463089943, + 0.006265377625823021, + 0.01590084470808506, + 0.002476603025570512, + 0.004031382501125336, + 0.0035717396531254053, + 0.00346016907133162, + 0.0034715156070888042, + 0.006860434543341398, + 0.002743878634646535, + 0.00268686399795115, + 0.01616382598876953, + 0.016543744131922722, + 0.023819224908947945, + 0.03289562091231346, + 0.01529927458614111, + 0.03595377132296562, + 0.04573611915111542, + 0.025260912254452705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.0059415861032903194 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly pans left, following a clown struggling forward. He wears a vibrant rainbow-hued afro wig, absurdly topped with an oversized blue bowler hat. His face is thickly smeared with white paint, and a dramatic red smile stretches from ear to ear, contrasting with his brow furrowed in intense concentration and his eyes widened. A painted-on blue teardrop dangles precariously beneath his right eye. He wears an extremely baggy, bright yellow jumpsuit covered in giant red and blue polka dots, and a cascading white ruffled neckpiece wraps around his neck. He is riding a shiny chrome unicycle, his legs laboriously pedaling in an uncoordinated yet determined rhythm, his upper body swaying violently from side to side in an effort to maintain balance, his arms fully extended, waving futilely in the air like wings. The bright afternoon sun shines on the wide asphalt road in the park, casting a vivid luster on his colorful clothes and dragging a clear shadow behind him that distorts and stretches as he sways. The whole scene unfolds in a joyful atmosphere full of comic tension.", + "seed": 22377, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_13/index_9668_seed_22377_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10232705250382423, + "vlm_logprob": -0.10232705250382423, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.10232705250382423, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07523182779550552 + ], + "opposite_logprob_tokens": [ + -2.9502317905426025 + ], + "logprob_reward": -0.07523182779550552, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.07523182779550552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18586444854736328, + 0.14991870522499084, + 0.13558383285999298, + 0.06736447662115097, + 0.0982699915766716, + 0.05702986568212509, + 0.06686175614595413, + 0.08050625771284103, + 0.08219434320926666, + 0.0764063373208046 + ], + "credit_map": [ + 0.01603356935083866, + 0.018689610064029694, + 0.017567714676260948, + 0.027845388278365135, + 0.025424005463719368, + 0.025456519797444344, + 0.0245448499917984, + 0.03030276857316494, + 0.01462031900882721, + 0.0181865431368351, + 0.015420579351484776, + 0.03855663165450096, + 0.020484108477830887, + 0.012516016140580177, + 0.02018827013671398, + 0.009946229867637157, + 0.016861412674188614, + 0.01416458934545517, + 0.012927936390042305, + 0.01989843137562275, + 0.01950988732278347, + 0.010434172116219997, + 0.018638091161847115, + 0.02314929850399494, + 0.008078143931925297, + 0.008151214569807053, + 0.009162217378616333, + 0.008650495670735836, + 0.010874984785914421, + 0.008755638264119625, + 0.008684847503900528, + 0.005006926599889994, + 0.012043281458318233, + 0.01060753408819437, + 0.016354570165276527, + 0.009893805719912052, + 0.012770461849868298, + 0.012705824337899685, + 0.010802856646478176, + 0.013091658242046833, + 0.006499802228063345, + 0.006629538256675005, + 0.00944582186639309, + 0.006722236052155495, + 0.007619865704327822, + 0.006727550178766251, + 0.0051865349523723125, + 0.008198514580726624, + 0.0077882194891572, + 0.007273090071976185, + 0.008927877061069012, + 0.009354771114885807, + 0.006822609342634678, + 0.007446355652064085, + 0.009555631317198277, + 0.009693198837339878, + 0.007916271686553955, + 0.011821846477687359, + 0.009925413876771927, + 0.008015311323106289, + 0.008545158430933952, + 0.008520172908902168, + 0.010488724336028099, + 0.015273356810212135, + 0.008908490650355816, + 0.008021669462323189, + 0.010826228186488152, + 0.016610242426395416, + 0.0073463753797113895, + 0.004878056235611439, + 0.019936127588152885, + 0.005667150020599365, + 0.0076082609593868256, + 0.01168997772037983, + 0.00963111873716116, + 0.01205391064286232, + 0.008394554257392883, + 0.01281613390892744, + 0.006508234888315201, + 0.0077041438780725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07523182779550552 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870486319065094 + ], + "opposite_logprob_tokens": [ + -1.137048602104187 + ], + "logprob_reward": -0.3870486319065094, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870486319065094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10560771077871323, + 0.03911712393164635, + 0.09089427441358566, + 0.06533022969961166, + 0.10656564682722092, + 0.11158037185668945, + 0.06195225566625595, + 0.04307330399751663, + 0.3027925193309784, + 0.07308659702539444 + ], + "credit_map": [ + 0.012311124242842197, + 0.01075949426740408, + 0.010444289073348045, + 0.019049404188990593, + 0.009576467797160149, + 0.013254603371024132, + 0.013117445632815361, + 0.017094874754548073, + 0.003676241496577859, + 0.007091016508638859, + 0.0034800670109689236, + 0.005857192911207676, + 0.005510516464710236, + 0.00439512450248003, + 0.004867789335548878, + 0.004239175468683243, + 0.011955835856497288, + 0.0065749650821089745, + 0.007234193850308657, + 0.00852490495890379, + 0.010817471891641617, + 0.007365272846072912, + 0.01981135830283165, + 0.018610280007123947, + 0.006267335265874863, + 0.007332662120461464, + 0.00740029476583004, + 0.00714351749047637, + 0.005668606609106064, + 0.011231670156121254, + 0.009169942699372768, + 0.011116196401417255, + 0.015020443126559258, + 0.0126756327226758, + 0.018528034910559654, + 0.011716865934431553, + 0.014159340411424637, + 0.011914157308638096, + 0.013426169753074646, + 0.00912499614059925, + 0.013870744034647942, + 0.007771821226924658, + 0.019124940037727356, + 0.014007813297212124, + 0.010845703072845936, + 0.011274725198745728, + 0.016242852434515953, + 0.01844177581369877, + 0.009245544672012329, + 0.004461235832422972, + 0.007379226386547089, + 0.006751169450581074, + 0.007204186636954546, + 0.011674000881612301, + 0.00566230108961463, + 0.009574593044817448, + 0.004352481570094824, + 0.0041539017111063, + 0.005543098319321871, + 0.005950151011347771, + 0.005447058007121086, + 0.00460792426019907, + 0.00865810364484787, + 0.00436058547347784, + 0.03307492658495903, + 0.025804119184613228, + 0.03909115120768547, + 0.0686202272772789, + 0.02168833091855049, + 0.01735701784491539, + 0.07198451459407806, + 0.025172239169478416, + 0.005600932985544205, + 0.012265735305845737, + 0.00550869619473815, + 0.015993284061551094, + 0.010832603089511395, + 0.007675076834857464, + 0.009066706523299217, + 0.006143560633063316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.3870486319065094 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978105328977108 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.00978105328977108, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978105328977108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16110260784626007, + 0.02686825394630432, + 0.0948202982544899, + 0.24349752068519592, + 0.11382171511650085, + 0.06538565456867218, + 0.10952577739953995, + 0.10263558477163315, + 0.04901689291000366, + 0.03332574665546417 + ], + "credit_map": [ + 0.020245106890797615, + 0.023696742951869965, + 0.01452052965760231, + 0.02242870070040226, + 0.026909014210104942, + 0.014742461033165455, + 0.02072206884622574, + 0.01783796213567257, + 0.0027823038399219513, + 0.0030174856074154377, + 0.0027560072485357523, + 0.0039661722257733345, + 0.0035576678346842527, + 0.0029647788032889366, + 0.003351739374920726, + 0.00447209645062685, + 0.010374226607382298, + 0.011752868071198463, + 0.007268592715263367, + 0.020679177716374397, + 0.015475805848836899, + 0.00702813733369112, + 0.014978332445025444, + 0.007263146340847015, + 0.024274485185742378, + 0.034090109169483185, + 0.04070037603378296, + 0.038283683359622955, + 0.02169429510831833, + 0.021209871396422386, + 0.04603607952594757, + 0.017208600416779518, + 0.01743045262992382, + 0.01454817596822977, + 0.024067511782050133, + 0.012846790254116058, + 0.011228478513658047, + 0.008962987922132015, + 0.012248854152858257, + 0.012488452717661858, + 0.010707901790738106, + 0.005229714326560497, + 0.010871920734643936, + 0.008658789098262787, + 0.005508846137672663, + 0.007661100011318922, + 0.009876993484795094, + 0.006870378274470568, + 0.014142691157758236, + 0.011606634594500065, + 0.019081760197877884, + 0.011396435089409351, + 0.014004118740558624, + 0.01669640652835369, + 0.013500764034688473, + 0.00909695215523243, + 0.012121403589844704, + 0.015485263429582119, + 0.013672084547579288, + 0.012381142005324364, + 0.013188812881708145, + 0.006590135395526886, + 0.015198948793113232, + 0.013997777365148067, + 0.005941971205174923, + 0.010630950331687927, + 0.007436806336045265, + 0.005333715118467808, + 0.00576873030513525, + 0.0037629385478794575, + 0.006682027596980333, + 0.0034597485791891813, + 0.003941797651350498, + 0.005015566013753414, + 0.004318437073379755, + 0.004437902010977268, + 0.003485544119030237, + 0.00535388570278883, + 0.0034654627088457346, + 0.0033071464858949184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.00978105328977108 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029802139848470688 + ], + "opposite_logprob_tokens": [ + -3.529802083969116 + ], + "logprob_reward": -0.029802139848470688, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029802139848470688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13250644505023956, + 0.05860808864235878, + 0.1158488392829895, + 0.1697925478219986, + 0.08401958644390106, + 0.13292008638381958, + 0.064379021525383, + 0.11464693397283554, + 0.08337364345788956, + 0.04390476644039154 + ], + "credit_map": [ + 0.017029212787747383, + 0.014071555808186531, + 0.012490705586969852, + 0.019035473465919495, + 0.01894395984709263, + 0.020918874070048332, + 0.016644692048430443, + 0.0133719677105546, + 0.009630157612264156, + 0.010137085802853107, + 0.006093357224017382, + 0.004464208614081144, + 0.006990955676883459, + 0.00792564544826746, + 0.008269616402685642, + 0.005097063258290291, + 0.012494047172367573, + 0.015941010788083076, + 0.011499437503516674, + 0.027814140543341637, + 0.019397877156734467, + 0.00779297947883606, + 0.012044363655149937, + 0.008864990435540676, + 0.017249301075935364, + 0.02563019096851349, + 0.02937605045735836, + 0.02392837591469288, + 0.019546296447515488, + 0.016845259815454483, + 0.02340126782655716, + 0.013815812766551971, + 0.013167856261134148, + 0.011967733502388, + 0.015937117859721184, + 0.009011618793010712, + 0.008678874000906944, + 0.008284109644591808, + 0.010236636735498905, + 0.006735654082149267, + 0.018163561820983887, + 0.010273001156747341, + 0.015654364600777626, + 0.014661958441138268, + 0.013120897114276886, + 0.014213860966265202, + 0.017777422443032265, + 0.029055031016469002, + 0.00797165185213089, + 0.0055134911090135574, + 0.008542635478079319, + 0.009595255367457867, + 0.006003426853567362, + 0.01289583370089531, + 0.008809511549770832, + 0.005047218408435583, + 0.017103929072618484, + 0.01158019807189703, + 0.018768668174743652, + 0.01689778082072735, + 0.013013296760618687, + 0.010344664566218853, + 0.01774553209543228, + 0.009192870929837227, + 0.00810928177088499, + 0.0064797899685800076, + 0.00924018956720829, + 0.013706252910196781, + 0.007737521082162857, + 0.010255569592118263, + 0.014231706038117409, + 0.013613349758088589, + 0.003590172855183482, + 0.006355029530823231, + 0.005763940047472715, + 0.009775790385901928, + 0.005583859048783779, + 0.0037721586413681507, + 0.004906528163701296, + 0.004157290793955326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.029802139848470688 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009771609678864479 + ], + "opposite_logprob_tokens": [ + -4.634771823883057 + ], + "logprob_reward": -0.009771609678864479, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009771609678864479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1560126692056656, + 0.04147922620177269, + 0.30620265007019043, + 0.04093741998076439, + 0.08584504574537277, + 0.15737687051296234, + 0.02409346029162407, + 0.02864878624677658, + 0.02015669271349907, + 0.13924714922904968 + ], + "credit_map": [ + 0.023412179201841354, + 0.01697351410984993, + 0.018246416002511978, + 0.024941369891166687, + 0.015320598147809505, + 0.026682933792471886, + 0.01658608950674534, + 0.013849581591784954, + 0.005355470813810825, + 0.0048997411504387856, + 0.005122768227010965, + 0.004404979757964611, + 0.005834648851305246, + 0.006949331145733595, + 0.003822666360065341, + 0.005089618731290102, + 0.040944620966911316, + 0.01711585372686386, + 0.030093664303421974, + 0.036667630076408386, + 0.02427280694246292, + 0.026366354897618294, + 0.04706862196326256, + 0.08367311209440231, + 0.003840259276330471, + 0.004403488244861364, + 0.0051002344116568565, + 0.006546302232891321, + 0.0057727196253836155, + 0.0034669239539653063, + 0.0073677669279277325, + 0.004439727868884802, + 0.012394985184073448, + 0.0081331767141819, + 0.013981210067868233, + 0.008275368250906467, + 0.011750743724405766, + 0.011250106617808342, + 0.011102041229605675, + 0.008957418613135815, + 0.019790824502706528, + 0.010699214413762093, + 0.021911168470978737, + 0.02510456182062626, + 0.011874119751155376, + 0.018317822366952896, + 0.022848384454846382, + 0.02683076821267605, + 0.002853044541552663, + 0.002354404656216502, + 0.003501088358461857, + 0.002805822528898716, + 0.002880160929635167, + 0.003986262250691652, + 0.003053805325180292, + 0.002658870304003358, + 0.0037454157136380672, + 0.0030010330956429243, + 0.003985651768743992, + 0.003406765405088663, + 0.003742920234799385, + 0.002106516622006893, + 0.003940201830118895, + 0.004720283672213554, + 0.0021786154247820377, + 0.003036617999896407, + 0.003006241051480174, + 0.0028719650581479073, + 0.0018317383946850896, + 0.0015798320528119802, + 0.0024870631750673056, + 0.003164620604366064, + 0.015607861801981926, + 0.009524247609078884, + 0.023656975477933884, + 0.014381887391209602, + 0.010076718404889107, + 0.030508851632475853, + 0.013775713741779327, + 0.02171490155160427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.009771609678864479 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly pans left, following a clown struggling forward. He wears a vibrant rainbow-hued afro wig, absurdly topped with an oversized blue bowler hat. His face is thickly smeared with white paint, and a dramatic red smile stretches from ear to ear, contrasting with his brow furrowed in intense concentration and his eyes widened. A painted-on blue teardrop dangles precariously beneath his right eye. He wears an extremely baggy, bright yellow jumpsuit covered in giant red and blue polka dots, and a cascading white ruffled neckpiece wraps around his neck. He is riding a shiny chrome unicycle, his legs laboriously pedaling in an uncoordinated yet determined rhythm, his upper body swaying violently from side to side in an effort to maintain balance, his arms fully extended, waving futilely in the air like wings. The bright afternoon sun shines on the wide asphalt road in the park, casting a vivid luster on his colorful clothes and dragging a clear shadow behind him that distorts and stretches as he sways. The whole scene unfolds in a joyful atmosphere full of comic tension.", + "seed": 22378, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_13/index_9668_seed_22378_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07752668717876077, + "vlm_logprob": -0.07752668717876077, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.07752668717876077, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05847778543829918 + ], + "opposite_logprob_tokens": [ + -3.3084778785705566 + ], + "logprob_reward": -0.05847778543829918, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.05847778543829918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12597551941871643, + 0.04035743698477745, + 0.04987719655036926, + 0.03874676674604416, + 0.46332767605781555, + 0.0789361447095871, + 0.03686562180519104, + 0.027276791632175446, + 0.026318999007344246, + 0.11231787502765656 + ], + "credit_map": [ + 0.014387107454240322, + 0.01649601384997368, + 0.029080335050821304, + 0.013182983733713627, + 0.011120396666228771, + 0.012142827734351158, + 0.011909136548638344, + 0.017656715586781502, + 0.004460221156477928, + 0.003835018491372466, + 0.004724529106169939, + 0.004651316907256842, + 0.008180991746485233, + 0.004431615583598614, + 0.0048422240652143955, + 0.005231516435742378, + 0.008642279542982578, + 0.00654720701277256, + 0.005946372635662556, + 0.00730569614097476, + 0.004297440871596336, + 0.005644151009619236, + 0.004702394362539053, + 0.006791654042899609, + 0.003235188778489828, + 0.006234516389667988, + 0.005474921781569719, + 0.006644520442932844, + 0.004685136955231428, + 0.0031392090022563934, + 0.0047179413959383965, + 0.004615333396941423, + 0.012933059595525265, + 0.03330366313457489, + 0.05385269969701767, + 0.02817753329873085, + 0.1325789988040924, + 0.02102290838956833, + 0.02090504579246044, + 0.16055378317832947, + 0.00795205868780613, + 0.011559254489839077, + 0.018704306334257126, + 0.006863331887871027, + 0.003935972228646278, + 0.00790439173579216, + 0.004908477887511253, + 0.01710834726691246, + 0.0025855570565909147, + 0.011056918650865555, + 0.0044570243917405605, + 0.004997268319129944, + 0.0034427742939442396, + 0.002037156606093049, + 0.0047844392247498035, + 0.003504482563585043, + 0.0029177716933190823, + 0.0055321077816188335, + 0.0028644565027207136, + 0.003946332726627588, + 0.002665191888809204, + 0.002170243998989463, + 0.003979207016527653, + 0.003201480722054839, + 0.0025255291257053614, + 0.002141860779374838, + 0.0025097711477428675, + 0.0028241053223609924, + 0.005354293156415224, + 0.004208587110042572, + 0.002516227075830102, + 0.004238627385348082, + 0.009730160236358643, + 0.006930728442966938, + 0.032404396682977676, + 0.007404777221381664, + 0.006898884195834398, + 0.004936964716762304, + 0.03347330540418625, + 0.010538646019995213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.05847778543829918 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25210461020469666 + ], + "opposite_logprob_tokens": [ + -1.502104640007019 + ], + "logprob_reward": -0.25210461020469666, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25210461020469666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14222218096256256, + 0.05371176823973656, + 0.05278102308511734, + 0.08873046934604645, + 0.2607654631137848, + 0.05031626299023628, + 0.07923109829425812, + 0.05928550660610199, + 0.05128994211554527, + 0.1616663634777069 + ], + "credit_map": [ + 0.015714311972260475, + 0.011813102290034294, + 0.02958109602332115, + 0.018225910142064095, + 0.011058726347982883, + 0.009514663368463516, + 0.020281797274947166, + 0.02603255771100521, + 0.00798695720732212, + 0.005280802492052317, + 0.0067524476908147335, + 0.006119827274233103, + 0.00418686680495739, + 0.006780508905649185, + 0.004344238433986902, + 0.012260115705430508, + 0.00835366640239954, + 0.005682176444679499, + 0.006443164777010679, + 0.006855968851596117, + 0.0048601278103888035, + 0.0060720425099134445, + 0.006082196719944477, + 0.008431674912571907, + 0.005427627824246883, + 0.008350387215614319, + 0.016867268830537796, + 0.011824971996247768, + 0.007948478683829308, + 0.006983812432736158, + 0.018611177802085876, + 0.012716732919216156, + 0.01792135089635849, + 0.03056146763265133, + 0.02805228903889656, + 0.0232799481600523, + 0.05522250011563301, + 0.014459368772804737, + 0.01804167591035366, + 0.07322683930397034, + 0.004471752792596817, + 0.003586612408980727, + 0.012223337776958942, + 0.0036643408238887787, + 0.00410454161465168, + 0.009457026608288288, + 0.0028851768001914024, + 0.00992346741259098, + 0.005270167253911495, + 0.02253151871263981, + 0.008993334136903286, + 0.009561189450323582, + 0.008795754984021187, + 0.007614464033395052, + 0.008892039768397808, + 0.00757262296974659, + 0.006736903917044401, + 0.010776698589324951, + 0.0060849725268781185, + 0.008640154264867306, + 0.004798602312803268, + 0.005042488686740398, + 0.00963326171040535, + 0.0075724171474576, + 0.006155970506370068, + 0.011668073944747448, + 0.004679650068283081, + 0.005967514123767614, + 0.005006786435842514, + 0.004561829846352339, + 0.007265993393957615, + 0.005984118673950434, + 0.012686452828347683, + 0.009239571169018745, + 0.0414469912648201, + 0.010913088917732239, + 0.010192167945206165, + 0.009567600674927235, + 0.0556536428630352, + 0.011966831050813198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.25210461020469666 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038097988814115524 + ], + "opposite_logprob_tokens": [ + -3.288098096847534 + ], + "logprob_reward": -0.038097988814115524, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038097988814115524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31189319491386414, + 0.048621781170368195, + 0.07579781860113144, + 0.054929573088884354, + 0.12919674813747406, + 0.20417076349258423, + 0.07751090824604034, + 0.0547674186527729, + 0.0061999037861824036, + 0.03691195696592331 + ], + "credit_map": [ + 0.03838847950100899, + 0.04231403395533562, + 0.05765106901526451, + 0.0582086481153965, + 0.031245145946741104, + 0.015184611082077026, + 0.037742625921964645, + 0.031158553436398506, + 0.0044846804812550545, + 0.0068951756693422794, + 0.005137916188687086, + 0.00821413192898035, + 0.006017055828124285, + 0.0035838496405631304, + 0.006342495791614056, + 0.007946472615003586, + 0.01136017870157957, + 0.01333625242114067, + 0.00880441814661026, + 0.011347143910825253, + 0.0075541227124631405, + 0.00881323404610157, + 0.006925700698047876, + 0.007656761910766363, + 0.006135240662842989, + 0.00625592889264226, + 0.01154717244207859, + 0.006626271177083254, + 0.008332290686666965, + 0.0050531188026070595, + 0.004700760822743177, + 0.006278787739574909, + 0.005003764759749174, + 0.016072828322649002, + 0.015468571335077286, + 0.01245142798870802, + 0.02572220377624035, + 0.0068802908062934875, + 0.009386167861521244, + 0.038211483508348465, + 0.017210494726896286, + 0.04694724828004837, + 0.04318223148584366, + 0.02053399570286274, + 0.011629522778093815, + 0.01855211891233921, + 0.015646517276763916, + 0.030468620359897614, + 0.0068297735415399075, + 0.0193917416036129, + 0.010300193913280964, + 0.007748544681817293, + 0.01068178378045559, + 0.008077922277152538, + 0.0065189930610358715, + 0.007961953058838844, + 0.007424768526107073, + 0.011734784580767155, + 0.005999456625431776, + 0.005749728996306658, + 0.004233452491462231, + 0.006362017244100571, + 0.005973244551569223, + 0.007289965637028217, + 0.0007872361456975341, + 0.000853173085488379, + 0.0007040663622319698, + 0.0007823057821951807, + 0.0006179104675538838, + 0.0008780758362263441, + 0.0007267797482199967, + 0.0008503558929078281, + 0.002780253067612648, + 0.008121847175061703, + 0.005634896457195282, + 0.004071441944688559, + 0.0033114615362137556, + 0.002523292787373066, + 0.006392640992999077, + 0.0040761204436421394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.038097988814115524 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370732069015503 + ], + "opposite_logprob_tokens": [ + -3.4087073802948 + ], + "logprob_reward": -0.03370732069015503, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370732069015503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2333431839942932, + 0.03434626758098602, + 0.05692874640226364, + 0.1415150761604309, + 0.35846754908561707, + 0.06633424013853073, + 0.017070192843675613, + 0.038657162338495255, + 0.02519208751618862, + 0.028145555406808853 + ], + "credit_map": [ + 0.03077395260334015, + 0.03180679678916931, + 0.03276151046156883, + 0.048993874341249466, + 0.02530701644718647, + 0.014051636680960655, + 0.031859688460826874, + 0.017788715660572052, + 0.003952622879296541, + 0.004151968751102686, + 0.0042844838462769985, + 0.004695450887084007, + 0.0037864057812839746, + 0.002916526747867465, + 0.004185604862868786, + 0.0063732038252055645, + 0.009651347994804382, + 0.007677474059164524, + 0.006916884332895279, + 0.007932232692837715, + 0.004861702676862478, + 0.007129765581339598, + 0.005845845211297274, + 0.0069134896621108055, + 0.01016176026314497, + 0.013506007380783558, + 0.02706940658390522, + 0.013907039538025856, + 0.019270433112978935, + 0.0172325000166893, + 0.020643625408411026, + 0.019724318757653236, + 0.011380541138350964, + 0.022475535050034523, + 0.039740677922964096, + 0.019399471580982208, + 0.11035185307264328, + 0.017883455380797386, + 0.012499979697167873, + 0.1247360110282898, + 0.003038739087060094, + 0.004334809724241495, + 0.022260267287492752, + 0.003955128602683544, + 0.002341527957469225, + 0.003035876201465726, + 0.00331046711653471, + 0.024057427421212196, + 0.0013075122842565179, + 0.003981555346399546, + 0.001737247221171856, + 0.002265491057187319, + 0.0023698583245277405, + 0.0017036673380061984, + 0.0018816132796928287, + 0.001823247643187642, + 0.004796909634023905, + 0.007807232439517975, + 0.00415525957942009, + 0.004974634852260351, + 0.0027225178200751543, + 0.0032877952326089144, + 0.005670958664268255, + 0.005241853650659323, + 0.002135687042027712, + 0.0033246742095798254, + 0.0017690228996798396, + 0.0030353576876223087, + 0.00447931420058012, + 0.003100000089034438, + 0.002691148780286312, + 0.0046568820253014565, + 0.0022213284391909838, + 0.00218398729339242, + 0.005405360367149115, + 0.002626722678542137, + 0.0021101676393300295, + 0.0017190489452332258, + 0.008642291650176048, + 0.0032366476953029633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.03370732069015503 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005245730746537447 + ], + "opposite_logprob_tokens": [ + -5.255245685577393 + ], + "logprob_reward": -0.005245730746537447, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005245730746537447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2410963624715805, + 0.04408928006887436, + 0.034435804933309555, + 0.01768380217254162, + 0.2988552451133728, + 0.09727387130260468, + 0.04259911924600601, + 0.05115180090069771, + 0.1235806792974472, + 0.04923401027917862 + ], + "credit_map": [ + 0.03641075640916824, + 0.02497735433280468, + 0.04293020814657211, + 0.025655481964349747, + 0.016723554581403732, + 0.02214314416050911, + 0.03249330073595047, + 0.03976254165172577, + 0.0074635762721300125, + 0.0038465557154268026, + 0.004925863351672888, + 0.0036944332532584667, + 0.00686973100528121, + 0.007142767310142517, + 0.003925691824406385, + 0.006220658775418997, + 0.006402784027159214, + 0.002967461012303829, + 0.0042435782961547375, + 0.0038958105724304914, + 0.004240513313561678, + 0.005551333539187908, + 0.0031828847713768482, + 0.0039514400996267796, + 0.0018149473471567035, + 0.002995718503370881, + 0.002199087990447879, + 0.002574013313278556, + 0.001759587088599801, + 0.00163710315246135, + 0.0018602306954562664, + 0.002843113848939538, + 0.009305420331656933, + 0.0184883214533329, + 0.03578921779990196, + 0.01702483557164669, + 0.08664567768573761, + 0.0143441678956151, + 0.012624576687812805, + 0.10463304072618484, + 0.007938648574054241, + 0.008829991333186626, + 0.026156332343816757, + 0.006087665446102619, + 0.006327308248728514, + 0.018249770626425743, + 0.004740408156067133, + 0.01894374005496502, + 0.003109398065134883, + 0.006160504184663296, + 0.00512482225894928, + 0.0038158069364726543, + 0.004816090222448111, + 0.007462689187377691, + 0.005147998686879873, + 0.006961809005588293, + 0.006814060267060995, + 0.00792617816478014, + 0.006054803729057312, + 0.005219379439949989, + 0.004762798082083464, + 0.007945298217236996, + 0.005038471892476082, + 0.0073908125050365925, + 0.007947569712996483, + 0.011230996809899807, + 0.00840111542493105, + 0.009813807904720306, + 0.031517889350652695, + 0.02217864617705345, + 0.012144078500568867, + 0.020346567034721375, + 0.0034582209773361683, + 0.005087950266897678, + 0.008615641854703426, + 0.0034025663044303656, + 0.005896553862839937, + 0.005596671253442764, + 0.008601297624409199, + 0.008575106039643288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.005245730746537447 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly pans left, following a clown struggling forward. He wears a vibrant rainbow-hued afro wig, absurdly topped with an oversized blue bowler hat. His face is thickly smeared with white paint, and a dramatic red smile stretches from ear to ear, contrasting with his brow furrowed in intense concentration and his eyes widened. A painted-on blue teardrop dangles precariously beneath his right eye. He wears an extremely baggy, bright yellow jumpsuit covered in giant red and blue polka dots, and a cascading white ruffled neckpiece wraps around his neck. He is riding a shiny chrome unicycle, his legs laboriously pedaling in an uncoordinated yet determined rhythm, his upper body swaying violently from side to side in an effort to maintain balance, his arms fully extended, waving futilely in the air like wings. The bright afternoon sun shines on the wide asphalt road in the park, casting a vivid luster on his colorful clothes and dragging a clear shadow behind him that distorts and stretches as he sways. The whole scene unfolds in a joyful atmosphere full of comic tension.", + "seed": 22379, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_13/index_9668_seed_22379_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07453113384544849, + "vlm_logprob": -0.07453113384544849, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.07453113384544849, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05709226429462433 + ], + "opposite_logprob_tokens": [ + -3.1820921897888184 + ], + "logprob_reward": -0.05709226429462433, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05709226429462433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037286922335624695, + 0.018600070849061012, + 0.1994515061378479, + 0.17667444050312042, + 0.24383850395679474, + 0.04789651557803154, + 0.13150081038475037, + 0.07637027651071548, + 0.041275423020124435, + 0.027105534449219704 + ], + "credit_map": [ + 0.003552280366420746, + 0.0029829293489456177, + 0.006095357239246368, + 0.004796840250492096, + 0.0032119012903422117, + 0.005622622556984425, + 0.004161420278251171, + 0.0068635717034339905, + 0.0021823833230882883, + 0.002443604404106736, + 0.001804065308533609, + 0.0020782819483429193, + 0.004030299372971058, + 0.0017649517394602299, + 0.0020764959044754505, + 0.002219989662989974, + 0.017668431624770164, + 0.021719826385378838, + 0.025416551157832146, + 0.031173216179013252, + 0.016584379598498344, + 0.029781002551317215, + 0.02558388188481331, + 0.031524211168289185, + 0.01397932879626751, + 0.02120702713727951, + 0.028475791215896606, + 0.022226208820939064, + 0.022108370438218117, + 0.015571081079542637, + 0.015505699440836906, + 0.037600934505462646, + 0.020769497379660606, + 0.006042879540473223, + 0.08450370281934738, + 0.011685442179441452, + 0.009894589893519878, + 0.018969310447573662, + 0.005584649741649628, + 0.08638843894004822, + 0.007796952500939369, + 0.0030579948797822, + 0.011405791155993938, + 0.0030289478600025177, + 0.002508532954379916, + 0.005696941167116165, + 0.0019309674389660358, + 0.012470386922359467, + 0.01253486331552267, + 0.004238825757056475, + 0.04833943769335747, + 0.007284246850758791, + 0.01052721869200468, + 0.008189103566110134, + 0.0031786803156137466, + 0.0372084304690361, + 0.005536149255931377, + 0.01060947310179472, + 0.013761607930064201, + 0.012000558897852898, + 0.005508252419531345, + 0.006819309201091528, + 0.007430143654346466, + 0.014704780653119087, + 0.00411821436136961, + 0.007056100759655237, + 0.00382999237626791, + 0.00548965111374855, + 0.005988773424178362, + 0.005936330184340477, + 0.004250115714967251, + 0.004606248810887337, + 0.0024208934046328068, + 0.002705082530155778, + 0.004984066355973482, + 0.002450012369081378, + 0.0034746634773910046, + 0.00465133273974061, + 0.0033692731522023678, + 0.0030502097215503454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.05709226429462433 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255776971578598 + ], + "opposite_logprob_tokens": [ + -1.600577712059021 + ], + "logprob_reward": -0.2255776971578598, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255776971578598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03921731561422348, + 0.018226085230708122, + 0.11779112368822098, + 0.0777626484632492, + 0.29002219438552856, + 0.05902506783604622, + 0.30323487520217896, + 0.03312607482075691, + 0.0370461642742157, + 0.024548493325710297 + ], + "credit_map": [ + 0.002953017596155405, + 0.0037699511740356684, + 0.007380958646535873, + 0.0045115347020328045, + 0.0032633699011057615, + 0.0037881562020629644, + 0.004983458202332258, + 0.00856686383485794, + 0.0021912280935794115, + 0.001382659305818379, + 0.002017610240727663, + 0.0018378040986135602, + 0.003965605981647968, + 0.0021636239252984524, + 0.0023197431582957506, + 0.002347811358049512, + 0.010892042890191078, + 0.011060512624680996, + 0.013062208890914917, + 0.017970556393265724, + 0.012535643763840199, + 0.016134703531861305, + 0.01652340777218342, + 0.019612044095993042, + 0.006083286367356777, + 0.008702649734914303, + 0.01423320360481739, + 0.009077112190425396, + 0.00815214030444622, + 0.006920396350324154, + 0.0062782359309494495, + 0.018315624445676804, + 0.024669423699378967, + 0.008803791366517544, + 0.09922018647193909, + 0.013655586168169975, + 0.013830550946295261, + 0.022123578935861588, + 0.006764122284948826, + 0.10095493495464325, + 0.01235367264598608, + 0.004680901765823364, + 0.011851741001009941, + 0.004180855583399534, + 0.0032083294354379177, + 0.006866023875772953, + 0.00378282368183136, + 0.012100715190172195, + 0.028693033382296562, + 0.006560220383107662, + 0.11546934396028519, + 0.016037514433264732, + 0.024274131283164024, + 0.018429487943649292, + 0.006836332380771637, + 0.08693481236696243, + 0.002918550744652748, + 0.003319127019494772, + 0.008445853367447853, + 0.0041600242257118225, + 0.0017681714380159974, + 0.003081694943830371, + 0.002050666604191065, + 0.0073819877579808235, + 0.0035876273177564144, + 0.005798227619379759, + 0.0049316552467644215, + 0.006046904716640711, + 0.004099100828170776, + 0.005411776248365641, + 0.003668620251119137, + 0.0035022457595914602, + 0.0025667815934866667, + 0.0028552026487886906, + 0.005563597660511732, + 0.003581770695745945, + 0.0027414467185735703, + 0.0024165769573301077, + 0.0023249273654073477, + 0.0024981873575598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 2 + ], + "selected_reward": -0.2255776971578598 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033698905259370804 + ], + "opposite_logprob_tokens": [ + -3.408698797225952 + ], + "logprob_reward": -0.033698905259370804, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033698905259370804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.010594816878437996, + 0.0053815399296581745, + 0.013388198800384998, + 0.008594206534326077, + 0.8464983701705933, + 0.017517851665616035, + 0.057118795812129974, + 0.00283479574136436, + 0.0045344955287873745, + 0.03353685885667801 + ], + "credit_map": [ + 0.0012181157944723964, + 0.0011374920140951872, + 0.0008827720885165036, + 0.0014365938259288669, + 0.0011185265611857176, + 0.0010804964695125818, + 0.0021698276977986097, + 0.001550992950797081, + 0.0004588172887451947, + 0.0006997904856689274, + 0.00045896932715550065, + 0.0005076570087112486, + 0.0012585677905008197, + 0.0005269719986245036, + 0.0007300105644389987, + 0.0007407561643049121, + 0.0012647478142753243, + 0.0015945435734465718, + 0.0016732050571590662, + 0.002203383482992649, + 0.0012311306782066822, + 0.0016047288663685322, + 0.0015821559354662895, + 0.0022343050222843885, + 0.0008760420023463666, + 0.0017171994550153613, + 0.0008563515148125589, + 0.0010549710132181644, + 0.0012965003261342645, + 0.0009128421079367399, + 0.000931690214201808, + 0.0009486108901910484, + 0.07219336926937103, + 0.020364850759506226, + 0.2969057261943817, + 0.03824286535382271, + 0.032329633831977844, + 0.0648927390575409, + 0.017307370901107788, + 0.30426186323165894, + 0.002422810299322009, + 0.0020372592844069004, + 0.0017494995845481753, + 0.003587916260585189, + 0.0011481018736958504, + 0.0018701473018154502, + 0.002328164642676711, + 0.002373953815549612, + 0.005319765768945217, + 0.0022278493270277977, + 0.02072818949818611, + 0.0034936589654535055, + 0.004380473867058754, + 0.0033565510530024767, + 0.001776939956471324, + 0.01583537831902504, + 0.0002536968095228076, + 0.0005684399511665106, + 0.0003530808899085969, + 0.00044702706509269774, + 0.0002558566629886627, + 0.00026281329337507486, + 0.00034746446181088686, + 0.00034641704405657947, + 0.0003125269431620836, + 0.0007316403789445758, + 0.00041080289520323277, + 0.00047954919864423573, + 0.0005866518476977944, + 0.0009673036984167993, + 0.0005081589915789664, + 0.0005378620699048042, + 0.0025959056802093983, + 0.003884644480422139, + 0.008010953664779663, + 0.004505680873990059, + 0.002958182478323579, + 0.005028329323977232, + 0.003731248900294304, + 0.002821916714310646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.033698905259370804 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866613447666168 + ], + "opposite_logprob_tokens": [ + -3.04866623878479 + ], + "logprob_reward": -0.04866613447666168, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866613447666168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09694136679172516, + 0.03682451695203781, + 0.08826881647109985, + 0.1317901760339737, + 0.13320916891098022, + 0.2188817262649536, + 0.13680672645568848, + 0.05013787001371384, + 0.0459904707968235, + 0.06114916503429413 + ], + "credit_map": [ + 0.0062800077721476555, + 0.007639511488378048, + 0.0172308050096035, + 0.011838210746645927, + 0.013465343043208122, + 0.007909982465207577, + 0.012680194340646267, + 0.01989731565117836, + 0.003741048276424408, + 0.003929931670427322, + 0.003356978064402938, + 0.004244223702698946, + 0.005955070722848177, + 0.0031890086829662323, + 0.007513550575822592, + 0.0048947082832455635, + 0.007589011453092098, + 0.006975171156227589, + 0.010279307141900063, + 0.015398185700178146, + 0.0094091622158885, + 0.009067225269973278, + 0.01417781412601471, + 0.01537293940782547, + 0.009465808048844337, + 0.013849923387169838, + 0.025170816108584404, + 0.01797277294099331, + 0.010654154233634472, + 0.010776863433420658, + 0.012903287075459957, + 0.030996564775705338, + 0.010515548288822174, + 0.006227067206054926, + 0.043672557920217514, + 0.009113152511417866, + 0.005930433049798012, + 0.010275749489665031, + 0.004514456260949373, + 0.04296020790934563, + 0.03109821490943432, + 0.01710774004459381, + 0.038246460258960724, + 0.030316457152366638, + 0.010809180326759815, + 0.02613104321062565, + 0.02080683782696724, + 0.04436579346656799, + 0.012529417872428894, + 0.006805460434406996, + 0.0486617237329483, + 0.007482263725250959, + 0.01037804689258337, + 0.008311006240546703, + 0.004824914038181305, + 0.03781390190124512, + 0.003420853754505515, + 0.006746676750481129, + 0.009625177830457687, + 0.007430601865053177, + 0.003716270439326763, + 0.0037125260569155216, + 0.004897779785096645, + 0.010587986558675766, + 0.004291847348213196, + 0.006595647428184748, + 0.007312503177672625, + 0.008037406019866467, + 0.006432986352592707, + 0.00576892402023077, + 0.004053010139614344, + 0.003498149337247014, + 0.005346838850528002, + 0.007913184352219105, + 0.008478975854814053, + 0.0053855604492127895, + 0.01026900950819254, + 0.008932318538427353, + 0.007911236956715584, + 0.006912045180797577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.04866613447666168 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot taken from a level angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera slowly panning left while following a clown struggling forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the clown wear a vibrant rainbow-hued afro wig topped with an oversized blue bowler hat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the clown riding a shiny chrome unicycle with legs pedaling and an upper body swaying violently?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene set on a wide asphalt road in a park under bright afternoon sun with a clear shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007620668038725853 + ], + "opposite_logprob_tokens": [ + -4.882620811462402 + ], + "logprob_reward": -0.007620668038725853, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007620668038725853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09480752050876617, + 0.04788820445537567, + 0.09280428290367126, + 0.186801478266716, + 0.11256121098995209, + 0.06799536198377609, + 0.3092929720878601, + 0.046591103076934814, + 0.02153753861784935, + 0.019720323383808136 + ], + "credit_map": [ + 0.012077169492840767, + 0.013150299899280071, + 0.013345308601856232, + 0.01242644153535366, + 0.00844061654061079, + 0.014369145967066288, + 0.009469401091337204, + 0.011529129929840565, + 0.006562733557075262, + 0.007842883467674255, + 0.007316019851714373, + 0.004947987850755453, + 0.006195178255438805, + 0.005914545152336359, + 0.0035807215608656406, + 0.0055281356908380985, + 0.008109170012176037, + 0.010098853148519993, + 0.0123679144307971, + 0.014764689840376377, + 0.00845824833959341, + 0.012351864948868752, + 0.013232685625553131, + 0.013420856557786465, + 0.016867229714989662, + 0.030131012201309204, + 0.02085614949464798, + 0.024826284497976303, + 0.031076479703187943, + 0.017936332151293755, + 0.018019123002886772, + 0.027088869363069534, + 0.009884717874228954, + 0.0037681341636925936, + 0.03588899224996567, + 0.005892264191061258, + 0.006212020758539438, + 0.009871719405055046, + 0.0032760947942733765, + 0.03776727616786957, + 0.010699847713112831, + 0.003899822011590004, + 0.016125228255987167, + 0.0032180012203752995, + 0.0035675140097737312, + 0.009275638498365879, + 0.0024911235086619854, + 0.018718184903264046, + 0.02792682684957981, + 0.011296198703348637, + 0.11347828805446625, + 0.01579897478222847, + 0.023297924548387527, + 0.018745511770248413, + 0.007751515135169029, + 0.09099774062633514, + 0.003985757939517498, + 0.002186124911531806, + 0.012192294001579285, + 0.003211545990779996, + 0.005367279518395662, + 0.0046053132973611355, + 0.004458851646631956, + 0.010583936236798763, + 0.003611285937950015, + 0.0021516364067792892, + 0.002150637563318014, + 0.0017395098693668842, + 0.00408406276255846, + 0.003244238905608654, + 0.002467283047735691, + 0.0020888829603791237, + 0.0018871432403102517, + 0.0022901438642293215, + 0.0036718701012432575, + 0.0023997880052775145, + 0.002221860224381089, + 0.0028927535749971867, + 0.0025457884185016155, + 0.0018109739758074284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 4 + ], + "selected_reward": -0.007620668038725853 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_14/index_tensor([8892, 8892])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_14/index_tensor([8892, 8892])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3ec2c3247e69bc8b548a68a524e3e86dc5883b6c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_14/index_tensor([8892, 8892])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, extreme wide-angle shot, trembling almost imperceptibly, simulates violent tectonic shifts, capturing the apocalyptic upheaval of a desolate alien world. Beneath a scarred sky swirling with purple and ochre clouds, the barren, amber-colored ground fractures violently, spiderweb-like cracks running across the wasteland before vast, glowing fissures tear open with catastrophic force. Dominating the horizon is a massive biomechanical foundation, an unholy fusion of petrified ivory-white bone and non-reflective obsidian-black metal. Its architecture defies human logic, with massive ribbed arches and towering spires scraping against the toxic clouds, their surfaces a disturbing blend of porous organic textures and cold alien metallurgy. As the ground tilts, the entire massive structure begins to groan and shift; the movements are agonizingly slow but immensely powerful, as the fused joints of bone and metal rub against each other, the entire base teetering precariously. A deafening, low-pitched protest echoes across the plain as the building tilts precariously, its immense weight causing the edges of the newly formed ravine to crumble. The scene is cast in the dim, eerie glow of a distant, dying sun, creating long, twisted shadows, while the fiery orange light from the cracks casts hellish highlights over the horrific forms of the base, creating a tense and awe-inspiring atmosphere of cosmic terror and imminent collapse.", + "seed": 30948, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_14/index_8892_seed_30948_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.015327214729040861, + "vlm_logprob": -0.015327214729040861, + "vlm_yes_no_margin": 5.175, + "vlm_token_credit": -0.015327214729040861, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02284400910139084 + ], + "opposite_logprob_tokens": [ + -4.272843837738037 + ], + "logprob_reward": -0.02284400910139084, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02284400910139084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49257779121398926, + 0.09297516196966171, + 0.02824820764362812, + 0.037281375378370285, + 0.009751670993864536, + 0.08823724091053009, + 0.025819990783929825, + 0.040124986320734024, + 0.02774866670370102, + 0.15723493695259094 + ], + "credit_map": [ + 0.02244582585990429, + 0.19347232580184937, + 0.014363281428813934, + 0.013841013424098492, + 0.02004219964146614, + 0.18752920627593994, + 0.012057800777256489, + 0.028826085850596428, + 0.01992308534681797, + 0.013307152315974236, + 0.01050508487969637, + 0.008706072345376015, + 0.006232335232198238, + 0.01606742851436138, + 0.006689606234431267, + 0.011544392444193363, + 0.005596050526946783, + 0.0033882656134665012, + 0.0024734155740588903, + 0.0027367365546524525, + 0.0024143122136592865, + 0.006953506264835596, + 0.0021332476753741503, + 0.002552674151957035, + 0.006886257324367762, + 0.0034497936721891165, + 0.00264312163926661, + 0.008044297806918621, + 0.00412936182692647, + 0.004854527302086353, + 0.003664766438305378, + 0.0036092496011406183, + 0.0015185592928901315, + 0.0009717615903355181, + 0.0009752034093253314, + 0.0011961829150095582, + 0.0019786099437624216, + 0.0013468011748045683, + 0.0008961937273852527, + 0.0008683587075211108, + 0.021433468908071518, + 0.006772355642169714, + 0.0033292188309133053, + 0.0033915897365659475, + 0.0046960897743701935, + 0.005166878458112478, + 0.01453909557312727, + 0.028908545151352882, + 0.0016504940576851368, + 0.00369175779633224, + 0.0025860972236841917, + 0.0028565272223204374, + 0.00510680116713047, + 0.002185163786634803, + 0.004511124454438686, + 0.003232021816074848, + 0.002879304811358452, + 0.003996389452368021, + 0.005127393640577793, + 0.0031287246383726597, + 0.006951126269996166, + 0.003019105177372694, + 0.009854537434875965, + 0.005168405827134848, + 0.002056871308013797, + 0.0032919899094849825, + 0.004493278916925192, + 0.0031243080738931894, + 0.0049430569633841515, + 0.003021887969225645, + 0.0027509534265846014, + 0.004066320601850748, + 0.008396215736865997, + 0.0058882106095552444, + 0.005728497169911861, + 0.004308898467570543, + 0.05771723762154579, + 0.01318090595304966, + 0.007088410202413797, + 0.054926563054323196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.02284400910139084 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010513499146327376 + ], + "opposite_logprob_tokens": [ + -6.876051425933838 + ], + "logprob_reward": -0.0010513499146327376, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010513499146327376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14508679509162903, + 0.06384558975696564, + 0.1186809167265892, + 0.0383431650698185, + 0.029393423348665237, + 0.12732407450675964, + 0.02855665795505047, + 0.023067843168973923, + 0.022155804559588432, + 0.40354570746421814 + ], + "credit_map": [ + 0.008349143899977207, + 0.04876674711704254, + 0.006439205724745989, + 0.006609257776290178, + 0.013156953267753124, + 0.04378297179937363, + 0.007517042104154825, + 0.010465491563081741, + 0.010337762534618378, + 0.010340877808630466, + 0.006450310815125704, + 0.004712716210633516, + 0.007113678380846977, + 0.010815352201461792, + 0.006649412214756012, + 0.0074254791252315044, + 0.0237102210521698, + 0.012973930686712265, + 0.008342161774635315, + 0.008975261822342873, + 0.00819464586675167, + 0.03863811492919922, + 0.008828411810100079, + 0.009018161334097385, + 0.005098676774650812, + 0.005236841272562742, + 0.004738328978419304, + 0.005840051919221878, + 0.0046156421303749084, + 0.005742373876273632, + 0.003692799713462591, + 0.0033784518018364906, + 0.004139621742069721, + 0.004105354659259319, + 0.0035212752409279346, + 0.0029825004749000072, + 0.005667147692292929, + 0.0031343726441264153, + 0.0034112930297851562, + 0.0024318580981343985, + 0.030079269781708717, + 0.010166826657950878, + 0.005101796705275774, + 0.0048101176507771015, + 0.0069459970109164715, + 0.007090663071721792, + 0.021165847778320312, + 0.041963547468185425, + 0.0017091145273298025, + 0.0030482569709420204, + 0.0021743953693658113, + 0.002817412605509162, + 0.006281696259975433, + 0.003239354584366083, + 0.0048494599759578705, + 0.0044369688257575035, + 0.002056561643257737, + 0.0024624892976135015, + 0.0029386223759502172, + 0.0020205972250550985, + 0.00449926033616066, + 0.003012758679687977, + 0.003161362139508128, + 0.00291619123890996, + 0.0015098577132448554, + 0.003563342848792672, + 0.0033733025193214417, + 0.002064441330730915, + 0.002807695884257555, + 0.0029238180723041296, + 0.002945551648736, + 0.0029677944257855415, + 0.018196862190961838, + 0.011735026724636555, + 0.00952854286879301, + 0.008293422870337963, + 0.15678925812244415, + 0.029329665005207062, + 0.01420137844979763, + 0.1554715633392334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.0010513499146327376 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864158570766449 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.00864158570766449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864158570766449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.718036949634552, + 0.033142562955617905, + 0.024533569812774658, + 0.027116134762763977, + 0.01434926688671112, + 0.0204938817769289, + 0.015693189576268196, + 0.025690456852316856, + 0.015065738931298256, + 0.10587818175554276 + ], + "credit_map": [ + 0.03195936232805252, + 0.2833894193172455, + 0.020128902047872543, + 0.015647156164050102, + 0.0313182957470417, + 0.2789744436740875, + 0.017218979075551033, + 0.039400454610586166, + 0.0035961640533059835, + 0.004760177806019783, + 0.002654881915077567, + 0.0026916980277746916, + 0.005354336928576231, + 0.006057855673134327, + 0.0044358945451676846, + 0.0035915574990212917, + 0.0034051164984703064, + 0.002929725917056203, + 0.00279175559990108, + 0.0031535951420664787, + 0.002969178603962064, + 0.003294356632977724, + 0.0029391159769147635, + 0.003050727304071188, + 0.004405719228088856, + 0.0026892225723713636, + 0.00310750724747777, + 0.0038336217403411865, + 0.0036786801647394896, + 0.003976351581513882, + 0.002794279484078288, + 0.0026307564694434404, + 0.0019331108778715134, + 0.0019816572312265635, + 0.0016314018284901977, + 0.002049168571829796, + 0.002164094243198633, + 0.0018308034632354975, + 0.0012533605331555009, + 0.0015056730480864644, + 0.002590738469734788, + 0.0037398235872387886, + 0.0016151039162650704, + 0.0020907623693346977, + 0.0037254774942994118, + 0.0017565751913934946, + 0.002239614725112915, + 0.0027357880026102066, + 0.0007520396611653268, + 0.0015485987532883883, + 0.001191231538541615, + 0.001590663567185402, + 0.0042046369053423405, + 0.001352773979306221, + 0.002807011129334569, + 0.002246235962957144, + 0.002265619346871972, + 0.0025487549137324095, + 0.003612722735852003, + 0.002721989993005991, + 0.004170670639723539, + 0.002982550533488393, + 0.004171489737927914, + 0.0032166624441742897, + 0.0015673395246267319, + 0.0014407350681722164, + 0.001820001401938498, + 0.0014354821760207415, + 0.0023293073754757643, + 0.0024871286004781723, + 0.0016546990955248475, + 0.0023310489486902952, + 0.004840467125177383, + 0.004090661182999611, + 0.00270664831623435, + 0.002389865694567561, + 0.04003868252038956, + 0.007988540455698967, + 0.0041998750530183315, + 0.03962346166372299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.00864158570766449 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010436094598844647 + ], + "opposite_logprob_tokens": [ + -6.876043796539307 + ], + "logprob_reward": -0.0010436094598844647, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010436094598844647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49510303139686584, + 0.04148833453655243, + 0.04288617894053459, + 0.016205208376049995, + 0.01506017241626978, + 0.18617242574691772, + 0.014659916050732136, + 0.017787063494324684, + 0.01895478554069996, + 0.15168289840221405 + ], + "credit_map": [ + 0.02184801921248436, + 0.19542036950588226, + 0.01416951883584261, + 0.011119507253170013, + 0.02141845040023327, + 0.19154663383960724, + 0.011777417734265327, + 0.027803102508187294, + 0.003536058822646737, + 0.008124302141368389, + 0.0037334521766752005, + 0.003914583008736372, + 0.003822549246251583, + 0.0104029206559062, + 0.004677759483456612, + 0.0032767108641564846, + 0.00550053920596838, + 0.005022110417485237, + 0.003858037758618593, + 0.004412532784044743, + 0.006654874887317419, + 0.006179481744766235, + 0.006635991856455803, + 0.0046226088888943195, + 0.0019780511502176523, + 0.0012991147814318538, + 0.00163153070025146, + 0.002868796931579709, + 0.0025690211914479733, + 0.002291516400873661, + 0.0017337411409243941, + 0.001833436544984579, + 0.0016213363269343972, + 0.0015993231208994985, + 0.0018847159808501601, + 0.0015168554382398725, + 0.0035416418686509132, + 0.00199307082220912, + 0.0015857481630519032, + 0.0013174803461879492, + 0.046121444553136826, + 0.013397865928709507, + 0.005961683113127947, + 0.006520126014947891, + 0.006696297787129879, + 0.01039770245552063, + 0.03188539296388626, + 0.06519192457199097, + 0.0009073048131540418, + 0.0018887569895014167, + 0.0010665530571714044, + 0.0011933058267459273, + 0.00362657243385911, + 0.0019393912516534328, + 0.0022488832473754883, + 0.0017891477327793837, + 0.0013261104468256235, + 0.0015170342521741986, + 0.0019475162262097, + 0.0014321983326226473, + 0.0034165738616138697, + 0.0016801949823275208, + 0.004118306562304497, + 0.0023491273168474436, + 0.0011943724239245057, + 0.0023598577827215195, + 0.0026387779507786036, + 0.0021260615903884172, + 0.0029922383837401867, + 0.0024677999317646027, + 0.002483768854290247, + 0.0026919094379991293, + 0.006528488826006651, + 0.004706940613687038, + 0.0042740898206830025, + 0.0032275421544909477, + 0.05749402940273285, + 0.011539501138031483, + 0.006128347013145685, + 0.05778396874666214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.0010436094598844647 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043055519461631775 + ], + "opposite_logprob_tokens": [ + -3.168055534362793 + ], + "logprob_reward": -0.043055519461631775, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043055519461631775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2497648447751999, + 0.058070696890354156, + 0.019112227484583855, + 0.019475016742944717, + 0.017886711284518242, + 0.21917538344860077, + 0.01573900133371353, + 0.02315639890730381, + 0.018144095316529274, + 0.35947564244270325 + ], + "credit_map": [ + 0.010572669096291065, + 0.09637150168418884, + 0.0068933493457734585, + 0.005668646190315485, + 0.011553236283361912, + 0.09361902624368668, + 0.00959545373916626, + 0.015490967780351639, + 0.006773464847356081, + 0.009078009985387325, + 0.0049032480455935, + 0.0036762843374162912, + 0.006572675425559282, + 0.013174638152122498, + 0.008674913085997105, + 0.005217458587139845, + 0.002889265539124608, + 0.0026163249276578426, + 0.001360368332825601, + 0.001271505723707378, + 0.001964422408491373, + 0.006534527987241745, + 0.0013707345351576805, + 0.0011050773318856955, + 0.002602630527690053, + 0.002145815407857299, + 0.0021607486996799707, + 0.0026698813308030367, + 0.00305586657486856, + 0.002980969613417983, + 0.0021958048455417156, + 0.001663300092332065, + 0.0020504884887486696, + 0.0024912646040320396, + 0.0020534154027700424, + 0.002120609860867262, + 0.0034606207627803087, + 0.0021449571941047907, + 0.0018042564624920487, + 0.0017610995564609766, + 0.051457297056913376, + 0.015745533630251884, + 0.008204449899494648, + 0.007411313708871603, + 0.010555200278759003, + 0.01183184515684843, + 0.037505507469177246, + 0.07646425813436508, + 0.0013314640382304788, + 0.0016140482621267438, + 0.001189032569527626, + 0.001566187129355967, + 0.0037298668175935745, + 0.0022816385608166456, + 0.0024164870847016573, + 0.0016102769877761602, + 0.001685671741142869, + 0.002302334411069751, + 0.0029247335623949766, + 0.002171470085158944, + 0.004390554968267679, + 0.0024688029661774635, + 0.004033604171127081, + 0.003179224906489253, + 0.00106744933873415, + 0.0020600741263478994, + 0.0025249607861042023, + 0.002505603479221463, + 0.0026708217337727547, + 0.002489999867975712, + 0.0020351209677755833, + 0.0027900643181055784, + 0.016485920175909996, + 0.010920843109488487, + 0.008542212657630444, + 0.007222260348498821, + 0.13906069099903107, + 0.027057329192757607, + 0.012887206859886646, + 0.13729915022850037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.043055519461631775 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, extreme wide-angle shot, trembling almost imperceptibly, simulates violent tectonic shifts, capturing the apocalyptic upheaval of a desolate alien world. Beneath a scarred sky swirling with purple and ochre clouds, the barren, amber-colored ground fractures violently, spiderweb-like cracks running across the wasteland before vast, glowing fissures tear open with catastrophic force. Dominating the horizon is a massive biomechanical foundation, an unholy fusion of petrified ivory-white bone and non-reflective obsidian-black metal. Its architecture defies human logic, with massive ribbed arches and towering spires scraping against the toxic clouds, their surfaces a disturbing blend of porous organic textures and cold alien metallurgy. As the ground tilts, the entire massive structure begins to groan and shift; the movements are agonizingly slow but immensely powerful, as the fused joints of bone and metal rub against each other, the entire base teetering precariously. A deafening, low-pitched protest echoes across the plain as the building tilts precariously, its immense weight causing the edges of the newly formed ravine to crumble. The scene is cast in the dim, eerie glow of a distant, dying sun, creating long, twisted shadows, while the fiery orange light from the cracks casts hellish highlights over the horrific forms of the base, creating a tense and awe-inspiring atmosphere of cosmic terror and imminent collapse.", + "seed": 30949, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_14/index_8892_seed_30949_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022145458310842515, + "vlm_logprob": -0.022145458310842515, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.022145458310842515, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033096957951784134 + ], + "opposite_logprob_tokens": [ + -4.408096790313721 + ], + "logprob_reward": -0.033096957951784134, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.033096957951784134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.149409681558609, + 0.10365378111600876, + 0.11926330626010895, + 0.052706487476825714, + 0.10751140117645264, + 0.07172991335391998, + 0.07891838997602463, + 0.1001802608370781, + 0.1233811303973198, + 0.09324561804533005 + ], + "credit_map": [ + 0.01530839130282402, + 0.016502313315868378, + 0.016540857031941414, + 0.024087319150567055, + 0.014236368238925934, + 0.014923037961125374, + 0.029411839321255684, + 0.018399566411972046, + 0.023570694029331207, + 0.008968840353190899, + 0.00726955384016037, + 0.009844312444329262, + 0.008319734595716, + 0.029643859714269638, + 0.007979384623467922, + 0.00805741362273693, + 0.02001884952187538, + 0.012523860670626163, + 0.009072743356227875, + 0.013248948380351067, + 0.02097555808722973, + 0.02352796122431755, + 0.011809094808995724, + 0.008086303249001503, + 0.009611966088414192, + 0.00534453708678484, + 0.00673313532024622, + 0.006048993207514286, + 0.006678256671875715, + 0.0090272706001997, + 0.004599844571202993, + 0.004662487655878067, + 0.019431790336966515, + 0.009213629178702831, + 0.011752587743103504, + 0.013137750327587128, + 0.009081417694687843, + 0.01032411027699709, + 0.00953963864594698, + 0.025030486285686493, + 0.005577485542744398, + 0.00329332216642797, + 0.005563537590205669, + 0.02936929278075695, + 0.0024547940120100975, + 0.001713125267997384, + 0.015260295942425728, + 0.008498065173625946, + 0.004143478348851204, + 0.005128776654601097, + 0.005328745115548372, + 0.008506865240633488, + 0.005738465115427971, + 0.005100460723042488, + 0.02371380850672722, + 0.02125779539346695, + 0.011549420654773712, + 0.012291643768548965, + 0.011108579114079475, + 0.012210057117044926, + 0.01583302579820156, + 0.009401959367096424, + 0.013700625859200954, + 0.014084958471357822, + 0.013106701895594597, + 0.017952602356672287, + 0.013758698478341103, + 0.01698778010904789, + 0.015926847234368324, + 0.012383966706693172, + 0.020769139751791954, + 0.012495399452745914, + 0.021987129002809525, + 0.014238130301237106, + 0.009181907400488853, + 0.007345949299633503, + 0.014018303714692593, + 0.009389814920723438, + 0.009359553456306458, + 0.007724829949438572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.033096957951784134 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022163845133036375 + ], + "opposite_logprob_tokens": [ + -6.127216339111328 + ], + "logprob_reward": -0.0022163845133036375, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022163845133036375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20987871289253235, + 0.06051456928253174, + 0.06023672968149185, + 0.05471288412809372, + 0.06806588172912598, + 0.2342730611562729, + 0.04651527851819992, + 0.05513972416520119, + 0.16408060491085052, + 0.04658254608511925 + ], + "credit_map": [ + 0.025041431188583374, + 0.024413421750068665, + 0.021704716607928276, + 0.034058064222335815, + 0.022756772115826607, + 0.019914746284484863, + 0.04068496450781822, + 0.02130458690226078, + 0.010800124146044254, + 0.004835500381886959, + 0.0068235271610319614, + 0.007269189227372408, + 0.005010160151869059, + 0.013700789771974087, + 0.005774755496531725, + 0.006300524342805147, + 0.009347077459096909, + 0.006140732206404209, + 0.006948474328964949, + 0.007299861870706081, + 0.01009335182607174, + 0.00916183553636074, + 0.005597132723778486, + 0.005648262798786163, + 0.006846074480563402, + 0.006396468263119459, + 0.005797740537673235, + 0.006507239304482937, + 0.007857779040932655, + 0.007880277000367641, + 0.008269267156720161, + 0.005158035550266504, + 0.011394091881811619, + 0.00561071652919054, + 0.009093508124351501, + 0.008571790531277657, + 0.006446309853345156, + 0.005235038697719574, + 0.006508413702249527, + 0.015206006355583668, + 0.009096752852201462, + 0.007872775197029114, + 0.009188725613057613, + 0.09367801249027252, + 0.008099979721009731, + 0.004321867134422064, + 0.08746075630187988, + 0.014554187655448914, + 0.002568064955994487, + 0.0034352594520896673, + 0.0025725539308041334, + 0.003952009603381157, + 0.003399861976504326, + 0.002765628043562174, + 0.014569181948900223, + 0.01325271651148796, + 0.004621763713657856, + 0.004999903496354818, + 0.0073026856407523155, + 0.010292801074683666, + 0.005403363611549139, + 0.0038792781997472048, + 0.006102180574089289, + 0.01253774669021368, + 0.019340014085173607, + 0.02196682244539261, + 0.02323528192937374, + 0.023189788684248924, + 0.016863154247403145, + 0.013431834056973457, + 0.030778072774410248, + 0.015275616198778152, + 0.010386155918240547, + 0.006087299436330795, + 0.004519897513091564, + 0.004200401715934277, + 0.006558120716363192, + 0.0048542628064751625, + 0.006183680146932602, + 0.003792723873630166 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.0022163845133036375 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549126520752907 + ], + "opposite_logprob_tokens": [ + -4.387548923492432 + ], + "logprob_reward": -0.012549126520752907, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549126520752907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08361907303333282, + 0.12258604913949966, + 0.06029873713850975, + 0.07207073271274567, + 0.05338728427886963, + 0.15709199011325836, + 0.1370517760515213, + 0.13715612888336182, + 0.13972507417201996, + 0.037013210356235504 + ], + "credit_map": [ + 0.008449284359812737, + 0.00836191512644291, + 0.007343105971813202, + 0.015776220709085464, + 0.010086284019052982, + 0.008761831559240818, + 0.01623101904988289, + 0.0086094094440341, + 0.03165064752101898, + 0.011835415847599506, + 0.009903361089527607, + 0.011692661792039871, + 0.007458159234374762, + 0.03764592111110687, + 0.00520533649250865, + 0.007194547913968563, + 0.012857964262366295, + 0.005018114112317562, + 0.004776990506798029, + 0.005409757606685162, + 0.007385543081909418, + 0.016222817823290825, + 0.004109527450054884, + 0.0045180246233940125, + 0.010571207851171494, + 0.0099138543009758, + 0.006491494830697775, + 0.006258466746658087, + 0.011608102358877659, + 0.013266519643366337, + 0.00730919511988759, + 0.006651886273175478, + 0.007904739119112492, + 0.006033885292708874, + 0.005491094663739204, + 0.006716344505548477, + 0.0069183106534183025, + 0.005099473521113396, + 0.005886530969291925, + 0.009336905553936958, + 0.006927927024662495, + 0.006753133609890938, + 0.006245087366551161, + 0.05289565399289131, + 0.00934081245213747, + 0.003618081798776984, + 0.058287687599658966, + 0.013023613020777702, + 0.011078125797212124, + 0.013465531170368195, + 0.00843823328614235, + 0.01469914149492979, + 0.013919386081397533, + 0.014575865119695663, + 0.032511062920093536, + 0.02836444228887558, + 0.01617245562374592, + 0.015919778496026993, + 0.017867615446448326, + 0.017538871616125107, + 0.021251196041703224, + 0.012183013372123241, + 0.014065035618841648, + 0.022158153355121613, + 0.01775469072163105, + 0.015270144678652287, + 0.02920456975698471, + 0.022725775837898254, + 0.014767914079129696, + 0.009082449600100517, + 0.018400901928544044, + 0.012518634088337421, + 0.0044056326150894165, + 0.004502694588154554, + 0.003675036132335663, + 0.004721777979284525, + 0.004795476794242859, + 0.004119921941310167, + 0.0070048170164227486, + 0.003787853056564927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.012549126520752907 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008173224050551653 + ], + "opposite_logprob_tokens": [ + -7.12581729888916 + ], + "logprob_reward": -0.0008173224050551653, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008173224050551653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1348864883184433, + 0.11575670540332794, + 0.06833325326442719, + 0.12054694443941116, + 0.04377460107207298, + 0.13489766418933868, + 0.1539810448884964, + 0.07046489417552948, + 0.12114562839269638, + 0.0362127386033535 + ], + "credit_map": [ + 0.014947810210287571, + 0.015269285999238491, + 0.011902141384780407, + 0.02295541577041149, + 0.017265692353248596, + 0.013527855277061462, + 0.024977872148156166, + 0.014040420763194561, + 0.027584541589021683, + 0.011417753994464874, + 0.011673755943775177, + 0.01181868277490139, + 0.010667580179870129, + 0.029692592099308968, + 0.007114950567483902, + 0.00578684825450182, + 0.007672888692468405, + 0.006448632571846247, + 0.007544991094619036, + 0.00865599513053894, + 0.011484572663903236, + 0.008357414044439793, + 0.012512686662375927, + 0.005656075663864613, + 0.021131645888090134, + 0.01817164570093155, + 0.007541932165622711, + 0.009338222444057465, + 0.01403474248945713, + 0.02701319195330143, + 0.011544253677129745, + 0.011771307326853275, + 0.006195638328790665, + 0.0036955152172595263, + 0.006774174980819225, + 0.007372255902737379, + 0.003917051944881678, + 0.0027114632539451122, + 0.0042967540211975574, + 0.00881174672394991, + 0.013544336892664433, + 0.005657806061208248, + 0.007448722142726183, + 0.05800018087029457, + 0.004161264281719923, + 0.0022294374648481607, + 0.036250315606594086, + 0.0076056006364524364, + 0.008086257614195347, + 0.00891655683517456, + 0.007321610581129789, + 0.011754878796637058, + 0.007092597428709269, + 0.007870679721236229, + 0.05505270138382912, + 0.04788576066493988, + 0.008581329137086868, + 0.008459928445518017, + 0.009301802143454552, + 0.010526089929044247, + 0.008272996172308922, + 0.006892218720167875, + 0.009811115451157093, + 0.008619414642453194, + 0.010585109703242779, + 0.01538664661347866, + 0.013187536969780922, + 0.017880301922559738, + 0.01225748285651207, + 0.01466979831457138, + 0.023984458297491074, + 0.013194301165640354, + 0.0067404815927147865, + 0.006814959924668074, + 0.002469878876581788, + 0.0038267418276518583, + 0.005163830239325762, + 0.0038607202004641294, + 0.004708698485046625, + 0.002627429086714983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.0008173224050551653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204750016331673 + ], + "opposite_logprob_tokens": [ + -2.8120474815368652 + ], + "logprob_reward": -0.06204750016331673, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204750016331673, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06703893840312958, + 0.22308753430843353, + 0.0906815305352211, + 0.11414504796266556, + 0.06108156964182854, + 0.07243616878986359, + 0.12097597867250443, + 0.10158312320709229, + 0.10875537246465683, + 0.040214747190475464 + ], + "credit_map": [ + 0.008212151937186718, + 0.008633493445813656, + 0.0059399777092039585, + 0.006882184650748968, + 0.01427612453699112, + 0.008830041624605656, + 0.00790507160127163, + 0.006359888706356287, + 0.05713561922311783, + 0.01927848346531391, + 0.01986500434577465, + 0.02062087692320347, + 0.014291255734860897, + 0.06853146106004715, + 0.010009725578129292, + 0.013355125673115253, + 0.008713961578905582, + 0.011983750388026237, + 0.008604400791227818, + 0.008269625715911388, + 0.022775424644351006, + 0.009850839152932167, + 0.013983748853206635, + 0.006499782204627991, + 0.016506671905517578, + 0.014142078347504139, + 0.01018364354968071, + 0.007157609332352877, + 0.018311047926545143, + 0.02136748842895031, + 0.012992586940526962, + 0.013483919203281403, + 0.009534191340208054, + 0.006554118357598782, + 0.0054268548265099525, + 0.006760151591151953, + 0.008671014569699764, + 0.004967732820659876, + 0.006518199574202299, + 0.012649305164813995, + 0.003615767229348421, + 0.004121803678572178, + 0.004295738879591227, + 0.024792587384581566, + 0.0039890180341899395, + 0.002750295912846923, + 0.022962255403399467, + 0.0059087020345032215, + 0.015566639602184296, + 0.014910316094756126, + 0.013698115013539791, + 0.01688346080482006, + 0.015203128568828106, + 0.016966404393315315, + 0.015578157268464565, + 0.012169756926596165, + 0.013349857181310654, + 0.013825816102325916, + 0.01239495724439621, + 0.012938898988068104, + 0.014017720706760883, + 0.008974537253379822, + 0.01003231294453144, + 0.01604902371764183, + 0.011571163311600685, + 0.01277969777584076, + 0.02215062640607357, + 0.019831668585538864, + 0.011250018142163754, + 0.008919168263673782, + 0.012109415605664253, + 0.010143614374101162, + 0.008339552208781242, + 0.005273226648569107, + 0.005196508951485157, + 0.00356174074113369, + 0.005664034280925989, + 0.003960000351071358, + 0.004920215345919132, + 0.0032994721550494432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.06204750016331673 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, extreme wide-angle shot, trembling almost imperceptibly, simulates violent tectonic shifts, capturing the apocalyptic upheaval of a desolate alien world. Beneath a scarred sky swirling with purple and ochre clouds, the barren, amber-colored ground fractures violently, spiderweb-like cracks running across the wasteland before vast, glowing fissures tear open with catastrophic force. Dominating the horizon is a massive biomechanical foundation, an unholy fusion of petrified ivory-white bone and non-reflective obsidian-black metal. Its architecture defies human logic, with massive ribbed arches and towering spires scraping against the toxic clouds, their surfaces a disturbing blend of porous organic textures and cold alien metallurgy. As the ground tilts, the entire massive structure begins to groan and shift; the movements are agonizingly slow but immensely powerful, as the fused joints of bone and metal rub against each other, the entire base teetering precariously. A deafening, low-pitched protest echoes across the plain as the building tilts precariously, its immense weight causing the edges of the newly formed ravine to crumble. The scene is cast in the dim, eerie glow of a distant, dying sun, creating long, twisted shadows, while the fiery orange light from the cracks casts hellish highlights over the horrific forms of the base, creating a tense and awe-inspiring atmosphere of cosmic terror and imminent collapse.", + "seed": 30950, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_14/index_8892_seed_30950_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.00980019880225882, + "vlm_logprob": -0.00980019880225882, + "vlm_yes_no_margin": 5.475, + "vlm_token_credit": -0.00980019880225882, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022847039625048637 + ], + "opposite_logprob_tokens": [ + -4.7728471755981445 + ], + "logprob_reward": -0.022847039625048637, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.022847039625048637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1205207109451294, + 0.028825106099247932, + 0.45982781052589417, + 0.07704818993806839, + 0.1177058219909668, + 0.030374230816960335, + 0.03963987156748772, + 0.042230911552906036, + 0.03126723691821098, + 0.052560098469257355 + ], + "credit_map": [ + 0.012649882584810257, + 0.015791289508342743, + 0.014004507102072239, + 0.01585427112877369, + 0.02651449292898178, + 0.011652089655399323, + 0.014727525413036346, + 0.00932665728032589, + 0.005557119380682707, + 0.0029151577036827803, + 0.003945073578506708, + 0.003100400557741523, + 0.0027668888214975595, + 0.00422258535400033, + 0.0031323470175266266, + 0.0031855341512709856, + 0.07627687603235245, + 0.02143663540482521, + 0.018148016184568405, + 0.029802635312080383, + 0.03356408700346947, + 0.038945186883211136, + 0.171326145529747, + 0.0703282430768013, + 0.00649726577103138, + 0.014478366822004318, + 0.010034448467195034, + 0.00763943325728178, + 0.01406881958246231, + 0.009723800234496593, + 0.007299530319869518, + 0.0073065198957920074, + 0.01254420354962349, + 0.013474704697728157, + 0.01009026076644659, + 0.014821506105363369, + 0.01611158438026905, + 0.02226017788052559, + 0.01449220348149538, + 0.013911173678934574, + 0.0050281500443816185, + 0.003390577854588628, + 0.003071586135774851, + 0.0031373652163892984, + 0.0036895619705319405, + 0.005354206543415785, + 0.0032587521709501743, + 0.003444031346589327, + 0.004342496395111084, + 0.005286808591336012, + 0.006287532392889261, + 0.004792042542248964, + 0.003680633148178458, + 0.005468603689223528, + 0.005070724990218878, + 0.0047110277228057384, + 0.004151413217186928, + 0.005365717224776745, + 0.005590066313743591, + 0.004743060562759638, + 0.004046442452818155, + 0.0064024850726127625, + 0.005842441692948341, + 0.0060892850160598755, + 0.0029681024607270956, + 0.003503932151943445, + 0.004116832744330168, + 0.003737430088222027, + 0.003372868523001671, + 0.0042959400452673435, + 0.004900747444480658, + 0.0043713850900530815, + 0.003156619146466255, + 0.0031284631695598364, + 0.002684368984773755, + 0.005670257844030857, + 0.013607999309897423, + 0.0039840396493673325, + 0.00464202743023634, + 0.01568632572889328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.022847039625048637 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001727756578475237 + ], + "opposite_logprob_tokens": [ + -6.37672758102417 + ], + "logprob_reward": -0.001727756578475237, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001727756578475237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16532832384109497, + 0.04578737914562225, + 0.30327463150024414, + 0.18224583566188812, + 0.10055822879076004, + 0.04488031566143036, + 0.039193566888570786, + 0.05052455514669418, + 0.04264482110738754, + 0.025562291964888573 + ], + "credit_map": [ + 0.01583772897720337, + 0.02352953888475895, + 0.019452566280961037, + 0.01991443894803524, + 0.03967767581343651, + 0.016237696632742882, + 0.020361822098493576, + 0.010316865518689156, + 0.0069828396663069725, + 0.004700873978435993, + 0.0056741079315543175, + 0.006132726091891527, + 0.005178174003958702, + 0.00679820915684104, + 0.005708990152925253, + 0.0046114614233374596, + 0.05055045336484909, + 0.012957513332366943, + 0.012225351296365261, + 0.018616944551467896, + 0.022138476371765137, + 0.024682559072971344, + 0.11517328023910522, + 0.046930063515901566, + 0.020485443994402885, + 0.025775928050279617, + 0.020729104056954384, + 0.012095042504370213, + 0.030353952199220657, + 0.0368981696665287, + 0.018282098695635796, + 0.01762610860168934, + 0.011049048975110054, + 0.00908723846077919, + 0.007437285501509905, + 0.005992820952087641, + 0.013807176612317562, + 0.027719726786017418, + 0.014170550741255283, + 0.011294389143586159, + 0.005503062624484301, + 0.0037316621746867895, + 0.004034013953059912, + 0.0036608402151614428, + 0.0070793828926980495, + 0.012022316455841064, + 0.003372054547071457, + 0.005476986523717642, + 0.0037092785350978374, + 0.006936745252460241, + 0.005543063394725323, + 0.005978439003229141, + 0.0036978248972445726, + 0.003947707358747721, + 0.005232695024460554, + 0.004147814586758614, + 0.004498296417295933, + 0.00519597390666604, + 0.004325285088270903, + 0.005095658823847771, + 0.004584594629704952, + 0.013224665075540543, + 0.007143278606235981, + 0.0064568063244223595, + 0.003833865048363805, + 0.0040954239666461945, + 0.004055675584822893, + 0.005586483981460333, + 0.0033952456433326006, + 0.004991058725863695, + 0.01033102534711361, + 0.006356045138090849, + 0.002787772100418806, + 0.004744958132505417, + 0.0020271269604563713, + 0.003623176133260131, + 0.0035965824499726295, + 0.002700590994209051, + 0.0026615511160343885, + 0.0034205366391688585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.001727756578475237 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007636283989995718 + ], + "opposite_logprob_tokens": [ + -4.882636070251465 + ], + "logprob_reward": -0.007636283989995718, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007636283989995718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07755585014820099, + 0.05585029348731041, + 0.15486444532871246, + 0.07617460191249847, + 0.2632688283920288, + 0.038573578000068665, + 0.05498339608311653, + 0.08475178480148315, + 0.10998768359422684, + 0.08398959040641785 + ], + "credit_map": [ + 0.008486566133797169, + 0.012463035993278027, + 0.01035286020487547, + 0.008467096835374832, + 0.011847393587231636, + 0.006158752832561731, + 0.012982798740267754, + 0.006797345355153084, + 0.007770287338644266, + 0.005387601908296347, + 0.00796467438340187, + 0.006554281804710627, + 0.009178439155220985, + 0.007969449274241924, + 0.006258085835725069, + 0.004767472855746746, + 0.02271418273448944, + 0.01340984646230936, + 0.009206091053783894, + 0.011906634084880352, + 0.013009976595640182, + 0.01865517906844616, + 0.04497228190302849, + 0.02099025808274746, + 0.008412368595600128, + 0.010008065961301327, + 0.006269574631005526, + 0.00784545112401247, + 0.010510577820241451, + 0.017423206940293312, + 0.008444949984550476, + 0.007260410580784082, + 0.0284042339771986, + 0.03652298077940941, + 0.021786246448755264, + 0.02339848130941391, + 0.051247209310531616, + 0.043230943381786346, + 0.02783258445560932, + 0.03084617108106613, + 0.003993628080934286, + 0.004966274835169315, + 0.0029776699375361204, + 0.005089585669338703, + 0.005700944922864437, + 0.00734303705394268, + 0.0035714975092560053, + 0.004930941388010979, + 0.00496015977114439, + 0.007566878106445074, + 0.008594558574259281, + 0.007130047306418419, + 0.005031587090343237, + 0.007204416673630476, + 0.007824961096048355, + 0.006670791655778885, + 0.007549125701189041, + 0.007564052473753691, + 0.007586431223899126, + 0.009351366199553013, + 0.0072732968255877495, + 0.022766780108213425, + 0.011969845741987228, + 0.010690882802009583, + 0.009586341679096222, + 0.010578818619251251, + 0.016900377348065376, + 0.012531118467450142, + 0.011138247326016426, + 0.012836079113185406, + 0.02278275415301323, + 0.013633931055665016, + 0.0056078326888382435, + 0.008412365801632404, + 0.005691773258149624, + 0.009654738940298557, + 0.017846085131168365, + 0.007016635034233332, + 0.0075063444674015045, + 0.022253815084695816 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 8 + ], + "selected_reward": -0.007636283989995718 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007211944903247058 + ], + "opposite_logprob_tokens": [ + -7.250720977783203 + ], + "logprob_reward": -0.0007211944903247058, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007211944903247058, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08711478859186172, + 0.02956322766840458, + 0.27179190516471863, + 0.0640350952744484, + 0.14119885861873627, + 0.11091160774230957, + 0.029812080785632133, + 0.1476135402917862, + 0.0905078575015068, + 0.027451032772660255 + ], + "credit_map": [ + 0.010003656148910522, + 0.011619382537901402, + 0.013260120525956154, + 0.0081675685942173, + 0.012465313076972961, + 0.012767894193530083, + 0.010341420769691467, + 0.008489435538649559, + 0.004466219339519739, + 0.0037913082633167505, + 0.003996273502707481, + 0.0028325975872576237, + 0.004934852942824364, + 0.0038539315573871136, + 0.003051812993362546, + 0.002636232180520892, + 0.047191355377435684, + 0.010807257145643234, + 0.011116145178675652, + 0.016345838084816933, + 0.01924995146691799, + 0.020961346104741096, + 0.10453420132398605, + 0.041585810482501984, + 0.007690504193305969, + 0.007065502926707268, + 0.004883105400949717, + 0.004735490772873163, + 0.010456493124365807, + 0.006046278402209282, + 0.017211826518177986, + 0.005945902317762375, + 0.015577176585793495, + 0.02031509205698967, + 0.010508484207093716, + 0.012023836374282837, + 0.022947443649172783, + 0.0279521644115448, + 0.015094837173819542, + 0.016779819503426552, + 0.011081847362220287, + 0.007208742201328278, + 0.006219330243766308, + 0.008478018455207348, + 0.021666986867785454, + 0.03271321579813957, + 0.008905173279345036, + 0.014638297259807587, + 0.0027770556043833494, + 0.00440114364027977, + 0.004649410489946604, + 0.0035412972792983055, + 0.0023245825432240963, + 0.004516545217484236, + 0.004166129045188427, + 0.0034359171986579895, + 0.013691076077520847, + 0.020259538665413857, + 0.015112114138901234, + 0.018019795417785645, + 0.015414239838719368, + 0.031566180288791656, + 0.019312534481287003, + 0.014238069765269756, + 0.007494876626878977, + 0.0066086966544389725, + 0.012425899505615234, + 0.01188134215772152, + 0.00905363168567419, + 0.010320429690182209, + 0.02139173075556755, + 0.011331247165799141, + 0.0026035320479422808, + 0.004437380004674196, + 0.0023549674078822136, + 0.003993041347712278, + 0.0032555789221078157, + 0.002929813228547573, + 0.003123188391327858, + 0.00475353142246604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.0007211944903247058 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160687193274498 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.0160687193274498, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160687193274498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09083042293787003, + 0.075650654733181, + 0.14914272725582123, + 0.1870145946741104, + 0.24570699036121368, + 0.033686213195323944, + 0.10762875527143478, + 0.04173491522669792, + 0.03347918391227722, + 0.03512544184923172 + ], + "credit_map": [ + 0.00921863317489624, + 0.010434994474053383, + 0.013091574423015118, + 0.009435528889298439, + 0.01512871216982603, + 0.012828751467168331, + 0.011813503690063953, + 0.008878737688064575, + 0.01267544087022543, + 0.009382644668221474, + 0.007441506255418062, + 0.010105857625603676, + 0.008959670551121235, + 0.01106189750134945, + 0.009066807106137276, + 0.006956844590604305, + 0.022089004516601562, + 0.008405760861933231, + 0.006791942287236452, + 0.010686838068068027, + 0.011603965424001217, + 0.01714116334915161, + 0.050211768597364426, + 0.02221229299902916, + 0.021549200639128685, + 0.03013187274336815, + 0.019383015111088753, + 0.01503017172217369, + 0.03304344788193703, + 0.018385382369160652, + 0.03190265968441963, + 0.017588874325156212, + 0.026026004925370216, + 0.03412378579378128, + 0.022323722019791603, + 0.020297115668654442, + 0.041502173990011215, + 0.04835337772965431, + 0.022706788033246994, + 0.03037404641509056, + 0.0032606518361717463, + 0.003068017540499568, + 0.002361447084695101, + 0.0036390258464962244, + 0.006862954702228308, + 0.006115978583693504, + 0.003849147353321314, + 0.004528991878032684, + 0.010152858681976795, + 0.0166060421615839, + 0.015257555991411209, + 0.015009861439466476, + 0.009469368495047092, + 0.01247149147093296, + 0.014755734242498875, + 0.01390585396438837, + 0.0029098934028297663, + 0.0028988965786993504, + 0.0026055071502923965, + 0.0028362602461129427, + 0.003450378542765975, + 0.012539591640233994, + 0.00513651967048645, + 0.009357874281704426, + 0.003201445797458291, + 0.0037379800342023373, + 0.0033240006305277348, + 0.004190594889223576, + 0.0032184028532356024, + 0.0059584747068583965, + 0.005541766528040171, + 0.004306521732360125, + 0.0025799537543207407, + 0.006119707599282265, + 0.0027887667529284954, + 0.0038199371192604303, + 0.0048493617214262486, + 0.004435866605490446, + 0.0034619420766830444, + 0.0070699118077754974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.0160687193274498 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, extreme wide-angle shot, trembling almost imperceptibly, simulates violent tectonic shifts, capturing the apocalyptic upheaval of a desolate alien world. Beneath a scarred sky swirling with purple and ochre clouds, the barren, amber-colored ground fractures violently, spiderweb-like cracks running across the wasteland before vast, glowing fissures tear open with catastrophic force. Dominating the horizon is a massive biomechanical foundation, an unholy fusion of petrified ivory-white bone and non-reflective obsidian-black metal. Its architecture defies human logic, with massive ribbed arches and towering spires scraping against the toxic clouds, their surfaces a disturbing blend of porous organic textures and cold alien metallurgy. As the ground tilts, the entire massive structure begins to groan and shift; the movements are agonizingly slow but immensely powerful, as the fused joints of bone and metal rub against each other, the entire base teetering precariously. A deafening, low-pitched protest echoes across the plain as the building tilts precariously, its immense weight causing the edges of the newly formed ravine to crumble. The scene is cast in the dim, eerie glow of a distant, dying sun, creating long, twisted shadows, while the fiery orange light from the cracks casts hellish highlights over the horrific forms of the base, creating a tense and awe-inspiring atmosphere of cosmic terror and imminent collapse.", + "seed": 30951, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_14/index_8892_seed_30951_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.012146811839193106, + "vlm_logprob": -0.012146811839193106, + "vlm_yes_no_margin": 5.25, + "vlm_token_credit": -0.012146811839193106, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022336216643452644 + ], + "opposite_logprob_tokens": [ + -4.397336006164551 + ], + "logprob_reward": -0.022336216643452644, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.022336216643452644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5713767409324646, + 0.0835883840918541, + 0.026465067639946938, + 0.021664177998900414, + 0.02890966087579727, + 0.013082088902592659, + 0.05383198708295822, + 0.012158731929957867, + 0.049983758479356766, + 0.13893945515155792 + ], + "credit_map": [ + 0.09656060487031937, + 0.025489289313554764, + 0.03571993485093117, + 0.1371075063943863, + 0.015439231880009174, + 0.021756278350949287, + 0.1300356537103653, + 0.1092681959271431, + 0.008343200199306011, + 0.005194865167140961, + 0.007867182604968548, + 0.005702256690710783, + 0.008110416121780872, + 0.009195318445563316, + 0.0207778699696064, + 0.018397271633148193, + 0.003015286521986127, + 0.002227739430963993, + 0.003727474482730031, + 0.0029918092768639326, + 0.003580411896109581, + 0.0031395868863910437, + 0.003024244448170066, + 0.00475851446390152, + 0.0021555547136813402, + 0.0021409480832517147, + 0.0024010546039789915, + 0.0026723071932792664, + 0.0025253386702388525, + 0.001557659707032144, + 0.0038622396532446146, + 0.004349072463810444, + 0.002931630238890648, + 0.001715671387501061, + 0.002833201549947262, + 0.0037482697516679764, + 0.0018736958736553788, + 0.0024123317562043667, + 0.006773041561245918, + 0.006621813867241144, + 0.0016787574859336019, + 0.0011441424721851945, + 0.0019343090243637562, + 0.001654809806495905, + 0.001379975350573659, + 0.0013878000900149345, + 0.002348671667277813, + 0.0015536214923486114, + 0.011779614724218845, + 0.0028605470433831215, + 0.0038594675716012716, + 0.0023132420610636473, + 0.001973322592675686, + 0.003747418290004134, + 0.005817146506160498, + 0.0214812234044075, + 0.0013280599378049374, + 0.0012529746163636446, + 0.0017772322753444314, + 0.001156297861598432, + 0.0016921551432460546, + 0.0014343013754114509, + 0.0018710825825110078, + 0.0016466261586174369, + 0.007148655131459236, + 0.0022624863777309656, + 0.0034500774927437305, + 0.0021368800662457943, + 0.0032953107729554176, + 0.0030067297630012035, + 0.019647939130663872, + 0.009035671129822731, + 0.014236420392990112, + 0.012177577242255211, + 0.006306661292910576, + 0.0030555101111531258, + 0.003329189494252205, + 0.013564301654696465, + 0.032916110008955, + 0.0533536821603775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.022336216643452644 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013455632142722607 + ], + "opposite_logprob_tokens": [ + -6.626345634460449 + ], + "logprob_reward": -0.0013455632142722607, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013455632142722607, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6822419166564941, + 0.05434451997280121, + 0.012498854659497738, + 0.01879151538014412, + 0.014122456312179565, + 0.009558668360114098, + 0.015491551719605923, + 0.007187562994658947, + 0.004042529966682196, + 0.18172042071819305 + ], + "credit_map": [ + 0.0779409259557724, + 0.02230633795261383, + 0.0327640026807785, + 0.14544224739074707, + 0.014135120436549187, + 0.02003682218492031, + 0.2091158628463745, + 0.160500630736351, + 0.004767013248056173, + 0.0060170185752213, + 0.0036354276817291975, + 0.00411473773419857, + 0.006546508986502886, + 0.005426133517175913, + 0.012460256926715374, + 0.011377430520951748, + 0.0018022967269644141, + 0.0013385697966441512, + 0.001845821039751172, + 0.0011024378472939134, + 0.0013962454395368695, + 0.0014005906414240599, + 0.0013350843219086528, + 0.0022778103593736887, + 0.0013646158622577786, + 0.0014280457980930805, + 0.0022030516993254423, + 0.001737984362989664, + 0.0018293539760634303, + 0.001260113320313394, + 0.0042948732152581215, + 0.004673478193581104, + 0.0014679764863103628, + 0.0007990856538526714, + 0.001695715938694775, + 0.0013032315764576197, + 0.001001268159598112, + 0.0013961561489850283, + 0.003432186786085367, + 0.0030268365517258644, + 0.0016745791072025895, + 0.0008175959810614586, + 0.0014535535592585802, + 0.00160304456949234, + 0.0007233249489217997, + 0.0008660068269819021, + 0.0012794623617082834, + 0.0011411021696403623, + 0.003214876865968108, + 0.0008362934459000826, + 0.0013331032823771238, + 0.000718806404620409, + 0.0006224398966878653, + 0.0011587012559175491, + 0.0015946856001392007, + 0.006012645550072193, + 0.0007301243604160845, + 0.0006863832823000848, + 0.0013399351155385375, + 0.0005223086918704212, + 0.0007831932744011283, + 0.0008992553921416402, + 0.0010822921758517623, + 0.00114407145883888, + 0.0005379593349061906, + 0.00030687873368151486, + 0.00044697956764139235, + 0.00027152185793966055, + 0.0004990649176761508, + 0.0004072923038620502, + 0.001063477131538093, + 0.0005093561485409737, + 0.018240751698613167, + 0.0159300547093153, + 0.007924092002213001, + 0.0038956869393587112, + 0.004229753743857145, + 0.01658337190747261, + 0.0443558506667614, + 0.07056088745594025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0013455632142722607 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009787074290215969 + ], + "opposite_logprob_tokens": [ + -4.634787082672119 + ], + "logprob_reward": -0.009787074290215969, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009787074290215969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6056016087532043, + 0.07306491583585739, + 0.016653109341859818, + 0.03610294684767723, + 0.031551044434309006, + 0.023459844291210175, + 0.016649965196847916, + 0.013741103932261467, + 0.045881472527980804, + 0.13729390501976013 + ], + "credit_map": [ + 0.05937441438436508, + 0.019292490556836128, + 0.02700597606599331, + 0.11719239503145218, + 0.012897422537207603, + 0.01801079325377941, + 0.19686679542064667, + 0.15496142208576202, + 0.007409805431962013, + 0.0045999642461538315, + 0.008260861039161682, + 0.006101272068917751, + 0.0067663658410310745, + 0.011765819042921066, + 0.014964105561375618, + 0.01319673378020525, + 0.0015234386082738638, + 0.0021349976304918528, + 0.002184984739869833, + 0.001798706711269915, + 0.0024280815850943327, + 0.0013911131536588073, + 0.0022681779228150845, + 0.002923613414168358, + 0.0032381645869463682, + 0.0041292826645076275, + 0.00694953091442585, + 0.006200919393450022, + 0.0037184797693043947, + 0.00323122413828969, + 0.003942915238440037, + 0.004692437592893839, + 0.0026179926935583353, + 0.0015177633613348007, + 0.0030280346982181072, + 0.00366200995631516, + 0.0013815152924507856, + 0.002434886060655117, + 0.008104577660560608, + 0.008804270066320896, + 0.0037151554133743048, + 0.0018480307189747691, + 0.004608877003192902, + 0.0032012800220400095, + 0.002020966960117221, + 0.0024510128423571587, + 0.0034464849159121513, + 0.0021680407226085663, + 0.0030130839440971613, + 0.0010077095357701182, + 0.0016747490735724568, + 0.0011548995971679688, + 0.0008381802472285926, + 0.0015271465526893735, + 0.0017649744404479861, + 0.005669225938618183, + 0.0016684798756614327, + 0.001268938765861094, + 0.0015387128805741668, + 0.001182260224595666, + 0.002159097231924534, + 0.0011068376479670405, + 0.0027964089531451464, + 0.0020203706808388233, + 0.006424971390515566, + 0.0021119392476975918, + 0.002592272125184536, + 0.0023267518263310194, + 0.00280356896109879, + 0.0028497104067355394, + 0.01807858981192112, + 0.008693675510585308, + 0.014454067684710026, + 0.011559453792870045, + 0.005809119902551174, + 0.0035201795399188995, + 0.0035698330029845238, + 0.012640519067645073, + 0.03252948448061943, + 0.0532112680375576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.009787074290215969 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009224927052855492 + ], + "opposite_logprob_tokens": [ + -7.000922679901123 + ], + "logprob_reward": -0.0009224927052855492, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009224927052855492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7312765121459961, + 0.08422013372182846, + 0.010005897842347622, + 0.015508262440562248, + 0.017367500811815262, + 0.0066288067027926445, + 0.017750708386301994, + 0.00981354434043169, + 0.0238958727568388, + 0.08353275060653687 + ], + "credit_map": [ + 0.07915937155485153, + 0.024433443322777748, + 0.036434974521398544, + 0.1876726895570755, + 0.017393235117197037, + 0.021352378651499748, + 0.22562462091445923, + 0.13920578360557556, + 0.008378781378269196, + 0.010680495761334896, + 0.007180657237768173, + 0.008329568430781364, + 0.012383314780890942, + 0.00795123353600502, + 0.014210241846740246, + 0.015105844475328922, + 0.0009016942931339145, + 0.0011251037940382957, + 0.0012722595129162073, + 0.0012190776178613305, + 0.0015485944459214807, + 0.0009302053949795663, + 0.0017129882471635938, + 0.0012959737796336412, + 0.0012471559457480907, + 0.002537406049668789, + 0.0014412940945476294, + 0.0015167311066761613, + 0.0014094258658587933, + 0.000987981678918004, + 0.002696237526834011, + 0.0036720300558954477, + 0.001243735896423459, + 0.0007032190333120525, + 0.0015765337739139795, + 0.000870902556926012, + 0.0009835578966885805, + 0.0012063803151249886, + 0.005311766639351845, + 0.005471404176205397, + 0.0007331231026910245, + 0.0005552286165766418, + 0.0006443090969696641, + 0.0006476555718109012, + 0.0010336042614653707, + 0.0011282069608569145, + 0.0011197358835488558, + 0.0007669429178349674, + 0.0038450995925813913, + 0.0009701424860395491, + 0.0012459318386390805, + 0.0007828144589439034, + 0.0007035694434307516, + 0.0013432574924081564, + 0.001936333137564361, + 0.006923559587448835, + 0.001110157580114901, + 0.0009833286749199033, + 0.0008653434924781322, + 0.0008022364345379174, + 0.001324534765444696, + 0.0009075371781364083, + 0.002396782860159874, + 0.0014236230636015534, + 0.003466222668066621, + 0.0011139045236632228, + 0.001790652284398675, + 0.0014429313596338034, + 0.0017306351801380515, + 0.0015386400045827031, + 0.008827963843941689, + 0.00398492394015193, + 0.008173832669854164, + 0.007450391072779894, + 0.0035942799877375364, + 0.0025158103089779615, + 0.0028666057623922825, + 0.007765891030430794, + 0.01976974867284298, + 0.03139619156718254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0009224927052855492 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle with an extreme wide-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a desolate alien world with barren, amber-colored ground and a scarred sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are vast, glowing fissures visible tearing open in the ground with spiderweb-like cracks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a massive biomechanical structure dominating the horizon, composed of ivory-white bone and obsidian-black metal?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the massive structure tilt and shift slowly while the ground fractures beneath it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026342712342739105 + ], + "opposite_logprob_tokens": [ + -3.6513426303863525 + ], + "logprob_reward": -0.026342712342739105, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026342712342739105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4587115943431854, + 0.16611211001873016, + 0.02726469375193119, + 0.031835418194532394, + 0.018954601138830185, + 0.009658286347985268, + 0.0189187154173851, + 0.01947254315018654, + 0.007102178875356913, + 0.24196982383728027 + ], + "credit_map": [ + 0.053803689777851105, + 0.017408350482583046, + 0.02783970534801483, + 0.0908006876707077, + 0.013268486596643925, + 0.01756158657371998, + 0.1321805864572525, + 0.10584855824708939, + 0.015376199036836624, + 0.02458067424595356, + 0.013486294075846672, + 0.015461727045476437, + 0.02559521608054638, + 0.01362944208085537, + 0.02773166447877884, + 0.030250906944274902, + 0.004066137131303549, + 0.00218402361497283, + 0.002948347944766283, + 0.003682164940983057, + 0.002125050639733672, + 0.003164532594382763, + 0.004048534203320742, + 0.005045903846621513, + 0.0018766603898257017, + 0.002201962983235717, + 0.0021397643722593784, + 0.0018604559591040015, + 0.0030925238970667124, + 0.0020434479229152203, + 0.008707648143172264, + 0.009912957437336445, + 0.0016880545299500227, + 0.0018732462776824832, + 0.0025631189346313477, + 0.002009063493460417, + 0.0022562339436262846, + 0.0023909045848995447, + 0.0034846377093344927, + 0.0026893431786447763, + 0.0009920882293954492, + 0.0007249322370626032, + 0.0014141388237476349, + 0.001484550186432898, + 0.0011588999768719077, + 0.0011356727918609977, + 0.0012806281447410583, + 0.0014673768309876323, + 0.0035419228952378035, + 0.0013939368072897196, + 0.001599763287231326, + 0.001140843378379941, + 0.001125057227909565, + 0.0018361627589911222, + 0.0018561379984021187, + 0.006424892693758011, + 0.002383021404966712, + 0.0022964696399867535, + 0.0015094912378117442, + 0.0016514938324689865, + 0.0027348296716809273, + 0.0019534260500222445, + 0.004143995698541403, + 0.0027998173609375954, + 0.0008969471091404557, + 0.0006462805904448032, + 0.0009323333506472409, + 0.0007750961231067777, + 0.0013288359623402357, + 0.0007520474027842283, + 0.0008663538028486073, + 0.0009042852907441556, + 0.02479882910847664, + 0.021494215354323387, + 0.010503746569156647, + 0.005561530124396086, + 0.005479426588863134, + 0.02181111089885235, + 0.05871731787919998, + 0.09360366314649582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.026342712342739105 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_15/index_tensor([3595, 3595])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_15/index_tensor([3595, 3595])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..20197a9962cf8b6ce6d7ce269c025a1a512f8a32 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_15/index_tensor([3595, 3595])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An epic, low-angle drone shot follows a knight as he gallops along the ruins of the ancient Great Wall. The knight is clad in a weathered, dark blue linen robe that flutters in the wind, revealing the hilt of a longsword slung across his back. His long, dark hair is simply tied up, a few strands whipped by the wind against his stern, angular face. He leans over and grips the reins, becoming one with the muscular, sweat-soaked dark brown steed beneath him. His hooves pound the massive, cracked, moss-covered stone bricks, sending up a cloud of yellow dust and rubble. The camera slowly zooms out, revealing the vast, majestic mountains in the background, silhouetted in varying shades against the golden sunset. The low-hanging sun casts long shadows on the city walls winding over the mountain ridge, giving the entire picture a warm and tragic tone. The dust in the air shimmers with golden particles in the light, creating a lonely, desolate and dynamic film atmosphere.", + "seed": 23807, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_15/index_3595_seed_23807_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05345461368560791, + "vlm_logprob": -0.05345461368560791, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.05345461368560791, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.167634978890419 + ], + "opposite_logprob_tokens": [ + -1.9176349639892578 + ], + "logprob_reward": -0.167634978890419, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.167634978890419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1299644559621811, + 0.11805729568004608, + 0.4152306020259857, + 0.08472153544425964, + 0.028712831437587738, + 0.04338759928941727, + 0.0231535192579031, + 0.03815385699272156, + 0.061412472277879715, + 0.05720588192343712 + ], + "credit_map": [ + 0.022015275433659554, + 0.024471189826726913, + 0.00991838239133358, + 0.014214481227099895, + 0.01119414996355772, + 0.02105378545820713, + 0.017886104062199593, + 0.009211087599396706, + 0.04082605615258217, + 0.025374839082360268, + 0.009575920179486275, + 0.0037824774626642466, + 0.0071634831838309765, + 0.005407614633440971, + 0.00408746674656868, + 0.021839434280991554, + 0.0848744586110115, + 0.1464325189590454, + 0.08847784250974655, + 0.01653948239982128, + 0.03521016612648964, + 0.011169792152941227, + 0.012145869433879852, + 0.020380476489663124, + 0.015729565173387527, + 0.014832163229584694, + 0.014551125466823578, + 0.0047712125815451145, + 0.012579397298395634, + 0.008262564428150654, + 0.009462532587349415, + 0.004532971885055304, + 0.0028504468500614166, + 0.003179417923092842, + 0.00393671402707696, + 0.0027637039311230183, + 0.003924225457012653, + 0.0034683176781982183, + 0.0043562003411352634, + 0.0042338063940405846, + 0.005284969229251146, + 0.0063611166551709175, + 0.0036654032301157713, + 0.004013334400951862, + 0.007955235429108143, + 0.005624373909085989, + 0.006107648368924856, + 0.004375517833977938, + 0.003141536843031645, + 0.0032900278456509113, + 0.001799918944016099, + 0.0023692194372415543, + 0.003743548644706607, + 0.002705028047785163, + 0.004030171781778336, + 0.002074066549539566, + 0.0065482924692332745, + 0.004782210104167461, + 0.0028105832170695066, + 0.0040785097517073154, + 0.006609862204641104, + 0.003952936269342899, + 0.004907834809273481, + 0.004463626071810722, + 0.00816928781569004, + 0.008388562127947807, + 0.007007380947470665, + 0.008623309433460236, + 0.009496212005615234, + 0.005552427843213081, + 0.008069563657045364, + 0.006105725187808275, + 0.006503297947347164, + 0.007467077579349279, + 0.006066181696951389, + 0.005689334124326706, + 0.011132778599858284, + 0.0044968873262405396, + 0.010958443395793438, + 0.004891882184892893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.167634978890419 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026373596861958504 + ], + "opposite_logprob_tokens": [ + -3.6513736248016357 + ], + "logprob_reward": -0.026373596861958504, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026373596861958504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.252705454826355, + 0.04381652548909187, + 0.24843525886535645, + 0.08682627975940704, + 0.05231295898556709, + 0.04157480597496033, + 0.07504290342330933, + 0.07206409424543381, + 0.06700877845287323, + 0.060212887823581696 + ], + "credit_map": [ + 0.047390926629304886, + 0.04532459378242493, + 0.03153657913208008, + 0.01632845401763916, + 0.0164021123200655, + 0.04490586742758751, + 0.023228546604514122, + 0.02758840285241604, + 0.005916144233196974, + 0.007602774538099766, + 0.006780126132071018, + 0.004456064198166132, + 0.005434359423816204, + 0.004379213321954012, + 0.00377787952311337, + 0.005469966679811478, + 0.041492804884910583, + 0.08266399055719376, + 0.06435977667570114, + 0.011089368723332882, + 0.021406715735793114, + 0.005949510261416435, + 0.007417426444590092, + 0.014055690728127956, + 0.012630250304937363, + 0.018243933096528053, + 0.012574448250234127, + 0.005642305128276348, + 0.015579954721033573, + 0.006078785285353661, + 0.0108591103926301, + 0.005217501427978277, + 0.005182003136724234, + 0.007028362713754177, + 0.0078021506778895855, + 0.004784342832863331, + 0.01031702570617199, + 0.004713357891887426, + 0.009269029833376408, + 0.0032166927121579647, + 0.004588949494063854, + 0.004907993599772453, + 0.003932594787329435, + 0.0038165694568306208, + 0.0073472256772220135, + 0.005601325072348118, + 0.006840130314230919, + 0.004540021996945143, + 0.013602768070995808, + 0.011937103234231472, + 0.0059793684631586075, + 0.006522550713270903, + 0.012795250862836838, + 0.008506279438734055, + 0.008988984860479832, + 0.006710600107908249, + 0.011620782315731049, + 0.008496991358697414, + 0.004816572181880474, + 0.007749032694846392, + 0.013544421643018723, + 0.008521905168890953, + 0.010130194015800953, + 0.007184201385825872, + 0.009890615940093994, + 0.014369355514645576, + 0.00694605428725481, + 0.004823015537112951, + 0.010942424647510052, + 0.007336577400565147, + 0.007643464952707291, + 0.005057272966951132, + 0.008705317974090576, + 0.012614228762686253, + 0.0065607186406850815, + 0.0052962233312428, + 0.010541588068008423, + 0.004941013641655445, + 0.006724267732352018, + 0.004829530604183674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.026373596861958504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369303047657013 + ], + "opposite_logprob_tokens": [ + -3.4086930751800537 + ], + "logprob_reward": -0.03369303047657013, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369303047657013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46094536781311035, + 0.03509310260415077, + 0.3311994969844818, + 0.036354709416627884, + 0.015353254973888397, + 0.017569266259670258, + 0.020216291770339012, + 0.04749540984630585, + 0.025613700971007347, + 0.010159378871321678 + ], + "credit_map": [ + 0.08379894495010376, + 0.08476084470748901, + 0.0617847815155983, + 0.030535168945789337, + 0.02885245718061924, + 0.0773075520992279, + 0.04314284771680832, + 0.05076276883482933, + 0.008998987264931202, + 0.006387624423950911, + 0.004021838773041964, + 0.0017165576573461294, + 0.004563066177070141, + 0.002443184144794941, + 0.0020929218735545874, + 0.0048689222894608974, + 0.04502926394343376, + 0.10458070039749146, + 0.09459611028432846, + 0.015314311720430851, + 0.03103533759713173, + 0.010694087482988834, + 0.009152209386229515, + 0.0207974836230278, + 0.006372438743710518, + 0.007570725865662098, + 0.00362273957580328, + 0.0019192678155377507, + 0.005248473957180977, + 0.005916161462664604, + 0.0030439922120422125, + 0.0026609094347804785, + 0.002105420222505927, + 0.0025502105709165335, + 0.0015924925683066249, + 0.0014611167134717107, + 0.002791974926367402, + 0.0016888173995539546, + 0.001969113713130355, + 0.0011941088596358895, + 0.0025422575417906046, + 0.0027448898181319237, + 0.0015613288851454854, + 0.001466390211135149, + 0.0029948814772069454, + 0.0017269743839278817, + 0.0028837532736361027, + 0.0016487909015268087, + 0.0027766029816120863, + 0.003031476167961955, + 0.002114506671205163, + 0.0025519176851958036, + 0.0025585361290723085, + 0.002834971761330962, + 0.002396578202024102, + 0.0019517035689204931, + 0.00916046742349863, + 0.006640030071139336, + 0.0036986451596021652, + 0.005163309630006552, + 0.005893807392567396, + 0.005457201972603798, + 0.0060529569163918495, + 0.0054289912804961205, + 0.0037651308812201023, + 0.004779814276844263, + 0.0029350798577070236, + 0.0031112506985664368, + 0.003383328439667821, + 0.002412155969068408, + 0.0023939048405736685, + 0.002833036705851555, + 0.0010975574841722846, + 0.0011769027914851904, + 0.0009574474534019828, + 0.0010398353915661573, + 0.002002091845497489, + 0.0009200487984344363, + 0.0019303145818412304, + 0.0010351805249229074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.03369303047657013 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029792651534080505 + ], + "opposite_logprob_tokens": [ + -3.529792547225952 + ], + "logprob_reward": -0.029792651534080505, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029792651534080505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1463700234889984, + 0.13784447312355042, + 0.3731992244720459, + 0.21381917595863342, + 0.026009762659668922, + 0.022746365517377853, + 0.016315635293722153, + 0.016733724623918533, + 0.028895271942019463, + 0.0180664099752903 + ], + "credit_map": [ + 0.0260528065264225, + 0.03012824058532715, + 0.013901314698159695, + 0.010682706721127033, + 0.012289110571146011, + 0.025662483647465706, + 0.016081715002655983, + 0.011571631766855717, + 0.031147580593824387, + 0.036127299070358276, + 0.01744183897972107, + 0.01008671149611473, + 0.015931086614727974, + 0.011082001961767673, + 0.00804831087589264, + 0.007979631423950195, + 0.04980085417628288, + 0.11422755569219589, + 0.10846397280693054, + 0.018654124811291695, + 0.0353650264441967, + 0.01104454044252634, + 0.011930129490792751, + 0.02371295727789402, + 0.04678170382976532, + 0.03321975842118263, + 0.04308426007628441, + 0.01191397849470377, + 0.0284504983574152, + 0.018844295293092728, + 0.02063683234155178, + 0.010887826792895794, + 0.0031268983148038387, + 0.003826508764177561, + 0.0026150865014642477, + 0.0033561585005372763, + 0.004455347545444965, + 0.002899936866015196, + 0.003248518565669656, + 0.0024813057389110327, + 0.003690467681735754, + 0.0030876200180500746, + 0.002232499420642853, + 0.0023517401423305273, + 0.0030655069276690483, + 0.0017237205756828189, + 0.0039023521821945906, + 0.0026924575213342905, + 0.002446749247610569, + 0.003042715834453702, + 0.0016087517142295837, + 0.00105875579174608, + 0.0029088028240948915, + 0.001821722718887031, + 0.0017047688597813249, + 0.0017233664402738214, + 0.002328032860532403, + 0.0017240582965314388, + 0.0011880981037393212, + 0.002035570563748479, + 0.0021281458903104067, + 0.0021014518570154905, + 0.002819907385855913, + 0.0024084565229713917, + 0.003608717815950513, + 0.004090048838406801, + 0.0027129976078867912, + 0.003177983919158578, + 0.004757729358971119, + 0.0033013520296663046, + 0.004395006690174341, + 0.0028514324221760035, + 0.002711440669372678, + 0.00256793899461627, + 0.002089916029945016, + 0.0018453204538673162, + 0.003177663078531623, + 0.0012493330286815763, + 0.002632654272019863, + 0.0017921407707035542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.029792651534080505 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778810665011406 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009778810665011406, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778810665011406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22138665616512299, + 0.16767361760139465, + 0.27850550413131714, + 0.06399238854646683, + 0.02744094654917717, + 0.025022275745868683, + 0.02947494387626648, + 0.0310757327824831, + 0.07401321083307266, + 0.08141469955444336 + ], + "credit_map": [ + 0.03874707967042923, + 0.030398735776543617, + 0.024663161486387253, + 0.01963813416659832, + 0.01779964566230774, + 0.03775086998939514, + 0.022663133218884468, + 0.029725929722189903, + 0.057249173521995544, + 0.03901512548327446, + 0.01345012616366148, + 0.006890296004712582, + 0.0105789415538311, + 0.008181868121027946, + 0.005918108392506838, + 0.02638999931514263, + 0.03126012906432152, + 0.07908502221107483, + 0.08555760979652405, + 0.014533457346260548, + 0.029270799830555916, + 0.010548267513513565, + 0.009290389716625214, + 0.018959863111376762, + 0.01377029623836279, + 0.009479474276304245, + 0.01354882214218378, + 0.00548310112208128, + 0.007855561561882496, + 0.004342667292803526, + 0.005890278611332178, + 0.003622194519266486, + 0.0036714135203510523, + 0.005862208548933268, + 0.003517893608659506, + 0.0034534460864961147, + 0.0035551786422729492, + 0.0023519762326031923, + 0.0031601195223629475, + 0.0018687135307118297, + 0.003184304339811206, + 0.0036798417568206787, + 0.0022212553303688765, + 0.0023564985021948814, + 0.00467626703903079, + 0.0035580650437623262, + 0.0037068193778395653, + 0.0016392257530242205, + 0.004035943653434515, + 0.00541370315477252, + 0.002694908529520035, + 0.0039029621984809637, + 0.0044898600317537785, + 0.002617687452584505, + 0.003359580645337701, + 0.002960299840196967, + 0.005888055078685284, + 0.005343738477677107, + 0.002430759835988283, + 0.0035658194683492184, + 0.0028460663743317127, + 0.003966005053371191, + 0.003385875141248107, + 0.0036494156811386347, + 0.011668930761516094, + 0.015147748403251171, + 0.006609310396015644, + 0.006988359149545431, + 0.01376941055059433, + 0.008374199271202087, + 0.00558104831725359, + 0.005874208640307188, + 0.011002886109054089, + 0.013216108083724976, + 0.00806126743555069, + 0.007735906634479761, + 0.014758359640836716, + 0.008117398247122765, + 0.009949891828000546, + 0.008572885766625404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.009778810665011406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, low-angle drone shot follows a knight as he gallops along the ruins of the ancient Great Wall. The knight is clad in a weathered, dark blue linen robe that flutters in the wind, revealing the hilt of a longsword slung across his back. His long, dark hair is simply tied up, a few strands whipped by the wind against his stern, angular face. He leans over and grips the reins, becoming one with the muscular, sweat-soaked dark brown steed beneath him. His hooves pound the massive, cracked, moss-covered stone bricks, sending up a cloud of yellow dust and rubble. The camera slowly zooms out, revealing the vast, majestic mountains in the background, silhouetted in varying shades against the golden sunset. The low-hanging sun casts long shadows on the city walls winding over the mountain ridge, giving the entire picture a warm and tragic tone. The dust in the air shimmers with golden particles in the light, creating a lonely, desolate and dynamic film atmosphere.", + "seed": 23808, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_15/index_3595_seed_23808_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10462904460728169, + "vlm_logprob": -0.10462904460728169, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.10462904460728169, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.059641629457473755 + ], + "opposite_logprob_tokens": [ + -3.0596415996551514 + ], + "logprob_reward": -0.059641629457473755, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.059641629457473755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16986405849456787, + 0.08999122679233551, + 0.13689684867858887, + 0.05177682265639305, + 0.04651152342557907, + 0.032234229147434235, + 0.08028178662061691, + 0.2948192059993744, + 0.08001235127449036, + 0.017611974850296974 + ], + "credit_map": [ + 0.014628935605287552, + 0.013821445405483246, + 0.016778849065303802, + 0.01972712203860283, + 0.031825076788663864, + 0.02621549740433693, + 0.025783617049455643, + 0.021083500236272812, + 0.006600010208785534, + 0.006781584117561579, + 0.00881476141512394, + 0.007349016610532999, + 0.009894530288875103, + 0.018884213641285896, + 0.008719613775610924, + 0.022947482764720917, + 0.005794934928417206, + 0.008599434979259968, + 0.010619202628731728, + 0.017593171447515488, + 0.00694715091958642, + 0.038156408816576004, + 0.010091367177665234, + 0.03909516707062721, + 0.003744683926925063, + 0.006321873050183058, + 0.007532029412686825, + 0.006127986125648022, + 0.005635359324514866, + 0.008437732234597206, + 0.00828692689538002, + 0.005690224468708038, + 0.003323204582557082, + 0.004851915407925844, + 0.005173533223569393, + 0.004638500511646271, + 0.011569657362997532, + 0.005166932474821806, + 0.007599111180752516, + 0.004188667517155409, + 0.0024297810159623623, + 0.002985896309837699, + 0.004328391049057245, + 0.004467810969799757, + 0.008029463700950146, + 0.0035951058380305767, + 0.0036745744291692972, + 0.0027232046704739332, + 0.011828159913420677, + 0.00833553820848465, + 0.009633234702050686, + 0.0081474082544446, + 0.00710982084274292, + 0.0114181162789464, + 0.00705668656155467, + 0.01675282046198845, + 0.07621191442012787, + 0.11108424514532089, + 0.017732955515384674, + 0.014766409993171692, + 0.024819830432534218, + 0.02148709073662758, + 0.015384434722363949, + 0.013332312926650047, + 0.00982570182532072, + 0.010004663839936256, + 0.008160034194588661, + 0.008597112260758877, + 0.013739758171141148, + 0.006379828788340092, + 0.013147850520908833, + 0.010157397948205471, + 0.0020260964520275593, + 0.0017895092023536563, + 0.003005451988428831, + 0.0020531173795461655, + 0.0018686559051275253, + 0.0020949319005012512, + 0.0029561445116996765, + 0.001818064134567976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.059641629457473755 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029821347445249557 + ], + "opposite_logprob_tokens": [ + -3.5298213958740234 + ], + "logprob_reward": -0.029821347445249557, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029821347445249557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0422755628824234, + 0.09898350387811661, + 0.26041075587272644, + 0.02271304279565811, + 0.0555892288684845, + 0.017612334340810776, + 0.024433303624391556, + 0.3224709928035736, + 0.08942068368196487, + 0.06609060615301132 + ], + "credit_map": [ + 0.003513762727379799, + 0.006005868781358004, + 0.005822266452014446, + 0.006507092621177435, + 0.0030342834070324898, + 0.00814651045948267, + 0.0029764834325760603, + 0.006269296631217003, + 0.004306340590119362, + 0.005421456415206194, + 0.008371439762413502, + 0.004822782706469297, + 0.005702742375433445, + 0.030888454988598824, + 0.006739625707268715, + 0.03273066505789757, + 0.01118380669504404, + 0.016995051875710487, + 0.022151309996843338, + 0.03941706568002701, + 0.010074649006128311, + 0.06944949179887772, + 0.018061069771647453, + 0.07307832688093185, + 0.0021794727072119713, + 0.002592573408037424, + 0.003624425269663334, + 0.0025430701207369566, + 0.0020814817398786545, + 0.00311374687589705, + 0.0031554591841995716, + 0.0034228130243718624, + 0.004392100032418966, + 0.005398703273385763, + 0.006251925136893988, + 0.005873324815183878, + 0.013785303570330143, + 0.005604815669357777, + 0.007532033603638411, + 0.00675102137029171, + 0.0011364874662831426, + 0.002155833877623081, + 0.0019311055075377226, + 0.0018605744699016213, + 0.0028429359663277864, + 0.002085984218865633, + 0.0025500180199742317, + 0.003049395279958844, + 0.0018072231905534863, + 0.0028183013200759888, + 0.004386723507195711, + 0.0050665587186813354, + 0.0028253193013370037, + 0.0021636285819113255, + 0.0027564961928874254, + 0.0026090529281646013, + 0.09775480628013611, + 0.13804669678211212, + 0.016066204756498337, + 0.010785550810396671, + 0.02176797203719616, + 0.016629710793495178, + 0.012318011373281479, + 0.009102034382522106, + 0.009447808377444744, + 0.011357754468917847, + 0.010070481337606907, + 0.01343625783920288, + 0.013427191413939, + 0.007433793041855097, + 0.013501237146556377, + 0.010746156796813011, + 0.006815620232373476, + 0.006467562168836594, + 0.008321781642735004, + 0.006406702566891909, + 0.012059595435857773, + 0.005551365669816732, + 0.013155516237020493, + 0.00731246592476964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.029821347445249557 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043046385049819946 + ], + "opposite_logprob_tokens": [ + -3.168046474456787 + ], + "logprob_reward": -0.043046385049819946, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043046385049819946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044743213802576065, + 0.08147319406270981, + 0.2564023733139038, + 0.07379035651683807, + 0.09445030242204666, + 0.011791730299592018, + 0.021159028634428978, + 0.3804854452610016, + 0.02752961404621601, + 0.008174709044396877 + ], + "credit_map": [ + 0.004960528574883938, + 0.005980950314551592, + 0.005531967617571354, + 0.005194955971091986, + 0.005582575686275959, + 0.008269227109849453, + 0.004654224496334791, + 0.0045687793754041195, + 0.003917154390364885, + 0.0036769660655409098, + 0.007055916357785463, + 0.003960669972002506, + 0.003374887630343437, + 0.027063950896263123, + 0.005785019136965275, + 0.02663862705230713, + 0.010817212052643299, + 0.015961308032274246, + 0.021973542869091034, + 0.0365426242351532, + 0.008737221360206604, + 0.07064509391784668, + 0.01725725084543228, + 0.07446814328432083, + 0.006179600954055786, + 0.0070505039766430855, + 0.01049263309687376, + 0.006249443162232637, + 0.005978398025035858, + 0.011364241130650043, + 0.0077615417540073395, + 0.018713993951678276, + 0.007058639079332352, + 0.009165055118501186, + 0.018521543592214584, + 0.023884611204266548, + 0.00576083455234766, + 0.010500318370759487, + 0.00915525108575821, + 0.010404043830931187, + 0.000998757896013558, + 0.001797327771782875, + 0.001974349608644843, + 0.0015211020363494754, + 0.0016055632149800658, + 0.0012081984896212816, + 0.0010580047965049744, + 0.0016284271841868758, + 0.0018868783954530954, + 0.001543462392874062, + 0.005097749177366495, + 0.001993671990931034, + 0.001875161426141858, + 0.00291939964517951, + 0.002143510850146413, + 0.0036991944070905447, + 0.12494399398565292, + 0.17021693289279938, + 0.016182726249098778, + 0.010079690255224705, + 0.022144149988889694, + 0.01833166554570198, + 0.012439451180398464, + 0.006146849133074284, + 0.0033739861100912094, + 0.003818887984380126, + 0.0031211788300424814, + 0.003371851285919547, + 0.004161233082413673, + 0.0021251130383461714, + 0.004299643449485302, + 0.0032577195670455694, + 0.0010240933625027537, + 0.0010706580942496657, + 0.001217696233652532, + 0.000815916049759835, + 0.0010595006169751287, + 0.001097876112908125, + 0.001263933489099145, + 0.0006250348524190485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.043046385049819946 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.387020468711853 + ], + "opposite_logprob_tokens": [ + -1.137020468711853 + ], + "logprob_reward": -0.387020468711853, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.387020468711853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028523873537778854, + 0.09153082221746445, + 0.3722826838493347, + 0.0291712898761034, + 0.043636504560709, + 0.016186947003006935, + 0.07178999483585358, + 0.20917335152626038, + 0.12007860839366913, + 0.01762591488659382 + ], + "credit_map": [ + 0.0024285055696964264, + 0.0022550260182470083, + 0.0036765197291970253, + 0.006643297150731087, + 0.0031676413491368294, + 0.005503296386450529, + 0.002449718536809087, + 0.0023998674005270004, + 0.005773589946329594, + 0.005665893666446209, + 0.008308212272822857, + 0.0055254618637263775, + 0.008368792943656445, + 0.027597498148679733, + 0.007401926908642054, + 0.02288944646716118, + 0.016427818685770035, + 0.025282736867666245, + 0.032517727464437485, + 0.05749404802918434, + 0.013502350077033043, + 0.09782052785158157, + 0.026800882071256638, + 0.10243657231330872, + 0.0034418089780956507, + 0.0026515598874539137, + 0.004140406381338835, + 0.003470285562798381, + 0.0035121336113661528, + 0.003763738088309765, + 0.004849685356020927, + 0.003341672709211707, + 0.002683841623365879, + 0.004940008278936148, + 0.0045689307153224945, + 0.004162007011473179, + 0.009605677798390388, + 0.00445583276450634, + 0.007606379687786102, + 0.005613827146589756, + 0.0014252921100705862, + 0.0015752778854221106, + 0.0022724438458681107, + 0.002314821584150195, + 0.002856912324205041, + 0.0016120673390105367, + 0.0023563357535749674, + 0.0017737956950441003, + 0.005608795676380396, + 0.0068193161860108376, + 0.012214874848723412, + 0.013620574958622456, + 0.0070928530767560005, + 0.009744229726493359, + 0.008223599754273891, + 0.008465748280286789, + 0.05757470801472664, + 0.06848429888486862, + 0.014675352722406387, + 0.012309936806559563, + 0.020462919026613235, + 0.014653227292001247, + 0.010811482556164265, + 0.010201438330113888, + 0.013036004267632961, + 0.01548177283257246, + 0.010582514107227325, + 0.013975704088807106, + 0.022066568955779076, + 0.01051308959722519, + 0.01750718243420124, + 0.01691577583551407, + 0.0017870849696919322, + 0.00245812744833529, + 0.0029315894935280085, + 0.0021176408044993877, + 0.0022224450949579477, + 0.001862078090198338, + 0.002445054706186056, + 0.0018018946284428239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.387020468711853 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036153923720121384 + ], + "opposite_logprob_tokens": [ + -5.628615379333496 + ], + "logprob_reward": -0.0036153923720121384, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036153923720121384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03888958320021629, + 0.02824905700981617, + 0.6137266755104065, + 0.01137805450707674, + 0.01249855849891901, + 0.007167903706431389, + 0.028018219396471977, + 0.24090149998664856, + 0.01211674977093935, + 0.007053717505186796 + ], + "credit_map": [ + 0.004293718840926886, + 0.004315578378736973, + 0.0037511291448026896, + 0.006606745067983866, + 0.0024677871260792017, + 0.0045504397712647915, + 0.002882015658542514, + 0.010022171773016453, + 0.0016186691354960203, + 0.001236448180861771, + 0.002367813605815172, + 0.0013690474443137646, + 0.002524049486964941, + 0.008417278528213501, + 0.0019841950852423906, + 0.008731554262340069, + 0.026618709787726402, + 0.03986814245581627, + 0.05298563092947006, + 0.0891607329249382, + 0.02143179625272751, + 0.16540168225765228, + 0.042289428412914276, + 0.17597050964832306, + 0.0011165576288476586, + 0.0012383676366880536, + 0.0019513913430273533, + 0.0011053751222789288, + 0.001079898327589035, + 0.0013734967214986682, + 0.0012029071804136038, + 0.002310059731826186, + 0.0010543143143877387, + 0.0013232898199930787, + 0.0013005832443013787, + 0.002040314255282283, + 0.0011898435186594725, + 0.002020624466240406, + 0.0010107075795531273, + 0.002558881649747491, + 0.0006734582129865885, + 0.001113464473746717, + 0.001429267111234367, + 0.0014888133155182004, + 0.0007258811965584755, + 0.0005496969097293913, + 0.0005831252783536911, + 0.0006041968008503318, + 0.003018636954948306, + 0.0019520194036886096, + 0.007478044368326664, + 0.0036724682431668043, + 0.0019666384905576706, + 0.004887374117970467, + 0.0027605360373854637, + 0.002282503293827176, + 0.07890108972787857, + 0.10598647594451904, + 0.010173605754971504, + 0.0071615842171013355, + 0.013857024721801281, + 0.012780947610735893, + 0.007677054964005947, + 0.004363724961876869, + 0.0013189888559281826, + 0.0011914689093828201, + 0.0012765309074893594, + 0.0016021509654819965, + 0.0021519134752452374, + 0.0009533781558275223, + 0.0019564793910831213, + 0.0016658389940857887, + 0.0006352294585667551, + 0.0007384284399449825, + 0.0007858112221583724, + 0.0007298819255083799, + 0.0012353401398286223, + 0.00072817801265046, + 0.001518207835033536, + 0.0006826401222497225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.0036153923720121384 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, low-angle drone shot follows a knight as he gallops along the ruins of the ancient Great Wall. The knight is clad in a weathered, dark blue linen robe that flutters in the wind, revealing the hilt of a longsword slung across his back. His long, dark hair is simply tied up, a few strands whipped by the wind against his stern, angular face. He leans over and grips the reins, becoming one with the muscular, sweat-soaked dark brown steed beneath him. His hooves pound the massive, cracked, moss-covered stone bricks, sending up a cloud of yellow dust and rubble. The camera slowly zooms out, revealing the vast, majestic mountains in the background, silhouetted in varying shades against the golden sunset. The low-hanging sun casts long shadows on the city walls winding over the mountain ridge, giving the entire picture a warm and tragic tone. The dust in the air shimmers with golden particles in the light, creating a lonely, desolate and dynamic film atmosphere.", + "seed": 23809, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_15/index_3595_seed_23809_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.054051202395930885, + "vlm_logprob": -0.054051202395930885, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.054051202395930885, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07777148485183716 + ], + "opposite_logprob_tokens": [ + -2.8277714252471924 + ], + "logprob_reward": -0.07777148485183716, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07777148485183716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2837851643562317, + 0.02371053770184517, + 0.02866847813129425, + 0.03854871168732643, + 0.2793711721897125, + 0.01783900149166584, + 0.16733933985233307, + 0.028161754831671715, + 0.05693695694208145, + 0.07563891261816025 + ], + "credit_map": [ + 0.026917720213532448, + 0.0322449691593647, + 0.042422372847795486, + 0.02747204340994358, + 0.05051488056778908, + 0.04415149614214897, + 0.02561013773083687, + 0.034451521933078766, + 0.0017833064775913954, + 0.0020628164056688547, + 0.0035170672927051783, + 0.0019067268585786223, + 0.005831110756844282, + 0.0035064530093222857, + 0.003089340403676033, + 0.002013716846704483, + 0.0021883712615817785, + 0.003538923105224967, + 0.0028749587945640087, + 0.0031507869716733694, + 0.00529100326821208, + 0.003164669731631875, + 0.005084923002868891, + 0.0033748429268598557, + 0.0030170620884746313, + 0.005755310412496328, + 0.0057053579948842525, + 0.004270161036401987, + 0.005115097854286432, + 0.0071706622838974, + 0.003753197845071554, + 0.0037618649657815695, + 0.07335962355136871, + 0.007514093071222305, + 0.015032822266221046, + 0.006080154329538345, + 0.008801659569144249, + 0.10058087855577469, + 0.020083963871002197, + 0.04791796952486038, + 0.0015328936278820038, + 0.002387161599472165, + 0.0016798075521364808, + 0.003349337261170149, + 0.001747736125253141, + 0.0019285008311271667, + 0.003367335069924593, + 0.0018462302396073937, + 0.02178821712732315, + 0.009007254615426064, + 0.0056486609391868114, + 0.009601011872291565, + 0.012403429485857487, + 0.07045318931341171, + 0.03285756707191467, + 0.005579996854066849, + 0.0027785932179540396, + 0.001817243406549096, + 0.0037184241227805614, + 0.002386520616710186, + 0.002766267629340291, + 0.008264959789812565, + 0.0035833006259053946, + 0.002846446121111512, + 0.007863333448767662, + 0.019178258255124092, + 0.004676803946495056, + 0.0037557787727564573, + 0.005347603932023048, + 0.008827109821140766, + 0.004418854136019945, + 0.0028692150954157114, + 0.008210084401071072, + 0.02640731818974018, + 0.025355011224746704, + 0.0030373926274478436, + 0.0032346914522349834, + 0.003440475556999445, + 0.002942526713013649, + 0.003011415246874094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.07777148485183716 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867272078990936 + ], + "opposite_logprob_tokens": [ + -3.048672676086426 + ], + "logprob_reward": -0.04867272078990936, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867272078990936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05678282305598259, + 0.02915109694004059, + 0.04967689514160156, + 0.03857387229800224, + 0.33989933133125305, + 0.05680152401328087, + 0.1684228628873825, + 0.051652029156684875, + 0.051621437072753906, + 0.15741805732250214 + ], + "credit_map": [ + 0.006185863632708788, + 0.005931824445724487, + 0.009439881891012192, + 0.006951331626623869, + 0.0060218507423996925, + 0.008507943712174892, + 0.007731596473604441, + 0.006012539379298687, + 0.002890269970521331, + 0.004406778607517481, + 0.003521873150020838, + 0.002507025608792901, + 0.005823747720569372, + 0.004161702934652567, + 0.002980562159791589, + 0.002859142841771245, + 0.005663110408931971, + 0.005880584474653006, + 0.007994599640369415, + 0.004072918090969324, + 0.006173727568238974, + 0.010271502658724785, + 0.005108517594635487, + 0.0045119477435946465, + 0.003385871881619096, + 0.0051287938840687275, + 0.004279622342437506, + 0.00474325567483902, + 0.004092107526957989, + 0.006733676418662071, + 0.006123648025095463, + 0.0040869018994271755, + 0.09148412197828293, + 0.009540583938360214, + 0.017732419073581696, + 0.007620465476065874, + 0.011044183745980263, + 0.12332245707511902, + 0.02402035892009735, + 0.05513476952910423, + 0.006160931661725044, + 0.004978120792657137, + 0.005801262333989143, + 0.009603315033018589, + 0.006297872867435217, + 0.007232030387967825, + 0.009760414250195026, + 0.006967585068196058, + 0.02011299505829811, + 0.011451877653598785, + 0.006823821924626827, + 0.010723577812314034, + 0.013551726937294006, + 0.06530823558568954, + 0.0333855114877224, + 0.007065142970532179, + 0.0038409887347370386, + 0.005869980901479721, + 0.008252603001892567, + 0.004356649704277515, + 0.010865310207009315, + 0.010722446255385876, + 0.0038452877197414637, + 0.003898768685758114, + 0.00874831061810255, + 0.016855599358677864, + 0.004625899717211723, + 0.004635585471987724, + 0.00492085749283433, + 0.00578549737110734, + 0.0039717815816402435, + 0.0020779117476195097, + 0.01689823716878891, + 0.05458454415202141, + 0.05308465659618378, + 0.006637323647737503, + 0.0063823191449046135, + 0.0071567995473742485, + 0.006300721317529678, + 0.0063734776340425014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.04867272078990936 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062037304043769836 + ], + "opposite_logprob_tokens": [ + -2.812037229537964 + ], + "logprob_reward": -0.062037304043769836, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062037304043769836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08270736783742905, + 0.05857196822762489, + 0.07657309621572495, + 0.050340183079242706, + 0.42364898324012756, + 0.021630356088280678, + 0.12988708913326263, + 0.026469636708498, + 0.014903503470122814, + 0.11526776850223541 + ], + "credit_map": [ + 0.009588897228240967, + 0.009862501174211502, + 0.014499313198029995, + 0.009254354052245617, + 0.008456747978925705, + 0.012634710408747196, + 0.008882688358426094, + 0.009528149850666523, + 0.005830002017319202, + 0.009141284041106701, + 0.00984997022897005, + 0.00475257309153676, + 0.008116375654935837, + 0.008120565675199032, + 0.0059659178368747234, + 0.006795275956392288, + 0.012309510260820389, + 0.009258100762963295, + 0.009286269545555115, + 0.006108630448579788, + 0.009893672540783882, + 0.008714260533452034, + 0.011172673664987087, + 0.00982997752726078, + 0.004739128053188324, + 0.008519033901393414, + 0.006146489642560482, + 0.0042273965664207935, + 0.004523774608969688, + 0.012035725638270378, + 0.003954371903091669, + 0.006194262765347958, + 0.09868253767490387, + 0.011068078689277172, + 0.02265474572777748, + 0.0111314095556736, + 0.01374359056353569, + 0.14816583693027496, + 0.033107317984104156, + 0.08509548753499985, + 0.003256558906286955, + 0.002095371950417757, + 0.002461269497871399, + 0.003064297139644623, + 0.0029501791577786207, + 0.0024970455560833216, + 0.0028764093294739723, + 0.002429225016385317, + 0.01954020746052265, + 0.006447745021432638, + 0.006417027674615383, + 0.007572146598249674, + 0.010568281635642052, + 0.04978495463728905, + 0.02179339900612831, + 0.007763340137898922, + 0.002311943331733346, + 0.0031084183137863874, + 0.003969946876168251, + 0.003347310470417142, + 0.004234699532389641, + 0.004742613527923822, + 0.0023567306343466043, + 0.002397972857579589, + 0.0022264544386416674, + 0.001932622049935162, + 0.0017897471552714705, + 0.0019592305179685354, + 0.0017311444971710443, + 0.002461452502757311, + 0.0015612214338034391, + 0.0012416315730661154, + 0.012864099815487862, + 0.040823519229888916, + 0.0391400083899498, + 0.004456677008420229, + 0.0042714923620224, + 0.005602887831628323, + 0.003790590912103653, + 0.0043184952810406685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.062037304043769836 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895738631486893 + ], + "opposite_logprob_tokens": [ + -2.5789573192596436 + ], + "logprob_reward": -0.07895738631486893, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895738631486893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16246779263019562, + 0.09022925794124603, + 0.03770078718662262, + 0.07923027873039246, + 0.4454323351383209, + 0.0188258346170187, + 0.09959664940834045, + 0.019183849915862083, + 0.0313958078622818, + 0.015937428921461105 + ], + "credit_map": [ + 0.017012661322951317, + 0.019361883401870728, + 0.023856429383158684, + 0.016789430752396584, + 0.029701843857765198, + 0.02530892752110958, + 0.012952489778399467, + 0.01748412288725376, + 0.009230461902916431, + 0.011494136415421963, + 0.01418543141335249, + 0.00797832291573286, + 0.015656478703022003, + 0.014469503425061703, + 0.00974978692829609, + 0.007465135306119919, + 0.003290127497166395, + 0.003074054140597582, + 0.0048110513016581535, + 0.004304382484406233, + 0.006582471542060375, + 0.006382949184626341, + 0.00542727205902338, + 0.003828477580100298, + 0.005355394445359707, + 0.009888224303722382, + 0.01046085637062788, + 0.011685502715408802, + 0.008832592517137527, + 0.01700923964381218, + 0.00847018975764513, + 0.007528282701969147, + 0.11298297345638275, + 0.011466673575341702, + 0.023931782692670822, + 0.009625962935388088, + 0.013954130001366138, + 0.15946701169013977, + 0.03288114815950394, + 0.08112264424562454, + 0.001500518061220646, + 0.0027081314474344254, + 0.002124815946444869, + 0.0031689347233623266, + 0.0020409408025443554, + 0.0023372264113277197, + 0.002821612171828747, + 0.0021236541215330362, + 0.009238318540155888, + 0.007345976773649454, + 0.004992413334548473, + 0.0057526035234332085, + 0.0073328264988958836, + 0.037033457309007645, + 0.023358481004834175, + 0.004542575683444738, + 0.002503576222807169, + 0.0011531159980222583, + 0.0027723771054297686, + 0.0012199798366054893, + 0.0014356335159391165, + 0.006642933934926987, + 0.0017499526729807258, + 0.0017062799306586385, + 0.004409409128129482, + 0.010057761333882809, + 0.0024604222271591425, + 0.0024523946922272444, + 0.00290170987136662, + 0.0053106495179235935, + 0.002316450234502554, + 0.0014870127197355032, + 0.0017338587203994393, + 0.004861136432737112, + 0.004352685995399952, + 0.0011764495866373181, + 0.001047878758981824, + 0.0010258888360112906, + 0.0008272404666058719, + 0.0009122900082729757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.07895738631486893 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002817115979269147 + ], + "opposite_logprob_tokens": [ + -5.877817153930664 + ], + "logprob_reward": -0.002817115979269147, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002817115979269147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12055599689483643, + 0.012645479291677475, + 0.0246439091861248, + 0.01514438632875681, + 0.20219382643699646, + 0.02861252799630165, + 0.4658690392971039, + 0.033872172236442566, + 0.04348747432231903, + 0.052975207567214966 + ], + "credit_map": [ + 0.009851285256445408, + 0.015408463776111603, + 0.0158099178224802, + 0.01338031142950058, + 0.019657963886857033, + 0.016239212825894356, + 0.013208518736064434, + 0.017000321298837662, + 0.0016550117870792747, + 0.0017140523996204138, + 0.001936084940098226, + 0.0016489429399371147, + 0.0011861218372359872, + 0.0015935500850901008, + 0.0013426324585452676, + 0.0015690827276557684, + 0.0031431936658918858, + 0.00270820246078074, + 0.0036731893196702003, + 0.0022038945462554693, + 0.003641402116045356, + 0.0036735415924340487, + 0.0022936027962714434, + 0.0033068829216063023, + 0.0013321812730282545, + 0.0021736742928624153, + 0.00163063604850322, + 0.0018507558852434158, + 0.002637363038957119, + 0.0024052050430327654, + 0.001685727620497346, + 0.0014288430102169514, + 0.049506742507219315, + 0.005322667770087719, + 0.010830866172909737, + 0.005028171464800835, + 0.006087512709200382, + 0.07192697376012802, + 0.015189051628112793, + 0.038301847875118256, + 0.0034868076909333467, + 0.003633700544014573, + 0.0023116469383239746, + 0.0035438283812254667, + 0.00362573005259037, + 0.005201215855777264, + 0.0041053760796785355, + 0.002704222686588764, + 0.060464583337306976, + 0.01565130613744259, + 0.016116658225655556, + 0.023640567436814308, + 0.0338682159781456, + 0.20178493857383728, + 0.09979673475027084, + 0.014546060934662819, + 0.004648747853934765, + 0.00314133008942008, + 0.004500535316765308, + 0.0022095819003880024, + 0.0031663815025240183, + 0.011242700740695, + 0.002603273605927825, + 0.002359619364142418, + 0.006941916421055794, + 0.015697449445724487, + 0.004275158047676086, + 0.0046539586037397385, + 0.0027908210176974535, + 0.0040945508517324924, + 0.0030836258083581924, + 0.0019499956397339702, + 0.005541210528463125, + 0.01690496690571308, + 0.018075644969940186, + 0.00229020812548697, + 0.002199379727244377, + 0.0032233339734375477, + 0.0025188191793859005, + 0.002221645088866353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.002817115979269147 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An epic, low-angle drone shot follows a knight as he gallops along the ruins of the ancient Great Wall. The knight is clad in a weathered, dark blue linen robe that flutters in the wind, revealing the hilt of a longsword slung across his back. His long, dark hair is simply tied up, a few strands whipped by the wind against his stern, angular face. He leans over and grips the reins, becoming one with the muscular, sweat-soaked dark brown steed beneath him. His hooves pound the massive, cracked, moss-covered stone bricks, sending up a cloud of yellow dust and rubble. The camera slowly zooms out, revealing the vast, majestic mountains in the background, silhouetted in varying shades against the golden sunset. The low-hanging sun casts long shadows on the city walls winding over the mountain ridge, giving the entire picture a warm and tragic tone. The dust in the air shimmers with golden particles in the light, creating a lonely, desolate and dynamic film atmosphere.", + "seed": 23810, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_15/index_3595_seed_23810_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11416917075403035, + "vlm_logprob": -0.11416917075403035, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.11416917075403035, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05832834541797638 + ], + "opposite_logprob_tokens": [ + -3.05832839012146 + ], + "logprob_reward": -0.05832834541797638, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05832834541797638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1939823478460312, + 0.07458895444869995, + 0.15417097508907318, + 0.10952438414096832, + 0.11914340406656265, + 0.052680470049381256, + 0.0803973451256752, + 0.10852383822202682, + 0.045568108558654785, + 0.06142013892531395 + ], + "credit_map": [ + 0.024268988519906998, + 0.018708258867263794, + 0.018983885645866394, + 0.033452171832323074, + 0.029521901160478592, + 0.013522450812160969, + 0.03371244668960571, + 0.021812237799167633, + 0.006193241570144892, + 0.006349551025778055, + 0.0070057157427072525, + 0.012569371610879898, + 0.011651918292045593, + 0.008983835577964783, + 0.010633577592670918, + 0.011201747693121433, + 0.013990313746035099, + 0.03205854445695877, + 0.01155552826821804, + 0.019152944907546043, + 0.02819114178419113, + 0.008841446600854397, + 0.019496578723192215, + 0.02088446356356144, + 0.007471758872270584, + 0.019237224012613297, + 0.013755704276263714, + 0.011131848208606243, + 0.02154429629445076, + 0.011504347436130047, + 0.011637235060334206, + 0.013241971842944622, + 0.018980322405695915, + 0.020063787698745728, + 0.01461571641266346, + 0.012942010536789894, + 0.017648739740252495, + 0.013932265341281891, + 0.011795743368566036, + 0.009164811111986637, + 0.00713384198024869, + 0.00598515011370182, + 0.00746497604995966, + 0.005288776010274887, + 0.01106498297303915, + 0.0038998660165816545, + 0.00802314467728138, + 0.003819734090939164, + 0.007901850156486034, + 0.014578375034034252, + 0.01008661463856697, + 0.00854822900146246, + 0.009285318665206432, + 0.014706340618431568, + 0.006533108651638031, + 0.008757510222494602, + 0.015092028304934502, + 0.033321015536785126, + 0.014774901792407036, + 0.01178671047091484, + 0.008187757804989815, + 0.011666093952953815, + 0.007658347021788359, + 0.006036991253495216, + 0.005649890284985304, + 0.004787134937942028, + 0.005753658711910248, + 0.004399345722049475, + 0.005991257727146149, + 0.005165936890989542, + 0.007561155594885349, + 0.0062597268261015415, + 0.006407896988093853, + 0.007307086605578661, + 0.007507562171667814, + 0.0065310331992805, + 0.008677150122821331, + 0.007006453350186348, + 0.01095972303301096, + 0.00702323205769062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.05832834541797638 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308418184518814 + ], + "opposite_logprob_tokens": [ + -3.168084144592285 + ], + "logprob_reward": -0.04308418184518814, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308418184518814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2092359960079193, + 0.06216935068368912, + 0.13146521151065826, + 0.05780794844031334, + 0.05881022661924362, + 0.039937201887369156, + 0.05070175975561142, + 0.20467154681682587, + 0.06180128827691078, + 0.12339948862791061 + ], + "credit_map": [ + 0.018510399386286736, + 0.020158953964710236, + 0.016586629673838615, + 0.03490239009261131, + 0.0507861003279686, + 0.01649787463247776, + 0.024551771581172943, + 0.027241870760917664, + 0.004437415860593319, + 0.005798223428428173, + 0.004121069330722094, + 0.013847891241312027, + 0.009015564806759357, + 0.005888573359698057, + 0.01078264880925417, + 0.008277961984276772, + 0.010335917584598064, + 0.014286134392023087, + 0.01066008023917675, + 0.02275005169212818, + 0.021703237667679787, + 0.012399449944496155, + 0.02224656008183956, + 0.017083782702684402, + 0.004135232884436846, + 0.008770777843892574, + 0.008825820870697498, + 0.005049620755016804, + 0.010542045347392559, + 0.007075964938849211, + 0.005224756896495819, + 0.008183727040886879, + 0.005801420658826828, + 0.00834786519408226, + 0.005445271264761686, + 0.007291783113032579, + 0.01076956745237112, + 0.005377116613090038, + 0.01030611153692007, + 0.00547108706086874, + 0.006063584238290787, + 0.006068904884159565, + 0.005908098537474871, + 0.00319558079354465, + 0.00909251719713211, + 0.003710344200953841, + 0.003749685827642679, + 0.0021484876051545143, + 0.005077132489532232, + 0.009639216586947441, + 0.0058546108193695545, + 0.005379640031605959, + 0.006960046477615833, + 0.0063807908445596695, + 0.005459153093397617, + 0.005951169412583113, + 0.025325486436486244, + 0.03241217881441116, + 0.0431542843580246, + 0.018884574994444847, + 0.029096180573105812, + 0.02279401756823063, + 0.015973828732967377, + 0.017031002789735794, + 0.009930342435836792, + 0.007633585017174482, + 0.009946768172085285, + 0.006873372942209244, + 0.005524280481040478, + 0.007926583290100098, + 0.007699525915086269, + 0.006266830023378134, + 0.01471920870244503, + 0.015951508656144142, + 0.015092071145772934, + 0.012249634601175785, + 0.015932651236653328, + 0.022004641592502594, + 0.017756912857294083, + 0.009692863561213017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.04308418184518814 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808846324682236 + ], + "opposite_logprob_tokens": [ + -3.28808856010437 + ], + "logprob_reward": -0.03808846324682236, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808846324682236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5181085467338562, + 0.06726135313510895, + 0.03989243507385254, + 0.021120566874742508, + 0.045793116092681885, + 0.05372767522931099, + 0.1050160825252533, + 0.07101331651210785, + 0.05898946151137352, + 0.01907745935022831 + ], + "credit_map": [ + 0.049547817558050156, + 0.06508192420005798, + 0.035733018070459366, + 0.10889630019664764, + 0.08262045681476593, + 0.042364586144685745, + 0.06957070529460907, + 0.06429373472929001, + 0.00685923034325242, + 0.00753195071592927, + 0.0062269726768136024, + 0.011815309524536133, + 0.01179733406752348, + 0.008680793456733227, + 0.006663964595645666, + 0.0076857940293848515, + 0.0038091507740318775, + 0.007146643009036779, + 0.003465728834271431, + 0.0038285700138658285, + 0.0073769777081906796, + 0.004095214419066906, + 0.0029081229586154222, + 0.007262029219418764, + 0.001892448985017836, + 0.0032704025506973267, + 0.0023909020237624645, + 0.002007789444178343, + 0.0042662532068789005, + 0.003176855156198144, + 0.0015299419173970819, + 0.002585973823443055, + 0.00735050905495882, + 0.007900013588368893, + 0.005708788987249136, + 0.0036722547374665737, + 0.006173658184707165, + 0.005403601098805666, + 0.004346467088907957, + 0.005237823352217674, + 0.005006253719329834, + 0.007045416627079248, + 0.007781577296555042, + 0.006755520589649677, + 0.011617365293204784, + 0.004869959782809019, + 0.006421741098165512, + 0.004229835234582424, + 0.010953869670629501, + 0.011419425718486309, + 0.019505886361002922, + 0.009136837907135487, + 0.015653830021619797, + 0.01754792220890522, + 0.007466489914804697, + 0.01333182118833065, + 0.009316651150584221, + 0.023922715336084366, + 0.015424872748553753, + 0.0056866914965212345, + 0.004621190018951893, + 0.004742366727441549, + 0.0037169791758060455, + 0.0035818531177937984, + 0.005741622298955917, + 0.00652813958004117, + 0.008459254167973995, + 0.006969289388507605, + 0.0074307662434875965, + 0.011800318025052547, + 0.0066284374333918095, + 0.005431631114333868, + 0.001949600176885724, + 0.0018952187383547425, + 0.0017027892172336578, + 0.00202659098431468, + 0.0025786554906517267, + 0.003545178798958659, + 0.003172613214701414, + 0.0022068126127123833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.03808846324682236 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42885470390319824 + ], + "opposite_logprob_tokens": [ + -1.0538547039031982 + ], + "logprob_reward": -0.42885470390319824, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42885470390319824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22066742181777954, + 0.0500732883810997, + 0.05695711076259613, + 0.031150618568062782, + 0.025035949423909187, + 0.027861418202519417, + 0.1783631145954132, + 0.313400536775589, + 0.019935080781579018, + 0.07655546814203262 + ], + "credit_map": [ + 0.021758902817964554, + 0.026866033673286438, + 0.016073256731033325, + 0.04974936321377754, + 0.0308430977165699, + 0.01875217817723751, + 0.030656542629003525, + 0.025968046858906746, + 0.003958233166486025, + 0.004681915510445833, + 0.004439963959157467, + 0.009300696663558483, + 0.008090304210782051, + 0.0056313760578632355, + 0.0069634923711419106, + 0.0070073045790195465, + 0.004892448894679546, + 0.009557154029607773, + 0.0036702912766486406, + 0.00700059998780489, + 0.012774928472936153, + 0.0041695861145854, + 0.007395884022116661, + 0.007496212143450975, + 0.0033805386628955603, + 0.003990905359387398, + 0.004957594908773899, + 0.0032206461764872074, + 0.00471948366612196, + 0.003366870107129216, + 0.0033887706231325865, + 0.004125809296965599, + 0.002422669669613242, + 0.0035910094156861305, + 0.003187449648976326, + 0.003017412731423974, + 0.004447379615157843, + 0.0020420646760612726, + 0.004180519841611385, + 0.002147443126887083, + 0.0030195799190551043, + 0.0034682215191423893, + 0.004017656669020653, + 0.004110363312065601, + 0.004367928020656109, + 0.0037488483358174562, + 0.002684357576072216, + 0.0024444619193673134, + 0.01908191293478012, + 0.023907959461212158, + 0.028849760070443153, + 0.014916960150003433, + 0.028133779764175415, + 0.02692926488816738, + 0.012223361060023308, + 0.024320119991898537, + 0.03417488560080528, + 0.08989217877388, + 0.0804443210363388, + 0.025892820209264755, + 0.026263495907187462, + 0.022279877215623856, + 0.01682998426258564, + 0.017622996121644974, + 0.002287298906594515, + 0.003213905030861497, + 0.0020790286362171173, + 0.0021128342486917973, + 0.003295533824712038, + 0.0029380314517766237, + 0.0022115104366093874, + 0.0017969394102692604, + 0.008208622224628925, + 0.00894564762711525, + 0.008814826607704163, + 0.0074530295096337795, + 0.01287307869642973, + 0.010094905272126198, + 0.012978353537619114, + 0.007186997216194868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.42885470390319824 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a knight riding a dark brown steed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the knight wearing a weathered dark blue linen robe with a longsword hilt visible on his back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting the ruins of the ancient Great Wall with massive cracked moss-covered stone bricks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a low-angle shot and slowly zoom out to reveal mountains in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a golden sunset casting long shadows and creating yellow dust clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024901593569666147 + ], + "opposite_logprob_tokens": [ + -6.002490043640137 + ], + "logprob_reward": -0.0024901593569666147, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024901593569666147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3732089400291443, + 0.1629115641117096, + 0.07207053899765015, + 0.029606174677610397, + 0.06623177975416183, + 0.06395557522773743, + 0.09903188049793243, + 0.06595051288604736, + 0.028577109798789024, + 0.03845600783824921 + ], + "credit_map": [ + 0.04076094925403595, + 0.05014561489224434, + 0.027825066819787025, + 0.08581086993217468, + 0.044140756130218506, + 0.034925904124975204, + 0.04361419007182121, + 0.04598555713891983, + 0.015446272678673267, + 0.02268318273127079, + 0.015830237418413162, + 0.028751976788043976, + 0.02232697792351246, + 0.01718359999358654, + 0.01790643110871315, + 0.022782858461141586, + 0.007692660205066204, + 0.0171134565025568, + 0.00613637687638402, + 0.006603545509278774, + 0.01162461657077074, + 0.008174550719559193, + 0.004851450677961111, + 0.009873872622847557, + 0.002101547084748745, + 0.005061401054263115, + 0.0025932269636541605, + 0.003216928569599986, + 0.006166187115013599, + 0.0025837684515863657, + 0.003965170122683048, + 0.003917939495295286, + 0.009076699614524841, + 0.010084523819386959, + 0.008231605403125286, + 0.005836882628500462, + 0.00898189377039671, + 0.0071725198067724705, + 0.008441835641860962, + 0.008405812084674835, + 0.009274713695049286, + 0.00877339020371437, + 0.011050206609070301, + 0.006402334664016962, + 0.011008924804627895, + 0.0077040051110088825, + 0.005309583153575659, + 0.004432414658367634, + 0.009469876065850258, + 0.010774180293083191, + 0.016509678214788437, + 0.009028424508869648, + 0.015068828128278255, + 0.015176527202129364, + 0.007617518771439791, + 0.015386836603283882, + 0.00860342476516962, + 0.008889353834092617, + 0.010625804774463177, + 0.007560699246823788, + 0.007147348485887051, + 0.00909467414021492, + 0.006479383911937475, + 0.007549817208200693, + 0.0032258599530905485, + 0.0028325060848146677, + 0.003380205947905779, + 0.003915911540389061, + 0.004883014131337404, + 0.003864648286253214, + 0.0037463156040757895, + 0.0027286470867693424, + 0.003488296875730157, + 0.004148378968238831, + 0.005520558916032314, + 0.005582356825470924, + 0.004557537846267223, + 0.004246129654347897, + 0.0066376072354614735, + 0.004275136161595583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0024901593569666147 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_2/index_tensor([7201, 7201])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_2/index_tensor([7201, 7201])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..751f5ec361cc599d10526123c9cff175acda7b32 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_2/index_tensor([7201, 7201])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up, shot from a slightly tilted angle, the camera remains still, focusing on a red apple suspended from a gnarled, mossy branch. The apple, perfectly spherical and with a rich, vibrant crimson skin, is dotted with shimmering dewdrops on its smooth surface, each one acting like a miniature magnifying glass, reflecting the surrounding light. As an invisible breeze sweeps by, the apple, its short, firm, dark brown stalk, begins to sway rhythmically and gently. This movement causes the dewdrops on its surface to roll gracefully and slowly, gliding along the apple's smooth curves. Some coalesce into larger droplets, shimmering in their movement. The first soft rays of morning sunlight, filtering through the blurred emerald green leaves in the background, strike the apple, casting a warm golden highlight on one side while casting the other in soft shadow. Each rolling dewdrop captures and shimmers captivatingly, creating an atmosphere of profound tranquility, freshness, and vitality.", + "seed": 22011, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_2/index_7201_seed_22011_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19767069146037103, + "vlm_logprob": -0.19767069146037103, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.19767069146037103, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012908587232232094 + ], + "opposite_logprob_tokens": [ + -4.887908458709717 + ], + "logprob_reward": -0.012908587232232094, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.012908587232232094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12326208502054214, + 0.10206015408039093, + 0.2935258448123932, + 0.0716250091791153, + 0.042384278029203415, + 0.057665709406137466, + 0.0794965922832489, + 0.07718884199857712, + 0.08110866695642471, + 0.07168282568454742 + ], + "credit_map": [ + 0.015422841534018517, + 0.008296959102153778, + 0.0167341697961092, + 0.010581796988844872, + 0.011189360171556473, + 0.011667333543300629, + 0.02831672504544258, + 0.021052900701761246, + 0.01458656694740057, + 0.006033666431903839, + 0.02986183948814869, + 0.008923218585550785, + 0.0071486239321529865, + 0.006709038745611906, + 0.006540107075124979, + 0.02225709706544876, + 0.0227134320884943, + 0.009515534155070782, + 0.11022855341434479, + 0.017199192196130753, + 0.01232756394892931, + 0.012486467137932777, + 0.016188548877835274, + 0.09286656230688095, + 0.006831501144915819, + 0.00783360842615366, + 0.007891901768743992, + 0.014726595021784306, + 0.009309792891144753, + 0.00933416560292244, + 0.006688535213470459, + 0.00900890864431858, + 0.005018226802349091, + 0.005352084059268236, + 0.008046679198741913, + 0.006303077097982168, + 0.004868551157414913, + 0.003526112297549844, + 0.003750920295715332, + 0.005518628749996424, + 0.006659429986029863, + 0.0043775117956101894, + 0.009745806455612183, + 0.00640785600990057, + 0.007545590866357088, + 0.010592669248580933, + 0.0043656351044774055, + 0.007971207611262798, + 0.01026237104088068, + 0.010412095114588737, + 0.012209358625113964, + 0.016161136329174042, + 0.004485235549509525, + 0.004465751349925995, + 0.00902886688709259, + 0.012471779249608517, + 0.007330626714974642, + 0.008055477403104305, + 0.011800573207437992, + 0.01526830904185772, + 0.00655263289809227, + 0.007462791167199612, + 0.008144640363752842, + 0.012573792599141598, + 0.0048363590613007545, + 0.0036760626826435328, + 0.028110774233937263, + 0.005202957894653082, + 0.003508994821459055, + 0.0033525885082781315, + 0.004610896110534668, + 0.02781003713607788, + 0.010835100896656513, + 0.005333007778972387, + 0.017193956300616264, + 0.007625999394804239, + 0.005381423979997635, + 0.005579011980444193, + 0.006810341961681843, + 0.012923981994390488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.012908587232232094 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158348977565765 + ], + "opposite_logprob_tokens": [ + -1.7015835046768188 + ], + "logprob_reward": -0.20158348977565765, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158348977565765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0907059907913208, + 0.18154551088809967, + 0.07429574429988861, + 0.08171997219324112, + 0.09168073534965515, + 0.03372771665453911, + 0.17458388209342957, + 0.044632453471422195, + 0.2045740783214569, + 0.022533847019076347 + ], + "credit_map": [ + 0.012480201199650764, + 0.007651044987142086, + 0.011941362172365189, + 0.012460465542972088, + 0.010217862203717232, + 0.013188790529966354, + 0.00960641447454691, + 0.01315984781831503, + 0.014112289063632488, + 0.00750076724216342, + 0.06753519177436829, + 0.010388602502644062, + 0.007607149425894022, + 0.006993903778493404, + 0.009111873805522919, + 0.058295730501413345, + 0.00943258311599493, + 0.008721806108951569, + 0.011418240144848824, + 0.014523345045745373, + 0.006689967587590218, + 0.004625850822776556, + 0.010169538669288158, + 0.008714410476386547, + 0.008182273246347904, + 0.011111869476735592, + 0.009381670504808426, + 0.01865151710808277, + 0.007063445635139942, + 0.007790814153850079, + 0.009426464326679707, + 0.010111922398209572, + 0.012012967839837074, + 0.010580685921013355, + 0.016227491199970245, + 0.012632408179342747, + 0.01055690087378025, + 0.008376489393413067, + 0.009797262027859688, + 0.01149652898311615, + 0.004674138501286507, + 0.0028470067773014307, + 0.0048865885473787785, + 0.003376961452886462, + 0.004983060993254185, + 0.0054235574789345264, + 0.0027507354971021414, + 0.004785670433193445, + 0.01224344503134489, + 0.011070561595261097, + 0.05898062139749527, + 0.016968173906207085, + 0.006090016569942236, + 0.00715684425085783, + 0.010505281388759613, + 0.051568951457738876, + 0.004477942828088999, + 0.005677740089595318, + 0.0049521056935191154, + 0.009597948752343655, + 0.00391934160143137, + 0.00412413477897644, + 0.00654412480071187, + 0.005339112598448992, + 0.016222428530454636, + 0.0062660654075443745, + 0.08346539735794067, + 0.007054313085973263, + 0.005005059763789177, + 0.007090386003255844, + 0.007043996825814247, + 0.07242643088102341, + 0.002872215351089835, + 0.002338224556297064, + 0.004215519409626722, + 0.0030483687296509743, + 0.0020927691366523504, + 0.002060249913483858, + 0.0025403816252946854, + 0.003366118762642145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.20158348977565765 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813691198825836 + ], + "opposite_logprob_tokens": [ + -1.4063690900802612 + ], + "logprob_reward": -0.2813691198825836, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813691198825836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16919656097888947, + 0.07945716381072998, + 0.2923707664012909, + 0.0781845822930336, + 0.03513603284955025, + 0.033317599445581436, + 0.11961283534765244, + 0.14180055260658264, + 0.02923194132745266, + 0.021691981703042984 + ], + "credit_map": [ + 0.013554122298955917, + 0.009135699830949306, + 0.04087989404797554, + 0.013060905039310455, + 0.0116175701841712, + 0.012265903875231743, + 0.021966872736811638, + 0.04671558365225792, + 0.01003855187445879, + 0.006396353244781494, + 0.015946850180625916, + 0.013251112774014473, + 0.00397511338815093, + 0.005588421132415533, + 0.008694812655448914, + 0.015565941110253334, + 0.02470661886036396, + 0.012589989230036736, + 0.10182764381170273, + 0.024377597495913506, + 0.011702420189976692, + 0.010178604163229465, + 0.017881667241454124, + 0.08910621702671051, + 0.008083903230726719, + 0.010242516174912453, + 0.00978990737348795, + 0.01997552067041397, + 0.005382586736232042, + 0.006081951782107353, + 0.00885101966559887, + 0.009777172468602657, + 0.004400994628667831, + 0.004195020534098148, + 0.005903573706746101, + 0.005304877646267414, + 0.003615163266658783, + 0.0032880615908652544, + 0.004151876550167799, + 0.004276466555893421, + 0.005167634226381779, + 0.002783380914479494, + 0.004870404023677111, + 0.003769480623304844, + 0.004434909205883741, + 0.004839862231165171, + 0.0025465372018516064, + 0.004905392415821552, + 0.007053349167108536, + 0.011591058224439621, + 0.033059991896152496, + 0.017639288678765297, + 0.004788972437381744, + 0.006120475009083748, + 0.008875601924955845, + 0.03048408403992653, + 0.013883642852306366, + 0.017377756536006927, + 0.01772404834628105, + 0.03120378591120243, + 0.012367673218250275, + 0.011699647642672062, + 0.016144149005413055, + 0.021399840712547302, + 0.0025474363937973976, + 0.0012693903408944607, + 0.010533724911510944, + 0.0013432196574285626, + 0.0017449657898396254, + 0.0019572121091187, + 0.0013969714054837823, + 0.008439022116363049, + 0.002709172200411558, + 0.0020666467025876045, + 0.004407347179949284, + 0.0022055013105273247, + 0.0018685274990275502, + 0.0023053770419210196, + 0.0017179421847686172, + 0.00441146781668067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.2813691198825836 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743040204048157 + ], + "opposite_logprob_tokens": [ + -0.9743040204048157 + ], + "logprob_reward": -0.4743040204048157, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743040204048157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16713643074035645, + 0.10683608800172806, + 0.0958009734749794, + 0.1487218290567398, + 0.16729243099689484, + 0.08762847632169724, + 0.05233187600970268, + 0.06083786487579346, + 0.06292104721069336, + 0.05049298703670502 + ], + "credit_map": [ + 0.010900428518652916, + 0.008469300344586372, + 0.04973654821515083, + 0.010111968964338303, + 0.009222482331097126, + 0.010503732599318027, + 0.015179533511400223, + 0.053012438118457794, + 0.009409007616341114, + 0.006312226410955191, + 0.030736785382032394, + 0.01628842204809189, + 0.005524476524442434, + 0.005210299976170063, + 0.008020770736038685, + 0.025334101170301437, + 0.018060822039842606, + 0.010260627605021, + 0.01413874514400959, + 0.01477177906781435, + 0.011023048311471939, + 0.0065732249058783054, + 0.010663773864507675, + 0.010308957658708096, + 0.013728435151278973, + 0.02031349390745163, + 0.01632755808532238, + 0.03954887390136719, + 0.012479083612561226, + 0.010412299074232578, + 0.017075367271900177, + 0.018836719915270805, + 0.017849018797278404, + 0.019894205033779144, + 0.031507283449172974, + 0.0334501639008522, + 0.014386696740984917, + 0.010414022020995617, + 0.01608205959200859, + 0.02370898239314556, + 0.011632008478045464, + 0.009440477006137371, + 0.012791155837476254, + 0.014951338991522789, + 0.008785969577729702, + 0.009096387773752213, + 0.010376013815402985, + 0.010555125772953033, + 0.004654414486140013, + 0.004477147478610277, + 0.015029619447886944, + 0.007307273801416159, + 0.0022786841727793217, + 0.002620596904307604, + 0.003975306171923876, + 0.011988832615315914, + 0.009272739291191101, + 0.009108300320804119, + 0.008608280681073666, + 0.010322557762265205, + 0.0027520719449967146, + 0.003585587255656719, + 0.008417009375989437, + 0.008771320804953575, + 0.0062011429108679295, + 0.003423700574785471, + 0.02359086275100708, + 0.003831814508885145, + 0.001977574545890093, + 0.0021740386728197336, + 0.003641989314928651, + 0.01807992532849312, + 0.0047616539523005486, + 0.005205164197832346, + 0.0116965863853693, + 0.007354192901402712, + 0.0024741424713283777, + 0.0025405886117368937, + 0.0050931088626384735, + 0.011367550119757652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.4743040204048157 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018188240006566048 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.018188240006566048, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018188240006566048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2240651398897171, + 0.10090311616659164, + 0.2856975197792053, + 0.037843480706214905, + 0.036452338099479675, + 0.06078539788722992, + 0.04981150105595589, + 0.07724583894014359, + 0.10161702334880829, + 0.025578662753105164 + ], + "credit_map": [ + 0.032973237335681915, + 0.010958361439406872, + 0.032860200852155685, + 0.0167393796145916, + 0.018439356237649918, + 0.010773630812764168, + 0.06712164729833603, + 0.03419933840632439, + 0.015153352171182632, + 0.0077997129410505295, + 0.022350553423166275, + 0.013134697452187538, + 0.007109054829925299, + 0.006274528335779905, + 0.01169044803828001, + 0.017390767112374306, + 0.02270856313407421, + 0.007521760184317827, + 0.11670473217964172, + 0.009768581949174404, + 0.005992763675749302, + 0.010069613344967365, + 0.014880181290209293, + 0.09805132448673248, + 0.004064404871314764, + 0.003338329493999481, + 0.004958915524184704, + 0.004397188778966665, + 0.00625145249068737, + 0.0059972782619297504, + 0.00389447552151978, + 0.004941433668136597, + 0.0057566165924072266, + 0.003678098786622286, + 0.007252243347465992, + 0.004929279442876577, + 0.003285181475803256, + 0.002960850251838565, + 0.00312302028760314, + 0.005467045586556196, + 0.011039366945624352, + 0.004353016149252653, + 0.01295521017163992, + 0.006596297491341829, + 0.006298167631030083, + 0.0069674039259552956, + 0.004462097305804491, + 0.00811383780092001, + 0.006743902340531349, + 0.0030753901228308678, + 0.014205592684447765, + 0.004668897483497858, + 0.0025921654887497425, + 0.0044975923374295235, + 0.0037113674916327, + 0.010316591709852219, + 0.01182447001338005, + 0.004537546541541815, + 0.013741792179644108, + 0.0061016809195280075, + 0.007878845557570457, + 0.009037457406520844, + 0.0070150261744856834, + 0.017109012231230736, + 0.008520757779479027, + 0.002583214081823826, + 0.04066065326333046, + 0.0038565390277653933, + 0.002675593364983797, + 0.004230000078678131, + 0.003961376380175352, + 0.035128895193338394, + 0.0037370468489825726, + 0.0024094018153846264, + 0.005282619968056679, + 0.0026902309618890285, + 0.001612229971215129, + 0.0017915175994858146, + 0.00277667841874063, + 0.005278938449919224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.018188240006566048 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, shot from a slightly tilted angle, the camera remains still, focusing on a red apple suspended from a gnarled, mossy branch. The apple, perfectly spherical and with a rich, vibrant crimson skin, is dotted with shimmering dewdrops on its smooth surface, each one acting like a miniature magnifying glass, reflecting the surrounding light. As an invisible breeze sweeps by, the apple, its short, firm, dark brown stalk, begins to sway rhythmically and gently. This movement causes the dewdrops on its surface to roll gracefully and slowly, gliding along the apple's smooth curves. Some coalesce into larger droplets, shimmering in their movement. The first soft rays of morning sunlight, filtering through the blurred emerald green leaves in the background, strike the apple, casting a warm golden highlight on one side while casting the other in soft shadow. Each rolling dewdrop captures and shimmers captivatingly, creating an atmosphere of profound tranquility, freshness, and vitality.", + "seed": 22012, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_2/index_7201_seed_22012_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.4238782852888107, + "vlm_logprob": -0.4238782852888107, + "vlm_yes_no_margin": 1.7249999523162842, + "vlm_token_credit": -0.4238782852888107, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0268409326672554 + ], + "opposite_logprob_tokens": [ + -4.026841163635254 + ], + "logprob_reward": -0.0268409326672554, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.0268409326672554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15567973256111145, + 0.11284246295690536, + 0.2713223397731781, + 0.03293516859412193, + 0.07264845818281174, + 0.03524855896830559, + 0.06571494042873383, + 0.07446750998497009, + 0.11824951320886612, + 0.06089131906628609 + ], + "credit_map": [ + 0.01594403386116028, + 0.011798146180808544, + 0.024260370060801506, + 0.017443694174289703, + 0.014779745601117611, + 0.026618706062436104, + 0.02343226969242096, + 0.02140275575220585, + 0.013818354345858097, + 0.007795718498528004, + 0.014660556800663471, + 0.017785340547561646, + 0.012603426352143288, + 0.012746011838316917, + 0.016081662848591805, + 0.017351390793919563, + 0.05036396160721779, + 0.02282167412340641, + 0.03742599859833717, + 0.023274676874279976, + 0.026688894256949425, + 0.03493224456906319, + 0.03561808541417122, + 0.04019678756594658, + 0.004102783743292093, + 0.002855139784514904, + 0.005609096027910709, + 0.004678002092987299, + 0.002558851847425103, + 0.005407033488154411, + 0.003644075244665146, + 0.004080183804035187, + 0.01064865943044424, + 0.004676850978285074, + 0.013316933065652847, + 0.00995839573442936, + 0.004961227998137474, + 0.0084189148619771, + 0.008120529353618622, + 0.012546946294605732, + 0.003104808973148465, + 0.0022896488662809134, + 0.005414212588220835, + 0.004971626680344343, + 0.004086020402610302, + 0.003033213084563613, + 0.006278881803154945, + 0.006070147268474102, + 0.008822237141430378, + 0.004677773453295231, + 0.011270563118159771, + 0.009592846035957336, + 0.0058602322824299335, + 0.006486683152616024, + 0.008109020069241524, + 0.01089558657258749, + 0.01296326331794262, + 0.00768500380218029, + 0.009974739514291286, + 0.009019218385219574, + 0.006594400387257338, + 0.00813220627605915, + 0.009674904868006706, + 0.010423778556287289, + 0.023080360144376755, + 0.008234015665948391, + 0.019690964370965958, + 0.014943609945476055, + 0.007496316451579332, + 0.01225235778838396, + 0.010482040233910084, + 0.022069847211241722, + 0.007806376088410616, + 0.004220127128064632, + 0.009539292193949223, + 0.006565696559846401, + 0.004564322996884584, + 0.010531756095588207, + 0.007032770197838545, + 0.010630976408720016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.0268409326672554 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159927010536194 + ], + "opposite_logprob_tokens": [ + -1.7015992403030396 + ], + "logprob_reward": -0.20159927010536194, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159927010536194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09917981177568436, + 0.14288116991519928, + 0.3958812952041626, + 0.0838509276509285, + 0.030052268877625465, + 0.02431846410036087, + 0.05156315118074417, + 0.06635960191488266, + 0.06665310263633728, + 0.03926018998026848 + ], + "credit_map": [ + 0.011536811478435993, + 0.014002500101923943, + 0.01263502798974514, + 0.013444640673696995, + 0.008968492038547993, + 0.010094020515680313, + 0.014126874506473541, + 0.014371444471180439, + 0.01253172755241394, + 0.00909500289708376, + 0.033697791397571564, + 0.02010110765695572, + 0.009902586229145527, + 0.01347503624856472, + 0.016274692490696907, + 0.027803221717476845, + 0.07770255953073502, + 0.03445792198181152, + 0.06241738423705101, + 0.03450930118560791, + 0.028401345014572144, + 0.052904631942510605, + 0.041053738445043564, + 0.06443437933921814, + 0.016240274533629417, + 0.007361870724707842, + 0.013996833935379982, + 0.010839860886335373, + 0.0056309630163013935, + 0.007670804392546415, + 0.009557712823152542, + 0.012552604079246521, + 0.003831699024885893, + 0.002575993537902832, + 0.004512574523687363, + 0.0028243393171578646, + 0.0029574159998446703, + 0.0036678723990917206, + 0.004315128084272146, + 0.005367242731153965, + 0.00282391719520092, + 0.0017905115382745862, + 0.003953400533646345, + 0.0027295537292957306, + 0.002055724849924445, + 0.00271451473236084, + 0.004630009178072214, + 0.0036208308301866055, + 0.0036163327749818563, + 0.002151717897504568, + 0.013814711011946201, + 0.0037181496154516935, + 0.003786195768043399, + 0.004119007848203182, + 0.0059548369608819485, + 0.014402194879949093, + 0.011287428438663483, + 0.007347078528255224, + 0.0084206722676754, + 0.007241077721118927, + 0.006134830415248871, + 0.0051849414594471455, + 0.01040032971650362, + 0.010343234986066818, + 0.012628056108951569, + 0.004612463992089033, + 0.011524473316967487, + 0.006566242314875126, + 0.00390625, + 0.007476233411580324, + 0.009115800261497498, + 0.010823572985827923, + 0.006767392624169588, + 0.002756151370704174, + 0.007331141736358404, + 0.004368093330413103, + 0.003098641289398074, + 0.004946365021169186, + 0.003964775241911411, + 0.0060276249423623085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.20159927010536194 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8984221816062927 + ], + "opposite_logprob_tokens": [ + -0.5234221816062927 + ], + "logprob_reward": -0.8984221816062927, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8984221816062927, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1251700520515442, + 0.1242462694644928, + 0.2283524125814438, + 0.19170929491519928, + 0.0853951945900917, + 0.0513942651450634, + 0.04836892709136009, + 0.05729901045560837, + 0.033665455877780914, + 0.05439907684922218 + ], + "credit_map": [ + 0.015305736102163792, + 0.00797955971211195, + 0.021423498168587685, + 0.015389482490718365, + 0.012525360099971294, + 0.015085744671523571, + 0.019115878269076347, + 0.018344802781939507, + 0.010815484449267387, + 0.007742227055132389, + 0.01962447538971901, + 0.01665210910141468, + 0.014737609773874283, + 0.010561513714492321, + 0.023735273629426956, + 0.02037758193910122, + 0.034804828464984894, + 0.01811736635863781, + 0.03139512985944748, + 0.021953744813799858, + 0.02634800225496292, + 0.02695954404771328, + 0.031020285561680794, + 0.03775350749492645, + 0.034352097660303116, + 0.014422672800719738, + 0.035694390535354614, + 0.015341478399932384, + 0.00978411827236414, + 0.022213751450181007, + 0.014848432503640652, + 0.04505236819386482, + 0.009900677017867565, + 0.004256508778780699, + 0.019943542778491974, + 0.007756698410958052, + 0.005249482113867998, + 0.008234985172748566, + 0.006783238146454096, + 0.02327006310224533, + 0.005552561022341251, + 0.00456448458135128, + 0.0075171589851379395, + 0.007270129397511482, + 0.0045289546251297, + 0.009496734477579594, + 0.004786495119333267, + 0.007677746936678886, + 0.003025141777470708, + 0.001757764839567244, + 0.013049593195319176, + 0.004135213792324066, + 0.004409351851791143, + 0.002548584481701255, + 0.0066689904779195786, + 0.01277428213506937, + 0.010682452470064163, + 0.003612042870372534, + 0.00821971707046032, + 0.004025832284241915, + 0.006218707654625177, + 0.008880151435732841, + 0.0074725691229105, + 0.008187535218894482, + 0.005085546523332596, + 0.001962201204150915, + 0.00617480231449008, + 0.0027037279214709997, + 0.0029885834082961082, + 0.0037723989225924015, + 0.0046939607709646225, + 0.006284233182668686, + 0.004704429302364588, + 0.0030289713758975267, + 0.009909966960549355, + 0.004501777235418558, + 0.004874269012361765, + 0.00471182307228446, + 0.010127715766429901, + 0.012540122494101524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.8984221816062927 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743280410766602 + ], + "opposite_logprob_tokens": [ + -0.47432804107666016 + ], + "logprob_reward": -0.9743280410766602, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743280410766602, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14927388727664948, + 0.15917491912841797, + 0.23391738533973694, + 0.054329365491867065, + 0.08483456820249557, + 0.03655964508652687, + 0.07518596947193146, + 0.052674151957035065, + 0.11084497720003128, + 0.0432051345705986 + ], + "credit_map": [ + 0.02770419418811798, + 0.011345828883349895, + 0.025941453874111176, + 0.012680625542998314, + 0.009428728371858597, + 0.023242393508553505, + 0.016454918310046196, + 0.022475728765130043, + 0.013232641853392124, + 0.011099129915237427, + 0.03632683679461479, + 0.022340133786201477, + 0.014298669062554836, + 0.015422211959958076, + 0.022178981453180313, + 0.02427630126476288, + 0.039754610508680344, + 0.018952148035168648, + 0.0370117723941803, + 0.020476747304201126, + 0.02032054029405117, + 0.031244119629263878, + 0.0283488929271698, + 0.037808530032634735, + 0.004849148914217949, + 0.0029233817476779222, + 0.011232600547373295, + 0.007435384672135115, + 0.0047030881978571415, + 0.006087517831474543, + 0.00782864447683096, + 0.009269589558243752, + 0.00896391086280346, + 0.006275834050029516, + 0.012742976658046246, + 0.00889989547431469, + 0.0068410104140639305, + 0.009457623586058617, + 0.015582159161567688, + 0.016071157529950142, + 0.0039931004866957664, + 0.003288727253675461, + 0.0051895505748689175, + 0.0048286207020282745, + 0.003842475125566125, + 0.004375848453491926, + 0.005789316259324551, + 0.005252004135400057, + 0.006371759809553623, + 0.003748818300664425, + 0.018792595714330673, + 0.006012328900396824, + 0.006189546547830105, + 0.005389397963881493, + 0.007779939100146294, + 0.020901571959257126, + 0.010224719531834126, + 0.004161725286394358, + 0.008200069889426231, + 0.0050370702520012856, + 0.004695916082710028, + 0.006529503036290407, + 0.005646595731377602, + 0.008178548887372017, + 0.021679896861314774, + 0.007824805565178394, + 0.02114204503595829, + 0.010231507010757923, + 0.006726352497935295, + 0.012124542146921158, + 0.011055447161197662, + 0.020060379058122635, + 0.008749306201934814, + 0.0030481850262731314, + 0.008710217662155628, + 0.0037517754826694727, + 0.0016884743236005306, + 0.0038166630547493696, + 0.0026860239449888468, + 0.010754484683275223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.9743280410766602 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01820100098848343 + ], + "opposite_logprob_tokens": [ + -4.018200874328613 + ], + "logprob_reward": -0.01820100098848343, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01820100098848343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07494118809700012, + 0.17048615217208862, + 0.2797996997833252, + 0.09960202872753143, + 0.11236941069364548, + 0.03430620953440666, + 0.03440738096833229, + 0.11772933602333069, + 0.047841403633356094, + 0.028517121449112892 + ], + "credit_map": [ + 0.009583156555891037, + 0.004525245167315006, + 0.01068995799869299, + 0.01019563339650631, + 0.007970021106302738, + 0.008959299884736538, + 0.011040851473808289, + 0.011977040208876133, + 0.014792468398809433, + 0.010992951691150665, + 0.04446277767419815, + 0.021573930978775024, + 0.010502350516617298, + 0.015577712096273899, + 0.016758223995566368, + 0.03582574799656868, + 0.05928986519575119, + 0.024795176461338997, + 0.047093041241168976, + 0.022259866818785667, + 0.019547468051314354, + 0.03480672463774681, + 0.028452211990952492, + 0.04355538636445999, + 0.018493635579943657, + 0.007060348987579346, + 0.01753348670899868, + 0.007443942129611969, + 0.007047436665743589, + 0.010127909481525421, + 0.011989177204668522, + 0.019906101748347282, + 0.012759514153003693, + 0.011267478577792645, + 0.013679825700819492, + 0.012735682539641857, + 0.010478496551513672, + 0.015558612532913685, + 0.020986037328839302, + 0.014903778210282326, + 0.0035407792311161757, + 0.00459033390507102, + 0.006161536555737257, + 0.003935710992664099, + 0.00272116856649518, + 0.004300098866224289, + 0.0033209742978215218, + 0.0057356106117367744, + 0.0034901010803878307, + 0.003364024218171835, + 0.0042303381487727165, + 0.005047422368079424, + 0.0036543882451951504, + 0.005397094879299402, + 0.0053001949563622475, + 0.003923818934708834, + 0.02045413851737976, + 0.01049274392426014, + 0.017590342089533806, + 0.014854944311082363, + 0.008360679261386395, + 0.014451300725340843, + 0.013749691657721996, + 0.017775515094399452, + 0.008419567719101906, + 0.003418356878682971, + 0.008535779081285, + 0.005333052482455969, + 0.0031266966834664345, + 0.004451449494808912, + 0.003709023352712393, + 0.0108474837616086, + 0.0033601969480514526, + 0.0025033196434378624, + 0.0039941067807376385, + 0.003617912530899048, + 0.0028980039060115814, + 0.003952548839151859, + 0.004321027081459761, + 0.00387001084163785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.01820100098848343 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "No", + "Yes" + ] + }, + { + "prompt": "stunning close-up, shot from a slightly tilted angle, the camera remains still, focusing on a red apple suspended from a gnarled, mossy branch. The apple, perfectly spherical and with a rich, vibrant crimson skin, is dotted with shimmering dewdrops on its smooth surface, each one acting like a miniature magnifying glass, reflecting the surrounding light. As an invisible breeze sweeps by, the apple, its short, firm, dark brown stalk, begins to sway rhythmically and gently. This movement causes the dewdrops on its surface to roll gracefully and slowly, gliding along the apple's smooth curves. Some coalesce into larger droplets, shimmering in their movement. The first soft rays of morning sunlight, filtering through the blurred emerald green leaves in the background, strike the apple, casting a warm golden highlight on one side while casting the other in soft shadow. Each rolling dewdrop captures and shimmers captivatingly, creating an atmosphere of profound tranquility, freshness, and vitality.", + "seed": 22013, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_2/index_7201_seed_22013_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14150154814124108, + "vlm_logprob": -0.14150154814124108, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.14150154814124108, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012282331474125385 + ], + "opposite_logprob_tokens": [ + -5.387282371520996 + ], + "logprob_reward": -0.012282331474125385, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.012282331474125385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11662351340055466, + 0.0816623643040657, + 0.14728900790214539, + 0.03986211493611336, + 0.03891638293862343, + 0.05584816634654999, + 0.027552681043744087, + 0.03395385667681694, + 0.36395639181137085, + 0.09433553367853165 + ], + "credit_map": [ + 0.014455516822636127, + 0.008821125142276287, + 0.01355205662548542, + 0.01288291160017252, + 0.017238210886716843, + 0.01803395338356495, + 0.015070566907525063, + 0.016569165512919426, + 0.007847701199352741, + 0.010830102488398552, + 0.011444698087871075, + 0.007975282147526741, + 0.01073485892266035, + 0.012070395052433014, + 0.009979167953133583, + 0.01078015100210905, + 0.016065966337919235, + 0.0107087017968297, + 0.01741013489663601, + 0.021694952622056007, + 0.014432935044169426, + 0.030439462512731552, + 0.018801845610141754, + 0.017734991386532784, + 0.005142682231962681, + 0.0033812792971730232, + 0.004266512114554644, + 0.003816517535597086, + 0.005424723494797945, + 0.008770600892603397, + 0.004768043756484985, + 0.004291751887649298, + 0.004810113459825516, + 0.0028447431977838278, + 0.0026206152979284525, + 0.00352273928001523, + 0.0058824969455599785, + 0.0054822140373289585, + 0.008318291045725346, + 0.005435167346149683, + 0.006758505012840033, + 0.004346886184066534, + 0.008073927834630013, + 0.0049555543810129166, + 0.007153472397476435, + 0.006812821142375469, + 0.00991764385253191, + 0.007829356007277966, + 0.0019485928351059556, + 0.003336066845804453, + 0.002141887554898858, + 0.0034879534505307674, + 0.003980324137955904, + 0.005286174360662699, + 0.004123752936720848, + 0.003247926477342844, + 0.004895142279565334, + 0.0036825521383434534, + 0.004148401785641909, + 0.002953974297270179, + 0.003728592535480857, + 0.004720012191683054, + 0.005226426757872105, + 0.004598753992468119, + 0.024765171110630035, + 0.008473368361592293, + 0.13823601603507996, + 0.014020548202097416, + 0.007740091532468796, + 0.019071122631430626, + 0.012318079359829426, + 0.13933196663856506, + 0.005742660257965326, + 0.011829799041152, + 0.011005723848938942, + 0.01817058026790619, + 0.010708775371313095, + 0.011812353506684303, + 0.012179126031696796, + 0.012886514887213707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.012282331474125385 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521108388900757 + ], + "opposite_logprob_tokens": [ + -1.5021108388900757 + ], + "logprob_reward": -0.2521108388900757, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521108388900757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09494688361883163, + 0.03487567603588104, + 0.2676161825656891, + 0.055299676954746246, + 0.054661862552165985, + 0.12432346493005753, + 0.07080578804016113, + 0.033618614077568054, + 0.22553719580173492, + 0.03831470385193825 + ], + "credit_map": [ + 0.014343410730361938, + 0.00709722051396966, + 0.01354173943400383, + 0.012705502100288868, + 0.01055228989571333, + 0.012399232015013695, + 0.010413721203804016, + 0.01389377098530531, + 0.00260557746514678, + 0.0033106179907917976, + 0.0033088033087551594, + 0.0037376026157289743, + 0.007678279187530279, + 0.005440995097160339, + 0.004791101440787315, + 0.004002699628472328, + 0.032387010753154755, + 0.023025263100862503, + 0.03531267121434212, + 0.03552311286330223, + 0.030955595895648003, + 0.05657771974802017, + 0.020553747192025185, + 0.03328106924891472, + 0.006063125096261501, + 0.0030202108900994062, + 0.006418326403945684, + 0.005158505868166685, + 0.00795307382941246, + 0.010879561305046082, + 0.008076639845967293, + 0.007730236276984215, + 0.0056820157915353775, + 0.006776198744773865, + 0.005017612129449844, + 0.004524427466094494, + 0.008466595783829689, + 0.007911399938166142, + 0.00838108267635107, + 0.007902530953288078, + 0.025470485910773277, + 0.007905405014753342, + 0.021978897973895073, + 0.007962310686707497, + 0.01014618668705225, + 0.0170064065605402, + 0.013580169528722763, + 0.02027360536158085, + 0.005194846540689468, + 0.00404806062579155, + 0.0053382618352770805, + 0.004361083265393972, + 0.014012006111443043, + 0.01454242691397667, + 0.012945800088346004, + 0.010363299399614334, + 0.0034967041574418545, + 0.0028262531850486994, + 0.002817356027662754, + 0.0021850387565791607, + 0.005578568670898676, + 0.006775558926165104, + 0.0055451844818890095, + 0.004393951501697302, + 0.014554629102349281, + 0.005170306656509638, + 0.08688892424106598, + 0.009082543663680553, + 0.0041285110637545586, + 0.009290998801589012, + 0.007078518625348806, + 0.08934276551008224, + 0.0030119260773062706, + 0.005574462935328484, + 0.004282396286725998, + 0.008287280797958374, + 0.0037029350642114878, + 0.002917145611718297, + 0.004747143480926752, + 0.005791413597762585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 5 + ], + "selected_reward": -0.2521108388900757 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521384358406067 + ], + "opposite_logprob_tokens": [ + -1.502138376235962 + ], + "logprob_reward": -0.2521384358406067, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521384358406067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09314542263746262, + 0.09804072976112366, + 0.13951121270656586, + 0.15747064352035522, + 0.09965416043996811, + 0.11260471493005753, + 0.053947124630212784, + 0.12252089381217957, + 0.06936773657798767, + 0.053737372159957886 + ], + "credit_map": [ + 0.012674118392169476, + 0.007612327579408884, + 0.010274921543896198, + 0.015975723043084145, + 0.010424495674669743, + 0.012032117694616318, + 0.01092859823256731, + 0.013223120011389256, + 0.013160787522792816, + 0.013603989966213703, + 0.01493724063038826, + 0.015510660596191883, + 0.009701518341898918, + 0.007926695980131626, + 0.011110885068774223, + 0.012088948860764503, + 0.01885255239903927, + 0.012064531445503235, + 0.028514327481389046, + 0.01671559363603592, + 0.014830983243882656, + 0.010824176482856274, + 0.013086672872304916, + 0.024622377008199692, + 0.018709978088736534, + 0.010783184319734573, + 0.013204853981733322, + 0.00955397542566061, + 0.02983401156961918, + 0.03664218261837959, + 0.019663283601403236, + 0.01907918229699135, + 0.012347149662673473, + 0.009156177751719952, + 0.006655486766248941, + 0.008450278080999851, + 0.013137610629200935, + 0.010043550282716751, + 0.022438982501626015, + 0.017424924299120903, + 0.009581808932125568, + 0.006499813869595528, + 0.019353266805410385, + 0.00891736801713705, + 0.013647828251123428, + 0.01046798750758171, + 0.030153948813676834, + 0.013982689008116722, + 0.004215413238853216, + 0.007938534021377563, + 0.00591933261603117, + 0.00753391720354557, + 0.006606029346585274, + 0.007449350319802761, + 0.007736573927104473, + 0.006547971628606319, + 0.01622125506401062, + 0.018271420150995255, + 0.00971902534365654, + 0.0146425636485219, + 0.01604553312063217, + 0.016139650717377663, + 0.016191063448786736, + 0.015290381386876106, + 0.00491819903254509, + 0.0029357990715652704, + 0.023075031116604805, + 0.004124708008021116, + 0.003338649868965149, + 0.00421083765104413, + 0.0042813923209905624, + 0.022483117878437042, + 0.004826448857784271, + 0.006522015668451786, + 0.006043151021003723, + 0.010689579881727695, + 0.006504701916128397, + 0.004769385792315006, + 0.006658914498984814, + 0.007723173592239618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 7 + ], + "selected_reward": -0.2521384358406067 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17990021407604218 + ], + "opposite_logprob_tokens": [ + -1.8049001693725586 + ], + "logprob_reward": -0.17990021407604218, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17990021407604218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13580723106861115, + 0.04919242858886719, + 0.14018873870372772, + 0.11675863713026047, + 0.027210090309381485, + 0.16608451306819916, + 0.04169676452875137, + 0.04234487563371658, + 0.2481500655412674, + 0.0325666218996048 + ], + "credit_map": [ + 0.019449524581432343, + 0.010630073957145214, + 0.015148651786148548, + 0.014694559387862682, + 0.022972319275140762, + 0.020031476393342018, + 0.017297202721238136, + 0.01558343879878521, + 0.004183050245046616, + 0.005290883127599955, + 0.004836014937609434, + 0.00537117850035429, + 0.00835556909441948, + 0.008372636511921883, + 0.006923993583768606, + 0.005859109107404947, + 0.019284380599856377, + 0.015262734144926071, + 0.018130788579583168, + 0.019542409107089043, + 0.01879347488284111, + 0.017994090914726257, + 0.014891180209815502, + 0.016289694234728813, + 0.013054938986897469, + 0.008853220380842686, + 0.009523346088826656, + 0.010321203619241714, + 0.020720724016427994, + 0.02651500143110752, + 0.015579447150230408, + 0.012190770357847214, + 0.0022812453098595142, + 0.004349244758486748, + 0.0021707434207201004, + 0.0033600395545363426, + 0.0025268655735999346, + 0.0026375846937298775, + 0.005041252821683884, + 0.004843117203563452, + 0.021616054698824883, + 0.013109585270285606, + 0.026675576344132423, + 0.011507414281368256, + 0.020435480400919914, + 0.01607726514339447, + 0.03556577488780022, + 0.021097369492053986, + 0.0033574083354324102, + 0.005063974764198065, + 0.004319088999181986, + 0.0064669265411794186, + 0.005870351102203131, + 0.005859586410224438, + 0.005564483348280191, + 0.00519494479522109, + 0.0035377931781113148, + 0.005340938456356525, + 0.0031519136391580105, + 0.0056555490009486675, + 0.008030882105231285, + 0.006708137225359678, + 0.004933029878884554, + 0.0049866363406181335, + 0.018813740462064743, + 0.006272647529840469, + 0.08741161227226257, + 0.012560801580548286, + 0.007285482250154018, + 0.01653447188436985, + 0.010057502426207066, + 0.08921383321285248, + 0.002676198724657297, + 0.004183140583336353, + 0.0028161134105175734, + 0.00452941982075572, + 0.005736965220421553, + 0.0042010024189949036, + 0.004631747957319021, + 0.003792033065110445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.17990021407604218 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075920425355434 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011075920425355434, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075920425355434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1063607707619667, + 0.03805623948574066, + 0.14363151788711548, + 0.16442933678627014, + 0.03757976368069649, + 0.11355630308389664, + 0.029648883268237114, + 0.05370420590043068, + 0.29494497179985046, + 0.018087999895215034 + ], + "credit_map": [ + 0.009410317987203598, + 0.011953798122704029, + 0.01059671025723219, + 0.018467310816049576, + 0.009987916797399521, + 0.014792855829000473, + 0.015219083987176418, + 0.015932783484458923, + 0.0028102979995310307, + 0.0048902216367423534, + 0.006397970020771027, + 0.005028554238379002, + 0.004635353572666645, + 0.004296824336051941, + 0.004941965918987989, + 0.0050550526939332485, + 0.021917302161455154, + 0.010242833755910397, + 0.025800514966249466, + 0.023040594533085823, + 0.0098876953125, + 0.015499919652938843, + 0.013665433041751385, + 0.02357722818851471, + 0.01940365694463253, + 0.013191699981689453, + 0.01309936959296465, + 0.014050998724997044, + 0.03094879724085331, + 0.034581635147333145, + 0.020950922742486, + 0.018202263861894608, + 0.001889024511910975, + 0.0026344843208789825, + 0.0019265338778495789, + 0.0027659600600600243, + 0.009109039790928364, + 0.008253800682723522, + 0.00636714044958353, + 0.004633782431483269, + 0.021867794916033745, + 0.006788183934986591, + 0.02450142242014408, + 0.006060041021555662, + 0.00830589048564434, + 0.013185125775635242, + 0.009936928749084473, + 0.022910911589860916, + 0.0034029032103717327, + 0.00240314775146544, + 0.002926381304860115, + 0.00352405640296638, + 0.003778662532567978, + 0.0039344970136880875, + 0.0057294536381959915, + 0.003949781414121389, + 0.007404714822769165, + 0.005961183924227953, + 0.0044285207986831665, + 0.0050579337403178215, + 0.006870586425065994, + 0.008938307873904705, + 0.008216286078095436, + 0.006826676893979311, + 0.01899074949324131, + 0.00904938206076622, + 0.09947450459003448, + 0.02659100480377674, + 0.008655285462737083, + 0.014911175705492496, + 0.016709811985492706, + 0.10056305676698685, + 0.0019216870423406363, + 0.003071799408644438, + 0.0016870093531906605, + 0.0027704290114343166, + 0.0021182119380682707, + 0.0018755688797682524, + 0.002377696568146348, + 0.0022655969951301813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.011075920425355434 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, shot from a slightly tilted angle, the camera remains still, focusing on a red apple suspended from a gnarled, mossy branch. The apple, perfectly spherical and with a rich, vibrant crimson skin, is dotted with shimmering dewdrops on its smooth surface, each one acting like a miniature magnifying glass, reflecting the surrounding light. As an invisible breeze sweeps by, the apple, its short, firm, dark brown stalk, begins to sway rhythmically and gently. This movement causes the dewdrops on its surface to roll gracefully and slowly, gliding along the apple's smooth curves. Some coalesce into larger droplets, shimmering in their movement. The first soft rays of morning sunlight, filtering through the blurred emerald green leaves in the background, strike the apple, casting a warm golden highlight on one side while casting the other in soft shadow. Each rolling dewdrop captures and shimmers captivatingly, creating an atmosphere of profound tranquility, freshness, and vitality.", + "seed": 22014, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_2/index_7201_seed_22014_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1544161716476083, + "vlm_logprob": -0.1544161716476083, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.1544161716476083, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014374868012964725 + ], + "opposite_logprob_tokens": [ + -4.889374732971191 + ], + "logprob_reward": -0.014374868012964725, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.014374868012964725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17889750003814697, + 0.06069798022508621, + 0.09119708836078644, + 0.11254830658435822, + 0.05377601459622383, + 0.07721559703350067, + 0.03602764382958412, + 0.13336002826690674, + 0.21433787047863007, + 0.04194200038909912 + ], + "credit_map": [ + 0.029378240928053856, + 0.019844036549329758, + 0.027239613234996796, + 0.01607067883014679, + 0.019499093294143677, + 0.018945755437016487, + 0.021309277042746544, + 0.026610808447003365, + 0.006399313919246197, + 0.00554427457973361, + 0.00893378909677267, + 0.008333913050591946, + 0.0057104662992060184, + 0.010089840739965439, + 0.008614949882030487, + 0.007071434985846281, + 0.008385511115193367, + 0.009606282226741314, + 0.012500474229454994, + 0.013304109685122967, + 0.009245185181498528, + 0.010418440215289593, + 0.01959165930747986, + 0.008145422674715519, + 0.006820050999522209, + 0.010476414114236832, + 0.01835416443645954, + 0.01581641286611557, + 0.013301486149430275, + 0.0072927698493003845, + 0.029723113402724266, + 0.010763898491859436, + 0.007088294252753258, + 0.005694129038602114, + 0.008232612162828445, + 0.006310407072305679, + 0.00635863933712244, + 0.005735132377594709, + 0.006987676955759525, + 0.007369122933596373, + 0.010665342211723328, + 0.007245715707540512, + 0.011674816720187664, + 0.009518373757600784, + 0.008458299562335014, + 0.007031585089862347, + 0.01513388566672802, + 0.007487576454877853, + 0.0034242249093949795, + 0.002777410438284278, + 0.007204348687082529, + 0.006056535989046097, + 0.0038798637688159943, + 0.0034297839738428593, + 0.005584146361798048, + 0.0036713306326419115, + 0.012732302770018578, + 0.014521722681820393, + 0.023623032495379448, + 0.014578021131455898, + 0.010452414862811565, + 0.010108652524650097, + 0.03265637531876564, + 0.014687508344650269, + 0.020839432254433632, + 0.017484288662672043, + 0.03868125006556511, + 0.024282975122332573, + 0.023419471457600594, + 0.03558932989835739, + 0.03137413039803505, + 0.022666994482278824, + 0.006024247966706753, + 0.003699101973325014, + 0.006210054270923138, + 0.004669348709285259, + 0.005547358188778162, + 0.005112271290272474, + 0.005641980562359095, + 0.005037636961787939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.014374868012964725 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158885419368744 + ], + "opposite_logprob_tokens": [ + -1.7015888690948486 + ], + "logprob_reward": -0.20158885419368744, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158885419368744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13932880759239197, + 0.09229300916194916, + 0.08197861164808273, + 0.06584406644105911, + 0.07844404131174088, + 0.1343519687652588, + 0.12440790981054306, + 0.07219410687685013, + 0.1774003505706787, + 0.03375714644789696 + ], + "credit_map": [ + 0.01552511379122734, + 0.018319077789783478, + 0.020557023584842682, + 0.015308467671275139, + 0.020055430009961128, + 0.01667289063334465, + 0.0155921820551157, + 0.017298618331551552, + 0.010886636562645435, + 0.007900744676589966, + 0.016129154711961746, + 0.013223739340901375, + 0.009341288357973099, + 0.015039597637951374, + 0.012289196252822876, + 0.007482653483748436, + 0.00841948576271534, + 0.008995544165372849, + 0.012183072976768017, + 0.011350525543093681, + 0.008931945078074932, + 0.005978522356599569, + 0.018537895753979683, + 0.00758162559941411, + 0.009508493356406689, + 0.007158521097153425, + 0.00936792977154255, + 0.008405257947742939, + 0.008923191577196121, + 0.007090374361723661, + 0.007073649205267429, + 0.008316651917994022, + 0.010223417542874813, + 0.007445422932505608, + 0.014628060162067413, + 0.009339646436274052, + 0.007564311381429434, + 0.009457225911319256, + 0.01020447351038456, + 0.009581482037901878, + 0.013812561519443989, + 0.016342753544449806, + 0.018184127286076546, + 0.01752215065062046, + 0.01510731503367424, + 0.015548136085271835, + 0.024907048791646957, + 0.012927875854074955, + 0.012059701606631279, + 0.012443684972822666, + 0.018538236618041992, + 0.016628554090857506, + 0.012148941867053509, + 0.01777171716094017, + 0.02230173349380493, + 0.012515351176261902, + 0.007123447954654694, + 0.007176424842327833, + 0.011223854497075081, + 0.009155317209661007, + 0.00703622167930007, + 0.005319648887962103, + 0.01611083373427391, + 0.009048357605934143, + 0.016834788024425507, + 0.016040390357375145, + 0.02964446321129799, + 0.02183625102043152, + 0.021741069853305817, + 0.028945306316018105, + 0.025111166760325432, + 0.017246924340724945, + 0.004180229734629393, + 0.003410482546314597, + 0.004669961519539356, + 0.00391851831227541, + 0.0032066130079329014, + 0.0066374740563333035, + 0.003609601641073823, + 0.004124268423765898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.20158885419368744 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38709962368011475 + ], + "opposite_logprob_tokens": [ + -1.1370996236801147 + ], + "logprob_reward": -0.38709962368011475, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38709962368011475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12620675563812256, + 0.07753553986549377, + 0.043511055409908295, + 0.03731163591146469, + 0.07781218737363815, + 0.08884746581315994, + 0.04357164725661278, + 0.1900061070919037, + 0.17322072386741638, + 0.14197692275047302 + ], + "credit_map": [ + 0.011438172310590744, + 0.012489883229136467, + 0.018939444795250893, + 0.021062588319182396, + 0.016104131937026978, + 0.007860232144594193, + 0.025915278121829033, + 0.012397019192576408, + 0.008133778348565102, + 0.007521117106080055, + 0.010978331789374352, + 0.011267896741628647, + 0.010011319071054459, + 0.007404951844364405, + 0.015407637692987919, + 0.0068105109967291355, + 0.004392731003463268, + 0.0067297182977199554, + 0.005941230803728104, + 0.0048152790404856205, + 0.006107531022280455, + 0.004503255244344473, + 0.006879392545670271, + 0.004141917452216148, + 0.0039019850082695484, + 0.005103074014186859, + 0.006016548722982407, + 0.006314618978649378, + 0.003868441330268979, + 0.0023117337841540575, + 0.0063586970791220665, + 0.0034365372266620398, + 0.007645081728696823, + 0.011232597753405571, + 0.011141415685415268, + 0.00995553471148014, + 0.011954331770539284, + 0.0062588294968008995, + 0.011261786334216595, + 0.008362606167793274, + 0.008227281272411346, + 0.007753225974738598, + 0.011382045224308968, + 0.00814241822808981, + 0.011629292741417885, + 0.010767029598355293, + 0.017387863248586655, + 0.01355830393731594, + 0.003730985801666975, + 0.003983655944466591, + 0.009048913605511189, + 0.006188234779983759, + 0.00394570454955101, + 0.005547035951167345, + 0.0075898063369095325, + 0.0035373077262192965, + 0.018145445734262466, + 0.01694728620350361, + 0.032514236867427826, + 0.019376138225197792, + 0.013642873615026474, + 0.016285734251141548, + 0.05297050625085831, + 0.020123885944485664, + 0.017611810937523842, + 0.014842324890196323, + 0.03256421163678169, + 0.01959775574505329, + 0.018997954204678535, + 0.027343740686774254, + 0.025025086477398872, + 0.0172378309071064, + 0.010875987820327282, + 0.009943249635398388, + 0.022392893210053444, + 0.022302167490124702, + 0.023165859282016754, + 0.015917779877781868, + 0.02292376570403576, + 0.014455214142799377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.38709962368011475 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037805378437042 + ], + "opposite_logprob_tokens": [ + -1.910378098487854 + ], + "logprob_reward": -0.16037805378437042, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037805378437042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1465688943862915, + 0.0588533990085125, + 0.06070102006196976, + 0.057367801666259766, + 0.028156854212284088, + 0.07618594169616699, + 0.03712790086865425, + 0.08211987465620041, + 0.3913135826587677, + 0.06160467490553856 + ], + "credit_map": [ + 0.019378269091248512, + 0.019831890240311623, + 0.018777772784233093, + 0.013714049011468887, + 0.023473534733057022, + 0.016936466097831726, + 0.01606268808245659, + 0.018394216895103455, + 0.0040292320773005486, + 0.006023990456014872, + 0.008795239962637424, + 0.009356523863971233, + 0.006298802327364683, + 0.005876988638192415, + 0.01185471098870039, + 0.006617911625653505, + 0.006537762004882097, + 0.009681581519544125, + 0.008981157094240189, + 0.007058342453092337, + 0.006550638470798731, + 0.009299998171627522, + 0.006758426316082478, + 0.005833113566040993, + 0.006963614374399185, + 0.0055268509313464165, + 0.0077753616496920586, + 0.007035552524030209, + 0.007074754685163498, + 0.0064257425256073475, + 0.009047964587807655, + 0.007517960853874683, + 0.0033914558589458466, + 0.003397450316697359, + 0.003865279722958803, + 0.004117409233003855, + 0.003235794370993972, + 0.0029344186186790466, + 0.003981395624577999, + 0.003233649767935276, + 0.00834262277930975, + 0.006708931177854538, + 0.009487434290349483, + 0.007576192729175091, + 0.009229960851371288, + 0.007158765569329262, + 0.019640842452645302, + 0.00804118812084198, + 0.00331056397408247, + 0.005342683754861355, + 0.005272923968732357, + 0.0038877076003700495, + 0.0034982888028025627, + 0.006163118407130241, + 0.004781453870236874, + 0.0048711588606238365, + 0.0061482335440814495, + 0.006354552693665028, + 0.0140983359888196, + 0.00840594433248043, + 0.006265830248594284, + 0.010788253508508205, + 0.01971050724387169, + 0.010348218493163586, + 0.040294475853443146, + 0.025106916204094887, + 0.0715687945485115, + 0.041353315114974976, + 0.04430219531059265, + 0.07431227713823318, + 0.05481237173080444, + 0.039563219994306564, + 0.00655672000721097, + 0.00532550411298871, + 0.010828549973666668, + 0.009461059235036373, + 0.007776056416332722, + 0.007198558654636145, + 0.008803989738225937, + 0.005654234904795885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.16037805378437042 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a red apple suspended from a gnarled, mossy branch?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly tilted and does the camera remain still?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the apple have a perfectly spherical shape with vibrant crimson skin dotted with shimmering dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the apple sway rhythmically while the dewdrops roll and coalesce on its surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by soft morning sunlight filtering through blurred emerald green leaves, casting a warm golden highlight and soft shadow on the apple?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639458566904068 + ], + "opposite_logprob_tokens": [ + -4.758639335632324 + ], + "logprob_reward": -0.008639458566904068, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639458566904068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1354684978723526, + 0.10326899588108063, + 0.13173894584178925, + 0.027944840490818024, + 0.06443663686513901, + 0.1096479743719101, + 0.0871625766158104, + 0.07339899241924286, + 0.20221088826656342, + 0.0647217333316803 + ], + "credit_map": [ + 0.021066326647996902, + 0.011487436480820179, + 0.019436169415712357, + 0.01997191272675991, + 0.012884397991001606, + 0.013682909309864044, + 0.019171079620718956, + 0.017768260091543198, + 0.016114739701151848, + 0.009623833931982517, + 0.016771849244832993, + 0.013640917837619781, + 0.010080152191221714, + 0.013278195634484291, + 0.01215835940092802, + 0.011600936762988567, + 0.01658739149570465, + 0.01744876615703106, + 0.020972443744540215, + 0.016312044113874435, + 0.011845031753182411, + 0.013118338771164417, + 0.01984509639441967, + 0.015609830617904663, + 0.0033873887732625008, + 0.0026007352862507105, + 0.004650556482374668, + 0.003353603184223175, + 0.003234256524592638, + 0.002087854780256748, + 0.004319610074162483, + 0.0043108342215418816, + 0.007806363981217146, + 0.0056522502563893795, + 0.011632288806140423, + 0.00864496361464262, + 0.005791177973151207, + 0.007622945122420788, + 0.0074626184068620205, + 0.009824025444686413, + 0.0104152737185359, + 0.010598104447126389, + 0.013742172159254551, + 0.01100960373878479, + 0.018430454656481743, + 0.015461703762412071, + 0.020125867798924446, + 0.00986478105187416, + 0.0069267638027668, + 0.012594493106007576, + 0.015004913322627544, + 0.012792651541531086, + 0.007345534861087799, + 0.008309031836688519, + 0.0164559967815876, + 0.007733182981610298, + 0.009711228311061859, + 0.005263037979602814, + 0.012007443234324455, + 0.0075882612727582455, + 0.005887898150831461, + 0.009606665931642056, + 0.012478414922952652, + 0.010856040753424168, + 0.021162182092666626, + 0.0244577769190073, + 0.03816543146967888, + 0.024809127673506737, + 0.020589297637343407, + 0.0270723607391119, + 0.023876884952187538, + 0.02207779698073864, + 0.010475716553628445, + 0.00651799701154232, + 0.010617120191454887, + 0.010812543332576752, + 0.005659742280840874, + 0.007128126919269562, + 0.006359501741826534, + 0.007150984834879637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.008639458566904068 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_3/index_tensor([2205, 2205])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_3/index_tensor([2205, 2205])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5266a382c83cfcb1911bdd34d7959c2ffce33234 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_3/index_tensor([2205, 2205])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunning close-up, the camera smoothly follows and locks onto the magnificent head of the mythical bird, Suzaku. Each of its feathers is insubstantial, a fusion of blazing vermilion light and vibrant golden flames, forming a flaming, ever-shifting crest. Its eyes, like two miniature stars, burn with pure golden light, revealing an ancient and majestic divinity. Its beak, sharp and pure, is composed of crystallized light, gleaming like a diamond. The background is the vast expanse of space, filled with slowly rotating nebulae composed of deep purple and indigo gas, within which countless tiny stars twinkle like dust. Suzaku gracefully raises its slender neck, its light-forged beak slowly opening, and utters a cry that resounds through the heavens yet carries visually through the vacuum of space. A golden, circular wave of energy suddenly emanates from its mouth, instantly illuminating the surrounding nebula. At the climax of this chirping, its body, composed entirely of flames and light, began to become illusory and translucent, its fiery outline gradually dissolved and blurred, and finally merged perfectly with the surrounding colorful nebula. Through its almost transparent body, one can clearly see the distant stars and dust belts slowly flowing behind it, creating a grand, sacred visual illusion that swayed between reality and illusion, full of ethereality and awe.", + "seed": 14669, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_3/index_2205_seed_14669_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.050416575465351346, + "vlm_logprob": -0.050416575465351346, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.050416575465351346, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1029801145195961 + ], + "opposite_logprob_tokens": [ + -2.477980136871338 + ], + "logprob_reward": -0.1029801145195961, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1029801145195961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09551730751991272, + 0.07020610570907593, + 0.2139471024274826, + 0.04580136388540268, + 0.07544798403978348, + 0.21018828451633453, + 0.11356192827224731, + 0.05103205516934395, + 0.07531996071338654, + 0.04897787421941757 + ], + "credit_map": [ + 0.011928973719477654, + 0.009425556287169456, + 0.013874883763492107, + 0.012466882355511189, + 0.01452880259603262, + 0.012197831645607948, + 0.010160975158214569, + 0.010933397337794304, + 0.008931173011660576, + 0.007216617465019226, + 0.01226082257926464, + 0.007810723967850208, + 0.008571716025471687, + 0.010277941823005676, + 0.00729892635717988, + 0.007838190533220768, + 0.020993968471884727, + 0.008937927894294262, + 0.06598077714443207, + 0.014536078087985516, + 0.012609211727976799, + 0.03352928161621094, + 0.009298896417021751, + 0.048060957342386246, + 0.006112202070653439, + 0.005157801788300276, + 0.00793386809527874, + 0.006513564381748438, + 0.0038545653223991394, + 0.006841297727078199, + 0.004072427283972502, + 0.005315639544278383, + 0.0073831039480865, + 0.005042391363531351, + 0.018683021888136864, + 0.007768141105771065, + 0.006895605940371752, + 0.008076656609773636, + 0.005716320592910051, + 0.015882737934589386, + 0.02180289477109909, + 0.012082711793482304, + 0.04049534350633621, + 0.016100097447633743, + 0.018557609990239143, + 0.056153640151023865, + 0.01279012393206358, + 0.03220587596297264, + 0.017359083518385887, + 0.010146980173885822, + 0.01905214600265026, + 0.011308071203529835, + 0.008815434761345387, + 0.019741905853152275, + 0.00809149257838726, + 0.01904681883752346, + 0.013439669273793697, + 0.004579941742122173, + 0.006998111493885517, + 0.003400897141546011, + 0.0039090728387236595, + 0.006098642945289612, + 0.0043181804940104485, + 0.008287536911666393, + 0.007835502736270428, + 0.006421663798391819, + 0.012073100544512272, + 0.0094349579885602, + 0.008040420711040497, + 0.01394741702824831, + 0.008499911986291409, + 0.009066992439329624, + 0.0044041406363248825, + 0.006174888927489519, + 0.007171413861215115, + 0.00606101518496871, + 0.0061169336549937725, + 0.005570732057094574, + 0.006553261540830135, + 0.006925488356500864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 6 + ], + "selected_reward": -0.1029801145195961 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076632569544017315 + ], + "opposite_logprob_tokens": [ + -4.882663249969482 + ], + "logprob_reward": -0.0076632569544017315, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076632569544017315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08945667743682861, + 0.029815347865223885, + 0.12998425960540771, + 0.028956541791558266, + 0.11352025717496872, + 0.42827826738357544, + 0.04985877498984337, + 0.04778103530406952, + 0.03569163382053375, + 0.04665716737508774 + ], + "credit_map": [ + 0.013170216232538223, + 0.0071155233308672905, + 0.021720673888921738, + 0.006941804196685553, + 0.01482397224754095, + 0.010389354079961777, + 0.0069664050824940205, + 0.008328728377819061, + 0.0022820981685072184, + 0.005705965217202902, + 0.003506579203531146, + 0.003368425415828824, + 0.004808604717254639, + 0.0026775372680276632, + 0.003459593513980508, + 0.004006543196737766, + 0.011705405078828335, + 0.011797585524618626, + 0.018375633284449577, + 0.010683566331863403, + 0.009979572147130966, + 0.04052053391933441, + 0.00764845497906208, + 0.019273502752184868, + 0.0029851107392460108, + 0.003305757185444236, + 0.003573773428797722, + 0.005251083988696337, + 0.0038225771859288216, + 0.003935455344617367, + 0.002730622421950102, + 0.0033521605655550957, + 0.010200652293860912, + 0.004925610963255167, + 0.039811551570892334, + 0.008967830799520016, + 0.005220664199441671, + 0.011375486850738525, + 0.00508820079267025, + 0.027930259704589844, + 0.047940000891685486, + 0.009081675671041012, + 0.17865562438964844, + 0.02749628573656082, + 0.010877806693315506, + 0.0437358133494854, + 0.010062308982014656, + 0.10042877495288849, + 0.006493088323622942, + 0.004919902887195349, + 0.014651785604655743, + 0.004899195861071348, + 0.003216787939891219, + 0.0038760339375585318, + 0.0027708697598427534, + 0.009031114168465137, + 0.011980404146015644, + 0.0038948282599449158, + 0.006733590271323919, + 0.004479486029595137, + 0.0045669106766581535, + 0.0049069710075855255, + 0.003559048054739833, + 0.007659796625375748, + 0.0032973680645227432, + 0.001847335952334106, + 0.00577617809176445, + 0.004405525978654623, + 0.004077814985066652, + 0.006136332638561726, + 0.005112337414175272, + 0.005038740113377571, + 0.0032983128912746906, + 0.005931907799094915, + 0.005280010402202606, + 0.010162349790334702, + 0.004745598882436752, + 0.005563725251704454, + 0.005927399732172489, + 0.005747863557189703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.0076632569544017315 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0891101285815239 + ], + "opposite_logprob_tokens": [ + -2.4641101360321045 + ], + "logprob_reward": -0.0891101285815239, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0891101285815239, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14496709406375885, + 0.1779664009809494, + 0.26657283306121826, + 0.029573997482657433, + 0.08427207916975021, + 0.061519041657447815, + 0.03350182622671127, + 0.057654302567243576, + 0.06953364610671997, + 0.07443875819444656 + ], + "credit_map": [ + 0.026585182175040245, + 0.013813717290759087, + 0.02132859267294407, + 0.012680678628385067, + 0.021526630967855453, + 0.015377046540379524, + 0.011023147031664848, + 0.022632107138633728, + 0.01880001462996006, + 0.018925251439213753, + 0.030950868502259254, + 0.020072536543011665, + 0.022686375305056572, + 0.027830075472593307, + 0.019420381635427475, + 0.01928088441491127, + 0.02516844868659973, + 0.019982285797595978, + 0.03676336258649826, + 0.02397214062511921, + 0.021179353818297386, + 0.08627164363861084, + 0.016480982303619385, + 0.03675459325313568, + 0.004815632943063974, + 0.0027499436400830746, + 0.004765165504068136, + 0.0030078659765422344, + 0.003181847045198083, + 0.0045887744054198265, + 0.0027714227326214314, + 0.0036933457013219595, + 0.00947485025972128, + 0.0036022462882101536, + 0.026915863156318665, + 0.006792173720896244, + 0.005579404532909393, + 0.00847076065838337, + 0.004030735231935978, + 0.019406044855713844, + 0.006969383452087641, + 0.00230217631906271, + 0.02040034346282482, + 0.004542493727058172, + 0.001987530617043376, + 0.0116618312895298, + 0.0026637401897460222, + 0.010991539806127548, + 0.004786411300301552, + 0.003459623083472252, + 0.005799694452434778, + 0.0038319076411426067, + 0.0037397746928036213, + 0.0042842477560043335, + 0.0032796929590404034, + 0.004320472478866577, + 0.011777585372328758, + 0.005951716098934412, + 0.007879442535340786, + 0.005103881936520338, + 0.00499549088999629, + 0.00895383208990097, + 0.004477985668927431, + 0.008514368906617165, + 0.007352698128670454, + 0.0031595262698829174, + 0.006951232440769672, + 0.009137793444097042, + 0.007515390869230032, + 0.015912607312202454, + 0.011680417694151402, + 0.007823978550732136, + 0.00867719016969204, + 0.010082748718559742, + 0.010793928056955338, + 0.009652922861278057, + 0.007253940217196941, + 0.009482747875154018, + 0.009636121802031994, + 0.008859161287546158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0891101285815239 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014207405038177967 + ], + "opposite_logprob_tokens": [ + -4.264207363128662 + ], + "logprob_reward": -0.014207405038177967, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014207405038177967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0622735470533371, + 0.2219889611005783, + 0.177449032664299, + 0.025550007820129395, + 0.07318522036075592, + 0.1616731882095337, + 0.07760053128004074, + 0.13564489781856537, + 0.04429590329527855, + 0.020338721573352814 + ], + "credit_map": [ + 0.011083661578595638, + 0.005752574652433395, + 0.012268122285604477, + 0.004410764668136835, + 0.007670738268643618, + 0.00995947327464819, + 0.00560201657935977, + 0.005526198074221611, + 0.021867409348487854, + 0.022097276523709297, + 0.03818482160568237, + 0.024486280977725983, + 0.03194192424416542, + 0.03570292517542839, + 0.02251337096095085, + 0.025194957852363586, + 0.01670253463089466, + 0.012071624398231506, + 0.02755705639719963, + 0.01701466180384159, + 0.013788780197501183, + 0.05406384542584419, + 0.011297387070953846, + 0.02495313435792923, + 0.0033494955860078335, + 0.003475538920611143, + 0.002395804040133953, + 0.0033240048214793205, + 0.004303023684769869, + 0.0029716421850025654, + 0.00283107184804976, + 0.002899427432566881, + 0.0061154961585998535, + 0.0037826572079211473, + 0.023820068687200546, + 0.006958295591175556, + 0.0034541471395641565, + 0.006697703618556261, + 0.0048163230530917645, + 0.017540521919727325, + 0.017758864909410477, + 0.008143981918692589, + 0.019968362525105476, + 0.012992885895073414, + 0.008763611316680908, + 0.057799339294433594, + 0.010024632327258587, + 0.02622150257229805, + 0.015894494950771332, + 0.010172292590141296, + 0.014072580263018608, + 0.008503171615302563, + 0.006352333817631006, + 0.006123407278209925, + 0.004854286089539528, + 0.011627965606749058, + 0.019729074090719223, + 0.018339406698942184, + 0.018892252817749977, + 0.016137517988681793, + 0.01976294443011284, + 0.011872312985360622, + 0.014081763103604317, + 0.016829626634716988, + 0.004365420900285244, + 0.00411121966317296, + 0.008013023994863033, + 0.005547989625483751, + 0.004894932731986046, + 0.008141012862324715, + 0.004472250118851662, + 0.004750049207359552, + 0.001983271911740303, + 0.0031691209878772497, + 0.0025195791386067867, + 0.002755376510322094, + 0.002360385376960039, + 0.0021999534219503403, + 0.002698680153116584, + 0.002652354771271348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.014207405038177967 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812197223305702 + ], + "opposite_logprob_tokens": [ + -3.2881219387054443 + ], + "logprob_reward": -0.03812197223305702, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812197223305702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11983134597539902, + 0.05562226101756096, + 0.132554292678833, + 0.04075450077652931, + 0.06911076605319977, + 0.13849501311779022, + 0.12750771641731262, + 0.07141725718975067, + 0.16517925262451172, + 0.07952757924795151 + ], + "credit_map": [ + 0.018003322184085846, + 0.011178758926689625, + 0.02672259509563446, + 0.012493036687374115, + 0.012921405024826527, + 0.011945761740207672, + 0.011160217225551605, + 0.01540625374764204, + 0.006596995517611504, + 0.00556138763204217, + 0.010369936935603619, + 0.00541450921446085, + 0.007755085360258818, + 0.008739667944610119, + 0.005585604812949896, + 0.0055990698747336864, + 0.013886761851608753, + 0.007893499918282032, + 0.029591049998998642, + 0.010869535617530346, + 0.008953018113970757, + 0.029483461752533913, + 0.007939101196825504, + 0.023937858641147614, + 0.004428907763212919, + 0.005385750904679298, + 0.007643197197467089, + 0.004767471458762884, + 0.0034034207928925753, + 0.006518458016216755, + 0.003559278091415763, + 0.005048015154898167, + 0.006310746539384127, + 0.003981107845902443, + 0.0217282734811306, + 0.006322566419839859, + 0.0051810890436172485, + 0.006698633544147015, + 0.0043502626940608025, + 0.014538091607391834, + 0.014367803931236267, + 0.005151854362338781, + 0.043871939182281494, + 0.008887790143489838, + 0.005927310790866613, + 0.028705880045890808, + 0.0059103346429765224, + 0.02567209117114544, + 0.014907054603099823, + 0.008884887211024761, + 0.040623392909765244, + 0.010906401090323925, + 0.005018712021410465, + 0.012177543714642525, + 0.005562945734709501, + 0.029426777735352516, + 0.019252628087997437, + 0.007385730277746916, + 0.011184712871909142, + 0.005560688208788633, + 0.006752903573215008, + 0.006087522953748703, + 0.0048181526362895966, + 0.010374920442700386, + 0.01476867962628603, + 0.013062793761491776, + 0.02661329135298729, + 0.02028076723217964, + 0.017234107479453087, + 0.03545498102903366, + 0.018290234729647636, + 0.01947440207004547, + 0.005484393797814846, + 0.009389985352754593, + 0.013444330543279648, + 0.015281733125448227, + 0.007980568334460258, + 0.009148307144641876, + 0.010365412570536137, + 0.008432851172983646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 2 + ], + "selected_reward": -0.03812197223305702 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera smoothly follows and locks onto the magnificent head of the mythical bird, Suzaku. Each of its feathers is insubstantial, a fusion of blazing vermilion light and vibrant golden flames, forming a flaming, ever-shifting crest. Its eyes, like two miniature stars, burn with pure golden light, revealing an ancient and majestic divinity. Its beak, sharp and pure, is composed of crystallized light, gleaming like a diamond. The background is the vast expanse of space, filled with slowly rotating nebulae composed of deep purple and indigo gas, within which countless tiny stars twinkle like dust. Suzaku gracefully raises its slender neck, its light-forged beak slowly opening, and utters a cry that resounds through the heavens yet carries visually through the vacuum of space. A golden, circular wave of energy suddenly emanates from its mouth, instantly illuminating the surrounding nebula. At the climax of this chirping, its body, composed entirely of flames and light, began to become illusory and translucent, its fiery outline gradually dissolved and blurred, and finally merged perfectly with the surrounding colorful nebula. Through its almost transparent body, one can clearly see the distant stars and dust belts slowly flowing behind it, creating a grand, sacred visual illusion that swayed between reality and illusion, full of ethereality and awe.", + "seed": 14670, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_3/index_2205_seed_14670_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017468959419056775, + "vlm_logprob": -0.017468959419056775, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.017468959419056775, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.052362505346536636 + ], + "opposite_logprob_tokens": [ + -3.0523624420166016 + ], + "logprob_reward": -0.052362505346536636, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.052362505346536636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21008023619651794, + 0.09434482455253601, + 0.03645950183272362, + 0.05626629665493965, + 0.023865049704909325, + 0.050534069538116455, + 0.21187293529510498, + 0.10656782239675522, + 0.07688114047050476, + 0.13312813639640808 + ], + "credit_map": [ + 0.050815705209970474, + 0.029546473175287247, + 0.02807241678237915, + 0.018513448536396027, + 0.013419986702501774, + 0.020674074068665504, + 0.01874469593167305, + 0.030293431133031845, + 0.013442306779325008, + 0.0098950220271945, + 0.009815667755901814, + 0.009180454537272453, + 0.01431713905185461, + 0.015281140804290771, + 0.008254059590399265, + 0.014159039594233036, + 0.009381203912198544, + 0.0055190459825098515, + 0.004349999595433474, + 0.0034734492655843496, + 0.0032020770013332367, + 0.0032234801910817623, + 0.002250027609989047, + 0.005060216411948204, + 0.007092892657965422, + 0.005589575972408056, + 0.007931703701615334, + 0.008508614264428616, + 0.00618169829249382, + 0.007513315416872501, + 0.0057220617309212685, + 0.007726436946541071, + 0.003318739589303732, + 0.0018712850287556648, + 0.0021911561489105225, + 0.0030336047057062387, + 0.0038416048046201468, + 0.0030219017062336206, + 0.0029817030299454927, + 0.0036050553899258375, + 0.0049307276494801044, + 0.0030350128654390574, + 0.009588385932147503, + 0.009898997843265533, + 0.005063357297331095, + 0.0058082579635083675, + 0.0051441192626953125, + 0.007065212354063988, + 0.025423821061849594, + 0.04385879635810852, + 0.028429685160517693, + 0.02346039190888405, + 0.033536896109580994, + 0.015945250168442726, + 0.019154079258441925, + 0.022064007818698883, + 0.008813449181616306, + 0.012253946624696255, + 0.008566252887248993, + 0.0069520664401352406, + 0.009376286529004574, + 0.033284012228250504, + 0.007428152486681938, + 0.019893651828169823, + 0.010072181932628155, + 0.006223674397915602, + 0.014868803322315216, + 0.008424588479101658, + 0.010853809304535389, + 0.008060195483267307, + 0.010416959412395954, + 0.007960928604006767, + 0.01735556870698929, + 0.012951279059052467, + 0.025771232321858406, + 0.01513457577675581, + 0.010730103589594364, + 0.014655680395662785, + 0.01620711386203766, + 0.02032257616519928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.052362505346536636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003214908065274358 + ], + "opposite_logprob_tokens": [ + -5.7532148361206055 + ], + "logprob_reward": -0.003214908065274358, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003214908065274358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11978422850370407, + 0.07777194678783417, + 0.11579272896051407, + 0.04647516459226608, + 0.02999136969447136, + 0.19465741515159607, + 0.04171789810061455, + 0.032285887748003006, + 0.1356445848941803, + 0.20587879419326782 + ], + "credit_map": [ + 0.032213930040597916, + 0.018310900777578354, + 0.013202864676713943, + 0.01025993749499321, + 0.011556200683116913, + 0.014517099596560001, + 0.006693183910101652, + 0.01303010992705822, + 0.007443026639521122, + 0.00763365626335144, + 0.008489290252327919, + 0.00780725572258234, + 0.012925976887345314, + 0.012267020530998707, + 0.007608334068208933, + 0.013597384095191956, + 0.020459657534956932, + 0.015334803611040115, + 0.014324653893709183, + 0.010671797208487988, + 0.013939538970589638, + 0.01665772870182991, + 0.008496766909956932, + 0.01590777188539505, + 0.00528295012190938, + 0.006570708472281694, + 0.006955580785870552, + 0.005149614531546831, + 0.004634995479136705, + 0.00555213401094079, + 0.0046790349297225475, + 0.007650150451809168, + 0.004781142808496952, + 0.002324641216546297, + 0.0030698152258992195, + 0.00318127335049212, + 0.00399177148938179, + 0.003912508022040129, + 0.004235723055899143, + 0.00449449522420764, + 0.012125689536333084, + 0.014967055991292, + 0.04382852092385292, + 0.033116962760686874, + 0.0169206615537405, + 0.027359090745449066, + 0.018280182033777237, + 0.02805924415588379, + 0.0058623189106583595, + 0.006101496517658234, + 0.005950686987489462, + 0.004543719347566366, + 0.004201036877930164, + 0.0049635968171060085, + 0.0044002775102853775, + 0.005694764666259289, + 0.005680008791387081, + 0.00416473438963294, + 0.0038419507909566164, + 0.003148512914776802, + 0.003317333525046706, + 0.0046659172512590885, + 0.0034105449449270964, + 0.004056883510202169, + 0.015306727029383183, + 0.012679137289524078, + 0.021787049248814583, + 0.02047555521130562, + 0.014671936631202698, + 0.014990831725299358, + 0.018914910033345222, + 0.016818435862660408, + 0.014525004662573338, + 0.01825018599629402, + 0.02135528437793255, + 0.028503665700554848, + 0.02939751371741295, + 0.03907829895615578, + 0.026837846264243126, + 0.027930987998843193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.003214908065274358 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024922997690737247 + ], + "opposite_logprob_tokens": [ + -6.002492427825928 + ], + "logprob_reward": -0.0024922997690737247, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024922997690737247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12583020329475403, + 0.12639294564723969, + 0.08826218545436859, + 0.06037886440753937, + 0.029906487092375755, + 0.13916917145252228, + 0.06613907217979431, + 0.07309748977422714, + 0.15148581564426422, + 0.1393376737833023 + ], + "credit_map": [ + 0.01769394241273403, + 0.0170889925211668, + 0.01450598705559969, + 0.014183911494910717, + 0.015921056270599365, + 0.01896948739886284, + 0.011153495870530605, + 0.016313323751091957, + 0.023099631071090698, + 0.01564326137304306, + 0.013430717401206493, + 0.014714778400957584, + 0.015318631194531918, + 0.013727187179028988, + 0.011206544935703278, + 0.019252195954322815, + 0.022655298933386803, + 0.012088606134057045, + 0.009948981925845146, + 0.006119113881140947, + 0.008883222006261349, + 0.010738352313637733, + 0.005800287704914808, + 0.01202832069247961, + 0.006364389322698116, + 0.006778760347515345, + 0.007410191465169191, + 0.008561046794056892, + 0.008404706604778767, + 0.008203230798244476, + 0.005921319127082825, + 0.008735219947993755, + 0.004702568054199219, + 0.0035724337212741375, + 0.002382023958489299, + 0.002773728221654892, + 0.005784350447356701, + 0.0033228660468012094, + 0.003561378223821521, + 0.003807137720286846, + 0.02297944761812687, + 0.018046893179416656, + 0.021320879459381104, + 0.015941213816404343, + 0.015446895733475685, + 0.014834900386631489, + 0.014361802488565445, + 0.01623714715242386, + 0.009039046242833138, + 0.010159551165997982, + 0.009493106976151466, + 0.007487742230296135, + 0.00795245822519064, + 0.007676750421524048, + 0.0057759638875722885, + 0.008554451167583466, + 0.009602965787053108, + 0.014206920750439167, + 0.007528366986662149, + 0.007380632683634758, + 0.01137856300920248, + 0.009575010277330875, + 0.006742872763425112, + 0.00668215611949563, + 0.028331482782959938, + 0.016609055921435356, + 0.018447846174240112, + 0.018860245123505592, + 0.01647689752280712, + 0.016130225732922554, + 0.020334094762802124, + 0.016295962035655975, + 0.018679296597838402, + 0.011771185323596, + 0.0178863313049078, + 0.010736457072198391, + 0.023656707257032394, + 0.023141775280237198, + 0.014930738136172295, + 0.018535176292061806 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.0024922997690737247 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080989614129066 + ], + "opposite_logprob_tokens": [ + -4.511081218719482 + ], + "logprob_reward": -0.011080989614129066, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080989614129066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029345914721488953, + 0.05254193767905235, + 0.02474336139857769, + 0.019740544259548187, + 0.022488947957754135, + 0.4127258062362671, + 0.04504755139350891, + 0.03618789464235306, + 0.07357890158891678, + 0.2835991084575653 + ], + "credit_map": [ + 0.004942229017615318, + 0.0041474103927612305, + 0.003954135347157717, + 0.0032569009345024824, + 0.002154788700863719, + 0.004332867916673422, + 0.0027539217844605446, + 0.0038036624900996685, + 0.00849505327641964, + 0.006006275769323111, + 0.006324756890535355, + 0.005259873811155558, + 0.004819913301616907, + 0.00683964230120182, + 0.005927346181124449, + 0.008869077078998089, + 0.003112412989139557, + 0.002719879848882556, + 0.0041410718113183975, + 0.003139809239655733, + 0.0030230877455323935, + 0.002726460574194789, + 0.002092868322506547, + 0.003787772264331579, + 0.002534714061766863, + 0.002036088379099965, + 0.0027879253029823303, + 0.0024996011052280664, + 0.0021947433706372976, + 0.002657645847648382, + 0.0021090370137244463, + 0.002920790808275342, + 0.0027400103863328695, + 0.0014851009473204613, + 0.0028113548178225756, + 0.0028000439051538706, + 0.002160493517294526, + 0.002094453666359186, + 0.004996738396584988, + 0.00340075115673244, + 0.026127872988581657, + 0.028007736429572105, + 0.09234638512134552, + 0.08194534480571747, + 0.028747782111167908, + 0.06154775246977806, + 0.03290237486362457, + 0.061100564897060394, + 0.004229679238051176, + 0.004618288483470678, + 0.00445776479318738, + 0.0036006325390189886, + 0.005631802137941122, + 0.010313935577869415, + 0.005602920427918434, + 0.006592531688511372, + 0.004068337846547365, + 0.005994394887238741, + 0.003252799389883876, + 0.004563494585454464, + 0.004501881077885628, + 0.006086729001253843, + 0.0033646235242486, + 0.004355634097009897, + 0.007908723317086697, + 0.006438004784286022, + 0.011065809987485409, + 0.010914265178143978, + 0.008373808115720749, + 0.00995913241058588, + 0.008105902001261711, + 0.010813256725668907, + 0.02436661906540394, + 0.02029474452137947, + 0.03167714923620224, + 0.022217607125639915, + 0.053909335285425186, + 0.06395769864320755, + 0.026133622974157333, + 0.04104235768318176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.011080989614129066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819409430027008 + ], + "opposite_logprob_tokens": [ + -4.018194198608398 + ], + "logprob_reward": -0.01819409430027008, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819409430027008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05397375673055649, + 0.03458147123456001, + 0.04311076179146767, + 0.021690599620342255, + 0.016768405213952065, + 0.5287696719169617, + 0.01790662109851837, + 0.05807352811098099, + 0.10227702558040619, + 0.12284820526838303 + ], + "credit_map": [ + 0.014233105815947056, + 0.009171395562589169, + 0.005975073203444481, + 0.004822049289941788, + 0.005001821555197239, + 0.005748382769525051, + 0.00341235613450408, + 0.005609572399407625, + 0.004461926408112049, + 0.004318871069699526, + 0.0043288483284413815, + 0.005020263604819775, + 0.0038964173290878534, + 0.004342463333159685, + 0.0026520169340074062, + 0.005560663063079119, + 0.012867129407823086, + 0.006557769607752562, + 0.005055837333202362, + 0.005372880958020687, + 0.002365772845223546, + 0.002653059083968401, + 0.0018674195744097233, + 0.006370891816914082, + 0.0018085031770169735, + 0.002547844313085079, + 0.002861638553440571, + 0.003821961348876357, + 0.0017190893413498998, + 0.003523992607370019, + 0.00214497372508049, + 0.0032625971361994743, + 0.003562785917893052, + 0.0015796783845871687, + 0.0017933162162080407, + 0.0012434856034815311, + 0.0020536514930427074, + 0.0021116072311997414, + 0.0020105629228055477, + 0.002413318259641528, + 0.03276170790195465, + 0.03774239122867584, + 0.11887110769748688, + 0.10413467884063721, + 0.03831254318356514, + 0.07650990784168243, + 0.042414430528879166, + 0.07802289724349976, + 0.002239102963358164, + 0.0020997084211558104, + 0.002547767711803317, + 0.0017067276639863849, + 0.0017723131459206343, + 0.0030314824543893337, + 0.0021140447352081537, + 0.002395474351942539, + 0.007040645461529493, + 0.010822206735610962, + 0.006520441733300686, + 0.005277808755636215, + 0.007197789382189512, + 0.007011505775153637, + 0.0055378093384206295, + 0.00866532139480114, + 0.015715831890702248, + 0.011104926466941833, + 0.012959067709743977, + 0.013980015181005001, + 0.010990300215780735, + 0.011010054498910904, + 0.014366596937179565, + 0.01215022336691618, + 0.013668496161699295, + 0.01661735214293003, + 0.018140092492103577, + 0.010072944685816765, + 0.018305424600839615, + 0.013537511229515076, + 0.01721843145787716, + 0.015287946909666061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.01819409430027008 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera smoothly follows and locks onto the magnificent head of the mythical bird, Suzaku. Each of its feathers is insubstantial, a fusion of blazing vermilion light and vibrant golden flames, forming a flaming, ever-shifting crest. Its eyes, like two miniature stars, burn with pure golden light, revealing an ancient and majestic divinity. Its beak, sharp and pure, is composed of crystallized light, gleaming like a diamond. The background is the vast expanse of space, filled with slowly rotating nebulae composed of deep purple and indigo gas, within which countless tiny stars twinkle like dust. Suzaku gracefully raises its slender neck, its light-forged beak slowly opening, and utters a cry that resounds through the heavens yet carries visually through the vacuum of space. A golden, circular wave of energy suddenly emanates from its mouth, instantly illuminating the surrounding nebula. At the climax of this chirping, its body, composed entirely of flames and light, began to become illusory and translucent, its fiery outline gradually dissolved and blurred, and finally merged perfectly with the surrounding colorful nebula. Through its almost transparent body, one can clearly see the distant stars and dust belts slowly flowing behind it, creating a grand, sacred visual illusion that swayed between reality and illusion, full of ethereality and awe.", + "seed": 14671, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_3/index_2205_seed_14671_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0195654870942235, + "vlm_logprob": -0.0195654870942235, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.0195654870942235, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03398067504167557 + ], + "opposite_logprob_tokens": [ + -3.65898060798645 + ], + "logprob_reward": -0.03398067504167557, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03398067504167557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06741461902856827, + 0.138023242354393, + 0.20827490091323853, + 0.24104806780815125, + 0.06009931489825249, + 0.016160814091563225, + 0.028017576783895493, + 0.06035373732447624, + 0.03531638905405998, + 0.14529132843017578 + ], + "credit_map": [ + 0.007873165421187878, + 0.008836658671498299, + 0.0110551742836833, + 0.00840065162628889, + 0.008604944683611393, + 0.006635637022554874, + 0.007613983005285263, + 0.008394408039748669, + 0.012599091045558453, + 0.027490530163049698, + 0.01650395803153515, + 0.015097551047801971, + 0.01927986554801464, + 0.01681380346417427, + 0.012945042923092842, + 0.017293397337198257, + 0.02062482386827469, + 0.01252515148371458, + 0.029157640412449837, + 0.009744023904204369, + 0.017782552167773247, + 0.038320139050483704, + 0.012998761609196663, + 0.06712181121110916, + 0.015687311068177223, + 0.022329142317175865, + 0.0707244873046875, + 0.016805527731776237, + 0.01767485961318016, + 0.018688039854168892, + 0.011001696810126305, + 0.06813700497150421, + 0.00544431246817112, + 0.012542324140667915, + 0.0039934734813869, + 0.008097223937511444, + 0.00861845351755619, + 0.007695503532886505, + 0.006312153302133083, + 0.007395871914923191, + 0.004649654496461153, + 0.002107897074893117, + 0.002010682597756386, + 0.0016932138241827488, + 0.0009969339007511735, + 0.0012092655524611473, + 0.0010827534133568406, + 0.0024104139301925898, + 0.004544536117464304, + 0.00528333755210042, + 0.002864038571715355, + 0.003620800329372287, + 0.0030454066582024097, + 0.0020011398009955883, + 0.002295931801199913, + 0.004362387116998434, + 0.005765213631093502, + 0.011651642620563507, + 0.00430833175778389, + 0.009156537242233753, + 0.009690627455711365, + 0.004922014661133289, + 0.007562668994069099, + 0.007296693045645952, + 0.003250419395044446, + 0.004391878377646208, + 0.006829292047768831, + 0.00427884329110384, + 0.003419766202569008, + 0.00360822188667953, + 0.0037554611917585135, + 0.005782505497336388, + 0.010359104722738266, + 0.007184417452663183, + 0.028224032372236252, + 0.007634459063410759, + 0.010430101305246353, + 0.008400374092161655, + 0.0511530339717865, + 0.0219058059155941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.03398067504167557 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005274783819913864 + ], + "opposite_logprob_tokens": [ + -5.255274772644043 + ], + "logprob_reward": -0.005274783819913864, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005274783819913864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09117579460144043, + 0.06029567867517471, + 0.34847554564476013, + 0.049244754016399384, + 0.03744716942310333, + 0.03630761057138443, + 0.038437630981206894, + 0.0373249277472496, + 0.08426681160926819, + 0.2170240432024002 + ], + "credit_map": [ + 0.008949952200055122, + 0.01196963619440794, + 0.015215576626360416, + 0.009185636416077614, + 0.010650137439370155, + 0.010595344007015228, + 0.011430120095610619, + 0.01317939069122076, + 0.0059864395298063755, + 0.007668154779821634, + 0.012095886282622814, + 0.0066207353956997395, + 0.007868574932217598, + 0.0065945531241595745, + 0.005706630647182465, + 0.007754703983664513, + 0.026861373335123062, + 0.029000816866755486, + 0.045720748603343964, + 0.011635677888989449, + 0.026292042806744576, + 0.058490898460149765, + 0.021163586527109146, + 0.12931041419506073, + 0.006462601479142904, + 0.006522315554320812, + 0.007138629909604788, + 0.0050825923681259155, + 0.007613382767885923, + 0.006327680312097073, + 0.003591306274756789, + 0.006506248377263546, + 0.0059346482157707214, + 0.007031125947833061, + 0.003426825860515237, + 0.0036951194051653147, + 0.0039053403306752443, + 0.003382069058716297, + 0.004631310701370239, + 0.005440731067210436, + 0.008973628282546997, + 0.005231834016740322, + 0.004035940393805504, + 0.0032713173422962427, + 0.003043007105588913, + 0.00258596520870924, + 0.002910063834860921, + 0.006255855318158865, + 0.008971095085144043, + 0.005385981872677803, + 0.003172754542902112, + 0.002645952859893441, + 0.0034823112655431032, + 0.00392565643414855, + 0.004073732998222113, + 0.006780147086828947, + 0.0037956838496029377, + 0.006746719125658274, + 0.0031828389037400484, + 0.0056247529573738575, + 0.005037193652242422, + 0.003219887614250183, + 0.005266582127660513, + 0.004451270215213299, + 0.008128573186695576, + 0.010004334151744843, + 0.009023856371641159, + 0.010187768377363682, + 0.011845780536532402, + 0.01124837901443243, + 0.010431970469653606, + 0.013396140187978745, + 0.05425012856721878, + 0.015055481344461441, + 0.04267401993274689, + 0.015499290078878403, + 0.015914782881736755, + 0.022817056626081467, + 0.029084257781505585, + 0.021729031577706337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.005274783819913864 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004644677042961121 + ], + "opposite_logprob_tokens": [ + -5.379644870758057 + ], + "logprob_reward": -0.004644677042961121, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004644677042961121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06258026510477066, + 0.07464688271284103, + 0.39962852001190186, + 0.10395444184541702, + 0.04164323955774307, + 0.05421055480837822, + 0.04287801682949066, + 0.047986432909965515, + 0.08563607931137085, + 0.08683553338050842 + ], + "credit_map": [ + 0.011254771612584591, + 0.010921597480773926, + 0.009373469278216362, + 0.006622134707868099, + 0.007327670697122812, + 0.004882226698100567, + 0.006335146725177765, + 0.0058632586151361465, + 0.011905313469469547, + 0.010951011441648006, + 0.012262486852705479, + 0.008821418508887291, + 0.009877657517790794, + 0.007641936186701059, + 0.0049987309612333775, + 0.00818832777440548, + 0.031685855239629745, + 0.02958361804485321, + 0.04948227107524872, + 0.01275603286921978, + 0.02787487395107746, + 0.06902778893709183, + 0.022599313408136368, + 0.15661878883838654, + 0.008765901438891888, + 0.008677261881530285, + 0.029752131551504135, + 0.006264118943363428, + 0.00754957739263773, + 0.00831758975982666, + 0.0044461204670369625, + 0.030181752517819405, + 0.004686849191784859, + 0.006426430772989988, + 0.006398127414286137, + 0.00395140890032053, + 0.0039614103734493256, + 0.003775537246838212, + 0.0031772644724696875, + 0.009266216307878494, + 0.011523816734552383, + 0.007830370217561722, + 0.005281310062855482, + 0.005440800916403532, + 0.004920952953398228, + 0.004715878516435623, + 0.007547945715487003, + 0.006949486210942268, + 0.008155944757163525, + 0.008037866093218327, + 0.004508649464696646, + 0.005538344848901033, + 0.004195225890725851, + 0.003499791957437992, + 0.0030772676691412926, + 0.005864930339157581, + 0.009628156200051308, + 0.006386020220816135, + 0.00529849249869585, + 0.005941540468484163, + 0.004563563968986273, + 0.0036905300803482533, + 0.003963838331401348, + 0.008514296263456345, + 0.014202576130628586, + 0.0067571368999779224, + 0.021574044600129128, + 0.008952110074460506, + 0.005164529196918011, + 0.005130756180733442, + 0.007265319582074881, + 0.016589613631367683, + 0.01806187629699707, + 0.006055506411939859, + 0.006227490026503801, + 0.007069113664329052, + 0.007265032269060612, + 0.00724900234490633, + 0.026593968272209167, + 0.008313556201756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -0.004644677042961121 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256284959614277 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256284959614277, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256284959614277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06135424226522446, + 0.08108845353126526, + 0.08477392792701721, + 0.04186197370290756, + 0.09982620179653168, + 0.057748839259147644, + 0.041632894426584244, + 0.020652228966355324, + 0.13572992384433746, + 0.3753312826156616 + ], + "credit_map": [ + 0.013627407141029835, + 0.007676968351006508, + 0.010351394303143024, + 0.00617544911801815, + 0.005385715514421463, + 0.0063321287743747234, + 0.005703796166926622, + 0.006101380102336407, + 0.006842717062681913, + 0.018243471160531044, + 0.008442705497145653, + 0.010880650021135807, + 0.010266168974339962, + 0.008624855428934097, + 0.006633146200329065, + 0.011154741048812866, + 0.009642905555665493, + 0.013856669887900352, + 0.008555863983929157, + 0.004000927787274122, + 0.007550933863967657, + 0.013550403527915478, + 0.005993912927806377, + 0.021622309461236, + 0.006134644150733948, + 0.005173457320779562, + 0.00807307567447424, + 0.003644058480858803, + 0.0027837848756462336, + 0.002583883935585618, + 0.0022846211213618517, + 0.011184451170265675, + 0.013541271910071373, + 0.017846792936325073, + 0.00943241361528635, + 0.011896099895238876, + 0.011007608845829964, + 0.009271513670682907, + 0.011600778438150883, + 0.0152297243475914, + 0.01084045972675085, + 0.007939894683659077, + 0.004852645564824343, + 0.00886786263436079, + 0.005417210515588522, + 0.004394537303596735, + 0.00775873102247715, + 0.007677497807890177, + 0.005779835395514965, + 0.006756494753062725, + 0.002575316932052374, + 0.003502251347526908, + 0.006850393023341894, + 0.006378508638590574, + 0.0044379993341863155, + 0.0053520966321229935, + 0.003348949598148465, + 0.003688518423587084, + 0.002221326343715191, + 0.002544982358813286, + 0.002653544070199132, + 0.001825977466069162, + 0.0017842170782387257, + 0.002584712812677026, + 0.010985582135617733, + 0.013250806368887424, + 0.024136409163475037, + 0.014110726304352283, + 0.012132876552641392, + 0.01591171696782112, + 0.017168914899230003, + 0.028032897040247917, + 0.07040689885616302, + 0.023619256913661957, + 0.11247795075178146, + 0.02486559934914112, + 0.02433231845498085, + 0.03895304352045059, + 0.04398283734917641, + 0.03669336438179016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.005256284959614277 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867101460695267 + ], + "opposite_logprob_tokens": [ + -3.048671007156372 + ], + "logprob_reward": -0.04867101460695267, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867101460695267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05811277776956558, + 0.04543907195329666, + 0.07933614403009415, + 0.07900562882423401, + 0.08664579689502716, + 0.03127431496977806, + 0.06107070669531822, + 0.04411377012729645, + 0.16892914474010468, + 0.3460726737976074 + ], + "credit_map": [ + 0.014139887876808643, + 0.006976269651204348, + 0.009384551085531712, + 0.005646968260407448, + 0.0064095486886799335, + 0.0053506894037127495, + 0.004605492576956749, + 0.005599361378699541, + 0.004729569889605045, + 0.008052176795899868, + 0.005347943864762783, + 0.00468063447624445, + 0.005550728645175695, + 0.005492077209055424, + 0.005074703134596348, + 0.0065112304873764515, + 0.007394169922918081, + 0.008055015467107296, + 0.012815049849450588, + 0.003815671196207404, + 0.007781844586133957, + 0.014024006202816963, + 0.004988343920558691, + 0.02046203799545765, + 0.0075398655608296394, + 0.006249910220503807, + 0.021918417885899544, + 0.0054672821424901485, + 0.003943913616240025, + 0.005845421925187111, + 0.0041297078132629395, + 0.023911096155643463, + 0.011622406542301178, + 0.015837358310818672, + 0.009170997887849808, + 0.008857757784426212, + 0.008915570564568043, + 0.009264499880373478, + 0.00784193817526102, + 0.015135258436203003, + 0.0066191088408231735, + 0.00402836874127388, + 0.0037650451995432377, + 0.0023521564435213804, + 0.003279326716437936, + 0.003223706502467394, + 0.002482254756614566, + 0.005524343810975552, + 0.013313685543835163, + 0.007706735748797655, + 0.005488941445946693, + 0.00532887177541852, + 0.006368686910718679, + 0.006374876946210861, + 0.005793705116957426, + 0.010695195756852627, + 0.005339281167834997, + 0.007188855670392513, + 0.004958895966410637, + 0.00655582407489419, + 0.005500039551407099, + 0.004311665892601013, + 0.005675135646015406, + 0.00458406750112772, + 0.01887509785592556, + 0.01201513409614563, + 0.04942414537072182, + 0.01707877591252327, + 0.0089359600096941, + 0.011464214883744717, + 0.012026934884488583, + 0.03910885751247406, + 0.051103103905916214, + 0.019889872521162033, + 0.09901964664459229, + 0.022351963445544243, + 0.01952667534351349, + 0.026452118530869484, + 0.08331285417079926, + 0.024416418746113777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.04867101460695267 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning close-up, the camera smoothly follows and locks onto the magnificent head of the mythical bird, Suzaku. Each of its feathers is insubstantial, a fusion of blazing vermilion light and vibrant golden flames, forming a flaming, ever-shifting crest. Its eyes, like two miniature stars, burn with pure golden light, revealing an ancient and majestic divinity. Its beak, sharp and pure, is composed of crystallized light, gleaming like a diamond. The background is the vast expanse of space, filled with slowly rotating nebulae composed of deep purple and indigo gas, within which countless tiny stars twinkle like dust. Suzaku gracefully raises its slender neck, its light-forged beak slowly opening, and utters a cry that resounds through the heavens yet carries visually through the vacuum of space. A golden, circular wave of energy suddenly emanates from its mouth, instantly illuminating the surrounding nebula. At the climax of this chirping, its body, composed entirely of flames and light, began to become illusory and translucent, its fiery outline gradually dissolved and blurred, and finally merged perfectly with the surrounding colorful nebula. Through its almost transparent body, one can clearly see the distant stars and dust belts slowly flowing behind it, creating a grand, sacred visual illusion that swayed between reality and illusion, full of ethereality and awe.", + "seed": 14672, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_3/index_2205_seed_14672_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0182079772464931, + "vlm_logprob": -0.0182079772464931, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.0182079772464931, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03747916221618652 + ], + "opposite_logprob_tokens": [ + -3.5374791622161865 + ], + "logprob_reward": -0.03747916221618652, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03747916221618652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1185070276260376, + 0.16062721610069275, + 0.030875956639647484, + 0.28466835618019104, + 0.1173916682600975, + 0.0321795791387558, + 0.04786005616188049, + 0.08372226357460022, + 0.08480639010667801, + 0.03936149179935455 + ], + "credit_map": [ + 0.013050359673798084, + 0.0159226693212986, + 0.0191046055406332, + 0.010221977718174458, + 0.019957292824983597, + 0.015494722872972488, + 0.007749143056571484, + 0.01700626313686371, + 0.04286697134375572, + 0.01853954792022705, + 0.023906316608190536, + 0.012581326998770237, + 0.01343251671642065, + 0.020803987979888916, + 0.008492175489664078, + 0.02000436931848526, + 0.002355609554797411, + 0.004732299130409956, + 0.0037571918219327927, + 0.003907147794961929, + 0.0038922608364373446, + 0.002979614306241274, + 0.0035662304144352674, + 0.005685601849108934, + 0.020382124930620193, + 0.05700595676898956, + 0.03458398953080177, + 0.03074265830218792, + 0.03959224373102188, + 0.038252681493759155, + 0.02849169448018074, + 0.03561701253056526, + 0.009030896238982677, + 0.019401896744966507, + 0.01299897488206625, + 0.011711683124303818, + 0.011684637516736984, + 0.011469142511487007, + 0.010233406908810139, + 0.030861029401421547, + 0.0038441389333456755, + 0.0033689578995108604, + 0.004465584177523851, + 0.0033984235487878323, + 0.0052172220312058926, + 0.003992288839071989, + 0.00346835027448833, + 0.004424613900482655, + 0.004137704614549875, + 0.007667824625968933, + 0.0062879351899027824, + 0.007714518811553717, + 0.006002019625157118, + 0.0060283346101641655, + 0.004125145263969898, + 0.005896576680243015, + 0.00948729645460844, + 0.007303155027329922, + 0.01088312454521656, + 0.017649704590439796, + 0.006766492035239935, + 0.017417563125491142, + 0.005294839385896921, + 0.008920084685087204, + 0.007987039163708687, + 0.009165244176983833, + 0.011361456476151943, + 0.01605519838631153, + 0.009270938113331795, + 0.014767651446163654, + 0.008897369727492332, + 0.0073014902882277966, + 0.004915389232337475, + 0.0025339459534734488, + 0.005752220284193754, + 0.0063860416412353516, + 0.003849111497402191, + 0.00972993578761816, + 0.0028409853111952543, + 0.0033538630232214928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.03747916221618652 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003214194905012846 + ], + "opposite_logprob_tokens": [ + -5.753214359283447 + ], + "logprob_reward": -0.003214194905012846, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003214194905012846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0803818628191948, + 0.17361925542354584, + 0.03693072497844696, + 0.27828022837638855, + 0.07722169160842896, + 0.04713922366499901, + 0.02434374764561653, + 0.07035087794065475, + 0.131307452917099, + 0.08042489737272263 + ], + "credit_map": [ + 0.009276892989873886, + 0.01064030360430479, + 0.014288113452494144, + 0.007592462003231049, + 0.01120032463222742, + 0.008887756615877151, + 0.006288770120590925, + 0.012207238003611565, + 0.045384105294942856, + 0.02058578096330166, + 0.02552257478237152, + 0.013865898363292217, + 0.013245316222310066, + 0.022572213783860207, + 0.01050913892686367, + 0.02193421684205532, + 0.0030991733074188232, + 0.006445870269089937, + 0.0035715755075216293, + 0.004899265710264444, + 0.004580321256071329, + 0.003464052453637123, + 0.006386845372617245, + 0.004483621101826429, + 0.050589289516210556, + 0.04402395710349083, + 0.04055829346179962, + 0.022565828636288643, + 0.025159813463687897, + 0.03636366128921509, + 0.024052396416664124, + 0.03496699780225754, + 0.007542829029262066, + 0.005385379306972027, + 0.01067833136767149, + 0.006411639973521233, + 0.005223523825407028, + 0.013112579472362995, + 0.005709552671760321, + 0.02315785363316536, + 0.0049873776733875275, + 0.003195173107087612, + 0.0049246870912611485, + 0.008803741075098515, + 0.0034267008304595947, + 0.006630822084844112, + 0.005602104123681784, + 0.009568614885210991, + 0.0031192058231681585, + 0.003554942552000284, + 0.003416424384340644, + 0.0036437902599573135, + 0.0025131083093583584, + 0.002204533899202943, + 0.002192723099142313, + 0.0036990176886320114, + 0.008483820594847202, + 0.005425147712230682, + 0.011636648327112198, + 0.013328335247933865, + 0.006739303003996611, + 0.012139737606048584, + 0.004822362679988146, + 0.007775524631142616, + 0.01292829867452383, + 0.019580058753490448, + 0.02176336571574211, + 0.01873837225139141, + 0.01545544434338808, + 0.018617909401655197, + 0.010862449184060097, + 0.013361564837396145, + 0.010251977480947971, + 0.007609307300299406, + 0.01214400865137577, + 0.011479088105261326, + 0.0067765372805297375, + 0.018201638013124466, + 0.006406520958989859, + 0.007555820979177952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.003214194905012846 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009788726456463337 + ], + "opposite_logprob_tokens": [ + -4.634788513183594 + ], + "logprob_reward": -0.009788726456463337, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009788726456463337, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03749429062008858, + 0.06707359850406647, + 0.05664823576807976, + 0.5543149709701538, + 0.051990896463394165, + 0.02032437175512314, + 0.027614975348114967, + 0.1365431249141693, + 0.02095903642475605, + 0.027036482468247414 + ], + "credit_map": [ + 0.004023897461593151, + 0.004222871270030737, + 0.00471887830644846, + 0.003611010732129216, + 0.004991741385310888, + 0.006532535422593355, + 0.003965587820857763, + 0.005427772179245949, + 0.01578254997730255, + 0.008976565673947334, + 0.008126327767968178, + 0.00634420569986105, + 0.006486816797405481, + 0.009089138358831406, + 0.004448161460459232, + 0.007819835096597672, + 0.004587387200444937, + 0.008368519134819508, + 0.005348220933228731, + 0.007008917164057493, + 0.007493284065276384, + 0.0064785960130393505, + 0.00981238391250372, + 0.007550934329628944, + 0.10432378947734833, + 0.08327402919530869, + 0.0842091515660286, + 0.042332496494054794, + 0.04784039407968521, + 0.06969132274389267, + 0.05099412798881531, + 0.07164964079856873, + 0.006019065156579018, + 0.004024103749543428, + 0.007164510432630777, + 0.004186734091490507, + 0.00391533225774765, + 0.008023558184504509, + 0.004096659831702709, + 0.01456092856824398, + 0.0031902724876999855, + 0.0028945808298885822, + 0.0027301928494125605, + 0.002288935473188758, + 0.0019464967772364616, + 0.002859208034351468, + 0.0019399886950850487, + 0.002474696608260274, + 0.002001195214688778, + 0.0030031625647097826, + 0.00533368531614542, + 0.004691841080784798, + 0.0025267302989959717, + 0.003602231154218316, + 0.00195797230117023, + 0.0044981553219258785, + 0.015785865485668182, + 0.009449806995689869, + 0.012287622317671776, + 0.034708306193351746, + 0.007802370935678482, + 0.038990676403045654, + 0.00842896569520235, + 0.009089495055377483, + 0.0020842074882239103, + 0.0019748599734157324, + 0.003517056116834283, + 0.0037753984797745943, + 0.0017497778171673417, + 0.004424919839948416, + 0.001627707271836698, + 0.0018051082734018564, + 0.00495593948289752, + 0.002057427540421486, + 0.002075761556625366, + 0.004823039285838604, + 0.001892140950076282, + 0.007424585986882448, + 0.001573403482325375, + 0.0022341825533658266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.009788726456463337 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419964898377657 + ], + "opposite_logprob_tokens": [ + -4.264199733734131 + ], + "logprob_reward": -0.01419964898377657, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419964898377657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06120805814862251, + 0.2279200255870819, + 0.11124350130558014, + 0.10132532566785812, + 0.09271962195634842, + 0.06696894019842148, + 0.061802301555871964, + 0.10049238055944443, + 0.14248710870742798, + 0.03383270278573036 + ], + "credit_map": [ + 0.0058168466202914715, + 0.009867975488305092, + 0.011184699833393097, + 0.004735991824418306, + 0.008429177105426788, + 0.00742625305429101, + 0.0051010604947805405, + 0.008646056987345219, + 0.05863451212644577, + 0.026129545643925667, + 0.0320306159555912, + 0.021743619814515114, + 0.017296409234404564, + 0.027039865031838417, + 0.01788097433745861, + 0.02716449461877346, + 0.009239329025149345, + 0.018721019849181175, + 0.01178872399032116, + 0.015954283997416496, + 0.01567966863512993, + 0.006425831932574511, + 0.020549075677990913, + 0.0128855649381876, + 0.018195051699876785, + 0.015877770259976387, + 0.015470409765839577, + 0.006844459567219019, + 0.009771889075636864, + 0.008821348659694195, + 0.011179391294717789, + 0.015165004879236221, + 0.011150562204420567, + 0.01149943470954895, + 0.009722219780087471, + 0.00779657531529665, + 0.006967916153371334, + 0.01785694807767868, + 0.006294482387602329, + 0.02143148146569729, + 0.008497107774019241, + 0.007343736477196217, + 0.005610431078821421, + 0.010587763041257858, + 0.007177744992077351, + 0.014413155615329742, + 0.006817701738327742, + 0.006521303206682205, + 0.006375027820467949, + 0.004530318081378937, + 0.008634470403194427, + 0.00991120096296072, + 0.004413097631186247, + 0.01575382612645626, + 0.003876857692375779, + 0.008307503536343575, + 0.00959410797804594, + 0.011742603965103626, + 0.01716005802154541, + 0.016730163246393204, + 0.019450530409812927, + 0.008681679144501686, + 0.00953699927777052, + 0.0075962417759001255, + 0.013319285586476326, + 0.018420716747641563, + 0.03268870711326599, + 0.01853610947728157, + 0.019120462238788605, + 0.012522472068667412, + 0.012325873598456383, + 0.015553484670817852, + 0.0035789271350950003, + 0.0030146166682243347, + 0.004068370442837477, + 0.005541767925024033, + 0.0036764454562216997, + 0.008065414614975452, + 0.002634397940710187, + 0.0032527619041502476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 2 + ], + "selected_reward": -0.01419964898377657 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical bird named Suzaku with a close-up framing of its head?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the bird's feathers depicted as insubstantial, blazing vermilion light, and vibrant golden flames?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background the vast expanse of space containing slowly rotating nebulae of deep purple and indigo gas with twinkling stars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the bird raise its slender neck, open its crystallized light beak, and emit a golden circular wave of energy?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the bird's body become illusory and translucent, merging with the surrounding nebula to reveal distant stars behind it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635815367102623 + ], + "opposite_logprob_tokens": [ + -3.651358127593994 + ], + "logprob_reward": -0.02635815367102623, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635815367102623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06760591268539429, + 0.18564584851264954, + 0.07148240506649017, + 0.17813754081726074, + 0.048434771597385406, + 0.05533630773425102, + 0.02521083690226078, + 0.14572274684906006, + 0.07948228716850281, + 0.1429414302110672 + ], + "credit_map": [ + 0.006778206676244736, + 0.00810408964753151, + 0.011121703311800957, + 0.006804345641285181, + 0.010049899108707905, + 0.007675583008676767, + 0.00614852923899889, + 0.010923553258180618, + 0.04852212965488434, + 0.02157391607761383, + 0.026852795854210854, + 0.015604037791490555, + 0.01393462996929884, + 0.02418464981019497, + 0.01202187966555357, + 0.02295178361237049, + 0.005101166199892759, + 0.009311147965490818, + 0.008381422609090805, + 0.011832231655716896, + 0.008230469189584255, + 0.00468017952516675, + 0.014723099768161774, + 0.009222682565450668, + 0.03582293167710304, + 0.026542816311120987, + 0.02648167870938778, + 0.012056229636073112, + 0.015915488824248314, + 0.021940333768725395, + 0.015179084613919258, + 0.02419896610081196, + 0.006109671667218208, + 0.00582348695024848, + 0.005019521806389093, + 0.004503339529037476, + 0.005646107718348503, + 0.009664991870522499, + 0.005056013353168964, + 0.006611637305468321, + 0.006971022114157677, + 0.005517294164747, + 0.004764783661812544, + 0.00781705230474472, + 0.004927984904497862, + 0.015830157324671745, + 0.004910601302981377, + 0.004597407765686512, + 0.002421792596578598, + 0.0025079224724322557, + 0.0046151853166520596, + 0.00329050631262362, + 0.002442210679873824, + 0.002815416781231761, + 0.0025553107261657715, + 0.004562491085380316, + 0.017734840512275696, + 0.010231195017695427, + 0.018199192360043526, + 0.031228845939040184, + 0.011907117441296577, + 0.03213033825159073, + 0.009742869064211845, + 0.014548334293067455, + 0.007533575873821974, + 0.009248743765056133, + 0.006769017316401005, + 0.0179117601364851, + 0.006545647047460079, + 0.019893670454621315, + 0.005688184406608343, + 0.005891682114452124, + 0.016701756045222282, + 0.015854841098189354, + 0.02229422517120838, + 0.021102169528603554, + 0.013861103914678097, + 0.02784520573914051, + 0.011384010314941406, + 0.013898109085857868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.02635815367102623 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_4/index_tensor([733, 733])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_4/index_tensor([733, 733])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..896e743732e407518dca3460aba698a677ba8f1d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_4/index_tensor([733, 733])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures the dramatic action of a sleek, modern centrifuge, its body constructed from brushed stainless steel and high-gloss black plastic, perched atop a polished white crystal countertop. Against the backdrop of a pristine kitchen and a gleaming white subway tile backsplash, a vibrant stream of bright orange pulp bursts from the machine. The moist, fibrous pulp gushes violently from the side trough in a powerful, continuous arc, ejecting fine droplets and textured shards of carrot and fruit matter that catch the bright, clean light of the room. The stream flows swiftly and decisively, landing with a satisfying effect within its designated clear plastic container and quickly accumulating in a colorful, richly textured mound. The atmosphere of the scene is a study in controlled chaos, with the pulp's dynamic, almost violent movement creating a powerful visual contrast to the immaculate, static order of the kitchen, highlighting the machine's raw power and efficiency.", + "seed": 30466, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_4/index_733_seed_30466_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.29206664389930664, + "vlm_logprob": -0.29206664389930664, + "vlm_yes_no_margin": 2.475, + "vlm_token_credit": -0.29206664389930664, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5352659225463867 + ], + "opposite_logprob_tokens": [ + -0.9102659225463867 + ], + "logprob_reward": -0.5352659225463867, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5352659225463867, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19076065719127655, + 0.12396440654993057, + 0.1057431697845459, + 0.12379366904497147, + 0.05461553856730461, + 0.10431656986474991, + 0.16054673492908478, + 0.08380931615829468, + 0.028894545510411263, + 0.023555463179945946 + ], + "credit_map": [ + 0.02176677994430065, + 0.02930930070579052, + 0.01454702578485012, + 0.0163858812302351, + 0.01725698634982109, + 0.02553480491042137, + 0.029072709381580353, + 0.03688716143369675, + 0.015112414956092834, + 0.01856011152267456, + 0.008668560534715652, + 0.010419081896543503, + 0.014048365876078606, + 0.015734033659100533, + 0.01880667172372341, + 0.022615160793066025, + 0.007157324347645044, + 0.010112841613590717, + 0.0053438786417245865, + 0.008906811475753784, + 0.013460119254887104, + 0.01597372628748417, + 0.023475931957364082, + 0.02131253108382225, + 0.0067915539257228374, + 0.007327507250010967, + 0.0029679471626877785, + 0.007724696304649115, + 0.016968611627817154, + 0.018902186304330826, + 0.04096279665827751, + 0.022148367017507553, + 0.005961963906884193, + 0.006575924810022116, + 0.004984966479241848, + 0.006157625466585159, + 0.007559338118880987, + 0.007155154831707478, + 0.008725396357476711, + 0.007495168596506119, + 0.012627318501472473, + 0.01188246626406908, + 0.012036982923746109, + 0.009627105668187141, + 0.010686278343200684, + 0.013195851817727089, + 0.017182642593979836, + 0.017077917233109474, + 0.013729863800108433, + 0.023451922461390495, + 0.011941039003431797, + 0.0218354519456625, + 0.02764902636408806, + 0.02780728042125702, + 0.017285969108343124, + 0.01684616692364216, + 0.009489871561527252, + 0.010736451484262943, + 0.006421526428312063, + 0.00549812288954854, + 0.018817782402038574, + 0.014715923927724361, + 0.011659854091703892, + 0.006469776853919029, + 0.002517216606065631, + 0.003839351935312152, + 0.002345892833545804, + 0.0034921159967780113, + 0.005286418367177248, + 0.004556396976113319, + 0.0032874024473130703, + 0.0035697491839528084, + 0.002203220035880804, + 0.005096854642033577, + 0.001498707919381559, + 0.0024435599334537983, + 0.003455572994425893, + 0.0024120130110532045, + 0.0032152293715626, + 0.0032303042244166136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.5352659225463867 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700511634349823 + ], + "opposite_logprob_tokens": [ + -2.6950511932373047 + ], + "logprob_reward": -0.0700511634349823, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700511634349823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12244907766580582, + 0.11878558993339539, + 0.05273790657520294, + 0.09332628548145294, + 0.04192338511347771, + 0.2651011049747467, + 0.0801190510392189, + 0.16456802189350128, + 0.03522377088665962, + 0.0257657989859581 + ], + "credit_map": [ + 0.01165733952075243, + 0.017458943650126457, + 0.009479357860982418, + 0.008356948383152485, + 0.011793493293225765, + 0.018085330724716187, + 0.020628023892641068, + 0.024989649653434753, + 0.012688962742686272, + 0.011509189382195473, + 0.006644160021096468, + 0.006949048954993486, + 0.015606324188411236, + 0.016606101766228676, + 0.022042520344257355, + 0.026739291846752167, + 0.004324291367083788, + 0.007091861683875322, + 0.0025669210590422153, + 0.0045082843862473965, + 0.007612087298184633, + 0.006877799052745104, + 0.013283682987093925, + 0.0064729806035757065, + 0.004835565574467182, + 0.010434980504214764, + 0.0025718763936311007, + 0.00593850202858448, + 0.012800115160644054, + 0.013744601048529148, + 0.029782185330986977, + 0.013218458741903305, + 0.003698482643812895, + 0.004859956447035074, + 0.0021080453880131245, + 0.00459749810397625, + 0.006185042671859264, + 0.005713350605219603, + 0.007171098608523607, + 0.007589913904666901, + 0.022129440680146217, + 0.03325425833463669, + 0.016532061621546745, + 0.022826090455055237, + 0.02792752906680107, + 0.03540061041712761, + 0.05988606438040733, + 0.04714507237076759, + 0.005699407309293747, + 0.011623511090874672, + 0.004198594484478235, + 0.00503231305629015, + 0.019970515742897987, + 0.019406581297516823, + 0.007415443658828735, + 0.006772682070732117, + 0.009649091400206089, + 0.01909405179321766, + 0.007610668428242207, + 0.012687197886407375, + 0.03969147801399231, + 0.027607610449194908, + 0.024809226393699646, + 0.023418694734573364, + 0.0031308522447943687, + 0.00335781485773623, + 0.00284604262560606, + 0.004056131467223167, + 0.004587930627167225, + 0.004160586278885603, + 0.007377777248620987, + 0.005706636235117912, + 0.0020847332198172808, + 0.0048093972727656364, + 0.0012667587725445628, + 0.0031759196426719427, + 0.0038908401038497686, + 0.0030490937642753124, + 0.0039648087695240974, + 0.003524248255416751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.0700511634349823 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026360709220170975 + ], + "opposite_logprob_tokens": [ + -3.6513607501983643 + ], + "logprob_reward": -0.026360709220170975, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026360709220170975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05117691680788994, + 0.03703910857439041, + 0.17340008914470673, + 0.3368711769580841, + 0.0470341295003891, + 0.1960548311471939, + 0.06654264777898788, + 0.04867127537727356, + 0.014590893872082233, + 0.02861895225942135 + ], + "credit_map": [ + 0.004060986451804638, + 0.00691568199545145, + 0.005808372050523758, + 0.007592311594635248, + 0.00575511297211051, + 0.006088670343160629, + 0.008130775764584541, + 0.0068250079639256, + 0.003549088491126895, + 0.003945923876017332, + 0.0019659544341266155, + 0.0025512240827083588, + 0.0054803346283733845, + 0.004921427462249994, + 0.006053964141756296, + 0.008571190759539604, + 0.015462938696146011, + 0.022219184786081314, + 0.012254421599209309, + 0.01468418724834919, + 0.02196255698800087, + 0.020822327584028244, + 0.035466089844703674, + 0.030528372153639793, + 0.016859913244843483, + 0.032334305346012115, + 0.01050452422350645, + 0.021615449339151382, + 0.045509278774261475, + 0.05056874826550484, + 0.11140752583742142, + 0.04807142913341522, + 0.002772330306470394, + 0.0046902853064239025, + 0.001955813728272915, + 0.005345502402633429, + 0.004854502622038126, + 0.006592425983399153, + 0.011436368338763714, + 0.00938690360635519, + 0.014323300682008266, + 0.0262313075363636, + 0.010406405664980412, + 0.017304640263319016, + 0.01942531019449234, + 0.0254325270652771, + 0.04961273819208145, + 0.033318620175123215, + 0.003655651118606329, + 0.00640934007242322, + 0.002037403639405966, + 0.0032657848205417395, + 0.02014712616801262, + 0.01997951976954937, + 0.006408417597413063, + 0.004639404825866222, + 0.003188785631209612, + 0.004522108472883701, + 0.0018685467075556517, + 0.0026526569854468107, + 0.014713946729898453, + 0.009144698269665241, + 0.006640908773988485, + 0.005939623340964317, + 0.001730623422190547, + 0.001780958380550146, + 0.0012813886860385537, + 0.0014249869855120778, + 0.0021482110023498535, + 0.0018640654161572456, + 0.0026191414799541235, + 0.0017415194306522608, + 0.0029598940163850784, + 0.00692292395979166, + 0.001723331050015986, + 0.003144853049889207, + 0.0038450134452432394, + 0.0029945785645395517, + 0.003996142651885748, + 0.003032215405255556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.026360709220170975 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024876620154827833 + ], + "opposite_logprob_tokens": [ + -6.002487659454346 + ], + "logprob_reward": -0.0024876620154827833, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024876620154827833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07265769690275192, + 0.04130421206355095, + 0.40530723333358765, + 0.17013300955295563, + 0.0390046164393425, + 0.10246792435646057, + 0.058675773441791534, + 0.05705852434039116, + 0.01741258054971695, + 0.03597850352525711 + ], + "credit_map": [ + 0.006989691872149706, + 0.011318805627524853, + 0.00636633113026619, + 0.007529781199991703, + 0.008530424907803535, + 0.009492562152445316, + 0.010603444650769234, + 0.011826641857624054, + 0.004133884329348803, + 0.004165743477642536, + 0.0024347840808331966, + 0.003381476504728198, + 0.006146462168544531, + 0.005348450504243374, + 0.007399543188512325, + 0.008293862454593182, + 0.019825195893645287, + 0.041417788714170456, + 0.01146194152534008, + 0.026602094992995262, + 0.049391571432352066, + 0.06154939904808998, + 0.1216110810637474, + 0.07344808429479599, + 0.014330138452351093, + 0.009211158379912376, + 0.005205743946135044, + 0.007729189936071634, + 0.01607847586274147, + 0.02806980349123478, + 0.04281225800514221, + 0.04669621214270592, + 0.002629808383062482, + 0.004046318586915731, + 0.0017991780769079924, + 0.004913737066090107, + 0.004121118225157261, + 0.005252577364444733, + 0.010227962397038937, + 0.006013911217451096, + 0.008052255026996136, + 0.013075163587927818, + 0.0047849444672465324, + 0.008937934413552284, + 0.011673100292682648, + 0.014643242582678795, + 0.02302914671599865, + 0.018272116780281067, + 0.004051965195685625, + 0.007349798455834389, + 0.003515289630740881, + 0.0054448796436190605, + 0.013529488816857338, + 0.013443617150187492, + 0.0068340725265443325, + 0.004506657365709543, + 0.0041727349162101746, + 0.006154832895845175, + 0.002645423635840416, + 0.004033406265079975, + 0.012713386677205563, + 0.008309584110975266, + 0.012101950123906136, + 0.006927201058715582, + 0.0013668097089976072, + 0.0020027044229209423, + 0.001300157979130745, + 0.0020531725604087114, + 0.002498274901881814, + 0.0021148964297026396, + 0.0032742961775511503, + 0.0028022658079862595, + 0.0017271633259952068, + 0.002444427227601409, + 0.0015127529622986913, + 0.002876784885302186, + 0.007829766720533371, + 0.00635526841506362, + 0.00566548528149724, + 0.007566846441477537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.0024876620154827833 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261677622795105 + ], + "opposite_logprob_tokens": [ + -0.5761677622795105 + ], + "logprob_reward": -0.8261677622795105, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261677622795105, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07214339077472687, + 0.04756347835063934, + 0.20181940495967865, + 0.1761871725320816, + 0.09867795556783676, + 0.08916256576776505, + 0.21447554230690002, + 0.04431885480880737, + 0.025266826152801514, + 0.0303848497569561 + ], + "credit_map": [ + 0.007038302253931761, + 0.011641096323728561, + 0.005309130530804396, + 0.0054499926045536995, + 0.006864933297038078, + 0.009431595914065838, + 0.015546033158898354, + 0.01086230855435133, + 0.003854637034237385, + 0.005923538934439421, + 0.002804200630635023, + 0.004050352610647678, + 0.004851765930652618, + 0.006923059467226267, + 0.00919264554977417, + 0.00996327679604292, + 0.0134693942964077, + 0.026803046464920044, + 0.008103807456791401, + 0.01773495227098465, + 0.0321706086397171, + 0.026278572157025337, + 0.04280998185276985, + 0.03444904461503029, + 0.00882007461041212, + 0.01314282976090908, + 0.0047764526680111885, + 0.011290413327515125, + 0.022790461778640747, + 0.0264294296503067, + 0.05943116918206215, + 0.029506340622901917, + 0.010111689567565918, + 0.014883863739669323, + 0.006038582883775234, + 0.009953387081623077, + 0.015255329199135303, + 0.014010281302034855, + 0.011514652520418167, + 0.016910159960389137, + 0.009374990127980709, + 0.012148590758442879, + 0.007667273748666048, + 0.007476651109755039, + 0.009985595941543579, + 0.01167396642267704, + 0.017335159704089165, + 0.013500331901013851, + 0.01390515360981226, + 0.01601104438304901, + 0.00764609407633543, + 0.008665353991091251, + 0.07078678160905838, + 0.06918969750404358, + 0.013215109705924988, + 0.0150563083589077, + 0.0036016812082380056, + 0.003655195003375411, + 0.0020925828721374273, + 0.00268412078730762, + 0.013271051459014416, + 0.008566346019506454, + 0.0055366079322993755, + 0.004911270458251238, + 0.0028954055160284042, + 0.0036300516221672297, + 0.0026074484921991825, + 0.002075907541438937, + 0.0047862776555120945, + 0.003861336736008525, + 0.002794470638036728, + 0.0026159284170717, + 0.0035307290963828564, + 0.004326896276324987, + 0.002681834390386939, + 0.0030207030940800905, + 0.004763271659612656, + 0.0034988997504115105, + 0.003976287320256233, + 0.004586229100823402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.8261677622795105 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, eye-level medium shot captures the dramatic action of a sleek, modern centrifuge, its body constructed from brushed stainless steel and high-gloss black plastic, perched atop a polished white crystal countertop. Against the backdrop of a pristine kitchen and a gleaming white subway tile backsplash, a vibrant stream of bright orange pulp bursts from the machine. The moist, fibrous pulp gushes violently from the side trough in a powerful, continuous arc, ejecting fine droplets and textured shards of carrot and fruit matter that catch the bright, clean light of the room. The stream flows swiftly and decisively, landing with a satisfying effect within its designated clear plastic container and quickly accumulating in a colorful, richly textured mound. The atmosphere of the scene is a study in controlled chaos, with the pulp's dynamic, almost violent movement creating a powerful visual contrast to the immaculate, static order of the kitchen, highlighting the machine's raw power and efficiency.", + "seed": 30467, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_4/index_733_seed_30467_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2434011209756136, + "vlm_logprob": -0.2434011209756136, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.2434011209756136, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.357324481010437 + ], + "opposite_logprob_tokens": [ + -1.232324481010437 + ], + "logprob_reward": -0.357324481010437, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.357324481010437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19046904146671295, + 0.11122138053178787, + 0.08276505768299103, + 0.07198604196310043, + 0.12115427851676941, + 0.016796482726931572, + 0.03346114605665207, + 0.3059358298778534, + 0.05252773314714432, + 0.01368299126625061 + ], + "credit_map": [ + 0.017578572034835815, + 0.036875322461128235, + 0.022384705021977425, + 0.021400466561317444, + 0.0167263001203537, + 0.028336556628346443, + 0.011199338361620903, + 0.03596777468919754, + 0.01179586909711361, + 0.011065885424613953, + 0.016427986323833466, + 0.01740976981818676, + 0.015572603791952133, + 0.013667477294802666, + 0.011703272350132465, + 0.013578513637185097, + 0.01537540927529335, + 0.011307725682854652, + 0.012113307602703571, + 0.011875836178660393, + 0.007945477962493896, + 0.007142555899918079, + 0.006078410893678665, + 0.010926336981356144, + 0.013007818721234798, + 0.011629900895059109, + 0.008623897098004818, + 0.013846139423549175, + 0.006125266198068857, + 0.006682798732072115, + 0.005689733661711216, + 0.006380485370755196, + 0.016915040090680122, + 0.017113836482167244, + 0.01563490368425846, + 0.014379340223968029, + 0.01202314905822277, + 0.011715204454958439, + 0.011632980778813362, + 0.02173982374370098, + 0.002305499278008938, + 0.0014164894819259644, + 0.003521359758451581, + 0.0035674222745001316, + 0.0010389630915597081, + 0.002042322186753154, + 0.0012717201607301831, + 0.0016327062621712685, + 0.003953905776143074, + 0.0025627193972468376, + 0.008053010329604149, + 0.006463619414716959, + 0.0023902312386780977, + 0.003931008279323578, + 0.003179074265062809, + 0.0029275771230459213, + 0.03984096273779869, + 0.02993125468492508, + 0.02823970653116703, + 0.08553414046764374, + 0.018648115918040276, + 0.05869273841381073, + 0.014951351098716259, + 0.030097564682364464, + 0.003195238532498479, + 0.005772364791482687, + 0.008327553980052471, + 0.009405775927007198, + 0.005567033309489489, + 0.007650899235159159, + 0.005424585659056902, + 0.007184276822954416, + 0.0011624580947682261, + 0.0018478077836334705, + 0.0016060560010373592, + 0.002001103712245822, + 0.0012116202851757407, + 0.001220547128468752, + 0.002548620104789734, + 0.0020847793202847242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.357324481010437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901742309331894 + ], + "opposite_logprob_tokens": [ + -2.579017400741577 + ], + "logprob_reward": -0.07901742309331894, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901742309331894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03162948414683342, + 0.033984992653131485, + 0.02231750264763832, + 0.014291666448116302, + 0.0755777657032013, + 0.040960703045129776, + 0.0168935377150774, + 0.5303138494491577, + 0.20784269273281097, + 0.02618779055774212 + ], + "credit_map": [ + 0.004283797927200794, + 0.005578330717980862, + 0.004592304117977619, + 0.0033348326105624437, + 0.0028156903572380543, + 0.004339998587965965, + 0.0019092519069090486, + 0.004775280598551035, + 0.003176972270011902, + 0.003690720535814762, + 0.003987540025264025, + 0.0038537068758159876, + 0.004754664376378059, + 0.004693980328738689, + 0.005494200624525547, + 0.004333209246397018, + 0.0034841839224100113, + 0.0033321527298539877, + 0.003387848846614361, + 0.0032468426506966352, + 0.002365963300690055, + 0.0021212466526776552, + 0.001694478909485042, + 0.002684785518795252, + 0.002211926504969597, + 0.0015696431510150433, + 0.002101234160363674, + 0.0030245743691921234, + 0.0015417997492477298, + 0.0014355044113472104, + 0.001249391702003777, + 0.0011575942626222968, + 0.010724055580794811, + 0.010384812951087952, + 0.010159220546483994, + 0.010755542665719986, + 0.0073107266798615456, + 0.007411262486129999, + 0.005772233474999666, + 0.013059922493994236, + 0.0046654329635202885, + 0.0047973087057471275, + 0.004926458466798067, + 0.00838626828044653, + 0.0030653472058475018, + 0.005337441340088844, + 0.004356822930276394, + 0.005425627809017897, + 0.0018757812213152647, + 0.0015751490136608481, + 0.003795723197981715, + 0.0029858003836125135, + 0.0012208180269226432, + 0.0021305156406015158, + 0.0017162433359771967, + 0.0015935100382193923, + 0.0635748878121376, + 0.05413256585597992, + 0.04802427813410759, + 0.14514857530593872, + 0.03555189445614815, + 0.10565488040447235, + 0.02576906606554985, + 0.05245772749185562, + 0.013803531415760517, + 0.022346502169966698, + 0.03409320488572121, + 0.03317144513130188, + 0.02290567383170128, + 0.029312919825315475, + 0.021587928757071495, + 0.030621502548456192, + 0.0021022488363087177, + 0.002723949495702982, + 0.004406111314892769, + 0.004475962836295366, + 0.003359769703820348, + 0.0025582322850823402, + 0.003079696325585246, + 0.003481820924207568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.07901742309331894 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204937033355236 + ], + "opposite_logprob_tokens": [ + -4.264204978942871 + ], + "logprob_reward": -0.014204937033355236, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204937033355236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13379843533039093, + 0.05764765292406082, + 0.08190026134252548, + 0.07411802560091019, + 0.07908875495195389, + 0.06526700407266617, + 0.027459587901830673, + 0.3932596743106842, + 0.05415235087275505, + 0.03330820053815842 + ], + "credit_map": [ + 0.015478359535336494, + 0.02645973302423954, + 0.016993703320622444, + 0.016566403210163116, + 0.01242815237492323, + 0.016944292932748795, + 0.01061323843896389, + 0.018314555287361145, + 0.0056235832162201405, + 0.008086520247161388, + 0.00878154020756483, + 0.009085600264370441, + 0.007175705395638943, + 0.006948182359337807, + 0.005321296863257885, + 0.00662523228675127, + 0.010379772633314133, + 0.008477767929434776, + 0.011388241313397884, + 0.012938207015395164, + 0.009539352729916573, + 0.006909899413585663, + 0.011592129245400429, + 0.010674899443984032, + 0.011959131807088852, + 0.008708360604941845, + 0.010559214279055595, + 0.012309412471950054, + 0.0076047745533287525, + 0.006156107876449823, + 0.008553673513233662, + 0.008267360739409924, + 0.007891149260103703, + 0.010656380094587803, + 0.008311595767736435, + 0.011646188795566559, + 0.007170223630964756, + 0.009435664862394333, + 0.014394695870578289, + 0.009582864120602608, + 0.0057004098780453205, + 0.005724532529711723, + 0.008232112042605877, + 0.009671671316027641, + 0.006627903785556555, + 0.007465560454875231, + 0.006285551935434341, + 0.015559264458715916, + 0.003741299035027623, + 0.0017292405245825648, + 0.006240771617740393, + 0.004841201938688755, + 0.00237169093452394, + 0.0030313602183014154, + 0.0029939725063741207, + 0.0025100556667894125, + 0.04337659478187561, + 0.04960252717137337, + 0.03303562104701996, + 0.08949145674705505, + 0.038260504603385925, + 0.06858568638563156, + 0.021730635315179825, + 0.04917670413851738, + 0.005898131523281336, + 0.006655221339315176, + 0.010792908258736134, + 0.006532733794301748, + 0.004240220878273249, + 0.00678658252581954, + 0.004923188127577305, + 0.008323372341692448, + 0.00267536542378366, + 0.00424849521368742, + 0.004323284141719341, + 0.005324084777384996, + 0.00429375609382987, + 0.0026065469719469547, + 0.005303329788148403, + 0.004533342085778713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.014204937033355236 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638276718556881 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638276718556881, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638276718556881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03982401266694069, + 0.03564159944653511, + 0.03763942793011665, + 0.017151495441794395, + 0.0667591392993927, + 0.022633368149399757, + 0.05101301521062851, + 0.41269010305404663, + 0.1311158388853073, + 0.18553197383880615 + ], + "credit_map": [ + 0.0034035658463835716, + 0.00712868245318532, + 0.004814064130187035, + 0.003995017148554325, + 0.0036345485132187605, + 0.006127238739281893, + 0.0027649218682199717, + 0.00795598141849041, + 0.0029399862978607416, + 0.00275967875495553, + 0.004029878415167332, + 0.003927759360522032, + 0.006880185566842556, + 0.005328125786036253, + 0.0036084596067667007, + 0.0061675310134887695, + 0.00694397184997797, + 0.004596162121742964, + 0.006932793650776148, + 0.004358941689133644, + 0.004789901431649923, + 0.003109523793682456, + 0.00312480004504323, + 0.003783336840569973, + 0.0023929052986204624, + 0.0025043459609150887, + 0.0021088342182338238, + 0.0034901839680969715, + 0.001810339861549437, + 0.0014296137960627675, + 0.0015383251011371613, + 0.0018769489834085107, + 0.008824261836707592, + 0.00831645354628563, + 0.008807036094367504, + 0.009352194145321846, + 0.006938784383237362, + 0.007093964144587517, + 0.004730524495244026, + 0.012695929035544395, + 0.0031711768824607134, + 0.002929150126874447, + 0.004401219077408314, + 0.0029821584466844797, + 0.0021031240466982126, + 0.0022482394706457853, + 0.0023883224930614233, + 0.0024099801667034626, + 0.006069771014153957, + 0.0048896921798586845, + 0.012513884343206882, + 0.007918058894574642, + 0.00385353434830904, + 0.005040750838816166, + 0.006218953989446163, + 0.0045083737932145596, + 0.05078558623790741, + 0.034189529716968536, + 0.04149939492344856, + 0.12374092638492584, + 0.021262411028146744, + 0.08968132734298706, + 0.01927761919796467, + 0.03225334361195564, + 0.009687474928796291, + 0.01978982612490654, + 0.022846585139632225, + 0.020756037905812263, + 0.011662930250167847, + 0.01583588309586048, + 0.012671440839767456, + 0.017865683883428574, + 0.0176878422498703, + 0.021802503615617752, + 0.02096724882721901, + 0.029637478291988373, + 0.028487160801887512, + 0.01691335439682007, + 0.025888416916131973, + 0.02414800226688385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.008638276718556881 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578204870223999 + ], + "opposite_logprob_tokens": [ + -0.6328204870223999 + ], + "logprob_reward": -0.7578204870223999, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578204870223999, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19866476953029633, + 0.08774852752685547, + 0.0401606485247612, + 0.044548399746418, + 0.20093855261802673, + 0.04801340028643608, + 0.0434446856379509, + 0.2215133160352707, + 0.09299935400485992, + 0.021968308836221695 + ], + "credit_map": [ + 0.023907218128442764, + 0.02196928858757019, + 0.0313287079334259, + 0.02040996216237545, + 0.02173185721039772, + 0.030380476266145706, + 0.016127975657582283, + 0.0328092984855175, + 0.00971261877566576, + 0.005264191422611475, + 0.009292908944189548, + 0.007200279273092747, + 0.017329951748251915, + 0.013288023881614208, + 0.009747292846441269, + 0.01591327041387558, + 0.008155925199389458, + 0.007221234496682882, + 0.007023902609944344, + 0.004943557549268007, + 0.004579116124659777, + 0.0027664259541779757, + 0.003055803244933486, + 0.0024146868381649256, + 0.007102272007614374, + 0.004273680970072746, + 0.007455138489603996, + 0.007262993138283491, + 0.004774200264364481, + 0.003795516910031438, + 0.0054139443673193455, + 0.004470654763281345, + 0.024501314386725426, + 0.027168473228812218, + 0.02310793101787567, + 0.024293826892971992, + 0.020185675472021103, + 0.020258402451872826, + 0.02777528576552868, + 0.033647675067186356, + 0.005823059938848019, + 0.005407240707427263, + 0.009613141417503357, + 0.007311368361115456, + 0.005208552349358797, + 0.004571017809212208, + 0.0039781793020665646, + 0.0061008441261947155, + 0.004937357734888792, + 0.004061829298734665, + 0.007924682460725307, + 0.00772971473634243, + 0.003729642601683736, + 0.004684776999056339, + 0.006865225732326508, + 0.0035114565398544073, + 0.0317213274538517, + 0.02463332563638687, + 0.021038325503468513, + 0.05333658307790756, + 0.016785040497779846, + 0.03450804576277733, + 0.012319732457399368, + 0.027170952409505844, + 0.007264020387083292, + 0.015281972475349903, + 0.019817829132080078, + 0.011551888659596443, + 0.007560861762613058, + 0.01119677722454071, + 0.007540921214967966, + 0.012785092927515507, + 0.0021008853800594807, + 0.0024117701686918736, + 0.0035894219763576984, + 0.0036518012639135122, + 0.002772860461845994, + 0.002263558330014348, + 0.0024717438500374556, + 0.0027062706649303436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.7578204870223999 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, eye-level medium shot captures the dramatic action of a sleek, modern centrifuge, its body constructed from brushed stainless steel and high-gloss black plastic, perched atop a polished white crystal countertop. Against the backdrop of a pristine kitchen and a gleaming white subway tile backsplash, a vibrant stream of bright orange pulp bursts from the machine. The moist, fibrous pulp gushes violently from the side trough in a powerful, continuous arc, ejecting fine droplets and textured shards of carrot and fruit matter that catch the bright, clean light of the room. The stream flows swiftly and decisively, landing with a satisfying effect within its designated clear plastic container and quickly accumulating in a colorful, richly textured mound. The atmosphere of the scene is a study in controlled chaos, with the pulp's dynamic, almost violent movement creating a powerful visual contrast to the immaculate, static order of the kitchen, highlighting the machine's raw power and efficiency.", + "seed": 30468, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_4/index_733_seed_30468_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.12715263646095992, + "vlm_logprob": -0.12715263646095992, + "vlm_yes_no_margin": 2.674999988079071, + "vlm_token_credit": -0.12715263646095992, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15319214761257172 + ], + "opposite_logprob_tokens": [ + -2.0281920433044434 + ], + "logprob_reward": -0.15319214761257172, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.15319214761257172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1354350745677948, + 0.17093127965927124, + 0.047389909625053406, + 0.024415597319602966, + 0.12189003825187683, + 0.019960055127739906, + 0.06822068244218826, + 0.08985602110624313, + 0.2415982484817505, + 0.0803031250834465 + ], + "credit_map": [ + 0.015510794706642628, + 0.01393519900739193, + 0.027191491797566414, + 0.02291475050151348, + 0.011069887317717075, + 0.015485044568777084, + 0.014605515636503696, + 0.014722400344908237, + 0.013105778023600578, + 0.007105108350515366, + 0.007410586811602116, + 0.0056259906850755215, + 0.007444043643772602, + 0.06294992566108704, + 0.007112483028322458, + 0.06017735227942467, + 0.00839821994304657, + 0.008404753170907497, + 0.0067506032064557076, + 0.007008654996752739, + 0.004190527833998203, + 0.0035910247825086117, + 0.005786289926618338, + 0.0032598315738141537, + 0.002392441499978304, + 0.003217005170881748, + 0.0026670689694583416, + 0.0044725751504302025, + 0.0024804463610053062, + 0.003210973460227251, + 0.0037004577461630106, + 0.0022746275644749403, + 0.01749252714216709, + 0.012074314057826996, + 0.02251109480857849, + 0.018102869391441345, + 0.010599511675536633, + 0.01950955018401146, + 0.013249917887151241, + 0.008350253105163574, + 0.0026362978387624025, + 0.002473813947290182, + 0.0027586338110268116, + 0.002706277882680297, + 0.0012393099023029208, + 0.004358706064522266, + 0.0018720893422141671, + 0.0019149266881868243, + 0.006308631971478462, + 0.005505141336470842, + 0.011427495628595352, + 0.006178557872772217, + 0.004381957929581404, + 0.01670357771217823, + 0.008620891720056534, + 0.009094424545764923, + 0.008228759281337261, + 0.008141925558447838, + 0.012347426265478134, + 0.015522809699177742, + 0.010744113475084305, + 0.009430929087102413, + 0.011488930322229862, + 0.013951128348708153, + 0.03867397457361221, + 0.016406696289777756, + 0.046302612870931625, + 0.046078041195869446, + 0.018299279734492302, + 0.027576427906751633, + 0.02048904448747635, + 0.027772173285484314, + 0.013531012460589409, + 0.008278322406113148, + 0.020927157253026962, + 0.013358136638998985, + 0.004352513700723648, + 0.007494334131479263, + 0.007833271287381649, + 0.004528377670794725 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.15319214761257172 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496862530708313 + ], + "opposite_logprob_tokens": [ + -2.9299685955047607 + ], + "logprob_reward": -0.05496862530708313, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496862530708313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10720956325531006, + 0.4115872383117676, + 0.023401321843266487, + 0.019228195771574974, + 0.06677436828613281, + 0.026575997471809387, + 0.06534658372402191, + 0.03980245813727379, + 0.1569175273180008, + 0.0831567719578743 + ], + "credit_map": [ + 0.01669725589454174, + 0.01689307391643524, + 0.019261499866843224, + 0.01564806140959263, + 0.007140143308788538, + 0.011079941876232624, + 0.010749696753919125, + 0.009739880450069904, + 0.032853201031684875, + 0.01625923253595829, + 0.018599696457386017, + 0.013920889236032963, + 0.015543929301202297, + 0.14976492524147034, + 0.01942676119506359, + 0.14521858096122742, + 0.002599551808089018, + 0.0020012317690998316, + 0.0038103112019598484, + 0.0026670286897569895, + 0.003448892617598176, + 0.002392892260104418, + 0.0026674007531255484, + 0.003814011812210083, + 0.0024628150276839733, + 0.0028066555969417095, + 0.002467026934027672, + 0.004373339936137199, + 0.0015129493549466133, + 0.0017583246808499098, + 0.0021309973672032356, + 0.001716085709631443, + 0.005893774330615997, + 0.007387916557490826, + 0.013462498784065247, + 0.014319345355033875, + 0.005744610447436571, + 0.005934227257966995, + 0.009158764034509659, + 0.004873224068433046, + 0.0035913533065468073, + 0.0029028751887381077, + 0.003835089271888137, + 0.003645629156380892, + 0.0019186506979167461, + 0.0053701638244092464, + 0.0033419476822018623, + 0.0019702832214534283, + 0.004546530544757843, + 0.0037502034101635218, + 0.010116392746567726, + 0.005374122876673937, + 0.004406869877129793, + 0.019552532583475113, + 0.007273195311427116, + 0.010326728224754333, + 0.0038130376487970352, + 0.0034185524564236403, + 0.006763349287211895, + 0.007063375785946846, + 0.003715180791914463, + 0.004284590482711792, + 0.004453034605830908, + 0.006291331723332405, + 0.03092087060213089, + 0.020454514771699905, + 0.022362247109413147, + 0.02226930856704712, + 0.013487095944583416, + 0.016074445098638535, + 0.011490453034639359, + 0.019858578220009804, + 0.012483722530305386, + 0.009390746243298054, + 0.015341176651418209, + 0.011988646350800991, + 0.007974458858370781, + 0.007009878754615784, + 0.008796032518148422, + 0.010172108188271523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.05496862530708313 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635478787124157 + ], + "opposite_logprob_tokens": [ + -3.6513547897338867 + ], + "logprob_reward": -0.02635478787124157, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635478787124157, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1014426201581955, + 0.10261579602956772, + 0.03319888934493065, + 0.02021215856075287, + 0.037137195467948914, + 0.028430018573999405, + 0.0394367054104805, + 0.17114980518817902, + 0.2325579971075058, + 0.23381882905960083 + ], + "credit_map": [ + 0.017751887440681458, + 0.018020229414105415, + 0.011509600095450878, + 0.01365711260586977, + 0.008380241692066193, + 0.011910248547792435, + 0.010847842320799828, + 0.009365462698042393, + 0.012569176964461803, + 0.009405706077814102, + 0.0075177536346018314, + 0.007615138776600361, + 0.00791163183748722, + 0.02553633041679859, + 0.008374171331524849, + 0.023685898631811142, + 0.003149032825604081, + 0.002803047653287649, + 0.0028903158381581306, + 0.0027938818093389273, + 0.004628952592611313, + 0.0038345190696418285, + 0.0035600035917013884, + 0.009539137594401836, + 0.0023517196532338858, + 0.0022998156491667032, + 0.002739422023296356, + 0.0027624850627034903, + 0.002181835239753127, + 0.0026271981187164783, + 0.0020348725374788046, + 0.003214811673387885, + 0.004943686071783304, + 0.004087068140506744, + 0.0043017189018428326, + 0.005036492366343737, + 0.004808636382222176, + 0.005760866217315197, + 0.0038594913203269243, + 0.00433923676609993, + 0.0028915544971823692, + 0.003440035507082939, + 0.003400261513888836, + 0.00410574721172452, + 0.0033242639619857073, + 0.00373946619220078, + 0.0031313530635088682, + 0.004397339187562466, + 0.0041903224773705006, + 0.0031462840270251036, + 0.004341986030340195, + 0.004136243369430304, + 0.0031525844242423773, + 0.009699493646621704, + 0.0038763529155403376, + 0.006893439218401909, + 0.025289153680205345, + 0.013490295968949795, + 0.02484457939863205, + 0.02582407556474209, + 0.017095347866415977, + 0.02390240505337715, + 0.018522191792726517, + 0.022181767970323563, + 0.04317235201597214, + 0.03511976823210716, + 0.028119809925556183, + 0.033601343631744385, + 0.025130756199359894, + 0.023945357650518417, + 0.017612596973776817, + 0.025856023654341698, + 0.018463628366589546, + 0.020227786153554916, + 0.06639327853918076, + 0.038438085466623306, + 0.019269928336143494, + 0.02621924690902233, + 0.025875983759760857, + 0.01893090456724167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.02635478787124157 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189423061907291 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.014189423061907291, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189423061907291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12436958402395248, + 0.1711486577987671, + 0.06990701705217361, + 0.03660484030842781, + 0.03525516018271446, + 0.06616700440645218, + 0.02974853664636612, + 0.1556636095046997, + 0.15156516432762146, + 0.15957048535346985 + ], + "credit_map": [ + 0.017989134415984154, + 0.02354784496128559, + 0.017320754006505013, + 0.015547381713986397, + 0.010402478277683258, + 0.01392384897917509, + 0.013817801140248775, + 0.011820338666439056, + 0.011965283192694187, + 0.008701666258275509, + 0.012771840207278728, + 0.009799132123589516, + 0.00820012018084526, + 0.05679236724972725, + 0.010760556906461716, + 0.052157700061798096, + 0.007775759324431419, + 0.007467110641300678, + 0.009216120466589928, + 0.007438600994646549, + 0.01178287249058485, + 0.006841341964900494, + 0.009096870198845863, + 0.010288338176906109, + 0.006181191187351942, + 0.0054455529898405075, + 0.004154197406023741, + 0.007022304926067591, + 0.0034179191570729017, + 0.0034941122867166996, + 0.003351347055286169, + 0.0035382125061005354, + 0.00317230517975986, + 0.004729054868221283, + 0.004405410494655371, + 0.006094337906688452, + 0.0030707628466188908, + 0.0050508431158959866, + 0.005022488534450531, + 0.003709957003593445, + 0.007761008106172085, + 0.006589087657630444, + 0.007079532835632563, + 0.0069779218174517155, + 0.006339770741760731, + 0.014072073623538017, + 0.008843333460390568, + 0.008504273369908333, + 0.0029071534518152475, + 0.00276663969270885, + 0.005390780046582222, + 0.004577543120831251, + 0.0025815542321652174, + 0.003913817927241325, + 0.0043611228466033936, + 0.0032499246299266815, + 0.013854462653398514, + 0.013506053946912289, + 0.017319252714514732, + 0.02453000284731388, + 0.0154051398858428, + 0.017970966175198555, + 0.016537899151444435, + 0.036539848893880844, + 0.027085119858384132, + 0.01833091303706169, + 0.019607072696089745, + 0.01889907382428646, + 0.014925711788237095, + 0.01463954709470272, + 0.018683329224586487, + 0.019394386559724808, + 0.023175127804279327, + 0.012857111170887947, + 0.027426138520240784, + 0.022878022864460945, + 0.014554983004927635, + 0.0168216060847044, + 0.023931356146931648, + 0.01792614348232746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.014189423061907291 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38705819845199585 + ], + "opposite_logprob_tokens": [ + -1.1370582580566406 + ], + "logprob_reward": -0.38705819845199585, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38705819845199585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21465277671813965, + 0.15380194783210754, + 0.06115527078509331, + 0.07086905092000961, + 0.06740293651819229, + 0.07948362827301025, + 0.07190822809934616, + 0.0859423503279686, + 0.14357119798660278, + 0.05121267959475517 + ], + "credit_map": [ + 0.027676250785589218, + 0.03764825314283371, + 0.025179164484143257, + 0.02722511813044548, + 0.017281340435147285, + 0.034519825130701065, + 0.022746803238987923, + 0.02237599715590477, + 0.010881194844841957, + 0.01191966887563467, + 0.010489716194570065, + 0.008963728323578835, + 0.008961205370724201, + 0.04661384969949722, + 0.011632202193140984, + 0.04434036836028099, + 0.007539304438978434, + 0.008659363724291325, + 0.00807380210608244, + 0.0067194788716733456, + 0.008161214180290699, + 0.005971166305243969, + 0.006593405269086361, + 0.009437530301511288, + 0.012584330514073372, + 0.009465647861361504, + 0.0080765625461936, + 0.01135945413261652, + 0.0077569312416017056, + 0.006344446446746588, + 0.007781993132084608, + 0.0074996803887188435, + 0.00990151148289442, + 0.005803680047392845, + 0.008195934817194939, + 0.005911075510084629, + 0.005966976284980774, + 0.017050938680768013, + 0.007379369810223579, + 0.007193443831056356, + 0.007374660111963749, + 0.011132891289889812, + 0.00925445556640625, + 0.009067115373909473, + 0.011621364392340183, + 0.013127650134265423, + 0.008518774062395096, + 0.009386710822582245, + 0.00492639048025012, + 0.00424746610224247, + 0.008938007988035679, + 0.005028863437473774, + 0.005017697345465422, + 0.02197064645588398, + 0.007643008138984442, + 0.014136134646832943, + 0.006371845491230488, + 0.005826251115649939, + 0.008394255302846432, + 0.012401904910802841, + 0.009517492726445198, + 0.010091029107570648, + 0.006837221793830395, + 0.026502346619963646, + 0.02723909355700016, + 0.022890888154506683, + 0.01802903786301613, + 0.021185535937547684, + 0.013582850806415081, + 0.014914306811988354, + 0.009626544080674648, + 0.016102934256196022, + 0.0043134237639606, + 0.00402256241068244, + 0.007655915338546038, + 0.006979404017329216, + 0.006609883159399033, + 0.007238675840198994, + 0.00693759648129344, + 0.007455216255038977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.38705819845199585 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures the dramatic action of a sleek, modern centrifuge, its body constructed from brushed stainless steel and high-gloss black plastic, perched atop a polished white crystal countertop. Against the backdrop of a pristine kitchen and a gleaming white subway tile backsplash, a vibrant stream of bright orange pulp bursts from the machine. The moist, fibrous pulp gushes violently from the side trough in a powerful, continuous arc, ejecting fine droplets and textured shards of carrot and fruit matter that catch the bright, clean light of the room. The stream flows swiftly and decisively, landing with a satisfying effect within its designated clear plastic container and quickly accumulating in a colorful, richly textured mound. The atmosphere of the scene is a study in controlled chaos, with the pulp's dynamic, almost violent movement creating a powerful visual contrast to the immaculate, static order of the kitchen, highlighting the machine's raw power and efficiency.", + "seed": 30469, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_4/index_733_seed_30469_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.21649713041260837, + "vlm_logprob": -0.21649713041260837, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.21649713041260837, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23676994442939758 + ], + "opposite_logprob_tokens": [ + -1.6117699146270752 + ], + "logprob_reward": -0.23676994442939758, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.23676994442939758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10785754024982452, + 0.07551585882902145, + 0.10539736598730087, + 0.24765287339687347, + 0.057118743658065796, + 0.03904939815402031, + 0.1271924525499344, + 0.051026325672864914, + 0.12406038492918015, + 0.06512899696826935 + ], + "credit_map": [ + 0.011178532615303993, + 0.012096754275262356, + 0.009662564843893051, + 0.012283521704375744, + 0.011799678206443787, + 0.022414397448301315, + 0.013890935108065605, + 0.014531166292726994, + 0.0052046990022063255, + 0.00872387271374464, + 0.005582973826676607, + 0.006168618332594633, + 0.01091267354786396, + 0.01804501935839653, + 0.009334113448858261, + 0.01154389325529337, + 0.004630309995263815, + 0.008348507806658745, + 0.005257955752313137, + 0.009092547930777073, + 0.009243311360478401, + 0.016239801421761513, + 0.035643935203552246, + 0.016941012814641, + 0.008413939736783504, + 0.00989626720547676, + 0.019373036921024323, + 0.0044393776915967464, + 0.04439767450094223, + 0.05755719915032387, + 0.007856197655200958, + 0.09571918100118637, + 0.011858542449772358, + 0.007644137367606163, + 0.0068726371973752975, + 0.006233489606529474, + 0.004838966298848391, + 0.006089458707720041, + 0.008009953424334526, + 0.00557156465947628, + 0.003398443106561899, + 0.004474378656595945, + 0.003468095324933529, + 0.004644636996090412, + 0.0045361691154539585, + 0.007529671769589186, + 0.005305849015712738, + 0.0056921569630503654, + 0.012576879933476448, + 0.012675979174673557, + 0.021509330719709396, + 0.021574724465608597, + 0.01477908343076706, + 0.010012294165790081, + 0.02034495212137699, + 0.01371921319514513, + 0.0038276512641459703, + 0.00501682423055172, + 0.0038807957898825407, + 0.004629164002835751, + 0.00764365354552865, + 0.007494772784411907, + 0.010800019837915897, + 0.007733448408544064, + 0.018516503274440765, + 0.012130037881433964, + 0.017237013205885887, + 0.013144243508577347, + 0.014600669965147972, + 0.01418932992964983, + 0.017802264541387558, + 0.016440337523818016, + 0.012134570628404617, + 0.005043203476816416, + 0.013535111211240292, + 0.006762675475329161, + 0.005068493541330099, + 0.005964326206594706, + 0.00710761034861207, + 0.009513008408248425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.23676994442939758 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430745892226696 + ], + "opposite_logprob_tokens": [ + -3.168074607849121 + ], + "logprob_reward": -0.0430745892226696, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430745892226696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18936394155025482, + 0.06839341670274734, + 0.1389257311820984, + 0.22568929195404053, + 0.017060574144124985, + 0.007910148240625858, + 0.016525886952877045, + 0.006782373413443565, + 0.3197519779205322, + 0.009596644900739193 + ], + "credit_map": [ + 0.01786285638809204, + 0.026979446411132812, + 0.015262279659509659, + 0.021228255704045296, + 0.019931284710764885, + 0.04498985409736633, + 0.016943734139204025, + 0.026166245341300964, + 0.00639300188049674, + 0.005792201030999422, + 0.006584739312529564, + 0.004893632605671883, + 0.008492780849337578, + 0.015271198935806751, + 0.010196324437856674, + 0.010769541375339031, + 0.007503263186663389, + 0.01479843445122242, + 0.007673807442188263, + 0.011485389433801174, + 0.014792750589549541, + 0.019546989351511, + 0.042347267270088196, + 0.020777825266122818, + 0.010396435856819153, + 0.010810819454491138, + 0.017824117094278336, + 0.007784965913742781, + 0.03993120789527893, + 0.05278334394097328, + 0.010949467308819294, + 0.07520896196365356, + 0.0029569710604846478, + 0.0027117666322737932, + 0.0016646989388391376, + 0.002552941208705306, + 0.001847239676862955, + 0.0017823922680690885, + 0.002184461336582899, + 0.0013601038372144103, + 0.0005961825954727829, + 0.0005857822834514081, + 0.000539929314982146, + 0.0005899243406020105, + 0.0009665475226938725, + 0.0017250932287424803, + 0.001495828852057457, + 0.0014108616160228848, + 0.001486799563281238, + 0.002001569140702486, + 0.0020508170127868652, + 0.0019666419830173254, + 0.0027178642340004444, + 0.0015425169840455055, + 0.0026963348500430584, + 0.0020633446983993053, + 0.000780214206315577, + 0.0006776437512598932, + 0.0006041068118065596, + 0.0005919899558648467, + 0.0009915161645039916, + 0.0010114596225321293, + 0.001146335038356483, + 0.0009791086195036769, + 0.04237950220704079, + 0.036573734134435654, + 0.040726397186517715, + 0.03202933445572853, + 0.04692602530121803, + 0.03592327982187271, + 0.04475895315408707, + 0.04043475165963173, + 0.0015161741757765412, + 0.0007188076269812882, + 0.001713538309559226, + 0.001024983823299408, + 0.0008736944873817265, + 0.0014420300722122192, + 0.0010508913546800613, + 0.001256525400094688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.0430745892226696 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810510411858559 + ], + "opposite_logprob_tokens": [ + -3.288105010986328 + ], + "logprob_reward": -0.03810510411858559, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810510411858559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1288003921508789, + 0.060212355107069016, + 0.05959193781018257, + 0.1545075625181198, + 0.04818785935640335, + 0.05876760929822922, + 0.056974947452545166, + 0.3735506534576416, + 0.047639813274145126, + 0.011766860261559486 + ], + "credit_map": [ + 0.012763354927301407, + 0.019741695374250412, + 0.011954495683312416, + 0.015026976354420185, + 0.01616121456027031, + 0.024671901017427444, + 0.011848986148834229, + 0.016631776466965675, + 0.0024927270133048296, + 0.004999146331101656, + 0.003006234997883439, + 0.0037853997200727463, + 0.006921025924384594, + 0.01629449799656868, + 0.012265964411199093, + 0.01044735312461853, + 0.006560022011399269, + 0.008809552527964115, + 0.004892223048955202, + 0.005575697869062424, + 0.009716424159705639, + 0.006194345187395811, + 0.009772301651537418, + 0.008071375079452991, + 0.008404705673456192, + 0.0072896466590464115, + 0.019795173779129982, + 0.0051216003485023975, + 0.024931879714131355, + 0.02779136411845684, + 0.005992246326059103, + 0.05518094450235367, + 0.006942131090909243, + 0.010063890367746353, + 0.006348085589706898, + 0.005790360737591982, + 0.004896124359220266, + 0.004055097699165344, + 0.005282992962747812, + 0.0048091779462993145, + 0.0029432091396301985, + 0.004405722953379154, + 0.003175833960995078, + 0.003802581923082471, + 0.00783095508813858, + 0.01309190969914198, + 0.011654853820800781, + 0.011862543411552906, + 0.006446779705584049, + 0.00731069128960371, + 0.008756828494369984, + 0.009717443026602268, + 0.007789070252329111, + 0.00462228013202548, + 0.007162220776081085, + 0.005169634707272053, + 0.01920941472053528, + 0.03076142817735672, + 0.04028955101966858, + 0.025386760011315346, + 0.06031433865427971, + 0.06654560565948486, + 0.06329361349344254, + 0.06774994730949402, + 0.0074852826073765755, + 0.004867964889854193, + 0.00654420955106616, + 0.004947388544678688, + 0.005379112437367439, + 0.00526250246912241, + 0.00666596507653594, + 0.006487390026450157, + 0.0015991369727998972, + 0.0012604687362909317, + 0.0016000918112695217, + 0.0011624377220869064, + 0.0015569828683510423, + 0.0015306874411180615, + 0.0015794399660080671, + 0.001477614976465702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.03810510411858559 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733110640197992 + ], + "opposite_logprob_tokens": [ + -5.006732940673828 + ], + "logprob_reward": -0.006733110640197992, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733110640197992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08543624728918076, + 0.04088471084833145, + 0.0507671982049942, + 0.5324167013168335, + 0.04501134157180786, + 0.08582937717437744, + 0.030396288260817528, + 0.07000024616718292, + 0.049705006182193756, + 0.009552867151796818 + ], + "credit_map": [ + 0.009551571682095528, + 0.010828586295247078, + 0.007891716435551643, + 0.009708615019917488, + 0.0098752835765481, + 0.015941385179758072, + 0.00975051335990429, + 0.011888574808835983, + 0.0033503093291074038, + 0.00473138689994812, + 0.003021783893927932, + 0.002757798181846738, + 0.005158304702490568, + 0.008143060840666294, + 0.007473047822713852, + 0.0062490166164934635, + 0.0049347844906151295, + 0.00796276144683361, + 0.0030991153325885534, + 0.00481799291446805, + 0.007353107910603285, + 0.006019126158207655, + 0.009771455079317093, + 0.0068088555708527565, + 0.020830227062106133, + 0.02065117470920086, + 0.05252399295568466, + 0.011785276234149933, + 0.09437917917966843, + 0.1105317547917366, + 0.016739875078201294, + 0.2049752026796341, + 0.0030449656769633293, + 0.006665954366326332, + 0.0027001043781638145, + 0.004469164181500673, + 0.00593909015879035, + 0.005137635860592127, + 0.009752592071890831, + 0.007301833014935255, + 0.0038996210787445307, + 0.005072324071079493, + 0.004852407146245241, + 0.005406441632658243, + 0.011459705419838428, + 0.017227675765752792, + 0.023415585979819298, + 0.014495615847408772, + 0.0026298221200704575, + 0.004396340809762478, + 0.002186748431995511, + 0.0035558189265429974, + 0.0049988022074103355, + 0.003960934467613697, + 0.004778992384672165, + 0.0038888289127498865, + 0.004576343111693859, + 0.005744356196373701, + 0.012296744622290134, + 0.00428943894803524, + 0.01003003865480423, + 0.009103000164031982, + 0.007660608272999525, + 0.016299717128276825, + 0.005544275976717472, + 0.006391291506588459, + 0.005903245881199837, + 0.004881781060248613, + 0.008395188488066196, + 0.006492922082543373, + 0.0061339628882706165, + 0.005962338298559189, + 0.0013027258682996035, + 0.0010881329653784633, + 0.0011518339160829782, + 0.0010369736701250076, + 0.001327195088379085, + 0.0015455106040462852, + 0.0010417597368359566, + 0.0010587353026494384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.006733110640197992 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a sleek, modern centrifuge made of brushed stainless steel and high-gloss black plastic?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the centrifuge perched on a polished white crystal countertop with a white subway tile backsplash?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a vibrant stream of bright orange pulp burst violently from the machine in a continuous arc?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the orange pulp land in a clear plastic container and accumulate into a colorful mound?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static and at eye-level?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578029036521912 + ], + "opposite_logprob_tokens": [ + -0.6328029036521912 + ], + "logprob_reward": -0.7578029036521912, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578029036521912, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.14108480513095856, + 0.13847017288208008, + 0.049152057617902756, + 0.23857846856117249, + 0.04972793534398079, + 0.02900414541363716, + 0.027620738372206688, + 0.10879502445459366, + 0.17671406269073486, + 0.04085256904363632 + ], + "credit_map": [ + 0.012258348055183887, + 0.016876855865120888, + 0.011343425139784813, + 0.013345536775887012, + 0.0149505864828825, + 0.02979818917810917, + 0.022325865924358368, + 0.020186008885502815, + 0.010228739120066166, + 0.012866990640759468, + 0.009669464081525803, + 0.00988121796399355, + 0.02116118185222149, + 0.019478967413306236, + 0.03126896172761917, + 0.02391466498374939, + 0.003902277909219265, + 0.006215380970388651, + 0.004911401774734259, + 0.007107352372258902, + 0.004845664836466312, + 0.005829351954162121, + 0.010579275898635387, + 0.0057613602839410305, + 0.011202708818018436, + 0.009600169956684113, + 0.02378934808075428, + 0.00621013343334198, + 0.04154428839683533, + 0.049176815897226334, + 0.0071480656042695045, + 0.0899069681763649, + 0.006601698696613312, + 0.008679820224642754, + 0.006958745885640383, + 0.0038108928129076958, + 0.0060211410745978355, + 0.0046384818851947784, + 0.006055386271327734, + 0.006961774546653032, + 0.0029707239009439945, + 0.0043229046277701855, + 0.0030187126249074936, + 0.002757137641310692, + 0.003130999393761158, + 0.0051065171137452126, + 0.0033629946410655975, + 0.004334160592406988, + 0.001621416537091136, + 0.0037101898342370987, + 0.002300563734024763, + 0.004429523833096027, + 0.0036151709500700235, + 0.0038588144816458225, + 0.004779207520186901, + 0.003305854042991996, + 0.006761202588677406, + 0.006423272658139467, + 0.023503435775637627, + 0.007165946997702122, + 0.01692117564380169, + 0.011615568771958351, + 0.009240425191819668, + 0.027164002880454063, + 0.022428086027503014, + 0.021624542772769928, + 0.021562989801168442, + 0.017091060057282448, + 0.028743328526616096, + 0.0212894044816494, + 0.022772029042243958, + 0.02120264247059822, + 0.004077865742146969, + 0.0032129951287060976, + 0.005657711997628212, + 0.003888783510774374, + 0.005353537388145924, + 0.008990363217890263, + 0.004161234945058823, + 0.005510079674422741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.7578029036521912 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_5/index_tensor([4017, 4017])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_5/index_tensor([4017, 4017])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bcd5303bb1d400af80e3546f482243c4e0c55a84 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_5/index_tensor([4017, 4017])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, static shot, shot from a slightly upward angle, focuses firmly on the dark gray tiled eaves of a traditional Chinese building. The scene is bathed in the soft, diffuse light of an overcast sky, evenly illuminating every detail of the two sparrows locked in fierce confrontation. Their feathers bristle in anger, making their small bodies appear unusually fluffy and menacing. Their brown and gray feathers, especially the signature black patch on their throats, appear disorganized. Their gleaming black eyes lock onto each other. In a tense confrontation, they alternately and violently open their tiny beaks, releasing a series of sharp, piercing, almost air-shattering calls. Their movements are not static, but rather provocative, with an explosive rhythm: one suddenly uses its slender legs to launch a swift, short leap forward, and the other immediately responds in kind, forming a rapid, back-and-forth dance of primal aggression across the rustic wooden beams of the eaves. The entire picture is painted in a calm tone, dominated by the earthy brown of the sparrows, the dark gray of the tiles in the background, and the dark brown of the wood, creating a serious and tense atmosphere, rendering the fierce conflict in this tiny world as solemn and dramatic as an epic.", + "seed": 3699, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_5/index_4017_seed_3699_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.16026243083178998, + "vlm_logprob": -0.16026243083178998, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.16026243083178998, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02230031043291092 + ], + "opposite_logprob_tokens": [ + -4.0223002433776855 + ], + "logprob_reward": -0.02230031043291092, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02230031043291092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3355172574520111, + 0.10404245555400848, + 0.1155446246266365, + 0.08284805715084076, + 0.03676846995949745, + 0.04906412959098816, + 0.12927788496017456, + 0.021968059241771698, + 0.0532904788851738, + 0.07167854905128479 + ], + "credit_map": [ + 0.0269179567694664, + 0.04564274102449417, + 0.030280275270342827, + 0.032934993505477905, + 0.07647214084863663, + 0.03585175797343254, + 0.032464854419231415, + 0.05495256930589676, + 0.008283276110887527, + 0.015787595883011818, + 0.008805164135992527, + 0.011300080455839634, + 0.028636259958148003, + 0.008984328247606754, + 0.008870827034115791, + 0.013374934904277325, + 0.00887882336974144, + 0.015269333496689796, + 0.008430467918515205, + 0.019193608313798904, + 0.01695934869349003, + 0.014874541200697422, + 0.014320752583444118, + 0.017617758363485336, + 0.006987844593822956, + 0.011944876052439213, + 0.008171770721673965, + 0.01404285617172718, + 0.00798021163791418, + 0.007856011390686035, + 0.009096956811845303, + 0.016767535358667374, + 0.0030574030242860317, + 0.005194625351577997, + 0.0036989047657698393, + 0.004567076452076435, + 0.005622773431241512, + 0.005767990369349718, + 0.002894127508625388, + 0.0059655713848769665, + 0.0035209297202527523, + 0.0055649420246481895, + 0.004609496798366308, + 0.01037284079939127, + 0.0072174216620624065, + 0.008189558982849121, + 0.003978415858000517, + 0.005610525608062744, + 0.009174654260277748, + 0.011946377344429493, + 0.010302034206688404, + 0.01816600374877453, + 0.01977454498410225, + 0.023722192272543907, + 0.012258675880730152, + 0.02393341064453125, + 0.0021433443762362003, + 0.0032007035333663225, + 0.0019720541313290596, + 0.0023802495561540127, + 0.003589687868952751, + 0.003096597967669368, + 0.001779290963895619, + 0.0038061344530433416, + 0.005245687440037727, + 0.006201450247317553, + 0.005684281699359417, + 0.006347542628645897, + 0.009384003467857838, + 0.008369139395654202, + 0.005097958259284496, + 0.006960422731935978, + 0.005064231809228659, + 0.007221038918942213, + 0.0071176099590957165, + 0.009834389202296734, + 0.00908580981194973, + 0.01326671801507473, + 0.007914625108242035, + 0.012174132280051708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.02230031043291092 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373844176530838 + ], + "opposite_logprob_tokens": [ + -3.408738374710083 + ], + "logprob_reward": -0.03373844176530838, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373844176530838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16284722089767456, + 0.08397616446018219, + 0.08642721176147461, + 0.08201920986175537, + 0.049876291304826736, + 0.07366728037595749, + 0.28867045044898987, + 0.031551193445920944, + 0.06794658303260803, + 0.07301830500364304 + ], + "credit_map": [ + 0.01645413227379322, + 0.03003973700106144, + 0.015927577391266823, + 0.01645873486995697, + 0.02716005966067314, + 0.020929086953401566, + 0.011737621389329433, + 0.024140289053320885, + 0.0055582644417881966, + 0.010025544092059135, + 0.006231010891497135, + 0.009069148451089859, + 0.02237403392791748, + 0.008737499825656414, + 0.008447305299341679, + 0.01353337150067091, + 0.006638640537858009, + 0.012791277840733528, + 0.009245386347174644, + 0.01621788740158081, + 0.013279672712087631, + 0.007501449901610613, + 0.008095555938780308, + 0.012657346203923225, + 0.006804471369832754, + 0.011604085564613342, + 0.008367355912923813, + 0.013434973545372486, + 0.008320360444486141, + 0.006985409650951624, + 0.008966537192463875, + 0.017536021769046783, + 0.004333017393946648, + 0.0082549424842, + 0.006026734132319689, + 0.007184337824583054, + 0.006808998063206673, + 0.006086402572691441, + 0.003590092295780778, + 0.007591772358864546, + 0.004400353878736496, + 0.006638776510953903, + 0.005139584187418222, + 0.015543979592621326, + 0.012641645967960358, + 0.01386819314211607, + 0.006185133010149002, + 0.009249623864889145, + 0.020033225417137146, + 0.023081328719854355, + 0.019781922921538353, + 0.045434121042490005, + 0.045545369386672974, + 0.05676684528589249, + 0.029582444578409195, + 0.04844524338841438, + 0.0033020784612745047, + 0.005275041330605745, + 0.0031611514277756214, + 0.0033868071623146534, + 0.005573215428739786, + 0.004289661534130573, + 0.0025556061882525682, + 0.004007635172456503, + 0.007046558428555727, + 0.007483991794288158, + 0.007141626439988613, + 0.009341598488390446, + 0.012127882800996304, + 0.00837633479386568, + 0.008056944236159325, + 0.00837165117263794, + 0.005512005649507046, + 0.007414951920509338, + 0.006931924261152744, + 0.010370596311986446, + 0.010191202163696289, + 0.015673136338591576, + 0.008517321199178696, + 0.008407175540924072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.03373844176530838 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933751106262207 + ], + "opposite_logprob_tokens": [ + -0.6933751106262207 + ], + "logprob_reward": -0.6933751106262207, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933751106262207, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09646664559841156, + 0.13790304958820343, + 0.07149095088243484, + 0.07638488709926605, + 0.06771593540906906, + 0.09577978402376175, + 0.2608223259449005, + 0.036184556782245636, + 0.07176503539085388, + 0.08548689633607864 + ], + "credit_map": [ + 0.007967090234160423, + 0.013072474859654903, + 0.00876778457313776, + 0.014709758572280407, + 0.014660889282822609, + 0.010221616365015507, + 0.010763098485767841, + 0.016303932294249535, + 0.010857611894607544, + 0.01888970099389553, + 0.012485401704907417, + 0.017465628683567047, + 0.03736850991845131, + 0.011776886880397797, + 0.011013462208211422, + 0.018045848235487938, + 0.006725999526679516, + 0.008547177538275719, + 0.006766174454241991, + 0.012114685028791428, + 0.010934455320239067, + 0.007339999545365572, + 0.008696642704308033, + 0.010365824215114117, + 0.008098077960312366, + 0.012502540834248066, + 0.009725429117679596, + 0.012564080767333508, + 0.011292804963886738, + 0.005407970864325762, + 0.006401959806680679, + 0.010392017662525177, + 0.005932668689638376, + 0.011305557563900948, + 0.009409312158823013, + 0.00875573419034481, + 0.006846894510090351, + 0.00788662675768137, + 0.005736321210861206, + 0.011842820793390274, + 0.005614939611405134, + 0.009837426245212555, + 0.005729314871132374, + 0.020483465865254402, + 0.016664432361721992, + 0.018079042434692383, + 0.007439580280333757, + 0.01193157397210598, + 0.01908857561647892, + 0.022095713764429092, + 0.017257800325751305, + 0.06095573306083679, + 0.03257572278380394, + 0.043356869369745255, + 0.028207838535308838, + 0.03728405013680458, + 0.0028764919843524694, + 0.004141147714108229, + 0.003643076866865158, + 0.005057898815721273, + 0.00535840168595314, + 0.00600530905649066, + 0.0036600942257791758, + 0.00544213829562068, + 0.0069418298080563545, + 0.008240971714258194, + 0.0074545843526721, + 0.010932859033346176, + 0.010446177795529366, + 0.009618342854082584, + 0.007012338377535343, + 0.011117924936115742, + 0.005551918409764767, + 0.01004716008901596, + 0.007477992679923773, + 0.013787555508315563, + 0.012033304199576378, + 0.013185340911149979, + 0.009481026791036129, + 0.013922592625021935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.6933751106262207 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033712390810251236 + ], + "opposite_logprob_tokens": [ + -3.408712387084961 + ], + "logprob_reward": -0.033712390810251236, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033712390810251236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4668266773223877, + 0.0739385113120079, + 0.045675504952669144, + 0.07718826085329056, + 0.09962648898363113, + 0.016008470207452774, + 0.08560184389352798, + 0.023969998583197594, + 0.0397099107503891, + 0.07145430892705917 + ], + "credit_map": [ + 0.03727371618151665, + 0.062183596193790436, + 0.043965209275484085, + 0.050463754683732986, + 0.11145216971635818, + 0.04032135382294655, + 0.044854700565338135, + 0.07631215453147888, + 0.0058488850481808186, + 0.011623673141002655, + 0.007408163044601679, + 0.008395050652325153, + 0.01693691499531269, + 0.007281877566128969, + 0.006267134565860033, + 0.010176820680499077, + 0.0037572297733277082, + 0.006005526520311832, + 0.005426481831818819, + 0.006115125026553869, + 0.006707233842462301, + 0.005066498648375273, + 0.004169110208749771, + 0.008428298868238926, + 0.006570149678736925, + 0.012103285640478134, + 0.007725000847131014, + 0.012017582543194294, + 0.009448392316699028, + 0.006997099611908197, + 0.008153832517564297, + 0.014172922819852829, + 0.008977213874459267, + 0.015783367678523064, + 0.017820199951529503, + 0.010445717722177505, + 0.014039632864296436, + 0.011283095926046371, + 0.007491247728466988, + 0.013786010444164276, + 0.001310426276177168, + 0.0022109781857579947, + 0.0016786751803010702, + 0.002962410217151046, + 0.0019834439735859632, + 0.0019327906193211675, + 0.0013872762210667133, + 0.0025424703489989042, + 0.0064583346247673035, + 0.008867116644978523, + 0.006353946402668953, + 0.023150384426116943, + 0.009460285305976868, + 0.011991449631750584, + 0.008310126140713692, + 0.011010201647877693, + 0.0021900020074099302, + 0.0030965169426053762, + 0.0020986651070415974, + 0.0033832332119345665, + 0.003924391698092222, + 0.002864635083824396, + 0.002827793126925826, + 0.0035847609397023916, + 0.0031548175029456615, + 0.003885256592184305, + 0.0049498723819851875, + 0.005313018802553415, + 0.004733377136290073, + 0.005349974613636732, + 0.003421675181016326, + 0.008901916444301605, + 0.004896427039057016, + 0.00571487145498395, + 0.005187963135540485, + 0.008602156303822994, + 0.009569672867655754, + 0.0208575539290905, + 0.00576362106949091, + 0.010862043127417564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.033712390810251236 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185900524258614 + ], + "opposite_logprob_tokens": [ + -4.018186092376709 + ], + "logprob_reward": -0.018185900524258614, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018185900524258614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26633259654045105, + 0.09057269245386124, + 0.08995486050844193, + 0.08145155757665634, + 0.1277788281440735, + 0.05312956124544144, + 0.14400970935821533, + 0.023143524304032326, + 0.08264629542827606, + 0.04098039120435715 + ], + "credit_map": [ + 0.02157234400510788, + 0.0358794704079628, + 0.022408565506339073, + 0.023465320467948914, + 0.04797527566552162, + 0.03547641262412071, + 0.02707383967936039, + 0.05248138681054115, + 0.008388938382267952, + 0.015137444250285625, + 0.008582381531596184, + 0.011394307017326355, + 0.01480976864695549, + 0.007462397217750549, + 0.007851311936974525, + 0.01694614626467228, + 0.007668963633477688, + 0.010579712688922882, + 0.011053171940147877, + 0.01739005371928215, + 0.013179293833673, + 0.007390223443508148, + 0.009450696408748627, + 0.013242747634649277, + 0.007786544505506754, + 0.010425146669149399, + 0.010304966941475868, + 0.012273559346795082, + 0.012738409452140331, + 0.009814651682972908, + 0.007409823592752218, + 0.010698454454541206, + 0.010747172869741917, + 0.01926565356552601, + 0.015135128051042557, + 0.017755046486854553, + 0.015558378770947456, + 0.017808271571993828, + 0.011194479651749134, + 0.020314689725637436, + 0.0034438115544617176, + 0.005332363769412041, + 0.0037860977463424206, + 0.009712462313473225, + 0.009107525460422039, + 0.008832620456814766, + 0.004355749115347862, + 0.008558927103877068, + 0.009945031255483627, + 0.012427274137735367, + 0.010440360754728317, + 0.022554408758878708, + 0.023199131712317467, + 0.02697785757482052, + 0.014864678494632244, + 0.023600971326231956, + 0.0020671782549470663, + 0.0029053299222141504, + 0.0023241571616381407, + 0.002859491156414151, + 0.0037519163452088833, + 0.004420666955411434, + 0.002145240781828761, + 0.0026695439592003822, + 0.0072220549918711185, + 0.007543749175965786, + 0.01111559383571148, + 0.00959780067205429, + 0.009981700219213963, + 0.009428183548152447, + 0.007183894980698824, + 0.020573321729898453, + 0.002922693034633994, + 0.004204639233648777, + 0.004182998090982437, + 0.004818167071789503, + 0.0054550389759242535, + 0.0057698446325957775, + 0.004577599465847015, + 0.00904940813779831 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.018185900524258614 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot, shot from a slightly upward angle, focuses firmly on the dark gray tiled eaves of a traditional Chinese building. The scene is bathed in the soft, diffuse light of an overcast sky, evenly illuminating every detail of the two sparrows locked in fierce confrontation. Their feathers bristle in anger, making their small bodies appear unusually fluffy and menacing. Their brown and gray feathers, especially the signature black patch on their throats, appear disorganized. Their gleaming black eyes lock onto each other. In a tense confrontation, they alternately and violently open their tiny beaks, releasing a series of sharp, piercing, almost air-shattering calls. Their movements are not static, but rather provocative, with an explosive rhythm: one suddenly uses its slender legs to launch a swift, short leap forward, and the other immediately responds in kind, forming a rapid, back-and-forth dance of primal aggression across the rustic wooden beams of the eaves. The entire picture is painted in a calm tone, dominated by the earthy brown of the sparrows, the dark gray of the tiles in the background, and the dark brown of the wood, creating a serious and tense atmosphere, rendering the fierce conflict in this tiny world as solemn and dramatic as an epic.", + "seed": 3700, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_5/index_4017_seed_3700_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06447371765971184, + "vlm_logprob": -0.06447371765971184, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.06447371765971184, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019442228600382805 + ], + "opposite_logprob_tokens": [ + -4.269442081451416 + ], + "logprob_reward": -0.019442228600382805, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.019442228600382805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1793527603149414, + 0.031917233020067215, + 0.022066844627261162, + 0.10623015463352203, + 0.1685226559638977, + 0.059785280376672745, + 0.04984834045171738, + 0.21634483337402344, + 0.03998345136642456, + 0.1259484738111496 + ], + "credit_map": [ + 0.015195823274552822, + 0.05982040986418724, + 0.007215799298137426, + 0.009263835847377777, + 0.015209381468594074, + 0.06044137105345726, + 0.0048578400164842606, + 0.007348308339715004, + 0.00453373696655035, + 0.005797184072434902, + 0.002496557543054223, + 0.0030192884150892496, + 0.004001283552497625, + 0.007879022508859634, + 0.0024267195258289576, + 0.00176344090141356, + 0.0025767714250832796, + 0.002929534064605832, + 0.00235388264991343, + 0.003176588797941804, + 0.00209994288161397, + 0.00425043934956193, + 0.0015770839527249336, + 0.00310260197147727, + 0.009770892560482025, + 0.007771601434797049, + 0.006971386261284351, + 0.03357616439461708, + 0.005336811766028404, + 0.008373876102268696, + 0.0066571347415447235, + 0.02777228131890297, + 0.017955053597688675, + 0.009306555613875389, + 0.01583923026919365, + 0.04701639339327812, + 0.009154928848147392, + 0.009979243390262127, + 0.01219644583761692, + 0.04707478731870651, + 0.017138255760073662, + 0.0057975188829004765, + 0.0062081594951450825, + 0.004076563287526369, + 0.013158814050257206, + 0.004220024216920137, + 0.004144882783293724, + 0.005041060503572226, + 0.007958542555570602, + 0.012409509159624577, + 0.004125158302485943, + 0.004873754922300577, + 0.006059993989765644, + 0.00584701681509614, + 0.0036790533922612667, + 0.004895311314612627, + 0.03686286881566048, + 0.023625196889042854, + 0.022743673995137215, + 0.02028806507587433, + 0.02388230711221695, + 0.020991912111639977, + 0.021069301292300224, + 0.046881504356861115, + 0.005619996227324009, + 0.0041252984665334225, + 0.003559282049536705, + 0.0033304907847195864, + 0.009598908945918083, + 0.0037587217520922422, + 0.005250442307442427, + 0.004740310367196798, + 0.02304871194064617, + 0.017499012872576714, + 0.007790584582835436, + 0.007482951041311026, + 0.021041754633188248, + 0.010884413495659828, + 0.009386624209582806, + 0.028814420104026794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.019442228600382805 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026364656165242195 + ], + "opposite_logprob_tokens": [ + -3.65136456489563 + ], + "logprob_reward": -0.026364656165242195, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026364656165242195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27404242753982544, + 0.10819877684116364, + 0.024031654000282288, + 0.06968110054731369, + 0.19611947238445282, + 0.03625905141234398, + 0.0327138788998127, + 0.12572787702083588, + 0.10456517338752747, + 0.02866056002676487 + ], + "credit_map": [ + 0.02227051556110382, + 0.0930400863289833, + 0.010331954807043076, + 0.019566912204027176, + 0.025029165670275688, + 0.0857529267668724, + 0.006860993802547455, + 0.01118988636881113, + 0.007907649502158165, + 0.02678017131984234, + 0.0066852751187980175, + 0.0050072395242750645, + 0.018366070464253426, + 0.03089294210076332, + 0.006139310076832771, + 0.00642011035233736, + 0.0024581875186413527, + 0.00288242194801569, + 0.0024140600580722094, + 0.003506644628942013, + 0.0023089824244379997, + 0.005289517808705568, + 0.0017827388364821672, + 0.0033891003113240004, + 0.007061032112687826, + 0.006039705127477646, + 0.00489999121055007, + 0.02182101272046566, + 0.003950260579586029, + 0.005653038155287504, + 0.004211774095892906, + 0.016044287011027336, + 0.032862305641174316, + 0.012625664472579956, + 0.01613488607108593, + 0.04430625960230827, + 0.01121851708739996, + 0.010717696510255337, + 0.012624427676200867, + 0.055629707872867584, + 0.008165085688233376, + 0.004509396385401487, + 0.004390416666865349, + 0.0032566378358751535, + 0.003911737352609634, + 0.0032265298068523407, + 0.003494462464004755, + 0.005304786842316389, + 0.005789515562355518, + 0.006746718194335699, + 0.0030226788949221373, + 0.0031740712001919746, + 0.0031787941697984934, + 0.004412291571497917, + 0.0016472820425406098, + 0.004742528777569532, + 0.022246502339839935, + 0.013941488228738308, + 0.012280555441975594, + 0.011510670185089111, + 0.014630863443017006, + 0.012490689754486084, + 0.011677642352879047, + 0.026949482038617134, + 0.016600482165813446, + 0.010320683941245079, + 0.00913326721638441, + 0.00847950205206871, + 0.0269634947180748, + 0.01022013183683157, + 0.01122900191694498, + 0.011618614196777344, + 0.0052064512856304646, + 0.003863764461129904, + 0.002194708911702037, + 0.002375807613134384, + 0.005724150687456131, + 0.0032152023632079363, + 0.0028282024431973696, + 0.0032522736582905054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.026364656165242195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156069099903107 + ], + "opposite_logprob_tokens": [ + -1.7015607357025146 + ], + "logprob_reward": -0.20156069099903107, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156069099903107, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21686367690563202, + 0.0514201819896698, + 0.04063591733574867, + 0.08026914298534393, + 0.0821654349565506, + 0.0653141662478447, + 0.09806237369775772, + 0.07632822543382645, + 0.08200336992740631, + 0.2069375216960907 + ], + "credit_map": [ + 0.017598725855350494, + 0.0610148087143898, + 0.016172518953680992, + 0.02446209453046322, + 0.01695895753800869, + 0.060735028237104416, + 0.008404207415878773, + 0.011517323553562164, + 0.007219339720904827, + 0.006458813324570656, + 0.0040991948917508125, + 0.005037185736000538, + 0.008507698774337769, + 0.011945808306336403, + 0.003846415551379323, + 0.00430572172626853, + 0.004495273809880018, + 0.0052121346816420555, + 0.004039895720779896, + 0.007895713672041893, + 0.006289418321102858, + 0.004582781810313463, + 0.003534484189003706, + 0.00458621745929122, + 0.007252319250255823, + 0.005429496522992849, + 0.005138887092471123, + 0.02520095556974411, + 0.004349683411419392, + 0.006306420546025038, + 0.004621860571205616, + 0.021969525143504143, + 0.02807045914232731, + 0.008876505307853222, + 0.005694571882486343, + 0.010319429449737072, + 0.006116408389061689, + 0.0044926973059773445, + 0.006438698619604111, + 0.012156656011939049, + 0.01082997303456068, + 0.008062625303864479, + 0.008725062012672424, + 0.006098791025578976, + 0.014896578155457973, + 0.0063132839277386665, + 0.0037274344358593225, + 0.006660417653620243, + 0.016728824004530907, + 0.021651916205883026, + 0.01115417294204235, + 0.010050658136606216, + 0.011741746217012405, + 0.010134829208254814, + 0.00590495252981782, + 0.010695278644561768, + 0.013106409460306168, + 0.008779894560575485, + 0.008120710961520672, + 0.0077451434917747974, + 0.013222780078649521, + 0.008169439621269703, + 0.006771992426365614, + 0.01041185948997736, + 0.013411363586783409, + 0.0089103439822793, + 0.006398244295269251, + 0.005791631527245045, + 0.020117145031690598, + 0.00944338459521532, + 0.008339052088558674, + 0.009592199698090553, + 0.03852849826216698, + 0.03718463331460953, + 0.011906544677913189, + 0.014709681272506714, + 0.030441006645560265, + 0.018766168504953384, + 0.01826687529683113, + 0.03713412210345268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.20156069099903107 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048658184707164764 + ], + "opposite_logprob_tokens": [ + -3.0486581325531006 + ], + "logprob_reward": -0.048658184707164764, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048658184707164764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3025400936603546, + 0.05704129859805107, + 0.019539445638656616, + 0.1426626443862915, + 0.12078951299190521, + 0.042544737458229065, + 0.07244547456502914, + 0.18456357717514038, + 0.016057029366493225, + 0.041816163808107376 + ], + "credit_map": [ + 0.022945281118154526, + 0.11342210322618484, + 0.008952699601650238, + 0.011952439323067665, + 0.02328985184431076, + 0.10455113649368286, + 0.00694273691624403, + 0.010483835823833942, + 0.004386856686323881, + 0.015514730475842953, + 0.004501222632825375, + 0.004859670531004667, + 0.0050178817473351955, + 0.01704523153603077, + 0.002561951754614711, + 0.0031537539325654507, + 0.0026264123152941465, + 0.00212799827568233, + 0.0022636745125055313, + 0.0031392371747642756, + 0.002054736716672778, + 0.0018549916567280889, + 0.002451145090162754, + 0.00302124978043139, + 0.013199036940932274, + 0.01023185532540083, + 0.008608480915427208, + 0.044590968638658524, + 0.0069153099320828915, + 0.011778146959841251, + 0.008254303596913815, + 0.0390845350921154, + 0.013235695660114288, + 0.007138528861105442, + 0.012002402916550636, + 0.03254994377493858, + 0.00708736339583993, + 0.006699892692267895, + 0.00841970182955265, + 0.03365599364042282, + 0.006111534312367439, + 0.005284740589559078, + 0.005998621694743633, + 0.0036230941768735647, + 0.007366675417870283, + 0.005795736331492662, + 0.004374913405627012, + 0.003989421762526035, + 0.01050854753702879, + 0.016435973346233368, + 0.009085439145565033, + 0.009184485301375389, + 0.007763626519590616, + 0.007543860003352165, + 0.004571006633341312, + 0.007352533750236034, + 0.030928852036595345, + 0.021389711648225784, + 0.017826367169618607, + 0.01925196498632431, + 0.02267437055706978, + 0.01795854978263378, + 0.017642801627516747, + 0.03689095750451088, + 0.0025233731139451265, + 0.0013948639389127493, + 0.0015245783142745495, + 0.0014313351130113006, + 0.0036376186180859804, + 0.0015292807947844267, + 0.002200237475335598, + 0.0018157429294660687, + 0.0077711958438158035, + 0.006829741410911083, + 0.003126077586784959, + 0.002975745126605034, + 0.005995289422571659, + 0.0036264415830373764, + 0.0031282640993595123, + 0.008363408036530018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.048658184707164764 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026342827826738358 + ], + "opposite_logprob_tokens": [ + -3.6513428688049316 + ], + "logprob_reward": -0.026342827826738358, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026342827826738358, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25915196537971497, + 0.07110525667667389, + 0.04138312488794327, + 0.039721108973026276, + 0.4392149746417999, + 0.023689603433012962, + 0.02458254061639309, + 0.04136702045798302, + 0.009535389952361584, + 0.050248995423316956 + ], + "credit_map": [ + 0.020585810765624046, + 0.09297528862953186, + 0.010998846031725407, + 0.020950403064489365, + 0.02218063920736313, + 0.07744498550891876, + 0.005629579070955515, + 0.008386429399251938, + 0.00984835997223854, + 0.017495963722467422, + 0.0053080408833920956, + 0.007643757853657007, + 0.009079009294509888, + 0.014754053205251694, + 0.003241895930841565, + 0.0037341760471463203, + 0.0058084032498300076, + 0.005281987600028515, + 0.0045673721469938755, + 0.0065443990752100945, + 0.004728866741061211, + 0.004497688729315996, + 0.002986760577186942, + 0.006967644672840834, + 0.006238632369786501, + 0.004079536069184542, + 0.0030938885174691677, + 0.0045537021942436695, + 0.0036425585858523846, + 0.008484356105327606, + 0.002960960380733013, + 0.006667476613074541, + 0.12848693132400513, + 0.041983503848314285, + 0.02751314640045166, + 0.051507268100976944, + 0.030040370300412178, + 0.023962559178471565, + 0.03213142603635788, + 0.10358980298042297, + 0.006974812131375074, + 0.0025482194032520056, + 0.0030956403352320194, + 0.001986578106880188, + 0.0032513763289898634, + 0.0016132245073094964, + 0.0015670390566810966, + 0.0026527156587690115, + 0.004073351621627808, + 0.004109167493879795, + 0.0029175870586186647, + 0.002099526347592473, + 0.0037545820232480764, + 0.0028304795268923044, + 0.0015516309067606926, + 0.0032462168019264936, + 0.007396711502224207, + 0.004051693249493837, + 0.0040403394959867, + 0.003348698141053319, + 0.005591970402747393, + 0.004148479551076889, + 0.003979599103331566, + 0.00880952924489975, + 0.0015462104929611087, + 0.0010616309009492397, + 0.0010229909094050527, + 0.0010138589423149824, + 0.001089947298169136, + 0.001048843958415091, + 0.001091968035325408, + 0.0016599397640675306, + 0.00936790369451046, + 0.008230374194681644, + 0.003586480626836419, + 0.003304276382550597, + 0.006800161674618721, + 0.004589783493429422, + 0.0039167157374322414, + 0.010453298687934875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.026342827826738358 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot, shot from a slightly upward angle, focuses firmly on the dark gray tiled eaves of a traditional Chinese building. The scene is bathed in the soft, diffuse light of an overcast sky, evenly illuminating every detail of the two sparrows locked in fierce confrontation. Their feathers bristle in anger, making their small bodies appear unusually fluffy and menacing. Their brown and gray feathers, especially the signature black patch on their throats, appear disorganized. Their gleaming black eyes lock onto each other. In a tense confrontation, they alternately and violently open their tiny beaks, releasing a series of sharp, piercing, almost air-shattering calls. Their movements are not static, but rather provocative, with an explosive rhythm: one suddenly uses its slender legs to launch a swift, short leap forward, and the other immediately responds in kind, forming a rapid, back-and-forth dance of primal aggression across the rustic wooden beams of the eaves. The entire picture is painted in a calm tone, dominated by the earthy brown of the sparrows, the dark gray of the tiles in the background, and the dark brown of the wood, creating a serious and tense atmosphere, rendering the fierce conflict in this tiny world as solemn and dramatic as an epic.", + "seed": 3701, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_5/index_4017_seed_3701_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1520555593073368, + "vlm_logprob": -0.1520555593073368, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.1520555593073368, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05450307950377464 + ], + "opposite_logprob_tokens": [ + -3.0545029640197754 + ], + "logprob_reward": -0.05450307950377464, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05450307950377464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13050691783428192, + 0.032561641186475754, + 0.04558222368359566, + 0.05646241083741188, + 0.1243743896484375, + 0.04937224090099335, + 0.04883521422743797, + 0.03541452810168266, + 0.0890825167298317, + 0.3878079354763031 + ], + "credit_map": [ + 0.011749126017093658, + 0.026330750435590744, + 0.010472094640135765, + 0.013581909239292145, + 0.017266344279050827, + 0.023570125922560692, + 0.017135076224803925, + 0.010401468724012375, + 0.004113682080060244, + 0.004036775324493647, + 0.0042314608581364155, + 0.004869821015745401, + 0.004399686120450497, + 0.0035857590846717358, + 0.004349997267127037, + 0.002974457573145628, + 0.003983330447226763, + 0.007058657705783844, + 0.003166027832776308, + 0.005234598647803068, + 0.005484797526150942, + 0.01226517092436552, + 0.004768493119627237, + 0.0036211435217410326, + 0.007296642754226923, + 0.007635361049324274, + 0.005082228220999241, + 0.005742114968597889, + 0.01025992352515459, + 0.0074259573593735695, + 0.006704286672174931, + 0.006315893027931452, + 0.010564562864601612, + 0.008142311125993729, + 0.00680333748459816, + 0.012607436627149582, + 0.01774030365049839, + 0.03679304197430611, + 0.02465650625526905, + 0.007066878490149975, + 0.004435935989022255, + 0.005670857150107622, + 0.004748872481286526, + 0.004243412986397743, + 0.005194975528866053, + 0.01380237191915512, + 0.00644349679350853, + 0.0048323143273591995, + 0.004254288971424103, + 0.009926866739988327, + 0.0034242658875882626, + 0.00549409631639719, + 0.007373353000730276, + 0.007374407723546028, + 0.00657269824296236, + 0.004415229428559542, + 0.003375803818926215, + 0.006728106644004583, + 0.002495366148650646, + 0.002865849994122982, + 0.003922112286090851, + 0.005439327098429203, + 0.007188715506345034, + 0.003399244975298643, + 0.008578546345233917, + 0.015232940204441547, + 0.0064345067366957664, + 0.01085895486176014, + 0.011537203565239906, + 0.007258088793605566, + 0.01982942968606949, + 0.00935283675789833, + 0.03327423334121704, + 0.08054549992084503, + 0.0210788045078516, + 0.03252532705664635, + 0.025965364649891853, + 0.02356833592057228, + 0.11625851690769196, + 0.054591819643974304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.05450307950377464 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331218123435974 + ], + "opposite_logprob_tokens": [ + -3.7733120918273926 + ], + "logprob_reward": -0.02331218123435974, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331218123435974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32680490612983704, + 0.05968989059329033, + 0.04761924594640732, + 0.03865371644496918, + 0.07428744435310364, + 0.020499233156442642, + 0.05413585528731346, + 0.01965337246656418, + 0.04025774449110031, + 0.3183985650539398 + ], + "credit_map": [ + 0.02729102410376072, + 0.03849055618047714, + 0.022264353930950165, + 0.03259877860546112, + 0.02577374316751957, + 0.08837885409593582, + 0.05618039518594742, + 0.03582720458507538, + 0.0070344675332307816, + 0.011584454216063023, + 0.005724230781197548, + 0.0064527555368840694, + 0.006686507258564234, + 0.010998997837305069, + 0.005727960262447596, + 0.005480515770614147, + 0.00399530166760087, + 0.00610696105286479, + 0.003011559834703803, + 0.0052121300250291824, + 0.005673667415976524, + 0.014278348535299301, + 0.004149254877120256, + 0.005192025564610958, + 0.004453497938811779, + 0.005104490555822849, + 0.003839137963950634, + 0.004332769196480513, + 0.004757150076329708, + 0.009402629919350147, + 0.003223697654902935, + 0.0035403419751673937, + 0.005233442410826683, + 0.0035104979760944843, + 0.002600450301542878, + 0.004938104189932346, + 0.010081043466925621, + 0.023460905998945236, + 0.019979223608970642, + 0.004483778029680252, + 0.0018564083147794008, + 0.002434166381135583, + 0.0017110038315877318, + 0.0015304374974220991, + 0.002972506918013096, + 0.006087156478315592, + 0.0021069254726171494, + 0.0018006297759711742, + 0.004501078277826309, + 0.014014925807714462, + 0.002386347623541951, + 0.0032128114253282547, + 0.003477222751826048, + 0.004760819487273693, + 0.015340818092226982, + 0.006441829726099968, + 0.0019965104293078184, + 0.004488973878324032, + 0.0015757214277982712, + 0.0019226822769269347, + 0.002163071185350418, + 0.002055432880297303, + 0.003989037591964006, + 0.001461942563764751, + 0.0033241896890103817, + 0.00474550062790513, + 0.004470969084650278, + 0.00675560999661684, + 0.007288283668458462, + 0.0050006103701889515, + 0.0036050130147486925, + 0.005067566409707069, + 0.028834179043769836, + 0.07229706645011902, + 0.014709251932799816, + 0.022876476868987083, + 0.01867944747209549, + 0.015906646847724915, + 0.10079015046358109, + 0.04430532082915306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.02331218123435974 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328317523002625 + ], + "opposite_logprob_tokens": [ + -0.7578317523002625 + ], + "logprob_reward": -0.6328317523002625, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328317523002625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.290001779794693, + 0.02225830964744091, + 0.04280678927898407, + 0.015500928275287151, + 0.09693095088005066, + 0.022466354072093964, + 0.07060819864273071, + 0.034083712846040726, + 0.0944853276014328, + 0.3108576536178589 + ], + "credit_map": [ + 0.01954330876469612, + 0.035531871020793915, + 0.016517365351319313, + 0.02647150494158268, + 0.022908208891749382, + 0.07554411143064499, + 0.05216231942176819, + 0.04132305830717087, + 0.002502827439457178, + 0.0040443758480250835, + 0.0020453950855880976, + 0.0021517935674637556, + 0.0031970771960914135, + 0.004906033165752888, + 0.0018766768043860793, + 0.001534128445200622, + 0.00329045788384974, + 0.006049104500561953, + 0.0029592097271233797, + 0.005510421935468912, + 0.004582151770591736, + 0.01377007458359003, + 0.003374082036316395, + 0.003271283581852913, + 0.0016233277274295688, + 0.002980692544952035, + 0.0014287400990724564, + 0.0018480075523257256, + 0.0020655891858041286, + 0.0029214550741016865, + 0.0011707027442753315, + 0.0014624124160036445, + 0.007112451363354921, + 0.004544743802398443, + 0.0036782287061214447, + 0.006419407203793526, + 0.01367794256657362, + 0.031017392873764038, + 0.025169923901557922, + 0.0053108567371964455, + 0.0014459462836384773, + 0.002590175485238433, + 0.0017073756316676736, + 0.0025940248742699623, + 0.0025346907787024975, + 0.007278041448444128, + 0.002519471338018775, + 0.0017966254381462932, + 0.006219295784831047, + 0.019233081489801407, + 0.0025993301533162594, + 0.003856629366055131, + 0.004571243189275265, + 0.005557157099246979, + 0.020094070583581924, + 0.00847738329321146, + 0.00317218410782516, + 0.00844939611852169, + 0.0025707574095577, + 0.004764965735375881, + 0.002796349348500371, + 0.0023080965038388968, + 0.007728551980108023, + 0.002293406520038843, + 0.009972763247787952, + 0.012460735626518726, + 0.007077581249177456, + 0.010454057715833187, + 0.015930194407701492, + 0.011822563596069813, + 0.016340363770723343, + 0.010427064262330532, + 0.02739548496901989, + 0.06270209699869156, + 0.016990063712000847, + 0.02539283223450184, + 0.021113799884915352, + 0.018513210117816925, + 0.09458425641059875, + 0.04416589438915253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.6328317523002625 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287026211619377 + ], + "opposite_logprob_tokens": [ + -3.773287057876587 + ], + "logprob_reward": -0.023287026211619377, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287026211619377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2527913749217987, + 0.0730084702372551, + 0.020058486610651016, + 0.02277970127761364, + 0.09104233235120773, + 0.0473690889775753, + 0.08390830457210541, + 0.14763356745243073, + 0.0559547021985054, + 0.2054539918899536 + ], + "credit_map": [ + 0.02258695289492607, + 0.02622317150235176, + 0.017789267003536224, + 0.032759543508291245, + 0.01800774596631527, + 0.056541603058576584, + 0.04565444961190224, + 0.03322862461209297, + 0.009081767871975899, + 0.015760384500026703, + 0.004904003813862801, + 0.006924166344106197, + 0.006212085019797087, + 0.011749566532671452, + 0.011187958531081676, + 0.007188545539975166, + 0.0019828928634524345, + 0.0030029849149286747, + 0.00157321582082659, + 0.0021362928673624992, + 0.0023284158669412136, + 0.0040594544261693954, + 0.002570676151663065, + 0.002404553582891822, + 0.002389142056927085, + 0.0024062078446149826, + 0.00237507582642138, + 0.0026465102564543486, + 0.004168887622654438, + 0.00499365571886301, + 0.002085351385176182, + 0.001714868238195777, + 0.007729366887360811, + 0.006707943044602871, + 0.003262702841311693, + 0.006309029646217823, + 0.01154827605932951, + 0.02351454459130764, + 0.02456297166645527, + 0.0074074966832995415, + 0.005558607168495655, + 0.006375526078045368, + 0.003783802269026637, + 0.003672834485769272, + 0.007330985739827156, + 0.010838509537279606, + 0.005089938640594482, + 0.004718882963061333, + 0.008804885670542717, + 0.014732040464878082, + 0.0069101168774068356, + 0.007333682384341955, + 0.012137136422097683, + 0.02060559019446373, + 0.005610139574855566, + 0.007774718571454287, + 0.014431405812501907, + 0.03803948685526848, + 0.0117275919765234, + 0.019748231396079063, + 0.01197704579681158, + 0.01154892798513174, + 0.029907338321208954, + 0.010253536514937878, + 0.004634105134755373, + 0.009565315209329128, + 0.005315843038260937, + 0.010740205645561218, + 0.0057388124987483025, + 0.0070974198170006275, + 0.008570684120059013, + 0.004292316269129515, + 0.01990414969623089, + 0.055109575390815735, + 0.007733486592769623, + 0.013313558883965015, + 0.011625087819993496, + 0.012236927635967731, + 0.06150755286216736, + 0.024023663252592087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.023287026211619377 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026343757286667824 + ], + "opposite_logprob_tokens": [ + -3.651343822479248 + ], + "logprob_reward": -0.026343757286667824, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026343757286667824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17317242920398712, + 0.06418916583061218, + 0.041505005210638046, + 0.03075532801449299, + 0.05348353460431099, + 0.06223844736814499, + 0.09693536907434464, + 0.04852473735809326, + 0.1421509087085724, + 0.2870450019836426 + ], + "credit_map": [ + 0.012191865593194962, + 0.02749781683087349, + 0.00928368791937828, + 0.012511491775512695, + 0.016367115080356598, + 0.036211490631103516, + 0.02839641086757183, + 0.030712563544511795, + 0.00892920047044754, + 0.00868933368474245, + 0.004662064835429192, + 0.006608820520341396, + 0.010072803124785423, + 0.010497980751097202, + 0.006535891909152269, + 0.00819307565689087, + 0.0039510419592261314, + 0.0038918310310691595, + 0.0020122486166656017, + 0.003129491349682212, + 0.005461747758090496, + 0.004332511220127344, + 0.010547866113483906, + 0.008178274147212505, + 0.0032197367399930954, + 0.002570183016359806, + 0.0034965535160154104, + 0.003647630102932453, + 0.005592310335487127, + 0.007224380504339933, + 0.0021051315125077963, + 0.0028994048479944468, + 0.00644699577242136, + 0.010354678146541119, + 0.003260547760874033, + 0.005516866687685251, + 0.009135152213275433, + 0.005445611663162708, + 0.00726320780813694, + 0.006060480605810881, + 0.0054466514848172665, + 0.008237696252763271, + 0.004804772324860096, + 0.007258376572281122, + 0.00558527372777462, + 0.020394057035446167, + 0.005674170330166817, + 0.004837455693632364, + 0.007482979912310839, + 0.015951773151755333, + 0.004875772632658482, + 0.00838741660118103, + 0.009309441782534122, + 0.02227885276079178, + 0.018707208335399628, + 0.00994192622601986, + 0.004553577397018671, + 0.006245769560337067, + 0.004383519757539034, + 0.006290394347161055, + 0.003234869334846735, + 0.005416668951511383, + 0.010519756004214287, + 0.007880187593400478, + 0.01625721901655197, + 0.026277601718902588, + 0.008341819979250431, + 0.01433144137263298, + 0.016989463940262794, + 0.013165044598281384, + 0.03136877715587616, + 0.01541954930871725, + 0.026658112183213234, + 0.07253206521272659, + 0.011103971861302853, + 0.018509088084101677, + 0.015469211153686047, + 0.014509755186736584, + 0.09046053886413574, + 0.03780228644609451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.026343757286667824 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot, shot from a slightly upward angle, focuses firmly on the dark gray tiled eaves of a traditional Chinese building. The scene is bathed in the soft, diffuse light of an overcast sky, evenly illuminating every detail of the two sparrows locked in fierce confrontation. Their feathers bristle in anger, making their small bodies appear unusually fluffy and menacing. Their brown and gray feathers, especially the signature black patch on their throats, appear disorganized. Their gleaming black eyes lock onto each other. In a tense confrontation, they alternately and violently open their tiny beaks, releasing a series of sharp, piercing, almost air-shattering calls. Their movements are not static, but rather provocative, with an explosive rhythm: one suddenly uses its slender legs to launch a swift, short leap forward, and the other immediately responds in kind, forming a rapid, back-and-forth dance of primal aggression across the rustic wooden beams of the eaves. The entire picture is painted in a calm tone, dominated by the earthy brown of the sparrows, the dark gray of the tiles in the background, and the dark brown of the wood, creating a serious and tense atmosphere, rendering the fierce conflict in this tiny world as solemn and dramatic as an epic.", + "seed": 3702, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_5/index_4017_seed_3702_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.061151765286922455, + "vlm_logprob": -0.061151765286922455, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.061151765286922455, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06791101396083832 + ], + "opposite_logprob_tokens": [ + -2.81791090965271 + ], + "logprob_reward": -0.06791101396083832, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06791101396083832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16847160458564758, + 0.0755281150341034, + 0.0799737423658371, + 0.02933833748102188, + 0.019802572205662727, + 0.05004715919494629, + 0.09606388956308365, + 0.15858198702335358, + 0.14457587897777557, + 0.17761671543121338 + ], + "credit_map": [ + 0.01783175766468048, + 0.01809663511812687, + 0.017759913578629494, + 0.02701355703175068, + 0.022108888253569603, + 0.019180061295628548, + 0.02988096885383129, + 0.01659982092678547, + 0.010123591870069504, + 0.009797685779631138, + 0.010141684673726559, + 0.009765229187905788, + 0.009877613745629787, + 0.009595627896487713, + 0.009184296242892742, + 0.007042382378131151, + 0.009719181805849075, + 0.007356358226388693, + 0.006162387318909168, + 0.017011001706123352, + 0.008401885628700256, + 0.008441622368991375, + 0.010093381628394127, + 0.012787924148142338, + 0.0032212489750236273, + 0.0035010837018489838, + 0.002806125907227397, + 0.004350972827523947, + 0.0037992510478943586, + 0.0024558459408581257, + 0.003713377984240651, + 0.005490430630743504, + 0.002167630475014448, + 0.0016647990560159087, + 0.0014976819511502981, + 0.0032484838739037514, + 0.002379263751208782, + 0.003168688388541341, + 0.00207372079603374, + 0.0036023028660565615, + 0.006925051566213369, + 0.004926183260977268, + 0.004249822814017534, + 0.009907232597470284, + 0.005888407118618488, + 0.006680216174572706, + 0.004689700901508331, + 0.006780544295907021, + 0.008915075100958347, + 0.005215196404606104, + 0.00463856803253293, + 0.00778439873829484, + 0.004333526361733675, + 0.004628372378647327, + 0.027494177222251892, + 0.033054567873477936, + 0.018196964636445045, + 0.031163353472948074, + 0.01411166787147522, + 0.023211367428302765, + 0.01665610633790493, + 0.016380349174141884, + 0.02079497091472149, + 0.018067212775349617, + 0.027342524379491806, + 0.015072314999997616, + 0.015363716520369053, + 0.021247882395982742, + 0.01726107858121395, + 0.020728612318634987, + 0.01416920218616724, + 0.013390540145337582, + 0.018413346260786057, + 0.01973167434334755, + 0.016182849183678627, + 0.029853736981749535, + 0.021218059584498405, + 0.028872938826680183, + 0.021376173943281174, + 0.021967938169836998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.06791101396083832 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016087019816040993 + ], + "opposite_logprob_tokens": [ + -4.141087055206299 + ], + "logprob_reward": -0.016087019816040993, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016087019816040993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07610148191452026, + 0.07306130975484848, + 0.0533304437994957, + 0.049818430095911026, + 0.03330089524388313, + 0.057078808546066284, + 0.07709962874650955, + 0.10897748172283173, + 0.16433347761631012, + 0.30689799785614014 + ], + "credit_map": [ + 0.01327307615429163, + 0.006984231993556023, + 0.006445496808737516, + 0.009665240533649921, + 0.009160538204014301, + 0.009749773889780045, + 0.013147728517651558, + 0.007675400003790855, + 0.01311899907886982, + 0.006487633567303419, + 0.00761334178969264, + 0.01037549413740635, + 0.006437958218157291, + 0.009173793718218803, + 0.008456270210444927, + 0.011397823691368103, + 0.005608473438769579, + 0.006435947027057409, + 0.004262218717485666, + 0.008940824307501316, + 0.0070971897803246975, + 0.007224224507808685, + 0.006211759988218546, + 0.007549806032329798, + 0.0045965504832565784, + 0.005400469526648521, + 0.0038500726222991943, + 0.006921134423464537, + 0.009446186944842339, + 0.00528675876557827, + 0.0069289980456233025, + 0.00738826859742403, + 0.003307912964373827, + 0.0027175305876880884, + 0.002880421234294772, + 0.004423134494572878, + 0.003979552537202835, + 0.0062173171900212765, + 0.0044946917332708836, + 0.005280336365103722, + 0.009952344000339508, + 0.006531416904181242, + 0.005168351344764233, + 0.007905532605946064, + 0.006936719641089439, + 0.006208873353898525, + 0.005205689929425716, + 0.009169884957373142, + 0.006946976762264967, + 0.0044035883620381355, + 0.003987809177488089, + 0.0075105526484549046, + 0.0045358045026659966, + 0.004110895097255707, + 0.020338188856840134, + 0.02526581659913063, + 0.01304853055626154, + 0.021384984254837036, + 0.010592085309326649, + 0.015773488208651543, + 0.014057054184377193, + 0.011916521936655045, + 0.012017082422971725, + 0.010187745094299316, + 0.027083180844783783, + 0.017306819558143616, + 0.01792963594198227, + 0.024729233235120773, + 0.016677362844347954, + 0.026512151584029198, + 0.016378778964281082, + 0.017716318368911743, + 0.028992189094424248, + 0.03421512618660927, + 0.025321725755929947, + 0.04972197115421295, + 0.03750242292881012, + 0.053696420043706894, + 0.03778580203652382, + 0.03966238349676132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.016087019816040993 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603671908378601 + ], + "opposite_logprob_tokens": [ + -1.9103672504425049 + ], + "logprob_reward": -0.1603671908378601, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603671908378601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26812943816185, + 0.09558108448982239, + 0.10057058930397034, + 0.02909988909959793, + 0.05691596493124962, + 0.043427180498838425, + 0.08764953166246414, + 0.08328280597925186, + 0.0880722776055336, + 0.14727114140987396 + ], + "credit_map": [ + 0.03172280639410019, + 0.0271071195602417, + 0.026391522958874702, + 0.051663946360349655, + 0.03622730076313019, + 0.03198700770735741, + 0.03235660865902901, + 0.030673179775476456, + 0.01740453578531742, + 0.00809941440820694, + 0.011258088052272797, + 0.014070729725062847, + 0.011615614406764507, + 0.012747547589242458, + 0.008748591877520084, + 0.011636569164693356, + 0.0076886387541890144, + 0.010142330080270767, + 0.010219949297606945, + 0.02726018615067005, + 0.008751948364078999, + 0.010192395187914371, + 0.011281488463282585, + 0.01503367256373167, + 0.0029217349365353584, + 0.0034186970442533493, + 0.0022674526553601027, + 0.004661878105252981, + 0.003789035137742758, + 0.0035554631613194942, + 0.004042011220008135, + 0.004443620331585407, + 0.006891361903399229, + 0.0047721462324261665, + 0.004917455837130547, + 0.008128929883241653, + 0.004985733423382044, + 0.010543561540544033, + 0.007154018618166447, + 0.00952276773750782, + 0.007647011429071426, + 0.0054284376092255116, + 0.0035812421701848507, + 0.005389923229813576, + 0.004507454577833414, + 0.005541973281651735, + 0.004576723091304302, + 0.006754424888640642, + 0.008231396786868572, + 0.005075122695416212, + 0.004528514575213194, + 0.010082965716719627, + 0.005035680253058672, + 0.005611253436654806, + 0.02365618757903576, + 0.025428427383303642, + 0.010229913517832756, + 0.010443043895065784, + 0.008677538484334946, + 0.011998273432254791, + 0.011479828506708145, + 0.008963515982031822, + 0.009569582529366016, + 0.011921127326786518, + 0.01403836626559496, + 0.008466250263154507, + 0.008549767546355724, + 0.014419511891901493, + 0.009358997456729412, + 0.01415441557765007, + 0.008291756734251976, + 0.010793222114443779, + 0.015022579580545425, + 0.016830746084451675, + 0.011725821532309055, + 0.023299288004636765, + 0.016270484775304794, + 0.029477419331669807, + 0.015416497364640236, + 0.01922832988202572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.1603671908378601 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810785710811615 + ], + "opposite_logprob_tokens": [ + -3.2881078720092773 + ], + "logprob_reward": -0.03810785710811615, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810785710811615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34182995557785034, + 0.16733352839946747, + 0.061394743621349335, + 0.04577424004673958, + 0.02933489717543125, + 0.03654738515615463, + 0.04405413568019867, + 0.07022924721240997, + 0.07016615569591522, + 0.13333570957183838 + ], + "credit_map": [ + 0.04467430338263512, + 0.033151231706142426, + 0.03878169506788254, + 0.03798580914735794, + 0.05400902032852173, + 0.03840677812695503, + 0.05470690131187439, + 0.04011420160531998, + 0.02648838423192501, + 0.03453736752271652, + 0.014883805997669697, + 0.0157234575599432, + 0.014810607768595219, + 0.01689649559557438, + 0.02179674431681633, + 0.022196665406227112, + 0.010299978777766228, + 0.0063164434395730495, + 0.005999675020575523, + 0.012034851126372814, + 0.0055441963486373425, + 0.00683201989158988, + 0.007464305497705936, + 0.006903268862515688, + 0.004334483295679092, + 0.004005633760243654, + 0.0062959035858511925, + 0.007179716136306524, + 0.00507695646956563, + 0.004800027702003717, + 0.005607170052826405, + 0.008474348112940788, + 0.0042181764729321, + 0.0031970241107046604, + 0.002053407719358802, + 0.004093112424015999, + 0.0029448713175952435, + 0.005009456537663937, + 0.003281785873696208, + 0.0045370617881417274, + 0.006507398094981909, + 0.003376938868314028, + 0.003528662957251072, + 0.006351246498525143, + 0.003978653345257044, + 0.004042831715196371, + 0.0034624517429620028, + 0.005299201235175133, + 0.003695896826684475, + 0.0023531606420874596, + 0.0020765222143381834, + 0.0028009479865431786, + 0.003059942042455077, + 0.002759123919531703, + 0.011644060723483562, + 0.01566447876393795, + 0.008948245085775852, + 0.00906322617083788, + 0.010204654186964035, + 0.009777993895113468, + 0.008546422235667706, + 0.0056334068067371845, + 0.008263952098786831, + 0.00979134626686573, + 0.009885457344353199, + 0.007524408400058746, + 0.00808699056506157, + 0.012517529539763927, + 0.006159178912639618, + 0.011152542196214199, + 0.007140450179576874, + 0.007699599955230951, + 0.014366837218403816, + 0.015177479945123196, + 0.012029786594212055, + 0.022706057876348495, + 0.016887715086340904, + 0.023352686315774918, + 0.013399092480540276, + 0.015416057780385017 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.03810785710811615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up focusing on two sparrows on the eaves of a traditional Chinese building?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the sparrows brown and gray with signature black patches on their throats?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the sparrows engage in a rapid back-and-forth leaping dance across rustic wooden beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of dark gray tiled eaves under an overcast sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static from a slightly upward angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023285744711756706 + ], + "opposite_logprob_tokens": [ + -3.7732856273651123 + ], + "logprob_reward": -0.023285744711756706, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023285744711756706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15031014382839203, + 0.1538430154323578, + 0.07689189165830612, + 0.07187458127737045, + 0.027069473639130592, + 0.04739079624414444, + 0.039755914360284805, + 0.07265474647283554, + 0.13210700452327728, + 0.22810247540473938 + ], + "credit_map": [ + 0.026245975866913795, + 0.019760631024837494, + 0.011844287626445293, + 0.022430989891290665, + 0.013432417996227741, + 0.01623104140162468, + 0.02670624665915966, + 0.013658546842634678, + 0.02061963640153408, + 0.01927412673830986, + 0.02141750045120716, + 0.023658525198698044, + 0.0151804955676198, + 0.016698414459824562, + 0.01564444974064827, + 0.02134985290467739, + 0.008974511176347733, + 0.006850740872323513, + 0.005587012507021427, + 0.01753070577979088, + 0.008969400078058243, + 0.009505091235041618, + 0.007998884655535221, + 0.011475537903606892, + 0.007433158345520496, + 0.006499412003904581, + 0.004420943558216095, + 0.009669741615653038, + 0.01574883610010147, + 0.012046199291944504, + 0.007203118875622749, + 0.008853169158101082, + 0.0025656826328486204, + 0.0023457370698451996, + 0.002309056231752038, + 0.004166450817137957, + 0.0032105217687785625, + 0.00551844947040081, + 0.00324730365537107, + 0.0037062678020447493, + 0.007756775710731745, + 0.006086524575948715, + 0.0036607803776860237, + 0.006563177332282066, + 0.006295420229434967, + 0.0058874147944152355, + 0.005040869582444429, + 0.0060998303815722466, + 0.0034819492138922215, + 0.003094262443482876, + 0.00275945826433599, + 0.00683442922309041, + 0.00689969165250659, + 0.005792531184852123, + 0.004059810657054186, + 0.006833774968981743, + 0.007700162939727306, + 0.007375857327133417, + 0.008066418580710888, + 0.01002784725278616, + 0.009479088708758354, + 0.009252993389964104, + 0.009072219021618366, + 0.01168015692383051, + 0.017365433275699615, + 0.012999004684388638, + 0.01137460209429264, + 0.017099492251873016, + 0.015065115876495838, + 0.015473752282559872, + 0.01856837421655655, + 0.024161221459507942, + 0.022163521498441696, + 0.024102019146084785, + 0.017459873110055923, + 0.036962974816560745, + 0.025173399597406387, + 0.03309629485011101, + 0.035310667008161545, + 0.03383370116353035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.023285744711756706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_6/index_tensor([1563, 1563])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_6/index_tensor([1563, 1563])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2a92daa954086fca5a79b3a9b75aa7dc168ca16c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_6/index_tensor([1563, 1563])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, handheld tracking shot, vibrating with every thumping footstep, follows a frantic chase along the winding, jagged ramparts of a massive ancient stone wall. The leading figure, a look of desperate panic on their face, stumbles along the weathered gray flagstones. Their long black hair whips wildly behind them, their flowing crimson silk robes billowing in the wind. Just one step behind, a second figure pursues relentlessly, their strides powerful yet measured. This pursuer, clad in a deep green silk robe wrapped around their legs, their expression one of cold, unwavering determination as they steadily close the distance. The wall itself, constructed of massive blocks of moss-speckled granite, stretches into the hazy distance, where it curves out of sight. Far below, a vast, misty valley can be seen, adding a sense of vertigo and danger. The entire scene is illuminated by the harsh golden light of the setting sun, casting long, twisting shadows and making the silk of the robes shimmer, creating an atmosphere of tense, breathless epicness.", + "seed": 20681, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_6/index_1563_seed_20681_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.36638834662735464, + "vlm_logprob": -0.36638834662735464, + "vlm_yes_no_margin": 1.825, + "vlm_token_credit": -0.36638834662735464, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3425450921058655 + ], + "opposite_logprob_tokens": [ + -1.4675450325012207 + ], + "logprob_reward": -0.3425450921058655, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.3425450921058655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2057274729013443, + 0.11925895512104034, + 0.104310043156147, + 0.06741608679294586, + 0.0518958643078804, + 0.019596658647060394, + 0.09345581382513046, + 0.11590409278869629, + 0.09733480215072632, + 0.1251002550125122 + ], + "credit_map": [ + 0.018950097262859344, + 0.02375994250178337, + 0.0275272224098444, + 0.035175107419490814, + 0.019464317709207535, + 0.034298110753297806, + 0.022312013432383537, + 0.024240640923380852, + 0.013216828927397728, + 0.013238326646387577, + 0.012202520854771137, + 0.016143465414643288, + 0.010378680191934109, + 0.026822159066796303, + 0.011902593076229095, + 0.015354377217590809, + 0.01052928064018488, + 0.01230610441416502, + 0.014508763328194618, + 0.01392059214413166, + 0.014593509025871754, + 0.010946261696517467, + 0.014161231927573681, + 0.013344296254217625, + 0.008055801503360271, + 0.009824615903198719, + 0.008987723849713802, + 0.006060115061700344, + 0.008430154994130135, + 0.007562657352536917, + 0.009246771223843098, + 0.009248248301446438, + 0.006905869580805302, + 0.007603522855788469, + 0.007314520888030529, + 0.006161614321172237, + 0.0050667994655668736, + 0.005296526942402124, + 0.008162995800375938, + 0.005384011659771204, + 0.0026682799216359854, + 0.0027291567530483007, + 0.0037648852448910475, + 0.0022207521833479404, + 0.002224881434813142, + 0.001574010937474668, + 0.0025377736892551184, + 0.0018769194139167666, + 0.01900182105600834, + 0.012072829529643059, + 0.012187132611870766, + 0.008076903410255909, + 0.010774649679660797, + 0.01284931693226099, + 0.007480295840650797, + 0.011012861505150795, + 0.013552463613450527, + 0.012793200090527534, + 0.03056720271706581, + 0.0067789084278047085, + 0.009508924558758736, + 0.023140328004956245, + 0.012288345955312252, + 0.007274723146110773, + 0.009256544522941113, + 0.013747524470090866, + 0.013001097366213799, + 0.010535352863371372, + 0.01232905127108097, + 0.013419306837022305, + 0.012001441791653633, + 0.013044471852481365, + 0.01855815388262272, + 0.017594551667571068, + 0.01979880779981613, + 0.01468399353325367, + 0.015050766989588737, + 0.012644406408071518, + 0.01845616288483143, + 0.008313402533531189 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.3425450921058655 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134866952896118 + ], + "opposite_logprob_tokens": [ + -0.3134867250919342 + ], + "logprob_reward": -1.3134866952896118, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134866952896118, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16874384880065918, + 0.06874425709247589, + 0.14397087693214417, + 0.05928491801023483, + 0.08903948217630386, + 0.05060110241174698, + 0.1026039719581604, + 0.10962339490652084, + 0.12226283550262451, + 0.08512532711029053 + ], + "credit_map": [ + 0.021120548248291016, + 0.02471405640244484, + 0.01997997798025608, + 0.025154294446110725, + 0.016204040497541428, + 0.02157917432487011, + 0.02277451381087303, + 0.0172172449529171, + 0.008592230267822742, + 0.008864734321832657, + 0.008916180580854416, + 0.010170546360313892, + 0.005913530942052603, + 0.010250849649310112, + 0.008041487075388432, + 0.0079946955665946, + 0.012634340673685074, + 0.010447303764522076, + 0.015404902398586273, + 0.0258643738925457, + 0.019965078681707382, + 0.017899589613080025, + 0.030262509360909462, + 0.011492766439914703, + 0.008144560270011425, + 0.007721857633441687, + 0.007056610658764839, + 0.007588161621242762, + 0.005149441305547953, + 0.007074599154293537, + 0.008027295581996441, + 0.008522388525307178, + 0.008441749028861523, + 0.009461824782192707, + 0.008672282099723816, + 0.010792107321321964, + 0.015796130523085594, + 0.00883758906275034, + 0.016729727387428284, + 0.010308065451681614, + 0.005031592212617397, + 0.005532355047762394, + 0.005430145654827356, + 0.004850981291383505, + 0.00682466384023428, + 0.0049165100790560246, + 0.0066969143226742744, + 0.011317942291498184, + 0.014237177558243275, + 0.01199109572917223, + 0.012233826331794262, + 0.010319744236767292, + 0.011181202717125416, + 0.015409107320010662, + 0.013779304921627045, + 0.013452514074742794, + 0.01621168851852417, + 0.02310864068567753, + 0.016061361879110336, + 0.007771458942443132, + 0.012638835236430168, + 0.013059709221124649, + 0.011979196220636368, + 0.008792499080300331, + 0.01542428694665432, + 0.021879570558667183, + 0.01129530742764473, + 0.012839371338486671, + 0.010745702311396599, + 0.019957344979047775, + 0.012265277095139027, + 0.017855968326330185, + 0.012342577800154686, + 0.010695978067815304, + 0.011308216489851475, + 0.01290876790881157, + 0.010553768835961819, + 0.00845283456146717, + 0.009434400126338005, + 0.009428782388567924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -1.3134866952896118 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020608583465218544 + ], + "opposite_logprob_tokens": [ + -3.89560866355896 + ], + "logprob_reward": -0.020608583465218544, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020608583465218544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1402571201324463, + 0.0957207903265953, + 0.048280030488967896, + 0.09626509249210358, + 0.20937703549861908, + 0.0268594641238451, + 0.03425842151045799, + 0.065635547041893, + 0.1211690828204155, + 0.1621774137020111 + ], + "credit_map": [ + 0.013124639168381691, + 0.01601109467446804, + 0.012246872298419476, + 0.01709207147359848, + 0.01973540149629116, + 0.027846191078424454, + 0.017330538481473923, + 0.01687031053006649, + 0.01065990049391985, + 0.013157122768461704, + 0.013085372745990753, + 0.01643223874270916, + 0.00882797408849001, + 0.01313376147300005, + 0.009395482018589973, + 0.011028933338820934, + 0.005599822849035263, + 0.004979622550308704, + 0.005659477319568396, + 0.004990819375962019, + 0.00819128192961216, + 0.005902098957449198, + 0.009208442643284798, + 0.003748466493561864, + 0.014048962853848934, + 0.014963612891733646, + 0.013157924637198448, + 0.008628925308585167, + 0.013792072422802448, + 0.009827690199017525, + 0.009492526762187481, + 0.01235336996614933, + 0.023334737867116928, + 0.027937646955251694, + 0.021381022408604622, + 0.022494520992040634, + 0.03514320030808449, + 0.019970610737800598, + 0.036046549677848816, + 0.023068752139806747, + 0.0032859293278306723, + 0.003798840567469597, + 0.0035945228300988674, + 0.0037135789170861244, + 0.0027645498048514128, + 0.0022167200222611427, + 0.004780481103807688, + 0.002704842947423458, + 0.004106857348233461, + 0.003629917511716485, + 0.004314842168241739, + 0.004773680120706558, + 0.005491571966558695, + 0.0034376869443804026, + 0.00491480203345418, + 0.0035890648141503334, + 0.014916184358298779, + 0.00885030347853899, + 0.011952601373195648, + 0.004680993501096964, + 0.00715238694101572, + 0.008331190794706345, + 0.005518884863704443, + 0.004233000334352255, + 0.015234313905239105, + 0.026930632069706917, + 0.011859084479510784, + 0.013860651291906834, + 0.009015833958983421, + 0.0156695656478405, + 0.011626037769019604, + 0.01697295717895031, + 0.017006946727633476, + 0.01829645410180092, + 0.01854858361184597, + 0.019182413816452026, + 0.02458852156996727, + 0.01619822531938553, + 0.03517764061689377, + 0.01317862793803215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.020608583465218544 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497325211763382 + ], + "opposite_logprob_tokens": [ + -2.9299733638763428 + ], + "logprob_reward": -0.05497325211763382, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497325211763382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1264151930809021, + 0.06590422242879868, + 0.23175100982189178, + 0.047538843005895615, + 0.042755112051963806, + 0.027788609266281128, + 0.027045348659157753, + 0.04068019613623619, + 0.05948827788233757, + 0.33063316345214844 + ], + "credit_map": [ + 0.012822752818465233, + 0.014605676755309105, + 0.007229125592857599, + 0.01031166035681963, + 0.02299891598522663, + 0.024163249880075455, + 0.01881391741335392, + 0.015469906851649284, + 0.006686537526547909, + 0.006014945916831493, + 0.007145555689930916, + 0.00974329188466072, + 0.006072983145713806, + 0.01325717568397522, + 0.009728687815368176, + 0.007255044300109148, + 0.023499181494116783, + 0.03513569012284279, + 0.035801514983177185, + 0.036166008561849594, + 0.026319056749343872, + 0.02458254061639309, + 0.019923105835914612, + 0.030323918908834457, + 0.004551613237708807, + 0.004998638294637203, + 0.004052110947668552, + 0.004706615582108498, + 0.009306039661169052, + 0.005796876735985279, + 0.007299283519387245, + 0.006827667821198702, + 0.006869426928460598, + 0.005511988885700703, + 0.0050002639181911945, + 0.004898742772638798, + 0.00583562720566988, + 0.0043877544812858105, + 0.005574679467827082, + 0.00467663025483489, + 0.003343688789755106, + 0.004056344740092754, + 0.003938750363886356, + 0.00349565502256155, + 0.004284876864403486, + 0.002397279953584075, + 0.003754840698093176, + 0.0025171737652271986, + 0.004068995360285044, + 0.0031676373910158873, + 0.0035677372943609953, + 0.002096805488690734, + 0.002711246022954583, + 0.0034225161653012037, + 0.0028800794389098883, + 0.005130333825945854, + 0.00782831385731697, + 0.006024325732141733, + 0.005457252264022827, + 0.0029283587355166674, + 0.004819178022444248, + 0.005756609607487917, + 0.004319935105741024, + 0.0035462251398712397, + 0.0064432453364133835, + 0.008548271842300892, + 0.007702256087213755, + 0.007030963897705078, + 0.005938492249697447, + 0.008630676195025444, + 0.006493303459137678, + 0.008701075799763203, + 0.03666654974222183, + 0.036403149366378784, + 0.03996440768241882, + 0.036012712866067886, + 0.04926954209804535, + 0.03392898291349411, + 0.07012619823217392, + 0.028261661529541016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.05497325211763382 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032811015844345 + ], + "opposite_logprob_tokens": [ + -2.350328207015991 + ], + "logprob_reward": -0.10032811015844345, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032811015844345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2790047228336334, + 0.2664092481136322, + 0.06512147933244705, + 0.03417166322469711, + 0.027946103364229202, + 0.029908640310168266, + 0.03304936736822128, + 0.1113559827208519, + 0.06468508392572403, + 0.08834777027368546 + ], + "credit_map": [ + 0.031601183116436005, + 0.04229729622602463, + 0.042039766907691956, + 0.04362865909934044, + 0.02434476837515831, + 0.03784409537911415, + 0.02828642539680004, + 0.028962478041648865, + 0.041391950100660324, + 0.03446243703365326, + 0.025194725021719933, + 0.03551442548632622, + 0.030986705794930458, + 0.043785981833934784, + 0.02271856740117073, + 0.03235441446304321, + 0.006328878924250603, + 0.0077786254696547985, + 0.008681323379278183, + 0.010276105254888535, + 0.006136591080576181, + 0.008083898574113846, + 0.006160486023873091, + 0.011675565503537655, + 0.005428348667919636, + 0.005174280144274235, + 0.003918248228728771, + 0.004273776896297932, + 0.0030696727335453033, + 0.00371245457790792, + 0.003945366479456425, + 0.0046495129354298115, + 0.004043033812195063, + 0.003074799431487918, + 0.0041084447875618935, + 0.002997081261128187, + 0.003159942338243127, + 0.002651969902217388, + 0.004556598607450724, + 0.0033542325254529715, + 0.004383380524814129, + 0.004583686590194702, + 0.005160083062946796, + 0.00398597214370966, + 0.003047588746994734, + 0.002477998146787286, + 0.0028959279879927635, + 0.0033740003127604723, + 0.006375879980623722, + 0.004789203405380249, + 0.004497174173593521, + 0.002644563326612115, + 0.004732423461973667, + 0.0037968838587403297, + 0.0025135567411780357, + 0.003699677996337414, + 0.020877929404377937, + 0.017355553805828094, + 0.02503473125398159, + 0.00652637192979455, + 0.006607120390981436, + 0.01124597992748022, + 0.012340015731751919, + 0.011368270963430405, + 0.008773268200457096, + 0.010508476756513119, + 0.00694408779963851, + 0.011261564679443836, + 0.005934441462159157, + 0.006965022999793291, + 0.006784482393413782, + 0.007513733115047216, + 0.0140644870698452, + 0.01281043142080307, + 0.015628665685653687, + 0.011770571582019329, + 0.010252663865685463, + 0.007090007420629263, + 0.009302941150963306, + 0.007428003009408712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.10032811015844345 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, handheld tracking shot, vibrating with every thumping footstep, follows a frantic chase along the winding, jagged ramparts of a massive ancient stone wall. The leading figure, a look of desperate panic on their face, stumbles along the weathered gray flagstones. Their long black hair whips wildly behind them, their flowing crimson silk robes billowing in the wind. Just one step behind, a second figure pursues relentlessly, their strides powerful yet measured. This pursuer, clad in a deep green silk robe wrapped around their legs, their expression one of cold, unwavering determination as they steadily close the distance. The wall itself, constructed of massive blocks of moss-speckled granite, stretches into the hazy distance, where it curves out of sight. Far below, a vast, misty valley can be seen, adding a sense of vertigo and danger. The entire scene is illuminated by the harsh golden light of the setting sun, casting long, twisting shadows and making the silk of the robes shimmer, creating an atmosphere of tense, breathless epicness.", + "seed": 20682, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_6/index_1563_seed_20682_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2886002141982317, + "vlm_logprob": -0.2886002141982317, + "vlm_yes_no_margin": 2.075, + "vlm_token_credit": -0.2886002141982317, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3141375482082367 + ], + "opposite_logprob_tokens": [ + -1.564137578010559 + ], + "logprob_reward": -0.3141375482082367, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.3141375482082367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18673163652420044, + 0.22827492654323578, + 0.0838402807712555, + 0.06350959092378616, + 0.11770714819431305, + 0.11321417987346649, + 0.06032981351017952, + 0.059845469892024994, + 0.034391824156045914, + 0.05215512588620186 + ], + "credit_map": [ + 0.019567091017961502, + 0.016154952347278595, + 0.017167439684271812, + 0.02229730412364006, + 0.026940245181322098, + 0.03375742584466934, + 0.026140861213207245, + 0.024706315249204636, + 0.01620432548224926, + 0.01103348657488823, + 0.01424388773739338, + 0.012992849573493004, + 0.0183151476085186, + 0.02306513860821724, + 0.06527084112167358, + 0.06714924424886703, + 0.011845171451568604, + 0.006734856870025396, + 0.004189937841147184, + 0.006270669400691986, + 0.017062559723854065, + 0.014867492951452732, + 0.010705113410949707, + 0.012164478190243244, + 0.004322362132370472, + 0.004394769202917814, + 0.0030593827832490206, + 0.006116097792983055, + 0.005753671284765005, + 0.008666968904435635, + 0.019252700731158257, + 0.011943640187382698, + 0.011624066159129143, + 0.00811417493969202, + 0.00897514633834362, + 0.011738617904484272, + 0.01654747501015663, + 0.00851931981742382, + 0.029288915917277336, + 0.022899437695741653, + 0.011616966687142849, + 0.00946518499404192, + 0.0064237164333462715, + 0.012440614402294159, + 0.009137971326708794, + 0.03238243609666824, + 0.015569751150906086, + 0.01617753691971302, + 0.012960890308022499, + 0.009806585498154163, + 0.007722019217908382, + 0.004203336779028177, + 0.0042659505270421505, + 0.009815637953579426, + 0.004103043116629124, + 0.0074523501098155975, + 0.005218033213168383, + 0.006821011658757925, + 0.007305591832846403, + 0.007909342646598816, + 0.010612149722874165, + 0.008456655777990818, + 0.008396695367991924, + 0.005125988740473986, + 0.003990147728472948, + 0.002954113995656371, + 0.00459030456840992, + 0.0037128115072846413, + 0.004761385731399059, + 0.003765673143789172, + 0.004578313324600458, + 0.006039073690772057, + 0.006063269916921854, + 0.006238170433789492, + 0.006654943805187941, + 0.004758440423756838, + 0.005247332155704498, + 0.00807239580899477, + 0.007007679436355829, + 0.008112894371151924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.3141375482082367 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.974357008934021 + ], + "opposite_logprob_tokens": [ + -0.474357008934021 + ], + "logprob_reward": -0.974357008934021, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.974357008934021, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13599172234535217, + 0.1703862100839615, + 0.05268173664808273, + 0.05138106271624565, + 0.1023976281285286, + 0.2536783218383789, + 0.050350502133369446, + 0.053877655416727066, + 0.07070031017065048, + 0.05855486914515495 + ], + "credit_map": [ + 0.012381755746901035, + 0.016528883948922157, + 0.015675421804189682, + 0.014136144891381264, + 0.01768466643989086, + 0.016551068052649498, + 0.022409502416849136, + 0.020624274387955666, + 0.01079257670789957, + 0.013666831888258457, + 0.010781586170196533, + 0.014562874101102352, + 0.02127867378294468, + 0.030280154198408127, + 0.03687738999724388, + 0.032146111130714417, + 0.0070487745106220245, + 0.0054060472175478935, + 0.0030609890818595886, + 0.00506920600309968, + 0.00893424917012453, + 0.010229140520095825, + 0.006042675580829382, + 0.006890647578984499, + 0.005683210212737322, + 0.00694448733702302, + 0.0037514767609536648, + 0.0066566672176122665, + 0.005270762834697962, + 0.006877506151795387, + 0.009541377425193787, + 0.006655571050941944, + 0.011145352385938168, + 0.010080559179186821, + 0.010090448893606663, + 0.012459308840334415, + 0.010384678840637207, + 0.007950187660753727, + 0.019164403900504112, + 0.021122677251696587, + 0.038730278611183167, + 0.04201169312000275, + 0.02441140078008175, + 0.02672729268670082, + 0.027541369199752808, + 0.034402839839458466, + 0.02645617537200451, + 0.033397264778614044, + 0.008427903987467289, + 0.0077293915674090385, + 0.006574896164238453, + 0.004529264289885759, + 0.003820945043116808, + 0.008320576511323452, + 0.0051667834632098675, + 0.005780735053122044, + 0.00505865179002285, + 0.0067526050843298435, + 0.0037488897796720266, + 0.013813179917633533, + 0.005275354720652103, + 0.0062887463718652725, + 0.008864926174283028, + 0.004075296688824892, + 0.007739110849797726, + 0.00685827387496829, + 0.0058830115012824535, + 0.007560959551483393, + 0.006930679548531771, + 0.009535384364426136, + 0.013962988741695881, + 0.012229901738464832, + 0.007954953238368034, + 0.007838246412575245, + 0.007399560417979956, + 0.006374443881213665, + 0.005160221830010414, + 0.006561566609889269, + 0.008760317228734493, + 0.008505556732416153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.974357008934021 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016088662669062614 + ], + "opposite_logprob_tokens": [ + -4.141088485717773 + ], + "logprob_reward": -0.016088662669062614, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016088662669062614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13970662653446198, + 0.15633708238601685, + 0.09646520018577576, + 0.048510175198316574, + 0.15201465785503387, + 0.07950133085250854, + 0.07992862164974213, + 0.07329489290714264, + 0.10495474189519882, + 0.06928668171167374 + ], + "credit_map": [ + 0.012980086728930473, + 0.016695983707904816, + 0.008788836188614368, + 0.013595282100141048, + 0.021106837317347527, + 0.026227101683616638, + 0.02393542230129242, + 0.016377072781324387, + 0.014174438081681728, + 0.010863215662539005, + 0.011329182423651218, + 0.013340034522116184, + 0.014335596933960915, + 0.019526973366737366, + 0.03749823570251465, + 0.03526940569281578, + 0.010198003612458706, + 0.010457715950906277, + 0.005338906776160002, + 0.014572028070688248, + 0.015614298172295094, + 0.01218588836491108, + 0.014727785252034664, + 0.013370572589337826, + 0.0036874113138765097, + 0.0035957377403974533, + 0.0033502746373414993, + 0.006184975616633892, + 0.00621976051479578, + 0.006683748215436935, + 0.012052622623741627, + 0.00673564663156867, + 0.016531240195035934, + 0.016525693237781525, + 0.013961035758256912, + 0.02216644212603569, + 0.01756478287279606, + 0.013496946543455124, + 0.03188400715589523, + 0.0198845062404871, + 0.009180999360978603, + 0.010314247570931911, + 0.007406395860016346, + 0.010822710581123829, + 0.008527455851435661, + 0.009875645861029625, + 0.010940300300717354, + 0.012433578260242939, + 0.012981507927179337, + 0.010561609640717506, + 0.011609229259192944, + 0.007288407534360886, + 0.00885728932917118, + 0.012610668316483498, + 0.007655874360352755, + 0.008364036679267883, + 0.00889098085463047, + 0.008696029894053936, + 0.007117156405001879, + 0.01365441270172596, + 0.007097410969436169, + 0.012302249670028687, + 0.010089400224387646, + 0.005447254050523043, + 0.013293561525642872, + 0.011631862260401249, + 0.012579156085848808, + 0.01111646555364132, + 0.008794786408543587, + 0.009842483326792717, + 0.013125102035701275, + 0.024571331217885017, + 0.006446154788136482, + 0.005288609769195318, + 0.00684948917478323, + 0.008647696115076542, + 0.0070433118380606174, + 0.008933463133871555, + 0.015326014719903469, + 0.010751944966614246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.016088662669062614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810602053999901 + ], + "opposite_logprob_tokens": [ + -3.2881059646606445 + ], + "logprob_reward": -0.03810602053999901, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810602053999901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18994781374931335, + 0.13852132856845856, + 0.10556862503290176, + 0.053019288927316666, + 0.07322733104228973, + 0.16418734192848206, + 0.04664085805416107, + 0.052057649940252304, + 0.13739719986915588, + 0.039432551711797714 + ], + "credit_map": [ + 0.0243973471224308, + 0.017838412895798683, + 0.011798493564128876, + 0.013237116858363152, + 0.029614346101880074, + 0.03668699041008949, + 0.031242307275533676, + 0.0251327957957983, + 0.0109109440818429, + 0.011964105069637299, + 0.00773571478202939, + 0.013370112515985966, + 0.015580304898321629, + 0.02764042280614376, + 0.02254570834338665, + 0.028774011880159378, + 0.016090059652924538, + 0.012494094669818878, + 0.006469178944826126, + 0.011585218831896782, + 0.01765548437833786, + 0.01727336272597313, + 0.011741910129785538, + 0.012259311974048615, + 0.0037402426823973656, + 0.0041940826922655106, + 0.0025549286510795355, + 0.005815243814140558, + 0.007469758857041597, + 0.009976315312087536, + 0.011537835001945496, + 0.0077308835461735725, + 0.007990619167685509, + 0.007784055545926094, + 0.00735744833946228, + 0.011814909987151623, + 0.00785725750029087, + 0.00789924617856741, + 0.01124769076704979, + 0.011276100762188435, + 0.014540565200150013, + 0.013719622045755386, + 0.008502661250531673, + 0.02192332223057747, + 0.013588136993348598, + 0.042334191501140594, + 0.0261959470808506, + 0.02338290773332119, + 0.006433295551687479, + 0.007295527029782534, + 0.00512903556227684, + 0.004383768420666456, + 0.005038298200815916, + 0.00708150677382946, + 0.005637786351144314, + 0.0056416429579257965, + 0.006335376761853695, + 0.00623953016474843, + 0.004929252900183201, + 0.005718957167118788, + 0.00631969328969717, + 0.010241171345114708, + 0.007110999897122383, + 0.005162667483091354, + 0.014425277709960938, + 0.011529752984642982, + 0.013448287732899189, + 0.024520499631762505, + 0.008551844395697117, + 0.015265760011970997, + 0.018701370805501938, + 0.030954405665397644, + 0.004709343891590834, + 0.0051232981495559216, + 0.004156044684350491, + 0.004771363455802202, + 0.0035111529286950827, + 0.0036224829964339733, + 0.007737781852483749, + 0.00580108305439353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.03810602053999901 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031183063983917 + ], + "opposite_logprob_tokens": [ + -2.350311756134033 + ], + "logprob_reward": -0.10031183063983917, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031183063983917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11902035027742386, + 0.16940133273601532, + 0.16318358480930328, + 0.023943757638335228, + 0.06100764498114586, + 0.1808302402496338, + 0.03718935325741768, + 0.04029523953795433, + 0.09069307893514633, + 0.11443540453910828 + ], + "credit_map": [ + 0.01025871280580759, + 0.012791188433766365, + 0.009295807220041752, + 0.012963556684553623, + 0.01844191737473011, + 0.022120416164398193, + 0.019254695624113083, + 0.013894070871174335, + 0.012478038668632507, + 0.007828332483768463, + 0.012018938548862934, + 0.006802727468311787, + 0.010603022761642933, + 0.01497197337448597, + 0.04988116770982742, + 0.05481714382767677, + 0.023008955642580986, + 0.014745483174920082, + 0.008353974670171738, + 0.012596690095961094, + 0.03375082090497017, + 0.030569955706596375, + 0.01863480731844902, + 0.02152291312813759, + 0.0019491238053888083, + 0.002687309868633747, + 0.0011123979929834604, + 0.0030155808199197054, + 0.0021545018535107374, + 0.003015197580680251, + 0.004640049301087856, + 0.005369599908590317, + 0.006164392456412315, + 0.005013061687350273, + 0.004960180725902319, + 0.007773844990879297, + 0.008169986307621002, + 0.005933390464633703, + 0.012586575001478195, + 0.010406220331788063, + 0.02416202425956726, + 0.013868987560272217, + 0.011777350679039955, + 0.01859232224524021, + 0.0162338949739933, + 0.04761269688606262, + 0.023741673678159714, + 0.024841301143169403, + 0.0067726313136518, + 0.006499644368886948, + 0.004192935302853584, + 0.00277884304523468, + 0.002858138643205166, + 0.005936048924922943, + 0.003832450369372964, + 0.004318666644394398, + 0.004192052874714136, + 0.004695730283856392, + 0.004752918146550655, + 0.008219962008297443, + 0.004644827451556921, + 0.005146954208612442, + 0.005355017259716988, + 0.003287780797109008, + 0.011609278619289398, + 0.008365826681256294, + 0.006540924310684204, + 0.006211504805833101, + 0.005946268327534199, + 0.010374635457992554, + 0.012373690493404865, + 0.0292709581553936, + 0.018553702160716057, + 0.018519176170229912, + 0.011740924790501595, + 0.011734116822481155, + 0.008606546558439732, + 0.010680443607270718, + 0.019950145855545998, + 0.014650351367890835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.10031183063983917 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, handheld tracking shot, vibrating with every thumping footstep, follows a frantic chase along the winding, jagged ramparts of a massive ancient stone wall. The leading figure, a look of desperate panic on their face, stumbles along the weathered gray flagstones. Their long black hair whips wildly behind them, their flowing crimson silk robes billowing in the wind. Just one step behind, a second figure pursues relentlessly, their strides powerful yet measured. This pursuer, clad in a deep green silk robe wrapped around their legs, their expression one of cold, unwavering determination as they steadily close the distance. The wall itself, constructed of massive blocks of moss-speckled granite, stretches into the hazy distance, where it curves out of sight. Far below, a vast, misty valley can be seen, adding a sense of vertigo and danger. The entire scene is illuminated by the harsh golden light of the setting sun, casting long, twisting shadows and making the silk of the robes shimmer, creating an atmosphere of tense, breathless epicness.", + "seed": 20683, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_6/index_1563_seed_20683_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2419360551983118, + "vlm_logprob": -0.2419360551983118, + "vlm_yes_no_margin": 2.575, + "vlm_token_credit": -0.2419360551983118, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24401772022247314 + ], + "opposite_logprob_tokens": [ + -1.8690177202224731 + ], + "logprob_reward": -0.24401772022247314, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.24401772022247314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23225003480911255, + 0.04346967488527298, + 0.05494558811187744, + 0.06065284088253975, + 0.05762800574302673, + 0.024792298674583435, + 0.026065386831760406, + 0.0736764669418335, + 0.05622154846787453, + 0.37029820680618286 + ], + "credit_map": [ + 0.02831631898880005, + 0.018493181094527245, + 0.018975798040628433, + 0.020845646038651466, + 0.024660062044858932, + 0.029551208019256592, + 0.03707994893193245, + 0.0543278343975544, + 0.0038335213903337717, + 0.005910771898925304, + 0.005522782914340496, + 0.007402071263641119, + 0.005510352086275816, + 0.005019778851419687, + 0.0046559595502913, + 0.005614431109279394, + 0.005982598289847374, + 0.007627094630151987, + 0.005647717043757439, + 0.00685892254114151, + 0.008975110948085785, + 0.006429015193134546, + 0.00579497404396534, + 0.00763014517724514, + 0.004562802612781525, + 0.0055701397359371185, + 0.005591813940554857, + 0.007131199352443218, + 0.0051668863743543625, + 0.007707425858825445, + 0.014196519739925861, + 0.01072604488581419, + 0.005290579050779343, + 0.01074165664613247, + 0.004187923390418291, + 0.006024315021932125, + 0.005863426718860865, + 0.00710903899744153, + 0.006848351564258337, + 0.011562708765268326, + 0.0021834399085491896, + 0.0021917582489550114, + 0.003119388362392783, + 0.002740329597145319, + 0.0030031567439436913, + 0.0025666935835033655, + 0.0031747876200824976, + 0.00581273902207613, + 0.0025758659467101097, + 0.0037760117556899786, + 0.002456813119351864, + 0.002902811858803034, + 0.0035360781475901604, + 0.0035872941371053457, + 0.00393063435330987, + 0.003299873322248459, + 0.00579017074778676, + 0.00797437597066164, + 0.006738030351698399, + 0.00650364626199007, + 0.008558019064366817, + 0.015607931651175022, + 0.014067834243178368, + 0.008436446078121662, + 0.0036094170063734055, + 0.006093377713114023, + 0.005041317082941532, + 0.014023464173078537, + 0.0067919110879302025, + 0.0061012799851596355, + 0.007923999801278114, + 0.006636770907789469, + 0.020470958203077316, + 0.10426980257034302, + 0.011421430855989456, + 0.020953064784407616, + 0.029673896729946136, + 0.1187792643904686, + 0.02169627696275711, + 0.043033480644226074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.24401772022247314 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8984063863754272 + ], + "opposite_logprob_tokens": [ + -0.5234063863754272 + ], + "logprob_reward": -0.8984063863754272, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8984063863754272, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.05397455766797066, + 0.03978126868605614, + 0.039087213575839996, + 0.05968685820698738, + 0.038528770208358765, + 0.02564235031604767, + 0.04668129235506058, + 0.02856852300465107, + 0.3170638382434845, + 0.3509852886199951 + ], + "credit_map": [ + 0.008695248514413834, + 0.006550834979861975, + 0.006825187709182501, + 0.005373150110244751, + 0.006166298408061266, + 0.006208624690771103, + 0.006595145910978317, + 0.00756006920710206, + 0.0033481239806860685, + 0.004978845827281475, + 0.003971086349338293, + 0.0072824484668672085, + 0.005197723396122456, + 0.004911709111183882, + 0.005701414775103331, + 0.004389923065900803, + 0.0028057885356247425, + 0.005113269668072462, + 0.004047734662890434, + 0.007593559566885233, + 0.004996072500944138, + 0.0038963868282735348, + 0.005157196428626776, + 0.005477207247167826, + 0.004277231637388468, + 0.007001806516200304, + 0.005872348323464394, + 0.006433436181396246, + 0.007900373078882694, + 0.009548409841954708, + 0.010636789724230766, + 0.008016466163098812, + 0.0029009380377829075, + 0.0049293325282633305, + 0.0035558617673814297, + 0.0047543044202029705, + 0.004605299327522516, + 0.004854037892073393, + 0.005945513024926186, + 0.006983486469835043, + 0.0012359196553006768, + 0.004287715069949627, + 0.0023170483764261007, + 0.0042138067074120045, + 0.003590641776099801, + 0.003115659346804023, + 0.0037153742741793394, + 0.0031661870889365673, + 0.007991346530616283, + 0.00515183387324214, + 0.005491521675139666, + 0.004049140494316816, + 0.004925393033772707, + 0.005738559644669294, + 0.006459371652454138, + 0.006874124053865671, + 0.00218785903416574, + 0.0024996157735586166, + 0.0020702879410237074, + 0.002163498429581523, + 0.003759040730074048, + 0.007963277399539948, + 0.004615965764969587, + 0.0033089795615524054, + 0.02157786674797535, + 0.03781724348664284, + 0.02232098951935768, + 0.09239239990711212, + 0.046164315193891525, + 0.03132203593850136, + 0.03173499181866646, + 0.033734001219272614, + 0.021423719823360443, + 0.11045393347740173, + 0.010127943009138107, + 0.014281322248280048, + 0.03358868509531021, + 0.10806523263454437, + 0.02238072082400322, + 0.030663756653666496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.8984063863754272 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011097259819507599 + ], + "opposite_logprob_tokens": [ + -4.511097431182861 + ], + "logprob_reward": -0.011097259819507599, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011097259819507599, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0604134239256382, + 0.048102159053087234, + 0.030079061165452003, + 0.07539739459753036, + 0.04727933928370476, + 0.04727538675069809, + 0.034098122268915176, + 0.03607301041483879, + 0.08416704833507538, + 0.5371150970458984 + ], + "credit_map": [ + 0.007228872273117304, + 0.008751986548304558, + 0.00490628881379962, + 0.006862335838377476, + 0.008368593640625477, + 0.00632287235930562, + 0.007450167555361986, + 0.010522306896746159, + 0.0044946069829165936, + 0.005984561052173376, + 0.004468815866857767, + 0.00501528475433588, + 0.007099186070263386, + 0.005962701980024576, + 0.008684649132192135, + 0.006392352748662233, + 0.0034139580093324184, + 0.0037993991281837225, + 0.0032119450625032187, + 0.003043248550966382, + 0.005117000080645084, + 0.003718713065609336, + 0.0035045379772782326, + 0.00427025742828846, + 0.003885847982019186, + 0.007968710735440254, + 0.006906208116561174, + 0.014119166880846024, + 0.00784300547093153, + 0.011179222725331783, + 0.011785749346017838, + 0.011709478683769703, + 0.0034512558486312628, + 0.006735760718584061, + 0.004189465660601854, + 0.006159900687634945, + 0.005331115331500769, + 0.008226838894188404, + 0.007971891202032566, + 0.005213112570345402, + 0.0042424011044204235, + 0.009156721644103527, + 0.00405028136447072, + 0.006324640940874815, + 0.005163780879229307, + 0.0050260149873793125, + 0.005798512138426304, + 0.007513031363487244, + 0.0036923682782799006, + 0.004098550882190466, + 0.0030417228117585182, + 0.004279382061213255, + 0.005850386340171099, + 0.004258880857378244, + 0.0054054418578743935, + 0.0034713882487267256, + 0.0032120842952281237, + 0.0038937798235565424, + 0.004339850042015314, + 0.0036429560277611017, + 0.006494521629065275, + 0.004848822019994259, + 0.005720686633139849, + 0.003920310642570257, + 0.004979384131729603, + 0.006962508894503117, + 0.0073655047453939915, + 0.02899322286248207, + 0.009871718473732471, + 0.006586266215890646, + 0.0095206992700696, + 0.009887737222015858, + 0.04011573642492294, + 0.18417039513587952, + 0.01821763999760151, + 0.028641963377594948, + 0.06348148733377457, + 0.13204504549503326, + 0.03331650048494339, + 0.03712627291679382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.011097259819507599 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026359083130955696 + ], + "opposite_logprob_tokens": [ + -3.6513590812683105 + ], + "logprob_reward": -0.026359083130955696, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026359083130955696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0517900325357914, + 0.03545320779085159, + 0.030495265498757362, + 0.030653774738311768, + 0.036201659590005875, + 0.018920237198472023, + 0.023213375359773636, + 0.06969433277845383, + 0.04311374947428703, + 0.660464346408844 + ], + "credit_map": [ + 0.005984602961689234, + 0.006790021434426308, + 0.0048380340449512005, + 0.005603657104074955, + 0.008368652313947678, + 0.0062619163654744625, + 0.006062716711312532, + 0.007880430668592453, + 0.003286753548309207, + 0.0047565000131726265, + 0.004477779846638441, + 0.004215186927467585, + 0.0037489214446395636, + 0.0037264495622366667, + 0.0059705921448767185, + 0.005271020345389843, + 0.003121286164969206, + 0.0038693868555128574, + 0.0026054191403090954, + 0.003141289111226797, + 0.005768720526248217, + 0.003818579949438572, + 0.003991263452917337, + 0.004179318901151419, + 0.0019327328773215413, + 0.003375005442649126, + 0.0029235759284347296, + 0.0033348656725138426, + 0.0040858048014342785, + 0.004839790519326925, + 0.00586505513638258, + 0.004296939820051193, + 0.0031126090325415134, + 0.006583559792488813, + 0.0026733980048447847, + 0.004510693717747927, + 0.0040048351511359215, + 0.005853857845067978, + 0.005055570974946022, + 0.004407132510095835, + 0.0017420966178178787, + 0.0025986325927078724, + 0.0023793750442564487, + 0.002550034085288644, + 0.002285287482663989, + 0.0020743426866829395, + 0.0022893850691616535, + 0.003001083852723241, + 0.002170366467908025, + 0.002690580440685153, + 0.0027581884060055017, + 0.0029256828129291534, + 0.0029429825954139233, + 0.003037736052647233, + 0.003480922430753708, + 0.0032069168519228697, + 0.005883613135665655, + 0.008569233119487762, + 0.0048803710378706455, + 0.006333349738270044, + 0.00926106609404087, + 0.01844404824078083, + 0.010997497476637363, + 0.005325156729668379, + 0.0034222318790853024, + 0.004847061354666948, + 0.0036140261217951775, + 0.009501230902969837, + 0.005887145176529884, + 0.005303322337567806, + 0.005420027766376734, + 0.005118698813021183, + 0.03311292827129364, + 0.21470756828784943, + 0.0209719967097044, + 0.03245724365115166, + 0.07115275412797928, + 0.19282905757427216, + 0.03736128658056259, + 0.05787152051925659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.026359083130955696 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799826443195343 + ], + "opposite_logprob_tokens": [ + -3.5297999382019043 + ], + "logprob_reward": -0.029799826443195343, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799826443195343, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09199556708335876, + 0.15750819444656372, + 0.26287561655044556, + 0.12032279372215271, + 0.0261849258095026, + 0.06346388161182404, + 0.021511584520339966, + 0.04992450773715973, + 0.053321581333875656, + 0.15289129316806793 + ], + "credit_map": [ + 0.009462786838412285, + 0.010180758312344551, + 0.01131375040858984, + 0.011885483749210835, + 0.013850905001163483, + 0.011616501957178116, + 0.012394865043461323, + 0.011290515772998333, + 0.014751577749848366, + 0.0216955728828907, + 0.016538003459572792, + 0.020255329087376595, + 0.023821014910936356, + 0.020675156265497208, + 0.0233949925750494, + 0.016376547515392303, + 0.023226577788591385, + 0.028904762119054794, + 0.01925247348845005, + 0.03194858133792877, + 0.04266241937875748, + 0.03249378129839897, + 0.03965461626648903, + 0.04473240673542023, + 0.010213273577392101, + 0.008027318865060806, + 0.013500489294528961, + 0.00819788221269846, + 0.01313725020736456, + 0.016915734857320786, + 0.02711537852883339, + 0.02321547456085682, + 0.0024637773167341948, + 0.0031003279145807028, + 0.002655317774042487, + 0.0033348323777318, + 0.0033151041716337204, + 0.004152051638811827, + 0.0038096189964562654, + 0.0033538956195116043, + 0.0063844420947134495, + 0.004413486458361149, + 0.011173154227435589, + 0.005666440352797508, + 0.007433424703776836, + 0.00843882281333208, + 0.006221216171979904, + 0.013732895255088806, + 0.002408545697107911, + 0.002251444384455681, + 0.0034226758871227503, + 0.002078880090266466, + 0.0022220287937670946, + 0.002960142446681857, + 0.0027366119902580976, + 0.0034312554635107517, + 0.003948837984353304, + 0.00473977904766798, + 0.005545462481677532, + 0.003976474981755018, + 0.0043400730937719345, + 0.013507629744708538, + 0.008978554978966713, + 0.0048876977525651455, + 0.003826417727395892, + 0.005143600981682539, + 0.003711574710905552, + 0.011864688247442245, + 0.006540143862366676, + 0.008054534904658794, + 0.008177265524864197, + 0.0060033537447452545, + 0.006868455559015274, + 0.036258064210414886, + 0.004884714726358652, + 0.007702955510467291, + 0.009116962552070618, + 0.05382862687110901, + 0.010150967165827751, + 0.024080554023385048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.029799826443195343 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, handheld tracking shot, vibrating with every thumping footstep, follows a frantic chase along the winding, jagged ramparts of a massive ancient stone wall. The leading figure, a look of desperate panic on their face, stumbles along the weathered gray flagstones. Their long black hair whips wildly behind them, their flowing crimson silk robes billowing in the wind. Just one step behind, a second figure pursues relentlessly, their strides powerful yet measured. This pursuer, clad in a deep green silk robe wrapped around their legs, their expression one of cold, unwavering determination as they steadily close the distance. The wall itself, constructed of massive blocks of moss-speckled granite, stretches into the hazy distance, where it curves out of sight. Far below, a vast, misty valley can be seen, adding a sense of vertigo and danger. The entire scene is illuminated by the harsh golden light of the setting sun, casting long, twisting shadows and making the silk of the robes shimmer, creating an atmosphere of tense, breathless epicness.", + "seed": 20684, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_6/index_1563_seed_20684_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.39785022661089897, + "vlm_logprob": -0.39785022661089897, + "vlm_yes_no_margin": 1.675, + "vlm_token_credit": -0.39785022661089897, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4505142271518707 + ], + "opposite_logprob_tokens": [ + -1.2005141973495483 + ], + "logprob_reward": -0.4505142271518707, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.4505142271518707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14821629226207733, + 0.0848417803645134, + 0.07626321166753769, + 0.08294246345758438, + 0.06273382157087326, + 0.31370753049850464, + 0.05692196264863014, + 0.042897794395685196, + 0.08784162998199463, + 0.04363350570201874 + ], + "credit_map": [ + 0.01673160493373871, + 0.022037312388420105, + 0.018097322434186935, + 0.013995452784001827, + 0.013476766645908356, + 0.021265216171741486, + 0.021821262314915657, + 0.02079135552048683, + 0.012426275759935379, + 0.009636112488806248, + 0.00984311755746603, + 0.00860145129263401, + 0.012149833142757416, + 0.009901056066155434, + 0.0073625994846224785, + 0.014921333640813828, + 0.01472022756934166, + 0.008812344633042812, + 0.01201207097619772, + 0.006901040207594633, + 0.006793953478336334, + 0.008724281564354897, + 0.006684556603431702, + 0.011614745482802391, + 0.01503242738544941, + 0.0122971897944808, + 0.01380414143204689, + 0.008527755737304688, + 0.008668722584843636, + 0.0076844776049256325, + 0.007253363728523254, + 0.009674389846622944, + 0.01085338182747364, + 0.00621839938685298, + 0.010570764541625977, + 0.006820726674050093, + 0.005242160987108946, + 0.008059434592723846, + 0.006405151914805174, + 0.008563803508877754, + 0.0688176080584526, + 0.010420643724501133, + 0.008314193226397038, + 0.012148488312959671, + 0.016775846481323242, + 0.1306023746728897, + 0.024948105216026306, + 0.04168030247092247, + 0.010075788013637066, + 0.0063278027810156345, + 0.0124321598559618, + 0.004822675138711929, + 0.004807938355952501, + 0.007606314495205879, + 0.004496711306273937, + 0.006352575030177832, + 0.006068601738661528, + 0.003777815029025078, + 0.005000435747206211, + 0.00260351225733757, + 0.007288492750376463, + 0.004287592601031065, + 0.007314629852771759, + 0.006556715816259384, + 0.020661428570747375, + 0.006544058211147785, + 0.006668413057923317, + 0.004280018620193005, + 0.008248117752373219, + 0.027830690145492554, + 0.00670105405151844, + 0.006907857954502106, + 0.005821376573294401, + 0.0036906374152749777, + 0.0054021128453314304, + 0.0036647438537329435, + 0.004472484812140465, + 0.007390022277832031, + 0.005298183765262365, + 0.007893948815762997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.4505142271518707 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2237173318862915 + ], + "opposite_logprob_tokens": [ + -0.3487173020839691 + ], + "logprob_reward": -1.2237173318862915, + "yes_no_margin_reward": -0.875, + "token_credit_reward": -1.2237173318862915, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10786569118499756, + 0.04874777793884277, + 0.029610751196742058, + 0.026480237022042274, + 0.05135594680905342, + 0.5253216028213501, + 0.02788630872964859, + 0.029529262334108353, + 0.12128240615129471, + 0.03192000836133957 + ], + "credit_map": [ + 0.013387417420744896, + 0.01428975723683834, + 0.012929020449519157, + 0.013281730934977531, + 0.01600460335612297, + 0.012122862040996552, + 0.010045386850833893, + 0.015804922208189964, + 0.005864751525223255, + 0.0052767591550946236, + 0.0060316030867397785, + 0.006884685717523098, + 0.005320875905454159, + 0.005317555274814367, + 0.006387954577803612, + 0.007663595024496317, + 0.005829906556755304, + 0.0032072863541543484, + 0.0033685173839330673, + 0.0029782920610159636, + 0.003776474157348275, + 0.0034131885040551424, + 0.0033330495934933424, + 0.003704037517309189, + 0.0040658144280314445, + 0.00337585317902267, + 0.0028002269100397825, + 0.0026822597719728947, + 0.00408373586833477, + 0.002588593866676092, + 0.0037318493705242872, + 0.0031519050244241953, + 0.008437533862888813, + 0.007159724831581116, + 0.007629773113876581, + 0.00485313730314374, + 0.005055552814155817, + 0.006036831997334957, + 0.006373940967023373, + 0.005809454247355461, + 0.11766674369573593, + 0.016048481687903404, + 0.01327610109001398, + 0.01970217376947403, + 0.027076356112957, + 0.22076182067394257, + 0.040790658444166183, + 0.06999927759170532, + 0.0048409150913357735, + 0.003340004710480571, + 0.004424508195370436, + 0.0019586049020290375, + 0.0023000503424555063, + 0.004411887843161821, + 0.0028353682719171047, + 0.0037749717012047768, + 0.003937756177037954, + 0.002164283534511924, + 0.003017611103132367, + 0.0020432269666343927, + 0.003662124741822481, + 0.007105053868144751, + 0.003204489126801491, + 0.004394718445837498, + 0.015375702641904354, + 0.015072338283061981, + 0.014329969882965088, + 0.008905020542442799, + 0.017431756481528282, + 0.024754444137215614, + 0.011904842220246792, + 0.013508344069123268, + 0.003999561537057161, + 0.006686156615614891, + 0.004533290397375822, + 0.003126393537968397, + 0.0045205275528132915, + 0.003365191863849759, + 0.002387341344729066, + 0.0033015485387295485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -1.2237173318862915 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01423114538192749 + ], + "opposite_logprob_tokens": [ + -4.264231204986572 + ], + "logprob_reward": -0.01423114538192749, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01423114538192749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14148163795471191, + 0.27424609661102295, + 0.06442292779684067, + 0.028190799057483673, + 0.13050764799118042, + 0.0859561413526535, + 0.029751628637313843, + 0.040746141225099564, + 0.09812874346971512, + 0.10656823217868805 + ], + "credit_map": [ + 0.0162343829870224, + 0.02043466828763485, + 0.018575774505734444, + 0.024622540920972824, + 0.015652567148208618, + 0.015562503598630428, + 0.013725735247135162, + 0.016673455014824867, + 0.03879844397306442, + 0.037229910492897034, + 0.024683885276317596, + 0.03679681569337845, + 0.029759962111711502, + 0.028041431680321693, + 0.03868060186505318, + 0.040255047380924225, + 0.007999857887625694, + 0.009835412725806236, + 0.005455139093101025, + 0.010916580446064472, + 0.010746221989393234, + 0.006618578918278217, + 0.008180814795196056, + 0.004670323338359594, + 0.00400752667337656, + 0.005381838884204626, + 0.003147434676066041, + 0.00300149479880929, + 0.003258425509557128, + 0.002989603905007243, + 0.0031736623495817184, + 0.003230809001252055, + 0.018109386786818504, + 0.017440181225538254, + 0.0170365609228611, + 0.014102310873568058, + 0.013540680520236492, + 0.01928040198981762, + 0.017013469710946083, + 0.013984655030071735, + 0.014503002166748047, + 0.007900899276137352, + 0.005928934086114168, + 0.006979547441005707, + 0.005557873751968145, + 0.027703864499926567, + 0.007025268394500017, + 0.010356748476624489, + 0.005001127254217863, + 0.003046905854716897, + 0.0033729923889040947, + 0.0024889707565307617, + 0.004499644506722689, + 0.004418826662003994, + 0.0032103005796670914, + 0.0037128597032278776, + 0.0040659415535628796, + 0.0034628387074917555, + 0.004986518528312445, + 0.0025948297698050737, + 0.004466171842068434, + 0.008713494054973125, + 0.004444682504981756, + 0.00801166146993637, + 0.020440394058823586, + 0.008034860715270042, + 0.007046021521091461, + 0.005995971150696278, + 0.010119443759322166, + 0.0337827168405056, + 0.006047967821359634, + 0.0066613694652915, + 0.015377267263829708, + 0.015589619055390358, + 0.01615219935774803, + 0.01425083726644516, + 0.010713678784668446, + 0.012100251391530037, + 0.008011172525584698, + 0.014373211190104485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.01423114538192749 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867305979132652 + ], + "opposite_logprob_tokens": [ + -3.048673152923584 + ], + "logprob_reward": -0.04867305979132652, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867305979132652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029266059398651123, + 0.05431576073169708, + 0.009439578279852867, + 0.01655004732310772, + 0.05726597458124161, + 0.6271820664405823, + 0.0071889194659888744, + 0.0233119186013937, + 0.07402271777391434, + 0.1014569103717804 + ], + "credit_map": [ + 0.0034351854119449854, + 0.0037194315809756517, + 0.003910220693796873, + 0.003960535395890474, + 0.004143307916820049, + 0.0033364463597536087, + 0.0031478977762162685, + 0.0036130347289144993, + 0.00812567863613367, + 0.007225342560559511, + 0.005355179775506258, + 0.006820117589086294, + 0.00635426165536046, + 0.005723078269511461, + 0.006450857501477003, + 0.008261242881417274, + 0.001324641052633524, + 0.001308363862335682, + 0.0009429503115825355, + 0.0011171553051099181, + 0.001861098106019199, + 0.0009634016896598041, + 0.0011066243750974536, + 0.0008153443341143429, + 0.0018207104876637459, + 0.0017741889460012317, + 0.0019596468191593885, + 0.0021837959066033363, + 0.002563560614362359, + 0.0016980108339339495, + 0.002712325192987919, + 0.0018378091044723988, + 0.0071095856837928295, + 0.0071644047275185585, + 0.00713261729106307, + 0.005055195651948452, + 0.008119231089949608, + 0.008662864565849304, + 0.008168007247149944, + 0.005854066926985979, + 0.13948236405849457, + 0.018160557374358177, + 0.01594453491270542, + 0.024231910705566406, + 0.032468535006046295, + 0.2641010880470276, + 0.04895094037055969, + 0.08384218066930771, + 0.0009842851432040334, + 0.0006907467031851411, + 0.001406444818712771, + 0.000570554519072175, + 0.0006859591812826693, + 0.0009579819743521512, + 0.000806915108114481, + 0.0010860313195735216, + 0.002964562037959695, + 0.00242929277010262, + 0.002227320335805416, + 0.0012559691676869988, + 0.0025787698104977608, + 0.0050912401638925076, + 0.0023194195237010717, + 0.004445345606654882, + 0.01833874173462391, + 0.004513356368988752, + 0.004307205323129892, + 0.0036647075321525335, + 0.0073808240704238415, + 0.027719883248209953, + 0.00354399043135345, + 0.004554011393338442, + 0.013305420987308025, + 0.013302771374583244, + 0.017677202820777893, + 0.013957132585346699, + 0.009768842719495296, + 0.01196067500859499, + 0.007859635166823864, + 0.013625227846205235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.04867305979132652 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera movement handheld and vibrating with every footstep?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the leading figure wearing flowing crimson silk robes and has long black hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the pursuing figure wearing a deep green silk robe and displaying an expression of cold determination?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by harsh golden light casting long, twisting shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521153688430786 + ], + "opposite_logprob_tokens": [ + -1.5021153688430786 + ], + "logprob_reward": -0.2521153688430786, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521153688430786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2388480007648468, + 0.16248773038387299, + 0.09277550131082535, + 0.04065791517496109, + 0.06664419919252396, + 0.10102279484272003, + 0.03268060460686684, + 0.04037148505449295, + 0.16929984092712402, + 0.05521199107170105 + ], + "credit_map": [ + 0.03295450285077095, + 0.026512401178479195, + 0.0425729863345623, + 0.024842729791998863, + 0.02302183397114277, + 0.027235915884375572, + 0.028124505653977394, + 0.033583104610443115, + 0.0235913023352623, + 0.013084663078188896, + 0.02493319660425186, + 0.013888844288885593, + 0.012398706749081612, + 0.01838979870080948, + 0.021877916529774666, + 0.03432328999042511, + 0.024657834321260452, + 0.013340935111045837, + 0.014565292745828629, + 0.0070724026300013065, + 0.00794976670295, + 0.0082334503531456, + 0.00819126982241869, + 0.00876455008983612, + 0.00942555908113718, + 0.005667985882610083, + 0.00584195414558053, + 0.0034279220271855593, + 0.0035174645017832518, + 0.003912434913218021, + 0.003572990885004401, + 0.005291602108627558, + 0.009920698590576649, + 0.009667065925896168, + 0.009729625657200813, + 0.007119255140423775, + 0.006998586934059858, + 0.007698060479015112, + 0.008482289500534534, + 0.007028613705188036, + 0.021854043006896973, + 0.004574881866574287, + 0.005215369164943695, + 0.004682449623942375, + 0.00597856380045414, + 0.03911949694156647, + 0.007722892332822084, + 0.011875084601342678, + 0.0041318973526358604, + 0.0032388365361839533, + 0.006326140370219946, + 0.0024329721927642822, + 0.0021829968318343163, + 0.00676735071465373, + 0.0028168167918920517, + 0.004783593583852053, + 0.005818442441523075, + 0.004586288705468178, + 0.00550735741853714, + 0.0030771340243518353, + 0.0037112305872142315, + 0.007565733045339584, + 0.004571161698549986, + 0.005534135736525059, + 0.0384761318564415, + 0.013623062521219254, + 0.008005999028682709, + 0.0072606224566698074, + 0.01755411922931671, + 0.06264374405145645, + 0.009024128317832947, + 0.012712018564343452, + 0.007555668242275715, + 0.010531977750360966, + 0.006913254037499428, + 0.008374111726880074, + 0.005019178614020348, + 0.006449648179113865, + 0.004007876385003328, + 0.006360273342579603 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.2521153688430786 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_7/index_tensor([4178, 4178])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_7/index_tensor([4178, 4178])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6d7900dbf8532cc1d27f9cfa736390203e0a193b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_7/index_tensor([4178, 4178])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera slowly and steadily zooms in, transporting us to an oriental ink painting imbued with the spirit of freehand brushwork. Beginning in a mid-ground, the scene gradually focuses on a fully-winged angel, at the apex of a mighty sword swing. The angel's form is slender and graceful, its massive wings rendered with bold, sweeping strokes. Each feather is outlined in varying shades of ink, their edges blending hazily into the abstract, light-and-dark background of abstracted space. The angel is clad in a flowing white robe, its folds traced with simple yet powerful calligraphic lines, fluttering with his movements. His focused, resolute expression radiates immense power. He clutches a simple, ancient sword in both hands, his body twisting into a tense bow. The blade slices through the air, leaving a powerful, sharp arc composed of both feibai and splashed ink techniques. The blade itself gleams with a sharp, cold gleam, punctuated by the use of white space. The entire scene is based on the subtle texture of rice paper, and the main colors are classic black, white and gray. The light does not come from an external light source, but is completely shaped by the strong contrast between large areas of white space and thick ink, creating a solemn, grand and dynamic atmosphere, as if an eternal battle moment is frozen on the scroll.", + "seed": 19896, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_7/index_4178_seed_19896_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05085400410462171, + "vlm_logprob": -0.05085400410462171, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.05085400410462171, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032794286962598562 + ], + "opposite_logprob_tokens": [ + -6.253279209136963 + ], + "logprob_reward": -0.0032794286962598562, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0032794286962598562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09453434497117996, + 0.06459439545869827, + 0.03682331740856171, + 0.13038934767246246, + 0.30957022309303284, + 0.06393688917160034, + 0.04444262757897377, + 0.030674749985337257, + 0.1427539736032486, + 0.08228018879890442 + ], + "credit_map": [ + 0.017218030989170074, + 0.006815874483436346, + 0.013842413201928139, + 0.008126867935061455, + 0.007821286097168922, + 0.012425461784005165, + 0.020321987569332123, + 0.007962421514093876, + 0.014801446348428726, + 0.00521030742675066, + 0.00753943482413888, + 0.007235926575958729, + 0.004488895647227764, + 0.006769959349185228, + 0.013916184194386005, + 0.00463224109262228, + 0.005906396545469761, + 0.0032006134279072285, + 0.006425421219319105, + 0.002768471837043762, + 0.00424604769796133, + 0.0029897401109337807, + 0.00748934131115675, + 0.00379728595726192, + 0.02936127409338951, + 0.007174911443144083, + 0.012371859513223171, + 0.00693407142534852, + 0.00982602871954441, + 0.006528478115797043, + 0.05094720423221588, + 0.007245518732815981, + 0.09983198344707489, + 0.008644824847579002, + 0.013808008283376694, + 0.009827040135860443, + 0.014593561179935932, + 0.009878813289105892, + 0.12840217351913452, + 0.024583814665675163, + 0.00913552287966013, + 0.006353654433041811, + 0.00884505920112133, + 0.012330915778875351, + 0.005818118341267109, + 0.005197392776608467, + 0.011208835989236832, + 0.005047386512160301, + 0.006189519539475441, + 0.0032840720377862453, + 0.008939909748733044, + 0.005275658797472715, + 0.003158503444865346, + 0.0038189301267266273, + 0.0060628075152635574, + 0.007713225204497576, + 0.005371473729610443, + 0.002550712088122964, + 0.003641190007328987, + 0.004916130565106869, + 0.002659701509401202, + 0.0030085083562880754, + 0.0050499956123530865, + 0.003477037651464343, + 0.025611409917473793, + 0.010888333432376385, + 0.021116431802511215, + 0.015342825092375278, + 0.01087107602506876, + 0.012114254757761955, + 0.02249927632510662, + 0.024310361593961716, + 0.012252910062670708, + 0.004714720882475376, + 0.00798773393034935, + 0.006303006783127785, + 0.007460908032953739, + 0.012023698538541794, + 0.013085287064313889, + 0.018451916053891182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 3 + ], + "selected_reward": -0.0032794286962598562 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032401978969574 + ], + "opposite_logprob_tokens": [ + -2.3503239154815674 + ], + "logprob_reward": -0.10032401978969574, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032401978969574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16443653404712677, + 0.08231421560049057, + 0.07157386094331741, + 0.05936996638774872, + 0.1967419683933258, + 0.14421969652175903, + 0.14275334775447845, + 0.043483439832925797, + 0.06497391313314438, + 0.03013300709426403 + ], + "credit_map": [ + 0.03439155966043472, + 0.01193446759134531, + 0.013286258094012737, + 0.014063728973269463, + 0.01359995175153017, + 0.018539398908615112, + 0.048216164112091064, + 0.010404999367892742, + 0.02249003015458584, + 0.003976579289883375, + 0.008475411683321, + 0.00587102398276329, + 0.004336516372859478, + 0.006305274087935686, + 0.023399686440825462, + 0.007459686603397131, + 0.006679885555058718, + 0.005248982459306717, + 0.014401543885469437, + 0.00681076617911458, + 0.008623400703072548, + 0.006775259505957365, + 0.015838805586099625, + 0.007195211946964264, + 0.011226269416511059, + 0.0033499784767627716, + 0.006547922734171152, + 0.00436315406113863, + 0.004617976490408182, + 0.0038693691603839397, + 0.0211245846003294, + 0.004270712845027447, + 0.05651593580842018, + 0.007860196754336357, + 0.012224488891661167, + 0.008297817781567574, + 0.012251610867679119, + 0.007959095761179924, + 0.07670730352401733, + 0.014925517141819, + 0.017296435311436653, + 0.021211130544543266, + 0.018390219658613205, + 0.02132897451519966, + 0.01663304679095745, + 0.014098825864493847, + 0.016596095636487007, + 0.01866496540606022, + 0.023207901045680046, + 0.008359706960618496, + 0.019579846411943436, + 0.032662272453308105, + 0.009879285469651222, + 0.010144286789000034, + 0.023510433733463287, + 0.01540960930287838, + 0.005750893149524927, + 0.0029296954162418842, + 0.007254912983626127, + 0.008907800540328026, + 0.003150865901261568, + 0.0034163224045187235, + 0.005595099180936813, + 0.00647784722968936, + 0.011176597326993942, + 0.00407106289640069, + 0.007582283578813076, + 0.005329804494976997, + 0.005040483083575964, + 0.005769969895482063, + 0.01013120450079441, + 0.01587250828742981, + 0.003709484124556184, + 0.0023631597869098186, + 0.005577767733484507, + 0.0025553361047059298, + 0.0024679037742316723, + 0.0037952151615172625, + 0.004698315169662237, + 0.004965823609381914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.10032401978969574 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059540290385484695 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.0059540290385484695, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059540290385484695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.191740944981575, + 0.12144121527671814, + 0.06279020756483078, + 0.0785381942987442, + 0.31864801049232483, + 0.046194542199373245, + 0.06766722351312637, + 0.030633030459284782, + 0.055999547243118286, + 0.026347069069743156 + ], + "credit_map": [ + 0.03874851390719414, + 0.013360675424337387, + 0.017315585166215897, + 0.021757300943136215, + 0.016508672386407852, + 0.009847494773566723, + 0.06311257183551788, + 0.011090124025940895, + 0.01538858748972416, + 0.012250399217009544, + 0.012105676345527172, + 0.012778174132108688, + 0.010974242351949215, + 0.018072089180350304, + 0.019775893539190292, + 0.020096154883503914, + 0.0054018753580749035, + 0.0037843361496925354, + 0.019037529826164246, + 0.005322952754795551, + 0.006483147386461496, + 0.004591494332998991, + 0.013031089678406715, + 0.0051377820782363415, + 0.017803989350795746, + 0.0033188131637871265, + 0.014232376590371132, + 0.003789359936490655, + 0.005171054974198341, + 0.004437778145074844, + 0.024549424648284912, + 0.005235391203314066, + 0.10285893082618713, + 0.00985308550298214, + 0.012285024859011173, + 0.010366160422563553, + 0.01568685658276081, + 0.008521211333572865, + 0.13555759191513062, + 0.02351916953921318, + 0.005276248324662447, + 0.006828885991126299, + 0.006659638602286577, + 0.007282289210706949, + 0.004160469863563776, + 0.004327194765210152, + 0.006407223641872406, + 0.005252588540315628, + 0.009243948385119438, + 0.006167084444314241, + 0.009416270069777966, + 0.01272053737193346, + 0.00628516124561429, + 0.005814477801322937, + 0.010551434941589832, + 0.00746830552816391, + 0.004596387036144733, + 0.0022266439627856016, + 0.004530815407633781, + 0.0058478121645748615, + 0.0019373864633962512, + 0.0025951138231903315, + 0.00451169116422534, + 0.004387181717902422, + 0.00857587344944477, + 0.004077679477632046, + 0.010109981521964073, + 0.007055507507175207, + 0.005094835069030523, + 0.006372560281306505, + 0.008232565596699715, + 0.006480540614575148, + 0.003122989321127534, + 0.0029649848584085703, + 0.004111341666430235, + 0.0030664480291306973, + 0.0028417326975613832, + 0.0037461339961737394, + 0.002831942168995738, + 0.0036614942364394665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0059540290385484695 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001942058908753097 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.001942058908753097, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001942058908753097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05113934352993965, + 0.03171432390809059, + 0.03714687377214432, + 0.05473298951983452, + 0.6174134612083435, + 0.024973973631858826, + 0.026275843381881714, + 0.03455575555562973, + 0.0555410273373127, + 0.06650634855031967 + ], + "credit_map": [ + 0.00888892263174057, + 0.0045172288082540035, + 0.007754357531666756, + 0.005538435187190771, + 0.005328900646418333, + 0.007060205563902855, + 0.006860034540295601, + 0.005191262345761061, + 0.009289244189858437, + 0.0017910185270011425, + 0.0026445435360074043, + 0.0028797488193958998, + 0.0019781130831688643, + 0.0019193370826542377, + 0.009586517699062824, + 0.0016258021350950003, + 0.005895179696381092, + 0.0026481107342988253, + 0.005423069465905428, + 0.0040067946538329124, + 0.0042114658281207085, + 0.003065103432163596, + 0.0070185489021241665, + 0.004878601990640163, + 0.01409138273447752, + 0.0027680506464093924, + 0.0031352767255157232, + 0.0023051747120916843, + 0.0041193291544914246, + 0.003129997057840228, + 0.02185116708278656, + 0.0033326088450849056, + 0.1939483880996704, + 0.01974911242723465, + 0.03334468603134155, + 0.019900407642126083, + 0.029722532257437706, + 0.01660083420574665, + 0.25848251581192017, + 0.04566498100757599, + 0.0033789139706641436, + 0.002558899810537696, + 0.004915249068289995, + 0.0033965464681386948, + 0.0025242625270038843, + 0.001830488326959312, + 0.004126937128603458, + 0.002242675982415676, + 0.004328092560172081, + 0.0017186980694532394, + 0.002900652354583144, + 0.0057588121853768826, + 0.002220624592155218, + 0.0018778764642775059, + 0.0044012535363435745, + 0.003069833619520068, + 0.005405242554843426, + 0.002185070188716054, + 0.005575803108513355, + 0.003243660321459174, + 0.002471777144819498, + 0.0027885688468813896, + 0.00553321884945035, + 0.007352413143962622, + 0.00924515537917614, + 0.0031368231866508722, + 0.010945025831460953, + 0.004954198841005564, + 0.0040940651670098305, + 0.007997597567737103, + 0.008448943495750427, + 0.0067192139104008675, + 0.00776768010109663, + 0.007002194877713919, + 0.009120306000113487, + 0.004525975324213505, + 0.0067492262460291386, + 0.007461107801645994, + 0.007739073131233454, + 0.016140788793563843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.001942058908753097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277048408985138 + ], + "opposite_logprob_tokens": [ + -2.017770528793335 + ], + "logprob_reward": -0.14277048408985138, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14277048408985138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.229139044880867, + 0.062410205602645874, + 0.04310848191380501, + 0.10879115760326385, + 0.17227371037006378, + 0.044522952288389206, + 0.10700083523988724, + 0.03176749125123024, + 0.11530385166406631, + 0.085682213306427 + ], + "credit_map": [ + 0.042064011096954346, + 0.010927142575383186, + 0.018491491675376892, + 0.028365768492221832, + 0.015389395877718925, + 0.010847640223801136, + 0.09157641232013702, + 0.011477184481918812, + 0.01407287735491991, + 0.003619717201218009, + 0.007563048042356968, + 0.005274470429867506, + 0.004673120100051165, + 0.005053615663200617, + 0.016315199434757233, + 0.005838159937411547, + 0.007245458196848631, + 0.0035706369671970606, + 0.007136162370443344, + 0.0045452238991856575, + 0.0033164224587380886, + 0.006053629331290722, + 0.005779219325631857, + 0.005461729131639004, + 0.02054588310420513, + 0.00680627441033721, + 0.014706168323755264, + 0.007251049391925335, + 0.00751396594569087, + 0.005889988970011473, + 0.0404658280313015, + 0.005612000823020935, + 0.05299181863665581, + 0.005727518815547228, + 0.014897745102643967, + 0.005115443840622902, + 0.007816057652235031, + 0.005221550352871418, + 0.06770065426826477, + 0.012802924029529095, + 0.0068212044425308704, + 0.0039156656712293625, + 0.007922867313027382, + 0.006019724067300558, + 0.0037874686531722546, + 0.0037016698624938726, + 0.006919228006154299, + 0.005435124970972538, + 0.018771013244986534, + 0.007920398376882076, + 0.01635546050965786, + 0.019551541656255722, + 0.006508145947009325, + 0.00784690398722887, + 0.016583915799856186, + 0.0134634580463171, + 0.004538760054856539, + 0.0026027837302535772, + 0.00384901138022542, + 0.004628101829439402, + 0.002090799156576395, + 0.00291144079528749, + 0.005694165825843811, + 0.005452428944408894, + 0.018743250519037247, + 0.00696669053286314, + 0.020074384286999702, + 0.013217800296843052, + 0.010496428236365318, + 0.013866724446415901, + 0.017510706558823586, + 0.014427871443331242, + 0.009542317129671574, + 0.006030064541846514, + 0.012862339615821838, + 0.007189578376710415, + 0.00849977321922779, + 0.010930375196039677, + 0.012065590359270573, + 0.01856217533349991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.14277048408985138 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera slowly and steadily zooms in, transporting us to an oriental ink painting imbued with the spirit of freehand brushwork. Beginning in a mid-ground, the scene gradually focuses on a fully-winged angel, at the apex of a mighty sword swing. The angel's form is slender and graceful, its massive wings rendered with bold, sweeping strokes. Each feather is outlined in varying shades of ink, their edges blending hazily into the abstract, light-and-dark background of abstracted space. The angel is clad in a flowing white robe, its folds traced with simple yet powerful calligraphic lines, fluttering with his movements. His focused, resolute expression radiates immense power. He clutches a simple, ancient sword in both hands, his body twisting into a tense bow. The blade slices through the air, leaving a powerful, sharp arc composed of both feibai and splashed ink techniques. The blade itself gleams with a sharp, cold gleam, punctuated by the use of white space. The entire scene is based on the subtle texture of rice paper, and the main colors are classic black, white and gray. The light does not come from an external light source, but is completely shaped by the strong contrast between large areas of white space and thick ink, creating a solemn, grand and dynamic atmosphere, as if an eternal battle moment is frozen on the scroll.", + "seed": 19897, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_7/index_4178_seed_19897_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02272109971381724, + "vlm_logprob": -0.02272109971381724, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.02272109971381724, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028981848154217005 + ], + "opposite_logprob_tokens": [ + -6.752898216247559 + ], + "logprob_reward": -0.0028981848154217005, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0028981848154217005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16608430445194244, + 0.10507006198167801, + 0.06110461428761482, + 0.13704480230808258, + 0.07273664325475693, + 0.16494305431842804, + 0.08701327443122864, + 0.023181255906820297, + 0.08778930455446243, + 0.09503265470266342 + ], + "credit_map": [ + 0.026879506185650826, + 0.014109797775745392, + 0.0286538228392601, + 0.011069958098232746, + 0.011420540511608124, + 0.016438117250800133, + 0.015511113218963146, + 0.04200145602226257, + 0.018240220844745636, + 0.008945141918957233, + 0.016227902844548225, + 0.007284601218998432, + 0.007806858979165554, + 0.010458242148160934, + 0.008601452223956585, + 0.02750563435256481, + 0.011341877281665802, + 0.004202232230454683, + 0.013250479474663734, + 0.004599628038704395, + 0.0025515547022223473, + 0.003635124070569873, + 0.005680497735738754, + 0.0158432275056839, + 0.021401504054665565, + 0.011918298900127411, + 0.029436448588967323, + 0.013326689600944519, + 0.009930947795510292, + 0.015604817308485508, + 0.012137744575738907, + 0.023288358002901077, + 0.01473433431237936, + 0.004764608573168516, + 0.024918779730796814, + 0.005775819532573223, + 0.003966034855693579, + 0.006666542496532202, + 0.004257661290466785, + 0.007652861997485161, + 0.03234298899769783, + 0.014254525303840637, + 0.031567465513944626, + 0.017378544434905052, + 0.01306536141782999, + 0.011865762993693352, + 0.012155629694461823, + 0.032312776893377304, + 0.012325224466621876, + 0.005410134792327881, + 0.014363698661327362, + 0.011508324183523655, + 0.005081973969936371, + 0.010006706230342388, + 0.007773656863719225, + 0.020543554797768593, + 0.004410623572766781, + 0.002099723555147648, + 0.003621137933805585, + 0.002674689283594489, + 0.0013449657708406448, + 0.0019297536928206682, + 0.002069347072392702, + 0.005031014326959848, + 0.012667777948081493, + 0.006367544177919626, + 0.019294708967208862, + 0.00776904821395874, + 0.006513574626296759, + 0.016804136335849762, + 0.006897293031215668, + 0.011475227773189545, + 0.019885307177901268, + 0.00941796600818634, + 0.017060566693544388, + 0.008165557868778706, + 0.006774440407752991, + 0.011698557995259762, + 0.007830494083464146, + 0.014199766330420971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.0028981848154217005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430745892226696 + ], + "opposite_logprob_tokens": [ + -3.168074607849121 + ], + "logprob_reward": -0.0430745892226696, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430745892226696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21376536786556244, + 0.11399371176958084, + 0.15524087846279144, + 0.07506878674030304, + 0.05455045774579048, + 0.07905947417020798, + 0.08439867198467255, + 0.04653925448656082, + 0.07534537464380264, + 0.10203805565834045 + ], + "credit_map": [ + 0.0310627780854702, + 0.028972821310162544, + 0.03443178907036781, + 0.018262874335050583, + 0.02046108804643154, + 0.023684991523623466, + 0.02320314198732376, + 0.033685892820358276, + 0.018007127568125725, + 0.018936892971396446, + 0.01931619457900524, + 0.013435931876301765, + 0.009088252671062946, + 0.009460754692554474, + 0.011049349792301655, + 0.014699219726026058, + 0.027515610679984093, + 0.017276179045438766, + 0.028341596946120262, + 0.013772428035736084, + 0.011348199099302292, + 0.010725212283432484, + 0.012630748562514782, + 0.03363090008497238, + 0.012574655935168266, + 0.004778691567480564, + 0.019008858129382133, + 0.009125911630690098, + 0.005325514357537031, + 0.009301049634814262, + 0.005669381469488144, + 0.009284723550081253, + 0.009397722780704498, + 0.005961372517049313, + 0.014500080607831478, + 0.005564866121858358, + 0.0034160923678427935, + 0.004543640650808811, + 0.003645262448117137, + 0.007521420251578093, + 0.015417475253343582, + 0.007337482646107674, + 0.014647473581135273, + 0.0058771646581590176, + 0.007810327224433422, + 0.008772581815719604, + 0.005074407439678907, + 0.014122562482953072, + 0.014274575747549534, + 0.006501454394310713, + 0.010990859940648079, + 0.006663394160568714, + 0.008809269405901432, + 0.014471685513854027, + 0.007260688114911318, + 0.015426741912961006, + 0.007103822194039822, + 0.004335650242865086, + 0.007086918689310551, + 0.0044244434684515, + 0.0035759895108640194, + 0.0042043412104249, + 0.004577143583446741, + 0.011230943724513054, + 0.010877146385610104, + 0.007496441248804331, + 0.015019548125565052, + 0.00563799636438489, + 0.005797276739031076, + 0.013194173574447632, + 0.005174093879759312, + 0.012148695066571236, + 0.014956382103264332, + 0.00941765122115612, + 0.015282880514860153, + 0.011305541731417179, + 0.008981580846011639, + 0.012547267600893974, + 0.011773914098739624, + 0.017772836610674858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0430745892226696 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099181387573481 + ], + "opposite_logprob_tokens": [ + -5.504099369049072 + ], + "logprob_reward": -0.004099181387573481, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099181387573481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24935071170330048, + 0.10532521456480026, + 0.13426262140274048, + 0.09182929992675781, + 0.14060167968273163, + 0.06789036095142365, + 0.04786873236298561, + 0.030734725296497345, + 0.06783708930015564, + 0.06429953128099442 + ], + "credit_map": [ + 0.040333956480026245, + 0.028544249013066292, + 0.04967889189720154, + 0.02112281136214733, + 0.02409336157143116, + 0.031170807778835297, + 0.01958482898771763, + 0.03482178971171379, + 0.01587190106511116, + 0.017581770196557045, + 0.015564140863716602, + 0.011957007460296154, + 0.009652574546635151, + 0.009070794098079205, + 0.01154081616550684, + 0.014086216688156128, + 0.02295728027820587, + 0.015172596089541912, + 0.022383008152246475, + 0.011017566546797752, + 0.006699558347463608, + 0.009508789516985416, + 0.01463255099952221, + 0.03189128637313843, + 0.017498871311545372, + 0.007890049368143082, + 0.019666507840156555, + 0.009356033056974411, + 0.008807600475847721, + 0.005345871206372976, + 0.012264501303434372, + 0.01099986582994461, + 0.02689552493393421, + 0.009778267703950405, + 0.040441613644361496, + 0.014310588128864765, + 0.008124446496367455, + 0.010855034925043583, + 0.010471089743077755, + 0.01972511224448681, + 0.009998137131333351, + 0.006598905194550753, + 0.009300806559622288, + 0.008015716448426247, + 0.007714713923633099, + 0.008520268835127354, + 0.008529070764780045, + 0.009212748147547245, + 0.007908080704510212, + 0.0030998943839222193, + 0.007929195649921894, + 0.005260970443487167, + 0.0035470090806484222, + 0.0039327931590378284, + 0.004321434535086155, + 0.011869355104863644, + 0.004452653229236603, + 0.002710509579628706, + 0.004486636724323034, + 0.004470532294362783, + 0.0031355605460703373, + 0.0028134144376963377, + 0.003588390536606312, + 0.005077028647065163, + 0.013298355042934418, + 0.004626183770596981, + 0.009791486896574497, + 0.009178156033158302, + 0.004581819288432598, + 0.006475645583122969, + 0.0074757738038897514, + 0.01240966934710741, + 0.01282308716326952, + 0.006977815646678209, + 0.011570942588150501, + 0.005045009311288595, + 0.004960377234965563, + 0.008053654804825783, + 0.0059640491381287575, + 0.008904599584639072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.004099181387573481 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001512932125478983 + ], + "opposite_logprob_tokens": [ + -6.5015130043029785 + ], + "logprob_reward": -0.001512932125478983, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001512932125478983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2038997858762741, + 0.11069213598966599, + 0.054378949105739594, + 0.08020076155662537, + 0.06870187819004059, + 0.22118248045444489, + 0.03835725039243698, + 0.04595230892300606, + 0.059313543140888214, + 0.11732084304094315 + ], + "credit_map": [ + 0.031798917800188065, + 0.02155858464539051, + 0.036810170859098434, + 0.016136180609464645, + 0.014310254715383053, + 0.025669053196907043, + 0.020814206451177597, + 0.03680243715643883, + 0.020000165328383446, + 0.011941041797399521, + 0.013376490212976933, + 0.010948432609438896, + 0.008188292384147644, + 0.009578410536050797, + 0.01202411949634552, + 0.024635199457406998, + 0.009427779354155064, + 0.005872088484466076, + 0.010037084110081196, + 0.005361844785511494, + 0.0043664053082466125, + 0.004012777470052242, + 0.005449378862977028, + 0.009851599112153053, + 0.013288154266774654, + 0.007352891378104687, + 0.01720293238759041, + 0.00894896313548088, + 0.007049737498164177, + 0.006593476049602032, + 0.009641293436288834, + 0.010123318992555141, + 0.013277750462293625, + 0.003469164716079831, + 0.022334523499011993, + 0.00719834491610527, + 0.0033883738797158003, + 0.005332235246896744, + 0.004513851832598448, + 0.009187642484903336, + 0.048989538103342056, + 0.016646867617964745, + 0.05608902499079704, + 0.026655826717615128, + 0.010153144598007202, + 0.01530974917113781, + 0.01921786367893219, + 0.02812049351632595, + 0.007118921726942062, + 0.003333346452564001, + 0.005345167592167854, + 0.005719678942114115, + 0.002869626274332404, + 0.0028801115695387125, + 0.004659101366996765, + 0.006431298796087503, + 0.007713423576205969, + 0.0046616955660283566, + 0.00653769401833415, + 0.006858568172901869, + 0.0036484531592577696, + 0.003305248450487852, + 0.004087231587618589, + 0.009139997884631157, + 0.011148808524012566, + 0.004482918884605169, + 0.009966875426471233, + 0.00670793978497386, + 0.003768038935959339, + 0.0066599659621715546, + 0.005918606650084257, + 0.010660397820174694, + 0.020073704421520233, + 0.013580979779362679, + 0.020305901765823364, + 0.010645335540175438, + 0.009817744605243206, + 0.016300790011882782, + 0.011765982024371624, + 0.014830422587692738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.001512932125478983 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202061101794243 + ], + "opposite_logprob_tokens": [ + -2.8120205402374268 + ], + "logprob_reward": -0.06202061101794243, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202061101794243, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0834977999329567, + 0.049641434103250504, + 0.07447408884763718, + 0.0563623271882534, + 0.09048425406217575, + 0.2233656495809555, + 0.056790176779031754, + 0.030015723779797554, + 0.14093995094299316, + 0.19442857801914215 + ], + "credit_map": [ + 0.0129745127633214, + 0.012052967213094234, + 0.014401579275727272, + 0.009444505907595158, + 0.006542891729623079, + 0.007615535520017147, + 0.009183231741189957, + 0.011282571591436863, + 0.007402381394058466, + 0.006379434373229742, + 0.006973616313189268, + 0.005477225407958031, + 0.004141778685152531, + 0.004819545429199934, + 0.00576274236664176, + 0.008684707805514336, + 0.01305978000164032, + 0.0068291425704956055, + 0.012038741260766983, + 0.005746893584728241, + 0.004172810819000006, + 0.005956663284450769, + 0.007293990347534418, + 0.019376063719391823, + 0.008453022688627243, + 0.005109624937176704, + 0.014877242036163807, + 0.005615534260869026, + 0.0040900008752942085, + 0.005587192717939615, + 0.005183160770684481, + 0.007446549367159605, + 0.014777353033423424, + 0.005247018299996853, + 0.025163454934954643, + 0.008016795851290226, + 0.0057161590084433556, + 0.007485182955861092, + 0.00662854453548789, + 0.017449747771024704, + 0.05182372406125069, + 0.019582215696573257, + 0.05962404981255531, + 0.021064553409814835, + 0.012107151560485363, + 0.016985338181257248, + 0.016443388536572456, + 0.02573522739112377, + 0.012270738370716572, + 0.00481311185285449, + 0.008875296451151371, + 0.006677825003862381, + 0.003915864042937756, + 0.005471804179251194, + 0.004185405094176531, + 0.010580130852758884, + 0.0045361602678895, + 0.0035898182541131973, + 0.004024076275527477, + 0.0033733933232724667, + 0.003272398840636015, + 0.0031771205831319094, + 0.0031104562804102898, + 0.004932298790663481, + 0.025444552302360535, + 0.012605085037648678, + 0.02163698710501194, + 0.01363389939069748, + 0.011194007471203804, + 0.021929413080215454, + 0.013519838452339172, + 0.02097615785896778, + 0.030236756429076195, + 0.02352427877485752, + 0.02702607586979866, + 0.018549172207713127, + 0.016613919287919998, + 0.027863679453730583, + 0.02305692248046398, + 0.02755778469145298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.06202061101794243 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera slowly and steadily zooms in, transporting us to an oriental ink painting imbued with the spirit of freehand brushwork. Beginning in a mid-ground, the scene gradually focuses on a fully-winged angel, at the apex of a mighty sword swing. The angel's form is slender and graceful, its massive wings rendered with bold, sweeping strokes. Each feather is outlined in varying shades of ink, their edges blending hazily into the abstract, light-and-dark background of abstracted space. The angel is clad in a flowing white robe, its folds traced with simple yet powerful calligraphic lines, fluttering with his movements. His focused, resolute expression radiates immense power. He clutches a simple, ancient sword in both hands, his body twisting into a tense bow. The blade slices through the air, leaving a powerful, sharp arc composed of both feibai and splashed ink techniques. The blade itself gleams with a sharp, cold gleam, punctuated by the use of white space. The entire scene is based on the subtle texture of rice paper, and the main colors are classic black, white and gray. The light does not come from an external light source, but is completely shaped by the strong contrast between large areas of white space and thick ink, creating a solemn, grand and dynamic atmosphere, as if an eternal battle moment is frozen on the scroll.", + "seed": 19898, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_7/index_4178_seed_19898_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04388399999588728, + "vlm_logprob": -0.04388399999588728, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.04388399999588728, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008678695186972618 + ], + "opposite_logprob_tokens": [ + -4.883678913116455 + ], + "logprob_reward": -0.008678695186972618, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008678695186972618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15976767241954803, + 0.08268695324659348, + 0.19909371435642242, + 0.08946439623832703, + 0.06998424977064133, + 0.039549026638269424, + 0.09840665012598038, + 0.10052558779716492, + 0.09914146363735199, + 0.06138034909963608 + ], + "credit_map": [ + 0.02289985492825508, + 0.03120872564613819, + 0.020434334874153137, + 0.016299927607178688, + 0.0206407830119133, + 0.013622470200061798, + 0.01981475204229355, + 0.01484681572765112, + 0.008069567382335663, + 0.009609431959688663, + 0.010155003517866135, + 0.008140071295201778, + 0.012989355251193047, + 0.007395188324153423, + 0.011456742882728577, + 0.014871593564748764, + 0.041205357760190964, + 0.01936034858226776, + 0.020778829231858253, + 0.011092160828411579, + 0.014334626495838165, + 0.013111261650919914, + 0.015395398251712322, + 0.06381572037935257, + 0.023135727271437645, + 0.008502195589244366, + 0.007904769852757454, + 0.00845322199165821, + 0.00841639842838049, + 0.00732989190146327, + 0.009425664320588112, + 0.01629652827978134, + 0.006761656142771244, + 0.00828793365508318, + 0.006791313178837299, + 0.012260359711945057, + 0.011637978255748749, + 0.007001315243542194, + 0.01067346055060625, + 0.006570233032107353, + 0.008292913436889648, + 0.0039490992203354836, + 0.005733247380703688, + 0.004205933306366205, + 0.003828093409538269, + 0.003193766577169299, + 0.006162742618471384, + 0.0041832332499325275, + 0.01619354821741581, + 0.00817224383354187, + 0.018307624384760857, + 0.010284808464348316, + 0.010587162338197231, + 0.010416864417493343, + 0.0134585527703166, + 0.010985839180648327, + 0.011987917125225067, + 0.013140389695763588, + 0.013719107024371624, + 0.013814818114042282, + 0.009876919910311699, + 0.01841059699654579, + 0.010938551276922226, + 0.008637290447950363, + 0.007674250286072493, + 0.008831657469272614, + 0.024846266955137253, + 0.010713133029639721, + 0.009325433522462845, + 0.017177218571305275, + 0.008694441989064217, + 0.011879055760800838, + 0.00728224590420723, + 0.006400736514478922, + 0.008212057873606682, + 0.007359822280704975, + 0.006862703710794449, + 0.011502780020236969, + 0.008332572877407074, + 0.00542742945253849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.008678695186972618 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033393651247025 + ], + "opposite_logprob_tokens": [ + -2.3503339290618896 + ], + "logprob_reward": -0.10033393651247025, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033393651247025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2750113010406494, + 0.04984079301357269, + 0.12849710881710052, + 0.09331612288951874, + 0.028295274823904037, + 0.04959815740585327, + 0.0588800273835659, + 0.053321268409490585, + 0.20645904541015625, + 0.05678096041083336 + ], + "credit_map": [ + 0.0700758844614029, + 0.02443542703986168, + 0.016147661954164505, + 0.021884234622120857, + 0.03608536347746849, + 0.022130219265818596, + 0.028071997687220573, + 0.056180521845817566, + 0.005575072485953569, + 0.008163467049598694, + 0.006545558571815491, + 0.004322686232626438, + 0.0061274124309420586, + 0.0037525335792452097, + 0.006483686622232199, + 0.008870377205312252, + 0.027520030736923218, + 0.009776095859706402, + 0.012775745242834091, + 0.008636347949504852, + 0.009233859367668629, + 0.007600835524499416, + 0.011291851289570332, + 0.04166233167052269, + 0.0191267728805542, + 0.006801857613027096, + 0.013364247977733612, + 0.005766279064118862, + 0.006409203168004751, + 0.0049987491220235825, + 0.00645813625305891, + 0.030390875414013863, + 0.0026475877966731787, + 0.0026598814874887466, + 0.002400551224127412, + 0.004775587469339371, + 0.005594030953943729, + 0.003962302580475807, + 0.003965431358665228, + 0.002289901254698634, + 0.005641448311507702, + 0.0057825190015137196, + 0.006868848577141762, + 0.0072152577340602875, + 0.006448117084801197, + 0.0047175344079732895, + 0.007438818458467722, + 0.0054856110364198685, + 0.007939552888274193, + 0.0068588051944971085, + 0.010068259201943874, + 0.006904334295541048, + 0.007327537052333355, + 0.007980331778526306, + 0.006316808518022299, + 0.005484404042363167, + 0.006472085602581501, + 0.005264097359031439, + 0.006816832348704338, + 0.007344220299273729, + 0.00574804050847888, + 0.01072691660374403, + 0.0065053654834628105, + 0.004443712066859007, + 0.016931956633925438, + 0.02048945426940918, + 0.04863481596112251, + 0.021976258605718613, + 0.021108942106366158, + 0.03525286912918091, + 0.023633036762475967, + 0.018431706354022026, + 0.005525735206902027, + 0.0061177355237305164, + 0.011710633523762226, + 0.007072205655276775, + 0.007218608632683754, + 0.008593269623816013, + 0.005876364652067423, + 0.004666401073336601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.10033393651247025 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076371124014258385 + ], + "opposite_logprob_tokens": [ + -4.882637023925781 + ], + "logprob_reward": -0.0076371124014258385, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076371124014258385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30143243074417114, + 0.10969220846891403, + 0.08684772998094559, + 0.09293750673532486, + 0.030257537961006165, + 0.032401371747255325, + 0.1357703059911728, + 0.049333810806274414, + 0.07513061910867691, + 0.0861964076757431 + ], + "credit_map": [ + 0.08436404913663864, + 0.03341355547308922, + 0.013242226094007492, + 0.020330386236310005, + 0.03815094754099846, + 0.018673719838261604, + 0.02059776335954666, + 0.07265980541706085, + 0.014262601733207703, + 0.015593981370329857, + 0.017127977684140205, + 0.01222419273108244, + 0.012067807838320732, + 0.009285005740821362, + 0.011517084203660488, + 0.017613552510738373, + 0.017160017043352127, + 0.0055618793703615665, + 0.0106308339163661, + 0.004903286229819059, + 0.005511739756911993, + 0.004869271535426378, + 0.006745968014001846, + 0.03146474063396454, + 0.031805314123630524, + 0.006596669089049101, + 0.004982039798051119, + 0.005568626802414656, + 0.005370134022086859, + 0.004875535611063242, + 0.004814925603568554, + 0.028924262151122093, + 0.004242556169629097, + 0.003918510861694813, + 0.0032527672592550516, + 0.0031617851927876472, + 0.004204933065921068, + 0.0034281443804502487, + 0.004262909758836031, + 0.003785931970924139, + 0.005868886597454548, + 0.003004658967256546, + 0.005343271419405937, + 0.003679391462355852, + 0.0032943671103566885, + 0.002938778605312109, + 0.003948185592889786, + 0.004323828965425491, + 0.017791302874684334, + 0.013709958642721176, + 0.0248007420450449, + 0.014282256364822388, + 0.015496369451284409, + 0.017096376046538353, + 0.014835190959274769, + 0.017758097499608994, + 0.005120700690895319, + 0.0062955450266599655, + 0.007271416950970888, + 0.0069644274190068245, + 0.005666026845574379, + 0.008972210809588432, + 0.005167401395738125, + 0.003876078873872757, + 0.006066954229027033, + 0.0066049410961568356, + 0.01786080189049244, + 0.007687760517001152, + 0.007050883490592241, + 0.013253564015030861, + 0.00976548157632351, + 0.0068402281031012535, + 0.008476595394313335, + 0.01179494522511959, + 0.012019295245409012, + 0.010025423020124435, + 0.009181851521134377, + 0.01877041906118393, + 0.009549831971526146, + 0.00637804763391614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0076371124014258385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024909917265176773 + ], + "opposite_logprob_tokens": [ + -6.002490997314453 + ], + "logprob_reward": -0.0024909917265176773, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024909917265176773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25982674956321716, + 0.07820110023021698, + 0.10638809204101562, + 0.14279483258724213, + 0.04016225039958954, + 0.05360642820596695, + 0.06477686762809753, + 0.08073177933692932, + 0.07953231036663055, + 0.09397962689399719 + ], + "credit_map": [ + 0.06296005845069885, + 0.03784187138080597, + 0.015233646146953106, + 0.01854041777551174, + 0.03337534889578819, + 0.013812167569994926, + 0.020725874230265617, + 0.05733734741806984, + 0.010242351330816746, + 0.011210745200514793, + 0.009599114768207073, + 0.008799196220934391, + 0.010752537287771702, + 0.005957759916782379, + 0.010233779437839985, + 0.011405616998672485, + 0.014106573536992073, + 0.012353169731795788, + 0.014750228263437748, + 0.00841834768652916, + 0.009104335680603981, + 0.007508929818868637, + 0.00898776389658451, + 0.03115873597562313, + 0.04567403346300125, + 0.007536592427641153, + 0.00990174151957035, + 0.00632341718301177, + 0.008293112739920616, + 0.007305474486202002, + 0.007142168469727039, + 0.05061827972531319, + 0.005325196776539087, + 0.005370901431888342, + 0.0049613588489592075, + 0.007008427754044533, + 0.004982606507837772, + 0.0033622777555137873, + 0.005472921300679445, + 0.0036785616539418697, + 0.009190871380269527, + 0.00665487302467227, + 0.008990363217890263, + 0.006112373899668455, + 0.005677572451531887, + 0.005136117339134216, + 0.005959140136837959, + 0.005885120015591383, + 0.007611572276800871, + 0.0072310869581997395, + 0.012696539051830769, + 0.008281407877802849, + 0.006722118705511093, + 0.008327126502990723, + 0.007260535378009081, + 0.006646482273936272, + 0.010106127709150314, + 0.010189607739448547, + 0.010826502926647663, + 0.00985618494451046, + 0.007828012108802795, + 0.0137802017852664, + 0.009190019220113754, + 0.008955125696957111, + 0.009042208082973957, + 0.007455420680344105, + 0.015138518996536732, + 0.007249207701534033, + 0.009335590526461601, + 0.014355052262544632, + 0.009870757348835468, + 0.007085555698722601, + 0.011088504455983639, + 0.011087572202086449, + 0.012798183597624302, + 0.012197745032608509, + 0.008600576780736446, + 0.022240685299038887, + 0.010015488602221012, + 0.005950873717665672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0024909917265176773 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10027926415205002 + ], + "opposite_logprob_tokens": [ + -2.3502793312072754 + ], + "logprob_reward": -0.10027926415205002, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10027926415205002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14411909878253937, + 0.09154273569583893, + 0.17444831132888794, + 0.10025670379400253, + 0.0611475333571434, + 0.021404599770903587, + 0.052881475538015366, + 0.09645022451877594, + 0.09035754203796387, + 0.1673918515443802 + ], + "credit_map": [ + 0.01409331988543272, + 0.019718782976269722, + 0.012210335582494736, + 0.02066868543624878, + 0.025055890902876854, + 0.013704949989914894, + 0.027126995846629143, + 0.011540131643414497, + 0.012915678322315216, + 0.013728401623666286, + 0.007764946669340134, + 0.012401843443512917, + 0.014262489974498749, + 0.0089617520570755, + 0.011556075885891914, + 0.009951547719538212, + 0.03756462410092354, + 0.011770961806178093, + 0.020065298303961754, + 0.010237899608910084, + 0.012571639381349087, + 0.01033700630068779, + 0.01287857536226511, + 0.05902230367064476, + 0.023847708478569984, + 0.00710649648681283, + 0.01094969641417265, + 0.00754745164886117, + 0.007262127939611673, + 0.004953343421220779, + 0.0056939744390547276, + 0.03289590775966644, + 0.010417262092232704, + 0.008176565170288086, + 0.00602960167452693, + 0.008236553519964218, + 0.008742636069655418, + 0.005185998976230621, + 0.007078772876411676, + 0.007280141115188599, + 0.0023879685904830694, + 0.00257366755977273, + 0.002882248256355524, + 0.003997518680989742, + 0.002607528353109956, + 0.002218098845332861, + 0.0029371054843068123, + 0.0018004622543230653, + 0.007311946712434292, + 0.005831074435263872, + 0.009926874190568924, + 0.00606969790533185, + 0.005568333901464939, + 0.006329353433102369, + 0.005945409182459116, + 0.005898786708712578, + 0.015179452486336231, + 0.01093991007655859, + 0.010711684823036194, + 0.012059424072504044, + 0.011700604110956192, + 0.013992365449666977, + 0.012075865641236305, + 0.009790921583771706, + 0.008202310651540756, + 0.008323369547724724, + 0.025197241455316544, + 0.00965170282870531, + 0.007549128495156765, + 0.017083123326301575, + 0.008137903176248074, + 0.006212760228663683, + 0.016689443960785866, + 0.020157380029559135, + 0.030297039076685905, + 0.020027311518788338, + 0.015875687822699547, + 0.0391298271715641, + 0.015758605673909187, + 0.009456555359065533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.10027926415205002 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera slowly and steadily zooms in, transporting us to an oriental ink painting imbued with the spirit of freehand brushwork. Beginning in a mid-ground, the scene gradually focuses on a fully-winged angel, at the apex of a mighty sword swing. The angel's form is slender and graceful, its massive wings rendered with bold, sweeping strokes. Each feather is outlined in varying shades of ink, their edges blending hazily into the abstract, light-and-dark background of abstracted space. The angel is clad in a flowing white robe, its folds traced with simple yet powerful calligraphic lines, fluttering with his movements. His focused, resolute expression radiates immense power. He clutches a simple, ancient sword in both hands, his body twisting into a tense bow. The blade slices through the air, leaving a powerful, sharp arc composed of both feibai and splashed ink techniques. The blade itself gleams with a sharp, cold gleam, punctuated by the use of white space. The entire scene is based on the subtle texture of rice paper, and the main colors are classic black, white and gray. The light does not come from an external light source, but is completely shaped by the strong contrast between large areas of white space and thick ink, creating a solemn, grand and dynamic atmosphere, as if an eternal battle moment is frozen on the scroll.", + "seed": 19899, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_7/index_4178_seed_19899_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.041061178385280075, + "vlm_logprob": -0.041061178385280075, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.041061178385280075, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006870810873806477 + ], + "opposite_logprob_tokens": [ + -5.131870746612549 + ], + "logprob_reward": -0.006870810873806477, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.006870810873806477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09094344824552536, + 0.08185455203056335, + 0.059074703603982925, + 0.05140465870499611, + 0.5006991028785706, + 0.07740146666765213, + 0.07120872288942337, + 0.026495128870010376, + 0.025467032566666603, + 0.015451156534254551 + ], + "credit_map": [ + 0.01326711568981409, + 0.00782271008938551, + 0.015269028954207897, + 0.010325276292860508, + 0.007729800883680582, + 0.009676539339125156, + 0.01156715489923954, + 0.01528581790626049, + 0.0120319165289402, + 0.006400285754352808, + 0.011062795296311378, + 0.01116968784481287, + 0.007107989396899939, + 0.007391733583062887, + 0.008056226186454296, + 0.018633918836712837, + 0.00910479761660099, + 0.0030444683507084846, + 0.010102775879204273, + 0.007077325601130724, + 0.00309211784042418, + 0.004684274084866047, + 0.008241048082709312, + 0.01372789777815342, + 0.007796724326908588, + 0.0034315416123718023, + 0.011086752638220787, + 0.007910579442977905, + 0.002762577263638377, + 0.003801692044362426, + 0.0053815897554159164, + 0.009233200922608376, + 0.1772303730249405, + 0.026282038539648056, + 0.02365328185260296, + 0.01369745098054409, + 0.01429383922368288, + 0.046819012612104416, + 0.017444105818867683, + 0.18127897381782532, + 0.013544694520533085, + 0.006656554527580738, + 0.011811841279268265, + 0.0058074528351426125, + 0.003516370663419366, + 0.007739863358438015, + 0.006811336148530245, + 0.021513361483812332, + 0.009959331713616848, + 0.004025732167065144, + 0.015485928393900394, + 0.007814677432179451, + 0.003141355700790882, + 0.006498297676444054, + 0.01115945540368557, + 0.013123942539095879, + 0.0032405804377049208, + 0.0026853198651224375, + 0.003938630688935518, + 0.002634903183206916, + 0.0020885958801954985, + 0.002421486657112837, + 0.002834164770320058, + 0.006651447620242834, + 0.003179281484335661, + 0.003162410343065858, + 0.003048723330721259, + 0.0028510242700576782, + 0.0020998534746468067, + 0.0024325884878635406, + 0.0029745122883468866, + 0.005718638189136982, + 0.0016807098872959614, + 0.002136184833943844, + 0.002142498269677162, + 0.0016767570050433278, + 0.0010643650311976671, + 0.0015706440899521112, + 0.0012514828704297543, + 0.003928516060113907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.006870810873806477 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206789240241051 + ], + "opposite_logprob_tokens": [ + -2.812067985534668 + ], + "logprob_reward": -0.06206789240241051, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206789240241051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06075136363506317, + 0.01640826277434826, + 0.023453207686543465, + 0.025009827688336372, + 0.7952882647514343, + 0.02079349383711815, + 0.014920251443982124, + 0.016917560249567032, + 0.011890939436852932, + 0.014566781930625439 + ], + "credit_map": [ + 0.008826485835015774, + 0.00521467998623848, + 0.012590520083904266, + 0.00589978601783514, + 0.004934503696858883, + 0.004922785330563784, + 0.006017982494086027, + 0.012344625778496265, + 0.002339442726224661, + 0.0012792687630280852, + 0.0023529566824436188, + 0.002005442278459668, + 0.0012565740616992116, + 0.0016495073214173317, + 0.001908053644001484, + 0.0036170175299048424, + 0.0031758311670273542, + 0.001728268340229988, + 0.0059330458752810955, + 0.0024508764035999775, + 0.0017805063398554921, + 0.0017105605220422149, + 0.002225511008873582, + 0.00444860989227891, + 0.004592457786202431, + 0.0014358002226799726, + 0.002749537117779255, + 0.0036577731370925903, + 0.001849374850280583, + 0.0012663525994867086, + 0.005753094796091318, + 0.0037054417189210653, + 0.2804836332798004, + 0.042472824454307556, + 0.036147456616163254, + 0.021119946613907814, + 0.020779158920049667, + 0.07590850442647934, + 0.027935605496168137, + 0.29044124484062195, + 0.003123403526842594, + 0.0013700638664886355, + 0.005288915242999792, + 0.00225981418043375, + 0.0010357338469475508, + 0.0020524344872683287, + 0.001934362342581153, + 0.00372876669280231, + 0.0029927303548902273, + 0.0010475985473021865, + 0.002476175781339407, + 0.0010754737304523587, + 0.000665194122120738, + 0.0010640051914379, + 0.0017513127531856298, + 0.0038477610796689987, + 0.0018548687221482396, + 0.0021610455587506294, + 0.0024220068007707596, + 0.0011626899940893054, + 0.0014483702834695578, + 0.0017427962739020586, + 0.0015332141192629933, + 0.004592569544911385, + 0.0012624217197299004, + 0.0015901169972494245, + 0.0012199004413560033, + 0.00199498631991446, + 0.00119362014811486, + 0.001152715878561139, + 0.0017417740309610963, + 0.0017354049487039447, + 0.0022859799209982157, + 0.0023718317970633507, + 0.0014184300089254975, + 0.0013327813940122724, + 0.0012386098969727755, + 0.00246215658262372, + 0.001137327984906733, + 0.002319667488336563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.06206789240241051 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631670217961073 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631670217961073, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631670217961073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06628033518791199, + 0.0170126985758543, + 0.01300500612705946, + 0.0066978647373616695, + 0.8636934161186218, + 0.00874189380556345, + 0.007315392605960369, + 0.0068355584517121315, + 0.007317334413528442, + 0.003100577974691987 + ], + "credit_map": [ + 0.009650256484746933, + 0.0053803035989403725, + 0.015325067564845085, + 0.005610454827547073, + 0.00570349907502532, + 0.005511852912604809, + 0.005521835759282112, + 0.013577069155871868, + 0.0023156022652983665, + 0.0012744000414386392, + 0.0041459049098193645, + 0.001964933006092906, + 0.0010805395431816578, + 0.0017247488722205162, + 0.0017991577042266726, + 0.002707415260374546, + 0.001966324634850025, + 0.0009511896641924977, + 0.0019402755424380302, + 0.001525325351394713, + 0.0009569530375301838, + 0.0011355545138940215, + 0.0016953317681327462, + 0.0028340518474578857, + 0.0009868674678727984, + 0.0005372361629270017, + 0.0010339203290641308, + 0.0008035040809772909, + 0.0005311152781359851, + 0.0005172000965103507, + 0.0008930293843150139, + 0.0013949923450127244, + 0.30500441789627075, + 0.045516740530729294, + 0.040388792753219604, + 0.022437863051891327, + 0.023230625316500664, + 0.0810818299651146, + 0.030343448743224144, + 0.31568968296051025, + 0.0012733896728605032, + 0.0011638773139566183, + 0.001265592174604535, + 0.001043059746734798, + 0.0006524730124510825, + 0.0011083028512075543, + 0.000911026494577527, + 0.001324173528701067, + 0.0010835299035534263, + 0.0007567050051875412, + 0.0015253361780196428, + 0.000620734179392457, + 0.00040615155012346804, + 0.0005902043194510043, + 0.0006878409185446799, + 0.0016448909882456064, + 0.00157981738448143, + 0.000785923795774579, + 0.0008626052876934409, + 0.0005676490836776793, + 0.00043320065014995635, + 0.0006802688003517687, + 0.0005536141688935459, + 0.001372479717247188, + 0.0007636986556462944, + 0.0008072059135884047, + 0.0008294438011944294, + 0.001222548307850957, + 0.0006162478821352124, + 0.0008185909828171134, + 0.0011760372435674071, + 0.0010835619177669287, + 0.00041880185017362237, + 0.00043155645835213363, + 0.00041735870763659477, + 0.00031755215604789555, + 0.00023322914785239846, + 0.0002306922251591459, + 0.0003357152745593339, + 0.0007156722131185234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.007631670217961073 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017137139802798629 + ], + "opposite_logprob_tokens": [ + -6.376713752746582 + ], + "logprob_reward": -0.0017137139802798629, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017137139802798629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10974405705928802, + 0.07529475539922714, + 0.04797065630555153, + 0.03242816403508186, + 0.31855398416519165, + 0.1390458047389984, + 0.10427161306142807, + 0.13103100657463074, + 0.024185247719287872, + 0.01747470162808895 + ], + "credit_map": [ + 0.014899441972374916, + 0.011794046498835087, + 0.0201826561242342, + 0.011730295605957508, + 0.011165923438966274, + 0.009911036118865013, + 0.011070516891777515, + 0.018990134820342064, + 0.010586893185973167, + 0.006003861781209707, + 0.012108268216252327, + 0.010747626423835754, + 0.007063588127493858, + 0.006892967503517866, + 0.01151149533689022, + 0.010380050167441368, + 0.006288587115705013, + 0.003422819310799241, + 0.009388430044054985, + 0.006778439041227102, + 0.0036117543932050467, + 0.003456579754129052, + 0.005564094986766577, + 0.00945995096117258, + 0.004896934609860182, + 0.002703365869820118, + 0.004683276172727346, + 0.003123961854726076, + 0.0023669330403208733, + 0.0030118930153548717, + 0.004172499291598797, + 0.007469302508980036, + 0.11087566614151001, + 0.0177221167832613, + 0.01595773547887802, + 0.008996703661978245, + 0.008879147469997406, + 0.03042680211365223, + 0.012006980367004871, + 0.11368884146213531, + 0.027304528281092644, + 0.012201501056551933, + 0.01624961942434311, + 0.00927125383168459, + 0.006190996151417494, + 0.015376358292996883, + 0.010918185114860535, + 0.04153336212038994, + 0.023142492398619652, + 0.009647788479924202, + 0.013716144487261772, + 0.007776541169732809, + 0.005760836880654097, + 0.008972854353487492, + 0.0074949017725884914, + 0.027760053053498268, + 0.01985028199851513, + 0.013069842010736465, + 0.020934773609042168, + 0.006609478034079075, + 0.007342630531638861, + 0.012928858399391174, + 0.008325450122356415, + 0.04196968674659729, + 0.004859511740505695, + 0.0021752105094492435, + 0.0037409018259495497, + 0.0018835895461961627, + 0.001391131547279656, + 0.0023748106323182583, + 0.00205293670296669, + 0.00570715544745326, + 0.0036682169884443283, + 0.0018684208625927567, + 0.00243161479011178, + 0.0012940488522872329, + 0.000986306695267558, + 0.0020993088837713003, + 0.0012131407856941223, + 0.003913643304258585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 7 + ], + "selected_reward": -0.0017137139802798629 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style that of an oriental ink painting with freehand brushwork?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a fully-winged angel performing a sword swing?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the angel wearing a flowing white robe with calligraphic lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered primarily in black, white, and gray on a rice paper texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly and steadily zoom in from a mid-ground to the angel?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702180445194244 + ], + "opposite_logprob_tokens": [ + -2.1270217895507812 + ], + "logprob_reward": -0.12702180445194244, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702180445194244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13085652887821198, + 0.02656569518148899, + 0.01718141883611679, + 0.022839875891804695, + 0.6365987658500671, + 0.1272847205400467, + 0.013089881278574467, + 0.006045529618859291, + 0.0025575056206434965, + 0.0169801227748394 + ], + "credit_map": [ + 0.0190386064350605, + 0.011082966811954975, + 0.03135886415839195, + 0.01111671794205904, + 0.012013446539640427, + 0.010962859727442265, + 0.011215530335903168, + 0.024067532271146774, + 0.0034834500402212143, + 0.001783727086149156, + 0.009063616394996643, + 0.002847558120265603, + 0.0016033536521717906, + 0.0021460424177348614, + 0.002554928185418248, + 0.0030830171890556812, + 0.0021992200054228306, + 0.0012632786529138684, + 0.0036218080203980207, + 0.0018482916057109833, + 0.0011858524521812797, + 0.0014317677123472095, + 0.002489532344043255, + 0.003141667926684022, + 0.003512352705001831, + 0.0014161098515614867, + 0.004327182658016682, + 0.0030324673280119896, + 0.0018498724093660712, + 0.0016674038488417864, + 0.0032539761159569025, + 0.0037805121392011642, + 0.22541193664073944, + 0.03396925702691078, + 0.029737289994955063, + 0.01674170233309269, + 0.016817284747958183, + 0.059985995292663574, + 0.022668659687042236, + 0.23126660287380219, + 0.033868078142404556, + 0.008883804082870483, + 0.025704167783260345, + 0.009316056035459042, + 0.0063734520226716995, + 0.010825217701494694, + 0.009221462532877922, + 0.0230924803763628, + 0.0026460543740540743, + 0.0011473536724224687, + 0.002008261624723673, + 0.0009090736857615411, + 0.0007150769233703613, + 0.0012064657639712095, + 0.001163235167041421, + 0.0032943591941148043, + 0.0008669125381857157, + 0.0009380152332596481, + 0.000690027023665607, + 0.0007468011463060975, + 0.0005341938231140375, + 0.0006129061221145093, + 0.0005875423667021096, + 0.0010691315401345491, + 0.0002989296044688672, + 0.00022431697288993746, + 0.00039331536390818655, + 0.00024166412185877562, + 0.00014284327335190028, + 0.0002410337037872523, + 0.00021577274310402572, + 0.0007996297790668905, + 0.0030649839900434017, + 0.0020389833953231573, + 0.0024368928279727697, + 0.0012815421214327216, + 0.0010605494026094675, + 0.0020694646518677473, + 0.001326591125689447, + 0.003701114794239402 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.12702180445194244 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_8/index_tensor([7181, 7181])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_8/index_tensor([7181, 7181])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c52157616af9d418658c321c6eeb6fec3d3d5ad8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_8/index_tensor([7181, 7181])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic slow-motion shot, captured from a low angle, tracks a majestic tiger as it moves with calculated grace through the dense, humid jungle undergrowth. The slow-motion effect dramatically amplifies the trembling contractions and flexes of its powerful shoulder and leg muscles beneath its vibrant orange-and-black striped fur, making every hair visible. Its body lies flat against the ground, its massive paws delicately and silently tapping the moist earth covered with giant ferns and fallen leaves. Each step is imbued with a sense of poised energy, its tail trailing steadily behind it, only a slight twitch at the end. Dappled sunlight filters through the layers of treetops, casting shifting shadows on its form. Its amber eyes remain fixed, pupils constricted, locked with unwavering, deadly focus on its prey, just beyond the camera's view. The scene, dominated by a palette of deep greens and warm oranges, is imbued with a quiet tension and the primal spirit of the hunt.", + "seed": 24125, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_8/index_7181_seed_24125_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.21443409081548453, + "vlm_logprob": -0.21443409081548453, + "vlm_yes_no_margin": 2.249999988079071, + "vlm_token_credit": -0.21443409081548453, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.593569815158844 + ], + "opposite_logprob_tokens": [ + -0.843569815158844 + ], + "logprob_reward": -0.593569815158844, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.593569815158844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3504129648208618, + 0.11812864243984222, + 0.04416436329483986, + 0.042932163923978806, + 0.03431480750441551, + 0.19272837042808533, + 0.03396853059530258, + 0.12454340606927872, + 0.03752007707953453, + 0.02128666080534458 + ], + "credit_map": [ + 0.06826931983232498, + 0.022777525708079338, + 0.028842616826295853, + 0.01341159362345934, + 0.028898946940898895, + 0.14530068635940552, + 0.009740627370774746, + 0.03317166864871979, + 0.01159910298883915, + 0.006324952933937311, + 0.01903534308075905, + 0.005741302855312824, + 0.007628557737916708, + 0.04904710873961449, + 0.004153350368142128, + 0.014598929323256016, + 0.007986540906131268, + 0.0038423596415668726, + 0.012526591308414936, + 0.002310957293957472, + 0.003049395279958844, + 0.006033184472471476, + 0.0025267554447054863, + 0.005888580344617367, + 0.012239854782819748, + 0.0035460502840578556, + 0.013109605759382248, + 0.0025241817347705364, + 0.0027920654974877834, + 0.0037378103006631136, + 0.00191965838894248, + 0.0030629346147179604, + 0.006081960629671812, + 0.004885529167950153, + 0.004977331962436438, + 0.003633742220699787, + 0.004231282044202089, + 0.0034027353394776583, + 0.0028565905522555113, + 0.0042456346563994884, + 0.027918972074985504, + 0.01583828777074814, + 0.037389129400253296, + 0.009574259631335735, + 0.013932604342699051, + 0.06773511320352554, + 0.006610732059925795, + 0.013729259371757507, + 0.005164270289242268, + 0.0027492770459502935, + 0.007117800880223513, + 0.0023711256217211485, + 0.0038798630703240633, + 0.0035655912943184376, + 0.002615530276671052, + 0.006505073979496956, + 0.015234964899718761, + 0.02163182944059372, + 0.010947367176413536, + 0.011643623933196068, + 0.017932569608092308, + 0.01151453610509634, + 0.017952876165509224, + 0.01768563687801361, + 0.004667920991778374, + 0.009931785054504871, + 0.0040186066180467606, + 0.0028599021025002003, + 0.0038031323347240686, + 0.0036314609460532665, + 0.005217788740992546, + 0.0033894807565957308, + 0.0026690331287682056, + 0.0022884083446115255, + 0.00284258252941072, + 0.0019756441470235586, + 0.002996287075802684, + 0.00245230202563107, + 0.0023142576683312654, + 0.0037481451872736216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.593569815158844 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870019853115082 + ], + "opposite_logprob_tokens": [ + -3.0487000942230225 + ], + "logprob_reward": -0.04870019853115082, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870019853115082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4401242733001709, + 0.10301162302494049, + 0.04040313884615898, + 0.04495840519666672, + 0.04058096930384636, + 0.12694327533245087, + 0.05188724771142006, + 0.09657318890094757, + 0.023607593029737473, + 0.03191031888127327 + ], + "credit_map": [ + 0.08323337137699127, + 0.03003513626754284, + 0.03980698063969612, + 0.020014503970742226, + 0.035440631210803986, + 0.177141934633255, + 0.015442050993442535, + 0.03900967165827751, + 0.009715595282614231, + 0.0052704354748129845, + 0.02013377472758293, + 0.004332603421062231, + 0.00724644074216485, + 0.03997371718287468, + 0.005098708439618349, + 0.011240347288548946, + 0.006746750790625811, + 0.0052786581218242645, + 0.005740894004702568, + 0.0038388455286622047, + 0.0038407654501497746, + 0.005807226989418268, + 0.003083746414631605, + 0.006066251080483198, + 0.013449027203023434, + 0.0026953211054205894, + 0.01705838181078434, + 0.0024520596489310265, + 0.00202031503431499, + 0.0030363809783011675, + 0.0019404796184971929, + 0.0023064406123012304, + 0.007025470491498709, + 0.00404858123511076, + 0.005688577890396118, + 0.0024556482676416636, + 0.005030875559896231, + 0.006994797848165035, + 0.0034093300346285105, + 0.005927690304815769, + 0.015739111229777336, + 0.011554013937711716, + 0.030118785798549652, + 0.007717563305050135, + 0.013411202467978, + 0.034539852291345596, + 0.005955816712230444, + 0.007906928658485413, + 0.005658639594912529, + 0.0036863237619400024, + 0.010030455887317657, + 0.002457340946421027, + 0.004792168270796537, + 0.006451842375099659, + 0.0031285963486880064, + 0.015681879594922066, + 0.014699960127472878, + 0.01265568658709526, + 0.020293211564421654, + 0.007772217039018869, + 0.009528916329145432, + 0.00904404278844595, + 0.010892342776060104, + 0.011686811223626137, + 0.0025397418066859245, + 0.00664302846416831, + 0.002649182453751564, + 0.0018260229844599962, + 0.002555565908551216, + 0.0018683277303352952, + 0.0019090535352006555, + 0.003616671310737729, + 0.005312205757945776, + 0.0032136342488229275, + 0.004134311806410551, + 0.003299885196611285, + 0.003581414697691798, + 0.0032524510752409697, + 0.003310844302177429, + 0.005805572960525751 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.04870019853115082 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004093378782272 + ], + "opposite_logprob_tokens": [ + -2.6950409412384033 + ], + "logprob_reward": -0.07004093378782272, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004093378782272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.297070175409317, + 0.2596968412399292, + 0.021612510085105896, + 0.02436443790793419, + 0.08095000684261322, + 0.06971113383769989, + 0.03282730653882027, + 0.14972586929798126, + 0.046144865453243256, + 0.017896799370646477 + ], + "credit_map": [ + 0.0578116849064827, + 0.02111952379345894, + 0.023599538952112198, + 0.01188517827540636, + 0.023821881040930748, + 0.12423713505268097, + 0.00786105077713728, + 0.026734158396720886, + 0.02439102716743946, + 0.011734775267541409, + 0.04590875655412674, + 0.009644851088523865, + 0.018288182094693184, + 0.11214133352041245, + 0.0073656183667480946, + 0.030222265049815178, + 0.0026877201162278652, + 0.0015595632139593363, + 0.006506238132715225, + 0.002161105629056692, + 0.0016100745415315032, + 0.0023502400144934654, + 0.0016837272560223937, + 0.0030538421124219894, + 0.003907753154635429, + 0.0024134793784469366, + 0.0038975693751126528, + 0.0021255011670291424, + 0.002979239448904991, + 0.004496739711612463, + 0.002468432066962123, + 0.0020757231395691633, + 0.0094065573066473, + 0.008642070926725864, + 0.011694199405610561, + 0.005159177351742983, + 0.009081791155040264, + 0.024272646754980087, + 0.006354123819619417, + 0.006339438259601593, + 0.008731660433113575, + 0.004278228618204594, + 0.008491748943924904, + 0.0025051378179341555, + 0.009858731180429459, + 0.02872627228498459, + 0.0023295332212001085, + 0.004789820406585932, + 0.005279034376144409, + 0.0024784153793007135, + 0.005194519180804491, + 0.0019185770070180297, + 0.002724479651078582, + 0.004169768188148737, + 0.001991771860048175, + 0.009070741944015026, + 0.02275187335908413, + 0.019858596846461296, + 0.0278402641415596, + 0.013438158668577671, + 0.014966888353228569, + 0.014105378650128841, + 0.017572449520230293, + 0.01919226534664631, + 0.0052610901184380054, + 0.013225439935922623, + 0.0047110929153859615, + 0.0033927320037037134, + 0.004162986297160387, + 0.0046955738216638565, + 0.005990392528474331, + 0.004705557134002447, + 0.0028133115265518427, + 0.0022877082228660583, + 0.002928720787167549, + 0.0018695403123274446, + 0.0016055864980444312, + 0.0017463704571127892, + 0.00156329944729805, + 0.0030822628177702427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.07004093378782272 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108358334749937 + ], + "opposite_logprob_tokens": [ + -4.511083602905273 + ], + "logprob_reward": -0.01108358334749937, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108358334749937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5432299971580505, + 0.17332713305950165, + 0.010414445772767067, + 0.017941830679774284, + 0.034881360828876495, + 0.13517339527606964, + 0.02512538991868496, + 0.02474113367497921, + 0.02302682027220726, + 0.01213841699063778 + ], + "credit_map": [ + 0.10863776504993439, + 0.038050323724746704, + 0.042870502918958664, + 0.020779982209205627, + 0.04370063543319702, + 0.2285989671945572, + 0.013309722766280174, + 0.04728216677904129, + 0.01686798594892025, + 0.008477438241243362, + 0.026024237275123596, + 0.006860470864921808, + 0.012186281383037567, + 0.07398078590631485, + 0.0053359875455498695, + 0.023593956604599953, + 0.002059376798570156, + 0.0009738894295878708, + 0.002021215623244643, + 0.0008870352758094668, + 0.0008189330692403018, + 0.0015543008921667933, + 0.0007015571463853121, + 0.0013981389347463846, + 0.0040680947713553905, + 0.0016618918161839247, + 0.003432379337027669, + 0.001702187000773847, + 0.0017213780665770173, + 0.002012085635215044, + 0.001798955607227981, + 0.0015448607737198472, + 0.004899118561297655, + 0.002984006190672517, + 0.00482237758114934, + 0.0020184689201414585, + 0.0043294914066791534, + 0.009397553279995918, + 0.0030468881595879793, + 0.0033834606874734163, + 0.017043527215719223, + 0.012051940895617008, + 0.03169620409607887, + 0.007641617674380541, + 0.012296623550355434, + 0.04034487158060074, + 0.005440617445856333, + 0.008658000268042088, + 0.0032352060079574585, + 0.002619639039039612, + 0.003962907008826733, + 0.0014675022102892399, + 0.004000719636678696, + 0.003005297388881445, + 0.0020760984625667334, + 0.0047580222599208355, + 0.0032549630850553513, + 0.0031792104709893465, + 0.0021758335642516613, + 0.0025479313917458057, + 0.0030805193819105625, + 0.0025501716881990433, + 0.00449704984202981, + 0.003455457277595997, + 0.0035172314383089542, + 0.005335415713489056, + 0.00256387353874743, + 0.0018787289736792445, + 0.0020742015913128853, + 0.0016229754546657205, + 0.0023244505282491446, + 0.003709944197908044, + 0.0016931361751630902, + 0.0013106161495670676, + 0.0017519716639071703, + 0.0010502758668735623, + 0.0013122701784595847, + 0.0009051976376213133, + 0.0012778814416378736, + 0.0028370690997689962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.01108358334749937 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3487759232521057 + ], + "opposite_logprob_tokens": [ + -1.223775863647461 + ], + "logprob_reward": -0.3487759232521057, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3487759232521057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2550554573535919, + 0.11204833537340164, + 0.06607092916965485, + 0.0372377373278141, + 0.12959852814674377, + 0.08789447695016861, + 0.03779171034693718, + 0.10454434156417847, + 0.1356731653213501, + 0.034085314720869064 + ], + "credit_map": [ + 0.048979997634887695, + 0.0211263969540596, + 0.02467449940741062, + 0.013309144414961338, + 0.020640477538108826, + 0.09693025052547455, + 0.009788650088012218, + 0.019606053829193115, + 0.011647901497781277, + 0.008056418970227242, + 0.014760165475308895, + 0.006980092264711857, + 0.008317749947309494, + 0.03866986930370331, + 0.0053770458325743675, + 0.01823909766972065, + 0.010897980071604252, + 0.005529324058443308, + 0.019657103344798088, + 0.005783993285149336, + 0.004562043584883213, + 0.007917776703834534, + 0.004402332007884979, + 0.0073203714564442635, + 0.004799524322152138, + 0.0040566702373325825, + 0.005348914302885532, + 0.0032220634166151285, + 0.005455245729535818, + 0.00696149654686451, + 0.004432408604770899, + 0.002961412537842989, + 0.02188989706337452, + 0.0104652289301157, + 0.022295242175459862, + 0.007945353165268898, + 0.011339451186358929, + 0.03407958522439003, + 0.008371835574507713, + 0.013211937621235847, + 0.010908698663115501, + 0.0069732507690787315, + 0.01318056508898735, + 0.0055702184326946735, + 0.00857408344745636, + 0.02944633923470974, + 0.0049884081818163395, + 0.00825291033834219, + 0.004558052401989698, + 0.004547958727926016, + 0.0067404876463115215, + 0.003813314950093627, + 0.004796445369720459, + 0.004322800785303116, + 0.0025807730853557587, + 0.006431879475712776, + 0.013870040886104107, + 0.017530903220176697, + 0.012373712845146656, + 0.009628629311919212, + 0.015026750043034554, + 0.00924897100776434, + 0.011617634445428848, + 0.015247696079313755, + 0.02368101291358471, + 0.01271407026797533, + 0.019812414422631264, + 0.012037576176226139, + 0.011078848503530025, + 0.017112305387854576, + 0.020735900849103928, + 0.01850103959441185, + 0.004780454561114311, + 0.004599170293658972, + 0.005804999265819788, + 0.0030483969021588564, + 0.0036190617829561234, + 0.0037480818573385477, + 0.0030273646116256714, + 0.0054577854461967945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.3487759232521057 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic slow-motion shot, captured from a low angle, tracks a majestic tiger as it moves with calculated grace through the dense, humid jungle undergrowth. The slow-motion effect dramatically amplifies the trembling contractions and flexes of its powerful shoulder and leg muscles beneath its vibrant orange-and-black striped fur, making every hair visible. Its body lies flat against the ground, its massive paws delicately and silently tapping the moist earth covered with giant ferns and fallen leaves. Each step is imbued with a sense of poised energy, its tail trailing steadily behind it, only a slight twitch at the end. Dappled sunlight filters through the layers of treetops, casting shifting shadows on its form. Its amber eyes remain fixed, pupils constricted, locked with unwavering, deadly focus on its prey, just beyond the camera's view. The scene, dominated by a palette of deep greens and warm oranges, is imbued with a quiet tension and the primal spirit of the hunt.", + "seed": 24126, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_8/index_7181_seed_24126_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16643305849283935, + "vlm_logprob": -0.16643305849283935, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.16643305849283935, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.40855976939201355 + ], + "opposite_logprob_tokens": [ + -1.158559799194336 + ], + "logprob_reward": -0.40855976939201355, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.40855976939201355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10642095655202866, + 0.045488812029361725, + 0.03282308205962181, + 0.028948744758963585, + 0.2964778542518616, + 0.07294712960720062, + 0.11499811708927155, + 0.09078923612833023, + 0.16406302154064178, + 0.04704301804304123 + ], + "credit_map": [ + 0.022071044892072678, + 0.01246598083525896, + 0.016990458592772484, + 0.009389882907271385, + 0.009585985913872719, + 0.011811287142336369, + 0.01016142312437296, + 0.013944892212748528, + 0.006720676552504301, + 0.0034117221366614103, + 0.00470882561057806, + 0.004355926997959614, + 0.00438023591414094, + 0.00989566184580326, + 0.0041594430804252625, + 0.007856318727135658, + 0.0043501281179487705, + 0.002606110880151391, + 0.0046356781385838985, + 0.003968866541981697, + 0.003362181130796671, + 0.00649307481944561, + 0.00354245537891984, + 0.003864588215947151, + 0.003826037049293518, + 0.0031487406231462955, + 0.004183757118880749, + 0.003556500654667616, + 0.0028133292216807604, + 0.0032657322008162737, + 0.0038519781082868576, + 0.004302669316530228, + 0.061758656054735184, + 0.018987655639648438, + 0.04897666722536087, + 0.014283192344009876, + 0.02257446199655533, + 0.06673231720924377, + 0.019851932302117348, + 0.043313004076480865, + 0.008014091290533543, + 0.007650486659258604, + 0.011763975024223328, + 0.008837143890559673, + 0.006137581542134285, + 0.012049908749759197, + 0.009107902646064758, + 0.009386034682393074, + 0.012994502671062946, + 0.010512309148907661, + 0.034569986164569855, + 0.014483919367194176, + 0.008784137666225433, + 0.010971126146614552, + 0.00857776589691639, + 0.014104361645877361, + 0.00999961607158184, + 0.012856932356953621, + 0.011459879577159882, + 0.009026770479977131, + 0.015055621974170208, + 0.013224661350250244, + 0.010847464203834534, + 0.008318289183080196, + 0.033356841653585434, + 0.02266629785299301, + 0.01818065531551838, + 0.016523590311408043, + 0.018932398408651352, + 0.014420167542994022, + 0.01856478862464428, + 0.021418292075395584, + 0.007114164996892214, + 0.006305600982159376, + 0.006017724052071571, + 0.004543777089565992, + 0.006034483201801777, + 0.005930708721280098, + 0.005574940238147974, + 0.0055216168984770775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.40855976939201355 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373844176530838 + ], + "opposite_logprob_tokens": [ + -3.408738374710083 + ], + "logprob_reward": -0.03373844176530838, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373844176530838, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08487682789564133, + 0.051307011395692825, + 0.022794710472226143, + 0.034244928508996964, + 0.22875210642814636, + 0.3687707781791687, + 0.032996710389852524, + 0.06256435066461563, + 0.07659119367599487, + 0.037101466208696365 + ], + "credit_map": [ + 0.011305542662739754, + 0.007815556600689888, + 0.018134931102395058, + 0.01137601025402546, + 0.006130247376859188, + 0.00989516917616129, + 0.008605721406638622, + 0.011613640002906322, + 0.007317058276385069, + 0.0040642861276865005, + 0.006503337994217873, + 0.004794689826667309, + 0.0053862822242081165, + 0.01217519212514162, + 0.0052132741548120975, + 0.005852884147316217, + 0.0037383600138127804, + 0.0019285932648926973, + 0.002699930453673005, + 0.0031102653592824936, + 0.0022945902310311794, + 0.004480073228478432, + 0.0016903416253626347, + 0.002852554665878415, + 0.0047422414645552635, + 0.0035791436675935984, + 0.004525663796812296, + 0.0035260722506791353, + 0.003266833024099469, + 0.005812689196318388, + 0.00427849730476737, + 0.004513782914727926, + 0.04706432670354843, + 0.014820070005953312, + 0.03305380791425705, + 0.01077357865869999, + 0.016406197100877762, + 0.055181678384542465, + 0.014386299066245556, + 0.03706611320376396, + 0.03712890297174454, + 0.05258147418498993, + 0.03366340696811676, + 0.04872526600956917, + 0.03351767733693123, + 0.057396095246076584, + 0.05497711896896362, + 0.05078078806400299, + 0.003182551823556423, + 0.0038962343242019415, + 0.005402233451604843, + 0.00501168891787529, + 0.0038544239941984415, + 0.003141094697639346, + 0.0036860588006675243, + 0.004822420421987772, + 0.008748934604227543, + 0.01049260888248682, + 0.008575168438255787, + 0.005646413192152977, + 0.009598176926374435, + 0.007232475094497204, + 0.006530656013637781, + 0.005739913322031498, + 0.013135774992406368, + 0.0098837586119771, + 0.00772344134747982, + 0.008748078718781471, + 0.009855661541223526, + 0.009819203987717628, + 0.007545433938503265, + 0.009879836812615395, + 0.004715743474662304, + 0.004623430781066418, + 0.004382130689918995, + 0.004580504726618528, + 0.004341246094554663, + 0.005128365941345692, + 0.004585289396345615, + 0.00474474998190999 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.03373844176530838 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205534189939499 + ], + "opposite_logprob_tokens": [ + -2.8120553493499756 + ], + "logprob_reward": -0.06205534189939499, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205534189939499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09005259722471237, + 0.06616686284542084, + 0.020709101110696793, + 0.09976234287023544, + 0.3416610360145569, + 0.1112276017665863, + 0.07369289547204971, + 0.07549024373292923, + 0.07463786005973816, + 0.04659948870539665 + ], + "credit_map": [ + 0.015133919194340706, + 0.009746396914124489, + 0.017653584480285645, + 0.008535546250641346, + 0.007973839528858662, + 0.012041440233588219, + 0.00831211730837822, + 0.010655753314495087, + 0.011141524650156498, + 0.006738973781466484, + 0.009032835252583027, + 0.007909491658210754, + 0.00874998141080141, + 0.007147481199353933, + 0.00929392408579588, + 0.006152649875730276, + 0.0027618450112640858, + 0.0020726537331938744, + 0.002381991595029831, + 0.0022918216418474913, + 0.002669272944331169, + 0.004007832147181034, + 0.002203131327405572, + 0.002320552710443735, + 0.01345523539930582, + 0.007988584227859974, + 0.010580667294561863, + 0.006656104698777199, + 0.012541379779577255, + 0.021320100873708725, + 0.010663891211152077, + 0.016556372866034508, + 0.07119443267583847, + 0.022854719310998917, + 0.04773653298616409, + 0.01650565303862095, + 0.02567962557077408, + 0.08573018014431, + 0.022508889436721802, + 0.04945101961493492, + 0.012631208635866642, + 0.0139614874497056, + 0.012351824901998043, + 0.013389628380537033, + 0.010586854070425034, + 0.017653455957770348, + 0.015482387505471706, + 0.015170764178037643, + 0.009644465520977974, + 0.0079368706792593, + 0.019156264141201973, + 0.00912944320589304, + 0.006919659674167633, + 0.006317676045000553, + 0.005779905244708061, + 0.008808606304228306, + 0.007844890467822552, + 0.012243682518601418, + 0.009715846739709377, + 0.006073893513530493, + 0.014905376359820366, + 0.008658944629132748, + 0.008384070359170437, + 0.007663540076464415, + 0.010485490784049034, + 0.011258947663009167, + 0.010659596882760525, + 0.007320031989365816, + 0.00968961138278246, + 0.0072177075780928135, + 0.007911170832812786, + 0.010095308534801006, + 0.004541024565696716, + 0.006592504680156708, + 0.005375474691390991, + 0.005426413845270872, + 0.006245228927582502, + 0.006091785151511431, + 0.0071281311102211475, + 0.0051989248022437096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.06205534189939499 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014202586375176907 + ], + "opposite_logprob_tokens": [ + -4.26420259475708 + ], + "logprob_reward": -0.014202586375176907, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014202586375176907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11291215568780899, + 0.05121670663356781, + 0.028800494968891144, + 0.02381371520459652, + 0.3255752921104431, + 0.10993987321853638, + 0.06187761574983597, + 0.10574746131896973, + 0.11950578540563583, + 0.06061091274023056 + ], + "credit_map": [ + 0.018496450036764145, + 0.01045076735317707, + 0.023084379732608795, + 0.010331695899367332, + 0.008118554018437862, + 0.012752133421599865, + 0.012258749455213547, + 0.017419423907995224, + 0.00863876473158598, + 0.0035738206934183836, + 0.010578634217381477, + 0.004673233721405268, + 0.004380149766802788, + 0.007992180995643139, + 0.004479395691305399, + 0.006900528445839882, + 0.006416452117264271, + 0.002062102546915412, + 0.0037210939917713404, + 0.0022212162148207426, + 0.0030456765089184046, + 0.0049522737972438335, + 0.0027083400636911392, + 0.0036733385641127825, + 0.0034616943448781967, + 0.0016124102985486388, + 0.003801513696089387, + 0.0018312358297407627, + 0.0024421741254627705, + 0.004114336334168911, + 0.0015953293768689036, + 0.004955020267516375, + 0.06541094928979874, + 0.02056017331779003, + 0.043203406035900116, + 0.017798613756895065, + 0.0251466017216444, + 0.08259335905313492, + 0.024014856666326523, + 0.04684734344482422, + 0.012708946131169796, + 0.014409192837774754, + 0.013600274920463562, + 0.014834832400083542, + 0.009803352877497673, + 0.01564587652683258, + 0.013460619375109673, + 0.015476783737540245, + 0.006620088592171669, + 0.006695395335555077, + 0.013333737850189209, + 0.008191155269742012, + 0.007786499336361885, + 0.0054792072623968124, + 0.006015571299940348, + 0.007755961734801531, + 0.015020912513136864, + 0.018612615764141083, + 0.015089195221662521, + 0.00966491550207138, + 0.015851536765694618, + 0.011743088252842426, + 0.01003467570990324, + 0.00973053090274334, + 0.027163071557879448, + 0.018488412722945213, + 0.012547350488603115, + 0.012523291632533073, + 0.013048206456005573, + 0.011465661227703094, + 0.008674276061356068, + 0.015595516189932823, + 0.005107135511934757, + 0.005183723289519548, + 0.005035201087594032, + 0.005422879941761494, + 0.007726038806140423, + 0.016031933948397636, + 0.006516226567327976, + 0.009587773121893406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.014202586375176907 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31360915303230286 + ], + "opposite_logprob_tokens": [ + -1.3136091232299805 + ], + "logprob_reward": -0.31360915303230286, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31360915303230286, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06430356949567795, + 0.028924718499183655, + 0.008581703528761864, + 0.016960518434643745, + 0.6066210269927979, + 0.037596769630908966, + 0.03362659737467766, + 0.05501919984817505, + 0.12095701694488525, + 0.027408864349126816 + ], + "credit_map": [ + 0.009126119315624237, + 0.006919550243765116, + 0.009691640734672546, + 0.00636222492903471, + 0.006294607184827328, + 0.008494148030877113, + 0.008797567337751389, + 0.008617712184786797, + 0.004959780722856522, + 0.0031664669513702393, + 0.003180839354172349, + 0.0030430322512984276, + 0.003263333812355995, + 0.004544982220977545, + 0.0033898113761097193, + 0.00337647064588964, + 0.0016618153313174844, + 0.0009827703470364213, + 0.001208797562867403, + 0.0009069002699106932, + 0.0008636934799142182, + 0.0008949632756412029, + 0.0008827365236356854, + 0.001180026913061738, + 0.002230691723525524, + 0.00168037589173764, + 0.0028764845337718725, + 0.0019205986754968762, + 0.0016479997429996729, + 0.0027158106677234173, + 0.001716243801638484, + 0.00217231223359704, + 0.13032902777194977, + 0.037557777017354965, + 0.07483107596635818, + 0.028431085869669914, + 0.045602429658174515, + 0.15830789506435394, + 0.04277164489030838, + 0.08879010379314423, + 0.004962010774761438, + 0.0037750445771962404, + 0.004578098654747009, + 0.004707768559455872, + 0.0037228353321552277, + 0.005595542956143618, + 0.00484901387244463, + 0.005406455602496862, + 0.004252045415341854, + 0.0035191960632801056, + 0.008408263325691223, + 0.003933440428227186, + 0.002987514715641737, + 0.0032226447947323322, + 0.0023240409791469574, + 0.0049794516526162624, + 0.0076716891489923, + 0.01008554082363844, + 0.007356784772127867, + 0.005200161132961512, + 0.008135327138006687, + 0.006196123082190752, + 0.0051632472313940525, + 0.00521032465621829, + 0.034024812281131744, + 0.02133564092218876, + 0.011875149793922901, + 0.010707598179578781, + 0.009972671046853065, + 0.00803693849593401, + 0.00908921379595995, + 0.015915000811219215, + 0.0028254659846425056, + 0.0037919001188129187, + 0.0028572569135576487, + 0.0024693801533430815, + 0.0035891588777303696, + 0.00573509419336915, + 0.0027578205335885286, + 0.0033827892038971186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.31360915303230286 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic slow-motion shot, captured from a low angle, tracks a majestic tiger as it moves with calculated grace through the dense, humid jungle undergrowth. The slow-motion effect dramatically amplifies the trembling contractions and flexes of its powerful shoulder and leg muscles beneath its vibrant orange-and-black striped fur, making every hair visible. Its body lies flat against the ground, its massive paws delicately and silently tapping the moist earth covered with giant ferns and fallen leaves. Each step is imbued with a sense of poised energy, its tail trailing steadily behind it, only a slight twitch at the end. Dappled sunlight filters through the layers of treetops, casting shifting shadows on its form. Its amber eyes remain fixed, pupils constricted, locked with unwavering, deadly focus on its prey, just beyond the camera's view. The scene, dominated by a palette of deep greens and warm oranges, is imbued with a quiet tension and the primal spirit of the hunt.", + "seed": 24127, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_8/index_7181_seed_24127_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.24937536269426347, + "vlm_logprob": -0.24937536269426347, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.24937536269426347, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6006869673728943 + ], + "opposite_logprob_tokens": [ + -0.8506869673728943 + ], + "logprob_reward": -0.6006869673728943, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.6006869673728943, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10771028697490692, + 0.028296280652284622, + 0.02004517801105976, + 0.01909448765218258, + 0.012581640854477882, + 0.016103602945804596, + 0.48700985312461853, + 0.015914782881736755, + 0.07246773689985275, + 0.22077609598636627 + ], + "credit_map": [ + 0.016066132113337517, + 0.012073475867509842, + 0.012477909214794636, + 0.021573448553681374, + 0.012278502807021141, + 0.010247382335364819, + 0.015902172774076462, + 0.0070912791416049, + 0.004263607785105705, + 0.0026995192747563124, + 0.005837822332978249, + 0.0030806108843535185, + 0.003500622697174549, + 0.002088526263833046, + 0.0034735749941319227, + 0.0033519999124109745, + 0.0034858090803027153, + 0.002621867461130023, + 0.004080981481820345, + 0.0018948111683130264, + 0.001802213373593986, + 0.0026275902055203915, + 0.0014410135336220264, + 0.002090892754495144, + 0.0034749156329780817, + 0.0019172156462445855, + 0.003325961995869875, + 0.002578874584287405, + 0.0018539105076342821, + 0.0024830990005284548, + 0.0017874418990686536, + 0.001673070597462356, + 0.0016370095545426011, + 0.0011338448384776711, + 0.0015257380437105894, + 0.0020814950112253428, + 0.0014301789924502373, + 0.0015226146206259727, + 0.0012353864731267095, + 0.0020153748337179422, + 0.001668261713348329, + 0.0022259072866290808, + 0.002172185108065605, + 0.0023472653701901436, + 0.0019294503144919872, + 0.0017164369346573949, + 0.0020343803334981203, + 0.002009717980399728, + 0.08212455362081528, + 0.03610711917281151, + 0.054128993302583694, + 0.07863543927669525, + 0.024396033957600594, + 0.02872484177350998, + 0.0265461727976799, + 0.15634673833847046, + 0.0025571982841938734, + 0.0013891762355342507, + 0.003219248028472066, + 0.002077092882245779, + 0.0015237083425745368, + 0.0011610538931563497, + 0.0017910725437104702, + 0.0021962341852486134, + 0.005196723155677319, + 0.00240488164126873, + 0.007881773635745049, + 0.002687417436391115, + 0.0040025836788117886, + 0.028753994032740593, + 0.0033728282433003187, + 0.01816754788160324, + 0.02425759844481945, + 0.029275676235556602, + 0.021195119246840477, + 0.014296685345470905, + 0.03448255732655525, + 0.06551361829042435, + 0.013671140186488628, + 0.018083715811371803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.6006869673728943 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890807956457138 + ], + "opposite_logprob_tokens": [ + -2.464080810546875 + ], + "logprob_reward": -0.0890807956457138, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890807956457138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1306198239326477, + 0.0625627264380455, + 0.036474332213401794, + 0.028903326019644737, + 0.03862353414297104, + 0.045410092920064926, + 0.12114027142524719, + 0.03664412721991539, + 0.35702967643737793, + 0.14259205758571625 + ], + "credit_map": [ + 0.017713401466608047, + 0.014575856737792492, + 0.017334263771772385, + 0.0275224931538105, + 0.013272440060973167, + 0.010550194419920444, + 0.019974445924162865, + 0.009676729328930378, + 0.010011348873376846, + 0.007059566210955381, + 0.009817566722631454, + 0.007709456142038107, + 0.006111292168498039, + 0.008247136138379574, + 0.00691925548017025, + 0.006687107961624861, + 0.006205606274306774, + 0.0048108939081430435, + 0.006613404024392366, + 0.005227560177445412, + 0.0031712243799120188, + 0.0031568396370857954, + 0.002847911324352026, + 0.004440891556441784, + 0.003918709233403206, + 0.0038785783108323812, + 0.004756217822432518, + 0.003459923667833209, + 0.0030421526171267033, + 0.003958138637244701, + 0.002581540495157242, + 0.0033080654684454203, + 0.00758590130135417, + 0.004614620003849268, + 0.005026457831263542, + 0.005760528147220612, + 0.003432448487728834, + 0.003182907123118639, + 0.004001231864094734, + 0.005019442178308964, + 0.006617696024477482, + 0.0064061046577990055, + 0.006022012326866388, + 0.003940701484680176, + 0.007012592628598213, + 0.005101141985505819, + 0.0038006992544978857, + 0.0065091447904706, + 0.01829097978770733, + 0.009657769463956356, + 0.04275226593017578, + 0.019191285595297813, + 0.008284670300781727, + 0.006130875088274479, + 0.007169318851083517, + 0.009663105010986328, + 0.006713593378663063, + 0.0038157999515533447, + 0.005045456811785698, + 0.005145900882780552, + 0.0033206255175173283, + 0.004127050284296274, + 0.0034415738191455603, + 0.005034126341342926, + 0.028614798560738564, + 0.01493728905916214, + 0.03830578550696373, + 0.020848747342824936, + 0.024920055642724037, + 0.12071195244789124, + 0.028162939473986626, + 0.0805281400680542, + 0.0138064781203866, + 0.01672060787677765, + 0.029963940382003784, + 0.01213350985199213, + 0.017285147681832314, + 0.022399527952075005, + 0.01384549867361784, + 0.016437342390418053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0890807956457138 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062071364372968674 + ], + "opposite_logprob_tokens": [ + -2.8120713233947754 + ], + "logprob_reward": -0.062071364372968674, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062071364372968674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11997982114553452, + 0.053706519305706024, + 0.033808644860982895, + 0.022618355229496956, + 0.055212780833244324, + 0.09146932512521744, + 0.3567686975002289, + 0.023951005190610886, + 0.039371903985738754, + 0.20311301946640015 + ], + "credit_map": [ + 0.018464867025613785, + 0.013251544907689095, + 0.016170307993888855, + 0.019264008849859238, + 0.01334415003657341, + 0.01372048445045948, + 0.015889663249254227, + 0.00987477321177721, + 0.007723954971879721, + 0.006845570635050535, + 0.007900484837591648, + 0.0064382776618003845, + 0.005480293184518814, + 0.007547367829829454, + 0.005954107269644737, + 0.005816458258777857, + 0.004520089831203222, + 0.003651717444881797, + 0.0067132618278265, + 0.002783177187666297, + 0.0036159560550004244, + 0.004566005896776915, + 0.00459054671227932, + 0.003367883153259754, + 0.003163965418934822, + 0.0023431049194186926, + 0.00312611972913146, + 0.002661058446392417, + 0.002848921110853553, + 0.003404261078685522, + 0.0021053915843367577, + 0.0029655296821147203, + 0.007128303404897451, + 0.005070357117801905, + 0.007468539755791426, + 0.008088955655694008, + 0.005211531650274992, + 0.0048691146075725555, + 0.009866737760603428, + 0.007509233895689249, + 0.012505457736551762, + 0.014362000860273838, + 0.011844620108604431, + 0.01259954646229744, + 0.012378370389342308, + 0.009161224588751793, + 0.009705967269837856, + 0.00891212746500969, + 0.060562312602996826, + 0.02054113708436489, + 0.05783413350582123, + 0.06221703812479973, + 0.018026523292064667, + 0.01938662678003311, + 0.018030276522040367, + 0.10017060488462448, + 0.004209225997328758, + 0.0019065019441768527, + 0.003435998922213912, + 0.00435006245970726, + 0.0019364957697689533, + 0.0017551082419231534, + 0.0033375404309481382, + 0.003020069794729352, + 0.004830809310078621, + 0.004146979656070471, + 0.008139795623719692, + 0.004657503217458725, + 0.004153676796704531, + 0.0041864668019115925, + 0.00492510013282299, + 0.004331569187343121, + 0.01778220944106579, + 0.02493821457028389, + 0.051172636449337006, + 0.01714339293539524, + 0.02879457175731659, + 0.023919444531202316, + 0.01840285025537014, + 0.020959677174687386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.062071364372968674 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059270069003105 + ], + "opposite_logprob_tokens": [ + -3.89559268951416 + ], + "logprob_reward": -0.02059270069003105, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059270069003105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0774700716137886, + 0.043004702776670456, + 0.03126148506999016, + 0.030156079679727554, + 0.01613483764231205, + 0.06721409410238266, + 0.29059499502182007, + 0.050257883965969086, + 0.20667679607868195, + 0.18722906708717346 + ], + "credit_map": [ + 0.013291561044752598, + 0.008329559117555618, + 0.009499064646661282, + 0.013640562072396278, + 0.007989964447915554, + 0.007830502465367317, + 0.008935499005019665, + 0.007953362539410591, + 0.005918871611356735, + 0.005214581731706858, + 0.008140974678099155, + 0.003735555335879326, + 0.004941959399729967, + 0.004480718169361353, + 0.005358368624001741, + 0.005213674157857895, + 0.007074575871229172, + 0.0025895810686051846, + 0.007932865060865879, + 0.0028022287879139185, + 0.002416410716250539, + 0.003204934997484088, + 0.0025059308391064405, + 0.0027349565643817186, + 0.004128586500883102, + 0.004300572443753481, + 0.0048055327497422695, + 0.003445935435593128, + 0.002514149295166135, + 0.004106649663299322, + 0.003382451366633177, + 0.003472201758995652, + 0.002754592802375555, + 0.0018345589051023126, + 0.0023996331728994846, + 0.0019750213250517845, + 0.0014574398519471288, + 0.001861143158748746, + 0.001636127708479762, + 0.0022163204848766327, + 0.010910230688750744, + 0.009588018991053104, + 0.008946070447564125, + 0.008542316034436226, + 0.009374997578561306, + 0.006457311101257801, + 0.006912573706358671, + 0.00648257602006197, + 0.04543618857860565, + 0.028073864057660103, + 0.029793012887239456, + 0.04339108616113663, + 0.017219673842191696, + 0.019276976585388184, + 0.017260288819670677, + 0.09014390408992767, + 0.013229054398834705, + 0.0030488022603094578, + 0.008488995023071766, + 0.007723914459347725, + 0.004324739798903465, + 0.0036687683314085007, + 0.0043527428060770035, + 0.005420865956693888, + 0.014631222002208233, + 0.006330476608127356, + 0.021752040833234787, + 0.006554726976901293, + 0.012399956583976746, + 0.08462686836719513, + 0.008106924593448639, + 0.052274592220783234, + 0.02174447663128376, + 0.018698597326874733, + 0.028706751763820648, + 0.011735630221664906, + 0.021832328289747238, + 0.05324165150523186, + 0.01726919785141945, + 0.014000432565808296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.02059270069003105 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4744449853897095 + ], + "opposite_logprob_tokens": [ + -0.9744449853897095 + ], + "logprob_reward": -0.4744449853897095, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4744449853897095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06762170791625977, + 0.05618836358189583, + 0.024525366723537445, + 0.041233718395233154, + 0.029638372361660004, + 0.0837097242474556, + 0.11227910220623016, + 0.02063136175274849, + 0.04850170388817787, + 0.5156705975532532 + ], + "credit_map": [ + 0.013319630175828934, + 0.008252635598182678, + 0.008888296782970428, + 0.01076713390648365, + 0.006381072103977203, + 0.006056939251720905, + 0.007771632168442011, + 0.006184370722621679, + 0.007992963306605816, + 0.007396969944238663, + 0.00923841167241335, + 0.0077640581876039505, + 0.006777991075068712, + 0.005418793298304081, + 0.005532263778150082, + 0.00606691325083375, + 0.0031480195466428995, + 0.0023833797313272953, + 0.003985207993537188, + 0.0022717388346791267, + 0.002895009471103549, + 0.0031864303164184093, + 0.0022863510530442, + 0.004369230940937996, + 0.006392688490450382, + 0.004678014665842056, + 0.006496398709714413, + 0.004641475155949593, + 0.004179902374744415, + 0.005784949753433466, + 0.004334721714258194, + 0.004725566133856773, + 0.0053542847745120525, + 0.003250239649787545, + 0.003419172717258334, + 0.005113947670906782, + 0.003107042284682393, + 0.0026422215159982443, + 0.003177824430167675, + 0.00357363885268569, + 0.009600863792002201, + 0.014233001507818699, + 0.011784317903220654, + 0.008978965692222118, + 0.0142869483679533, + 0.009680268354713917, + 0.00796031579375267, + 0.007185039110481739, + 0.018212316557765007, + 0.013074526563286781, + 0.026989657431840897, + 0.013750486075878143, + 0.010860906913876534, + 0.007870784029364586, + 0.010910000652074814, + 0.010610422119498253, + 0.003268948756158352, + 0.0022351325023919344, + 0.0022455849684774876, + 0.0025602676905691624, + 0.0020642606541514397, + 0.0029617580585181713, + 0.002422216348350048, + 0.0028731939382851124, + 0.004713086877018213, + 0.0039195227436721325, + 0.005460415501147509, + 0.0033929517958313227, + 0.0047418889589607716, + 0.013444497250020504, + 0.003268447471782565, + 0.009560895152390003, + 0.053542811423540115, + 0.07344690710306168, + 0.05204343795776367, + 0.04352220147848129, + 0.08218595385551453, + 0.12991395592689514, + 0.034528203308582306, + 0.046487126499414444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 5 + ], + "selected_reward": -0.4744449853897095 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic slow-motion shot, captured from a low angle, tracks a majestic tiger as it moves with calculated grace through the dense, humid jungle undergrowth. The slow-motion effect dramatically amplifies the trembling contractions and flexes of its powerful shoulder and leg muscles beneath its vibrant orange-and-black striped fur, making every hair visible. Its body lies flat against the ground, its massive paws delicately and silently tapping the moist earth covered with giant ferns and fallen leaves. Each step is imbued with a sense of poised energy, its tail trailing steadily behind it, only a slight twitch at the end. Dappled sunlight filters through the layers of treetops, casting shifting shadows on its form. Its amber eyes remain fixed, pupils constricted, locked with unwavering, deadly focus on its prey, just beyond the camera's view. The scene, dominated by a palette of deep greens and warm oranges, is imbued with a quiet tension and the primal spirit of the hunt.", + "seed": 24128, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_8/index_7181_seed_24128_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2190822869539261, + "vlm_logprob": -0.2190822869539261, + "vlm_yes_no_margin": 2.1749999642372133, + "vlm_token_credit": -0.2190822869539261, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4499651789665222 + ], + "opposite_logprob_tokens": [ + -1.074965238571167 + ], + "logprob_reward": -0.4499651789665222, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.4499651789665222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08412589132785797, + 0.08341410756111145, + 0.12568578124046326, + 0.07865393906831741, + 0.10658792406320572, + 0.048029713332653046, + 0.10313563793897629, + 0.10799602419137955, + 0.23968520760536194, + 0.02268579788506031 + ], + "credit_map": [ + 0.009057499468326569, + 0.008235848508775234, + 0.010433617979288101, + 0.007369235157966614, + 0.010323483496904373, + 0.013375815935432911, + 0.008916390128433704, + 0.01641399785876274, + 0.00479045370593667, + 0.0053512719459831715, + 0.009068498387932777, + 0.014144356362521648, + 0.0048864735290408134, + 0.009315474890172482, + 0.021924419328570366, + 0.013933159410953522, + 0.024191828444600105, + 0.016514483839273453, + 0.015006444416940212, + 0.01582501083612442, + 0.011402204632759094, + 0.0124423922970891, + 0.012790891341865063, + 0.017512520775198936, + 0.010318481363356113, + 0.008423852734267712, + 0.01414361223578453, + 0.007970535196363926, + 0.006704534869641066, + 0.012402253225445747, + 0.005246100015938282, + 0.013444569893181324, + 0.006623223423957825, + 0.006430347450077534, + 0.01690373569726944, + 0.011469928547739983, + 0.011570248752832413, + 0.025532126426696777, + 0.012613171711564064, + 0.015445133671164513, + 0.006365584209561348, + 0.004313501063734293, + 0.009155666455626488, + 0.00503980927169323, + 0.005431779660284519, + 0.006136526819318533, + 0.005195120349526405, + 0.006391724571585655, + 0.011281239800155163, + 0.010911188088357449, + 0.01596161536872387, + 0.010257944464683533, + 0.008018634282052517, + 0.008812510408461094, + 0.023404130712151527, + 0.014488376677036285, + 0.01707194373011589, + 0.010793578810989857, + 0.013421866111457348, + 0.0129787502810359, + 0.015330230817198753, + 0.012360170483589172, + 0.013465139083564281, + 0.012574346736073494, + 0.03734084218740463, + 0.028505122289061546, + 0.045580219477415085, + 0.03557364642620087, + 0.016013218089938164, + 0.01233107503503561, + 0.01843874156475067, + 0.04590234532952309, + 0.002687236526980996, + 0.002539978828281164, + 0.004276660270988941, + 0.0027393330819904804, + 0.0021925936453044415, + 0.0019209292950108647, + 0.0018308922881260514, + 0.0044981748796999454 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.4499651789665222 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870542138814926 + ], + "opposite_logprob_tokens": [ + -3.0487053394317627 + ], + "logprob_reward": -0.04870542138814926, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870542138814926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13423354923725128, + 0.06427960842847824, + 0.049008287489414215, + 0.04421541467308998, + 0.10448205471038818, + 0.03453608974814415, + 0.15215931832790375, + 0.2944059669971466, + 0.0939091220498085, + 0.028770603239536285 + ], + "credit_map": [ + 0.013677471317350864, + 0.012486332096159458, + 0.018144864588975906, + 0.012191086076200008, + 0.01599910296499729, + 0.023558370769023895, + 0.014033869840204716, + 0.02414245344698429, + 0.0041434867307543755, + 0.005980224814265966, + 0.005786714144051075, + 0.009469994343817234, + 0.005455808714032173, + 0.009958152659237385, + 0.012497209943830967, + 0.010988020338118076, + 0.003517491975799203, + 0.004865435883402824, + 0.00565924309194088, + 0.006565300282090902, + 0.004344054032117128, + 0.011363673023879528, + 0.006083565764129162, + 0.006609522737562656, + 0.005230877548456192, + 0.004476461559534073, + 0.009873120114207268, + 0.004319746047258377, + 0.004163397941738367, + 0.007025341037660837, + 0.002678401069715619, + 0.006448065396398306, + 0.005502906162291765, + 0.006468052975833416, + 0.01679396815598011, + 0.00924932025372982, + 0.014403768815100193, + 0.0265305545181036, + 0.010410253889858723, + 0.015123231336474419, + 0.002613846678286791, + 0.00263679469935596, + 0.006477024871855974, + 0.0039810058660805225, + 0.0028537707403302193, + 0.003921033348888159, + 0.005829287692904472, + 0.006223325617611408, + 0.012791786342859268, + 0.015438029542565346, + 0.019274935126304626, + 0.01667209155857563, + 0.016315238550305367, + 0.013595424592494965, + 0.036986835300922394, + 0.02108498476445675, + 0.02548099309206009, + 0.02295377105474472, + 0.03547608107328415, + 0.04127603769302368, + 0.034159183502197266, + 0.027913741767406464, + 0.07056213170289993, + 0.03658401593565941, + 0.01700812391936779, + 0.01235177181661129, + 0.01225795317441225, + 0.012252490036189556, + 0.007047237828373909, + 0.00651954859495163, + 0.011163940653204918, + 0.015308055095374584, + 0.002774737775325775, + 0.0024885053280740976, + 0.0044190664775669575, + 0.0033652405254542828, + 0.0030928535852581263, + 0.0028474368155002594, + 0.0027115538250654936, + 0.007071207277476788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.04870542138814926 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497156083583832 + ], + "opposite_logprob_tokens": [ + -2.92997145652771 + ], + "logprob_reward": -0.05497156083583832, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497156083583832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03468363732099533, + 0.02036399580538273, + 0.01885966584086418, + 0.060369208455085754, + 0.06869335472583771, + 0.029285268858075142, + 0.18899017572402954, + 0.5462937951087952, + 0.023514725267887115, + 0.008946157991886139 + ], + "credit_map": [ + 0.0033997525461018085, + 0.00522506982088089, + 0.004214140120893717, + 0.004491262603551149, + 0.004716084338724613, + 0.004930535331368446, + 0.0035925216507166624, + 0.00411427253857255, + 0.0018902120646089315, + 0.0027165513020008802, + 0.002639797981828451, + 0.0029568432364612818, + 0.0016402163309976459, + 0.0027619486209005117, + 0.0035259986761957407, + 0.0022324284072965384, + 0.0016222117701545358, + 0.0018569568637758493, + 0.0027765969280153513, + 0.0026604405138641596, + 0.0020381801296025515, + 0.0030980422161519527, + 0.0018894204404205084, + 0.0029178177937865257, + 0.007330435793846846, + 0.007345777936279774, + 0.011349027045071125, + 0.0058470736257731915, + 0.00585116958245635, + 0.008371922187507153, + 0.005077133420854807, + 0.009196674451231956, + 0.004318078514188528, + 0.004131052643060684, + 0.01199854165315628, + 0.006514905486255884, + 0.007962950505316257, + 0.017092788591980934, + 0.006613096687942743, + 0.010061941109597683, + 0.004240117035806179, + 0.0026607539039105177, + 0.0051313708536326885, + 0.003920823335647583, + 0.003080196911469102, + 0.003341519273817539, + 0.0032989955507218838, + 0.0036114926915615797, + 0.01779930479824543, + 0.0191628597676754, + 0.026693185791373253, + 0.01802685484290123, + 0.016990048810839653, + 0.015732474625110626, + 0.049688369035720825, + 0.024897076189517975, + 0.04505690559744835, + 0.047910287976264954, + 0.06227513402700424, + 0.07856868207454681, + 0.05796842277050018, + 0.05076489597558975, + 0.13484081625938416, + 0.06890865415334702, + 0.0033605603966861963, + 0.003080895869061351, + 0.004140871111303568, + 0.002825784496963024, + 0.002282434841617942, + 0.0014137951657176018, + 0.0020634979009628296, + 0.004346884787082672, + 0.0008922306005842984, + 0.0007595253409817815, + 0.0018502711318433285, + 0.0015293951146304607, + 0.000682448036968708, + 0.0007087980047799647, + 0.0009091683314181864, + 0.0016143206739798188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.05497156083583832 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197722733020782 + ], + "opposite_logprob_tokens": [ + -4.018197536468506 + ], + "logprob_reward": -0.018197722733020782, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018197722733020782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059871986508369446, + 0.052671048790216446, + 0.04418279975652695, + 0.03812001645565033, + 0.04089599475264549, + 0.05736793205142021, + 0.15646624565124512, + 0.3384642004966736, + 0.0980159118771553, + 0.11394388228654861 + ], + "credit_map": [ + 0.005774639546871185, + 0.005512392148375511, + 0.008841289207339287, + 0.005548178683966398, + 0.006606814451515675, + 0.01108553260564804, + 0.006891930941492319, + 0.00961120706051588, + 0.004816459957510233, + 0.008022054098546505, + 0.005164972972124815, + 0.006721040699630976, + 0.006045735906809568, + 0.006857701111584902, + 0.009302019141614437, + 0.005741065368056297, + 0.003380109556019306, + 0.004044415429234505, + 0.003965878859162331, + 0.005714595317840576, + 0.003593022935092449, + 0.012107022106647491, + 0.004438479896634817, + 0.006939276587218046, + 0.00440165726467967, + 0.003751523792743683, + 0.006061959080398083, + 0.0041830893605947495, + 0.004412773530930281, + 0.006453226786106825, + 0.003317698137834668, + 0.005538087338209152, + 0.0025356640107929707, + 0.0026633073575794697, + 0.006405877415090799, + 0.003152014221996069, + 0.005843308288604021, + 0.010112624615430832, + 0.004410001449286938, + 0.0057731992565095425, + 0.00711536081507802, + 0.005489525850862265, + 0.008777654729783535, + 0.006180326454341412, + 0.005202659871429205, + 0.007801475003361702, + 0.006654334254562855, + 0.010146595537662506, + 0.020445913076400757, + 0.01686123199760914, + 0.020187629386782646, + 0.016636429354548454, + 0.01374208927154541, + 0.013180868700146675, + 0.0359339602291584, + 0.01947813108563423, + 0.028496084734797478, + 0.03026282787322998, + 0.03946084529161453, + 0.05077192559838295, + 0.03770297393202782, + 0.0338018573820591, + 0.07586230337619781, + 0.04210537299513817, + 0.018095532432198524, + 0.012971722520887852, + 0.016194630414247513, + 0.010724104009568691, + 0.008637795224785805, + 0.00807806383818388, + 0.008913695812225342, + 0.014400371350347996, + 0.010070979595184326, + 0.008561593480408192, + 0.02098996378481388, + 0.014149473048746586, + 0.010473179630935192, + 0.009793235920369625, + 0.009934937581419945, + 0.02997051551938057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.018197722733020782 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle low, tracking a tiger moving through dense, humid jungle undergrowth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the tiger's fur vibrant orange and black with visible individual hairs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the ground covered with giant ferns and fallen leaves while the tiger's paws tap the moist earth?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the tiger's amber eyes fixed with constricted pupils, locked in focus on prey beyond the camera's view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5235715508460999 + ], + "opposite_logprob_tokens": [ + -0.8985715508460999 + ], + "logprob_reward": -0.5235715508460999, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5235715508460999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0827440395951271, + 0.08175760507583618, + 0.03591849282383919, + 0.08203539997339249, + 0.026325911283493042, + 0.07301042228937149, + 0.12343772500753403, + 0.3694380521774292, + 0.10317531228065491, + 0.022157052531838417 + ], + "credit_map": [ + 0.007132621482014656, + 0.00817374512553215, + 0.012069888412952423, + 0.007874744944274426, + 0.015150620602071285, + 0.011031052097678185, + 0.011120119132101536, + 0.010191244073212147, + 0.007263411767780781, + 0.012951248325407505, + 0.007353191729635, + 0.012404810637235641, + 0.008821331895887852, + 0.011706387624144554, + 0.012377550825476646, + 0.00887967273592949, + 0.0034087547101080418, + 0.0034158399794250727, + 0.003480451414361596, + 0.0036401806864887476, + 0.0033825794234871864, + 0.008654804900288582, + 0.0035340716131031513, + 0.006401813589036465, + 0.01016557402908802, + 0.014192651957273483, + 0.013365527614951134, + 0.008305846713483334, + 0.008649639785289764, + 0.009157299995422363, + 0.006598565261811018, + 0.011600296013057232, + 0.0022216755896806717, + 0.002344269771128893, + 0.0040257954970002174, + 0.003528977045789361, + 0.0034314896911382675, + 0.0035046315751969814, + 0.004816181026399136, + 0.0024528910871595144, + 0.008561912924051285, + 0.009659522213041782, + 0.00992301944643259, + 0.008557850494980812, + 0.007409651763737202, + 0.009300666861236095, + 0.008251679129898548, + 0.01134611014276743, + 0.018623823300004005, + 0.0145785016939044, + 0.015327299013733864, + 0.015796273946762085, + 0.011121480725705624, + 0.01065280381590128, + 0.023032458499073982, + 0.014305097982287407, + 0.032597269862890244, + 0.034674860537052155, + 0.041406359523534775, + 0.05705402046442032, + 0.04680255800485611, + 0.04056007042527199, + 0.07225174456834793, + 0.044091176241636276, + 0.020123256370425224, + 0.01109374314546585, + 0.01629200205206871, + 0.011276597157120705, + 0.009500831365585327, + 0.00735366391018033, + 0.00984150730073452, + 0.01769370399415493, + 0.0017245617927983403, + 0.0018178608734160662, + 0.004328488372266293, + 0.0027062047738581896, + 0.001757307443767786, + 0.001628201105631888, + 0.0018123831832781434, + 0.006382045801728964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.5235715508460999 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_9/index_tensor([8416, 8416])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_9/index_tensor([8416, 8416])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b51a634e70acbf8257a774b2b4e7647cb6fb5ebe --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_9/index_tensor([8416, 8416])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot upward from an extremely low angle, the camera follows a crystal-clear firefly in a slow, graceful spiral as it steadily and deliberately orbits an unnaturally large flower, a fantastical hybrid of a lotus and a lily. A warm, luminous yellow-green glow emanates from the firefly's abdomen, leaving a trail of soft, dreamy light that gradually dissipates in its wake, the only dynamic source of light in the enchanted forest at night. The enormous flower it encircles has layered petals, a translucent milky white, edged with a faint pink-purple halo and coated in a layer of delicate, shimmering dewdrops. A constant, soft golden glow emanates from the center of the pistil, illuminating the inner petals. In the background is a deep, almost inky night sky, dotted with a few distant, bright stars, and framed by the blurred outlines of ancient trees. The moving light of fireflies casts shifting, soft shadows on the huge petals, which complement the static golden light of the stamens. The entire picture is shrouded in a tranquility that is almost religious, creating a quiet, dreamy atmosphere full of the sacredness of life.", + "seed": 5648, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_9/index_8416_seed_5648_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1115783616900444, + "vlm_logprob": -0.1115783616900444, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.1115783616900444, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4427732527256012 + ], + "opposite_logprob_tokens": [ + -1.0677732229232788 + ], + "logprob_reward": -0.4427732527256012, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4427732527256012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1963091343641281, + 0.07484667748212814, + 0.09226778894662857, + 0.25071898102760315, + 0.10703740268945694, + 0.04016171395778656, + 0.12303748726844788, + 0.03824413940310478, + 0.041976749897003174, + 0.03539993613958359 + ], + "credit_map": [ + 0.015710441395640373, + 0.01118059828877449, + 0.014946596696972847, + 0.0758100152015686, + 0.01039817277342081, + 0.010197915136814117, + 0.027807412669062614, + 0.03025798685848713, + 0.008302170783281326, + 0.005242681596428156, + 0.008947473019361496, + 0.012663859874010086, + 0.0059547629207372665, + 0.007494657766073942, + 0.010491692461073399, + 0.0157493744045496, + 0.016743235290050507, + 0.005719243083149195, + 0.015288794413208961, + 0.01054534874856472, + 0.0030975614208728075, + 0.00692716334015131, + 0.0074043842032551765, + 0.02654205821454525, + 0.06847521662712097, + 0.015498057939112186, + 0.018545804545283318, + 0.0779004916548729, + 0.006253642030060291, + 0.012069514021277428, + 0.027342751622200012, + 0.024633513763546944, + 0.010634406469762325, + 0.005826202221214771, + 0.029739610850811005, + 0.0288983304053545, + 0.005697230808436871, + 0.00868151243776083, + 0.008828646503388882, + 0.008731462992727757, + 0.004616227000951767, + 0.003646651515737176, + 0.004910850897431374, + 0.00766491424292326, + 0.0031336811371147633, + 0.006067605689167976, + 0.003915115725249052, + 0.00620666891336441, + 0.030314644798636436, + 0.007760137319564819, + 0.012032880447804928, + 0.03509775921702385, + 0.004290115553885698, + 0.009005448780953884, + 0.015447534620761871, + 0.009088966995477676, + 0.005305033642798662, + 0.0029872888699173927, + 0.003932802006602287, + 0.006426359061151743, + 0.002625648397952318, + 0.006237500812858343, + 0.0030879999976605177, + 0.007641506381332874, + 0.004852852318435907, + 0.010098544880747795, + 0.004619085695594549, + 0.005672272760421038, + 0.003441530279815197, + 0.004773382563143969, + 0.004538849927484989, + 0.003980233799666166, + 0.004000572487711906, + 0.007543842773884535, + 0.004592882003635168, + 0.00456482358276844, + 0.0028649454470723867, + 0.004209044389426708, + 0.004415617790073156, + 0.003208206268027425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.4427732527256012 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043119464069604874 + ], + "opposite_logprob_tokens": [ + -3.168119430541992 + ], + "logprob_reward": -0.043119464069604874, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043119464069604874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08162510395050049, + 0.06397973001003265, + 0.13502942025661469, + 0.11857553571462631, + 0.12617632746696472, + 0.04548739269375801, + 0.22018440067768097, + 0.072739377617836, + 0.07271099835634232, + 0.06349170207977295 + ], + "credit_map": [ + 0.012187318876385689, + 0.007169957272708416, + 0.008961555548012257, + 0.010872536338865757, + 0.007266203872859478, + 0.008887680247426033, + 0.011297745630145073, + 0.01498210895806551, + 0.004763515666127205, + 0.0061002676375210285, + 0.006283256225287914, + 0.012196951545774937, + 0.0049983784556388855, + 0.006245279219001532, + 0.011592285707592964, + 0.011799793690443039, + 0.013873346149921417, + 0.007877534255385399, + 0.03407910466194153, + 0.04452630504965782, + 0.004672770854085684, + 0.0060081761330366135, + 0.011164121329784393, + 0.012828055769205093, + 0.01817280240356922, + 0.0084801921620965, + 0.010443136096000671, + 0.039819952100515366, + 0.004242270719259977, + 0.009823120199143887, + 0.013083230704069138, + 0.014510827139019966, + 0.011197352781891823, + 0.011746753938496113, + 0.03947531059384346, + 0.032615505158901215, + 0.007254340220242739, + 0.006252098362892866, + 0.007027723826467991, + 0.010607251897454262, + 0.008821016177535057, + 0.003842168254777789, + 0.004606465343385935, + 0.008375823497772217, + 0.003807040164247155, + 0.0051042381674051285, + 0.0050829132087528706, + 0.005847725551575422, + 0.05943378806114197, + 0.014365144073963165, + 0.019976574927568436, + 0.06140464171767235, + 0.007326990831643343, + 0.015758197754621506, + 0.027337225154042244, + 0.014581833966076374, + 0.009357711300253868, + 0.0038841485511511564, + 0.008972524665296078, + 0.012405217625200748, + 0.0034529927652329206, + 0.010768336243927479, + 0.008424018509685993, + 0.015474425628781319, + 0.00923356506973505, + 0.01756270043551922, + 0.006619337480515242, + 0.009329195134341717, + 0.006494938395917416, + 0.008956605568528175, + 0.007197279948741198, + 0.007317374460399151, + 0.007031411863863468, + 0.013756903819739819, + 0.010783018544316292, + 0.01054302603006363, + 0.004286172334104776, + 0.005494746379554272, + 0.006746650207787752, + 0.004849766381084919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.043119464069604874 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005967419594526291 + ], + "opposite_logprob_tokens": [ + -5.130967617034912 + ], + "logprob_reward": -0.005967419594526291, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005967419594526291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14880217611789703, + 0.07921404391527176, + 0.09673285484313965, + 0.08273573964834213, + 0.05268307775259018, + 0.06586898863315582, + 0.23103812336921692, + 0.08193705230951309, + 0.08979147672653198, + 0.07119642943143845 + ], + "credit_map": [ + 0.013450334779918194, + 0.010519147850573063, + 0.012254567816853523, + 0.04449401795864105, + 0.007772078737616539, + 0.011624020524322987, + 0.021272217854857445, + 0.027415785938501358, + 0.006273647770285606, + 0.006567821372300386, + 0.00762847438454628, + 0.01521983090788126, + 0.0061418479308485985, + 0.008301571011543274, + 0.014561863616108894, + 0.014518984593451023, + 0.01523812860250473, + 0.005727154668420553, + 0.021603696048259735, + 0.019535988569259644, + 0.004430315922945738, + 0.00546960299834609, + 0.006959767080843449, + 0.01776820607483387, + 0.022202033549547195, + 0.005975326988846064, + 0.007499407511204481, + 0.01814744807779789, + 0.004373110365122557, + 0.005639080423861742, + 0.008135633543133736, + 0.010763696394860744, + 0.005990870762616396, + 0.007512149401009083, + 0.011412354186177254, + 0.009927180595695972, + 0.0052231233566999435, + 0.0038290759548544884, + 0.004495937842875719, + 0.004292384721338749, + 0.010515456087887287, + 0.006988529581576586, + 0.005719207227230072, + 0.012473505921661854, + 0.0065284473821520805, + 0.007124816998839378, + 0.007484353147447109, + 0.009034669026732445, + 0.05964284390211105, + 0.014636071398854256, + 0.02115144021809101, + 0.06518435478210449, + 0.008085199631750584, + 0.016383515670895576, + 0.030640937387943268, + 0.015313739888370037, + 0.009869449771940708, + 0.005120063200592995, + 0.011738951317965984, + 0.01512821763753891, + 0.004502154421061277, + 0.011117205955088139, + 0.009191068820655346, + 0.015269939787685871, + 0.011408400721848011, + 0.023477734997868538, + 0.008701507933437824, + 0.011314758099615574, + 0.008284085430204868, + 0.010016566142439842, + 0.008358427323400974, + 0.008230001665651798, + 0.007498930208384991, + 0.014921455644071102, + 0.012653247453272343, + 0.01205406803637743, + 0.0046989573165774345, + 0.006003414746373892, + 0.007943891920149326, + 0.005422470159828663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.005967419594526291 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011080872267484665 + ], + "opposite_logprob_tokens": [ + -4.511080741882324 + ], + "logprob_reward": -0.011080872267484665, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011080872267484665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1337616890668869, + 0.07243778556585312, + 0.05377582088112831, + 0.19704611599445343, + 0.07453929632902145, + 0.07750757038593292, + 0.1874072104692459, + 0.07815714925527573, + 0.04072443023324013, + 0.08464299142360687 + ], + "credit_map": [ + 0.019106311723589897, + 0.008066918700933456, + 0.008571701124310493, + 0.03463807329535484, + 0.013850393705070019, + 0.009249551221728325, + 0.01713825762271881, + 0.02314048260450363, + 0.008263195864856243, + 0.007463338319212198, + 0.007413980085402727, + 0.0109790600836277, + 0.0058037699200212955, + 0.0065317777916789055, + 0.010850882157683372, + 0.015131774358451366, + 0.006866996176540852, + 0.003963802475482225, + 0.013427522964775562, + 0.0060850814916193485, + 0.0030500448774546385, + 0.005124574061483145, + 0.004648991394788027, + 0.01060880720615387, + 0.04770825803279877, + 0.016765033826231956, + 0.014101902022957802, + 0.051155466586351395, + 0.014210192486643791, + 0.012117789126932621, + 0.02151632308959961, + 0.01947113871574402, + 0.01685529574751854, + 0.0055216336622834206, + 0.010506059043109417, + 0.007832399569451809, + 0.0039019491523504257, + 0.0056162322871387005, + 0.005914283450692892, + 0.01839143969118595, + 0.009488922543823719, + 0.009046582505106926, + 0.008075111545622349, + 0.01263368222862482, + 0.007179340813308954, + 0.00751479109749198, + 0.008482872508466244, + 0.015086270868778229, + 0.04760788753628731, + 0.012014219537377357, + 0.018130680546164513, + 0.053148139268159866, + 0.006043919362127781, + 0.013760167174041271, + 0.023242643103003502, + 0.013459552079439163, + 0.010298656299710274, + 0.00705354567617178, + 0.009641552343964577, + 0.013550813309848309, + 0.005107348319143057, + 0.011449300684034824, + 0.007792980410158634, + 0.013262951746582985, + 0.00461367703974247, + 0.006447584368288517, + 0.006062841508537531, + 0.005219966173171997, + 0.0028983107767999172, + 0.005236170254647732, + 0.005420858971774578, + 0.004825018346309662, + 0.012197494506835938, + 0.01268703956156969, + 0.016567472368478775, + 0.013384087942540646, + 0.005049440544098616, + 0.009503561072051525, + 0.009250015951693058, + 0.0060038804076612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.011080872267484665 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495079979300499 + ], + "opposite_logprob_tokens": [ + -2.929950714111328 + ], + "logprob_reward": -0.05495079979300499, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495079979300499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.155080184340477, + 0.1229291558265686, + 0.05945077911019325, + 0.2306251972913742, + 0.09566125273704529, + 0.033712439239025116, + 0.19022321701049805, + 0.03609774261713028, + 0.038138497620821, + 0.038081519305706024 + ], + "credit_map": [ + 0.013086550869047642, + 0.009419660083949566, + 0.011748198419809341, + 0.055514346808195114, + 0.007538982201367617, + 0.010008394718170166, + 0.020340751856565475, + 0.027423301711678505, + 0.014694775454699993, + 0.009991973638534546, + 0.016027506440877914, + 0.020359797403216362, + 0.008992505259811878, + 0.009944788180291653, + 0.02039160206913948, + 0.022526198998093605, + 0.008975849486887455, + 0.004612240474671125, + 0.007948608137667179, + 0.00933013204485178, + 0.003088599070906639, + 0.005671885330229998, + 0.0042528738267719746, + 0.01557058934122324, + 0.05924902483820915, + 0.014054950326681137, + 0.016198256984353065, + 0.07742340862751007, + 0.00548534095287323, + 0.012164621613919735, + 0.026439771056175232, + 0.01960982382297516, + 0.012064511887729168, + 0.0073762559331953526, + 0.02021276392042637, + 0.009234746918082237, + 0.006029966752976179, + 0.009722466580569744, + 0.00911774393171072, + 0.021902795881032944, + 0.003447202267125249, + 0.00362350232899189, + 0.005488888826221228, + 0.004085713066160679, + 0.003507815068587661, + 0.00327071500942111, + 0.004234150052070618, + 0.006054453551769257, + 0.04195394366979599, + 0.01166812889277935, + 0.019107703119516373, + 0.055396951735019684, + 0.006535228341817856, + 0.014881309121847153, + 0.02515423484146595, + 0.015525706112384796, + 0.004518114496022463, + 0.0031870114617049694, + 0.005860242992639542, + 0.006874265149235725, + 0.002756115747615695, + 0.004250788129866123, + 0.0034179522190243006, + 0.005233253352344036, + 0.004440410062670708, + 0.0060958536341786385, + 0.004294215701520443, + 0.004839055240154266, + 0.0028828971553593874, + 0.006395557429641485, + 0.00428742216899991, + 0.004903085995465517, + 0.0042473310604691505, + 0.00726404320448637, + 0.005029955878853798, + 0.005583547055721283, + 0.0023993575014173985, + 0.005345056299120188, + 0.004701039753854275, + 0.0035111894831061363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.05495079979300499 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot upward from an extremely low angle, the camera follows a crystal-clear firefly in a slow, graceful spiral as it steadily and deliberately orbits an unnaturally large flower, a fantastical hybrid of a lotus and a lily. A warm, luminous yellow-green glow emanates from the firefly's abdomen, leaving a trail of soft, dreamy light that gradually dissipates in its wake, the only dynamic source of light in the enchanted forest at night. The enormous flower it encircles has layered petals, a translucent milky white, edged with a faint pink-purple halo and coated in a layer of delicate, shimmering dewdrops. A constant, soft golden glow emanates from the center of the pistil, illuminating the inner petals. In the background is a deep, almost inky night sky, dotted with a few distant, bright stars, and framed by the blurred outlines of ancient trees. The moving light of fireflies casts shifting, soft shadows on the huge petals, which complement the static golden light of the stamens. The entire picture is shrouded in a tranquility that is almost religious, creating a quiet, dreamy atmosphere full of the sacredness of life.", + "seed": 5649, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_9/index_8416_seed_5649_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1488490838557482, + "vlm_logprob": -0.1488490838557482, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.1488490838557482, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5861272215843201 + ], + "opposite_logprob_tokens": [ + -0.8361272215843201 + ], + "logprob_reward": -0.5861272215843201, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5861272215843201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09682812541723251, + 0.1017925962805748, + 0.05159372836351395, + 0.191531240940094, + 0.16467972099781036, + 0.06365251541137695, + 0.061233922839164734, + 0.0627737045288086, + 0.10980682820081711, + 0.09610766172409058 + ], + "credit_map": [ + 0.012672421522438526, + 0.013250280171632767, + 0.011592789553105831, + 0.014672357589006424, + 0.008615047670900822, + 0.008122292347252369, + 0.014008913189172745, + 0.01389401126652956, + 0.009117992594838142, + 0.00865041185170412, + 0.007044930011034012, + 0.018393369391560555, + 0.01406022533774376, + 0.0075921169482171535, + 0.01582220010459423, + 0.02111133374273777, + 0.0046084122732281685, + 0.003985988441854715, + 0.004756443668156862, + 0.0067291478626430035, + 0.004530774895101786, + 0.008687537163496017, + 0.007425310555845499, + 0.010870103724300861, + 0.029208356514573097, + 0.015249883756041527, + 0.0146950026974082, + 0.02235499583184719, + 0.018874600529670715, + 0.0317012295126915, + 0.024232162162661552, + 0.03521498665213585, + 0.012986907735466957, + 0.005861496087163687, + 0.00995628535747528, + 0.008676331490278244, + 0.009446383453905582, + 0.033610355108976364, + 0.06316957622766495, + 0.020972372964024544, + 0.0075649782083928585, + 0.0057273670099675655, + 0.008126635104417801, + 0.006006023846566677, + 0.008529381826519966, + 0.007791175041347742, + 0.009795483201742172, + 0.01011146605014801, + 0.004055575933307409, + 0.006220967974513769, + 0.0061307791620492935, + 0.010485820472240448, + 0.006434516049921513, + 0.008797586895525455, + 0.011116963811218739, + 0.00799170508980751, + 0.005082422401756048, + 0.006164028309285641, + 0.008923662826418877, + 0.008666237816214561, + 0.007796882651746273, + 0.006768296472728252, + 0.010287506505846977, + 0.009084664285182953, + 0.008251996710896492, + 0.007859306409955025, + 0.00522095151245594, + 0.006718088407069445, + 0.00800660066306591, + 0.03542102500796318, + 0.008357521146535873, + 0.029971323907375336, + 0.005815853364765644, + 0.0048726387321949005, + 0.008015754632651806, + 0.008732552640140057, + 0.007165808696299791, + 0.026735248044133186, + 0.01080289389938116, + 0.02396690472960472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.5861272215843201 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500112101435661 + ], + "opposite_logprob_tokens": [ + -2.9300010204315186 + ], + "logprob_reward": -0.05500112101435661, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500112101435661, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10342647135257721, + 0.24787376821041107, + 0.08165496587753296, + 0.1031511127948761, + 0.05659455806016922, + 0.06525437533855438, + 0.0785517543554306, + 0.09374334663152695, + 0.09394422918558121, + 0.07580534368753433 + ], + "credit_map": [ + 0.014245322905480862, + 0.015588609501719475, + 0.01153154019266367, + 0.013196991756558418, + 0.011726043187081814, + 0.008460083976387978, + 0.012507591396570206, + 0.016170291230082512, + 0.0239469762891531, + 0.020307103171944618, + 0.01953241042792797, + 0.054715048521757126, + 0.028966808691620827, + 0.01867862045764923, + 0.03291994333267212, + 0.04880683496594429, + 0.0067972345277667046, + 0.007046288810670376, + 0.006383113097399473, + 0.010453850962221622, + 0.013539480976760387, + 0.009134751744568348, + 0.015371153131127357, + 0.012929096817970276, + 0.017240943387150764, + 0.00825814250856638, + 0.009032861329615116, + 0.011717389337718487, + 0.007287986110895872, + 0.015012172982096672, + 0.01315593533217907, + 0.021445687860250473, + 0.004593307618051767, + 0.0036834189668297768, + 0.0041132159531116486, + 0.004075031727552414, + 0.005328402388840914, + 0.008533170446753502, + 0.015963848680257797, + 0.010304166004061699, + 0.008276933804154396, + 0.00589324813336134, + 0.00485235545784235, + 0.009559083729982376, + 0.007945085875689983, + 0.0064524016343057156, + 0.01095557026565075, + 0.01131969504058361, + 0.006154661998152733, + 0.009484566748142242, + 0.006124319974333048, + 0.01607404090464115, + 0.007990922778844833, + 0.009191203862428665, + 0.014971964061260223, + 0.008560075424611568, + 0.013034890405833721, + 0.010182315483689308, + 0.010173403657972813, + 0.009769353084266186, + 0.007523150648921728, + 0.01383277215063572, + 0.012324029579758644, + 0.016903430223464966, + 0.01209647674113512, + 0.011624183505773544, + 0.007107060868293047, + 0.008855193853378296, + 0.011137538589537144, + 0.019051186740398407, + 0.012726780958473682, + 0.011345804668962955, + 0.007578645832836628, + 0.00663729989901185, + 0.008055784739553928, + 0.012743317522108555, + 0.009044005535542965, + 0.010044422931969166, + 0.010785294696688652, + 0.010916574858129025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.05500112101435661 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009799469262361526 + ], + "opposite_logprob_tokens": [ + -4.634799480438232 + ], + "logprob_reward": -0.009799469262361526, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009799469262361526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07691514492034912, + 0.21553777158260345, + 0.05085822194814682, + 0.1424267590045929, + 0.08698149025440216, + 0.08045791834592819, + 0.0657903403043747, + 0.08689440786838531, + 0.10614454001188278, + 0.0879933089017868 + ], + "credit_map": [ + 0.008075634017586708, + 0.009407768957316875, + 0.009508337825536728, + 0.008929068222641945, + 0.0077241212129592896, + 0.004672793205827475, + 0.008509433828294277, + 0.020087994635105133, + 0.02116767317056656, + 0.018795950338244438, + 0.01532828714698553, + 0.04394005239009857, + 0.027814893051981926, + 0.01656102016568184, + 0.02944941259920597, + 0.042480506002902985, + 0.004094898700714111, + 0.004896705970168114, + 0.004423308651894331, + 0.007514906581491232, + 0.005706296302378178, + 0.005956051871180534, + 0.009331955574452877, + 0.008934104815125465, + 0.02288726530969143, + 0.012043518014252186, + 0.012174665927886963, + 0.01570345088839531, + 0.01086777076125145, + 0.019721150398254395, + 0.017806297168135643, + 0.03122265450656414, + 0.006662500556558371, + 0.00406348193064332, + 0.007040057331323624, + 0.005308609921485186, + 0.005527794361114502, + 0.014176350086927414, + 0.03132092207670212, + 0.01288178376853466, + 0.007151131518185139, + 0.009502868168056011, + 0.010626170784235, + 0.007746735122054815, + 0.014025377109646797, + 0.009480113163590431, + 0.011813657358288765, + 0.010111880488693714, + 0.004551668651401997, + 0.006615588907152414, + 0.00449922913685441, + 0.013548869639635086, + 0.007288024295121431, + 0.008159089833498001, + 0.013089455664157867, + 0.008038422092795372, + 0.012683265842497349, + 0.009645266458392143, + 0.009079379960894585, + 0.009988817386329174, + 0.00648135133087635, + 0.012886561453342438, + 0.011716742068529129, + 0.014413026161491871, + 0.009492029435932636, + 0.01041044108569622, + 0.007508392911404371, + 0.008094643242657185, + 0.010398996993899345, + 0.018787221983075142, + 0.012278378009796143, + 0.029174447059631348, + 0.007233313750475645, + 0.007652525324374437, + 0.00990581139922142, + 0.01343103963881731, + 0.009387106634676456, + 0.012741941958665848, + 0.013589882291853428, + 0.014051700010895729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.009799469262361526 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329634316265583 + ], + "opposite_logprob_tokens": [ + -3.773296356201172 + ], + "logprob_reward": -0.02329634316265583, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329634316265583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06411904841661453, + 0.08897141367197037, + 0.060232918709516525, + 0.0576445572078228, + 0.08590243011713028, + 0.09842012822628021, + 0.10005466639995575, + 0.12227904051542282, + 0.09492651373147964, + 0.22744925320148468 + ], + "credit_map": [ + 0.006488611921668053, + 0.008594938553869724, + 0.005999205633997917, + 0.010210844688117504, + 0.005664009600877762, + 0.00712725380435586, + 0.00929771363735199, + 0.010736473836004734, + 0.01258502621203661, + 0.007072869688272476, + 0.015618031844496727, + 0.015570240095257759, + 0.008900975808501244, + 0.009444277733564377, + 0.011988257989287376, + 0.007791746407747269, + 0.005227346438914537, + 0.005947619676589966, + 0.005728127900511026, + 0.01192962285131216, + 0.008464711718261242, + 0.00563410297036171, + 0.009309396147727966, + 0.007991996593773365, + 0.00893916841596365, + 0.005309196189045906, + 0.0054642208851873875, + 0.0074350424110889435, + 0.0044692265801131725, + 0.006926547270268202, + 0.007423516362905502, + 0.011677645146846771, + 0.006139729171991348, + 0.004575433675199747, + 0.007063364144414663, + 0.004690483212471008, + 0.005031770095229149, + 0.014271118678152561, + 0.031656283885240555, + 0.01247425377368927, + 0.009972978383302689, + 0.009416842833161354, + 0.013213369064033031, + 0.013908122666180134, + 0.012944920919835567, + 0.009284580126404762, + 0.016580475494265556, + 0.013098855502903461, + 0.007096187211573124, + 0.009006083942949772, + 0.010990442708134651, + 0.017407968640327454, + 0.010251357220113277, + 0.012073464691638947, + 0.01901901327073574, + 0.014210163615643978, + 0.01454886607825756, + 0.015185951255261898, + 0.017766619101166725, + 0.01558051910251379, + 0.01267626415938139, + 0.0134386345744133, + 0.015245135873556137, + 0.017837058752775192, + 0.007903484627604485, + 0.010044336318969727, + 0.00601049093529582, + 0.006978339049965143, + 0.01041845977306366, + 0.027101319283246994, + 0.007931556552648544, + 0.018538538366556168, + 0.014232073910534382, + 0.016460992395877838, + 0.01860673911869526, + 0.02322547137737274, + 0.021758126094937325, + 0.05174654349684715, + 0.03284165635704994, + 0.0485776849091053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.02329634316265583 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002126425504684 + ], + "opposite_logprob_tokens": [ + -2.695021152496338 + ], + "logprob_reward": -0.07002126425504684, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002126425504684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09034228324890137, + 0.05510400980710983, + 0.06002464145421982, + 0.07020968198776245, + 0.0932302251458168, + 0.06984204798936844, + 0.11005064845085144, + 0.04368232563138008, + 0.3009756803512573, + 0.10653845220804214 + ], + "credit_map": [ + 0.012076224200427532, + 0.011551848612725735, + 0.008085869252681732, + 0.01348742377012968, + 0.008165652863681316, + 0.006794273387640715, + 0.01101000141352415, + 0.019170990213751793, + 0.00715956324711442, + 0.004860046319663525, + 0.005481485743075609, + 0.01204938068985939, + 0.006470731925219297, + 0.004004158545285463, + 0.008386757224798203, + 0.0066918861120939255, + 0.006490900181233883, + 0.005689726211130619, + 0.007310955785214901, + 0.006888415198773146, + 0.0063180928118526936, + 0.009517272002995014, + 0.008468103595077991, + 0.009341173805296421, + 0.008077376522123814, + 0.00513096060603857, + 0.006780450232326984, + 0.008761596865952015, + 0.00635868776589632, + 0.0095338374376297, + 0.009392838925123215, + 0.016173940151929855, + 0.006216208450496197, + 0.0053371284157037735, + 0.008179266937077045, + 0.005998009815812111, + 0.007512531243264675, + 0.013962998986244202, + 0.030695626512169838, + 0.015328453853726387, + 0.008330082520842552, + 0.006340311840176582, + 0.007278739009052515, + 0.007103580515831709, + 0.009936383925378323, + 0.00815115962177515, + 0.010686289519071579, + 0.012015501037240028, + 0.006954571232199669, + 0.009602692909538746, + 0.010473552159965038, + 0.01825007237493992, + 0.011001615785062313, + 0.016408808529376984, + 0.020178301259875298, + 0.017181044444441795, + 0.004305463749915361, + 0.0056751989759504795, + 0.005874096415936947, + 0.005769675597548485, + 0.00495136296376586, + 0.004549234639853239, + 0.006808919366449118, + 0.005748376715928316, + 0.02137926034629345, + 0.016841717064380646, + 0.012327727861702442, + 0.01468905434012413, + 0.023184465244412422, + 0.10036919265985489, + 0.026183169335126877, + 0.08600107580423355, + 0.010392611846327782, + 0.008879434317350388, + 0.01047003548592329, + 0.009500346146523952, + 0.013822343200445175, + 0.022975971922278404, + 0.010363462381064892, + 0.02013424225151539 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.07002126425504684 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot upward from an extremely low angle, the camera follows a crystal-clear firefly in a slow, graceful spiral as it steadily and deliberately orbits an unnaturally large flower, a fantastical hybrid of a lotus and a lily. A warm, luminous yellow-green glow emanates from the firefly's abdomen, leaving a trail of soft, dreamy light that gradually dissipates in its wake, the only dynamic source of light in the enchanted forest at night. The enormous flower it encircles has layered petals, a translucent milky white, edged with a faint pink-purple halo and coated in a layer of delicate, shimmering dewdrops. A constant, soft golden glow emanates from the center of the pistil, illuminating the inner petals. In the background is a deep, almost inky night sky, dotted with a few distant, bright stars, and framed by the blurred outlines of ancient trees. The moving light of fireflies casts shifting, soft shadows on the huge petals, which complement the static golden light of the stamens. The entire picture is shrouded in a tranquility that is almost religious, creating a quiet, dreamy atmosphere full of the sacredness of life.", + "seed": 5650, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_9/index_8416_seed_5650_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1023440974764526, + "vlm_logprob": -0.1023440974764526, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.1023440974764526, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.44904080033302307 + ], + "opposite_logprob_tokens": [ + -1.0740407705307007 + ], + "logprob_reward": -0.44904080033302307, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.44904080033302307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11784356832504272, + 0.028780728578567505, + 0.12525968253612518, + 0.03388284891843796, + 0.24527379870414734, + 0.025613268837332726, + 0.0538526214659214, + 0.25396257638931274, + 0.09296702593564987, + 0.02256392128765583 + ], + "credit_map": [ + 0.026511110365390778, + 0.041265763342380524, + 0.008399619720876217, + 0.006514513865113258, + 0.00768978102132678, + 0.008435213938355446, + 0.010011788457632065, + 0.00901578739285469, + 0.004477931652218103, + 0.005399772897362709, + 0.002515422413125634, + 0.003229344030842185, + 0.0022319371346384287, + 0.002315567573532462, + 0.005279665347188711, + 0.003331088926643133, + 0.025721032172441483, + 0.03458648547530174, + 0.016479231417179108, + 0.013263779692351818, + 0.009882672689855099, + 0.00810274388641119, + 0.009504497051239014, + 0.0077192396856844425, + 0.006159077398478985, + 0.006659313570708036, + 0.004218531306833029, + 0.002824281807988882, + 0.0032701988238841295, + 0.003197838319465518, + 0.004047426860779524, + 0.0035061798989772797, + 0.04934290796518326, + 0.04134548455476761, + 0.022043384611606598, + 0.0173550546169281, + 0.01796363666653633, + 0.025905178859829903, + 0.021740438416600227, + 0.04957771673798561, + 0.006075698416680098, + 0.006018227431923151, + 0.0025624551344662905, + 0.0013352051610127091, + 0.0033671194687485695, + 0.001881826319731772, + 0.0026489486917853355, + 0.001723788445815444, + 0.009404169395565987, + 0.008289789780974388, + 0.007798183709383011, + 0.006249960511922836, + 0.006777927279472351, + 0.005586331710219383, + 0.005208291113376617, + 0.004537965636700392, + 0.09749995917081833, + 0.1055263802409172, + 0.00557737285271287, + 0.003446324495598674, + 0.014408369548618793, + 0.01682182401418686, + 0.006102333776652813, + 0.004579992964863777, + 0.02127784676849842, + 0.03870503231883049, + 0.0024002413265407085, + 0.001723134540952742, + 0.006431167013943195, + 0.018692178651690483, + 0.0018201905768364668, + 0.0019172338070347905, + 0.003733173944056034, + 0.004441424738615751, + 0.0018244738457724452, + 0.0015106943901628256, + 0.004932217765599489, + 0.004098345525562763, + 0.0010116728954017162, + 0.0010119181824848056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.44904080033302307 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006770882289856672 + ], + "opposite_logprob_tokens": [ + -5.006771087646484 + ], + "logprob_reward": -0.006770882289856672, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006770882289856672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07126370817422867, + 0.018115002661943436, + 0.06966135650873184, + 0.01743844896554947, + 0.03839550167322159, + 0.01158193126320839, + 0.028406428173184395, + 0.02338988706469536, + 0.6948486566543579, + 0.026899011805653572 + ], + "credit_map": [ + 0.01375966053456068, + 0.025579165667295456, + 0.006235252600163221, + 0.004538741894066334, + 0.0051131159998476505, + 0.004194288514554501, + 0.006264135241508484, + 0.00557934632524848, + 0.005136795341968536, + 0.0037671898026019335, + 0.0014576445100829005, + 0.0013350266963243484, + 0.001377985579892993, + 0.001185080735012889, + 0.0022647082805633545, + 0.0015905732288956642, + 0.014475935138761997, + 0.019496366381645203, + 0.009542634710669518, + 0.006854969076812267, + 0.005063916556537151, + 0.004464120604097843, + 0.004901722073554993, + 0.004861688707023859, + 0.0031230009626597166, + 0.003359849564731121, + 0.002180598909035325, + 0.0015239366330206394, + 0.0015894317766651511, + 0.0015553705161437392, + 0.0021151769906282425, + 0.0019910838454961777, + 0.006815810687839985, + 0.006462159566581249, + 0.003183259628713131, + 0.0052345143631100655, + 0.0021202571224421263, + 0.0042170574888587, + 0.0036681832280009985, + 0.006694258656352758, + 0.003005440579727292, + 0.002201326657086611, + 0.001914405613206327, + 0.000645255611743778, + 0.0009403001749888062, + 0.0010936945909634233, + 0.001057491172105074, + 0.0007240175036713481, + 0.0043698870576918125, + 0.003971021622419357, + 0.0038897553458809853, + 0.002107282169163227, + 0.006711653433740139, + 0.0032184580340981483, + 0.0022534101735800505, + 0.0018849606858566403, + 0.00459632370620966, + 0.004168386105448008, + 0.0020537693053483963, + 0.0012501139426603913, + 0.005127217620611191, + 0.00354947941377759, + 0.001390262390486896, + 0.001254336442798376, + 0.172890767455101, + 0.2729840874671936, + 0.018340209499001503, + 0.011448634788393974, + 0.07862398773431778, + 0.1138664036989212, + 0.013003971427679062, + 0.01369057409465313, + 0.0035558543168008327, + 0.006011126097291708, + 0.0014311638660728931, + 0.001561964163556695, + 0.006823597475886345, + 0.0047139269299805164, + 0.0012426120229065418, + 0.0015587660018354654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.006770882289856672 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005270752124488354 + ], + "opposite_logprob_tokens": [ + -5.255270957946777 + ], + "logprob_reward": -0.005270752124488354, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005270752124488354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2950705289840698, + 0.031544968485832214, + 0.07174186408519745, + 0.032006457448005676, + 0.042754076421260834, + 0.0213116817176342, + 0.020766204223036766, + 0.13856260478496552, + 0.2973216772079468, + 0.048919979482889175 + ], + "credit_map": [ + 0.06453859806060791, + 0.12661591172218323, + 0.02079874277114868, + 0.014667020179331303, + 0.022063741460442543, + 0.017269203439354897, + 0.01596287451684475, + 0.01315445825457573, + 0.0051574911922216415, + 0.006490266416221857, + 0.0031367253977805376, + 0.0026706841308623552, + 0.002722069388255477, + 0.0029527239967137575, + 0.00411297008395195, + 0.004302038345485926, + 0.012500057928264141, + 0.009965942241251469, + 0.0070655164308846, + 0.008081493899226189, + 0.0051203081384301186, + 0.0085985092446208, + 0.010187741369009018, + 0.010222292505204678, + 0.005649446044117212, + 0.007037670351564884, + 0.006117027252912521, + 0.0037077562883496284, + 0.00253550847992301, + 0.0018658334156498313, + 0.00308430683799088, + 0.002008910523727536, + 0.00792031642049551, + 0.007618311792612076, + 0.003895323956385255, + 0.004611841402947903, + 0.0025785486213862896, + 0.0048571256920695305, + 0.003783523105084896, + 0.007489085663110018, + 0.004361334256827831, + 0.00392255000770092, + 0.004255441017448902, + 0.0013073133304715157, + 0.002051324350759387, + 0.001831528265029192, + 0.0023882121313363314, + 0.001193979405798018, + 0.003659832989796996, + 0.002921616891399026, + 0.002979560988023877, + 0.0020431012380868196, + 0.002988163847476244, + 0.0017330202972516418, + 0.0029669031500816345, + 0.001474004820920527, + 0.046525295823812485, + 0.05022820830345154, + 0.003405034774914384, + 0.004154505208134651, + 0.013202660717070103, + 0.012723403051495552, + 0.004675289615988731, + 0.003648206125944853, + 0.06809588521718979, + 0.11902449280023575, + 0.0068336487747728825, + 0.005209683440625668, + 0.030110027641057968, + 0.05641287937760353, + 0.005492978263646364, + 0.006142090540379286, + 0.0055530075915157795, + 0.010729003697633743, + 0.0028668141458183527, + 0.002495039254426956, + 0.014183683320879936, + 0.008491341955959797, + 0.002342972904443741, + 0.0022581126540899277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.005270752124488354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544653378427029 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544653378427029, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544653378427029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2045736163854599, + 0.117658331990242, + 0.07451178878545761, + 0.041880398988723755, + 0.11262188106775284, + 0.03176002949476242, + 0.029860317707061768, + 0.25667256116867065, + 0.10541451722383499, + 0.025046562775969505 + ], + "credit_map": [ + 0.047807082533836365, + 0.0866713896393776, + 0.011799448169767857, + 0.009447449818253517, + 0.015089629217982292, + 0.012257466092705727, + 0.012613140046596527, + 0.00888801459223032, + 0.023994075134396553, + 0.03392272815108299, + 0.010928130708634853, + 0.00888347253203392, + 0.009681343100965023, + 0.007878050208091736, + 0.011487963609397411, + 0.010882565751671791, + 0.011578778736293316, + 0.01152735110372305, + 0.007537050172686577, + 0.008366278372704983, + 0.005097211338579655, + 0.008350816555321217, + 0.011820933781564236, + 0.010233367793262005, + 0.007048218976706266, + 0.006119719706475735, + 0.004129428416490555, + 0.003713709069415927, + 0.0028733527287840843, + 0.004526274278759956, + 0.00498891482129693, + 0.008480781689286232, + 0.02260732464492321, + 0.022097120061516762, + 0.009334392845630646, + 0.007408175151795149, + 0.007205433677881956, + 0.011946161277592182, + 0.009707578457891941, + 0.022315699607133865, + 0.006476933602243662, + 0.007367090322077274, + 0.002404776867479086, + 0.0015754017513245344, + 0.005600936245173216, + 0.0027095552068203688, + 0.0034089654218405485, + 0.0022163689136505127, + 0.004060337319970131, + 0.004203918855637312, + 0.004294916056096554, + 0.0029697497375309467, + 0.005643922835588455, + 0.002659894758835435, + 0.0036209954414516687, + 0.002406580373644829, + 0.095554418861866, + 0.10518881678581238, + 0.007587615866214037, + 0.005119749810546637, + 0.012231938540935516, + 0.01862543448805809, + 0.006113623734563589, + 0.006250960752367973, + 0.014719447121024132, + 0.019933175295591354, + 0.00573427090421319, + 0.007692377083003521, + 0.028500152751803398, + 0.018736042082309723, + 0.005478329490870237, + 0.004620728548616171, + 0.0034686606377363205, + 0.005747262388467789, + 0.0015060406876727939, + 0.0013830045936629176, + 0.00596120860427618, + 0.004519286565482616, + 0.001179703394882381, + 0.0012813934590667486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.012544653378427029 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809339925646782 + ], + "opposite_logprob_tokens": [ + -3.288093328475952 + ], + "logprob_reward": -0.03809339925646782, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809339925646782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.582571804523468, + 0.11444340646266937, + 0.11952012777328491, + 0.024487460032105446, + 0.013849695213139057, + 0.02773604355752468, + 0.006781233474612236, + 0.03410514444112778, + 0.04539189487695694, + 0.03111312724649906 + ], + "credit_map": [ + 0.12891055643558502, + 0.24653048813343048, + 0.0416811928153038, + 0.02921442501246929, + 0.042523887008428574, + 0.03513338044285774, + 0.03329920768737793, + 0.025278717279434204, + 0.02550472319126129, + 0.034854352474212646, + 0.011201021261513233, + 0.008427245542407036, + 0.010168921202421188, + 0.007053901441395283, + 0.009224716573953629, + 0.008008528500795364, + 0.023889319971203804, + 0.03289112448692322, + 0.01719352975487709, + 0.012461605481803417, + 0.00916665606200695, + 0.00791481789201498, + 0.008109751157462597, + 0.007893318310379982, + 0.004251217935234308, + 0.004294108133763075, + 0.0017384489765390754, + 0.0017301560146734118, + 0.0017980580450966954, + 0.0028873595874756575, + 0.002742788754403591, + 0.005045323166996241, + 0.002823473187163472, + 0.0031225814018398523, + 0.0014246037462726235, + 0.0008790440042503178, + 0.0009510017116554081, + 0.0013210210017859936, + 0.0009616798488423228, + 0.0023662904277443886, + 0.0054435329511761665, + 0.00511139165610075, + 0.002802690025418997, + 0.0019115626346319914, + 0.003821760416030884, + 0.002516272012144327, + 0.003704833798110485, + 0.0024239993654191494, + 0.000999767566099763, + 0.0008878559456206858, + 0.001024963683448732, + 0.000443280121544376, + 0.0015748682199046016, + 0.0005422842805273831, + 0.0008636012789793313, + 0.00044461232027970254, + 0.011899622157216072, + 0.01313213724642992, + 0.0012705023400485516, + 0.0006297677755355835, + 0.0029216266702860594, + 0.0024562571197748184, + 0.000812188140116632, + 0.0009830446215346456, + 0.010339992120862007, + 0.015259726904332638, + 0.0019597255159169436, + 0.001082573551684618, + 0.007726857904344797, + 0.0066293589770793915, + 0.0012150402180850506, + 0.0011786166578531265, + 0.0038747028447687626, + 0.0060517191886901855, + 0.001877804403193295, + 0.0012801119592040777, + 0.009724820032715797, + 0.00579924276098609, + 0.0013341641752049327, + 0.0011705622309818864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.03809339925646782 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot upward from an extremely low angle, the camera follows a crystal-clear firefly in a slow, graceful spiral as it steadily and deliberately orbits an unnaturally large flower, a fantastical hybrid of a lotus and a lily. A warm, luminous yellow-green glow emanates from the firefly's abdomen, leaving a trail of soft, dreamy light that gradually dissipates in its wake, the only dynamic source of light in the enchanted forest at night. The enormous flower it encircles has layered petals, a translucent milky white, edged with a faint pink-purple halo and coated in a layer of delicate, shimmering dewdrops. A constant, soft golden glow emanates from the center of the pistil, illuminating the inner petals. In the background is a deep, almost inky night sky, dotted with a few distant, bright stars, and framed by the blurred outlines of ancient trees. The moving light of fireflies casts shifting, soft shadows on the huge petals, which complement the static golden light of the stamens. The entire picture is shrouded in a tranquility that is almost religious, creating a quiet, dreamy atmosphere full of the sacredness of life.", + "seed": 5651, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000090/group_9/index_8416_seed_5651_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08649978516623377, + "vlm_logprob": -0.08649978516623377, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.08649978516623377, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.36152827739715576 + ], + "opposite_logprob_tokens": [ + -1.2365282773971558 + ], + "logprob_reward": -0.36152827739715576, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.36152827739715576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24758917093276978, + 0.06638641655445099, + 0.053559888154268265, + 0.3669925928115845, + 0.058089032769203186, + 0.031717076897621155, + 0.03977365046739578, + 0.04922175779938698, + 0.05903621017932892, + 0.027634194120764732 + ], + "credit_map": [ + 0.04255993291735649, + 0.046244073659181595, + 0.02035374566912651, + 0.02453770488500595, + 0.025324443355202675, + 0.02965615503489971, + 0.027166422456502914, + 0.03174670785665512, + 0.008991028182208538, + 0.012656842358410358, + 0.008099410682916641, + 0.005866325460374355, + 0.006008002441376448, + 0.011616519652307034, + 0.0058126868680119514, + 0.0073355999775230885, + 0.01215734425932169, + 0.007045646198093891, + 0.005654024891555309, + 0.004325782414525747, + 0.0061238291673362255, + 0.007190098986029625, + 0.005393174942582846, + 0.005669986363500357, + 0.07739029824733734, + 0.128104105591774, + 0.011372705921530724, + 0.01702580787241459, + 0.019573064520955086, + 0.07856904715299606, + 0.018467748537659645, + 0.016489816829562187, + 0.010460062883794308, + 0.007556494325399399, + 0.005176654551178217, + 0.00463997945189476, + 0.005777277518063784, + 0.015157700516283512, + 0.0045387581922113895, + 0.0047821057960391045, + 0.004888560622930527, + 0.002882471075281501, + 0.002525380812585354, + 0.0047652022913098335, + 0.0022238746751099825, + 0.006779327988624573, + 0.0045002531260252, + 0.0031520079355686903, + 0.006143780890852213, + 0.00517495721578598, + 0.002816800493746996, + 0.005348482634872198, + 0.0035980280954390764, + 0.005990945268422365, + 0.006807839032262564, + 0.003892818232998252, + 0.011467023752629757, + 0.0033650470431894064, + 0.004586077760905027, + 0.0049664732068777084, + 0.003441505366936326, + 0.012209701351821423, + 0.005046541802585125, + 0.004139389377087355, + 0.007941468618810177, + 0.005616908892989159, + 0.004976092837750912, + 0.012103486806154251, + 0.005112016573548317, + 0.008905119262635708, + 0.009363983757793903, + 0.0050171311013400555, + 0.0043804761953651905, + 0.0016297067049890757, + 0.0023950536269694567, + 0.0036607671063393354, + 0.0025379720609635115, + 0.006274472456425428, + 0.002918329555541277, + 0.003837414551526308 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.36152827739715576 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02062656544148922 + ], + "opposite_logprob_tokens": [ + -3.8956265449523926 + ], + "logprob_reward": -0.02062656544148922, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02062656544148922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3228014409542084, + 0.05261249467730522, + 0.031188206747174263, + 0.4639554023742676, + 0.040280409157276154, + 0.013180180452764034, + 0.022031422704458237, + 0.022714903578162193, + 0.0193322841078043, + 0.011903265491127968 + ], + "credit_map": [ + 0.06306964159011841, + 0.05603865161538124, + 0.030045824125409126, + 0.02860964462161064, + 0.026002774015069008, + 0.038444884121418, + 0.031608857214450836, + 0.04898115620017052, + 0.008537945337593555, + 0.00741846626624465, + 0.007123025134205818, + 0.005333656445145607, + 0.004328539129346609, + 0.008536585606634617, + 0.005543495994061232, + 0.005790784023702145, + 0.005737690255045891, + 0.005911792628467083, + 0.002758389338850975, + 0.0018397611565887928, + 0.0031696425285190344, + 0.006123203784227371, + 0.002630770206451416, + 0.0030169582460075617, + 0.10475581884384155, + 0.16774454712867737, + 0.01281767524778843, + 0.020724186673760414, + 0.023588603362441063, + 0.09595184028148651, + 0.021448833867907524, + 0.016923924908041954, + 0.006724161561578512, + 0.007416224107146263, + 0.00268784468062222, + 0.004003206267952919, + 0.0050580683164298534, + 0.00643357215449214, + 0.0035728539805859327, + 0.004384481813758612, + 0.002415840048342943, + 0.00122892577201128, + 0.0011100464034825563, + 0.0013791978126391768, + 0.0011054453207179904, + 0.003385895863175392, + 0.0011772429570555687, + 0.0013775876723229885, + 0.0035814892034977674, + 0.003291589207947254, + 0.0016651907935738564, + 0.0034596589393913746, + 0.002311830408871174, + 0.002729336731135845, + 0.0028445387724786997, + 0.002147788880392909, + 0.0055415015667676926, + 0.0019751605577766895, + 0.0021655000746250153, + 0.002136742463335395, + 0.002470561070367694, + 0.004818848334252834, + 0.00177211151458323, + 0.0018344786949455738, + 0.0023960107937455177, + 0.002137676114216447, + 0.0016422518528997898, + 0.0041078790090978146, + 0.001612375839613378, + 0.0017639144789427519, + 0.003792505944147706, + 0.0018796698423102498, + 0.0018179393373429775, + 0.0012200891505926847, + 0.0015532438410446048, + 0.0016762031009420753, + 0.0009344274876639247, + 0.0022600963711738586, + 0.0014763915678486228, + 0.0009648752748034894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.02062656544148922 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041068983264267445 + ], + "opposite_logprob_tokens": [ + -5.5041069984436035 + ], + "logprob_reward": -0.0041068983264267445, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041068983264267445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2840558588504791, + 0.09426558762788773, + 0.02416042797267437, + 0.4563460052013397, + 0.04170981049537659, + 0.0139007568359375, + 0.011281847953796387, + 0.02191663719713688, + 0.02121032029390335, + 0.031152762472629547 + ], + "credit_map": [ + 0.053872402757406235, + 0.04998748376965523, + 0.027320947498083115, + 0.025804072618484497, + 0.02237205021083355, + 0.0331730991601944, + 0.029139501973986626, + 0.042386289685964584, + 0.015200948342680931, + 0.025819147005677223, + 0.008598219603300095, + 0.006404898129403591, + 0.005249808542430401, + 0.01937473937869072, + 0.006253453437238932, + 0.007364371791481972, + 0.0038185042794793844, + 0.003753824159502983, + 0.002045226749032736, + 0.0019761123694479465, + 0.004064072389155626, + 0.0033604621421545744, + 0.003139324951916933, + 0.002002899767830968, + 0.09996313601732254, + 0.16627515852451324, + 0.013028500601649284, + 0.020008226856589317, + 0.023534102365374565, + 0.09710056334733963, + 0.019951345399022102, + 0.016484998166561127, + 0.006054264027625322, + 0.0035349479876458645, + 0.003386933822184801, + 0.0052985153160989285, + 0.0048760282807052135, + 0.007343764416873455, + 0.00522358063608408, + 0.00599177461117506, + 0.002870300319045782, + 0.0012696696212515235, + 0.0011211000382900238, + 0.0013522353256121278, + 0.0010444920044392347, + 0.0032733839470893145, + 0.0012002990115433931, + 0.0017692764522507787, + 0.0016688286559656262, + 0.0013623804552480578, + 0.000889078713953495, + 0.0016320800641551614, + 0.0014902815455570817, + 0.0015629841946065426, + 0.0014898583758622408, + 0.0011863565305247903, + 0.004655888304114342, + 0.0021494184620678425, + 0.0019985735416412354, + 0.002494390122592449, + 0.0022812634706497192, + 0.0033928717020899057, + 0.0029992905911058187, + 0.0019449400715529919, + 0.002242498332634568, + 0.00236895983107388, + 0.0017651923699304461, + 0.0048409635201096535, + 0.0019901543855667114, + 0.0018579409224912524, + 0.004095533862709999, + 0.002049077767878771, + 0.006175542715936899, + 0.0026260640006512403, + 0.0025773150846362114, + 0.0030487370677292347, + 0.002591728465631604, + 0.00892464630305767, + 0.0024487492628395557, + 0.0027599807363003492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.0041068983264267445 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542887590825558 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.012542887590825558, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542887590825558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28714483976364136, + 0.08394350111484528, + 0.05471954122185707, + 0.198931485414505, + 0.12353148311376572, + 0.02804558351635933, + 0.02059886045753956, + 0.08546824753284454, + 0.07153567671775818, + 0.046080779284238815 + ], + "credit_map": [ + 0.05725156143307686, + 0.05411890894174576, + 0.030258622020483017, + 0.024207569658756256, + 0.022726017981767654, + 0.03244209289550781, + 0.02822129987180233, + 0.03791877254843712, + 0.011064386926591396, + 0.016197849065065384, + 0.007846595719456673, + 0.008712217211723328, + 0.010654805228114128, + 0.011826208792626858, + 0.005665917415171862, + 0.011975526809692383, + 0.00953641813248396, + 0.007419677451252937, + 0.0051034362986683846, + 0.005230622831732035, + 0.008415930904448032, + 0.006631908006966114, + 0.007255523465573788, + 0.005126024596393108, + 0.047139011323451996, + 0.07105813920497894, + 0.0073722838424146175, + 0.008687484078109264, + 0.00922956969588995, + 0.038811009377241135, + 0.008518831804394722, + 0.008115150034427643, + 0.019190816208720207, + 0.019675634801387787, + 0.00678679347038269, + 0.014588837511837482, + 0.01813124492764473, + 0.01821349561214447, + 0.013964256271719933, + 0.012980407103896141, + 0.005730388220399618, + 0.0036727944388985634, + 0.0022834755945950747, + 0.0032068921718746424, + 0.002425211714580655, + 0.005739012733101845, + 0.0029028593562543392, + 0.0020849513821303844, + 0.003017673036083579, + 0.0023611821234226227, + 0.001327461446635425, + 0.002684319391846657, + 0.002563096582889557, + 0.0029660991858690977, + 0.003135770559310913, + 0.00254325894638896, + 0.01794901117682457, + 0.008109303191304207, + 0.005574815906584263, + 0.010845511220395565, + 0.00935310311615467, + 0.014446801505982876, + 0.010817511938512325, + 0.008372196927666664, + 0.01008154358714819, + 0.008068039081990719, + 0.005142404232174158, + 0.010682391002774239, + 0.00940672867000103, + 0.007945287972688675, + 0.012724341824650764, + 0.007484935689717531, + 0.010056530125439167, + 0.004385100677609444, + 0.0059516034089028835, + 0.007108715828508139, + 0.003935069777071476, + 0.005633690860122442, + 0.004848221316933632, + 0.00416184775531292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.012542887590825558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from an extremely low angle looking upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a firefly moving in a slow, graceful spiral orbiting a large flower?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the firefly emit a warm, luminous yellow-green glow from its abdomen that leaves a dissipating light trail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the flower a fantastical hybrid of a lotus and lily with translucent milky white petals edged in pink-purple and covered in dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a deep night sky with distant stars and blurred outlines of ancient trees?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033694297075271606 + ], + "opposite_logprob_tokens": [ + -3.408694267272949 + ], + "logprob_reward": -0.033694297075271606, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033694297075271606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11235551536083221, + 0.05595937371253967, + 0.21013256907463074, + 0.3458968698978424, + 0.0474068745970726, + 0.05465051904320717, + 0.03343658521771431, + 0.0808856412768364, + 0.042906515300273895, + 0.016369611024856567 + ], + "credit_map": [ + 0.01696176268160343, + 0.01648537814617157, + 0.01074916310608387, + 0.01298944279551506, + 0.012229615822434425, + 0.01320573315024376, + 0.01301621925085783, + 0.016718197613954544, + 0.008286525495350361, + 0.010186119005084038, + 0.00661968719214201, + 0.00686079915612936, + 0.00533561734482646, + 0.008761744014918804, + 0.0048133861273527145, + 0.005095494911074638, + 0.030648136511445045, + 0.036802973598241806, + 0.017993738874793053, + 0.01715213805437088, + 0.030887264758348465, + 0.036245595663785934, + 0.024447999894618988, + 0.015954699367284775, + 0.07580631971359253, + 0.12619253993034363, + 0.010540656745433807, + 0.014750951901078224, + 0.01722215861082077, + 0.07376644760370255, + 0.014612028375267982, + 0.013005759567022324, + 0.007225331384688616, + 0.004656387027353048, + 0.00445633614435792, + 0.005431154742836952, + 0.005335223861038685, + 0.01009744219481945, + 0.004570588935166597, + 0.005634411703795195, + 0.009353884495794773, + 0.006029817741364241, + 0.004254491534084082, + 0.004043001215904951, + 0.005039977841079235, + 0.01537000760436058, + 0.0053312345407903194, + 0.005228098947554827, + 0.007367610931396484, + 0.0033872700296342373, + 0.0022388554643839598, + 0.003965689335018396, + 0.002577613340690732, + 0.006720091216266155, + 0.004434631206095219, + 0.002744820201769471, + 0.018938573077321053, + 0.007203946355730295, + 0.004576004575937986, + 0.007361622992902994, + 0.0075262440368533134, + 0.020217452198266983, + 0.0065928176045417786, + 0.008468981832265854, + 0.00613329978659749, + 0.00351713877171278, + 0.003409066004678607, + 0.008377814665436745, + 0.0037052498664706945, + 0.007559609599411488, + 0.0058141485787928104, + 0.004390188958495855, + 0.0036421187687665224, + 0.0010510326828807592, + 0.00182265043258667, + 0.001758457045070827, + 0.0010373503901064396, + 0.0038936776109039783, + 0.0014048407319933176, + 0.0017594830133020878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.033694297075271606 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_8/index_tensor([9559, 9559])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_8/index_tensor([9559, 9559])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9086ba3995738a30e7b85ba6ae92f1020a6a9ffc --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_8/index_tensor([9559, 9559])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot follows a two-wheeled bronze chariot, pulled by four majestic black horses, galloping across a vast, dusty battlefield. Atop the chariot sits a majestic, middle-aged emperor, poised as a rock. He wears an ancient imperial crown with dangling beads and a heavy, deep-black silk dragon robe, delicately embroidered with dazzling gold thread, depicting five-clawed golden dragons with bared fangs and distinct scales. The wide cuffs dance wildly in the wind. His face is stern, his gaze piercing, as he gazes straight ahead. One hand firmly grips the rough leather reins, steadying the neighing steed, while the other rests on the hilt of the ancient sword at his waist. The chariot itself is covered in ancient cloud-thunder patterns and deep battle scars, its bronze gleaming in the setting sun, a dark, cold luster. The chariot's massive wheels whirred at high speed, stirring up a cloud of yellow sand along with the horses' hooves, almost obscuring the silhouettes of the clashing armies and the burning beacon fires in the distance. The setting sun's golden afterglow pierced the dust, tinging the entire scene with a warm and tragic hue, outlining the emperor's resolute profile and the shimmering golden threads of his dragon robe. The entire scene was imbued with a sense of epic power and unstoppable royal momentum, creating a solemn and tense atmosphere.", + "seed": 1308, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_8/index_9559_seed_1308_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04924780847504735, + "vlm_logprob": -0.04924780847504735, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.04924780847504735, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0948219746351242 + ], + "opposite_logprob_tokens": [ + -2.8448219299316406 + ], + "logprob_reward": -0.0948219746351242, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0948219746351242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29003775119781494, + 0.11028759181499481, + 0.055989768356084824, + 0.07609384506940842, + 0.06931385397911072, + 0.05743992701172829, + 0.09516944736242294, + 0.0824049636721611, + 0.036138493567705154, + 0.12712430953979492 + ], + "credit_map": [ + 0.07397260516881943, + 0.026486404240131378, + 0.07049814611673355, + 0.01954302191734314, + 0.036484748125076294, + 0.019912347197532654, + 0.024445438757538795, + 0.018695030361413956, + 0.018424157053232193, + 0.016993284225463867, + 0.018153294920921326, + 0.018355103209614754, + 0.013891112059354782, + 0.008323348127305508, + 0.010176756419241428, + 0.005970533937215805, + 0.011616294272243977, + 0.006036808714270592, + 0.012775924988090992, + 0.0071504488587379456, + 0.009092109277844429, + 0.003001323901116848, + 0.003266687272116542, + 0.0030501708388328552, + 0.011717839166522026, + 0.007567364256829023, + 0.009911531582474709, + 0.010366606526076794, + 0.005888175219297409, + 0.008563444018363953, + 0.009337317198514938, + 0.012741565704345703, + 0.01187102496623993, + 0.0075020682998001575, + 0.014166941866278648, + 0.009500382468104362, + 0.006376838311553001, + 0.005850270390510559, + 0.006671037524938583, + 0.007375288289040327, + 0.00729400385171175, + 0.0068765063770115376, + 0.01223667524755001, + 0.006800829898566008, + 0.007846149615943432, + 0.004330387804657221, + 0.006837222259491682, + 0.005218151491135359, + 0.01168377697467804, + 0.02137426845729351, + 0.012671242468059063, + 0.013339475728571415, + 0.011512591503560543, + 0.006495343055576086, + 0.007449408061802387, + 0.010643338784575462, + 0.0047827777452766895, + 0.0033948859199881554, + 0.031179377809166908, + 0.0084226718172431, + 0.006510620936751366, + 0.004282814450562, + 0.004248375538736582, + 0.019583437591791153, + 0.00736240716651082, + 0.003304421668872237, + 0.003869215492159128, + 0.004858838394284248, + 0.004691432695835829, + 0.004347058478742838, + 0.0030620296020060778, + 0.004643090069293976, + 0.022762391716241837, + 0.015397191978991032, + 0.012687820009887218, + 0.016156736761331558, + 0.019618934020400047, + 0.015248602256178856, + 0.010984323918819427, + 0.014268320985138416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.0948219746351242 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890762135386467 + ], + "opposite_logprob_tokens": [ + -2.464076280593872 + ], + "logprob_reward": -0.0890762135386467, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890762135386467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14874935150146484, + 0.3127153813838959, + 0.17230811715126038, + 0.05348770692944527, + 0.07619091868400574, + 0.030405249446630478, + 0.08096975088119507, + 0.053709059953689575, + 0.015453980304300785, + 0.056010544300079346 + ], + "credit_map": [ + 0.016324011608958244, + 0.018424853682518005, + 0.017346886917948723, + 0.01917034015059471, + 0.02780282311141491, + 0.020203765481710434, + 0.013351776637136936, + 0.016124865040183067, + 0.04375281184911728, + 0.03894354775547981, + 0.03865858539938927, + 0.052234090864658356, + 0.032463569194078445, + 0.022579479962587357, + 0.051856622099876404, + 0.03222664073109627, + 0.0563320517539978, + 0.010779199190437794, + 0.07034408301115036, + 0.010860137641429901, + 0.006445006467401981, + 0.0055190324783325195, + 0.008030849508941174, + 0.003997747786343098, + 0.007562587969005108, + 0.0061429706402122974, + 0.008194338530302048, + 0.011752728372812271, + 0.004426741506904364, + 0.004431797191500664, + 0.0055689821019768715, + 0.00540755782276392, + 0.011049078777432442, + 0.01290593110024929, + 0.014197374694049358, + 0.007542314473539591, + 0.008904734626412392, + 0.006152171641588211, + 0.00690617086365819, + 0.00853313785046339, + 0.004890714772045612, + 0.0036523251328617334, + 0.0032081359531730413, + 0.002786195371299982, + 0.003566667903214693, + 0.0026863112580031157, + 0.0036569093354046345, + 0.005957985296845436, + 0.009648749604821205, + 0.012716447934508324, + 0.010231819935142994, + 0.009198563173413277, + 0.017399510368704796, + 0.0065222675912082195, + 0.00653248094022274, + 0.008719905279576778, + 0.005535340867936611, + 0.005632917396724224, + 0.008488738909363747, + 0.006069886963814497, + 0.00650903582572937, + 0.0053812176920473576, + 0.004547364078462124, + 0.011544551700353622, + 0.0028121001087129116, + 0.0016908879624679685, + 0.0020889558363705873, + 0.0015143940690904856, + 0.002475559711456299, + 0.001547129824757576, + 0.0015988113591447473, + 0.0017261393368244171, + 0.008153003640472889, + 0.006603440269827843, + 0.005026094615459442, + 0.007141567301005125, + 0.010574149899184704, + 0.00620028842240572, + 0.006383857224136591, + 0.005928135477006435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0890762135386467 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306614026427269 + ], + "opposite_logprob_tokens": [ + -3.1680660247802734 + ], + "logprob_reward": -0.04306614026427269, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306614026427269, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12255113571882248, + 0.0744272992014885, + 0.4151766002178192, + 0.02610561065375805, + 0.1076144203543663, + 0.027051812037825584, + 0.1175067275762558, + 0.028287511318922043, + 0.025258615612983704, + 0.05602031946182251 + ], + "credit_map": [ + 0.03291958197951317, + 0.009731220081448555, + 0.03667387738823891, + 0.011593728326261044, + 0.008831695653498173, + 0.006764987949281931, + 0.008079614490270615, + 0.007956425659358501, + 0.011002341285347939, + 0.008505656383931637, + 0.00976915005594492, + 0.01079805288463831, + 0.006983124651014805, + 0.00594713818281889, + 0.013032772578299046, + 0.008389062248170376, + 0.1320611536502838, + 0.025677397847175598, + 0.16564825177192688, + 0.02653922326862812, + 0.01318220142275095, + 0.016090067103505135, + 0.020426645874977112, + 0.015551645308732986, + 0.004241109825670719, + 0.0028235330246388912, + 0.005020515061914921, + 0.004072097595781088, + 0.002440617186948657, + 0.0032124409917742014, + 0.002051898278295994, + 0.0022433986887335777, + 0.015169230289757252, + 0.010257583111524582, + 0.02566150762140751, + 0.010068531148135662, + 0.013840598985552788, + 0.010608495213091373, + 0.012156154029071331, + 0.009852319955825806, + 0.00353243644349277, + 0.0029670142102986574, + 0.004788298159837723, + 0.003076135879382491, + 0.002338303020223975, + 0.002398574724793434, + 0.0032409110572189093, + 0.004710139706730843, + 0.013105957768857479, + 0.014810210093855858, + 0.024634018540382385, + 0.014415809884667397, + 0.011402273550629616, + 0.013807730749249458, + 0.009871172718703747, + 0.015459559857845306, + 0.00221378356218338, + 0.0013077727053314447, + 0.009407268837094307, + 0.0028749967459589243, + 0.001264785067178309, + 0.001976198051124811, + 0.0017450503073632717, + 0.007497657090425491, + 0.0040895529091358185, + 0.001763619133271277, + 0.0038298191502690315, + 0.003187792841345072, + 0.003752233926206827, + 0.0025415965355932713, + 0.002521871356293559, + 0.0035721282474696636, + 0.006487728096544743, + 0.00562426820397377, + 0.005437726154923439, + 0.008241014555096626, + 0.01150934025645256, + 0.006096820812672377, + 0.007144071627408266, + 0.0054793511517345905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.04306614026427269 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539827264845371 + ], + "opposite_logprob_tokens": [ + -4.387539863586426 + ], + "logprob_reward": -0.012539827264845371, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539827264845371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2399825006723404, + 0.09075067937374115, + 0.05220263451337814, + 0.19989897310733795, + 0.12969520688056946, + 0.02638697251677513, + 0.06481047719717026, + 0.0706583708524704, + 0.0700930505990982, + 0.055521149188280106 + ], + "credit_map": [ + 0.0903940498828888, + 0.013334291987121105, + 0.09077298641204834, + 0.01122564822435379, + 0.01004584226757288, + 0.0094087990000844, + 0.0071762134321033955, + 0.007624683436006308, + 0.014484817162156105, + 0.01414796058088541, + 0.013158870860934258, + 0.016118943691253662, + 0.007924030534923077, + 0.005256582982838154, + 0.011361844837665558, + 0.008297644555568695, + 0.011969289742410183, + 0.0061319125816226006, + 0.01270696148276329, + 0.005971262231469154, + 0.006646955385804176, + 0.0029772156849503517, + 0.00323282228782773, + 0.002566218376159668, + 0.04466870054602623, + 0.0189520176500082, + 0.04311402887105942, + 0.026815518736839294, + 0.01445561833679676, + 0.013783149421215057, + 0.01940891146659851, + 0.018701031804084778, + 0.02005118690431118, + 0.013337042182683945, + 0.028507420793175697, + 0.013400740921497345, + 0.016450872644782066, + 0.012941151857376099, + 0.014406976290047169, + 0.01059981994330883, + 0.003304586512967944, + 0.002527585718780756, + 0.006625623907893896, + 0.002465201774612069, + 0.00280853477306664, + 0.002950652502477169, + 0.0031702863052487373, + 0.002534502651542425, + 0.009175801649689674, + 0.007920912466943264, + 0.006494239903986454, + 0.006793533451855183, + 0.014877412468194962, + 0.006681661121547222, + 0.005338296294212341, + 0.007528624497354031, + 0.0065808952786028385, + 0.005640738178044558, + 0.01666206866502762, + 0.008613948710262775, + 0.007290331646800041, + 0.007783845998346806, + 0.004200320225208998, + 0.013886225409805775, + 0.018554428592324257, + 0.006622963584959507, + 0.00785821583122015, + 0.008478476665914059, + 0.008495724759995937, + 0.010155607014894485, + 0.005121036898344755, + 0.004806602839380503, + 0.004586480092257261, + 0.006139977835118771, + 0.005245516542345285, + 0.00782102346420288, + 0.010631040669977665, + 0.005997849628329277, + 0.009683547541499138, + 0.005415716674178839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.012539827264845371 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734886672347784 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.006734886672347784, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734886672347784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3925783038139343, + 0.12816354632377625, + 0.11266936361789703, + 0.08755774050951004, + 0.060763075947761536, + 0.03194011375308037, + 0.06407871097326279, + 0.050528384745121, + 0.029065944254398346, + 0.042654771357774734 + ], + "credit_map": [ + 0.13099363446235657, + 0.03035653382539749, + 0.1308591365814209, + 0.02317235805094242, + 0.02426701970398426, + 0.020838819444179535, + 0.01730545610189438, + 0.014785396866500378, + 0.026235466822981834, + 0.01486616488546133, + 0.023927748203277588, + 0.015076735056936741, + 0.01092058327049017, + 0.009735832922160625, + 0.01377343200147152, + 0.013627598062157631, + 0.0264528077095747, + 0.011346247047185898, + 0.035449348390102386, + 0.010338306427001953, + 0.01348059717565775, + 0.005026042927056551, + 0.006056986283510923, + 0.004519036039710045, + 0.015934288501739502, + 0.009254946373403072, + 0.012810914777219296, + 0.013801331631839275, + 0.0065020048059523106, + 0.00865556113421917, + 0.009813616052269936, + 0.010785086080431938, + 0.011763195507228374, + 0.006445862352848053, + 0.012243950739502907, + 0.007112365216016769, + 0.004728381056338549, + 0.00560363195836544, + 0.006504858378320932, + 0.0063608321361243725, + 0.00544325215741992, + 0.0054197306744754314, + 0.004064253531396389, + 0.0034149433486163616, + 0.003550179535523057, + 0.003484379965811968, + 0.003287426196038723, + 0.0032759481109678745, + 0.008752112276852131, + 0.00983576849102974, + 0.006194246932864189, + 0.006056813523173332, + 0.014498034492135048, + 0.007426230702549219, + 0.0052664619870483875, + 0.0060490453615784645, + 0.004251384641975164, + 0.003324556862935424, + 0.015560726635158062, + 0.004869544412940741, + 0.0025867444928735495, + 0.0035772130358964205, + 0.0026351851411163807, + 0.013723034411668777, + 0.005805902648717165, + 0.004319739528000355, + 0.0023627416230738163, + 0.0037775267846882343, + 0.003991433884948492, + 0.003706588875502348, + 0.0023068622685968876, + 0.0027951544616371393, + 0.003016181057319045, + 0.0042755138128995895, + 0.0027950862422585487, + 0.003929808270186186, + 0.012881790287792683, + 0.005339581985026598, + 0.00556469801813364, + 0.004852114710956812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.006734886672347784 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a two-wheeled bronze chariot, pulled by four majestic black horses, galloping across a vast, dusty battlefield. Atop the chariot sits a majestic, middle-aged emperor, poised as a rock. He wears an ancient imperial crown with dangling beads and a heavy, deep-black silk dragon robe, delicately embroidered with dazzling gold thread, depicting five-clawed golden dragons with bared fangs and distinct scales. The wide cuffs dance wildly in the wind. His face is stern, his gaze piercing, as he gazes straight ahead. One hand firmly grips the rough leather reins, steadying the neighing steed, while the other rests on the hilt of the ancient sword at his waist. The chariot itself is covered in ancient cloud-thunder patterns and deep battle scars, its bronze gleaming in the setting sun, a dark, cold luster. The chariot's massive wheels whirred at high speed, stirring up a cloud of yellow sand along with the horses' hooves, almost obscuring the silhouettes of the clashing armies and the burning beacon fires in the distance. The setting sun's golden afterglow pierced the dust, tinging the entire scene with a warm and tragic hue, outlining the emperor's resolute profile and the shimmering golden threads of his dragon robe. The entire scene was imbued with a sense of epic power and unstoppable royal momentum, creating a solemn and tense atmosphere.", + "seed": 1309, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_8/index_9559_seed_1309_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.17254440672695637, + "vlm_logprob": -0.17254440672695637, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.17254440672695637, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.23605121672153473 + ], + "opposite_logprob_tokens": [ + -1.7360512018203735 + ], + "logprob_reward": -0.23605121672153473, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.23605121672153473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42165064811706543, + 0.09188902378082275, + 0.029375571757555008, + 0.03870931267738342, + 0.04854986071586609, + 0.0483732707798481, + 0.12444678694009781, + 0.04285522550344467, + 0.09334930032491684, + 0.0608009435236454 + ], + "credit_map": [ + 0.06457499414682388, + 0.037715665996074677, + 0.05982503294944763, + 0.16877977550029755, + 0.022095512598752975, + 0.013172279112040997, + 0.024323701858520508, + 0.031163781881332397, + 0.01623617857694626, + 0.02255805768072605, + 0.010736469179391861, + 0.01556637417525053, + 0.008189196698367596, + 0.005960844457149506, + 0.006110927555710077, + 0.00653099175542593, + 0.003144195768982172, + 0.004440099932253361, + 0.0038806393276900053, + 0.0052469768561422825, + 0.003195685800164938, + 0.0038307439535856247, + 0.002764918142929673, + 0.0028723159339278936, + 0.005368268117308617, + 0.004455831833183765, + 0.005715458653867245, + 0.007695579435676336, + 0.004954836331307888, + 0.004147379659116268, + 0.0028761811554431915, + 0.0034957854077219963, + 0.006300561595708132, + 0.005072897300124168, + 0.010644322261214256, + 0.005003389902412891, + 0.007419428788125515, + 0.004293021280318499, + 0.005241732113063335, + 0.004574514459818602, + 0.005229114089161158, + 0.0037596761249005795, + 0.013455929234623909, + 0.006073556374758482, + 0.005132335238158703, + 0.004160787910223007, + 0.005334812216460705, + 0.005227065645158291, + 0.018010104075074196, + 0.01701224595308304, + 0.014830517582595348, + 0.02368674799799919, + 0.016252921894192696, + 0.009484982118010521, + 0.012642397545278072, + 0.012526878155767918, + 0.008615299127995968, + 0.006359566003084183, + 0.006027877796441317, + 0.004645905457437038, + 0.00594065198674798, + 0.0041242982260882854, + 0.004062016028910875, + 0.0030796208884567022, + 0.013383913785219193, + 0.00906379334628582, + 0.011607041582465172, + 0.015893878415226936, + 0.01298914197832346, + 0.008126780390739441, + 0.012587013654410839, + 0.00969774555414915, + 0.009206751361489296, + 0.00676231412217021, + 0.010485372506082058, + 0.007844136096537113, + 0.005084389820694923, + 0.005060645751655102, + 0.008249429054558277, + 0.00810791552066803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.23605121672153473 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743354916572571 + ], + "opposite_logprob_tokens": [ + -0.9743354916572571 + ], + "logprob_reward": -0.4743354916572571, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743354916572571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40718981623649597, + 0.03868388384580612, + 0.033548127859830856, + 0.1132013350725174, + 0.0823611468076706, + 0.12438246607780457, + 0.048045191913843155, + 0.07957708835601807, + 0.05064080283045769, + 0.022370142862200737 + ], + "credit_map": [ + 0.0749146118760109, + 0.028589248657226562, + 0.06123340129852295, + 0.15951167047023773, + 0.01694406196475029, + 0.013140762224793434, + 0.02414197288453579, + 0.028714092448353767, + 0.00466211698949337, + 0.005394732113927603, + 0.004180479794740677, + 0.007929383777081966, + 0.006568353623151779, + 0.0033233880531042814, + 0.0029704503249377012, + 0.0036549814976751804, + 0.003195779165253043, + 0.003520422615110874, + 0.005066790152341127, + 0.005791509989649057, + 0.004356358665972948, + 0.0040541766211390495, + 0.004322339780628681, + 0.0032407515682280064, + 0.013411149382591248, + 0.01007372047752142, + 0.012790163978934288, + 0.02892417646944523, + 0.01910415105521679, + 0.00928095355629921, + 0.010555882006883621, + 0.009061138145625591, + 0.009577245451509953, + 0.010162963531911373, + 0.012473692186176777, + 0.010620372369885445, + 0.011546608991920948, + 0.006370760500431061, + 0.010541299358010292, + 0.01106820348650217, + 0.022559914737939835, + 0.007120201829820871, + 0.04498128220438957, + 0.01635574921965599, + 0.004906478337943554, + 0.0060903532430529594, + 0.010009965859353542, + 0.01235852763056755, + 0.005029398016631603, + 0.006214054301381111, + 0.00542127201333642, + 0.011810663156211376, + 0.005101845599710941, + 0.006076481193304062, + 0.0034964685328304768, + 0.004895011428743601, + 0.011530055664479733, + 0.012292827479541302, + 0.011069484986364841, + 0.012839178554713726, + 0.011233358643949032, + 0.005781048443168402, + 0.008297709748148918, + 0.006533428560942411, + 0.007848360575735569, + 0.00618283124640584, + 0.005915121641010046, + 0.006285741925239563, + 0.011500953696668148, + 0.004055566154420376, + 0.0047971270978450775, + 0.00405510189011693, + 0.0030616046860814095, + 0.002626363653689623, + 0.002189493738114834, + 0.0027500144205987453, + 0.003840623889118433, + 0.002389423083513975, + 0.00298325065523386, + 0.0025293694343417883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.4743354916572571 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296620965003967 + ], + "opposite_logprob_tokens": [ + -2.237966299057007 + ], + "logprob_reward": -0.11296620965003967, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296620965003967, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7033199667930603, + 0.06431719660758972, + 0.026951665058732033, + 0.024289753288030624, + 0.03679546341300011, + 0.03199677914381027, + 0.015592371113598347, + 0.023976296186447144, + 0.03350893035531044, + 0.03925161808729172 + ], + "credit_map": [ + 0.1250355988740921, + 0.045561283826828, + 0.10446498543024063, + 0.3002367913722992, + 0.02304551750421524, + 0.015623408369719982, + 0.03945480287075043, + 0.04989752918481827, + 0.007878606207668781, + 0.012800064869225025, + 0.006903717294335365, + 0.01953696273267269, + 0.004665810149163008, + 0.004274273291230202, + 0.004164904356002808, + 0.004092854913324118, + 0.003485337132588029, + 0.003134302794933319, + 0.004557549022138119, + 0.004886954557150602, + 0.003526984481140971, + 0.003416134975850582, + 0.002069938462227583, + 0.001874461886473, + 0.002656460041180253, + 0.0028492864221334457, + 0.003013166831806302, + 0.006180542521178722, + 0.0030883667059242725, + 0.002031076466664672, + 0.0020888454746454954, + 0.0023820060305297375, + 0.00480887945741415, + 0.004131795838475227, + 0.005366804543882608, + 0.005338076036423445, + 0.007229759823530912, + 0.0035860310308635235, + 0.003362999763339758, + 0.0029711141251027584, + 0.005897458642721176, + 0.0026695302221924067, + 0.006252285558730364, + 0.0038670534268021584, + 0.005073078908026218, + 0.0020804861560463905, + 0.0028765504248440266, + 0.003280332777649164, + 0.0018826165469363332, + 0.0017762422794476151, + 0.0016439218306913972, + 0.0040887570939958096, + 0.0016020550392568111, + 0.0016051986021921039, + 0.0014112107455730438, + 0.0015823680441826582, + 0.0033794173505157232, + 0.0028543462976813316, + 0.0023395386524498463, + 0.005635641515254974, + 0.0024617828894406557, + 0.0021302851382642984, + 0.002766568446531892, + 0.002408714033663273, + 0.003615947673097253, + 0.004079671110957861, + 0.006802642717957497, + 0.005195652600377798, + 0.006047500297427177, + 0.002432522363960743, + 0.002487743739038706, + 0.0028472470585256815, + 0.006377099547535181, + 0.004351959563791752, + 0.003576819784939289, + 0.005284840241074562, + 0.008340245112776756, + 0.0042971461080014706, + 0.0029237952549010515, + 0.004099710378795862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.11296620965003967 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081741079688072 + ], + "opposite_logprob_tokens": [ + -4.141081809997559 + ], + "logprob_reward": -0.016081741079688072, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081741079688072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27469128370285034, + 0.07740772515535355, + 0.0462099052965641, + 0.07602095603942871, + 0.08375595510005951, + 0.08072423934936523, + 0.08541760593652725, + 0.1245049312710762, + 0.05412911996245384, + 0.09713829308748245 + ], + "credit_map": [ + 0.051101747900247574, + 0.0240622628480196, + 0.041106343269348145, + 0.0887693539261818, + 0.017706800252199173, + 0.015749938786029816, + 0.016922926530241966, + 0.019271891564130783, + 0.010300472378730774, + 0.011159494519233704, + 0.009180059656500816, + 0.017685262486338615, + 0.010998683050274849, + 0.006663373205810785, + 0.005975526757538319, + 0.005444852635264397, + 0.005262379068881273, + 0.007943118922412395, + 0.006308780517429113, + 0.008620110340416431, + 0.006162614095956087, + 0.004481641575694084, + 0.003885843325406313, + 0.003545415820553899, + 0.01025971956551075, + 0.009354333393275738, + 0.01001761481165886, + 0.022067613899707794, + 0.005729177035391331, + 0.006606994196772575, + 0.005881614983081818, + 0.006103892810642719, + 0.012799583375453949, + 0.008276697248220444, + 0.013714658096432686, + 0.011525079607963562, + 0.010949353687465191, + 0.00803349632769823, + 0.008573628030717373, + 0.009883468970656395, + 0.011055590584874153, + 0.0068102432414889336, + 0.017003705725073814, + 0.01296265423297882, + 0.008928272873163223, + 0.006820264738053083, + 0.00780579075217247, + 0.009337718598544598, + 0.011577076278626919, + 0.014007705263793468, + 0.010443514212965965, + 0.016677699983119965, + 0.011763446033000946, + 0.00593339279294014, + 0.006973647978156805, + 0.008041124790906906, + 0.017289532348513603, + 0.01689002476632595, + 0.013885139487683773, + 0.028921915218234062, + 0.013778596185147762, + 0.010309426113963127, + 0.012697800993919373, + 0.01073249988257885, + 0.008136882446706295, + 0.006519572343677282, + 0.006358383223414421, + 0.0065246280282735825, + 0.012040174566209316, + 0.0042238081805408, + 0.0060294982977211475, + 0.004296171944588423, + 0.016048358753323555, + 0.012303536757826805, + 0.01000032015144825, + 0.011156141757965088, + 0.017335966229438782, + 0.01053967047482729, + 0.008519738912582397, + 0.011234559118747711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.016081741079688072 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287374526262283 + ], + "opposite_logprob_tokens": [ + -3.773287296295166 + ], + "logprob_reward": -0.023287374526262283, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287374526262283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4676115810871124, + 0.0292589720338583, + 0.0747973844408989, + 0.03955671563744545, + 0.039405759423971176, + 0.04111073538661003, + 0.03592389076948166, + 0.035218119621276855, + 0.19545599818229675, + 0.04166082665324211 + ], + "credit_map": [ + 0.09979770332574844, + 0.0275283083319664, + 0.06388047337532043, + 0.19424296915531158, + 0.01404296513646841, + 0.008487421087920666, + 0.025987740606069565, + 0.03364403545856476, + 0.004681989550590515, + 0.00638620788231492, + 0.003149111056700349, + 0.004690274596214294, + 0.0031449226662516594, + 0.0024177050217986107, + 0.0021603519562631845, + 0.0026284120976924896, + 0.008148999884724617, + 0.009386250749230385, + 0.01038780901581049, + 0.012290414422750473, + 0.01439871359616518, + 0.009634298272430897, + 0.005102948285639286, + 0.005447945557534695, + 0.00418206537142396, + 0.004204383119940758, + 0.005113889463245869, + 0.012817987240850925, + 0.004605840425938368, + 0.002602004213258624, + 0.0031801231671124697, + 0.002850423101335764, + 0.00574398273602128, + 0.005465852562338114, + 0.0068166786804795265, + 0.0050574676133692265, + 0.006569327786564827, + 0.0029878325294703245, + 0.003781246952712536, + 0.002983369864523411, + 0.00934675708413124, + 0.002433415735140443, + 0.011904358863830566, + 0.004410648252815008, + 0.002955728443339467, + 0.002397725125774741, + 0.003568136366084218, + 0.004093965515494347, + 0.007091104052960873, + 0.006668458227068186, + 0.004373040050268173, + 0.005070710089057684, + 0.0033479954581707716, + 0.0031983281951397657, + 0.0026725686620920897, + 0.0035016853362321854, + 0.007797536440193653, + 0.0037338670808821917, + 0.004812235478311777, + 0.0035340467002242804, + 0.005028792656958103, + 0.004253996070474386, + 0.002898180391639471, + 0.0031594650354236364, + 0.028438588604331017, + 0.01940237358212471, + 0.035114869475364685, + 0.030849797651171684, + 0.03013230301439762, + 0.012823394499719143, + 0.019463742151856422, + 0.019230924546718597, + 0.005739611107856035, + 0.004778598435223103, + 0.006935185752809048, + 0.004808085039258003, + 0.007967116311192513, + 0.0028558026533573866, + 0.0034786895848810673, + 0.0050977375358343124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.023287374526262283 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a two-wheeled bronze chariot, pulled by four majestic black horses, galloping across a vast, dusty battlefield. Atop the chariot sits a majestic, middle-aged emperor, poised as a rock. He wears an ancient imperial crown with dangling beads and a heavy, deep-black silk dragon robe, delicately embroidered with dazzling gold thread, depicting five-clawed golden dragons with bared fangs and distinct scales. The wide cuffs dance wildly in the wind. His face is stern, his gaze piercing, as he gazes straight ahead. One hand firmly grips the rough leather reins, steadying the neighing steed, while the other rests on the hilt of the ancient sword at his waist. The chariot itself is covered in ancient cloud-thunder patterns and deep battle scars, its bronze gleaming in the setting sun, a dark, cold luster. The chariot's massive wheels whirred at high speed, stirring up a cloud of yellow sand along with the horses' hooves, almost obscuring the silhouettes of the clashing armies and the burning beacon fires in the distance. The setting sun's golden afterglow pierced the dust, tinging the entire scene with a warm and tragic hue, outlining the emperor's resolute profile and the shimmering golden threads of his dragon robe. The entire scene was imbued with a sense of epic power and unstoppable royal momentum, creating a solemn and tense atmosphere.", + "seed": 1310, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_8/index_9559_seed_1310_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08198142386972904, + "vlm_logprob": -0.08198142386972904, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.08198142386972904, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1304350644350052 + ], + "opposite_logprob_tokens": [ + -2.505434989929199 + ], + "logprob_reward": -0.1304350644350052, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1304350644350052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1986270397901535, + 0.07195517420768738, + 0.02374131791293621, + 0.04947024583816528, + 0.01825508289039135, + 0.027173180133104324, + 0.3156623840332031, + 0.12349988520145416, + 0.0826234221458435, + 0.08899223059415817 + ], + "credit_map": [ + 0.031134236603975296, + 0.026813235133886337, + 0.043396152555942535, + 0.025187823921442032, + 0.02433937042951584, + 0.014400520361959934, + 0.021122204139828682, + 0.012233512476086617, + 0.00984959490597248, + 0.0065569630824029446, + 0.011669354513287544, + 0.008138623088598251, + 0.007773943245410919, + 0.008083597756922245, + 0.006686805747449398, + 0.013196298852562904, + 0.0022582721430808306, + 0.002371607581153512, + 0.005179794039577246, + 0.004534319508820772, + 0.0028003889601677656, + 0.0018858275143429637, + 0.0025539067573845387, + 0.0021572038531303406, + 0.0046686348505318165, + 0.0049742101691663265, + 0.00658149179071188, + 0.007095783017575741, + 0.007049182429909706, + 0.005799568258225918, + 0.005330205429345369, + 0.00797117780894041, + 0.002848009578883648, + 0.002278080675750971, + 0.0023578505497425795, + 0.0031111137941479683, + 0.002171556232497096, + 0.0023039153311401606, + 0.0016003731871023774, + 0.0015841855201870203, + 0.0031484835781157017, + 0.0028661435935646296, + 0.0029919471126049757, + 0.004300873260945082, + 0.00250580464489758, + 0.00261234724894166, + 0.0028925940860062838, + 0.00585499033331871, + 0.02393648773431778, + 0.1041044145822525, + 0.007387630175799131, + 0.01660071685910225, + 0.022712500765919685, + 0.10755536705255508, + 0.019842391833662987, + 0.01352289505302906, + 0.020477399230003357, + 0.0328412689268589, + 0.012022667564451694, + 0.014346904121339321, + 0.01579398475587368, + 0.01135850790888071, + 0.010660882107913494, + 0.005998286884278059, + 0.013494491577148438, + 0.014611227437853813, + 0.010968792252242565, + 0.011477955617010593, + 0.009356087073683739, + 0.005914750508964062, + 0.011067027226090431, + 0.005733097903430462, + 0.010322129353880882, + 0.010731961578130722, + 0.01914202608168125, + 0.014597739093005657, + 0.007738356478512287, + 0.009738950058817863, + 0.008481418713927269, + 0.008239662274718285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.1304350644350052 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17992311716079712 + ], + "opposite_logprob_tokens": [ + -1.8049230575561523 + ], + "logprob_reward": -0.17992311716079712, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17992311716079712, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38754335045814514, + 0.0913475975394249, + 0.03454117849469185, + 0.040912698954343796, + 0.04496266692876816, + 0.02892204001545906, + 0.0694514736533165, + 0.07667259871959686, + 0.12890329957008362, + 0.09674308449029922 + ], + "credit_map": [ + 0.05117237567901611, + 0.05858717858791351, + 0.08094063401222229, + 0.06273319572210312, + 0.0396980419754982, + 0.02380741573870182, + 0.03560515493154526, + 0.03499934822320938, + 0.011940806172788143, + 0.008870401419699192, + 0.015573793090879917, + 0.011717453598976135, + 0.008640741929411888, + 0.00909934751689434, + 0.008630967698991299, + 0.01687408797442913, + 0.0031011514365673065, + 0.003832148388028145, + 0.004403268918395042, + 0.005139179993420839, + 0.006143860053271055, + 0.004714103881269693, + 0.004172827582806349, + 0.0030346382409334183, + 0.005286368075758219, + 0.004177386406809092, + 0.008253593929111958, + 0.004786832258105278, + 0.005845574662089348, + 0.005066498648375273, + 0.004014863166958094, + 0.0034815792459994555, + 0.005989786237478256, + 0.006010600831359625, + 0.004941446706652641, + 0.008208109065890312, + 0.0057403021492064, + 0.005775901023298502, + 0.00539489695802331, + 0.0029016253538429737, + 0.003190320683643222, + 0.003636823734268546, + 0.002852048957720399, + 0.003327056998386979, + 0.006532014347612858, + 0.0027600987814366817, + 0.0036306697875261307, + 0.0029930081218481064, + 0.005523995496332645, + 0.021337032318115234, + 0.0024244138039648533, + 0.0035758623853325844, + 0.0050780074670910835, + 0.02345198206603527, + 0.004907290451228619, + 0.0031528857070952654, + 0.009210202842950821, + 0.01746414601802826, + 0.006511978339403868, + 0.007335144095122814, + 0.010658332146704197, + 0.01259169913828373, + 0.006757019553333521, + 0.006144077051430941, + 0.019457010552287102, + 0.016243452206254005, + 0.018836868926882744, + 0.015898121520876884, + 0.01844598911702633, + 0.012988501228392124, + 0.013786625117063522, + 0.013246742077171803, + 0.013632738031446934, + 0.0128323370590806, + 0.014133366756141186, + 0.011797457002103329, + 0.012874217703938484, + 0.013493889942765236, + 0.00806250050663948, + 0.009916579350829124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.17992311716079712 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206318736076355 + ], + "opposite_logprob_tokens": [ + -2.812063217163086 + ], + "logprob_reward": -0.06206318736076355, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206318736076355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12551189959049225, + 0.045663874596357346, + 0.05700794979929924, + 0.06432550400495529, + 0.02870328724384308, + 0.04434295743703842, + 0.4270908832550049, + 0.03900559991598129, + 0.09042984247207642, + 0.0779181569814682 + ], + "credit_map": [ + 0.01745808869600296, + 0.023184053599834442, + 0.02590222656726837, + 0.017220716923475266, + 0.013761045411229134, + 0.009306371212005615, + 0.010028869844973087, + 0.008650530129671097, + 0.005674850195646286, + 0.004047758411616087, + 0.008093050681054592, + 0.008210787549614906, + 0.0037014586851000786, + 0.005711376667022705, + 0.004118048120290041, + 0.006106548476964235, + 0.0062581016682088375, + 0.005315260495990515, + 0.007490630727261305, + 0.008650976233184338, + 0.006304502487182617, + 0.012129714712500572, + 0.005606387741863728, + 0.005252373404800892, + 0.007417047396302223, + 0.004822616931051016, + 0.01200934313237667, + 0.0063585685566067696, + 0.012074120342731476, + 0.007224533706903458, + 0.0077621000818908215, + 0.006657170597463846, + 0.0025861794129014015, + 0.0032212412916123867, + 0.0037330316845327616, + 0.0040876250714063644, + 0.005812210962176323, + 0.00313265947625041, + 0.0041422457434237, + 0.0019880926702171564, + 0.005941940937191248, + 0.00570321548730135, + 0.00846667867153883, + 0.005351027939468622, + 0.0069420961663126945, + 0.00409085676074028, + 0.004434718284755945, + 0.0034124215599149466, + 0.03301829844713211, + 0.14507904648780823, + 0.010304453782737255, + 0.019931050017476082, + 0.02494627609848976, + 0.1546105444431305, + 0.022792300209403038, + 0.016408948227763176, + 0.004191540647298098, + 0.008640211075544357, + 0.0042250677943229675, + 0.004360433667898178, + 0.0065561397932469845, + 0.00339142931625247, + 0.0038960082456469536, + 0.0037447724025696516, + 0.017193159088492393, + 0.018651572987437248, + 0.012651230208575726, + 0.009503070265054703, + 0.011137387715280056, + 0.008180497214198112, + 0.007648315746337175, + 0.005464612971991301, + 0.014373458921909332, + 0.014790123328566551, + 0.01176358200609684, + 0.007533674128353596, + 0.010856867767870426, + 0.006788460537791252, + 0.005448227282613516, + 0.0063637616112828255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.06206318736076355 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200471341609955 + ], + "opposite_logprob_tokens": [ + -4.264200687408447 + ], + "logprob_reward": -0.014200471341609955, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200471341609955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13296036422252655, + 0.036560267210006714, + 0.030420036986470222, + 0.031139208003878593, + 0.01289808563888073, + 0.014413970522582531, + 0.4720487892627716, + 0.027846943587064743, + 0.17465417087078094, + 0.06705820560455322 + ], + "credit_map": [ + 0.02140975557267666, + 0.01638488657772541, + 0.03145373612642288, + 0.02235831692814827, + 0.012315168976783752, + 0.009244512766599655, + 0.012454834766685963, + 0.007339161355048418, + 0.005019354168325663, + 0.003335326910018921, + 0.005188917741179466, + 0.0046180845238268375, + 0.0041367425583302975, + 0.0035872645676136017, + 0.0033079744316637516, + 0.007366601377725601, + 0.002750451909378171, + 0.0028986644465476274, + 0.004723335150629282, + 0.0040532490238547325, + 0.005217364989221096, + 0.004012002609670162, + 0.00308872084133327, + 0.003676247550174594, + 0.004215703811496496, + 0.00196249783039093, + 0.009199239313602448, + 0.003530150977894664, + 0.00302299577742815, + 0.003221847116947174, + 0.003248505527153611, + 0.0027382690459489822, + 0.0012952635297551751, + 0.001492304028943181, + 0.0018366548465564847, + 0.0021754237823188305, + 0.001922875875607133, + 0.001972324913367629, + 0.0013266807654872537, + 0.0008765570819377899, + 0.002020519459620118, + 0.0013885993976145983, + 0.001623131800442934, + 0.002033878117799759, + 0.001640173839405179, + 0.0017677528085187078, + 0.0018654137384146452, + 0.0020745026413351297, + 0.0374157652258873, + 0.16273722052574158, + 0.009585071355104446, + 0.022393887862563133, + 0.026709599420428276, + 0.17096459865570068, + 0.024433337152004242, + 0.017809297889471054, + 0.002637138357385993, + 0.004662328865379095, + 0.0032464300747960806, + 0.004117255564779043, + 0.00353636615909636, + 0.002775361528620124, + 0.004336514975875616, + 0.0025355478283017874, + 0.028338465839624405, + 0.026868285611271858, + 0.025304021313786507, + 0.020478203892707825, + 0.02340386062860489, + 0.017482565715909004, + 0.017296018078923225, + 0.015482746064662933, + 0.01050919946283102, + 0.012577056884765625, + 0.008897875435650349, + 0.0062743923626840115, + 0.009590347297489643, + 0.007555543445050716, + 0.00551304267719388, + 0.006140749901533127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.014200471341609955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232852790504694 + ], + "opposite_logprob_tokens": [ + -3.773285388946533 + ], + "logprob_reward": -0.0232852790504694, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232852790504694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25639432668685913, + 0.07785673439502716, + 0.024220077320933342, + 0.01950756274163723, + 0.021328436210751534, + 0.01898241974413395, + 0.29677310585975647, + 0.06833399832248688, + 0.1268211305141449, + 0.08978218585252762 + ], + "credit_map": [ + 0.03988061472773552, + 0.03516603633761406, + 0.061214104294776917, + 0.03950413689017296, + 0.02471558377146721, + 0.01705837994813919, + 0.017901897430419922, + 0.020953571423888206, + 0.011942539364099503, + 0.00947327446192503, + 0.01410831417888403, + 0.008855562657117844, + 0.016687843948602676, + 0.006430949084460735, + 0.006094371899962425, + 0.004263874609023333, + 0.0035964578855782747, + 0.0024279896169900894, + 0.003938309848308563, + 0.004868126008659601, + 0.0017364314990118146, + 0.0028896708972752094, + 0.0021695324685424566, + 0.0025935599114745855, + 0.0020106590818613768, + 0.0023274822160601616, + 0.0028236370999366045, + 0.0025072270072996616, + 0.002364855958148837, + 0.002503999276086688, + 0.0023104127030819654, + 0.002659290563315153, + 0.0021052479278296232, + 0.0029564187861979008, + 0.0028087503742426634, + 0.003843539860099554, + 0.0034781619906425476, + 0.0022538253106176853, + 0.002712985733523965, + 0.001169506344012916, + 0.003022518241778016, + 0.0023101840633898973, + 0.0029082326218485832, + 0.002128716791048646, + 0.0030197259038686752, + 0.0013979378854855895, + 0.0019737849943339825, + 0.0022213193587958813, + 0.023304475471377373, + 0.1022990494966507, + 0.006144057959318161, + 0.013646173290908337, + 0.017270147800445557, + 0.10721102356910706, + 0.015540164895355701, + 0.011357992887496948, + 0.013190198689699173, + 0.014060293324291706, + 0.006091342773288488, + 0.006181125994771719, + 0.007542678155004978, + 0.0116976797580719, + 0.004615341313183308, + 0.004955343436449766, + 0.020502354949712753, + 0.013852217234671116, + 0.016304250806570053, + 0.012250286526978016, + 0.019281094893813133, + 0.01547367312014103, + 0.014580617658793926, + 0.014576632529497147, + 0.014300120994448662, + 0.014637970365583897, + 0.017976850271224976, + 0.010673063807189465, + 0.008944674395024776, + 0.009545627050101757, + 0.007520922459661961, + 0.0061829607002437115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.0232852790504694 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot follows a two-wheeled bronze chariot, pulled by four majestic black horses, galloping across a vast, dusty battlefield. Atop the chariot sits a majestic, middle-aged emperor, poised as a rock. He wears an ancient imperial crown with dangling beads and a heavy, deep-black silk dragon robe, delicately embroidered with dazzling gold thread, depicting five-clawed golden dragons with bared fangs and distinct scales. The wide cuffs dance wildly in the wind. His face is stern, his gaze piercing, as he gazes straight ahead. One hand firmly grips the rough leather reins, steadying the neighing steed, while the other rests on the hilt of the ancient sword at his waist. The chariot itself is covered in ancient cloud-thunder patterns and deep battle scars, its bronze gleaming in the setting sun, a dark, cold luster. The chariot's massive wheels whirred at high speed, stirring up a cloud of yellow sand along with the horses' hooves, almost obscuring the silhouettes of the clashing armies and the burning beacon fires in the distance. The setting sun's golden afterglow pierced the dust, tinging the entire scene with a warm and tragic hue, outlining the emperor's resolute profile and the shimmering golden threads of his dragon robe. The entire scene was imbued with a sense of epic power and unstoppable royal momentum, creating a solemn and tense atmosphere.", + "seed": 1311, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_8/index_9559_seed_1311_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12027542404830456, + "vlm_logprob": -0.12027542404830456, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.12027542404830456, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11560703068971634 + ], + "opposite_logprob_tokens": [ + -2.6156070232391357 + ], + "logprob_reward": -0.11560703068971634, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.11560703068971634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10436758399009705, + 0.10426709800958633, + 0.04320110008120537, + 0.055350687354803085, + 0.059678252786397934, + 0.10168184340000153, + 0.0583648681640625, + 0.09280546009540558, + 0.07590091973543167, + 0.3043821454048157 + ], + "credit_map": [ + 0.015309828333556652, + 0.009194724261760712, + 0.015410187654197216, + 0.010675855912268162, + 0.008500495925545692, + 0.015164661221206188, + 0.012946479953825474, + 0.017165344208478928, + 0.01370738074183464, + 0.013281642459332943, + 0.018245864659547806, + 0.014611576683819294, + 0.009291199035942554, + 0.009444129653275013, + 0.011774694547057152, + 0.013910611160099506, + 0.0034889073576778173, + 0.0041267056949436665, + 0.005973717663437128, + 0.010409935377538204, + 0.0025671999901533127, + 0.005465172231197357, + 0.005811427254229784, + 0.005358037073165178, + 0.003838254138827324, + 0.00976677518337965, + 0.009497618302702904, + 0.008164319209754467, + 0.004127679858356714, + 0.007984619587659836, + 0.0044371639378368855, + 0.007534255273640156, + 0.004687595181167126, + 0.008497067727148533, + 0.009125799871981144, + 0.009070673026144505, + 0.0061603025533258915, + 0.005706604570150375, + 0.005538562312722206, + 0.01089165173470974, + 0.011378870345652103, + 0.010442058555781841, + 0.016820121556520462, + 0.017720269039273262, + 0.009228111244738102, + 0.010759555734694004, + 0.011207056231796741, + 0.014125807210803032, + 0.007526285480707884, + 0.005485400557518005, + 0.011744601652026176, + 0.006943305488675833, + 0.006263003684580326, + 0.007950284518301487, + 0.007394287269562483, + 0.005057699978351593, + 0.010270900093019009, + 0.007982204668223858, + 0.019928928464651108, + 0.014935978688299656, + 0.009256341494619846, + 0.011132546700537205, + 0.009281566366553307, + 0.010016999207437038, + 0.015273557044565678, + 0.004558630753308535, + 0.021130839362740517, + 0.012289421632885933, + 0.004599688109010458, + 0.004960430320352316, + 0.005933412816375494, + 0.007154939230531454, + 0.06757831573486328, + 0.020373007282614708, + 0.08610264211893082, + 0.04217701032757759, + 0.02188628539443016, + 0.02115771919488907, + 0.025100884959101677, + 0.020006263628602028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.11560703068971634 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38710305094718933 + ], + "opposite_logprob_tokens": [ + -1.1371030807495117 + ], + "logprob_reward": -0.38710305094718933, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38710305094718933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46693679690361023, + 0.14989173412322998, + 0.018186647444963455, + 0.08188018947839737, + 0.07443742454051971, + 0.04435238242149353, + 0.03951635956764221, + 0.046164464205503464, + 0.028341487050056458, + 0.05029254034161568 + ], + "credit_map": [ + 0.027051741257309914, + 0.016888419166207314, + 0.12144098430871964, + 0.01847654953598976, + 0.0615825392305851, + 0.17635369300842285, + 0.020348938181996346, + 0.024793943390250206, + 0.01587144285440445, + 0.021236078813672066, + 0.018512314185500145, + 0.023068031296133995, + 0.010285698808729649, + 0.015565370209515095, + 0.020728617906570435, + 0.024624180048704147, + 0.0032707895152270794, + 0.0027677963953465223, + 0.0034310477785766125, + 0.0018709837459027767, + 0.0012549922103062272, + 0.0018980297027155757, + 0.0012556705623865128, + 0.002437337301671505, + 0.005253246054053307, + 0.00690118782222271, + 0.015515218488872051, + 0.00901392288506031, + 0.007224437780678272, + 0.01929105632007122, + 0.0069560157135128975, + 0.0117251081392169, + 0.005286832340061665, + 0.008933098055422306, + 0.010447504930198193, + 0.008080419152975082, + 0.006824418902397156, + 0.007165631279349327, + 0.010753735899925232, + 0.01694578118622303, + 0.0033176783472299576, + 0.004194885026663542, + 0.008217657916247845, + 0.007220305968075991, + 0.005265512969344854, + 0.004618460778146982, + 0.005183117464184761, + 0.006334760691970587, + 0.004122545011341572, + 0.003064002376049757, + 0.006940095219761133, + 0.00818705279380083, + 0.00397831853479147, + 0.0033451758790761232, + 0.0054776230826973915, + 0.004401549696922302, + 0.004777860827744007, + 0.003933685831725597, + 0.0075096189975738525, + 0.009289048612117767, + 0.006626140791922808, + 0.004866699688136578, + 0.004870166536420584, + 0.004291245248168707, + 0.002486195880919695, + 0.0015861961292102933, + 0.003540879115462303, + 0.008397424593567848, + 0.0025676318909972906, + 0.002849260810762644, + 0.003673501079902053, + 0.003240399295464158, + 0.008859304711222649, + 0.003276909003034234, + 0.010838944464921951, + 0.006313606631010771, + 0.0045668440870940685, + 0.007528672460466623, + 0.0049963900819420815, + 0.00391186960041523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.38710305094718933 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700443834066391 + ], + "opposite_logprob_tokens": [ + -2.6950442790985107 + ], + "logprob_reward": -0.0700443834066391, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700443834066391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3662034273147583, + 0.09698303788900375, + 0.04855048656463623, + 0.1610756814479828, + 0.0768541693687439, + 0.043038737028837204, + 0.03358021378517151, + 0.028233736753463745, + 0.03596377745270729, + 0.10951676964759827 + ], + "credit_map": [ + 0.019412878900766373, + 0.012087468057870865, + 0.09249431639909744, + 0.015425286255776882, + 0.04852164164185524, + 0.14087368547916412, + 0.01752166450023651, + 0.0198664627969265, + 0.01664789579808712, + 0.008940305560827255, + 0.014818002469837666, + 0.014249568805098534, + 0.01147667970508337, + 0.010898231528699398, + 0.010641660541296005, + 0.009310695342719555, + 0.004437886644154787, + 0.007841160520911217, + 0.008436983451247215, + 0.006632269825786352, + 0.0034895355347543955, + 0.007456424180418253, + 0.0063648228533566, + 0.0038914023898541927, + 0.009849592112004757, + 0.016944188624620438, + 0.03387283533811569, + 0.018791314214468002, + 0.013346046209335327, + 0.034870076924562454, + 0.012818711809813976, + 0.020582908764481544, + 0.005102000664919615, + 0.005863265134394169, + 0.01151115819811821, + 0.009576559998095036, + 0.005265528801828623, + 0.007600628770887852, + 0.01175027433782816, + 0.020184753462672234, + 0.00488708121702075, + 0.005010594148188829, + 0.007619587704539299, + 0.007643871009349823, + 0.004662344232201576, + 0.0043912651017308235, + 0.004981005564332008, + 0.0038429838605225086, + 0.0038859304040670395, + 0.0027844151481986046, + 0.004834282677620649, + 0.0056879702024161816, + 0.003797811921685934, + 0.002476243767887354, + 0.006910096853971481, + 0.0032034628093242645, + 0.0030716657638549805, + 0.002643016166985035, + 0.00408966327086091, + 0.004529683850705624, + 0.0038401386700570583, + 0.0027774451300501823, + 0.0037600197829306126, + 0.0035221048165112734, + 0.004237924236804247, + 0.0021579761523753405, + 0.007532958872616291, + 0.006102783139795065, + 0.0033356528729200363, + 0.0045270309783518314, + 0.00483185239136219, + 0.00323759694583714, + 0.02490372769534588, + 0.006203794851899147, + 0.029481276869773865, + 0.012332630343735218, + 0.008580640889704227, + 0.008414887823164463, + 0.009174725972115993, + 0.010425087995827198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0700443834066391 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012549009174108505 + ], + "opposite_logprob_tokens": [ + -4.387548923492432 + ], + "logprob_reward": -0.012549009174108505, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012549009174108505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5108067393302917, + 0.06658618152141571, + 0.03514283895492554, + 0.03683485463261604, + 0.08490896970033646, + 0.04192914068698883, + 0.04334822669625282, + 0.026634996756911278, + 0.11341255903244019, + 0.040395479649305344 + ], + "credit_map": [ + 0.028610143810510635, + 0.015802746638655663, + 0.13718414306640625, + 0.021443486213684082, + 0.06727689504623413, + 0.1917111575603485, + 0.022693337872624397, + 0.026084834709763527, + 0.009973649866878986, + 0.007403313647955656, + 0.008980615995824337, + 0.01211415696889162, + 0.008192634209990501, + 0.007046239450573921, + 0.00753772770985961, + 0.005337846465408802, + 0.00250771502032876, + 0.0057403333485126495, + 0.006644377484917641, + 0.00850518885999918, + 0.0017011752352118492, + 0.0029974631033837795, + 0.004249550402164459, + 0.0027970371302217245, + 0.0029716340359300375, + 0.004798113834112883, + 0.009112421423196793, + 0.0046853842213749886, + 0.0035386967938393354, + 0.004726179409772158, + 0.002844849368557334, + 0.004157575313001871, + 0.005471731070429087, + 0.007571193855255842, + 0.011824702844023705, + 0.01062396727502346, + 0.0055739860981702805, + 0.008544971235096455, + 0.011066812090575695, + 0.024231603369116783, + 0.004760520067065954, + 0.004317824263125658, + 0.006850806996226311, + 0.006150631234049797, + 0.0036976581905037165, + 0.004478028044104576, + 0.003729494521394372, + 0.007944179698824883, + 0.004871890880167484, + 0.003598621813580394, + 0.009531470015645027, + 0.006289870943874121, + 0.005012974608689547, + 0.005420721136033535, + 0.005281701218336821, + 0.0033409767784178257, + 0.002719312207773328, + 0.0027760579250752926, + 0.005592776462435722, + 0.005947416182607412, + 0.00304482807405293, + 0.0018381038680672646, + 0.002162875374779105, + 0.0025536271277815104, + 0.020287884399294853, + 0.006521265488117933, + 0.0410306379199028, + 0.012181154452264309, + 0.008138536475598812, + 0.00947309099137783, + 0.009947889484465122, + 0.005832093767821789, + 0.005999907851219177, + 0.0034787512850016356, + 0.006190785206854343, + 0.007531946524977684, + 0.005359348841011524, + 0.004368648398667574, + 0.004732748027890921, + 0.0027333449106663465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.012549009174108505 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a tracking shot following a two-wheeled bronze chariot pulled by four black horses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a middle-aged emperor wearing a deep-black silk dragon robe with gold embroidery visible atop the chariot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the chariot covered in ancient cloud-thunder patterns and battle scars while stirring up a cloud of yellow sand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by a setting sun that tings the environment with a warm golden afterglow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073646023869514 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073646023869514, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073646023869514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6485422253608704, + 0.041937053203582764, + 0.038814157247543335, + 0.05016264319419861, + 0.0447363518178463, + 0.05546128749847412, + 0.024334030225872993, + 0.021027591079473495, + 0.03637295961380005, + 0.03861168399453163 + ], + "credit_map": [ + 0.03590324893593788, + 0.022789519280195236, + 0.17133240401744843, + 0.02785838395357132, + 0.08474257588386536, + 0.2409043163061142, + 0.02954624406993389, + 0.03546551242470741, + 0.005838583689182997, + 0.006206432357430458, + 0.005629677791148424, + 0.005084950476884842, + 0.0040126703679561615, + 0.003976861014962196, + 0.0051154145039618015, + 0.0060724616050720215, + 0.003057868452742696, + 0.007137846201658249, + 0.005705381277948618, + 0.010206669569015503, + 0.002033699071034789, + 0.0040756226517260075, + 0.00303850625641644, + 0.003558560274541378, + 0.0046192253939807415, + 0.005246937274932861, + 0.007756001316010952, + 0.004734023939818144, + 0.004297642968595028, + 0.012989403679966927, + 0.003545842133462429, + 0.006973563227802515, + 0.0038494467735290527, + 0.00688133854418993, + 0.005925226490944624, + 0.0054138461127877235, + 0.005157220643013716, + 0.005051870364695787, + 0.004450426436960697, + 0.008006979711353779, + 0.006424570456147194, + 0.004869701806455851, + 0.016622347757220268, + 0.006607442628592253, + 0.0047434307634830475, + 0.007724029943346977, + 0.004975815303623676, + 0.003493951167911291, + 0.0029689837247133255, + 0.0024541416205465794, + 0.005936095491051674, + 0.0025002469774335623, + 0.0029813882429152727, + 0.0026253298856317997, + 0.003310818923637271, + 0.0015570261748507619, + 0.0024773685727268457, + 0.0016944187227636576, + 0.003487229812890291, + 0.002379112411290407, + 0.0025291377678513527, + 0.002216954715549946, + 0.0031345379538834095, + 0.003108829725533724, + 0.004459809977561235, + 0.0024509329814463854, + 0.0077141523361206055, + 0.006706533022224903, + 0.002668709261342883, + 0.004234292544424534, + 0.004145327024161816, + 0.00399320013821125, + 0.00562285864725709, + 0.003019417868927121, + 0.004759832285344601, + 0.00592678552493453, + 0.004450351931154728, + 0.008045373484492302, + 0.003560320008546114, + 0.003226745640859008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.016073646023869514 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_0/index_tensor([8451, 8451])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_0/index_tensor([8451, 8451])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..147803703a7bb9c98555732bceeb0f1860ae7fa6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_0/index_tensor([8451, 8451])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot from a slightly upward angle, the camera held still, focusing on a queen. A magnificent, heavy golden crown, inlaid with crimson gems and intricate baroque carvings, weighed heavily on her forehead. A few cascading strands of raven hair cascaded from the crown's edge, draping around her pale, sharply defined face. She wore a deep purple velvet gown with intricate silver embroidery at the collar, which shimmered faintly in the dim light. Her setting was a grand but dimly lit Gothic palace hall, the only faint light coming from the cool moonlight filtering through the distant high windows and the flickering flames of a few candles in silver candlesticks beside her, casting long, twisted shadows on the massive stone pillars in the background. At first, her eyes were closed, her face calm, yet a subtle hint of fatigue lingered. Then, slowly, with an almost palpable weight, she lifted her long lashes. The warm candlelight traced a soft golden edge along the contours of her cheeks. When her eyes fully opened, a pair of icy blue, hawk-like eyes were revealed. Their gaze, without a trace of hesitation, pierced the camera directly and resolutely. The entire scene, interwoven with the cold moonlight and the warm candlelight, was filled with a solemn, awe-inspiring, and powerful atmosphere, as if a silent storm was brewing in the midst of her decision.", + "seed": 20629, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_0/index_8451_seed_20629_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.24833909571170806, + "vlm_logprob": -0.24833909571170806, + "vlm_yes_no_margin": 1.625, + "vlm_token_credit": -0.24833909571170806, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28670769929885864 + ], + "opposite_logprob_tokens": [ + -1.4117076396942139 + ], + "logprob_reward": -0.28670769929885864, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28670769929885864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2113519012928009, + 0.0481438934803009, + 0.06678485870361328, + 0.027704782783985138, + 0.042757075279951096, + 0.021173272281885147, + 0.4198265075683594, + 0.04595582187175751, + 0.021980181336402893, + 0.09432167559862137 + ], + "credit_map": [ + 0.020020371302962303, + 0.024018099531531334, + 0.02139974944293499, + 0.023025190457701683, + 0.037286289036273956, + 0.025478363037109375, + 0.03414245322346687, + 0.025981400161981583, + 0.00431782566010952, + 0.004107065964490175, + 0.004154583904892206, + 0.00523204542696476, + 0.007262933999300003, + 0.00935886800289154, + 0.008381562307476997, + 0.005329008214175701, + 0.008510704152286053, + 0.013446447439491749, + 0.005973769351840019, + 0.006574295926839113, + 0.010317947715520859, + 0.007991780526936054, + 0.00842808187007904, + 0.005541835445910692, + 0.0035078884102404118, + 0.0024280452635139227, + 0.003996803425252438, + 0.0028971850406378508, + 0.004003418143838644, + 0.003951983526349068, + 0.003098610555753112, + 0.003820847487077117, + 0.004857285879552364, + 0.005733449012041092, + 0.004625624977052212, + 0.0032858361955732107, + 0.006287849973887205, + 0.010229124687612057, + 0.0035665356554090977, + 0.004171372391283512, + 0.0023596195969730616, + 0.001669298973865807, + 0.002274682279676199, + 0.002414500107988715, + 0.0035195646341890097, + 0.003234982956200838, + 0.003289669519290328, + 0.002410955261439085, + 0.05239555239677429, + 0.07103081792593002, + 0.05259134992957115, + 0.04461165890097618, + 0.06214514747262001, + 0.06888511776924133, + 0.03868293762207031, + 0.029483938589692116, + 0.006663442123681307, + 0.003723717527464032, + 0.00583950849249959, + 0.0046373954974114895, + 0.0070238918997347355, + 0.006795243360102177, + 0.005743795074522495, + 0.005528830923140049, + 0.0019941639620810747, + 0.0030477719847112894, + 0.002779556903988123, + 0.002490090671926737, + 0.002811005339026451, + 0.0038072909228503704, + 0.0029314011335372925, + 0.0021189020480960608, + 0.010663272812962532, + 0.01964794099330902, + 0.010455157607793808, + 0.010579121299088001, + 0.01394728384912014, + 0.011983620002865791, + 0.011102980934083462, + 0.005942291114479303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.28670769929885864 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1604248732328415 + ], + "opposite_logprob_tokens": [ + -1.910424828529358 + ], + "logprob_reward": -0.1604248732328415, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1604248732328415, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08900458365678787, + 0.03359637409448624, + 0.08916330337524414, + 0.05056407302618027, + 0.05122246965765953, + 0.024295972660183907, + 0.5069004893302917, + 0.012093729339540005, + 0.012588308192789555, + 0.13057072460651398 + ], + "credit_map": [ + 0.009905503131449223, + 0.011813675984740257, + 0.008598929271101952, + 0.010103351436555386, + 0.01602335087954998, + 0.009640690870583057, + 0.012533726170659065, + 0.010385346598923206, + 0.005289741791784763, + 0.0036345007829368114, + 0.0036398833617568016, + 0.003782428102567792, + 0.003953668754547834, + 0.004738070536404848, + 0.0038999877870082855, + 0.004658093210309744, + 0.010876079089939594, + 0.018026160076260567, + 0.008052457123994827, + 0.008304590359330177, + 0.01499200239777565, + 0.010427646338939667, + 0.010778236202895641, + 0.007706123869866133, + 0.006220564246177673, + 0.006727362051606178, + 0.007298896089196205, + 0.006037638522684574, + 0.006145140156149864, + 0.007972907274961472, + 0.00519909942522645, + 0.00496246013790369, + 0.006403859704732895, + 0.006685086525976658, + 0.005625681485980749, + 0.0041435169987380505, + 0.0072948625311255455, + 0.012956102378666401, + 0.003705891314893961, + 0.004407460801303387, + 0.0024813516065478325, + 0.001586859580129385, + 0.0027722339145839214, + 0.0020343640353530645, + 0.0032519306987524033, + 0.006672551389783621, + 0.002821213100105524, + 0.002675465540960431, + 0.06609171628952026, + 0.08621856570243835, + 0.06106427311897278, + 0.05216522514820099, + 0.07377833873033524, + 0.08654201030731201, + 0.04526612535119057, + 0.03577416390180588, + 0.0018186146626248956, + 0.0009935385314747691, + 0.0018809180473908782, + 0.0009173729340545833, + 0.001697428640909493, + 0.0017591953510418534, + 0.0014796274481341243, + 0.0015470318030565977, + 0.0014571513747796416, + 0.0016853734850883484, + 0.0015567925292998552, + 0.0012302389368414879, + 0.0017364665400236845, + 0.0023781859781593084, + 0.0012448850320652127, + 0.0012992130359634757, + 0.014588659629225731, + 0.031137334182858467, + 0.013652716763317585, + 0.010024144314229488, + 0.021323032677173615, + 0.01764424331486225, + 0.012701497413218021, + 0.009499086067080498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 2 + ], + "selected_reward": -0.1604248732328415 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11300314962863922 + ], + "opposite_logprob_tokens": [ + -2.2380032539367676 + ], + "logprob_reward": -0.11300314962863922, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11300314962863922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1546282321214676, + 0.07870446890592575, + 0.05211872234940529, + 0.05576956272125244, + 0.05378665030002594, + 0.034278687089681625, + 0.4141600728034973, + 0.031303565949201584, + 0.03133951127529144, + 0.09391055256128311 + ], + "credit_map": [ + 0.01338366698473692, + 0.010456793010234833, + 0.019561396911740303, + 0.017811261117458344, + 0.02447257563471794, + 0.020504849031567574, + 0.023562613874673843, + 0.02487507462501526, + 0.00837083999067545, + 0.011202193796634674, + 0.00923804473131895, + 0.00867701880633831, + 0.009443875402212143, + 0.014430331997573376, + 0.008467313833534718, + 0.008874847553670406, + 0.007938870228827, + 0.009647198021411896, + 0.0048171221278607845, + 0.005710850935429335, + 0.007017431315034628, + 0.00552002526819706, + 0.005792527459561825, + 0.005674694199115038, + 0.0066815707832574844, + 0.007843350060284138, + 0.006931604351848364, + 0.0057238806039094925, + 0.007015768438577652, + 0.01102253794670105, + 0.005767658818513155, + 0.0047831907868385315, + 0.00606196653097868, + 0.006433118134737015, + 0.005621355026960373, + 0.004273237194865942, + 0.007372942753136158, + 0.014580757357180119, + 0.004238532856106758, + 0.0052047367207705975, + 0.0035029512364417315, + 0.0026779042091220617, + 0.0044876569882035255, + 0.0025029205717146397, + 0.0034789268393069506, + 0.011308000423014164, + 0.002551410812884569, + 0.003768918802961707, + 0.057715583592653275, + 0.06855874508619308, + 0.047036077827215195, + 0.04046371206641197, + 0.0594252273440361, + 0.07378192245960236, + 0.035553496330976486, + 0.03162529692053795, + 0.004684042651206255, + 0.0023192521184682846, + 0.0037563173100352287, + 0.0023782101925462484, + 0.003231616457924247, + 0.005451779346913099, + 0.0045387898571789265, + 0.004943558946251869, + 0.003202371532097459, + 0.005539226811379194, + 0.0031478183809667826, + 0.0028831884264945984, + 0.005737006664276123, + 0.004813906736671925, + 0.0030766783747822046, + 0.002939314115792513, + 0.0103157302364707, + 0.021197617053985596, + 0.00980356428772211, + 0.00803216639906168, + 0.015480438247323036, + 0.012782033532857895, + 0.009104377590119839, + 0.007194631267338991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.11300314962863922 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328769326210022 + ], + "opposite_logprob_tokens": [ + -0.7578769326210022 + ], + "logprob_reward": -0.6328769326210022, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328769326210022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10942541062831879, + 0.03432749956846237, + 0.054501112550497055, + 0.045382872223854065, + 0.04046352580189705, + 0.06598825752735138, + 0.35807913541793823, + 0.14502272009849548, + 0.06941995769739151, + 0.07738952338695526 + ], + "credit_map": [ + 0.01427683886140585, + 0.008387215435504913, + 0.01484221313148737, + 0.009473719634115696, + 0.021251114085316658, + 0.012163850478827953, + 0.01772794872522354, + 0.011302517727017403, + 0.0037076284643262625, + 0.0034849103540182114, + 0.003761457046493888, + 0.003941035829484463, + 0.005588205996900797, + 0.005355685483664274, + 0.004819865804165602, + 0.0036687094252556562, + 0.005868745036423206, + 0.010463528335094452, + 0.005140058696269989, + 0.005010074004530907, + 0.00863114558160305, + 0.007708040066063404, + 0.006213082000613213, + 0.005466436967253685, + 0.004920437466353178, + 0.0050697894766926765, + 0.005084760952740908, + 0.003421284956857562, + 0.006568118464201689, + 0.011279802769422531, + 0.0043788934126496315, + 0.004659782629460096, + 0.004448303021490574, + 0.004513473715633154, + 0.004078570753335953, + 0.0029926998540759087, + 0.005185725633054972, + 0.011403284966945648, + 0.0035842612851411104, + 0.00425720727071166, + 0.006198531948029995, + 0.0064896754920482635, + 0.008058945648372173, + 0.004667588975280523, + 0.009516277350485325, + 0.017830953001976013, + 0.006659689359366894, + 0.0065665896981954575, + 0.04857367277145386, + 0.061636362224817276, + 0.04228319227695465, + 0.03581739589571953, + 0.051634933799505234, + 0.059909965842962265, + 0.03292391821742058, + 0.025299692526459694, + 0.016632115468382835, + 0.01325917337089777, + 0.016491077840328217, + 0.016177024692296982, + 0.023060083389282227, + 0.024396274238824844, + 0.01896536350250244, + 0.016041606664657593, + 0.008326892741024494, + 0.004523963667452335, + 0.008461598306894302, + 0.010275406762957573, + 0.007942796684801579, + 0.011364854872226715, + 0.009381270967423916, + 0.00914317648857832, + 0.008669854141771793, + 0.01814749650657177, + 0.008488956838846207, + 0.006506255362182856, + 0.012436906807124615, + 0.01009173970669508, + 0.007663029711693525, + 0.005385279655456543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.6328769326210022 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868282377719879 + ], + "opposite_logprob_tokens": [ + -3.048682928085327 + ], + "logprob_reward": -0.04868282377719879, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868282377719879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15455728769302368, + 0.07252947986125946, + 0.07118073850870132, + 0.06822247803211212, + 0.11007668077945709, + 0.061907101422548294, + 0.3098808228969574, + 0.044637519866228104, + 0.06448934972286224, + 0.042518503963947296 + ], + "credit_map": [ + 0.019325001165270805, + 0.013896561227738857, + 0.020051531493663788, + 0.018850412219762802, + 0.023522112518548965, + 0.020757019519805908, + 0.018659984692931175, + 0.01949464902281761, + 0.008334707468748093, + 0.008241888135671616, + 0.009162954986095428, + 0.006624644622206688, + 0.007629411295056343, + 0.015509241260588169, + 0.008145533502101898, + 0.008881094865500927, + 0.00956298504024744, + 0.008189103566110134, + 0.0074559906497597694, + 0.006800960749387741, + 0.009072921238839626, + 0.011925514787435532, + 0.0075932457111775875, + 0.010580011643469334, + 0.008760364726185799, + 0.012866292148828506, + 0.008540446870028973, + 0.006512809544801712, + 0.008730888366699219, + 0.01092070434242487, + 0.006346757989376783, + 0.005544210784137249, + 0.019001172855496407, + 0.010157297365367413, + 0.01346761267632246, + 0.007858305238187313, + 0.00942141655832529, + 0.026886027306318283, + 0.009769315831363201, + 0.013515538536012173, + 0.010874554514884949, + 0.006260352209210396, + 0.006958732847124338, + 0.003942352253943682, + 0.005398572888225317, + 0.015764962881803513, + 0.0053092436864972115, + 0.007398330606520176, + 0.04164991155266762, + 0.054149042814970016, + 0.033591851592063904, + 0.031081296503543854, + 0.0471058264374733, + 0.05210711434483528, + 0.026548877358436584, + 0.023646896705031395, + 0.006097306497395039, + 0.002903963206335902, + 0.005151825025677681, + 0.005158178973942995, + 0.0054620010778307915, + 0.008310521952807903, + 0.0052601308561861515, + 0.006293590646237135, + 0.007099557667970657, + 0.009026549756526947, + 0.006895661819726229, + 0.008215589448809624, + 0.010799425654113293, + 0.00835573859512806, + 0.0073496331460773945, + 0.006747192237526178, + 0.004917693790048361, + 0.006623825524002314, + 0.005455044563859701, + 0.003873060457408428, + 0.004849523305892944, + 0.007437406573444605, + 0.003626295132562518, + 0.005735654849559069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.04868282377719879 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward angle, the camera held still, focusing on a queen. A magnificent, heavy golden crown, inlaid with crimson gems and intricate baroque carvings, weighed heavily on her forehead. A few cascading strands of raven hair cascaded from the crown's edge, draping around her pale, sharply defined face. She wore a deep purple velvet gown with intricate silver embroidery at the collar, which shimmered faintly in the dim light. Her setting was a grand but dimly lit Gothic palace hall, the only faint light coming from the cool moonlight filtering through the distant high windows and the flickering flames of a few candles in silver candlesticks beside her, casting long, twisted shadows on the massive stone pillars in the background. At first, her eyes were closed, her face calm, yet a subtle hint of fatigue lingered. Then, slowly, with an almost palpable weight, she lifted her long lashes. The warm candlelight traced a soft golden edge along the contours of her cheeks. When her eyes fully opened, a pair of icy blue, hawk-like eyes were revealed. Their gaze, without a trace of hesitation, pierced the camera directly and resolutely. The entire scene, interwoven with the cold moonlight and the warm candlelight, was filled with a solemn, awe-inspiring, and powerful atmosphere, as if a silent storm was brewing in the midst of her decision.", + "seed": 20630, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_0/index_8451_seed_20630_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.049451363272964954, + "vlm_logprob": -0.049451363272964954, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.049451363272964954, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019471338018774986 + ], + "opposite_logprob_tokens": [ + -4.269471168518066 + ], + "logprob_reward": -0.019471338018774986, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.019471338018774986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25735774636268616, + 0.034284334629774094, + 0.1793482005596161, + 0.03219433128833771, + 0.021569596603512764, + 0.0441446453332901, + 0.06747206300497055, + 0.059628184884786606, + 0.08016923069953918, + 0.2238316535949707 + ], + "credit_map": [ + 0.026313072070479393, + 0.022667881101369858, + 0.03161545470356941, + 0.03329949080944061, + 0.054425276815891266, + 0.027001235634088516, + 0.026309890672564507, + 0.03572545945644379, + 0.0031014708802103996, + 0.0027988317888230085, + 0.0033126375637948513, + 0.004296326544135809, + 0.008173356764018536, + 0.0032152042258530855, + 0.0040385667234659195, + 0.005347942002117634, + 0.014718987047672272, + 0.006142435595393181, + 0.018332675099372864, + 0.06891299784183502, + 0.009179801680147648, + 0.006020957604050636, + 0.007206798065453768, + 0.04883355647325516, + 0.0031201024539768696, + 0.002299358369782567, + 0.0031161389779299498, + 0.003855552291497588, + 0.007985623553395271, + 0.003042757511138916, + 0.0032931831665337086, + 0.005481617525219917, + 0.00279815006069839, + 0.0014126099413260818, + 0.002682174788787961, + 0.0022625622805207968, + 0.004165694583207369, + 0.0016650038305670023, + 0.002341430401429534, + 0.004241973627358675, + 0.005313172936439514, + 0.0036611519753932953, + 0.004825135227292776, + 0.005072007887065411, + 0.0066668190993368626, + 0.0052464809268713, + 0.0072952876798808575, + 0.006064592860639095, + 0.006130611523985863, + 0.0048937685787677765, + 0.01114529836922884, + 0.008800565265119076, + 0.009060142561793327, + 0.0060144150629639626, + 0.011462449096143246, + 0.009964818134903908, + 0.0065209995955228806, + 0.005910644773393869, + 0.009944252669811249, + 0.007065491750836372, + 0.005480748135596514, + 0.005863737780600786, + 0.008095241151750088, + 0.010747072286903858, + 0.0058797006495296955, + 0.013634009286761284, + 0.008377972058951855, + 0.007605390623211861, + 0.022413786500692368, + 0.0038087056018412113, + 0.0063246251083910465, + 0.012125054374337196, + 0.024185121059417725, + 0.0476330891251564, + 0.025285711511969566, + 0.021084874868392944, + 0.03913651034235954, + 0.01652926579117775, + 0.0209323912858963, + 0.029044710099697113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.019471338018774986 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012568551115691662 + ], + "opposite_logprob_tokens": [ + -4.387568473815918 + ], + "logprob_reward": -0.012568551115691662, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012568551115691662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12976112961769104, + 0.013797277584671974, + 0.11008965224027634, + 0.017221035435795784, + 0.013171506114304066, + 0.011767333373427391, + 0.013332556001842022, + 0.010705562308430672, + 0.36138585209846497, + 0.31876808404922485 + ], + "credit_map": [ + 0.013135344721376896, + 0.011871203780174255, + 0.01824192702770233, + 0.016529899090528488, + 0.027365073561668396, + 0.013500274159014225, + 0.012830538675189018, + 0.016286857426166534, + 0.0013677813112735748, + 0.0012909468496218324, + 0.0018645667005330324, + 0.0016179617960005999, + 0.002549470169469714, + 0.0013614118797704577, + 0.0018744069384410977, + 0.001870731357485056, + 0.00957274530082941, + 0.0036607482470571995, + 0.011766841635107994, + 0.04398494213819504, + 0.003947567194700241, + 0.003316650167107582, + 0.00457540201023221, + 0.029264749959111214, + 0.0014953602803871036, + 0.0012466266052797437, + 0.0018300513038411736, + 0.0018707327544689178, + 0.004243162460625172, + 0.0019886589143425226, + 0.001802946673706174, + 0.002743495162576437, + 0.0015759491361677647, + 0.0009626057581044734, + 0.001919597969390452, + 0.0017739643808454275, + 0.002385735744610429, + 0.0011412168387323618, + 0.0012126523070037365, + 0.002199782757088542, + 0.0013751696096733212, + 0.0008418665383942425, + 0.0015879619168117642, + 0.0012206026585772634, + 0.0017070688772946596, + 0.0008560580317862332, + 0.0016452023992314935, + 0.0025334034580737352, + 0.001592501299455762, + 0.0011899853125214577, + 0.002186940750107169, + 0.0015348674496635795, + 0.0016372199170291424, + 0.001313474727794528, + 0.0016424021450802684, + 0.0022351618390530348, + 0.0007910333224572241, + 0.0014496363000944257, + 0.0013217795640230179, + 0.0015059353318065405, + 0.0020327402744442225, + 0.0010214474750682712, + 0.0012993752025067806, + 0.0012836140813305974, + 0.025973303243517876, + 0.06597892940044403, + 0.031921904534101486, + 0.03755846619606018, + 0.10246778279542923, + 0.013466139324009418, + 0.030603813007473946, + 0.053415462374687195, + 0.03577839955687523, + 0.06630750000476837, + 0.03850109502673149, + 0.02913006953895092, + 0.05376873537898064, + 0.02365272119641304, + 0.029939958825707436, + 0.041689589619636536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.012568551115691662 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899660617113113 + ], + "opposite_logprob_tokens": [ + -2.5789966583251953 + ], + "logprob_reward": -0.07899660617113113, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899660617113113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17282724380493164, + 0.0321209654211998, + 0.5633719563484192, + 0.020886389538645744, + 0.017765434458851814, + 0.019866587594151497, + 0.024190951138734818, + 0.03732811659574509, + 0.05197295919060707, + 0.059669461101293564 + ], + "credit_map": [ + 0.014820603653788567, + 0.014437449164688587, + 0.02014266885817051, + 0.029162826016545296, + 0.015132570639252663, + 0.011732148937880993, + 0.018396496772766113, + 0.04900246113538742, + 0.0023942107800394297, + 0.001495435368269682, + 0.004737282637506723, + 0.004821395501494408, + 0.00368325412273407, + 0.0024241635110229254, + 0.0035571667831391096, + 0.009008053690195084, + 0.04820677638053894, + 0.01868082582950592, + 0.059028260409832, + 0.22557087242603302, + 0.01741652935743332, + 0.016542883589863777, + 0.021710705012083054, + 0.15621505677700043, + 0.0019502410432323813, + 0.0014333397848531604, + 0.0027508833445608616, + 0.0030940540600568056, + 0.0027901348657906055, + 0.0015845594462007284, + 0.0024970537051558495, + 0.004786121193319559, + 0.002153925132006407, + 0.001189754344522953, + 0.0019963192753493786, + 0.0024426307063549757, + 0.003208283567801118, + 0.001226835884153843, + 0.0020645938348025084, + 0.0034830898512154818, + 0.0025372453965246677, + 0.0013539348728954792, + 0.002780855167657137, + 0.002716290531679988, + 0.002157457871362567, + 0.0014531624037772417, + 0.0028125580865889788, + 0.0040550814010202885, + 0.002346112858504057, + 0.0023439398501068354, + 0.0027145622298121452, + 0.0034580593928694725, + 0.004131478723138571, + 0.0022425209172070026, + 0.0028334991075098515, + 0.004120775032788515, + 0.003814237890765071, + 0.004257846623659134, + 0.006891967263072729, + 0.004104319028556347, + 0.0043466174975037575, + 0.003714537713676691, + 0.003659622510895133, + 0.0065389652736485004, + 0.003645814722403884, + 0.009746504947543144, + 0.004586884286254644, + 0.00544148962944746, + 0.014646206982433796, + 0.0018857994582504034, + 0.004185682628303766, + 0.007834571413695812, + 0.007615603972226381, + 0.01122664287686348, + 0.007581640034914017, + 0.00474397512152791, + 0.010075563564896584, + 0.004595159087330103, + 0.005358419381082058, + 0.008472459390759468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.07899660617113113 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293341219425201 + ], + "opposite_logprob_tokens": [ + -2.237933397293091 + ], + "logprob_reward": -0.11293341219425201, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293341219425201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07471191138029099, + 0.016468049958348274, + 0.5393718481063843, + 0.04096142575144768, + 0.017531339079141617, + 0.020097270607948303, + 0.0142835583537817, + 0.022127976641058922, + 0.042719509452581406, + 0.2117270678281784 + ], + "credit_map": [ + 0.007772586774080992, + 0.007634065579622984, + 0.008123707957565784, + 0.009303889237344265, + 0.012823806144297123, + 0.007135175168514252, + 0.011933274567127228, + 0.009985414333641529, + 0.0012104063061997294, + 0.0011371830478310585, + 0.0017830216092988849, + 0.0021094793919473886, + 0.003444664180278778, + 0.001284988597035408, + 0.0032732521649450064, + 0.0022250558249652386, + 0.04569661244750023, + 0.017959484830498695, + 0.056272100657224655, + 0.21182984113693237, + 0.02362711727619171, + 0.016073551028966904, + 0.021627986803650856, + 0.14628516137599945, + 0.0035326762590557337, + 0.0030870609916746616, + 0.004104034509509802, + 0.005040829069912434, + 0.01093447208404541, + 0.002802734263241291, + 0.004028518218547106, + 0.007431103847920895, + 0.002182045951485634, + 0.0011820205254480243, + 0.0017205980839207768, + 0.0021907929331064224, + 0.004713871981948614, + 0.0014026113785803318, + 0.0016971672885119915, + 0.002442230936139822, + 0.001946679432876408, + 0.001966031501069665, + 0.00297989952377975, + 0.0022386987693607807, + 0.003276318544521928, + 0.0015805187867954373, + 0.002454141853377223, + 0.003654984524473548, + 0.0013233469799160957, + 0.001188731985166669, + 0.0022991630248725414, + 0.00211942195892334, + 0.0019769177306443453, + 0.0012067155912518501, + 0.002334326971322298, + 0.001834936672821641, + 0.0024325947742909193, + 0.002043468877673149, + 0.0036218296736478806, + 0.0030012528877705336, + 0.0022925171069800854, + 0.002029399387538433, + 0.0030008137691766024, + 0.003706101793795824, + 0.004106162115931511, + 0.004708010237663984, + 0.006648847833275795, + 0.005720200948417187, + 0.004852731246501207, + 0.003464932320639491, + 0.007399425841867924, + 0.005819205194711685, + 0.021648172289133072, + 0.044576577842235565, + 0.02448979578912258, + 0.018999693915247917, + 0.03928057476878166, + 0.014610899612307549, + 0.02017628215253353, + 0.027945097535848618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.11293341219425201 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023286908864974976 + ], + "opposite_logprob_tokens": [ + -3.773286819458008 + ], + "logprob_reward": -0.023286908864974976, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023286908864974976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19430840015411377, + 0.04672065004706383, + 0.07613368332386017, + 0.02952093631029129, + 0.044679705053567886, + 0.02931392937898636, + 0.025011280551552773, + 0.02983907237648964, + 0.32496029138565063, + 0.1995120644569397 + ], + "credit_map": [ + 0.017326470464468002, + 0.012814450077712536, + 0.037653207778930664, + 0.030397865921258926, + 0.014157479628920555, + 0.0195932537317276, + 0.03244662657380104, + 0.02991902269423008, + 0.005183181259781122, + 0.0032846182584762573, + 0.009750191122293472, + 0.0069059571251273155, + 0.004320470616221428, + 0.004812483675777912, + 0.006387439090758562, + 0.006076304242014885, + 0.005935110151767731, + 0.00257348595187068, + 0.0077056740410625935, + 0.02739315666258335, + 0.006996434647589922, + 0.0027806474827229977, + 0.003901250194758177, + 0.01884792000055313, + 0.003472990356385708, + 0.001925747375935316, + 0.004677603952586651, + 0.004655276890844107, + 0.0029881931841373444, + 0.003919529262930155, + 0.002587544033303857, + 0.005294048693031073, + 0.007804285269230604, + 0.0036698775365948677, + 0.007091848645359278, + 0.007172046694904566, + 0.005231541581451893, + 0.004324952606111765, + 0.003224983112886548, + 0.006160164717584848, + 0.0041578044183552265, + 0.0023821904323995113, + 0.004347339738160372, + 0.003472491167485714, + 0.0035646939650177956, + 0.0032958039082586765, + 0.0035823043435811996, + 0.004511296283453703, + 0.0026952086482197046, + 0.0019252229249104857, + 0.004869039170444012, + 0.0029847908299416304, + 0.0026236388366669416, + 0.0021899803541600704, + 0.0038673030212521553, + 0.0038560929242521524, + 0.00262438808567822, + 0.0030221992637962103, + 0.00429491838440299, + 0.0036373180337250233, + 0.005786473862826824, + 0.0027011441998183727, + 0.0032279924489557743, + 0.004544634371995926, + 0.02410763129591942, + 0.05774514377117157, + 0.028655491769313812, + 0.032323114573955536, + 0.09327900409698486, + 0.012798681855201721, + 0.02826281450688839, + 0.047788407653570175, + 0.02448612079024315, + 0.040553461760282516, + 0.023430995643138885, + 0.01796143874526024, + 0.03150465711951256, + 0.017590660601854324, + 0.01823442243039608, + 0.02575029619038105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.023286908864974976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward angle, the camera held still, focusing on a queen. A magnificent, heavy golden crown, inlaid with crimson gems and intricate baroque carvings, weighed heavily on her forehead. A few cascading strands of raven hair cascaded from the crown's edge, draping around her pale, sharply defined face. She wore a deep purple velvet gown with intricate silver embroidery at the collar, which shimmered faintly in the dim light. Her setting was a grand but dimly lit Gothic palace hall, the only faint light coming from the cool moonlight filtering through the distant high windows and the flickering flames of a few candles in silver candlesticks beside her, casting long, twisted shadows on the massive stone pillars in the background. At first, her eyes were closed, her face calm, yet a subtle hint of fatigue lingered. Then, slowly, with an almost palpable weight, she lifted her long lashes. The warm candlelight traced a soft golden edge along the contours of her cheeks. When her eyes fully opened, a pair of icy blue, hawk-like eyes were revealed. Their gaze, without a trace of hesitation, pierced the camera directly and resolutely. The entire scene, interwoven with the cold moonlight and the warm candlelight, was filled with a solemn, awe-inspiring, and powerful atmosphere, as if a silent storm was brewing in the midst of her decision.", + "seed": 20631, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_0/index_8451_seed_20631_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10762850120663643, + "vlm_logprob": -0.10762850120663643, + "vlm_yes_no_margin": 2.8, + "vlm_token_credit": -0.10762850120663643, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044781316071748734 + ], + "opposite_logprob_tokens": [ + -3.16978120803833 + ], + "logprob_reward": -0.044781316071748734, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.044781316071748734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44192057847976685, + 0.10390830039978027, + 0.0647907704114914, + 0.033855415880680084, + 0.0483844019472599, + 0.06363842636346817, + 0.05954132601618767, + 0.05216977000236511, + 0.06620315462350845, + 0.06558787077665329 + ], + "credit_map": [ + 0.04805225878953934, + 0.05568789318203926, + 0.05062698572874069, + 0.07231249660253525, + 0.05220884084701538, + 0.051810696721076965, + 0.05657166615128517, + 0.05464977025985718, + 0.007663329131901264, + 0.014378183521330357, + 0.008526512421667576, + 0.016875600442290306, + 0.0124769676476717, + 0.011504887603223324, + 0.019886836409568787, + 0.012595977634191513, + 0.005296004470437765, + 0.00994117185473442, + 0.007024690974503756, + 0.011114264838397503, + 0.007225340697914362, + 0.005431932397186756, + 0.009822199121117592, + 0.008935166522860527, + 0.003554491326212883, + 0.0059944037348032, + 0.004477976355701685, + 0.006010567769408226, + 0.0038476104382425547, + 0.00300102517940104, + 0.0029557079542428255, + 0.0040136342868208885, + 0.004141614306718111, + 0.007217623293399811, + 0.005335274152457714, + 0.007801173720508814, + 0.005754468031227589, + 0.0044625927694141865, + 0.007612669374793768, + 0.006058984436094761, + 0.006007168907672167, + 0.009572915732860565, + 0.007778936065733433, + 0.009503835812211037, + 0.006843233481049538, + 0.005399509333074093, + 0.009175431914627552, + 0.009357396513223648, + 0.004199451766908169, + 0.008145809173583984, + 0.006571083329617977, + 0.009554585441946983, + 0.008268548175692558, + 0.006898393388837576, + 0.006920552812516689, + 0.008982900530099869, + 0.005016689654439688, + 0.006947134155780077, + 0.008577213622629642, + 0.008385512046515942, + 0.00428879214450717, + 0.0032872792799025774, + 0.007994158193469048, + 0.007672993000596762, + 0.006171610206365585, + 0.011742526665329933, + 0.008469617925584316, + 0.0114121213555336, + 0.0068747494369745255, + 0.00610082084313035, + 0.006722799967974424, + 0.008708910085260868, + 0.004606053698807955, + 0.012872837483882904, + 0.004626331850886345, + 0.013107124716043472, + 0.006972345057874918, + 0.006184810306876898, + 0.009000915102660656, + 0.00821744929999113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.044781316071748734 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009809266775846481 + ], + "opposite_logprob_tokens": [ + -4.634809494018555 + ], + "logprob_reward": -0.009809266775846481, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009809266775846481, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33216965198516846, + 0.13630662858486176, + 0.18171007931232452, + 0.045069217681884766, + 0.06557730585336685, + 0.026530839502811432, + 0.05054595693945885, + 0.03912408649921417, + 0.07254653424024582, + 0.05041968449950218 + ], + "credit_map": [ + 0.02636873722076416, + 0.043176136910915375, + 0.03556166961789131, + 0.05343402177095413, + 0.04504884034395218, + 0.03704126924276352, + 0.04887542501091957, + 0.0426635667681694, + 0.008128424175083637, + 0.020401887595653534, + 0.01167280226945877, + 0.023167409002780914, + 0.021344056352972984, + 0.016946282237768173, + 0.019926410168409348, + 0.014719368889927864, + 0.010541679337620735, + 0.026984361931681633, + 0.012883903458714485, + 0.029841506853699684, + 0.028730306774377823, + 0.022092901170253754, + 0.026498856022953987, + 0.024136560037732124, + 0.004289529751986265, + 0.00812527071684599, + 0.0037591701839119196, + 0.008029717020690441, + 0.004588043317198753, + 0.0033824830316007137, + 0.006284622009843588, + 0.006610380485653877, + 0.004773381631821394, + 0.009262189269065857, + 0.0050385803915560246, + 0.010204320773482323, + 0.01289217546582222, + 0.009736106730997562, + 0.007893131114542484, + 0.005777421873062849, + 0.0025348670314997435, + 0.003994384780526161, + 0.0029211987275630236, + 0.004172933287918568, + 0.002674346324056387, + 0.0028372048400342464, + 0.0035599591210484505, + 0.003835945622995496, + 0.0023627832997590303, + 0.009703085757791996, + 0.004017032217234373, + 0.00894880946725607, + 0.008066434413194656, + 0.006024879869073629, + 0.0055547752417624, + 0.005868153180927038, + 0.004632786847651005, + 0.005198844708502293, + 0.0047629717737436295, + 0.006579356733709574, + 0.003978124354034662, + 0.003699720371514559, + 0.005091491155326366, + 0.005180789157748222, + 0.004829060286283493, + 0.01312259677797556, + 0.006143529899418354, + 0.016058316454291344, + 0.008288626559078693, + 0.007367713842540979, + 0.007568049244582653, + 0.009168642573058605, + 0.004132561385631561, + 0.0090979915112257, + 0.004714855924248695, + 0.009274458512663841, + 0.005284039303660393, + 0.005530680064111948, + 0.005250819958746433, + 0.007134278770536184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.009809266775846481 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20163513720035553 + ], + "opposite_logprob_tokens": [ + -1.7016351222991943 + ], + "logprob_reward": -0.20163513720035553, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20163513720035553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3364870846271515, + 0.10873057693243027, + 0.09106692671775818, + 0.04169093817472458, + 0.04336921498179436, + 0.056233830749988556, + 0.08958608657121658, + 0.0815376415848732, + 0.11576689034700394, + 0.03553076833486557 + ], + "credit_map": [ + 0.023061970248818398, + 0.04619619622826576, + 0.028790825977921486, + 0.04555357247591019, + 0.046474114060401917, + 0.049892399460077286, + 0.050313256680965424, + 0.04620475322008133, + 0.008437875658273697, + 0.02123621106147766, + 0.01186885591596365, + 0.017644809558987617, + 0.012559912167489529, + 0.013172145932912827, + 0.012858253903687, + 0.010952510870993137, + 0.010460486635565758, + 0.017359863966703415, + 0.010660646483302116, + 0.01407026220113039, + 0.010881618596613407, + 0.009712759405374527, + 0.009764030575752258, + 0.008157256059348583, + 0.0027842014096677303, + 0.005464135203510523, + 0.0028764770831912756, + 0.007316472940146923, + 0.005490274168550968, + 0.0043929447419941425, + 0.006165310274809599, + 0.0072011216543614864, + 0.0050513469614088535, + 0.007680806796997786, + 0.004758005030453205, + 0.007149047218263149, + 0.003931580577045679, + 0.003739218460395932, + 0.004937152378261089, + 0.006122055929154158, + 0.00857522338628769, + 0.011442601680755615, + 0.006033942103385925, + 0.009114718064665794, + 0.003927929792553186, + 0.005187261383980513, + 0.00577067956328392, + 0.0061814747750759125, + 0.0069913100451231, + 0.018955277279019356, + 0.008194797672331333, + 0.01576198637485504, + 0.010545505210757256, + 0.009843084961175919, + 0.010388289578258991, + 0.008905834518373013, + 0.010781077668070793, + 0.011598639190196991, + 0.012471134774386883, + 0.014019662514328957, + 0.0051766084507107735, + 0.004918870981782675, + 0.0097056208178401, + 0.012866036035120487, + 0.007700887508690357, + 0.016423411667346954, + 0.008519225753843784, + 0.02450689859688282, + 0.01377153117209673, + 0.013363570906221867, + 0.014722990803420544, + 0.016758373007178307, + 0.002552780555561185, + 0.005439904984086752, + 0.005313200410455465, + 0.005632774904370308, + 0.0031643209513276815, + 0.00318709434941411, + 0.005479658488184214, + 0.004761033225804567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.20163513720035553 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521163821220398 + ], + "opposite_logprob_tokens": [ + -1.5021164417266846 + ], + "logprob_reward": -0.2521163821220398, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521163821220398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.473544180393219, + 0.10791302472352982, + 0.05513931065797806, + 0.09878061711788177, + 0.04181471839547157, + 0.04014906659722328, + 0.030833370983600616, + 0.06980720907449722, + 0.027092142030596733, + 0.054926320910453796 + ], + "credit_map": [ + 0.054814405739307404, + 0.09607615321874619, + 0.056221380829811096, + 0.12434574961662292, + 0.029390983283519745, + 0.023350374773144722, + 0.049189940094947815, + 0.04015524685382843, + 0.007872183807194233, + 0.01913553848862648, + 0.009355124086141586, + 0.02414575032889843, + 0.01084898505359888, + 0.00868294294923544, + 0.014520999975502491, + 0.01335151307284832, + 0.006011737044900656, + 0.007504320703446865, + 0.00593601306900382, + 0.009067609906196594, + 0.005975707899779081, + 0.00398948835209012, + 0.008293621242046356, + 0.008360816165804863, + 0.005294359754770994, + 0.012560633942484856, + 0.006799762137234211, + 0.01649094559252262, + 0.014759240671992302, + 0.01044386625289917, + 0.018478702753782272, + 0.01395310927182436, + 0.0031665386632084846, + 0.006365758366882801, + 0.004125614184886217, + 0.0063750543631613255, + 0.004352228716015816, + 0.005547202657908201, + 0.005297412630170584, + 0.0065849097445607185, + 0.0029751514084637165, + 0.005465429276227951, + 0.004950737580657005, + 0.006528531201183796, + 0.003580404445528984, + 0.003940559457987547, + 0.005723242647945881, + 0.006985016632825136, + 0.0031143147498369217, + 0.005456387531012297, + 0.003640334587544203, + 0.005086044315248728, + 0.0025568988639861345, + 0.002781287068501115, + 0.004364565946161747, + 0.0038335423450917006, + 0.004812302067875862, + 0.010154839605093002, + 0.0073023210279643536, + 0.01375232357531786, + 0.006584437098354101, + 0.00711313309147954, + 0.008150603622198105, + 0.011937258765101433, + 0.0018966576317325234, + 0.005896762944757938, + 0.0029867617413401604, + 0.0045977262780070305, + 0.0028316061943769455, + 0.0020568061154335737, + 0.0029580004047602415, + 0.0038678233977407217, + 0.0036563610192388296, + 0.007805293425917625, + 0.005490853451192379, + 0.010226789861917496, + 0.004648189526051283, + 0.005064615048468113, + 0.006849290803074837, + 0.01118493266403675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.2521163821220398 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029800403863191605 + ], + "opposite_logprob_tokens": [ + -3.5298004150390625 + ], + "logprob_reward": -0.029800403863191605, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029800403863191605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18159860372543335, + 0.10995005071163177, + 0.07001187652349472, + 0.1034276932477951, + 0.09320307523012161, + 0.1733109951019287, + 0.07964767515659332, + 0.05411148443818092, + 0.06559620797634125, + 0.06914233416318893 + ], + "credit_map": [ + 0.018852047622203827, + 0.03475970774888992, + 0.029819248244166374, + 0.030654042959213257, + 0.010439401492476463, + 0.010454713366925716, + 0.020963136106729507, + 0.025656310841441154, + 0.02102491818368435, + 0.02176312729716301, + 0.015219500288367271, + 0.02192278765141964, + 0.00498222978785634, + 0.005666423588991165, + 0.009287738241255283, + 0.010083328001201153, + 0.007863011211156845, + 0.014021827839314938, + 0.012017856352031231, + 0.01369511615484953, + 0.004173490218818188, + 0.003632747568190098, + 0.0068956101313233376, + 0.007712218444794416, + 0.017337026074528694, + 0.024798521772027016, + 0.011143560521304607, + 0.02265164442360401, + 0.005174774676561356, + 0.00547215947881341, + 0.008509094826877117, + 0.008340906351804733, + 0.00880544912070036, + 0.021083690226078033, + 0.010657628998160362, + 0.017172567546367645, + 0.006767645012587309, + 0.006982072256505489, + 0.011448342353105545, + 0.010285676456987858, + 0.019594110548496246, + 0.04477936029434204, + 0.021679174154996872, + 0.03369325399398804, + 0.010692683979868889, + 0.014291536994278431, + 0.011478125117719173, + 0.017102748155593872, + 0.00987334456294775, + 0.01703246869146824, + 0.013502472080290318, + 0.015815675258636475, + 0.004740424454212189, + 0.004858026280999184, + 0.0062610008753836155, + 0.007564264815300703, + 0.005386484321206808, + 0.012795379385352135, + 0.006832460407167673, + 0.0091464938595891, + 0.003632825566455722, + 0.003750659991055727, + 0.004404707346111536, + 0.008162473328411579, + 0.0043212673626840115, + 0.01824747584760189, + 0.004707681946456432, + 0.014244250021874905, + 0.0056124161928892136, + 0.004378114826977253, + 0.007026824168860912, + 0.0070581804029643536, + 0.0075135002844035625, + 0.016408251598477364, + 0.00718333013355732, + 0.01394352875649929, + 0.004497456829994917, + 0.00445735827088356, + 0.006501099560409784, + 0.008637812919914722 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.029800403863191605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot from a slightly upward angle, the camera held still, focusing on a queen. A magnificent, heavy golden crown, inlaid with crimson gems and intricate baroque carvings, weighed heavily on her forehead. A few cascading strands of raven hair cascaded from the crown's edge, draping around her pale, sharply defined face. She wore a deep purple velvet gown with intricate silver embroidery at the collar, which shimmered faintly in the dim light. Her setting was a grand but dimly lit Gothic palace hall, the only faint light coming from the cool moonlight filtering through the distant high windows and the flickering flames of a few candles in silver candlesticks beside her, casting long, twisted shadows on the massive stone pillars in the background. At first, her eyes were closed, her face calm, yet a subtle hint of fatigue lingered. Then, slowly, with an almost palpable weight, she lifted her long lashes. The warm candlelight traced a soft golden edge along the contours of her cheeks. When her eyes fully opened, a pair of icy blue, hawk-like eyes were revealed. Their gaze, without a trace of hesitation, pierced the camera directly and resolutely. The entire scene, interwoven with the cold moonlight and the warm candlelight, was filled with a solemn, awe-inspiring, and powerful atmosphere, as if a silent storm was brewing in the midst of her decision.", + "seed": 20632, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_0/index_8451_seed_20632_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06937153367325663, + "vlm_logprob": -0.06937153367325663, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.06937153367325663, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0098484568297863 + ], + "opposite_logprob_tokens": [ + -4.884848594665527 + ], + "logprob_reward": -0.0098484568297863, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0098484568297863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03526793047785759, + 0.016721127554774284, + 0.02829594723880291, + 0.8292896151542664, + 0.014607178047299385, + 0.005993694998323917, + 0.0035403985530138016, + 0.007220913656055927, + 0.010727881453931332, + 0.048335324972867966 + ], + "credit_map": [ + 0.004876148886978626, + 0.002162007847800851, + 0.005760295782238245, + 0.0033619890455156565, + 0.004933512769639492, + 0.0027737822383642197, + 0.004540711175650358, + 0.0068594831973314285, + 0.005335132125765085, + 0.0008310348493978381, + 0.0038575208745896816, + 0.0009407068137079477, + 0.0010141154052689672, + 0.0006728199077770114, + 0.0010334636317566037, + 0.003036335576325655, + 0.008471119217574596, + 0.001576910843141377, + 0.003568798303604126, + 0.0016135871410369873, + 0.0016124347457662225, + 0.0018326456192880869, + 0.0011716712033376098, + 0.008448783308267593, + 0.2665921151638031, + 0.025540074333548546, + 0.06226713955402374, + 0.03466063365340233, + 0.014728033915162086, + 0.046882256865501404, + 0.01772232912480831, + 0.360897034406662, + 0.0022712843492627144, + 0.001208244008012116, + 0.0013224896974861622, + 0.0012500695884227753, + 0.001594610745087266, + 0.0027532526291906834, + 0.0017682057805359364, + 0.0024390225298702717, + 0.0012392563512548804, + 0.0005919626564718783, + 0.0009402918512932956, + 0.0008357476908713579, + 0.00046457984717562795, + 0.000629817193839699, + 0.0006649568676948547, + 0.0006270830053836107, + 0.0006228733691386878, + 0.000276029430096969, + 0.000619067344814539, + 0.00043906521750614047, + 0.0003123716451227665, + 0.00030032338690944016, + 0.0004645097942557186, + 0.0005061588017269969, + 0.00115675397682935, + 0.0006485733320005238, + 0.0007301142322830856, + 0.0007975767366588116, + 0.0013059276388958097, + 0.0006263455143198371, + 0.0008799224742688239, + 0.0010756996925920248, + 0.0017037821235135198, + 0.0009719998924992979, + 0.001036253059282899, + 0.0009758105734363198, + 0.0023574759252369404, + 0.0009693006868474185, + 0.0013933763839304447, + 0.0013198833912611008, + 0.006629998330026865, + 0.006345230154693127, + 0.005375278182327747, + 0.003889851039275527, + 0.012299858964979649, + 0.004514485597610474, + 0.005360017064958811, + 0.003920604940503836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.0098484568297863 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007654739078134298 + ], + "opposite_logprob_tokens": [ + -4.882654666900635 + ], + "logprob_reward": -0.007654739078134298, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007654739078134298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045253537595272064, + 0.03284262865781784, + 0.17521798610687256, + 0.6148988008499146, + 0.020452415570616722, + 0.016266439110040665, + 0.011099057272076607, + 0.011658872477710247, + 0.016023807227611542, + 0.05628647655248642 + ], + "credit_map": [ + 0.0069124153815209866, + 0.0026826204266399145, + 0.012401278130710125, + 0.0036439946852624416, + 0.005216887686401606, + 0.0033309601712971926, + 0.003946126438677311, + 0.007119257468730211, + 0.010148986242711544, + 0.0015199477784335613, + 0.008144460618495941, + 0.002016706857830286, + 0.002123879035934806, + 0.0016825104830786586, + 0.0017795275198295712, + 0.005426611285656691, + 0.05914419889450073, + 0.008393509313464165, + 0.030284292995929718, + 0.008492693305015564, + 0.005138298962265253, + 0.009140126407146454, + 0.004602610133588314, + 0.05002225562930107, + 0.19749639928340912, + 0.018893204629421234, + 0.04616956040263176, + 0.025776030495762825, + 0.011187843047082424, + 0.034863609820604324, + 0.013202855363488197, + 0.267309308052063, + 0.0025854597333818674, + 0.0017544865841045976, + 0.002780853770673275, + 0.0026466725394129753, + 0.0029095669742673635, + 0.0028831916861236095, + 0.0019486298551782966, + 0.0029435548931360245, + 0.0021071522496640682, + 0.001177553553134203, + 0.002394185634329915, + 0.0036275081802159548, + 0.0017181389266625047, + 0.001267093699425459, + 0.0020880447700619698, + 0.0018867634935304523, + 0.0019499524496495724, + 0.0011564330197870731, + 0.001540653989650309, + 0.001194775104522705, + 0.0008954851073212922, + 0.001619728747755289, + 0.0012060991721227765, + 0.0015359302051365376, + 0.001771958777680993, + 0.001034472486935556, + 0.0012389952316880226, + 0.0011060837423428893, + 0.0019519163761287928, + 0.0013500336790457368, + 0.0017148848855867982, + 0.001490528229624033, + 0.002878413302823901, + 0.0014691834803670645, + 0.001886368147097528, + 0.0015064359176903963, + 0.002624472603201866, + 0.00205258559435606, + 0.0015930694062262774, + 0.0020132793579250574, + 0.006938967388123274, + 0.00890442356467247, + 0.0064017330296337605, + 0.00372340832836926, + 0.014716516248881817, + 0.0052682566456496716, + 0.0063405223190784454, + 0.003992652520537376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 9 + ], + "selected_reward": -0.007654739078134298 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700620487332344 + ], + "opposite_logprob_tokens": [ + -2.6950621604919434 + ], + "logprob_reward": -0.0700620487332344, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700620487332344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054916832596063614, + 0.03690951317548752, + 0.1996411830186844, + 0.5886589288711548, + 0.020245730876922607, + 0.0561826266348362, + 0.008473383262753487, + 0.015175770968198776, + 0.012754814699292183, + 0.007041329517960548 + ], + "credit_map": [ + 0.00853021815419197, + 0.0032219253480434418, + 0.01627689227461815, + 0.006290839519351721, + 0.004093956667929888, + 0.003261241363361478, + 0.003527991008013487, + 0.009713760577142239, + 0.008527432568371296, + 0.0017785010859370232, + 0.006506696809083223, + 0.0018189529655501246, + 0.0015790985198691487, + 0.001733422395773232, + 0.0017496469663456082, + 0.013215751387178898, + 0.0666595846414566, + 0.00986921414732933, + 0.03749053552746773, + 0.009660324081778526, + 0.006450112443417311, + 0.01011768076568842, + 0.005806493107229471, + 0.053587209433317184, + 0.18892192840576172, + 0.018310435116291046, + 0.044244587421417236, + 0.02451445162296295, + 0.010655838064849377, + 0.03330346941947937, + 0.013141585513949394, + 0.2555665671825409, + 0.002804181771352887, + 0.0019607420545071363, + 0.002843528753146529, + 0.003745516762137413, + 0.0019221898401156068, + 0.0022244376596063375, + 0.0025828673969954252, + 0.0021622644271701574, + 0.007493885233998299, + 0.004943610168993473, + 0.007638310547918081, + 0.01413598284125328, + 0.004150143824517727, + 0.004781417548656464, + 0.00648755207657814, + 0.006551711820065975, + 0.0011918856762349606, + 0.0013976594200357795, + 0.0011758372420445085, + 0.0010672392090782523, + 0.0010051581775769591, + 0.0008487176964990795, + 0.0009020079742185771, + 0.0008848758297972381, + 0.0029291207902133465, + 0.0016452108975499868, + 0.0022379839792847633, + 0.0019418848678469658, + 0.0015485510230064392, + 0.0010883478680625558, + 0.001983235590159893, + 0.001801432459615171, + 0.0019556626211851835, + 0.001118816900998354, + 0.002260620705783367, + 0.0012683720560744405, + 0.0014098179526627064, + 0.0019815568812191486, + 0.0013006140943616629, + 0.0014593502273783088, + 0.0009313916089013219, + 0.0006150456611067057, + 0.001024071709252894, + 0.0006857796106487513, + 0.0012454978423193097, + 0.0008252496481873095, + 0.0010578961810097098, + 0.0006563953938893974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.0700620487332344 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22558902204036713 + ], + "opposite_logprob_tokens": [ + -1.6005890369415283 + ], + "logprob_reward": -0.22558902204036713, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22558902204036713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02825750783085823, + 0.02169843204319477, + 0.07387129217386246, + 0.8268172144889832, + 0.009510830044746399, + 0.004995609167963266, + 0.0029427988920360804, + 0.004602500237524509, + 0.0057051521725952625, + 0.02159862406551838 + ], + "credit_map": [ + 0.003450073068961501, + 0.001813371549360454, + 0.005926766432821751, + 0.00227747461758554, + 0.004825820680707693, + 0.002047674497589469, + 0.0035033454187214375, + 0.004412982612848282, + 0.007709972094744444, + 0.0009294022456742823, + 0.00488127488642931, + 0.0009881588630378246, + 0.0006909978110343218, + 0.000798999157268554, + 0.0008910376345738769, + 0.004808588419109583, + 0.024256229400634766, + 0.003736730432137847, + 0.011097920127213001, + 0.0036632628180086613, + 0.0024796200450509787, + 0.003884644014760852, + 0.002355418633669615, + 0.022397473454475403, + 0.2656816244125366, + 0.025375422090291977, + 0.061849016696214676, + 0.03449787572026253, + 0.01485172938555479, + 0.0467798188328743, + 0.01784718967974186, + 0.35993459820747375, + 0.0016923107905313373, + 0.0005975250387564301, + 0.000865442561917007, + 0.0007891066488809884, + 0.0012376136146485806, + 0.0013650251785293221, + 0.0012338138185441494, + 0.0017299914034083486, + 0.0006141742342151701, + 0.0004919394268654287, + 0.0006267799763008952, + 0.0007567001157440245, + 0.0004506170516833663, + 0.0008351196884177625, + 0.000460631592432037, + 0.000759646762162447, + 0.000469524587970227, + 0.00033666042145341635, + 0.0004297714913263917, + 0.0002961297577712685, + 0.00024294298782479018, + 0.0005110139027237892, + 0.000271113240160048, + 0.0003856426919810474, + 0.0008309431141242385, + 0.00042793049942702055, + 0.00070334249176085, + 0.000481971277622506, + 0.0004660353879444301, + 0.00028387794736772776, + 0.0005069619510322809, + 0.0009014373063109815, + 0.0008526769233867526, + 0.0006490775849670172, + 0.0006272454629652202, + 0.000693109177518636, + 0.0007402008050121367, + 0.0007924125529825687, + 0.0006670802249573171, + 0.0006833492079749703, + 0.0026741009205579758, + 0.003292613662779331, + 0.00235806149430573, + 0.0014619322028011084, + 0.0057447561994194984, + 0.0020476968493312597, + 0.002517137909308076, + 0.0015023243613541126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.22558902204036713 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of a queen with pale, sharply defined features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the queen wearing a heavy golden crown with crimson gems and baroque carvings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the queen wear a deep purple velvet gown with silver embroidery at the collar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene transition from the queen's eyes closed to her opening them to reveal icy blue, hawk-like eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a dimly lit Gothic palace hall with moonlight and candlelight casting shadows on stone pillars?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370340168476105 + ], + "opposite_logprob_tokens": [ + -3.408703327178955 + ], + "logprob_reward": -0.03370340168476105, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370340168476105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17629846930503845, + 0.13758201897144318, + 0.24091839790344238, + 0.17271235585212708, + 0.06790617853403091, + 0.03335285931825638, + 0.04368927702307701, + 0.03909868746995926, + 0.029710207134485245, + 0.058731559664011 + ], + "credit_map": [ + 0.026382766664028168, + 0.013793155550956726, + 0.04579891636967659, + 0.017993878573179245, + 0.010632256045937538, + 0.009880449622869492, + 0.012794766575098038, + 0.03902226686477661, + 0.03782888129353523, + 0.008012705482542515, + 0.016798989847302437, + 0.0065172454342246056, + 0.006439052056521177, + 0.00718392850831151, + 0.007779987528920174, + 0.047021232545375824, + 0.07819482684135437, + 0.012133443728089333, + 0.029229070991277695, + 0.012992629781365395, + 0.005896452348679304, + 0.01382338348776102, + 0.006967220455408096, + 0.08168136328458786, + 0.05507947877049446, + 0.005557953380048275, + 0.012530092149972916, + 0.007052673026919365, + 0.0038545429706573486, + 0.009591802954673767, + 0.004613069351762533, + 0.07443273067474365, + 0.010859837755560875, + 0.006444415543228388, + 0.010058712214231491, + 0.00490221893414855, + 0.005553300026804209, + 0.014176338911056519, + 0.0072633069939911366, + 0.008648045361042023, + 0.006996824871748686, + 0.0043883840553462505, + 0.006553925573825836, + 0.00313694984652102, + 0.0024270096328109503, + 0.00325310998596251, + 0.0030575899872928858, + 0.0035390653647482395, + 0.0067390152253210545, + 0.004890578333288431, + 0.0077200597152113914, + 0.00362544902600348, + 0.0041749607771635056, + 0.007553876843303442, + 0.0038970650639384985, + 0.005088271573185921, + 0.006785871926695108, + 0.0038248407654464245, + 0.00635898532345891, + 0.0039530787616968155, + 0.005097017623484135, + 0.002597235143184662, + 0.00531403161585331, + 0.00516762537881732, + 0.004790655802935362, + 0.0025378281716257334, + 0.00428245821967721, + 0.00304439221508801, + 0.004191084764897823, + 0.004081277176737785, + 0.004020736087113619, + 0.0027617723681032658, + 0.006224050186574459, + 0.008648638613522053, + 0.005869053769856691, + 0.00423461152240634, + 0.017466995865106583, + 0.004626022186130285, + 0.00810276623815298, + 0.0035594177898019552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.03370340168476105 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_1/index_tensor([7682, 7682])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_1/index_tensor([7682, 7682])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..055879add655d214045dcb6010d14a7b2a9b76c0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_1/index_tensor([7682, 7682])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow panning shot, from a horizontal perspective, sweeps across an ancient Chinese courtyard shrouded in mist. At the center of the frame, an elderly man with white hair and youthful features, dressed in a loose, clean, traditional white cotton and linen training suit, is immersed in a Tai Chi practice. The bluestone floor beneath his feet gleams faintly with moisture, while the backdrop is a dark wooden corridor with intricate carvings and a verdant bamboo forest, faintly visible through the mist. The elderly man's expression is focused and serene, his arms drawing smooth, full arcs in a perceptibly slow rhythm, his center of gravity shifting steadily and effortlessly between his legs in sync with his breathing. Each movement is as smooth and uninterrupted as flowing water. As he assumes a pushing-hands stance, his wide sleeves flutter gently with his movements, as if parting the milky morning mist that permeates the surrounding area. The soft diffuse light of the early morning penetrates the mist, giving the entire scene a tranquil and hazy texture. The light outlines the soft contours of his white clothes and the subtle luster of his silver hair. The main colors are white, gray and green, presenting an extremely simple, elegant and peaceful visual atmosphere as a whole.", + "seed": 20503, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_1/index_7682_seed_20503_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1251184671651572, + "vlm_logprob": -0.1251184671651572, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.1251184671651572, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5824032425880432 + ], + "opposite_logprob_tokens": [ + -0.8324032425880432 + ], + "logprob_reward": -0.5824032425880432, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5824032425880432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06778677552938461, + 0.07333637773990631, + 0.2396421879529953, + 0.0682484582066536, + 0.24710258841514587, + 0.1472827047109604, + 0.06932158768177032, + 0.02028176747262478, + 0.03709793835878372, + 0.029899654909968376 + ], + "credit_map": [ + 0.006342706736177206, + 0.010332105681300163, + 0.008480674587190151, + 0.009900704026222229, + 0.010930942371487617, + 0.007162891328334808, + 0.009515673853456974, + 0.005121082998812199, + 0.011785563081502914, + 0.010163386352360249, + 0.007920002564787865, + 0.008187311701476574, + 0.009790473617613316, + 0.013234486803412437, + 0.0059464615769684315, + 0.006308693438768387, + 0.007579956203699112, + 0.056742943823337555, + 0.012680763378739357, + 0.09690054506063461, + 0.03768988698720932, + 0.008822408504784107, + 0.013105177320539951, + 0.006120499689131975, + 0.008586015552282333, + 0.00933846551924944, + 0.00801252480596304, + 0.008867410942912102, + 0.009812391363084316, + 0.007126740645617247, + 0.01027378998696804, + 0.006231118459254503, + 0.014314860105514526, + 0.0337359756231308, + 0.021416496485471725, + 0.019346587359905243, + 0.08144684880971909, + 0.025813516229391098, + 0.03665956109762192, + 0.014368745498359203, + 0.012285374104976654, + 0.020811187103390694, + 0.012566879391670227, + 0.02371240220963955, + 0.020510539412498474, + 0.011266821064054966, + 0.03414169326424599, + 0.011987808160483837, + 0.0038686341140419245, + 0.015876388177275658, + 0.006502512842416763, + 0.007623479701578617, + 0.013728265650570393, + 0.005292515270411968, + 0.012230741791427135, + 0.004199047572910786, + 0.0022538539487868547, + 0.0031932478304952383, + 0.0035723019391298294, + 0.0033238118048757315, + 0.002579938154667616, + 0.0020570375490933657, + 0.0018592423293739557, + 0.0014423326356336474, + 0.004201636649668217, + 0.005783446133136749, + 0.004607349168509245, + 0.004451523069292307, + 0.006108549423515797, + 0.004733898211270571, + 0.004959961399435997, + 0.0022515738382935524, + 0.004618158098310232, + 0.004390777088701725, + 0.003982096444815397, + 0.004568195436149836, + 0.0035824410151690245, + 0.0026747500523924828, + 0.0044616758823394775, + 0.0016215640353038907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 5 + ], + "selected_reward": -0.5824032425880432 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075758263468742 + ], + "opposite_logprob_tokens": [ + -4.141075611114502 + ], + "logprob_reward": -0.016075758263468742, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075758263468742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15541766583919525, + 0.11525166779756546, + 0.24971726536750793, + 0.08579940348863602, + 0.13169580698013306, + 0.10041311383247375, + 0.08502164483070374, + 0.03368157893419266, + 0.022756939753890038, + 0.020244916900992393 + ], + "credit_map": [ + 0.011837596073746681, + 0.019367428496479988, + 0.01794595830142498, + 0.01829310692846775, + 0.030436716973781586, + 0.020286358892917633, + 0.0240632314234972, + 0.013187270611524582, + 0.013512387871742249, + 0.017123546451330185, + 0.008186926133930683, + 0.02180980145931244, + 0.021438784897327423, + 0.017957599833607674, + 0.00806169118732214, + 0.007160925772041082, + 0.012156744487583637, + 0.07686074078083038, + 0.013034346513450146, + 0.08028830587863922, + 0.030329229310154915, + 0.014420118182897568, + 0.01560300961136818, + 0.007024771999567747, + 0.012222767807543278, + 0.011838367208838463, + 0.010510077700018883, + 0.011898350901901722, + 0.012467315420508385, + 0.00969826802611351, + 0.010566440410912037, + 0.006597817409783602, + 0.00808029156178236, + 0.020301319658756256, + 0.015297851525247097, + 0.010851497761905193, + 0.02639654465019703, + 0.023673683404922485, + 0.015698282048106194, + 0.011396338231861591, + 0.006710932590067387, + 0.019236158579587936, + 0.012814518995583057, + 0.012027760036289692, + 0.012511012144386768, + 0.005206990987062454, + 0.026036100462079048, + 0.005869635380804539, + 0.005692030768841505, + 0.01847028359770775, + 0.007902860641479492, + 0.009923364967107773, + 0.01490116398781538, + 0.00783678237348795, + 0.012727826833724976, + 0.007567331660538912, + 0.002546248259022832, + 0.004820634610950947, + 0.0055246916599571705, + 0.005680374335497618, + 0.005126625765115023, + 0.0035553888883441687, + 0.004090654663741589, + 0.0023369600530713797, + 0.0019042263738811016, + 0.004620194435119629, + 0.002866141963750124, + 0.003187518334016204, + 0.003162720939144492, + 0.001536308554932475, + 0.00409146212041378, + 0.0013883669162169099, + 0.003363413969054818, + 0.0033129395451396704, + 0.0026828404515981674, + 0.0025327769108116627, + 0.0026263361796736717, + 0.0016578921349719167, + 0.0026445735711604357, + 0.0014241428580135107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.016075758263468742 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735715549439192 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735715549439192, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735715549439192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13201548159122467, + 0.1597314029932022, + 0.1643041968345642, + 0.10279690474271774, + 0.10677173733711243, + 0.11640074849128723, + 0.08505169302225113, + 0.04680618271231651, + 0.04076850041747093, + 0.04535314813256264 + ], + "credit_map": [ + 0.013266857713460922, + 0.027216073125600815, + 0.011883138678967953, + 0.01387158501893282, + 0.02316349186003208, + 0.016432233154773712, + 0.013410559855401516, + 0.0127715440467, + 0.020213715732097626, + 0.01587412692606449, + 0.013774983584880829, + 0.04139048606157303, + 0.025803234428167343, + 0.01609647087752819, + 0.014893945306539536, + 0.011684435419738293, + 0.008092125877737999, + 0.05505814775824547, + 0.01043375302106142, + 0.04665405675768852, + 0.017258765175938606, + 0.013365980237722397, + 0.009532892145216465, + 0.003908471670001745, + 0.012753419578075409, + 0.013243861496448517, + 0.013783694244921207, + 0.011324119754135609, + 0.01895146258175373, + 0.013711229898035526, + 0.011001941747963428, + 0.008027171716094017, + 0.008617464452981949, + 0.011811100877821445, + 0.015586013905704021, + 0.010913994163274765, + 0.015242551453411579, + 0.019915498793125153, + 0.012188445776700974, + 0.012496665120124817, + 0.008764354512095451, + 0.016774404793977737, + 0.013603580184280872, + 0.020055105909705162, + 0.01887056976556778, + 0.00840434804558754, + 0.021342147141695023, + 0.008586234413087368, + 0.010625218041241169, + 0.011744706891477108, + 0.013498741202056408, + 0.011737421154975891, + 0.011793132871389389, + 0.007734315004199743, + 0.011597410775721073, + 0.006320741027593613, + 0.004470625426620245, + 0.007350733503699303, + 0.007239583879709244, + 0.007339272648096085, + 0.007743733935058117, + 0.003736290382221341, + 0.005848997738212347, + 0.0030769440345466137, + 0.0034553715959191322, + 0.005715747829526663, + 0.008100328035652637, + 0.006403322331607342, + 0.004756211303174496, + 0.002950313501060009, + 0.0068369703367352486, + 0.0025502359494566917, + 0.005650217644870281, + 0.007984187453985214, + 0.006234418600797653, + 0.005725698545575142, + 0.005673739593476057, + 0.0043614571914076805, + 0.00711342366412282, + 0.002609999617561698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.006735715549439192 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018181217834353447 + ], + "opposite_logprob_tokens": [ + -4.018181324005127 + ], + "logprob_reward": -0.018181217834353447, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018181217834353447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12728329002857208, + 0.1250959187746048, + 0.1350019872188568, + 0.08166906237602234, + 0.16572314500808716, + 0.030132364481687546, + 0.15135557949543, + 0.10218272358179092, + 0.01987057365477085, + 0.06168530881404877 + ], + "credit_map": [ + 0.009926572442054749, + 0.01752331480383873, + 0.014705025590956211, + 0.015051268972456455, + 0.02365707792341709, + 0.0163718331605196, + 0.019690608605742455, + 0.010357584804296494, + 0.01628451608121395, + 0.011492447927594185, + 0.010653447359800339, + 0.028380578383803368, + 0.022046277299523354, + 0.014377523213624954, + 0.012993778102099895, + 0.008867341093719006, + 0.006532044615596533, + 0.03163403645157814, + 0.006888565607368946, + 0.042500536888837814, + 0.02715185098350048, + 0.006804791744798422, + 0.008797413669526577, + 0.0046927472576498985, + 0.008785367012023926, + 0.012407872825860977, + 0.009197703562676907, + 0.010390267707407475, + 0.014868974685668945, + 0.009874225594103336, + 0.009844728745520115, + 0.006299917586147785, + 0.009097157046198845, + 0.02148596942424774, + 0.014332453720271587, + 0.012851053848862648, + 0.05600390210747719, + 0.01723639667034149, + 0.025320621207356453, + 0.009395591914653778, + 0.0030048585031181574, + 0.004367857705801725, + 0.0032990560866892338, + 0.0049802390858531, + 0.0033881599083542824, + 0.0031287474557757378, + 0.004507556091994047, + 0.0034558901097625494, + 0.013011167757213116, + 0.032717764377593994, + 0.023854663595557213, + 0.021473506465554237, + 0.01778436079621315, + 0.011705800890922546, + 0.020649345591664314, + 0.010158964432775974, + 0.012244069948792458, + 0.017935099080204964, + 0.016626399010419846, + 0.01511518657207489, + 0.014228351414203644, + 0.009034842252731323, + 0.011030187830328941, + 0.0059685916639864445, + 0.0016543627716600895, + 0.004172212444245815, + 0.00365948467515409, + 0.002994574373587966, + 0.0021344979759305716, + 0.002157298382371664, + 0.001995024736970663, + 0.001103118760511279, + 0.00677045714110136, + 0.010567739605903625, + 0.00886351615190506, + 0.007565803360193968, + 0.007644142955541611, + 0.0041387504898011684, + 0.012792197056114674, + 0.003342703217640519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.018181217834353447 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021964015904814005 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021964015904814005, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021964015904814005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15123821794986725, + 0.0906398817896843, + 0.19389574229717255, + 0.06171081215143204, + 0.1357039213180542, + 0.12554699182510376, + 0.14764918386936188, + 0.04746594280004501, + 0.020716825500130653, + 0.025432446971535683 + ], + "credit_map": [ + 0.022153634577989578, + 0.027192668989300728, + 0.014070991426706314, + 0.014429048635065556, + 0.020105166360735893, + 0.013221260160207748, + 0.027043908834457397, + 0.013021553866565228, + 0.015263350680470467, + 0.007632451597601175, + 0.008442824706435204, + 0.01743367128074169, + 0.015562648884952068, + 0.013278725557029247, + 0.0065077212639153, + 0.006518496200442314, + 0.006560641806572676, + 0.05038286745548248, + 0.011023078113794327, + 0.07708911597728729, + 0.02546507678925991, + 0.008080870844423771, + 0.01084843184798956, + 0.004445680417120457, + 0.006217948626726866, + 0.007668809965252876, + 0.007409485522657633, + 0.007483767811208963, + 0.013633457943797112, + 0.006863709073513746, + 0.007482495158910751, + 0.0049511403776705265, + 0.008313219994306564, + 0.018446633592247963, + 0.013534230180084705, + 0.011483431793749332, + 0.03829735890030861, + 0.017545979470014572, + 0.01890541799366474, + 0.009177658706903458, + 0.008748767897486687, + 0.013814153149724007, + 0.011274830438196659, + 0.02281867526471615, + 0.016735544428229332, + 0.009043763391673565, + 0.03245649114251137, + 0.010654768906533718, + 0.00983883161097765, + 0.026505570858716965, + 0.02549429051578045, + 0.024848002940416336, + 0.015448001213371754, + 0.01311640627682209, + 0.019193824380636215, + 0.013204275630414486, + 0.003804420121014118, + 0.007928101345896721, + 0.006408144254237413, + 0.008225233294069767, + 0.007540733553469181, + 0.003353037405759096, + 0.006546382792294025, + 0.0036598918959498405, + 0.0014835016336292028, + 0.004771560430526733, + 0.0028480475302785635, + 0.0031086173839867115, + 0.002627637004479766, + 0.0017435726476833224, + 0.0029082312248647213, + 0.0012256597401574254, + 0.003076232736930251, + 0.004683907609432936, + 0.0034901690669357777, + 0.00363401067443192, + 0.0031047845259308815, + 0.0027744038961827755, + 0.0029827416874468327, + 0.0016861996846273541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.0021964015904814005 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow panning shot, from a horizontal perspective, sweeps across an ancient Chinese courtyard shrouded in mist. At the center of the frame, an elderly man with white hair and youthful features, dressed in a loose, clean, traditional white cotton and linen training suit, is immersed in a Tai Chi practice. The bluestone floor beneath his feet gleams faintly with moisture, while the backdrop is a dark wooden corridor with intricate carvings and a verdant bamboo forest, faintly visible through the mist. The elderly man's expression is focused and serene, his arms drawing smooth, full arcs in a perceptibly slow rhythm, his center of gravity shifting steadily and effortlessly between his legs in sync with his breathing. Each movement is as smooth and uninterrupted as flowing water. As he assumes a pushing-hands stance, his wide sleeves flutter gently with his movements, as if parting the milky morning mist that permeates the surrounding area. The soft diffuse light of the early morning penetrates the mist, giving the entire scene a tranquil and hazy texture. The light outlines the soft contours of his white clothes and the subtle luster of his silver hair. The main colors are white, gray and green, presenting an extremely simple, elegant and peaceful visual atmosphere as a whole.", + "seed": 20504, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_1/index_7682_seed_20504_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.28409235635772345, + "vlm_logprob": -0.28409235635772345, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.28409235635772345, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3190761804580688 + ], + "opposite_logprob_tokens": [ + -0.31907621026039124 + ], + "logprob_reward": -1.3190761804580688, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3190761804580688, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.020802706480026245, + 0.024076268076896667, + 0.9126430153846741, + 0.008116064593195915, + 0.004913023207336664, + 0.005025925114750862, + 0.007094319444149733, + 0.003041642252355814, + 0.011590833775699139, + 0.0026963050477206707 + ], + "credit_map": [ + 0.002108403481543064, + 0.0033751337323337793, + 0.002804860007017851, + 0.003419223939999938, + 0.0013954410096630454, + 0.0016161594539880753, + 0.003069872735068202, + 0.00301361083984375, + 0.0012622858630493283, + 0.002024285262450576, + 0.0012252273736521602, + 0.004330102354288101, + 0.0018306949641555548, + 0.0017323291394859552, + 0.007345268968492746, + 0.004326070658862591, + 0.02229868434369564, + 0.01562858372926712, + 0.009009345434606075, + 0.03124931827187538, + 0.028644395992159843, + 0.02333948202431202, + 0.3909701108932495, + 0.3915030062198639, + 0.0004754590627271682, + 0.0006453687092289329, + 0.0005049405735917389, + 0.0010804010089486837, + 0.0012522925389930606, + 0.0008331282297149301, + 0.0019192389445379376, + 0.0014052340993657708, + 0.00043658228241838515, + 0.0005162284360267222, + 0.00042671459959819913, + 0.0006831035716459155, + 0.0005264781648293138, + 0.0004295918915886432, + 0.0010269584599882364, + 0.0008673649863339961, + 0.00031540816416963935, + 0.0006480810116045177, + 0.0004505427787080407, + 0.0007331361412070692, + 0.0012253157328814268, + 0.0005036040092818439, + 0.0005654952838085592, + 0.0005843413528054953, + 0.000785674958024174, + 0.0012591627892106771, + 0.0008552885847166181, + 0.00076006114250049, + 0.0011639156145974994, + 0.0007125051924958825, + 0.0006957095465622842, + 0.0008620007429271936, + 0.00033140197047032416, + 0.00026111261104233563, + 0.0004069834540132433, + 0.0005421747919172049, + 0.00018974451813846827, + 0.0003576252202037722, + 0.00026534945936873555, + 0.0006872497615404427, + 0.001498960074968636, + 0.0008121239370666444, + 0.0009696098859421909, + 0.00320334336720407, + 0.0005893154884688556, + 0.00207544956356287, + 0.0010995272314175963, + 0.0013425027718767524, + 0.0002336455654585734, + 0.00028451072284951806, + 0.00029440352227538824, + 0.0003970802354160696, + 0.0002542663423810154, + 0.00033572089159861207, + 0.0004704862367361784, + 0.0004261911381036043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -1.3190761804580688 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594101399183273 + ], + "opposite_logprob_tokens": [ + -3.8955941200256348 + ], + "logprob_reward": -0.020594101399183273, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594101399183273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09341426193714142, + 0.1927088499069214, + 0.3317020833492279, + 0.029534583911299706, + 0.05525771528482437, + 0.014843201264739037, + 0.043023545295000076, + 0.1727435141801834, + 0.04981499910354614, + 0.016957206651568413 + ], + "credit_map": [ + 0.009089184924960136, + 0.011179493740200996, + 0.009084802120923996, + 0.01280573196709156, + 0.009699096903204918, + 0.01523375790566206, + 0.013751274906098843, + 0.01257092971354723, + 0.010348753072321415, + 0.01418505422770977, + 0.01107074785977602, + 0.034736186265945435, + 0.01491797249764204, + 0.014013082720339298, + 0.05812886357307434, + 0.03530821576714516, + 0.00786520354449749, + 0.005503121763467789, + 0.003658253699541092, + 0.011560727842152119, + 0.010946563445031643, + 0.008377433754503727, + 0.141648069024086, + 0.14214272797107697, + 0.001764260116033256, + 0.0022039201576262712, + 0.002376912161707878, + 0.003980854991823435, + 0.0041549778543412685, + 0.0032275267876684666, + 0.006986806634813547, + 0.004839327652007341, + 0.0038959276862442493, + 0.009245297871530056, + 0.005729042459279299, + 0.007042028475552797, + 0.005668167490512133, + 0.00475279102101922, + 0.01093429047614336, + 0.007990173064172268, + 0.0013122785603627563, + 0.0018688811687752604, + 0.001860850490629673, + 0.002095067873597145, + 0.002340941457077861, + 0.001805984997190535, + 0.0018319046357646585, + 0.0017272934783250093, + 0.004061731975525618, + 0.007434418890625238, + 0.0051385750994086266, + 0.004660237114876509, + 0.006693155970424414, + 0.004863736219704151, + 0.003782611805945635, + 0.006389078684151173, + 0.01950152777135372, + 0.01220723707228899, + 0.016204772517085075, + 0.014682089909911156, + 0.0164821557700634, + 0.039488889276981354, + 0.01334389392286539, + 0.04083295911550522, + 0.0033221845515072346, + 0.0038866624236106873, + 0.0049500796012580395, + 0.010435378178954124, + 0.0037518134340643883, + 0.013801339082419872, + 0.0044526900164783, + 0.005214857868850231, + 0.0017524226568639278, + 0.0019393712282180786, + 0.0023223417811095715, + 0.002214489970356226, + 0.0016946772811934352, + 0.0019157363567501307, + 0.003092891536653042, + 0.002025278750807047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 7 + ], + "selected_reward": -0.020594101399183273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058405987918377 + ], + "opposite_logprob_tokens": [ + -3.8955841064453125 + ], + "logprob_reward": -0.02058405987918377, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058405987918377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047637954354286194, + 0.08579121530056, + 0.5774813294410706, + 0.032340988516807556, + 0.024699833244085312, + 0.036698199808597565, + 0.01803099364042282, + 0.07023043185472488, + 0.08595721423625946, + 0.021131813526153564 + ], + "credit_map": [ + 0.0031027882359921932, + 0.0044676996767520905, + 0.004598147701472044, + 0.008475222624838352, + 0.004967526998370886, + 0.007636788301169872, + 0.00698653282597661, + 0.007403249386698008, + 0.005394991021603346, + 0.007479275576770306, + 0.004918893799185753, + 0.014711287803947926, + 0.006793896667659283, + 0.0064956326968967915, + 0.02496647648513317, + 0.01503076683729887, + 0.014237968251109123, + 0.009937752038240433, + 0.005817041266709566, + 0.020089609548449516, + 0.018308451399207115, + 0.014996524900197983, + 0.2471894472837448, + 0.24690458178520203, + 0.001979015301913023, + 0.0019202406983822584, + 0.002956007607281208, + 0.005282978527247906, + 0.003030759282410145, + 0.004175418522208929, + 0.007524728775024414, + 0.005471842363476753, + 0.0022260427940636873, + 0.0039763133972883224, + 0.002069227397441864, + 0.0038258852437138557, + 0.003393047722056508, + 0.0021861980203539133, + 0.003866675542667508, + 0.0031564447563141584, + 0.002403451129794121, + 0.004846272524446249, + 0.0036058055702596903, + 0.005091344937682152, + 0.0076049696654081345, + 0.0036830217577517033, + 0.005607894156128168, + 0.003855440765619278, + 0.0016753488453105092, + 0.0022538118064403534, + 0.003929577302187681, + 0.0018924531759694219, + 0.001991633791476488, + 0.002074452815577388, + 0.002228956203907728, + 0.0019847594667226076, + 0.0076351589523255825, + 0.005674411077052355, + 0.006855874788016081, + 0.006359749939292669, + 0.0062268078327178955, + 0.016259169206023216, + 0.005103005096316338, + 0.016116254031658173, + 0.006999961566179991, + 0.006060274317860603, + 0.007417822256684303, + 0.022674938663840294, + 0.005731485318392515, + 0.021920137107372284, + 0.008080591447651386, + 0.00707200076431036, + 0.001650547026656568, + 0.002800426445901394, + 0.002341860206797719, + 0.0023807513061910868, + 0.0028625265695154667, + 0.0017890139715746045, + 0.004973460920155048, + 0.002333228010684252 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.02058405987918377 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549577958881855 + ], + "opposite_logprob_tokens": [ + -2.929957866668701 + ], + "logprob_reward": -0.0549577958881855, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549577958881855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06641902774572372, + 0.11666414141654968, + 0.7477008104324341, + 0.014719790779054165, + 0.0063935243524611, + 0.005366194527596235, + 0.02103610709309578, + 0.010432995855808258, + 0.00487649068236351, + 0.0063908882439136505 + ], + "credit_map": [ + 0.004491122905164957, + 0.005872542038559914, + 0.008825975470244884, + 0.006534677930176258, + 0.007535357028245926, + 0.014847200363874435, + 0.009251213632524014, + 0.009060938842594624, + 0.006292716600000858, + 0.008901399560272694, + 0.005999139044433832, + 0.020612968131899834, + 0.009608520194888115, + 0.00884383823722601, + 0.035099320113658905, + 0.02130623534321785, + 0.01829501986503601, + 0.01293669082224369, + 0.007548249792307615, + 0.02572866901755333, + 0.02326100878417492, + 0.019115950912237167, + 0.32006362080574036, + 0.3207516074180603, + 0.0008975373348221183, + 0.0007453765720129013, + 0.0011249638628214598, + 0.002866431837901473, + 0.0014958061510697007, + 0.0016942539950832725, + 0.003560510464012623, + 0.0023349113762378693, + 0.0005086440360173583, + 0.0007689771591685712, + 0.0007294592796824872, + 0.0008066726732067764, + 0.0007425039075314999, + 0.0010046062525361776, + 0.0009727227734401822, + 0.0008599385037086904, + 0.000494726060424, + 0.0009876638650894165, + 0.0009153758292086422, + 0.0007206337759271264, + 0.0005648225196637213, + 0.0005514158983714879, + 0.0005756536265835166, + 0.0005559025448746979, + 0.0019818993750959635, + 0.0036084624007344246, + 0.0025481798220425844, + 0.0022522015497088432, + 0.0035989859607070684, + 0.0024308254942297935, + 0.001824757200665772, + 0.0027907956391572952, + 0.0010078782215714455, + 0.0014152915682643652, + 0.0014357679756358266, + 0.0009535800782032311, + 0.0009893951937556267, + 0.0019187827128916979, + 0.0009598927572369576, + 0.0017524081049486995, + 0.0004344000481069088, + 0.000647693988867104, + 0.0005996685940772295, + 0.0006750498432666063, + 0.0005727834650315344, + 0.0009894527029246092, + 0.0005372049054130912, + 0.0004202375712338835, + 0.0006654972676187754, + 0.0007768393261358142, + 0.0009478171123191714, + 0.0007109253783710301, + 0.0007378046866506338, + 0.0006553409621119499, + 0.0010020775953307748, + 0.0008945857989601791 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0549577958881855 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249644163995981 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249644163995981, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249644163995981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16580672562122345, + 0.2760438621044159, + 0.12068428844213486, + 0.06178979203104973, + 0.033905029296875, + 0.030726084485650063, + 0.07213587313890457, + 0.05429486930370331, + 0.14443553984165192, + 0.040177904069423676 + ], + "credit_map": [ + 0.016364090144634247, + 0.017256615683436394, + 0.018709277734160423, + 0.02191006951034069, + 0.01624070480465889, + 0.027709733694791794, + 0.025052452459931374, + 0.02256377600133419, + 0.01549523789435625, + 0.021755021065473557, + 0.014251839369535446, + 0.04885612428188324, + 0.022228719666600227, + 0.021006684750318527, + 0.08248740434646606, + 0.049962833523750305, + 0.004703945480287075, + 0.0041976324282586575, + 0.004706183448433876, + 0.0043192580342292786, + 0.006668211426585913, + 0.006483082193881273, + 0.04522811993956566, + 0.04437785595655441, + 0.003829365596175194, + 0.003192098578438163, + 0.005454945843666792, + 0.011025841347873211, + 0.005833806004375219, + 0.006915426347404718, + 0.015178299508988857, + 0.010360009968280792, + 0.0022381434682756662, + 0.005379009060561657, + 0.00283122924156487, + 0.006480943877249956, + 0.003535689553245902, + 0.0027192262932658195, + 0.005315589718520641, + 0.005405196454375982, + 0.0026433018501847982, + 0.004785406403243542, + 0.0049222903326153755, + 0.0049080438911914825, + 0.0033125774934887886, + 0.0037970449775457382, + 0.0027267958503216505, + 0.0036306255497038364, + 0.006590308155864477, + 0.012448753230273724, + 0.00896672997623682, + 0.007618492003530264, + 0.012134235352277756, + 0.008197478950023651, + 0.006549824494868517, + 0.00963005144149065, + 0.005013640969991684, + 0.00737351318821311, + 0.007981091737747192, + 0.004761963151395321, + 0.005832735914736986, + 0.008853243663907051, + 0.00543194031342864, + 0.009046736173331738, + 0.008993539959192276, + 0.009287478402256966, + 0.01354227028787136, + 0.03465878218412399, + 0.009366552345454693, + 0.041333165019750595, + 0.012221748009324074, + 0.015032006427645683, + 0.004231033846735954, + 0.004171981476247311, + 0.007197905797511339, + 0.0032471506856381893, + 0.004325301852077246, + 0.004027968272566795, + 0.006235769018530846, + 0.0067407977767288685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.005249644163995981 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow panning shot, from a horizontal perspective, sweeps across an ancient Chinese courtyard shrouded in mist. At the center of the frame, an elderly man with white hair and youthful features, dressed in a loose, clean, traditional white cotton and linen training suit, is immersed in a Tai Chi practice. The bluestone floor beneath his feet gleams faintly with moisture, while the backdrop is a dark wooden corridor with intricate carvings and a verdant bamboo forest, faintly visible through the mist. The elderly man's expression is focused and serene, his arms drawing smooth, full arcs in a perceptibly slow rhythm, his center of gravity shifting steadily and effortlessly between his legs in sync with his breathing. Each movement is as smooth and uninterrupted as flowing water. As he assumes a pushing-hands stance, his wide sleeves flutter gently with his movements, as if parting the milky morning mist that permeates the surrounding area. The soft diffuse light of the early morning penetrates the mist, giving the entire scene a tranquil and hazy texture. The light outlines the soft contours of his white clothes and the subtle luster of his silver hair. The main colors are white, gray and green, presenting an extremely simple, elegant and peaceful visual atmosphere as a whole.", + "seed": 20505, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_1/index_7682_seed_20505_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21680056042969226, + "vlm_logprob": -0.21680056042969226, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.21680056042969226, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9849371910095215 + ], + "opposite_logprob_tokens": [ + -0.4849371910095215 + ], + "logprob_reward": -0.9849371910095215, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9849371910095215, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08997532725334167, + 0.08388284593820572, + 0.04986456781625748, + 0.04745786264538765, + 0.17898517847061157, + 0.06734391301870346, + 0.05146150663495064, + 0.12117721885442734, + 0.24951821565628052, + 0.060333333909511566 + ], + "credit_map": [ + 0.011848955415189266, + 0.010316344909369946, + 0.009465200826525688, + 0.007378212176263332, + 0.013827282004058361, + 0.011347159743309021, + 0.015393821522593498, + 0.010398357175290585, + 0.009610513225197792, + 0.01102625485509634, + 0.007165960967540741, + 0.005903145764023066, + 0.010636398568749428, + 0.0119823282584548, + 0.019320650026202202, + 0.008237596601247787, + 0.004275868646800518, + 0.007298171054571867, + 0.003785591572523117, + 0.004335050471127033, + 0.005411006044596434, + 0.00718363281339407, + 0.0074592758901417255, + 0.010115977376699448, + 0.005160608794540167, + 0.006818057503551245, + 0.00334774237126112, + 0.0038699035067111254, + 0.00933203473687172, + 0.004744326695799828, + 0.006974693387746811, + 0.007210498210042715, + 0.02230316400527954, + 0.030135495588183403, + 0.016846057027578354, + 0.011348881758749485, + 0.02179606445133686, + 0.009203887544572353, + 0.05208718776702881, + 0.015264441259205341, + 0.013610550202429295, + 0.011435400694608688, + 0.006602594628930092, + 0.004571768920868635, + 0.008192818611860275, + 0.007750960532575846, + 0.009052811190485954, + 0.0061270142905414104, + 0.005144181195646524, + 0.009617133997380733, + 0.004203756805509329, + 0.006195671856403351, + 0.007229817099869251, + 0.006249654106795788, + 0.006762387230992317, + 0.00605890853330493, + 0.016023969277739525, + 0.012448292225599289, + 0.01245188619941473, + 0.018120288848876953, + 0.017772892490029335, + 0.01257331669330597, + 0.01186015922576189, + 0.019926421344280243, + 0.023719098418951035, + 0.04711931571364403, + 0.030887840315699577, + 0.03627315163612366, + 0.03822987899184227, + 0.027744900435209274, + 0.023431891575455666, + 0.022112160921096802, + 0.008081491105258465, + 0.006525084841996431, + 0.010242907330393791, + 0.006160826422274113, + 0.011176365427672863, + 0.0052970461547374725, + 0.0066629014909267426, + 0.00618671253323555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.9849371910095215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203527010977268 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.014203527010977268, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203527010977268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0756479874253273, + 0.05606026574969292, + 0.056157417595386505, + 0.04789390414953232, + 0.0670044869184494, + 0.037884313613176346, + 0.12413569539785385, + 0.11984833329916, + 0.2546424865722656, + 0.1607251763343811 + ], + "credit_map": [ + 0.012610222212970257, + 0.010103942826390266, + 0.007024068385362625, + 0.004779600538313389, + 0.011707888916134834, + 0.0072351437993347645, + 0.010139891877770424, + 0.01204722747206688, + 0.007030182983726263, + 0.008178544230759144, + 0.003782115876674652, + 0.0038300666492432356, + 0.009643400087952614, + 0.010789897292852402, + 0.006107254885137081, + 0.006698801647871733, + 0.0045990427024662495, + 0.007957844994962215, + 0.0043266755528748035, + 0.004679630044847727, + 0.005442214198410511, + 0.006604598835110664, + 0.009643332101404667, + 0.012904076837003231, + 0.004796809051185846, + 0.005890137981623411, + 0.0034018189180642366, + 0.002881542779505253, + 0.009460192173719406, + 0.005859944969415665, + 0.009646167047321796, + 0.005957289598882198, + 0.013333187438547611, + 0.0065079559572041035, + 0.0060219536535441875, + 0.003112224629148841, + 0.009266429580748081, + 0.003693882841616869, + 0.017951752990484238, + 0.007117103319615126, + 0.007527900859713554, + 0.003116547130048275, + 0.0032955750357359648, + 0.002823061076924205, + 0.004857798106968403, + 0.003820539452135563, + 0.008195447735488415, + 0.004247442353516817, + 0.013175169937312603, + 0.018957339227199554, + 0.016200479120016098, + 0.011546735651791096, + 0.015997404232621193, + 0.017428182065486908, + 0.01609349064528942, + 0.014736894518136978, + 0.015130222775042057, + 0.012243288569152355, + 0.013295437209308147, + 0.016790296882390976, + 0.012799008749425411, + 0.01158831175416708, + 0.010277088731527328, + 0.027724672108888626, + 0.019620392471551895, + 0.04981369525194168, + 0.020956484600901604, + 0.023468270897865295, + 0.06046278029680252, + 0.022927096113562584, + 0.03266787528991699, + 0.024725882336497307, + 0.023059751838445663, + 0.02282537706196308, + 0.018235383555293083, + 0.018682114779949188, + 0.03058125451207161, + 0.014382029883563519, + 0.01786162704229355, + 0.015097638592123985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 6 + ], + "selected_reward": -0.014203527010977268 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818566583096981 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.01818566583096981, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818566583096981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10932634025812149, + 0.057596899569034576, + 0.04295362904667854, + 0.04687000438570976, + 0.18056531250476837, + 0.047709494829177856, + 0.06974248588085175, + 0.08590525388717651, + 0.24619083106517792, + 0.1131398156285286 + ], + "credit_map": [ + 0.016040677204728127, + 0.011774496175348759, + 0.00808268878608942, + 0.006826549302786589, + 0.028457999229431152, + 0.015158731490373611, + 0.011019876226782799, + 0.011965316720306873, + 0.007819823920726776, + 0.007662796415388584, + 0.004402463324368, + 0.003742233384400606, + 0.010283011011779308, + 0.009540454484522343, + 0.0065824477933347225, + 0.007563672959804535, + 0.003623535158112645, + 0.005693182349205017, + 0.002678723307326436, + 0.0025792382657527924, + 0.005746226292103529, + 0.0055784680880606174, + 0.005410519428551197, + 0.011643736623227596, + 0.005069679580628872, + 0.0057469382882118225, + 0.002676456468179822, + 0.002838907763361931, + 0.00991468969732523, + 0.005495691206306219, + 0.008878631517291069, + 0.006249009631574154, + 0.02137952297925949, + 0.02758428081870079, + 0.011772135272622108, + 0.010847666300833225, + 0.02347419410943985, + 0.009162789210677147, + 0.060895562171936035, + 0.015449165366590023, + 0.007011949084699154, + 0.004906665068119764, + 0.003635178320109844, + 0.002816910156980157, + 0.00913975015282631, + 0.0078049469739198685, + 0.00708576338365674, + 0.005308331456035376, + 0.00781242223456502, + 0.008943400345742702, + 0.009744931943714619, + 0.007355680223554373, + 0.008213217370212078, + 0.0102333789691329, + 0.00890337023884058, + 0.008536087349057198, + 0.010116131976246834, + 0.00905249547213316, + 0.008228468708693981, + 0.012131537310779095, + 0.01227437611669302, + 0.008079593069851398, + 0.00805170089006424, + 0.01797095313668251, + 0.018150819465517998, + 0.04617820307612419, + 0.028812691569328308, + 0.034990984946489334, + 0.04556628316640854, + 0.022318875417113304, + 0.028599873185157776, + 0.021573081612586975, + 0.013445376418530941, + 0.015924954786896706, + 0.01717032678425312, + 0.01923912949860096, + 0.01387731172144413, + 0.01405927911400795, + 0.00918673723936081, + 0.010236697271466255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.01818566583096981 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204111501574516 + ], + "opposite_logprob_tokens": [ + -2.8120410442352295 + ], + "logprob_reward": -0.06204111501574516, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204111501574516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11781889945268631, + 0.052491869777441025, + 0.058894507586956024, + 0.051507603377103806, + 0.10346163064241409, + 0.07011440396308899, + 0.15721148252487183, + 0.10021024197340012, + 0.16786278784275055, + 0.12042660266160965 + ], + "credit_map": [ + 0.024872055277228355, + 0.0144810127094388, + 0.011096027679741383, + 0.009514079429209232, + 0.016344362869858742, + 0.011373146437108517, + 0.015396881848573685, + 0.014741318300366402, + 0.006244659889489412, + 0.0061536491848528385, + 0.004835979547351599, + 0.003389807417988777, + 0.0072549087926745415, + 0.011193987913429737, + 0.007390856742858887, + 0.00602801563218236, + 0.007712945342063904, + 0.007834980264306068, + 0.005151205696165562, + 0.008711234666407108, + 0.007539467420428991, + 0.0074309599585831165, + 0.006868058815598488, + 0.007645647972822189, + 0.008691229857504368, + 0.005691877566277981, + 0.004245101939886808, + 0.002712876768782735, + 0.008708212524652481, + 0.005813242867588997, + 0.010950970463454723, + 0.004694084636867046, + 0.015458053909242153, + 0.012592153623700142, + 0.006192232947796583, + 0.005272681824862957, + 0.012895542196929455, + 0.004647072404623032, + 0.03686102107167244, + 0.009542867541313171, + 0.009929881431162357, + 0.01195789035409689, + 0.005662387702614069, + 0.004729609005153179, + 0.008872759528458118, + 0.005182742606848478, + 0.01642410084605217, + 0.0073550185188651085, + 0.01998164691030979, + 0.029611436650156975, + 0.015018894337117672, + 0.01948673650622368, + 0.016370991244912148, + 0.01769069954752922, + 0.019084816798567772, + 0.019966239109635353, + 0.01588464342057705, + 0.009681214578449726, + 0.01090026181191206, + 0.015099202282726765, + 0.011338971555233002, + 0.011480162851512432, + 0.013036697171628475, + 0.012789081782102585, + 0.012818561866879463, + 0.027341609820723534, + 0.021456381306052208, + 0.028073741123080254, + 0.02608458511531353, + 0.017124487087130547, + 0.018256019800901413, + 0.016707394272089005, + 0.014080047607421875, + 0.011290805414319038, + 0.012332836166024208, + 0.012696528807282448, + 0.020550159737467766, + 0.018873386085033417, + 0.011134812608361244, + 0.01946801133453846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.06204111501574516 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004635303281247616 + ], + "opposite_logprob_tokens": [ + -5.379635334014893 + ], + "logprob_reward": -0.004635303281247616, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004635303281247616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1136053055524826, + 0.0687132179737091, + 0.07107110321521759, + 0.03771664574742317, + 0.17370888590812683, + 0.05400586873292923, + 0.04075389727950096, + 0.11960399895906448, + 0.21155638992786407, + 0.10926467925310135 + ], + "credit_map": [ + 0.019638441503047943, + 0.012980937026441097, + 0.01030732411891222, + 0.008841237984597683, + 0.019121112301945686, + 0.011593479663133621, + 0.014604144729673862, + 0.016518624499440193, + 0.0077889179810881615, + 0.00979855377227068, + 0.00480737304314971, + 0.004905402194708586, + 0.011215255595743656, + 0.011323153041303158, + 0.01199246570467949, + 0.006882098037749529, + 0.004767784848809242, + 0.00905095599591732, + 0.004504939541220665, + 0.004659508354961872, + 0.007161173969507217, + 0.00858356710523367, + 0.014501260593533516, + 0.017841916531324387, + 0.0038525003474205732, + 0.0036974677350372076, + 0.0026090219616889954, + 0.0021862266585230827, + 0.008518627844750881, + 0.005180951673537493, + 0.006999534089118242, + 0.004672313574701548, + 0.019398607313632965, + 0.028447838500142097, + 0.014637995511293411, + 0.010537020862102509, + 0.022773703560233116, + 0.008831207640469074, + 0.054412368685007095, + 0.014670132659375668, + 0.01310879085212946, + 0.008865490555763245, + 0.005048098508268595, + 0.0035223881714046, + 0.005412820726633072, + 0.004749628249555826, + 0.00783761590719223, + 0.00546103622764349, + 0.003992641344666481, + 0.005882792640477419, + 0.005019211210310459, + 0.0042647747322916985, + 0.007237032055854797, + 0.004852250684052706, + 0.005526446271687746, + 0.003978749271482229, + 0.01762429066002369, + 0.013015439733862877, + 0.012812978588044643, + 0.01683136075735092, + 0.014857969246804714, + 0.013386552222073078, + 0.011731773614883423, + 0.01934364065527916, + 0.01759457215666771, + 0.03808172047138214, + 0.027001196518540382, + 0.03295290842652321, + 0.031594134867191315, + 0.022083453834056854, + 0.020721018314361572, + 0.021527379751205444, + 0.012154251337051392, + 0.011237766593694687, + 0.009816428646445274, + 0.011938189156353474, + 0.023701321333646774, + 0.016165446490049362, + 0.01139663252979517, + 0.01285464409738779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.004635303281247616 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow panning shot, from a horizontal perspective, sweeps across an ancient Chinese courtyard shrouded in mist. At the center of the frame, an elderly man with white hair and youthful features, dressed in a loose, clean, traditional white cotton and linen training suit, is immersed in a Tai Chi practice. The bluestone floor beneath his feet gleams faintly with moisture, while the backdrop is a dark wooden corridor with intricate carvings and a verdant bamboo forest, faintly visible through the mist. The elderly man's expression is focused and serene, his arms drawing smooth, full arcs in a perceptibly slow rhythm, his center of gravity shifting steadily and effortlessly between his legs in sync with his breathing. Each movement is as smooth and uninterrupted as flowing water. As he assumes a pushing-hands stance, his wide sleeves flutter gently with his movements, as if parting the milky morning mist that permeates the surrounding area. The soft diffuse light of the early morning penetrates the mist, giving the entire scene a tranquil and hazy texture. The light outlines the soft contours of his white clothes and the subtle luster of his silver hair. The main colors are white, gray and green, presenting an extremely simple, elegant and peaceful visual atmosphere as a whole.", + "seed": 20506, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_1/index_7682_seed_20506_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.34406972397118807, + "vlm_logprob": -0.34406972397118807, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.34406972397118807, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6040990352630615 + ], + "opposite_logprob_tokens": [ + -0.22909903526306152 + ], + "logprob_reward": -1.6040990352630615, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6040990352630615, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07036053389310837, + 0.11715912818908691, + 0.1929258555173874, + 0.11119458824396133, + 0.12898635864257812, + 0.06729436665773392, + 0.18251670897006989, + 0.03320296108722687, + 0.05248425900936127, + 0.04387521371245384 + ], + "credit_map": [ + 0.006768427323549986, + 0.00971632543951273, + 0.006769832689315081, + 0.008263704366981983, + 0.01133810356259346, + 0.009142346680164337, + 0.01164417527616024, + 0.006717616226524115, + 0.013011028058826923, + 0.018204238265752792, + 0.016349708661437035, + 0.018021157011389732, + 0.016006356105208397, + 0.012193134985864162, + 0.01153473649173975, + 0.011838768608868122, + 0.024216629564762115, + 0.020328551530838013, + 0.014408189803361893, + 0.033339690417051315, + 0.023912156000733376, + 0.030633961781859398, + 0.024840310215950012, + 0.021246371790766716, + 0.013379676267504692, + 0.007653775159269571, + 0.005928301252424717, + 0.04164687544107437, + 0.007638279348611832, + 0.004980043042451143, + 0.025042761117219925, + 0.0049248673021793365, + 0.01718532294034958, + 0.023433055728673935, + 0.01409347914159298, + 0.0174239594489336, + 0.018121987581253052, + 0.00888090766966343, + 0.018418487161397934, + 0.011429155245423317, + 0.005456461571156979, + 0.016054604202508926, + 0.0053397356532514095, + 0.007289971690624952, + 0.011341902427375317, + 0.005159846972674131, + 0.011235740035772324, + 0.005416098982095718, + 0.017469920217990875, + 0.02071954868733883, + 0.02245955541729927, + 0.03050658106803894, + 0.02431277371942997, + 0.022404715418815613, + 0.02444380335509777, + 0.020199820399284363, + 0.004220863338559866, + 0.006468398030847311, + 0.003432498313486576, + 0.004265486728399992, + 0.0038550267927348614, + 0.0025940705090761185, + 0.0042917183600366116, + 0.004074897617101669, + 0.006549433805048466, + 0.005684513598680496, + 0.004477331414818764, + 0.00779426796361804, + 0.005948765203356743, + 0.00903221033513546, + 0.008910357020795345, + 0.004087379667907953, + 0.00652038911357522, + 0.0037769931368529797, + 0.0058448524214327335, + 0.005554089788347483, + 0.005468199960887432, + 0.00560537027195096, + 0.006623277440667152, + 0.004482043907046318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 4 + ], + "selected_reward": -1.6040990352630615 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029820770025253296 + ], + "opposite_logprob_tokens": [ + -3.529820680618286 + ], + "logprob_reward": -0.029820770025253296, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029820770025253296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12260694056749344, + 0.08495363593101501, + 0.20169828832149506, + 0.04795191064476967, + 0.07725952565670013, + 0.24427326023578644, + 0.098627008497715, + 0.03685890510678291, + 0.042815372347831726, + 0.042955152690410614 + ], + "credit_map": [ + 0.012989174574613571, + 0.014054917730391026, + 0.01626180112361908, + 0.01655278354883194, + 0.021686848253011703, + 0.010325294919312, + 0.01789778284728527, + 0.012838342227041721, + 0.008224177174270153, + 0.011662349104881287, + 0.011003718711435795, + 0.009668265469372272, + 0.015735110267996788, + 0.007317045703530312, + 0.011189348995685577, + 0.010153613053262234, + 0.019473228603601456, + 0.023494752123951912, + 0.017662685364484787, + 0.031203707680106163, + 0.03187134116888046, + 0.0300592053681612, + 0.027733784168958664, + 0.02019958570599556, + 0.0049467082135379314, + 0.004033008590340614, + 0.003396710380911827, + 0.015524886548519135, + 0.004447455983608961, + 0.003447166411206126, + 0.009807461872696877, + 0.0023485126439481974, + 0.009534585289657116, + 0.01618703454732895, + 0.009068180806934834, + 0.010903997346758842, + 0.010197809897363186, + 0.005505009088665247, + 0.009389449842274189, + 0.006473464425653219, + 0.02193385548889637, + 0.0491865873336792, + 0.023266170173883438, + 0.032187316566705704, + 0.034825973212718964, + 0.016977643594145775, + 0.04273028299212456, + 0.023165423423051834, + 0.010898672975599766, + 0.01016957126557827, + 0.013339824043214321, + 0.017132800072431564, + 0.011639476753771305, + 0.012602719478309155, + 0.011278943158686161, + 0.011564997024834156, + 0.00380593235604465, + 0.006942959502339363, + 0.002904831897467375, + 0.004727071616798639, + 0.005698549561202526, + 0.0032008786220103502, + 0.00565293338149786, + 0.003925744444131851, + 0.005488342139869928, + 0.004772727843374014, + 0.004697574768215418, + 0.005612005013972521, + 0.004865699447691441, + 0.007169241551309824, + 0.007127243559807539, + 0.003082536393776536, + 0.00661886902526021, + 0.0038487876299768686, + 0.005028311628848314, + 0.006220494396984577, + 0.005488166119903326, + 0.004679644014686346, + 0.006851093377918005, + 0.004219788126647472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.029820770025253296 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810877725481987 + ], + "opposite_logprob_tokens": [ + -3.2881088256835938 + ], + "logprob_reward": -0.03810877725481987, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810877725481987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1349639743566513, + 0.07472734153270721, + 0.2195756584405899, + 0.06412342935800552, + 0.16481998562812805, + 0.07149355113506317, + 0.08860199898481369, + 0.054877426475286484, + 0.08559729903936386, + 0.04121939092874527 + ], + "credit_map": [ + 0.011640344746410847, + 0.01722899079322815, + 0.011524870991706848, + 0.01928548328578472, + 0.026514695957303047, + 0.014371645636856556, + 0.020367005839943886, + 0.014030924998223782, + 0.007036570459604263, + 0.013060666620731354, + 0.0094198202714324, + 0.008522095158696175, + 0.014237593859434128, + 0.005975174251943827, + 0.008047458715736866, + 0.008427965454757214, + 0.02511673979461193, + 0.02384316362440586, + 0.0176080334931612, + 0.03458615764975548, + 0.030616963282227516, + 0.0354992114007473, + 0.028082774952054024, + 0.02422260120511055, + 0.005870572756975889, + 0.007828285917639732, + 0.005150963552296162, + 0.015163896605372429, + 0.008339856751263142, + 0.005714936647564173, + 0.012164799496531487, + 0.0038901129737496376, + 0.023524466902017593, + 0.02883199229836464, + 0.01659367047250271, + 0.021241653710603714, + 0.02465958707034588, + 0.010631917975842953, + 0.023974087089300156, + 0.015362605452537537, + 0.006419665180146694, + 0.011414656415581703, + 0.006510816514492035, + 0.010130536742508411, + 0.007903456687927246, + 0.00821166206151247, + 0.012064699083566666, + 0.008838050067424774, + 0.00785590149462223, + 0.010123525746166706, + 0.010266455821692944, + 0.015298767015337944, + 0.011218767613172531, + 0.012275606393814087, + 0.011123014613986015, + 0.010439963079988956, + 0.0068920510821044445, + 0.00905096810311079, + 0.00768598448485136, + 0.007545349188148975, + 0.007225301116704941, + 0.004450561478734016, + 0.006367085035890341, + 0.005660120397806168, + 0.00896135438233614, + 0.013301782310009003, + 0.007631837390363216, + 0.009632891975343227, + 0.010490634478628635, + 0.011938844807446003, + 0.016103854402899742, + 0.007536094170063734, + 0.0076018390245735645, + 0.003879884257912636, + 0.00564097985625267, + 0.005460924468934536, + 0.0039254105649888515, + 0.0044730911031365395, + 0.0058309962041676044, + 0.004406265914440155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.03810877725481987 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430665984749794 + ], + "opposite_logprob_tokens": [ + -3.1680665016174316 + ], + "logprob_reward": -0.0430665984749794, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430665984749794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18763458728790283, + 0.07480260729789734, + 0.10460737347602844, + 0.0664178803563118, + 0.055006250739097595, + 0.2425796091556549, + 0.09383072704076767, + 0.0528150238096714, + 0.052921704947948456, + 0.06938423216342926 + ], + "credit_map": [ + 0.02449219860136509, + 0.023473279550671577, + 0.02182810753583908, + 0.024950597435235977, + 0.028406288474798203, + 0.017039570957422256, + 0.024231869727373123, + 0.02321266196668148, + 0.007820742204785347, + 0.010164378210902214, + 0.010350559838116169, + 0.008774038404226303, + 0.010594346560537815, + 0.010484158992767334, + 0.00837886892259121, + 0.008235509507358074, + 0.009363416582345963, + 0.012245389632880688, + 0.009006143547594547, + 0.014541040174663067, + 0.014783889055252075, + 0.016808632761240005, + 0.01585845649242401, + 0.012000402435660362, + 0.005908388178795576, + 0.009864386171102524, + 0.006983344443142414, + 0.011703338474035263, + 0.00787501223385334, + 0.008420174941420555, + 0.009303870610892773, + 0.006359370891004801, + 0.008213545195758343, + 0.009122679010033607, + 0.006289668846875429, + 0.0057737501338124275, + 0.007529031950980425, + 0.004577605053782463, + 0.007245233282446861, + 0.006254741922020912, + 0.020820822566747665, + 0.052637979388237, + 0.013897009193897247, + 0.024344846606254578, + 0.046651896089315414, + 0.03669115900993347, + 0.02872695028781891, + 0.018808959051966667, + 0.009601873345673084, + 0.008610548451542854, + 0.012296628206968307, + 0.016149362549185753, + 0.010857515037059784, + 0.013616056181490421, + 0.011069982312619686, + 0.011628757230937481, + 0.005425024311989546, + 0.008837331086397171, + 0.007923166267573833, + 0.006662634667009115, + 0.006077687721699476, + 0.004121978767216206, + 0.008497058413922787, + 0.005270141642540693, + 0.006456549745053053, + 0.0067721144296228886, + 0.007646315731108189, + 0.005685293581336737, + 0.006013429723680019, + 0.009098390117287636, + 0.007047226652503014, + 0.0042023793794214725, + 0.010918944142758846, + 0.006471209228038788, + 0.007008995395153761, + 0.010226799175143242, + 0.00896427407860756, + 0.010238002054393291, + 0.008321129716932774, + 0.0072348760440945625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.0430665984749794 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow pan sweeping across the scene from a horizontal perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central subject an elderly man with white hair wearing a loose white cotton and linen training suit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man performing Tai Chi movements with smooth, full arcs and a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting an ancient Chinese courtyard with a bluestone floor, dark wooden corridor, and verdant bamboo forest visible through mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft diffuse early morning light creating a tranquil, hazy texture with white, gray, and green colors?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253438837826252 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253438837826252, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253438837826252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12697915732860565, + 0.10687819868326187, + 0.14149877429008484, + 0.07585228234529495, + 0.11261861026287079, + 0.05471718683838844, + 0.1971670240163803, + 0.05283384025096893, + 0.07047504186630249, + 0.060979925096035004 + ], + "credit_map": [ + 0.014476003125309944, + 0.01626075617969036, + 0.012364639900624752, + 0.015274737030267715, + 0.01698797196149826, + 0.017059193924069405, + 0.01767050288617611, + 0.016885336488485336, + 0.010326727293431759, + 0.016842951998114586, + 0.013721855357289314, + 0.014218748547136784, + 0.015094177797436714, + 0.012552321888506413, + 0.012045369483530521, + 0.012076037004590034, + 0.01660490594804287, + 0.014215401373803616, + 0.011655482463538647, + 0.0233449749648571, + 0.016648419201374054, + 0.022328482940793037, + 0.02031056582927704, + 0.016390519216656685, + 0.0076436870731413364, + 0.011694133281707764, + 0.00620027631521225, + 0.012093598023056984, + 0.010298021137714386, + 0.011839966289699078, + 0.01016221847385168, + 0.005920378025621176, + 0.014865342527627945, + 0.023354126140475273, + 0.009989559650421143, + 0.01421449426561594, + 0.017289118841290474, + 0.00730071309953928, + 0.015622053295373917, + 0.009983193129301071, + 0.00509314751252532, + 0.007038692943751812, + 0.004718074109405279, + 0.006134776398539543, + 0.00748644582927227, + 0.011913125403225422, + 0.007389825768768787, + 0.00494309701025486, + 0.019105637446045876, + 0.021167447790503502, + 0.02400139346718788, + 0.03193844109773636, + 0.026246987283229828, + 0.025352124124765396, + 0.026866033673286438, + 0.022488942369818687, + 0.007998104207217693, + 0.006229338236153126, + 0.0062722438015043736, + 0.009227459318935871, + 0.006442880257964134, + 0.004791508428752422, + 0.007061873562633991, + 0.004810430109500885, + 0.010067019611597061, + 0.007035290822386742, + 0.008206349797546864, + 0.007555128075182438, + 0.008506393060088158, + 0.013719682581722736, + 0.010298950597643852, + 0.005086225923150778, + 0.010672681033611298, + 0.004904864355921745, + 0.006381014361977577, + 0.009060016833245754, + 0.007618237752467394, + 0.007951225154101849, + 0.0074616409838199615, + 0.006930237635970116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.005253438837826252 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_10/index_tensor([5196, 5196])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_10/index_tensor([5196, 5196])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1bb49c57fec199caf95e813b402a77ca70932f97 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_10/index_tensor([5196, 5196])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a medium shot from a sideways angle, then slowly zooms in, ultimately settling on a serene close-up of her profile. In a simple, inviting bedroom, warm, golden afternoon sunlight slants through a large floor-to-ceiling window, creating a brilliant beam of light. Countless tiny particles of light slowly float, twirl, and dance in the air like golden elves. At the center of the frame, a young woman, dressed in a loose white cotton dress, gracefully stretches one arm overhead, her fingertips gently reaching toward the sky. The light illuminates her curves, making them appear exceptionally soft and slender. Her long, flowing hair, a delicate blend of soft rainbow hues—pale pink, light purple, sky blue, and mint green—cascades through the strands, casting a dreamy, almost translucent halo around the edges. As the camera zooms in, her focused and serene profile is captured clearly: her fair cheeks are sharply defined, her long lashes cast a subtle shadow across her eyelids, her lips slightly parted, her breathing steady and deep. The entire scene is imbued with the delicate light and shadow and dreamy atmosphere that are the hallmarks of Makoto Shinkai's animated films, with the dominant colors of warm gold and soft white creating a feeling of ultimate tranquility, purity, and vitality.", + "seed": 8614, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_10/index_5196_seed_8614_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.014028008375316859, + "vlm_logprob": -0.014028008375316859, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.014028008375316859, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029156623408198357 + ], + "opposite_logprob_tokens": [ + -3.6541566848754883 + ], + "logprob_reward": -0.029156623408198357, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.029156623408198357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13688841462135315, + 0.13232997059822083, + 0.033404041081666946, + 0.037962399423122406, + 0.09177376329898834, + 0.07545019686222076, + 0.41088080406188965, + 0.024884097278118134, + 0.02261299267411232, + 0.03381328657269478 + ], + "credit_map": [ + 0.017273874953389168, + 0.017142467200756073, + 0.013406949117779732, + 0.017264463007450104, + 0.013818377628922462, + 0.015259730629622936, + 0.023628156632184982, + 0.019094401970505714, + 0.02536948397755623, + 0.011274361051619053, + 0.006924645509570837, + 0.046100009232759476, + 0.009242535568773746, + 0.010654803365468979, + 0.011571281589567661, + 0.01119284424930811, + 0.005527718458324671, + 0.006860535126179457, + 0.003215154865756631, + 0.004296380560845137, + 0.003833688097074628, + 0.0031061808113008738, + 0.004018182400614023, + 0.002546201227232814, + 0.004252157174050808, + 0.0038371800910681486, + 0.0036486214958131313, + 0.007636894006282091, + 0.005985044874250889, + 0.002798910718411207, + 0.004217713605612516, + 0.005585875827819109, + 0.005223193671554327, + 0.006127957254648209, + 0.02116238884627819, + 0.032077446579933167, + 0.006751372944563627, + 0.005615920759737492, + 0.008326255716383457, + 0.006489228922873735, + 0.010008555836975574, + 0.007320981472730637, + 0.008049668744206429, + 0.011979318223893642, + 0.008000042289495468, + 0.0068280077539384365, + 0.008835072629153728, + 0.014428545720875263, + 0.03960154205560684, + 0.01477038487792015, + 0.13875392079353333, + 0.014672618359327316, + 0.028661442920565605, + 0.11026358604431152, + 0.03953874856233597, + 0.024618566036224365, + 0.002334935124963522, + 0.003001393284648657, + 0.002180033829063177, + 0.002623253734782338, + 0.003921077586710453, + 0.002266479656100273, + 0.004240323323756456, + 0.004316601902246475, + 0.002965456573292613, + 0.002819352550432086, + 0.002863300498574972, + 0.00320111820474267, + 0.003065922064706683, + 0.0028032350819557905, + 0.002627259586006403, + 0.002267348114401102, + 0.004325187765061855, + 0.00431424006819725, + 0.0031836279667913914, + 0.005717274732887745, + 0.0051474422216415405, + 0.002106711734086275, + 0.006481919437646866, + 0.0025368817150592804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.029156623408198357 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023305542767047882 + ], + "opposite_logprob_tokens": [ + -3.773305654525757 + ], + "logprob_reward": -0.023305542767047882, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023305542767047882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10114319622516632, + 0.16150489449501038, + 0.029442736878991127, + 0.027472222223877907, + 0.08050890266895294, + 0.04640107974410057, + 0.4592270851135254, + 0.028692277148365974, + 0.035388946533203125, + 0.030218660831451416 + ], + "credit_map": [ + 0.010713283903896809, + 0.01144951581954956, + 0.008136741816997528, + 0.009987362660467625, + 0.015611529350280762, + 0.010731575079262257, + 0.021617170423269272, + 0.012896024622023106, + 0.04150582477450371, + 0.009739499539136887, + 0.00629751943051815, + 0.0648004561662674, + 0.0057063838467001915, + 0.009053568355739117, + 0.011830075643956661, + 0.012571566738188267, + 0.0038358501624315977, + 0.004463654477149248, + 0.002699492033571005, + 0.0060419682413339615, + 0.003644948359578848, + 0.002342579187825322, + 0.0044392733834683895, + 0.0019749714992940426, + 0.003387053031474352, + 0.002499670023098588, + 0.002864450216293335, + 0.0035980294924229383, + 0.0028856508433818817, + 0.004483375232666731, + 0.0026993206702172756, + 0.0050546759739518166, + 0.004670283291488886, + 0.005742814391851425, + 0.01857393980026245, + 0.028700893744826317, + 0.006459561642259359, + 0.004826350603252649, + 0.006704764440655708, + 0.00483029754832387, + 0.004023279063403606, + 0.005141064990311861, + 0.004880364052951336, + 0.01052048522979021, + 0.004514617845416069, + 0.00400950014591217, + 0.00641649030148983, + 0.006895281840115786, + 0.045293137431144714, + 0.01517529971897602, + 0.15290352702140808, + 0.01717834733426571, + 0.028578368946909904, + 0.13890044391155243, + 0.04006931185722351, + 0.021128658205270767, + 0.0030641129706054926, + 0.0031710625626146793, + 0.004412801004946232, + 0.0027338676154613495, + 0.003832513466477394, + 0.003004168625921011, + 0.0033601534087210894, + 0.005113598890602589, + 0.0046027647331357, + 0.0032290355302393436, + 0.004555013496428728, + 0.003105939831584692, + 0.006558008026331663, + 0.005046552512794733, + 0.004140289034694433, + 0.0041513447649776936, + 0.0037985111121088266, + 0.0037825782783329487, + 0.002574437065050006, + 0.004258466884493828, + 0.004697149619460106, + 0.0017591334180906415, + 0.006174515467137098, + 0.003173871897161007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.023305542767047882 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007635810878127813 + ], + "opposite_logprob_tokens": [ + -4.882635593414307 + ], + "logprob_reward": -0.007635810878127813, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007635810878127813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03516778349876404, + 0.07339763641357422, + 0.013710006140172482, + 0.022165102884173393, + 0.019122088328003883, + 0.04196402430534363, + 0.7411454319953918, + 0.02006102167069912, + 0.015631377696990967, + 0.017635509371757507 + ], + "credit_map": [ + 0.004568808246403933, + 0.004278179723769426, + 0.0027262612711638212, + 0.005154484882950783, + 0.005115802865475416, + 0.0037671944592148066, + 0.0055219014175236225, + 0.004035147838294506, + 0.017658105120062828, + 0.00522775249555707, + 0.0031022957991808653, + 0.02657143957912922, + 0.004455311223864555, + 0.004106592852622271, + 0.0070032100193202496, + 0.005272928159683943, + 0.0015574832214042544, + 0.002763541415333748, + 0.0016090588178485632, + 0.0023822651710361242, + 0.0013086356921121478, + 0.0011484430870041251, + 0.0019258829997852445, + 0.0010146952699869871, + 0.002154194051399827, + 0.0030631867703050375, + 0.0024811485782265663, + 0.003602032084017992, + 0.002631934592500329, + 0.0019021229818463326, + 0.0025383238680660725, + 0.0037921604234725237, + 0.0019315803656354547, + 0.0026579888071864843, + 0.002455370035022497, + 0.002537692664191127, + 0.002892368007451296, + 0.0016295351088047028, + 0.002595663769170642, + 0.002421889454126358, + 0.004863422363996506, + 0.005194564815610647, + 0.004604650661349297, + 0.0054200091399252415, + 0.007014393340796232, + 0.003693496808409691, + 0.005853532813489437, + 0.005319956690073013, + 0.07748646289110184, + 0.02312259003520012, + 0.25524061918258667, + 0.02446015551686287, + 0.029947888106107712, + 0.23929031193256378, + 0.06042419373989105, + 0.031173251569271088, + 0.0025258127134293318, + 0.001605757512152195, + 0.002979680197313428, + 0.0014632140519097447, + 0.0021227989345788956, + 0.0029754871502518654, + 0.0019057474564760923, + 0.004482523538172245, + 0.001783059211447835, + 0.001604403369128704, + 0.002044719411060214, + 0.0015746362041682005, + 0.0020318280439823866, + 0.0020578664261847734, + 0.0017158127157017589, + 0.002819051733240485, + 0.00189129076898098, + 0.0018256268231198192, + 0.0011906606378033757, + 0.0023994375951588154, + 0.0018759705126285553, + 0.0025631170719861984, + 0.0028352898079901934, + 0.0030541159212589264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.007635810878127813 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005946089047938585 + ], + "opposite_logprob_tokens": [ + -5.130946159362793 + ], + "logprob_reward": -0.005946089047938585, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005946089047938585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04549913853406906, + 0.1837431788444519, + 0.041975025087594986, + 0.0779041051864624, + 0.35234788060188293, + 0.043952517211437225, + 0.18031881749629974, + 0.026157300919294357, + 0.02797023206949234, + 0.020131763070821762 + ], + "credit_map": [ + 0.00499649066478014, + 0.004573957063257694, + 0.0036222157068550587, + 0.004667120520025492, + 0.006137524265795946, + 0.004053634591400623, + 0.010271571576595306, + 0.007176624145358801, + 0.04665644094347954, + 0.012302283197641373, + 0.007638965733349323, + 0.07202906906604767, + 0.007451050914824009, + 0.009623647667467594, + 0.013258962891995907, + 0.01478276401758194, + 0.004404060076922178, + 0.00747993728145957, + 0.004493115935474634, + 0.008916111662983894, + 0.0048280437476933, + 0.002524329349398613, + 0.006033110432326794, + 0.0032963166013360023, + 0.008554408326745033, + 0.013933190144598484, + 0.009037005715072155, + 0.008966580964624882, + 0.01048524584621191, + 0.012146953493356705, + 0.006890516262501478, + 0.007890203036367893, + 0.01894761621952057, + 0.024471348151564598, + 0.0877922773361206, + 0.13375769555568695, + 0.023013779893517494, + 0.02121400088071823, + 0.028440235182642937, + 0.014710907824337482, + 0.005938136950135231, + 0.00531834177672863, + 0.004832134582102299, + 0.0071767535991966724, + 0.005452126730233431, + 0.0037059530150145292, + 0.004126791376620531, + 0.007402278017252684, + 0.019024662673473358, + 0.005397321656346321, + 0.06005498394370079, + 0.0068554687313735485, + 0.007619855459779501, + 0.058106549084186554, + 0.01516041811555624, + 0.00809955783188343, + 0.0032759513705968857, + 0.002910837298259139, + 0.005022798199206591, + 0.0023327735252678394, + 0.002858585212379694, + 0.0037345357704907656, + 0.002509607933461666, + 0.0035122111439704895, + 0.004036787431687117, + 0.003919141832739115, + 0.0039100199937820435, + 0.0037216227501630783, + 0.004298236686736345, + 0.0024786407593637705, + 0.0037585911341011524, + 0.0018471896182745695, + 0.003210618859156966, + 0.002081002341583371, + 0.0018997353035956621, + 0.0036931242793798447, + 0.002688782289624214, + 0.0011613831156864762, + 0.0036642353516072035, + 0.0017328807152807713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 6 + ], + "selected_reward": -0.005946089047938585 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095975775271654 + ], + "opposite_logprob_tokens": [ + -5.504096031188965 + ], + "logprob_reward": -0.004095975775271654, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095975775271654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08844751864671707, + 0.29267168045043945, + 0.0328960083425045, + 0.0644923523068428, + 0.24618178606033325, + 0.11459997296333313, + 0.0754944309592247, + 0.03650599345564842, + 0.021350812166929245, + 0.027359487488865852 + ], + "credit_map": [ + 0.009930213913321495, + 0.013261924497783184, + 0.011011400260031223, + 0.00948486477136612, + 0.009976528584957123, + 0.012132437899708748, + 0.013531322591006756, + 0.009118824265897274, + 0.07694201171398163, + 0.0179186649620533, + 0.01386954914778471, + 0.11387830227613449, + 0.010291390120983124, + 0.015691500157117844, + 0.02133513055741787, + 0.022745124995708466, + 0.004019056912511587, + 0.00391651876270771, + 0.004595689009875059, + 0.004232470877468586, + 0.005808990448713303, + 0.002584187313914299, + 0.005266297608613968, + 0.0024727978743612766, + 0.005806701257824898, + 0.012303872033953667, + 0.006725310813635588, + 0.013061415404081345, + 0.0076440549455583096, + 0.007180996239185333, + 0.006356453988701105, + 0.005413547158241272, + 0.01318584755063057, + 0.0182125773280859, + 0.04649513587355614, + 0.07864326983690262, + 0.031247789040207863, + 0.0158174280077219, + 0.022536063566803932, + 0.020043669268488884, + 0.0176747627556324, + 0.017145581543445587, + 0.011182409711182117, + 0.015925157815217972, + 0.014439117163419724, + 0.013472762890160084, + 0.011636300012469292, + 0.01312387827783823, + 0.007004681974649429, + 0.007341119460761547, + 0.017114805057644844, + 0.005913096014410257, + 0.008726791478693485, + 0.01254764199256897, + 0.010742409154772758, + 0.006103881634771824, + 0.005402426701039076, + 0.003952734172344208, + 0.006250977516174316, + 0.0037949641700834036, + 0.004786373116075993, + 0.003594465320929885, + 0.004346763249486685, + 0.004377290606498718, + 0.002862051595002413, + 0.002429954009130597, + 0.0022840576712042093, + 0.0023506172001361847, + 0.0033835440408438444, + 0.0026015860494226217, + 0.0028789835050702095, + 0.002560017164796591, + 0.0037236397620290518, + 0.0031964597292244434, + 0.0020402693189680576, + 0.004073674324899912, + 0.004121148027479649, + 0.002696719253435731, + 0.0039037903770804405, + 0.003603785764425993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 5 + ], + "selected_reward": -0.004095975775271654 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a medium shot from a sideways angle, then slowly zooms in, ultimately settling on a serene close-up of her profile. In a simple, inviting bedroom, warm, golden afternoon sunlight slants through a large floor-to-ceiling window, creating a brilliant beam of light. Countless tiny particles of light slowly float, twirl, and dance in the air like golden elves. At the center of the frame, a young woman, dressed in a loose white cotton dress, gracefully stretches one arm overhead, her fingertips gently reaching toward the sky. The light illuminates her curves, making them appear exceptionally soft and slender. Her long, flowing hair, a delicate blend of soft rainbow hues—pale pink, light purple, sky blue, and mint green—cascades through the strands, casting a dreamy, almost translucent halo around the edges. As the camera zooms in, her focused and serene profile is captured clearly: her fair cheeks are sharply defined, her long lashes cast a subtle shadow across her eyelids, her lips slightly parted, her breathing steady and deep. The entire scene is imbued with the delicate light and shadow and dreamy atmosphere that are the hallmarks of Makoto Shinkai's animated films, with the dominant colors of warm gold and soft white creating a feeling of ultimate tranquility, purity, and vitality.", + "seed": 8615, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_10/index_5196_seed_8615_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.030564704537391664, + "vlm_logprob": -0.030564704537391664, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.030564704537391664, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05450228974223137 + ], + "opposite_logprob_tokens": [ + -3.054502248764038 + ], + "logprob_reward": -0.05450228974223137, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05450228974223137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45438623428344727, + 0.06580427289009094, + 0.045249857008457184, + 0.06011709198355675, + 0.06903457641601562, + 0.07559751719236374, + 0.0664951428771019, + 0.05656490474939346, + 0.05202546343207359, + 0.054724857211112976 + ], + "credit_map": [ + 0.06045817211270332, + 0.09849036484956741, + 0.04049444571137428, + 0.0654478520154953, + 0.04544360563158989, + 0.04106083884835243, + 0.052656278014183044, + 0.050334710627794266, + 0.010248919017612934, + 0.0038225522730499506, + 0.009791500866413116, + 0.009208915755152702, + 0.007731294259428978, + 0.008542731404304504, + 0.008443678729236126, + 0.008014688268303871, + 0.007664162199944258, + 0.0026208977214992046, + 0.006970692425966263, + 0.008903590962290764, + 0.005209174007177353, + 0.0037384151946753263, + 0.004765646532177925, + 0.0053772819228470325, + 0.008123680017888546, + 0.00830396544188261, + 0.007908005267381668, + 0.012640842236578465, + 0.005384760443121195, + 0.007900048978626728, + 0.004845382180064917, + 0.005010412540286779, + 0.008128130808472633, + 0.006049414165318012, + 0.013625863008201122, + 0.011936893686652184, + 0.007700695656239986, + 0.006520070135593414, + 0.007729694712907076, + 0.007343822158873081, + 0.011981504037976265, + 0.011594143696129322, + 0.012581110931932926, + 0.008750570937991142, + 0.008838944137096405, + 0.008026647381484509, + 0.00802653469145298, + 0.005798074882477522, + 0.006639646831899881, + 0.008677355945110321, + 0.00926310382783413, + 0.011980598792433739, + 0.011166146025061607, + 0.0052548423409461975, + 0.009583323262631893, + 0.003930128179490566, + 0.009502388536930084, + 0.008542395196855068, + 0.008345643058419228, + 0.009486804716289043, + 0.007142157759517431, + 0.003498110920190811, + 0.006602175533771515, + 0.0034452383406460285, + 0.009387523867189884, + 0.010058026760816574, + 0.007289006374776363, + 0.004816242028027773, + 0.009360089898109436, + 0.0027947593480348587, + 0.005676692351698875, + 0.0026431262958794832, + 0.010160036385059357, + 0.009990465827286243, + 0.003496961435303092, + 0.0064401482231915, + 0.009026882238686085, + 0.004584158770740032, + 0.006050604395568371, + 0.004975602962076664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.05450228974223137 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700511634349823 + ], + "opposite_logprob_tokens": [ + -2.6950511932373047 + ], + "logprob_reward": -0.0700511634349823, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700511634349823, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37124550342559814, + 0.09090369194746017, + 0.06979948282241821, + 0.05276356264948845, + 0.08072810620069504, + 0.07841785997152328, + 0.11212284862995148, + 0.052209511399269104, + 0.061587683856487274, + 0.0302217248827219 + ], + "credit_map": [ + 0.06303074955940247, + 0.04913725703954697, + 0.04685135930776596, + 0.06715076416730881, + 0.0297516081482172, + 0.03234883397817612, + 0.03614765778183937, + 0.046827275305986404, + 0.011722002178430557, + 0.0040092165581882, + 0.009085413068532944, + 0.007830881513655186, + 0.01465262845158577, + 0.012153176590800285, + 0.0159904807806015, + 0.015459900721907616, + 0.008392185904085636, + 0.004392685834318399, + 0.01028076745569706, + 0.012933369725942612, + 0.008053943514823914, + 0.00723645556718111, + 0.009293519891798496, + 0.009216556325554848, + 0.004557000007480383, + 0.005363079719245434, + 0.006524054799228907, + 0.007287719752639532, + 0.006162502337247133, + 0.007522985339164734, + 0.006039276719093323, + 0.009306947700679302, + 0.008689896203577518, + 0.006596000865101814, + 0.011784299276769161, + 0.009685051627457142, + 0.011793536134064198, + 0.009555180557072163, + 0.011541507206857204, + 0.011082641780376434, + 0.010679696686565876, + 0.010259275324642658, + 0.008246568031609058, + 0.006154333241283894, + 0.012952852062880993, + 0.010333549231290817, + 0.011760849505662918, + 0.008030736818909645, + 0.010101786814630032, + 0.013616178184747696, + 0.01580539345741272, + 0.01741746999323368, + 0.020882869139313698, + 0.010621142573654652, + 0.01592489331960678, + 0.007753121666610241, + 0.0050454228185117245, + 0.006288777571171522, + 0.005643865559250116, + 0.005594766233116388, + 0.008973301388323307, + 0.0053548491559922695, + 0.008530128747224808, + 0.006778405979275703, + 0.007408027071505785, + 0.006881068926304579, + 0.006972203496843576, + 0.006475070025771856, + 0.009884942322969437, + 0.005379339214414358, + 0.012221838347613811, + 0.006365197245031595, + 0.004250239580869675, + 0.004781836178153753, + 0.0022742552682757378, + 0.002437272574752569, + 0.005803578067570925, + 0.0032885028049349785, + 0.0039927344769239426, + 0.0033933089580386877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0700511634349823 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086177080869675 + ], + "opposite_logprob_tokens": [ + -4.5110859870910645 + ], + "logprob_reward": -0.011086177080869675, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086177080869675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10880868881940842, + 0.047493427991867065, + 0.07278028130531311, + 0.06772197782993317, + 0.07576204836368561, + 0.06970633566379547, + 0.1095556989312172, + 0.13087183237075806, + 0.11234165728092194, + 0.2049580216407776 + ], + "credit_map": [ + 0.013625694438815117, + 0.012576581910252571, + 0.017694074660539627, + 0.012058097869157791, + 0.010936199687421322, + 0.010332752019166946, + 0.01604025438427925, + 0.015545033849775791, + 0.007212005089968443, + 0.0029591857455670834, + 0.007798722945153713, + 0.0054787928238511086, + 0.005559482146054506, + 0.00499500147998333, + 0.007417652290314436, + 0.006072586867958307, + 0.00950541254132986, + 0.0043462361209094524, + 0.009460633620619774, + 0.01619603857398033, + 0.00861989613622427, + 0.007141290232539177, + 0.007099190726876259, + 0.010411578230559826, + 0.005582488141953945, + 0.00666307657957077, + 0.013638099655508995, + 0.01021003071218729, + 0.005294693633913994, + 0.00999713595956564, + 0.005198249593377113, + 0.011138202622532845, + 0.00824169535189867, + 0.009418910369277, + 0.005696008447557688, + 0.004430609289556742, + 0.014141494408249855, + 0.01516763400286436, + 0.008680669590830803, + 0.009985026903450489, + 0.00808627624064684, + 0.008890857920050621, + 0.006112949922680855, + 0.0038608908653259277, + 0.01374234352260828, + 0.008726637810468674, + 0.012619887478649616, + 0.007666491437703371, + 0.008134320378303528, + 0.013046075589954853, + 0.01127243135124445, + 0.012236190028488636, + 0.020388364791870117, + 0.018678152933716774, + 0.014318590052425861, + 0.011481578461825848, + 0.009167619049549103, + 0.0164762195199728, + 0.0165802463889122, + 0.012906593270599842, + 0.034072112292051315, + 0.012972020544111729, + 0.017775196582078934, + 0.010921827517449856, + 0.02204255945980549, + 0.01957881636917591, + 0.013516944833099842, + 0.01389043778181076, + 0.014681363478302956, + 0.008047343231737614, + 0.015714552253484726, + 0.0048696366138756275, + 0.022595912218093872, + 0.026502611115574837, + 0.010941154323518276, + 0.018464067950844765, + 0.04232587665319443, + 0.021747523918747902, + 0.035806719213724136, + 0.026574160903692245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.011086177080869675 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464171078056097 + ], + "opposite_logprob_tokens": [ + -5.379641532897949 + ], + "logprob_reward": -0.00464171078056097, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464171078056097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14898765087127686, + 0.09263549745082855, + 0.09406894445419312, + 0.10139153897762299, + 0.0978161171078682, + 0.14471739530563354, + 0.1795383244752884, + 0.02750624157488346, + 0.08852705359458923, + 0.02481122687458992 + ], + "credit_map": [ + 0.022866740822792053, + 0.026560556143522263, + 0.019920358434319496, + 0.020663365721702576, + 0.01117218192666769, + 0.014587719924747944, + 0.010114498436450958, + 0.02310222201049328, + 0.01114620454609394, + 0.004652510862797499, + 0.010419595055282116, + 0.008507420308887959, + 0.010977860540151596, + 0.02380692958831787, + 0.010349972173571587, + 0.01277499832212925, + 0.014059696346521378, + 0.004075932316482067, + 0.014235702343285084, + 0.020481843501329422, + 0.009359206072986126, + 0.01026491355150938, + 0.012350101955235004, + 0.009241541847586632, + 0.007217989303171635, + 0.013556063175201416, + 0.02018238604068756, + 0.010534828528761864, + 0.010876706801354885, + 0.011888515204191208, + 0.00685055460780859, + 0.020284486934542656, + 0.009323830716311932, + 0.007635569199919701, + 0.015215697698295116, + 0.010101004503667355, + 0.013187797740101814, + 0.012028035707771778, + 0.011014677584171295, + 0.019309502094984055, + 0.019523805007338524, + 0.01954531855881214, + 0.015794949606060982, + 0.018157504498958588, + 0.02164272777736187, + 0.01710798218846321, + 0.01996731013059616, + 0.012977785430848598, + 0.016490889713168144, + 0.022666169330477715, + 0.021464288234710693, + 0.023843517526984215, + 0.03475828096270561, + 0.022899556905031204, + 0.022660385817289352, + 0.014755211770534515, + 0.005491434596478939, + 0.0045450725592672825, + 0.0030193275306373835, + 0.0035663412418216467, + 0.0036731830332428217, + 0.0017563562141731381, + 0.004045653156936169, + 0.0014088707976043224, + 0.016602247953414917, + 0.014331364072859287, + 0.008358890190720558, + 0.008459425531327724, + 0.015641383826732635, + 0.005290190223604441, + 0.015068098902702332, + 0.004775445442646742, + 0.004287408664822578, + 0.003839093493297696, + 0.0022526520770043135, + 0.00345554924570024, + 0.003928959369659424, + 0.0015469399513676763, + 0.0034576382022351027, + 0.0020429857540875673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.00464171078056097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542181648314 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.012542181648314, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542181648314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33242082595825195, + 0.10537639260292053, + 0.12203533947467804, + 0.06823714822530746, + 0.07246742397546768, + 0.06633245199918747, + 0.07998295873403549, + 0.064405657351017, + 0.04828858748078346, + 0.0404532253742218 + ], + "credit_map": [ + 0.05385684594511986, + 0.0346495695412159, + 0.048354022204875946, + 0.049917370080947876, + 0.026914555579423904, + 0.042234089225530624, + 0.03032856248319149, + 0.046165820211172104, + 0.024548785760998726, + 0.008989079855382442, + 0.012548195198178291, + 0.009809964336454868, + 0.006061812397092581, + 0.017798395827412605, + 0.00745587470009923, + 0.01816428266465664, + 0.015146031975746155, + 0.009176265448331833, + 0.01979980245232582, + 0.01927684061229229, + 0.013584435917437077, + 0.015071254223585129, + 0.016423316672444344, + 0.013557384721934795, + 0.006045117974281311, + 0.008938438259065151, + 0.014671655371785164, + 0.013170682825148106, + 0.0058952029794454575, + 0.008748439140617847, + 0.004263838287442923, + 0.006503771990537643, + 0.008001653477549553, + 0.006442453246563673, + 0.011629892513155937, + 0.009910491295158863, + 0.009481568820774555, + 0.007910690270364285, + 0.010178253054618835, + 0.00891241803765297, + 0.01092167105525732, + 0.008494684472680092, + 0.013452674262225628, + 0.007897977717220783, + 0.010524272918701172, + 0.00483899237588048, + 0.006514388602226973, + 0.0036877894308418036, + 0.009058615192770958, + 0.011419348418712616, + 0.009956062771379948, + 0.008201077580451965, + 0.016149763017892838, + 0.008952812291681767, + 0.009505204856395721, + 0.0067400760017335415, + 0.010604125447571278, + 0.010468505322933197, + 0.0066810669377446175, + 0.006295759696513414, + 0.01119889598339796, + 0.005198741797357798, + 0.00987187959253788, + 0.004086680710315704, + 0.008074008859694004, + 0.008208908140659332, + 0.00457545043900609, + 0.006767068989574909, + 0.0066022020764648914, + 0.003205740824341774, + 0.006990326102823019, + 0.003864882979542017, + 0.007675325497984886, + 0.005324774421751499, + 0.004121625795960426, + 0.0039306506514549255, + 0.005510380025953054, + 0.004972614347934723, + 0.004072688519954681, + 0.004845166578888893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.012542181648314 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a medium shot from a sideways angle, then slowly zooms in, ultimately settling on a serene close-up of her profile. In a simple, inviting bedroom, warm, golden afternoon sunlight slants through a large floor-to-ceiling window, creating a brilliant beam of light. Countless tiny particles of light slowly float, twirl, and dance in the air like golden elves. At the center of the frame, a young woman, dressed in a loose white cotton dress, gracefully stretches one arm overhead, her fingertips gently reaching toward the sky. The light illuminates her curves, making them appear exceptionally soft and slender. Her long, flowing hair, a delicate blend of soft rainbow hues—pale pink, light purple, sky blue, and mint green—cascades through the strands, casting a dreamy, almost translucent halo around the edges. As the camera zooms in, her focused and serene profile is captured clearly: her fair cheeks are sharply defined, her long lashes cast a subtle shadow across her eyelids, her lips slightly parted, her breathing steady and deep. The entire scene is imbued with the delicate light and shadow and dreamy atmosphere that are the hallmarks of Makoto Shinkai's animated films, with the dominant colors of warm gold and soft white creating a feeling of ultimate tranquility, purity, and vitality.", + "seed": 8616, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_10/index_5196_seed_8616_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.018704094202257694, + "vlm_logprob": -0.018704094202257694, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.018704094202257694, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048976749181747437 + ], + "opposite_logprob_tokens": [ + -3.1739766597747803 + ], + "logprob_reward": -0.048976749181747437, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.048976749181747437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07551153004169464, + 0.05368857458233833, + 0.05641297250986099, + 0.2282266467809677, + 0.15010999143123627, + 0.15797059237957, + 0.11853768676519394, + 0.08461747318506241, + 0.032158538699150085, + 0.04276600480079651 + ], + "credit_map": [ + 0.006812439300119877, + 0.006845899857580662, + 0.009863663464784622, + 0.006577599793672562, + 0.010244326665997505, + 0.011037720367312431, + 0.016994845122098923, + 0.007135030347853899, + 0.0060204709880054, + 0.004783390089869499, + 0.010240382514894009, + 0.004647030960768461, + 0.005866403225809336, + 0.008209017105400562, + 0.00685823755338788, + 0.007063641212880611, + 0.004734942689538002, + 0.0048287250101566315, + 0.008548073470592499, + 0.0029950521420687437, + 0.008343259803950787, + 0.008683635853230953, + 0.013917965814471245, + 0.004361317493021488, + 0.02145237661898136, + 0.02722480334341526, + 0.02859584242105484, + 0.012377668172121048, + 0.0444820411503315, + 0.022272812202572823, + 0.04113328456878662, + 0.030687810853123665, + 0.01888423040509224, + 0.016343174502253532, + 0.016626538708806038, + 0.008620803244411945, + 0.02123488299548626, + 0.01643730141222477, + 0.03754616528749466, + 0.01441689021885395, + 0.028172267600893974, + 0.014228849671781063, + 0.0325482003390789, + 0.01807946152985096, + 0.02543466165661812, + 0.012696079909801483, + 0.014368459582328796, + 0.012442604638636112, + 0.018502933904528618, + 0.014333401806652546, + 0.01413798052817583, + 0.016056668013334274, + 0.023504741489887238, + 0.008227464742958546, + 0.011821579188108444, + 0.011952920816838741, + 0.011715463362634182, + 0.0043120491318404675, + 0.010188719257712364, + 0.007840249687433243, + 0.011725195683538914, + 0.013153066858649254, + 0.011231633834540844, + 0.014451093040406704, + 0.0038741212338209152, + 0.0030814644414931536, + 0.0037481493782252073, + 0.002957722404971719, + 0.005286076106131077, + 0.004146687686443329, + 0.004602119792252779, + 0.004462195094674826, + 0.0055396053940057755, + 0.006805542856454849, + 0.004616750404238701, + 0.004577721003443003, + 0.008299626410007477, + 0.004684663377702236, + 0.00515409279614687, + 0.003088005119934678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.048976749181747437 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023309269919991493 + ], + "opposite_logprob_tokens": [ + -3.7733092308044434 + ], + "logprob_reward": -0.023309269919991493, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023309269919991493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10671927034854889, + 0.021472014486789703, + 0.025308195501565933, + 0.0511978343129158, + 0.03386951610445976, + 0.04977136850357056, + 0.4710301160812378, + 0.12884250283241272, + 0.051023609936237335, + 0.06076556816697121 + ], + "credit_map": [ + 0.009782186709344387, + 0.01800711080431938, + 0.01534486934542656, + 0.02033987082540989, + 0.011965776793658733, + 0.008321131579577923, + 0.016659284010529518, + 0.00629903469234705, + 0.0016524073434993625, + 0.0032746910583227873, + 0.0025518741458654404, + 0.003713805926963687, + 0.002315254881978035, + 0.00242388597689569, + 0.0033706058748066425, + 0.002169488463550806, + 0.0029287273064255714, + 0.004379142075777054, + 0.002156185219064355, + 0.002146537182852626, + 0.003495541401207447, + 0.004356900230050087, + 0.0027970881201326847, + 0.00304807280190289, + 0.005359826609492302, + 0.006596769671887159, + 0.005277401767671108, + 0.0035640988498926163, + 0.010123739019036293, + 0.004678791854530573, + 0.0077765025198459625, + 0.00782070029526949, + 0.003906410187482834, + 0.004655396565794945, + 0.00356275774538517, + 0.0031384401954710484, + 0.005901414901018143, + 0.004060022067278624, + 0.006144640035927296, + 0.002500435570254922, + 0.006507623475044966, + 0.004267864860594273, + 0.008591274730861187, + 0.006896303966641426, + 0.008548473007977009, + 0.004471229389309883, + 0.005446308292448521, + 0.005042295437306166, + 0.06034569814801216, + 0.059465501457452774, + 0.04832739010453224, + 0.06100224703550339, + 0.10518909990787506, + 0.0359715111553669, + 0.05162689462304115, + 0.04910176619887352, + 0.012628627941012383, + 0.007340013515204191, + 0.014369637705385685, + 0.008338158950209618, + 0.026646146550774574, + 0.023620055988430977, + 0.018723567947745323, + 0.01717628911137581, + 0.006195008754730225, + 0.005166677292436361, + 0.004534625913947821, + 0.004859596956521273, + 0.01216149888932705, + 0.005565151106566191, + 0.007835716009140015, + 0.004705330822616816, + 0.00464697415009141, + 0.005526505410671234, + 0.004390504211187363, + 0.0066596223041415215, + 0.015744438394904137, + 0.007407187484204769, + 0.008918064646422863, + 0.007472269702702761 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.023309269919991493 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081814765930176 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081814765930176, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081814765930176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07724493741989136, + 0.03802109509706497, + 0.045231521129608154, + 0.13744626939296722, + 0.23123785853385925, + 0.12612184882164001, + 0.13501189649105072, + 0.14440912008285522, + 0.036236561834812164, + 0.029038943350315094 + ], + "credit_map": [ + 0.006947533693164587, + 0.01200258918106556, + 0.012597315944731236, + 0.011422927491366863, + 0.008578243665397167, + 0.008324745111167431, + 0.01200052909553051, + 0.0053710551001131535, + 0.003618157934397459, + 0.0058591971173882484, + 0.004776143003255129, + 0.005365719087421894, + 0.003880593227222562, + 0.005473756697028875, + 0.004778315778821707, + 0.004269212018698454, + 0.006437757983803749, + 0.01076376810669899, + 0.00476138386875391, + 0.004421256482601166, + 0.0055121565237641335, + 0.0031982099171727896, + 0.005405801814049482, + 0.004731186665594578, + 0.015361213125288486, + 0.021386591717600822, + 0.016833310946822166, + 0.00964355282485485, + 0.02325712889432907, + 0.011174322105944157, + 0.0201865267008543, + 0.01960362121462822, + 0.030768558382987976, + 0.017446666955947876, + 0.02859937958419323, + 0.026060935109853745, + 0.03621654585003853, + 0.01639505848288536, + 0.06075156480073929, + 0.01499913726001978, + 0.020945489406585693, + 0.009872667491436005, + 0.024789853021502495, + 0.015018169768154621, + 0.02307709865272045, + 0.010527163743972778, + 0.01225266233086586, + 0.00963873416185379, + 0.02285083755850792, + 0.016594139859080315, + 0.017570100724697113, + 0.016486572101712227, + 0.02430886961519718, + 0.010282413102686405, + 0.013900754041969776, + 0.013018200173974037, + 0.02199571393430233, + 0.0066375103779137135, + 0.018959319218993187, + 0.011909577995538712, + 0.020488254725933075, + 0.027011966332793236, + 0.01919572986662388, + 0.018211042508482933, + 0.005736628547310829, + 0.00381761253811419, + 0.004808092024177313, + 0.0039027733728289604, + 0.005665589589625597, + 0.0036962218582630157, + 0.0034407670609652996, + 0.005168875679373741, + 0.0034197939094156027, + 0.004114414099603891, + 0.002624266780912876, + 0.002741189207881689, + 0.007397167384624481, + 0.0034861445892602205, + 0.0030457309912890196, + 0.0022102368529886007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.011081814765930176 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638276718556881 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638276718556881, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638276718556881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1591726690530777, + 0.07633538544178009, + 0.05053636059165001, + 0.06729280203580856, + 0.03395530581474304, + 0.052316825836896896, + 0.1546240597963333, + 0.13841596245765686, + 0.13648945093154907, + 0.13086117804050446 + ], + "credit_map": [ + 0.011347667314112186, + 0.02055639587342739, + 0.015964699909090996, + 0.025341833010315895, + 0.02263559214770794, + 0.010376968421041965, + 0.041564974933862686, + 0.011384527198970318, + 0.007255613338202238, + 0.00963825173676014, + 0.008441416546702385, + 0.012348946183919907, + 0.00896418932825327, + 0.01100245676934719, + 0.010007021948695183, + 0.008677477948367596, + 0.0049475072883069515, + 0.004325046204030514, + 0.007486424408853054, + 0.0038061465602368116, + 0.00834800023585558, + 0.006791416555643082, + 0.009884392842650414, + 0.0049474239349365234, + 0.006916097830981016, + 0.0067513673566281796, + 0.007509605959057808, + 0.0044145104475319386, + 0.017076360061764717, + 0.005722368601709604, + 0.013345294632017612, + 0.005557189229875803, + 0.004242862109094858, + 0.004205632023513317, + 0.004750596359372139, + 0.003766847075894475, + 0.005799243226647377, + 0.004394920077174902, + 0.0036307252012193203, + 0.0031644790433347225, + 0.008816277608275414, + 0.004172856919467449, + 0.01148219220340252, + 0.005028035957366228, + 0.007685335353016853, + 0.004759917035698891, + 0.0051536280661821365, + 0.0052185761742293835, + 0.01980833150446415, + 0.01931215636432171, + 0.017016854137182236, + 0.020215660333633423, + 0.031562868505716324, + 0.011857442557811737, + 0.016813473775982857, + 0.01803726516664028, + 0.014102381654083729, + 0.010335837490856647, + 0.014553132466971874, + 0.014597125351428986, + 0.02523772604763508, + 0.015024952590465546, + 0.02556312456727028, + 0.019001668319106102, + 0.015413591638207436, + 0.01731538400053978, + 0.01513647474348545, + 0.01442556269466877, + 0.024339882656931877, + 0.02162305638194084, + 0.015257852151989937, + 0.012977637350559235, + 0.01473651546984911, + 0.015697428956627846, + 0.013469447381794453, + 0.014627482742071152, + 0.025648467242717743, + 0.013993917964398861, + 0.019064420834183693, + 0.013623502105474472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.008638276718556881 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015143604250624776 + ], + "opposite_logprob_tokens": [ + -6.501514434814453 + ], + "logprob_reward": -0.0015143604250624776, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015143604250624776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11971424520015717, + 0.24359963834285736, + 0.05644726753234863, + 0.041345275938510895, + 0.10844723880290985, + 0.19071367383003235, + 0.052046068012714386, + 0.06602832674980164, + 0.06679362058639526, + 0.054864633828401566 + ], + "credit_map": [ + 0.00883158203214407, + 0.014000962488353252, + 0.014015068300068378, + 0.014269497245550156, + 0.017174864187836647, + 0.012989967130124569, + 0.03024210035800934, + 0.008190210908651352, + 0.026791607961058617, + 0.019589534029364586, + 0.041242972016334534, + 0.019905684515833855, + 0.027436261996626854, + 0.03726275637745857, + 0.03604022040963173, + 0.03533058241009712, + 0.005437174346297979, + 0.010451150126755238, + 0.006444786209613085, + 0.005136234685778618, + 0.007063369732350111, + 0.007098621688783169, + 0.010130797512829304, + 0.004685130901634693, + 0.004417573567479849, + 0.004281631670892239, + 0.005902986507862806, + 0.0025674982462078333, + 0.007169736083596945, + 0.004919227212667465, + 0.006508611608296633, + 0.005578012205660343, + 0.014771336689591408, + 0.01200095284730196, + 0.01729258894920349, + 0.0062934341840445995, + 0.016446366906166077, + 0.012909835204482079, + 0.020307593047618866, + 0.008425132371485233, + 0.033135272562503815, + 0.015260467305779457, + 0.03884829953312874, + 0.020031219348311424, + 0.03412444517016411, + 0.016471218317747116, + 0.018015597015619278, + 0.014827149920165539, + 0.004599170293658972, + 0.006145753897726536, + 0.003576116869226098, + 0.0059217712841928005, + 0.012132775969803333, + 0.00612898962572217, + 0.007185408379882574, + 0.006356078200042248, + 0.00862062443047762, + 0.005185094196349382, + 0.009212242439389229, + 0.00817572046071291, + 0.00959701742976904, + 0.006774192210286856, + 0.009360486641526222, + 0.00910294707864523, + 0.011983147822320461, + 0.008236282505095005, + 0.008331944234669209, + 0.00656578503549099, + 0.007562978193163872, + 0.006811663508415222, + 0.00602676160633564, + 0.011275057680904865, + 0.0078625762835145, + 0.007888300344347954, + 0.006582451052963734, + 0.004949069581925869, + 0.007725060917437077, + 0.005767289083451033, + 0.005311147775501013, + 0.008778736926615238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.0015143604250624776 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a medium shot from a sideways angle, then slowly zooms in, ultimately settling on a serene close-up of her profile. In a simple, inviting bedroom, warm, golden afternoon sunlight slants through a large floor-to-ceiling window, creating a brilliant beam of light. Countless tiny particles of light slowly float, twirl, and dance in the air like golden elves. At the center of the frame, a young woman, dressed in a loose white cotton dress, gracefully stretches one arm overhead, her fingertips gently reaching toward the sky. The light illuminates her curves, making them appear exceptionally soft and slender. Her long, flowing hair, a delicate blend of soft rainbow hues—pale pink, light purple, sky blue, and mint green—cascades through the strands, casting a dreamy, almost translucent halo around the edges. As the camera zooms in, her focused and serene profile is captured clearly: her fair cheeks are sharply defined, her long lashes cast a subtle shadow across her eyelids, her lips slightly parted, her breathing steady and deep. The entire scene is imbued with the delicate light and shadow and dreamy atmosphere that are the hallmarks of Makoto Shinkai's animated films, with the dominant colors of warm gold and soft white creating a feeling of ultimate tranquility, purity, and vitality.", + "seed": 8617, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_10/index_5196_seed_8617_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.006516500003635883, + "vlm_logprob": -0.006516500003635883, + "vlm_yes_no_margin": 5.375, + "vlm_token_credit": -0.006516500003635883, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014488139189779758 + ], + "opposite_logprob_tokens": [ + -4.389488220214844 + ], + "logprob_reward": -0.014488139189779758, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.014488139189779758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2636640667915344, + 0.10016554594039917, + 0.05663909390568733, + 0.04564310982823372, + 0.05893199145793915, + 0.08690603077411652, + 0.1003403440117836, + 0.08734554052352905, + 0.07443489134311676, + 0.12592937052249908 + ], + "credit_map": [ + 0.04429829493165016, + 0.03657059744000435, + 0.05292481556534767, + 0.025453360751271248, + 0.017734482884407043, + 0.0376090481877327, + 0.023746181279420853, + 0.0253272894769907, + 0.013216002844274044, + 0.008226354606449604, + 0.026212727651000023, + 0.01265387237071991, + 0.008094025775790215, + 0.014050769619643688, + 0.0072270343080163, + 0.010484766215085983, + 0.006032330449670553, + 0.006931581534445286, + 0.010139305144548416, + 0.0045167068019509315, + 0.004953890107572079, + 0.009219122119247913, + 0.0076569681987166405, + 0.0071891904808580875, + 0.00857633538544178, + 0.00562043022364378, + 0.009770089760422707, + 0.0034683526027947664, + 0.004221124108880758, + 0.005937823094427586, + 0.003572353394702077, + 0.00447660218924284, + 0.006458790972828865, + 0.012494303286075592, + 0.004926825407892466, + 0.010902042500674725, + 0.00807527918368578, + 0.00524898199364543, + 0.005934408865869045, + 0.0048913611099123955, + 0.020617036148905754, + 0.011933909729123116, + 0.010760190896689892, + 0.009227421134710312, + 0.010556008666753769, + 0.010273495689034462, + 0.005976072512567043, + 0.007561895530670881, + 0.019898872822523117, + 0.012171854265034199, + 0.012555012479424477, + 0.012614027597010136, + 0.00909368321299553, + 0.009448166936635971, + 0.012372320517897606, + 0.012186408042907715, + 0.010717990808188915, + 0.008976064622402191, + 0.015391522087156773, + 0.009892096742987633, + 0.007612504065036774, + 0.015540752559900284, + 0.006533635780215263, + 0.012680969201028347, + 0.010399787686765194, + 0.00839296355843544, + 0.016455523669719696, + 0.008225255645811558, + 0.008465430699288845, + 0.011147604323923588, + 0.0052701253443956375, + 0.006078199949115515, + 0.013571052812039852, + 0.020763728767633438, + 0.01759805530309677, + 0.012872403487563133, + 0.011245238594710827, + 0.014049106277525425, + 0.014009465463459492, + 0.0218203142285347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.014488139189779758 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007642672397196293 + ], + "opposite_logprob_tokens": [ + -4.88264274597168 + ], + "logprob_reward": -0.007642672397196293, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007642672397196293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1119406670331955, + 0.12174845486879349, + 0.04424988478422165, + 0.076172836124897, + 0.05326630175113678, + 0.047182466834783554, + 0.13082438707351685, + 0.0647236779332161, + 0.1206352561712265, + 0.22925610840320587 + ], + "credit_map": [ + 0.014626869931817055, + 0.011780187487602234, + 0.014768377877771854, + 0.007447694428265095, + 0.011671178042888641, + 0.017832212150096893, + 0.015862006694078445, + 0.017952144145965576, + 0.012696197256445885, + 0.014730933122336864, + 0.018846241757273674, + 0.010886387899518013, + 0.015250885859131813, + 0.017943672835826874, + 0.0189292524009943, + 0.012464881874620914, + 0.003752070013433695, + 0.004709848668426275, + 0.004895481746643782, + 0.004615734331309795, + 0.0060386876575648785, + 0.005599879194051027, + 0.006754451431334019, + 0.00788373313844204, + 0.01396877970546484, + 0.008508763276040554, + 0.017780570313334465, + 0.006629777140915394, + 0.005775931756943464, + 0.009902916848659515, + 0.005389355588704348, + 0.008216743357479572, + 0.005354033783078194, + 0.007056868635118008, + 0.006401180289685726, + 0.006037414539605379, + 0.009261095896363258, + 0.00516061345115304, + 0.006748206447809935, + 0.007246886380016804, + 0.008627424947917461, + 0.006362915970385075, + 0.006722071208059788, + 0.00511909881606698, + 0.006086886860430241, + 0.005060711409896612, + 0.004191990476101637, + 0.005011365283280611, + 0.021088309586048126, + 0.02077522873878479, + 0.01984989270567894, + 0.012755343690514565, + 0.018597986549139023, + 0.012760098092257977, + 0.013778505846858025, + 0.011219033040106297, + 0.011325376108288765, + 0.005117090418934822, + 0.012007520534098148, + 0.007150669582188129, + 0.007214418612420559, + 0.009476397186517715, + 0.004834800027310848, + 0.007597400341182947, + 0.015166204422712326, + 0.02438427321612835, + 0.008459096774458885, + 0.011579726822674274, + 0.017971396446228027, + 0.016439106315374374, + 0.014964376576244831, + 0.011671077460050583, + 0.02980375662446022, + 0.043611783534288406, + 0.027132200077176094, + 0.026625750586390495, + 0.02375008910894394, + 0.021074354648590088, + 0.03124050609767437, + 0.026017656549811363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.007642672397196293 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067406888119876385 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.0067406888119876385, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067406888119876385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15835049748420715, + 0.05958672985434532, + 0.047040168195962906, + 0.06253465265035629, + 0.07042073458433151, + 0.07055123895406723, + 0.04173321649432182, + 0.08845187723636627, + 0.1527976095676422, + 0.24853332340717316 + ], + "credit_map": [ + 0.016207648441195488, + 0.01901298388838768, + 0.029914909973740578, + 0.022815490141510963, + 0.016430305317044258, + 0.02208929881453514, + 0.01510284747928381, + 0.016777005046606064, + 0.009106435813009739, + 0.005269678309559822, + 0.012828640639781952, + 0.005421385634690523, + 0.006953235249966383, + 0.008528901264071465, + 0.004783546086400747, + 0.006694905459880829, + 0.006394132040441036, + 0.005178913939744234, + 0.0066717397421598434, + 0.003646907163783908, + 0.004856706131249666, + 0.00951092317700386, + 0.004937691148370504, + 0.0058431560173630714, + 0.00886479765176773, + 0.005567651707679033, + 0.020201196894049644, + 0.004283294081687927, + 0.006414076313376427, + 0.008042735978960991, + 0.005176547449082136, + 0.003984352573752403, + 0.010025130584836006, + 0.012606147676706314, + 0.008218859322369099, + 0.011838288977742195, + 0.007784044370055199, + 0.006724298931658268, + 0.007526029366999865, + 0.005697931163012981, + 0.017024492844939232, + 0.009392033331096172, + 0.010105443187057972, + 0.007387146819382906, + 0.007102596573531628, + 0.0069575984962284565, + 0.005658816080540419, + 0.00692311255261302, + 0.005707834847271442, + 0.004809330217540264, + 0.004954983945935965, + 0.005789205431938171, + 0.004019357729703188, + 0.005588151514530182, + 0.003875019494444132, + 0.006989333778619766, + 0.010206899605691433, + 0.00981096737086773, + 0.01809079945087433, + 0.011434648185968399, + 0.007878672331571579, + 0.013732796534895897, + 0.006583198439329863, + 0.010713892057538033, + 0.019999166950583458, + 0.025467760860919952, + 0.018961895257234573, + 0.018558800220489502, + 0.014980084262788296, + 0.020846372470259666, + 0.016135873273015022, + 0.017847662791609764, + 0.028493693098425865, + 0.03843636065721512, + 0.03000396303832531, + 0.03190824016928673, + 0.027231933549046516, + 0.028630752116441727, + 0.03294100612401962, + 0.03088737279176712 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.0067406888119876385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021969962399452925 + ], + "opposite_logprob_tokens": [ + -6.127196788787842 + ], + "logprob_reward": -0.0021969962399452925, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021969962399452925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07743749022483826, + 0.04740053787827492, + 0.04901620373129845, + 0.10827544331550598, + 0.06647337973117828, + 0.11168037354946136, + 0.10992464423179626, + 0.09580174088478088, + 0.10438010096549988, + 0.2296101152896881 + ], + "credit_map": [ + 0.013235506601631641, + 0.008223018608987331, + 0.013175783678889275, + 0.00630932254716754, + 0.006529492326080799, + 0.010741867125034332, + 0.011562963016331196, + 0.007659535855054855, + 0.0044293999671936035, + 0.0048762778751552105, + 0.009631717577576637, + 0.005275339353829622, + 0.004813352134078741, + 0.007679223082959652, + 0.005432190373539925, + 0.005263038445264101, + 0.003997523337602615, + 0.005187275819480419, + 0.009741386398673058, + 0.00443914532661438, + 0.004585078451782465, + 0.006876206956803799, + 0.006368221249431372, + 0.00782136432826519, + 0.015239983797073364, + 0.00974617712199688, + 0.029141025617718697, + 0.009375003166496754, + 0.010092055425047874, + 0.013771109282970428, + 0.00789695605635643, + 0.013013136573135853, + 0.007113732397556305, + 0.010974653996527195, + 0.007515724282711744, + 0.010410510003566742, + 0.009840713813900948, + 0.006407172419130802, + 0.006137392949312925, + 0.008073481731116772, + 0.023847078904509544, + 0.014662053436040878, + 0.016416387632489204, + 0.01180713064968586, + 0.013982316479086876, + 0.012423396110534668, + 0.009053678251802921, + 0.009488335810601711, + 0.017354903742671013, + 0.013765104115009308, + 0.0175866037607193, + 0.011808437295258045, + 0.013402723707258701, + 0.010394974611699581, + 0.012221050448715687, + 0.013390854001045227, + 0.014547064900398254, + 0.009783701971173286, + 0.02371559478342533, + 0.013795089907944202, + 0.008071289397776127, + 0.009716327302157879, + 0.007115451619029045, + 0.009057225659489632, + 0.013128959573805332, + 0.017492087557911873, + 0.01778321899473667, + 0.01641981117427349, + 0.01279508788138628, + 0.008559432812035084, + 0.009931981563568115, + 0.008269516751170158, + 0.03230680152773857, + 0.048304446041584015, + 0.037013113498687744, + 0.03238897770643234, + 0.02021886594593525, + 0.01589200645685196, + 0.025928126648068428, + 0.017557773739099503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.0021969962399452925 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a bedroom with a large floor-to-ceiling window?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young woman wearing a loose white cotton dress visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the woman have long hair with a blend of pale pink, light purple, sky blue, and mint green hues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a slow zoom in from a medium shot to a close-up of the woman's profile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are warm golden afternoon sunlight beams and floating light particles visible in the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015140033792704344 + ], + "opposite_logprob_tokens": [ + -6.501513957977295 + ], + "logprob_reward": -0.0015140033792704344, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015140033792704344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18811936676502228, + 0.09789160639047623, + 0.11480312794446945, + 0.04296249896287918, + 0.07835498452186584, + 0.10919796675443649, + 0.05869181081652641, + 0.06395269930362701, + 0.07680275291204453, + 0.16922329366207123 + ], + "credit_map": [ + 0.021390900015830994, + 0.019060058519244194, + 0.03316306322813034, + 0.013901209458708763, + 0.01524349208921194, + 0.031738732010126114, + 0.018173879012465477, + 0.035448018461465836, + 0.015005757100880146, + 0.00913966353982687, + 0.024045923724770546, + 0.009268391877412796, + 0.006207246799021959, + 0.017426231876015663, + 0.005868272390216589, + 0.01093011163175106, + 0.018533509224653244, + 0.021180830895900726, + 0.018424050882458687, + 0.009101240895688534, + 0.010292645543813705, + 0.013950196094810963, + 0.011446944437921047, + 0.011873693205416203, + 0.005366029217839241, + 0.004789501428604126, + 0.008090260438621044, + 0.004327711183577776, + 0.005742878187447786, + 0.006331769283860922, + 0.004426405765116215, + 0.0038879383355379105, + 0.01133959461003542, + 0.01701625995337963, + 0.009244202636182308, + 0.01193541381508112, + 0.008422144688665867, + 0.007499493658542633, + 0.00689194118604064, + 0.006005924195051193, + 0.02302536740899086, + 0.014056256972253323, + 0.015139363706111908, + 0.01142064668238163, + 0.01046073343604803, + 0.01239518541842699, + 0.011802096851170063, + 0.010898308828473091, + 0.008912507444620132, + 0.008710773661732674, + 0.007248070556670427, + 0.007179283071309328, + 0.00845989678055048, + 0.005159799009561539, + 0.00751549331471324, + 0.005505980923771858, + 0.010423354804515839, + 0.009905876591801643, + 0.010853207670152187, + 0.00857057049870491, + 0.005330231506377459, + 0.007376497145742178, + 0.005396961700171232, + 0.006095993332564831, + 0.010876805521547794, + 0.011992527171969414, + 0.008753121830523014, + 0.008054547943174839, + 0.009144052863121033, + 0.011861455626785755, + 0.00919885840266943, + 0.0069213807582855225, + 0.02046802081167698, + 0.028543595224618912, + 0.021724039688706398, + 0.019638726487755775, + 0.017954383045434952, + 0.018771009519696236, + 0.020617123693227768, + 0.021506374701857567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.0015140033792704344 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_11/index_tensor([7664, 7664])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_11/index_tensor([7664, 7664])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5aa64cf4f5dd51d88ac4c547896bf0d9522ce91b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_11/index_tensor([7664, 7664])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from a horizontal angle, the still, long shot captures a single, withered, yellow maple leaf, its edges slightly curled and its veins distinct. With a grace almost reminiscent of slow motion, it silently twirls, sways, and slowly descends through the air. Against its backdrop, an old wooden fence stretches endlessly from the foreground into the distance. Its faded white paint has peeled away, revealing the weathered, dark gray wood grain beneath, creating a powerful diagonal line that guides the composition. This line leads the viewer through the falling leaves in the foreground and into the depths of a desolate, leaf-filled field shrouded in a thin layer of mist. The entire scene is bathed in a soft, faded, desaturated light, with few distinct shadows. The ochre-yellow hues of the fallen leaves, the off-white of the fence, and the gray-brown of the background earth blend into a harmonious, melancholic hue, creating an atmosphere of profound tranquility, a timeless autumn scene imbued with a poetic, bleak, and tranquil beauty.", + "seed": 23888, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_11/index_7664_seed_23888_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.027743798960000277, + "vlm_logprob": -0.027743798960000277, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.027743798960000277, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00617998605594039 + ], + "opposite_logprob_tokens": [ + -5.2561798095703125 + ], + "logprob_reward": -0.00617998605594039, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00617998605594039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5555676817893982, + 0.0566461905837059, + 0.0814204141497612, + 0.019684460014104843, + 0.10127957910299301, + 0.021898068487644196, + 0.03653239086270332, + 0.03583100065588951, + 0.047242771834135056, + 0.04389745742082596 + ], + "credit_map": [ + 0.08109182864427567, + 0.0659889504313469, + 0.09464754164218903, + 0.02371412143111229, + 0.16644448041915894, + 0.08163097500801086, + 0.01791170798242092, + 0.02413804829120636, + 0.005720636807382107, + 0.006049932911992073, + 0.007010255474597216, + 0.004382503218948841, + 0.009308816865086555, + 0.014746162109076977, + 0.0056694732047617435, + 0.0037584123201668262, + 0.007645526900887489, + 0.007871712557971478, + 0.010642094537615776, + 0.005566190928220749, + 0.011420227587223053, + 0.013363673351705074, + 0.0183482114225626, + 0.006562777794897556, + 0.002372945426031947, + 0.003064120886847377, + 0.003122961148619652, + 0.001700091757811606, + 0.003306786995381117, + 0.0037687451113015413, + 0.00117260473780334, + 0.001176204765215516, + 0.011920103803277016, + 0.01928621158003807, + 0.009018471464514732, + 0.00650464603677392, + 0.020997177809476852, + 0.021194996312260628, + 0.004465936683118343, + 0.007892037741839886, + 0.0023642070591449738, + 0.004373178817331791, + 0.0025778496637940407, + 0.0014141168212518096, + 0.002943251980468631, + 0.004504395183175802, + 0.0011349301785230637, + 0.0025861398316919804, + 0.004229131620377302, + 0.008970821276307106, + 0.003965173382312059, + 0.002390229841694236, + 0.005205147434026003, + 0.0064881001599133015, + 0.0025189046282321215, + 0.0027648811228573322, + 0.0036509435158222914, + 0.006170228123664856, + 0.004453208297491074, + 0.002429343294352293, + 0.005037134047597647, + 0.008182902820408344, + 0.0018503698520362377, + 0.004056871868669987, + 0.004621278028935194, + 0.008556641638278961, + 0.004856485407799482, + 0.0024293798487633467, + 0.007571991533041, + 0.012296963483095169, + 0.00226778257638216, + 0.004642250016331673, + 0.0051546236500144005, + 0.00823031272739172, + 0.005505269393324852, + 0.0030801845714449883, + 0.006086769048124552, + 0.008434824645519257, + 0.0027174041606485844, + 0.00468807015568018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.00617998605594039 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294607818126678 + ], + "opposite_logprob_tokens": [ + -2.237946033477783 + ], + "logprob_reward": -0.11294607818126678, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294607818126678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7638216018676758, + 0.018757322803139687, + 0.035987306386232376, + 0.014494461007416248, + 0.05359211191534996, + 0.023569921031594276, + 0.0379030741751194, + 0.025324588641524315, + 0.00865868292748928, + 0.01789100281894207 + ], + "credit_map": [ + 0.05894424393773079, + 0.07196379452943802, + 0.19071528315544128, + 0.026819750666618347, + 0.3160806894302368, + 0.061971623450517654, + 0.02038220688700676, + 0.016943903639912605, + 0.0024857029784470797, + 0.0021897100377827883, + 0.0022512823343276978, + 0.0011311516864225268, + 0.00490416307002306, + 0.0031216940842568874, + 0.0015097069554030895, + 0.0011639107251539826, + 0.0031497343443334103, + 0.006074702832847834, + 0.003933365922421217, + 0.0019161050440743566, + 0.004584506154060364, + 0.009010843001306057, + 0.005239906255155802, + 0.0020781385246664286, + 0.0015814072685316205, + 0.0016769254580140114, + 0.0022387912031263113, + 0.0009235594770871103, + 0.0025020225439220667, + 0.0035936920903623104, + 0.0011304653016850352, + 0.0008475954527966678, + 0.006661530584096909, + 0.013527672737836838, + 0.004300353582948446, + 0.0037680475506931543, + 0.009114065207540989, + 0.009444314055144787, + 0.002661059843376279, + 0.0041150618344545364, + 0.0027431107591837645, + 0.00454364251345396, + 0.0022950898855924606, + 0.001726004877127707, + 0.0030153575353324413, + 0.004263727925717831, + 0.0012287534773349762, + 0.0037542309146374464, + 0.004697922617197037, + 0.008987209759652615, + 0.004052252043038607, + 0.001960529712960124, + 0.00513466028496623, + 0.007414404768496752, + 0.0020047544967383146, + 0.0036513374652713537, + 0.002638420555740595, + 0.006517347414046526, + 0.002538577653467655, + 0.0018937052227556705, + 0.003888696199283004, + 0.005120228510349989, + 0.0009982377523556352, + 0.0017293737037107348, + 0.001076343934983015, + 0.00148525252006948, + 0.0011055916547775269, + 0.00048289206461049616, + 0.001274004695005715, + 0.0021601063199341297, + 0.00040786463068798184, + 0.000666626263409853, + 0.0021246043033897877, + 0.0024987878277897835, + 0.0029335003346204758, + 0.001132334815338254, + 0.0027782106772065163, + 0.003684666007757187, + 0.0011026549618691206, + 0.0016362402820959687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.11294607818126678 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624453864991665 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624453864991665, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624453864991665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4026532769203186, + 0.0771559402346611, + 0.05552767589688301, + 0.06353812664747238, + 0.11657250672578812, + 0.07003851234912872, + 0.03682360798120499, + 0.04396255686879158, + 0.08650306612253189, + 0.0472247339785099 + ], + "credit_map": [ + 0.06982029229402542, + 0.0444781519472599, + 0.0462598092854023, + 0.013949621468782425, + 0.16945210099220276, + 0.03412281349301338, + 0.013586040586233139, + 0.010984446853399277, + 0.007685139309614897, + 0.01084815338253975, + 0.011272004805505276, + 0.006201355718076229, + 0.009223274886608124, + 0.018244784325361252, + 0.009337014518678188, + 0.0043442128226161, + 0.0048242174088954926, + 0.005925831384956837, + 0.007522440515458584, + 0.0033846632577478886, + 0.007544241845607758, + 0.011022137477993965, + 0.010905232280492783, + 0.004398909397423267, + 0.0070465682074427605, + 0.0085647227242589, + 0.010308783501386642, + 0.004752067849040031, + 0.01004529558122158, + 0.013023588806390762, + 0.004503235220909119, + 0.005293864756822586, + 0.014590185135602951, + 0.030519651249051094, + 0.009810660034418106, + 0.00788542814552784, + 0.0177563838660717, + 0.020721392706036568, + 0.006223294883966446, + 0.009065508842468262, + 0.00818802509456873, + 0.01843571849167347, + 0.006098917685449123, + 0.005063685122877359, + 0.01180238090455532, + 0.010986837558448315, + 0.002542646136134863, + 0.00692030182108283, + 0.003981101326644421, + 0.005853657145053148, + 0.003917950205504894, + 0.002473862376064062, + 0.00524282967671752, + 0.007603302597999573, + 0.003189362119883299, + 0.004561542067676783, + 0.00483374809846282, + 0.007104168180376291, + 0.005792052950710058, + 0.0021480179857462645, + 0.006942700129002333, + 0.011120278388261795, + 0.0021536650601774454, + 0.0038679272402077913, + 0.009343964979052544, + 0.012655467726290226, + 0.009942101314663887, + 0.003570926608517766, + 0.01522421557456255, + 0.024667127057909966, + 0.004346059635281563, + 0.006753199268132448, + 0.005094502586871386, + 0.00798012688755989, + 0.005594688002020121, + 0.003259408986195922, + 0.006960938218981028, + 0.011040258221328259, + 0.002884412184357643, + 0.004410398658365011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.007624453864991665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009772907942533493 + ], + "opposite_logprob_tokens": [ + -4.634772777557373 + ], + "logprob_reward": -0.009772907942533493, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009772907942533493, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26335856318473816, + 0.07835768908262253, + 0.04287741705775261, + 0.06852630525827408, + 0.21385487914085388, + 0.05294128134846687, + 0.10558333992958069, + 0.10560421645641327, + 0.01578688994050026, + 0.05310942605137825 + ], + "credit_map": [ + 0.025555849075317383, + 0.025196755304932594, + 0.06689700484275818, + 0.010227334685623646, + 0.10151750594377518, + 0.018725646659731865, + 0.007021086290478706, + 0.00821738038212061, + 0.007659066002815962, + 0.009182145819067955, + 0.011570981703698635, + 0.006753345485776663, + 0.009568296372890472, + 0.015784434974193573, + 0.011818368919193745, + 0.006021048408001661, + 0.0037459982559084892, + 0.004653919022530317, + 0.0050547560676932335, + 0.002810005098581314, + 0.005959678906947374, + 0.008006052114069462, + 0.008380433544516563, + 0.004266574513167143, + 0.007421078160405159, + 0.009099375456571579, + 0.010043449699878693, + 0.005170966498553753, + 0.011068734340369701, + 0.014393972232937813, + 0.004803559277206659, + 0.006525167264044285, + 0.026129581034183502, + 0.06088925525546074, + 0.018263952806591988, + 0.015371004119515419, + 0.03465943783521652, + 0.03388416767120361, + 0.008932473137974739, + 0.015725018456578255, + 0.006139391101896763, + 0.012280077673494816, + 0.004347262904047966, + 0.003041555406525731, + 0.008172000758349895, + 0.010614871978759766, + 0.0021162836346775293, + 0.006229836028069258, + 0.009758174419403076, + 0.019912438467144966, + 0.013113330118358135, + 0.006163568701595068, + 0.013767904601991177, + 0.02278512716293335, + 0.009451298043131828, + 0.010631494224071503, + 0.01118293683975935, + 0.016839677467942238, + 0.011114086955785751, + 0.005618810188025236, + 0.018177302554249763, + 0.025166315957903862, + 0.007631965912878513, + 0.009873121976852417, + 0.0018319262890145183, + 0.0025535617023706436, + 0.0020920138340443373, + 0.000956097268499434, + 0.0027206670492887497, + 0.003745855065062642, + 0.000772070896346122, + 0.0011146977776661515, + 0.005555700045078993, + 0.009733617305755615, + 0.006061364430934191, + 0.0037517992313951254, + 0.00888732261955738, + 0.010816353373229504, + 0.0028730251360684633, + 0.005430241581052542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.009772907942533493 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021955687552690506 + ], + "opposite_logprob_tokens": [ + -6.127195358276367 + ], + "logprob_reward": -0.0021955687552690506, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021955687552690506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4515034556388855, + 0.040996354073286057, + 0.03149539604783058, + 0.04088080674409866, + 0.16233834624290466, + 0.07663525640964508, + 0.08137273788452148, + 0.0451127327978611, + 0.04133192077279091, + 0.028333010151982307 + ], + "credit_map": [ + 0.07249417901039124, + 0.030040595680475235, + 0.13662825524806976, + 0.01862739957869053, + 0.123424232006073, + 0.04044477269053459, + 0.011108852922916412, + 0.018735168501734734, + 0.004723523277789354, + 0.006594107486307621, + 0.004435036797076464, + 0.003675316460430622, + 0.005019244272261858, + 0.008739015087485313, + 0.004089247435331345, + 0.003720864187926054, + 0.0033307145349681377, + 0.0036870327312499285, + 0.0038467056583613157, + 0.001693065045401454, + 0.004497147165238857, + 0.005907687824219465, + 0.005736934952437878, + 0.0027961069718003273, + 0.005373439285904169, + 0.005635004490613937, + 0.006262053269892931, + 0.003738363739103079, + 0.006644862238317728, + 0.006594637408852577, + 0.002933954354375601, + 0.0036984914913773537, + 0.01865713857114315, + 0.04785888269543648, + 0.014308865182101727, + 0.011392761021852493, + 0.025247961282730103, + 0.02723209746181965, + 0.007178185041993856, + 0.01046245452016592, + 0.010068271309137344, + 0.017888162285089493, + 0.007046267855912447, + 0.005667082499712706, + 0.011992321349680424, + 0.013330722227692604, + 0.004248208831995726, + 0.006394220981746912, + 0.00858093611896038, + 0.023557042703032494, + 0.007993675768375397, + 0.00542675843462348, + 0.010439366102218628, + 0.0148628456518054, + 0.004939565435051918, + 0.0055725425481796265, + 0.005491313524544239, + 0.007774244528263807, + 0.00567434960976243, + 0.0035769492387771606, + 0.00746354553848505, + 0.00879025086760521, + 0.002899685176089406, + 0.003442394779995084, + 0.0042617591097950935, + 0.006457298994064331, + 0.005320413038134575, + 0.0022888954263180494, + 0.006399523001164198, + 0.008660979568958282, + 0.0028351701330393553, + 0.005107880104333162, + 0.002787098055705428, + 0.0028349682688713074, + 0.0034123791847378016, + 0.002120095072314143, + 0.003818496596068144, + 0.007458142936229706, + 0.0023142960853874683, + 0.0035875316243618727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0021955687552690506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a horizontal angle, the still, long shot captures a single, withered, yellow maple leaf, its edges slightly curled and its veins distinct. With a grace almost reminiscent of slow motion, it silently twirls, sways, and slowly descends through the air. Against its backdrop, an old wooden fence stretches endlessly from the foreground into the distance. Its faded white paint has peeled away, revealing the weathered, dark gray wood grain beneath, creating a powerful diagonal line that guides the composition. This line leads the viewer through the falling leaves in the foreground and into the depths of a desolate, leaf-filled field shrouded in a thin layer of mist. The entire scene is bathed in a soft, faded, desaturated light, with few distinct shadows. The ochre-yellow hues of the fallen leaves, the off-white of the fence, and the gray-brown of the background earth blend into a harmonious, melancholic hue, creating an atmosphere of profound tranquility, a timeless autumn scene imbued with a poetic, bleak, and tranquil beauty.", + "seed": 23889, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_11/index_7664_seed_23889_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02915105186402798, + "vlm_logprob": -0.02915105186402798, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.02915105186402798, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008947635069489479 + ], + "opposite_logprob_tokens": [ + -5.008947849273682 + ], + "logprob_reward": -0.008947635069489479, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008947635069489479, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47046130895614624, + 0.10572977364063263, + 0.07778580486774445, + 0.13039632141590118, + 0.06619599461555481, + 0.03196902945637703, + 0.029994022101163864, + 0.025410210713744164, + 0.03450098633766174, + 0.027556583285331726 + ], + "credit_map": [ + 0.06146469712257385, + 0.05643639713525772, + 0.06206567585468292, + 0.06314272433519363, + 0.10830690711736679, + 0.041715625673532486, + 0.04282085970044136, + 0.03450840711593628, + 0.009274630807340145, + 0.011604566127061844, + 0.01944630593061447, + 0.014059078879654408, + 0.008216465823352337, + 0.009393610060214996, + 0.02560131996870041, + 0.008133786730468273, + 0.009321358986198902, + 0.009303265251219273, + 0.009755067527294159, + 0.013727458193898201, + 0.007392370607703924, + 0.01215705182403326, + 0.009134234860539436, + 0.006994988303631544, + 0.012912139296531677, + 0.016610004007816315, + 0.01574663445353508, + 0.02102571725845337, + 0.012292449362576008, + 0.026896430179476738, + 0.013983692973852158, + 0.010929238051176071, + 0.0063165463507175446, + 0.00689414469525218, + 0.01046153623610735, + 0.010231193155050278, + 0.005316623952239752, + 0.007783631328493357, + 0.008079338818788528, + 0.011112975887954235, + 0.0038390762638300657, + 0.0032752363476902246, + 0.0041549671441316605, + 0.004685358144342899, + 0.002655251882970333, + 0.004857223015278578, + 0.0034837396815419197, + 0.005018175113946199, + 0.005823961924761534, + 0.004450935870409012, + 0.0038451584987342358, + 0.005678583402186632, + 0.0019146332051604986, + 0.0029245459008961916, + 0.0026339420583099127, + 0.002722256351262331, + 0.003647213103249669, + 0.0030596773140132427, + 0.0030651187989860773, + 0.0034148474223911762, + 0.0019178747897967696, + 0.003502595005556941, + 0.0024527576752007008, + 0.004350122530013323, + 0.005364799406379461, + 0.004536283668130636, + 0.0046565476804971695, + 0.006904685404151678, + 0.001965554431080818, + 0.003520364174619317, + 0.0035070746671408415, + 0.004045671317726374, + 0.004361240658909082, + 0.0024765876587480307, + 0.004336066078394651, + 0.005121349357068539, + 0.0017539741238579154, + 0.0023891835007816553, + 0.0034715277142822742, + 0.003646652214229107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.008947635069489479 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295917630195618 + ], + "opposite_logprob_tokens": [ + -2.237959146499634 + ], + "logprob_reward": -0.11295917630195618, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295917630195618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46993744373321533, + 0.07147792726755142, + 0.12898679077625275, + 0.08539356291294098, + 0.0731302797794342, + 0.036946214735507965, + 0.04411974176764488, + 0.0338626392185688, + 0.035379476845264435, + 0.02076590433716774 + ], + "credit_map": [ + 0.058652203530073166, + 0.04801018908619881, + 0.06769295036792755, + 0.08536688983440399, + 0.08040264248847961, + 0.05483397841453552, + 0.0432516410946846, + 0.03172697126865387, + 0.012268689461052418, + 0.012393767014145851, + 0.011217629536986351, + 0.01056597474962473, + 0.004730789922177792, + 0.0068982369266450405, + 0.007548272144049406, + 0.005854577291756868, + 0.016817301511764526, + 0.02063867449760437, + 0.015744071453809738, + 0.024213628843426704, + 0.01018389593809843, + 0.013979515992105007, + 0.014987559989094734, + 0.012422154657542706, + 0.010824977420270443, + 0.011702217161655426, + 0.011899679899215698, + 0.015011662617325783, + 0.006946099456399679, + 0.010697579942643642, + 0.012161062099039555, + 0.006150293629616499, + 0.013925161212682724, + 0.011809999123215675, + 0.010306555777788162, + 0.014202471822500229, + 0.0043923575431108475, + 0.00713937496766448, + 0.004974997602403164, + 0.006379359867423773, + 0.0052522094920277596, + 0.0056878249160945415, + 0.0045533799566328526, + 0.005136927589774132, + 0.0029303652700036764, + 0.0063559431582689285, + 0.002828035969287157, + 0.004201531410217285, + 0.004938164260238409, + 0.005475468002259731, + 0.006169492844492197, + 0.007501113694161177, + 0.003044264158234, + 0.005778572056442499, + 0.005661166273057461, + 0.005551507230848074, + 0.004526039119809866, + 0.003902736585587263, + 0.005407352466136217, + 0.004351750947535038, + 0.0026630680076777935, + 0.0036009482573717833, + 0.0034369241911917925, + 0.005973820574581623, + 0.005743111949414015, + 0.004739810712635517, + 0.005313968751579523, + 0.004931627307087183, + 0.0020089142490178347, + 0.0030466183088719845, + 0.0043630884028971195, + 0.0052323411218822, + 0.003394166938960552, + 0.003051884239539504, + 0.0027431792113929987, + 0.0031331570353358984, + 0.0017529497854411602, + 0.002527246018871665, + 0.002495653461664915, + 0.00166766915936023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.11295917630195618 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817641779780388 + ], + "opposite_logprob_tokens": [ + -4.018176555633545 + ], + "logprob_reward": -0.01817641779780388, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817641779780388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09333129227161407, + 0.10376580059528351, + 0.11524421721696854, + 0.1565868854522705, + 0.16472876071929932, + 0.07107827812433243, + 0.07189556956291199, + 0.052927643060684204, + 0.04077223688364029, + 0.12966935336589813 + ], + "credit_map": [ + 0.011738310568034649, + 0.00999496877193451, + 0.012544585391879082, + 0.014014983549714088, + 0.015927940607070923, + 0.012151078321039677, + 0.009576508775353432, + 0.0073829153552651405, + 0.01841147430241108, + 0.02067660726606846, + 0.01564284972846508, + 0.01748676784336567, + 0.005457381717860699, + 0.005656038876622915, + 0.013545582070946693, + 0.006889091804623604, + 0.016948115080595016, + 0.018399864435195923, + 0.0175747312605381, + 0.020502861589193344, + 0.006665480323135853, + 0.008616052567958832, + 0.012859801761806011, + 0.013677305541932583, + 0.011815464124083519, + 0.01571577973663807, + 0.02203122153878212, + 0.021411024034023285, + 0.015041010454297066, + 0.03555496782064438, + 0.018600570037961006, + 0.016416851431131363, + 0.015840934589505196, + 0.013219616375863552, + 0.02647642232477665, + 0.02239273115992546, + 0.012766087427735329, + 0.03881995379924774, + 0.013374019414186478, + 0.02183898724615574, + 0.012772981077432632, + 0.009566283784806728, + 0.00971583928912878, + 0.011818557977676392, + 0.005253629293292761, + 0.010037831962108612, + 0.006330124102532864, + 0.005583026446402073, + 0.011549388989806175, + 0.00907494779676199, + 0.01103956252336502, + 0.011810806579887867, + 0.004902190528810024, + 0.009123949334025383, + 0.007758208084851503, + 0.0066365129314363, + 0.010517989285290241, + 0.006991235539317131, + 0.008777836337685585, + 0.009422745555639267, + 0.0026750953402370214, + 0.006187258288264275, + 0.003665479365736246, + 0.004690004047006369, + 0.007096393033862114, + 0.005383443087339401, + 0.006373973097652197, + 0.006254273001104593, + 0.0025904495269060135, + 0.003294625086709857, + 0.004173068795353174, + 0.00560601195320487, + 0.017185233533382416, + 0.013472400605678558, + 0.01936383917927742, + 0.022225504741072655, + 0.007578628603368998, + 0.015391169115900993, + 0.016590462997555733, + 0.017862122505903244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.01817641779780388 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003185794921591878 + ], + "opposite_logprob_tokens": [ + -5.753185749053955 + ], + "logprob_reward": -0.003185794921591878, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003185794921591878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23672454059123993, + 0.15024146437644958, + 0.07007037848234177, + 0.1574578732252121, + 0.12352451682090759, + 0.09215542674064636, + 0.04380428045988083, + 0.033077094703912735, + 0.03307613357901573, + 0.05986830219626427 + ], + "credit_map": [ + 0.050594646483659744, + 0.027722110971808434, + 0.03194933384656906, + 0.043117817491292953, + 0.021670764312148094, + 0.022577300667762756, + 0.020858822390437126, + 0.018233750015497208, + 0.02390228398144245, + 0.02830355055630207, + 0.022260993719100952, + 0.027915433049201965, + 0.007747963536530733, + 0.009467276744544506, + 0.020528381690382957, + 0.01011557038873434, + 0.009378273040056229, + 0.007935568690299988, + 0.00885770469903946, + 0.009180992841720581, + 0.006493753753602505, + 0.008748133666813374, + 0.009032266214489937, + 0.01044368464499712, + 0.014552446082234383, + 0.01826993003487587, + 0.020734544843435287, + 0.02402258850634098, + 0.013016223907470703, + 0.03153155744075775, + 0.020079944282770157, + 0.015250620432198048, + 0.012520702555775642, + 0.01148794312030077, + 0.01867809146642685, + 0.017178067937493324, + 0.010004210285842419, + 0.024446679279208183, + 0.011552253738045692, + 0.017656570300459862, + 0.015645796433091164, + 0.011656117625534534, + 0.013309343717992306, + 0.01621793769299984, + 0.006795842200517654, + 0.012546968646347523, + 0.009411030448973179, + 0.006572392303496599, + 0.005874872673302889, + 0.0047856466844677925, + 0.006663918495178223, + 0.007788449991494417, + 0.0034294240176677704, + 0.0065041352063417435, + 0.005045901983976364, + 0.003711929079145193, + 0.0036877202801406384, + 0.0031853890977799892, + 0.004794682841747999, + 0.006371072493493557, + 0.0028492070268839598, + 0.005341574549674988, + 0.0034466604702174664, + 0.003400790737941861, + 0.004630561452358961, + 0.00397971598431468, + 0.005608269479125738, + 0.005250716581940651, + 0.0020198212005198, + 0.003239547600969672, + 0.0033250043634325266, + 0.005022495985031128, + 0.006374734919518232, + 0.006323887966573238, + 0.009560356847941875, + 0.010847139172255993, + 0.003856442403048277, + 0.0076063573360443115, + 0.00803543534129858, + 0.007263949140906334 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.003185794921591878 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024862352292984724 + ], + "opposite_logprob_tokens": [ + -6.002486228942871 + ], + "logprob_reward": -0.0024862352292984724, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024862352292984724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3205423057079315, + 0.10667258501052856, + 0.0498419925570488, + 0.14523042738437653, + 0.11015841364860535, + 0.07498931139707565, + 0.07681697607040405, + 0.044319164007902145, + 0.04337306693196297, + 0.028055783361196518 + ], + "credit_map": [ + 0.04198765754699707, + 0.0369781032204628, + 0.04932793974876404, + 0.06802934408187866, + 0.031384676694869995, + 0.03599421679973602, + 0.03185669332742691, + 0.024983687326312065, + 0.02399948239326477, + 0.02038867585361004, + 0.013714218512177467, + 0.017630893737077713, + 0.006021017674356699, + 0.007133989594876766, + 0.009203419089317322, + 0.008580886758863926, + 0.005692771635949612, + 0.006409183144569397, + 0.007166403345763683, + 0.007961386814713478, + 0.004597789607942104, + 0.004850743804126978, + 0.00809419620782137, + 0.005069518927484751, + 0.015637870877981186, + 0.02094443514943123, + 0.01856778748333454, + 0.021551618352532387, + 0.012041481211781502, + 0.027025651186704636, + 0.01643911749124527, + 0.013022464700043201, + 0.011042473837733269, + 0.010927987284958363, + 0.015060908161103725, + 0.018949588760733604, + 0.009996889159083366, + 0.017142709344625473, + 0.011672969907522202, + 0.015364882536232471, + 0.009247968904674053, + 0.008227651007473469, + 0.009342649020254612, + 0.010628036223351955, + 0.005569240543991327, + 0.013497261330485344, + 0.009128421545028687, + 0.00934807863086462, + 0.015472625382244587, + 0.010576066561043262, + 0.010194972157478333, + 0.013523813337087631, + 0.003937366418540478, + 0.009805516339838505, + 0.006128096021711826, + 0.0071785226464271545, + 0.007641782518476248, + 0.0056727007031440735, + 0.006321259308606386, + 0.008524195291101933, + 0.0029112393967807293, + 0.005442503374069929, + 0.004227073397487402, + 0.0035784104838967323, + 0.009266102686524391, + 0.006455335300415754, + 0.00661199726164341, + 0.008097346872091293, + 0.0022261233534663916, + 0.003945353906601667, + 0.0031170581933110952, + 0.003653747495263815, + 0.0048731788992881775, + 0.004190988373011351, + 0.0032589687034487724, + 0.004117657896131277, + 0.0015628550900146365, + 0.003738538594916463, + 0.004015261307358742, + 0.002298334613442421 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.0024862352292984724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a horizontal angle, the still, long shot captures a single, withered, yellow maple leaf, its edges slightly curled and its veins distinct. With a grace almost reminiscent of slow motion, it silently twirls, sways, and slowly descends through the air. Against its backdrop, an old wooden fence stretches endlessly from the foreground into the distance. Its faded white paint has peeled away, revealing the weathered, dark gray wood grain beneath, creating a powerful diagonal line that guides the composition. This line leads the viewer through the falling leaves in the foreground and into the depths of a desolate, leaf-filled field shrouded in a thin layer of mist. The entire scene is bathed in a soft, faded, desaturated light, with few distinct shadows. The ochre-yellow hues of the fallen leaves, the off-white of the fence, and the gray-brown of the background earth blend into a harmonious, melancholic hue, creating an atmosphere of profound tranquility, a timeless autumn scene imbued with a poetic, bleak, and tranquil beauty.", + "seed": 23890, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_11/index_7664_seed_23890_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.006835101265460252, + "vlm_logprob": -0.006835101265460252, + "vlm_yes_no_margin": 5.375, + "vlm_token_credit": -0.006835101265460252, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0034585916437208652 + ], + "opposite_logprob_tokens": [ + -6.253458499908447 + ], + "logprob_reward": -0.0034585916437208652, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0034585916437208652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15984191000461578, + 0.2328735589981079, + 0.220977783203125, + 0.1464267373085022, + 0.06018146872520447, + 0.04606396332383156, + 0.059400033205747604, + 0.027157004922628403, + 0.02251093089580536, + 0.024566665291786194 + ], + "credit_map": [ + 0.020716678351163864, + 0.023902151733636856, + 0.019051311537623405, + 0.020891685038805008, + 0.02652396261692047, + 0.019659390673041344, + 0.015515536069869995, + 0.013581196777522564, + 0.034041572362184525, + 0.07835423201322556, + 0.014951803721487522, + 0.011965055949985981, + 0.0324271023273468, + 0.03174470737576485, + 0.01258903183043003, + 0.016800062730908394, + 0.025840286165475845, + 0.030299250036478043, + 0.029131026938557625, + 0.012699677608907223, + 0.053649235516786575, + 0.040867432951927185, + 0.01097804494202137, + 0.017512818798422813, + 0.015312351286411285, + 0.01840529963374138, + 0.012481686659157276, + 0.00675560999661684, + 0.025656666606664658, + 0.05146543309092522, + 0.004719773773103952, + 0.011629918590188026, + 0.008259306661784649, + 0.0056159463711082935, + 0.006595482584089041, + 0.005902569741010666, + 0.009754478931427002, + 0.01506786234676838, + 0.004687737673521042, + 0.004298075567930937, + 0.0049784365110099316, + 0.004508045967668295, + 0.0036662097554653883, + 0.002987784333527088, + 0.010842218063771725, + 0.01368098147213459, + 0.0026030398439615965, + 0.0027972443494945765, + 0.006262895185500383, + 0.006346444599330425, + 0.005187609698623419, + 0.0035695554688572884, + 0.013705630786716938, + 0.018627583980560303, + 0.002607843605801463, + 0.003092467784881592, + 0.0026198129635304213, + 0.002456480171531439, + 0.002631580922752619, + 0.0019229367608204484, + 0.006413556635379791, + 0.007885216735303402, + 0.0017907668370753527, + 0.0014366540126502514, + 0.001848262851126492, + 0.0023503319825977087, + 0.0017044725827872753, + 0.0015409726183861494, + 0.006943410728126764, + 0.005790226627141237, + 0.001273705274797976, + 0.0010595486965030432, + 0.0021794866770505905, + 0.002579383086413145, + 0.002426787745207548, + 0.002350413938984275, + 0.005417539272457361, + 0.00580710219219327, + 0.002004913752898574, + 0.0018010367639362812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0034585916437208652 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018195616081357002 + ], + "opposite_logprob_tokens": [ + -4.018195629119873 + ], + "logprob_reward": -0.018195616081357002, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018195616081357002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09519396722316742, + 0.6074960827827454, + 0.10426319390535355, + 0.07070769369602203, + 0.03183790668845177, + 0.030618013814091682, + 0.035891588777303696, + 0.007020655553787947, + 0.011008436791598797, + 0.005962455179542303 + ], + "credit_map": [ + 0.009350528009235859, + 0.012250694446265697, + 0.012779584154486656, + 0.009225928224623203, + 0.014330635778605938, + 0.010838743299245834, + 0.014310638420283794, + 0.012107208371162415, + 0.027213050052523613, + 0.18034209311008453, + 0.017152905464172363, + 0.024852313101291656, + 0.046743445098400116, + 0.2635270953178406, + 0.011737723834812641, + 0.03592749312520027, + 0.012207676656544209, + 0.01198405958712101, + 0.01764446310698986, + 0.006726242136210203, + 0.01862286403775215, + 0.012283174321055412, + 0.007320955395698547, + 0.01747375912964344, + 0.007569565903395414, + 0.009529643692076206, + 0.005870406050235033, + 0.004869429860264063, + 0.011016651056706905, + 0.022980663925409317, + 0.004046090878546238, + 0.004825239535421133, + 0.0032810098491609097, + 0.00310612958855927, + 0.002970625413581729, + 0.002295839600265026, + 0.006244799587875605, + 0.010257244110107422, + 0.0017998639959841967, + 0.0018823962891474366, + 0.0025977164041250944, + 0.0028973675798624754, + 0.0024097494315356016, + 0.001875498448498547, + 0.007450415752828121, + 0.00969697255641222, + 0.0018400776898488402, + 0.001850214204750955, + 0.003509875852614641, + 0.003746706759557128, + 0.0025574450846761465, + 0.0021832555066794157, + 0.008117454126477242, + 0.012204769998788834, + 0.001523647690191865, + 0.0020484330598264933, + 0.0007036211318336427, + 0.0007211361080408096, + 0.0007754754624329507, + 0.0008339135674759746, + 0.0011686530197039247, + 0.0014841087395325303, + 0.0006784171564504504, + 0.0006553306011483073, + 0.0014114495133981109, + 0.0014082696288824081, + 0.000872909789904952, + 0.000620190636254847, + 0.0019184150733053684, + 0.0029777411837130785, + 0.0008969675400294363, + 0.000902492378372699, + 0.00045580562436953187, + 0.0005583983147516847, + 0.0003433192614465952, + 0.0003212267765775323, + 0.0016789082437753677, + 0.0019151601009070873, + 0.00029724929481744766, + 0.0003923875920008868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.018195616081357002 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00594525970518589 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.00594525970518589, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00594525970518589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05316862091422081, + 0.6731503009796143, + 0.08624407649040222, + 0.034722041338682175, + 0.05491555854678154, + 0.025058411061763763, + 0.03360318765044212, + 0.011907916516065598, + 0.015985578298568726, + 0.0112442122772336 + ], + "credit_map": [ + 0.006379741244018078, + 0.006547999568283558, + 0.00871224794536829, + 0.005576177034527063, + 0.008327114395797253, + 0.006662291940301657, + 0.005958439316600561, + 0.005004617851227522, + 0.03127502277493477, + 0.19144849479198456, + 0.020006375387310982, + 0.02700222097337246, + 0.05458574369549751, + 0.2955409586429596, + 0.01159722451120615, + 0.041694335639476776, + 0.009422251023352146, + 0.011839533224701881, + 0.009552069008350372, + 0.006447675172239542, + 0.02134402096271515, + 0.01598113402724266, + 0.005851663649082184, + 0.005805738270282745, + 0.003963496536016464, + 0.0049417694099247456, + 0.0032804624643176794, + 0.0025397264398634434, + 0.005631147418171167, + 0.00946251954883337, + 0.001949536963365972, + 0.0029533880297094584, + 0.005420174449682236, + 0.0051288544200360775, + 0.0040231747552752495, + 0.0043623498640954494, + 0.013918312266469002, + 0.013722732663154602, + 0.004552213940769434, + 0.003787753637880087, + 0.0027890203054994345, + 0.0023768828250467777, + 0.0027622259221971035, + 0.0024270822759717703, + 0.0056956978514790535, + 0.0055937981233000755, + 0.0016091351862996817, + 0.0018045706674456596, + 0.002525239484384656, + 0.0029065171256661415, + 0.0021242713555693626, + 0.0012347475858405232, + 0.009692467749118805, + 0.012367397546768188, + 0.001202425337396562, + 0.0015501310117542744, + 0.0016223359853029251, + 0.001760271959938109, + 0.0013384080957621336, + 0.0009300176752731204, + 0.002139578340575099, + 0.002437079092487693, + 0.0008036743965931237, + 0.000876552308909595, + 0.0019773526582866907, + 0.0015244928654283285, + 0.002173690591007471, + 0.0010501567739993334, + 0.003952247090637684, + 0.003307353937998414, + 0.0009324470884166658, + 0.0010678402613848448, + 0.0015581682091578841, + 0.0014955097576603293, + 0.0013483522925525904, + 0.0008930160547606647, + 0.0019628871232271194, + 0.0021051226649433374, + 0.000970384746324271, + 0.0009107724763453007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.00594525970518589 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004088852554559708 + ], + "opposite_logprob_tokens": [ + -5.504088878631592 + ], + "logprob_reward": -0.004088852554559708, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004088852554559708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049146246165037155, + 0.5687204003334045, + 0.028211796656250954, + 0.10487108677625656, + 0.05939629301428795, + 0.037303898483514786, + 0.08796396106481552, + 0.016025882214307785, + 0.024148231372237206, + 0.024212250486016273 + ], + "credit_map": [ + 0.005269182845950127, + 0.0054966541938483715, + 0.006201964803040028, + 0.0063612195663154125, + 0.005719376727938652, + 0.005793754477053881, + 0.0061828745529055595, + 0.008121214807033539, + 0.02504902146756649, + 0.1729917824268341, + 0.01702963560819626, + 0.023433169350028038, + 0.04905525967478752, + 0.240226149559021, + 0.010728747583925724, + 0.030206644907593727, + 0.0034182616509497166, + 0.0035580501426011324, + 0.003913232125341892, + 0.0027599676977843046, + 0.005071642808616161, + 0.004790957551449537, + 0.002612261101603508, + 0.0020874235779047012, + 0.01174330897629261, + 0.01368503738194704, + 0.008648076094686985, + 0.00480841426178813, + 0.016578366979956627, + 0.03801275044679642, + 0.0037100245244801044, + 0.007685106713324785, + 0.004872064106166363, + 0.005177662242203951, + 0.00472813518717885, + 0.0044096107594668865, + 0.01709255576133728, + 0.015944598242640495, + 0.0036694647278636694, + 0.0035022005904465914, + 0.004041943233460188, + 0.004881618078798056, + 0.0033655629958957434, + 0.0025031210388988256, + 0.007327222730964422, + 0.010358079336583614, + 0.0021564774215221405, + 0.0026698738802224398, + 0.006314438302069902, + 0.007990222424268723, + 0.004328699316829443, + 0.0028084786608815193, + 0.025310233235359192, + 0.03324713930487633, + 0.004010586533695459, + 0.003954166080802679, + 0.0014021378010511398, + 0.0015796723309904337, + 0.0012760034296661615, + 0.0009367833845317364, + 0.00435060728341341, + 0.004907601047307253, + 0.0008133447845466435, + 0.0007597315707243979, + 0.0020207357592880726, + 0.0022703830618411303, + 0.0020666392520070076, + 0.0013793498510494828, + 0.007457703351974487, + 0.0064179482869803905, + 0.0011627764906734228, + 0.0013726965989917517, + 0.0016648436430841684, + 0.001856227987445891, + 0.0015476028202101588, + 0.0010593114420771599, + 0.007313069887459278, + 0.008703512139618397, + 0.0010058789048343897, + 0.0010618057567626238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 6 + ], + "selected_reward": -0.004088852554559708 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024871863424777985 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.0024871863424777985, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024871863424777985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053248483687639236, + 0.7071281671524048, + 0.08429200202226639, + 0.023225050419569016, + 0.03604169934988022, + 0.02521202340722084, + 0.03068423829972744, + 0.017751552164554596, + 0.011917836032807827, + 0.010498987510800362 + ], + "credit_map": [ + 0.005284278653562069, + 0.00785622838884592, + 0.00677851215004921, + 0.006194890011101961, + 0.011169333010911942, + 0.0066690267994999886, + 0.005156130064278841, + 0.004140084609389305, + 0.031298086047172546, + 0.22834613919258118, + 0.020532896742224693, + 0.029192926362156868, + 0.048894498497247696, + 0.30232328176498413, + 0.010708827525377274, + 0.03583147004246712, + 0.011229393072426319, + 0.010104711167514324, + 0.013135368004441261, + 0.005324849393218756, + 0.01806025765836239, + 0.01091080717742443, + 0.005371356848627329, + 0.010155262425541878, + 0.0021616716403514147, + 0.0025171409361064434, + 0.0030938407871872187, + 0.0015645303064957261, + 0.004441580735146999, + 0.004287183750420809, + 0.0014316217275336385, + 0.003727479837834835, + 0.004973449278622866, + 0.00391926197335124, + 0.00328724249266088, + 0.002610736759379506, + 0.00648588128387928, + 0.010605134069919586, + 0.002291486132889986, + 0.0018685022369027138, + 0.0027297071646898985, + 0.0022181563545018435, + 0.002382022561505437, + 0.0013022548519074917, + 0.006296283565461636, + 0.007707026321440935, + 0.0013256402453407645, + 0.001250932808034122, + 0.0028846149798482656, + 0.003107459284365177, + 0.002224454190582037, + 0.001480820239521563, + 0.0075587378814816475, + 0.010661913082003593, + 0.0013435797300189734, + 0.0014226592611521482, + 0.0018532241228967905, + 0.00193368224427104, + 0.0015988699160516262, + 0.0009561568149365485, + 0.003818241413682699, + 0.0055340733379125595, + 0.0011575412936508656, + 0.0008997631375677884, + 0.0008870100136846304, + 0.001183668035082519, + 0.0006100568571127951, + 0.0005808115238323808, + 0.004215240944176912, + 0.0034939805045723915, + 0.0004882521170657128, + 0.00045881635742262006, + 0.0015321274986490607, + 0.0013269252376630902, + 0.00091110187349841, + 0.0007713803788647056, + 0.0021943559404462576, + 0.002237569773569703, + 0.0007144645205698907, + 0.0008110625203698874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0024871863424777985 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a horizontal angle, the still, long shot captures a single, withered, yellow maple leaf, its edges slightly curled and its veins distinct. With a grace almost reminiscent of slow motion, it silently twirls, sways, and slowly descends through the air. Against its backdrop, an old wooden fence stretches endlessly from the foreground into the distance. Its faded white paint has peeled away, revealing the weathered, dark gray wood grain beneath, creating a powerful diagonal line that guides the composition. This line leads the viewer through the falling leaves in the foreground and into the depths of a desolate, leaf-filled field shrouded in a thin layer of mist. The entire scene is bathed in a soft, faded, desaturated light, with few distinct shadows. The ochre-yellow hues of the fallen leaves, the off-white of the fence, and the gray-brown of the background earth blend into a harmonious, melancholic hue, creating an atmosphere of profound tranquility, a timeless autumn scene imbued with a poetic, bleak, and tranquil beauty.", + "seed": 23891, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_11/index_7664_seed_23891_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0059135907096788285, + "vlm_logprob": -0.0059135907096788285, + "vlm_yes_no_margin": 5.625, + "vlm_token_credit": -0.0059135907096788285, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005974766332656145 + ], + "opposite_logprob_tokens": [ + -6.630974769592285 + ], + "logprob_reward": -0.005974766332656145, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.005974766332656145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01943495310842991, + 0.009973514825105667, + 0.19629672169685364, + 0.16736535727977753, + 0.05220205336809158, + 0.1645427644252777, + 0.006536734290421009, + 0.033541202545166016, + 0.19104912877082825, + 0.15905757248401642 + ], + "credit_map": [ + 0.002399720484390855, + 0.00463875662535429, + 0.0018051705555990338, + 0.0038718106225132942, + 0.0015804758295416832, + 0.002287098206579685, + 0.001646209741011262, + 0.0012057116255164146, + 0.0008499110117554665, + 0.0009353365167044103, + 0.0021337915677577257, + 0.0012259616050869226, + 0.0008281352929770947, + 0.00318030989728868, + 0.00045563653111457825, + 0.00036443237331695855, + 0.012989864684641361, + 0.012938641011714935, + 0.04841294512152672, + 0.018365928903222084, + 0.0173446387052536, + 0.07850135862827301, + 0.0038544749841094017, + 0.00388887501321733, + 0.010587782599031925, + 0.010491659864783287, + 0.0392831414937973, + 0.01566271111369133, + 0.017905468121170998, + 0.06734621524810791, + 0.002940473146736622, + 0.003147898940369487, + 0.003148350864648819, + 0.003460431005805731, + 0.012268655933439732, + 0.0044805919751524925, + 0.005528346169739962, + 0.021333161741495132, + 0.0009560050093568861, + 0.0010265138698741794, + 0.009885838255286217, + 0.011183859780430794, + 0.039704177528619766, + 0.013269676826894283, + 0.017400799319148064, + 0.06816795468330383, + 0.002360836137086153, + 0.002569611184298992, + 0.000651627779006958, + 0.0005188488285057247, + 0.0009977984009310603, + 0.0008279464673250914, + 0.0008196004200726748, + 0.0018396945670247078, + 0.0004843488859478384, + 0.0003968694363720715, + 0.0040543824434280396, + 0.003131514647975564, + 0.005791558884084225, + 0.0034082261845469475, + 0.004359177779406309, + 0.009239794686436653, + 0.0021473069209605455, + 0.001409241696819663, + 0.011591311544179916, + 0.013450817205011845, + 0.04409031569957733, + 0.016653746366500854, + 0.019834794104099274, + 0.07865344732999802, + 0.003014987101778388, + 0.003759710118174553, + 0.009775296784937382, + 0.011128696613013744, + 0.03672356158494949, + 0.01383421290665865, + 0.016291428357362747, + 0.06619229167699814, + 0.0023789447732269764, + 0.002733140252530575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 3 + ], + "selected_reward": -0.005974766332656145 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543829157948494 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543829157948494, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543829157948494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07416222989559174, + 0.010896620340645313, + 0.23701119422912598, + 0.41571155190467834, + 0.042527806013822556, + 0.07433707267045975, + 0.03263822942972183, + 0.026875995099544525, + 0.03665796294808388, + 0.04918130859732628 + ], + "credit_map": [ + 0.007550857495516539, + 0.021343562752008438, + 0.00682206591591239, + 0.012072568759322166, + 0.00660862447693944, + 0.012361184693872929, + 0.0044196294620633125, + 0.0029837419278919697, + 0.001016832422465086, + 0.0018046684563159943, + 0.0019248741446062922, + 0.0015153440181165934, + 0.000991127104498446, + 0.002540493616834283, + 0.0005740183987654746, + 0.0005292631103657186, + 0.01564066670835018, + 0.01590036414563656, + 0.05839671567082405, + 0.021773360669612885, + 0.021102391183376312, + 0.09490089118480682, + 0.004681074060499668, + 0.004615751560777426, + 0.02577931247651577, + 0.02622287906706333, + 0.09884685277938843, + 0.03835390880703926, + 0.04482205957174301, + 0.1687925159931183, + 0.005403649993240833, + 0.007490408141165972, + 0.002611310686916113, + 0.0028689373284578323, + 0.009762412868440151, + 0.0035948161967098713, + 0.004569276701658964, + 0.017054811120033264, + 0.0012870344799011946, + 0.0007792069809511304, + 0.004455673974007368, + 0.005036022514104843, + 0.017844369634985924, + 0.006044730544090271, + 0.007887840270996094, + 0.03076842799782753, + 0.0010407720692455769, + 0.0012592388084158301, + 0.0022624838165938854, + 0.002139362972229719, + 0.007773871999233961, + 0.003358264919370413, + 0.003238366451114416, + 0.0122070899233222, + 0.0008850830490700901, + 0.0007737062987871468, + 0.0028899069875478745, + 0.0021301573142409325, + 0.004749416373670101, + 0.0026079644449055195, + 0.0034096422605216503, + 0.00879478920251131, + 0.0013806544011458755, + 0.000913468946237117, + 0.0022457996383309364, + 0.0025251428596675396, + 0.008627239614725113, + 0.0031988518312573433, + 0.0037421591114252806, + 0.014750301837921143, + 0.0007268856279551983, + 0.0008415867341682315, + 0.003088564844802022, + 0.003530042478814721, + 0.011207463219761848, + 0.004130084533244371, + 0.005016221199184656, + 0.0205079298466444, + 0.0008032260229811072, + 0.0008977773250080645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.012543829157948494 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005248814355581999 + ], + "opposite_logprob_tokens": [ + -5.2552490234375 + ], + "logprob_reward": -0.005248814355581999, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005248814355581999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03059687837958336, + 0.01760796271264553, + 0.18925686180591583, + 0.0870872214436531, + 0.162734717130661, + 0.24309776723384857, + 0.06393706798553467, + 0.05274584889411926, + 0.08585873246192932, + 0.06707692891359329 + ], + "credit_map": [ + 0.003398504573851824, + 0.005977222695946693, + 0.0033593226689845324, + 0.005629919469356537, + 0.0028652192559093237, + 0.0035948927979916334, + 0.0031471310649067163, + 0.0026246646884828806, + 0.0014373272424563766, + 0.0023493021726608276, + 0.003494451753795147, + 0.0020092870108783245, + 0.0014752383576706052, + 0.005569756496697664, + 0.0007227188325487077, + 0.000549880787730217, + 0.012371919117867947, + 0.012234266847372055, + 0.04719829559326172, + 0.01775922067463398, + 0.016666032373905182, + 0.07552117109298706, + 0.0037674682680517435, + 0.0037384990137070417, + 0.005472755059599876, + 0.005610777996480465, + 0.020586909726262093, + 0.007878920994699001, + 0.009408239275217056, + 0.03529120236635208, + 0.0012497175484895706, + 0.0015886897454038262, + 0.010087912902235985, + 0.010918546468019485, + 0.03851500526070595, + 0.01406276784837246, + 0.017394091933965683, + 0.06669491529464722, + 0.0020080232061445713, + 0.0030534560792148113, + 0.014557013288140297, + 0.016376465559005737, + 0.05875582993030548, + 0.019594628363847733, + 0.02578173205256462, + 0.10070247948169708, + 0.0034590328577905893, + 0.0038705971091985703, + 0.003943542018532753, + 0.0041344305500388145, + 0.014846013858914375, + 0.007468382827937603, + 0.0063768713735044, + 0.02360011823475361, + 0.0016606246354058385, + 0.0019070813432335854, + 0.003625201992690563, + 0.004029362928122282, + 0.01143964845687151, + 0.004595059901475906, + 0.005769261159002781, + 0.02084789052605629, + 0.0014460515230894089, + 0.000993370427750051, + 0.006566339172422886, + 0.0060266051441431046, + 0.018960407003760338, + 0.007324173115193844, + 0.008595186285674572, + 0.034193530678749084, + 0.002180500188842416, + 0.002011998789384961, + 0.004403168335556984, + 0.004625255241990089, + 0.014658180065453053, + 0.006958510726690292, + 0.006678970530629158, + 0.02614256553351879, + 0.0017818673513829708, + 0.0018284044926986098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.005248814355581999 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004088733810931444 + ], + "opposite_logprob_tokens": [ + -5.504088878631592 + ], + "logprob_reward": -0.004088733810931444, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004088733810931444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03693627566099167, + 0.012306934222579002, + 0.10521478950977325, + 0.09609018266201019, + 0.1512424647808075, + 0.2954997718334198, + 0.05377049371600151, + 0.05514863133430481, + 0.1766793131828308, + 0.01711120270192623 + ], + "credit_map": [ + 0.0037832162342965603, + 0.008770245127379894, + 0.004158589988946915, + 0.007385252043604851, + 0.0027387484442442656, + 0.004912287928164005, + 0.0031179378274828196, + 0.002069994108751416, + 0.0010840281611308455, + 0.0008024899871088564, + 0.0029007052071392536, + 0.001500168815255165, + 0.0009133406565524638, + 0.004377203993499279, + 0.00036154917324893177, + 0.00036744671524502337, + 0.0069624027237296104, + 0.007001282647252083, + 0.026040205731987953, + 0.009757849387824535, + 0.009370717220008373, + 0.042002275586128235, + 0.0020733654964715242, + 0.0020066776778548956, + 0.006068211980164051, + 0.006154261063784361, + 0.022430410608649254, + 0.008971151895821095, + 0.010342681780457497, + 0.03878059238195419, + 0.0014889073790982366, + 0.001853956375271082, + 0.0096055306494236, + 0.010079636238515377, + 0.035571493208408356, + 0.013088214211165905, + 0.016169551759958267, + 0.061503030359745026, + 0.0025780233554542065, + 0.002646971959620714, + 0.017751522362232208, + 0.01997353509068489, + 0.0709642618894577, + 0.024081209674477577, + 0.031152846291661263, + 0.12179573625326157, + 0.004900840111076832, + 0.004879815503954887, + 0.007125822361558676, + 0.003581895725801587, + 0.009934942238032818, + 0.009805482812225819, + 0.004981669597327709, + 0.01242110040038824, + 0.003802952589467168, + 0.002116626361384988, + 0.009602714329957962, + 0.0062492177821695805, + 0.004852704703807831, + 0.007375102024525404, + 0.008569005876779556, + 0.008530648425221443, + 0.00505093764513731, + 0.004918296355754137, + 0.0114336758852005, + 0.012327006086707115, + 0.04006652161478996, + 0.01602066121995449, + 0.018350934609770775, + 0.07199470698833466, + 0.0030623760540038347, + 0.00342340930365026, + 0.0012738413643091917, + 0.0011591138318181038, + 0.003163049928843975, + 0.002503681695088744, + 0.001527242362499237, + 0.006020316854119301, + 0.0008038115338422358, + 0.0006601429777219892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.004088733810931444 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single, withered, yellow maple leaf with distinct veins and slightly curled edges?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown twirling, swaying, and slowly descending through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an old wooden fence with faded, peeling white paint revealing dark gray wood grain visible in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in a desolate, leaf-filled field shrouded in a thin layer of mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting soft, faded, and desaturated with few distinct shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017118098912760615 + ], + "opposite_logprob_tokens": [ + -6.376711845397949 + ], + "logprob_reward": -0.0017118098912760615, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017118098912760615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08166243880987167, + 0.042880307883024216, + 0.035021260380744934, + 0.21171514689922333, + 0.30682387948036194, + 0.027407100424170494, + 0.07159734517335892, + 0.027037549763917923, + 0.06179878115653992, + 0.1340562403202057 + ], + "credit_map": [ + 0.010577182285487652, + 0.01896062307059765, + 0.009477267041802406, + 0.016377495601773262, + 0.00649155443534255, + 0.009502043016254902, + 0.006523949094116688, + 0.003752324264496565, + 0.003550691530108452, + 0.0031033738050609827, + 0.009377611801028252, + 0.005443668458610773, + 0.0034915872383862734, + 0.014865833334624767, + 0.0017429339932277799, + 0.0013046098174527287, + 0.002746197395026684, + 0.0027758406940847635, + 0.007501494605094194, + 0.00387554126791656, + 0.00338589190505445, + 0.012679615058004856, + 0.001324417651630938, + 0.000732261105440557, + 0.013310004025697708, + 0.013447742909193039, + 0.050468869507312775, + 0.019509924575686455, + 0.022783251479268074, + 0.08575796335935593, + 0.0026144180446863174, + 0.003822966944426298, + 0.019353512674570084, + 0.01979806460440159, + 0.07275106012821198, + 0.02652023360133171, + 0.03244955092668533, + 0.12237873673439026, + 0.007652029860764742, + 0.005920677445828915, + 0.0019230227917432785, + 0.001891592750325799, + 0.006245526950806379, + 0.0022928735706955194, + 0.0028431781101971865, + 0.010420836508274078, + 0.0011711390689015388, + 0.000618933467194438, + 0.005502820480614901, + 0.00463873939588666, + 0.016009243205189705, + 0.008483564481139183, + 0.006779662799090147, + 0.024037033319473267, + 0.003971394617110491, + 0.0021748885046690702, + 0.0033092389348894358, + 0.002493355656042695, + 0.00450587784871459, + 0.002592426259070635, + 0.0035920152440667152, + 0.006764566525816917, + 0.0018808802124112844, + 0.001899189199320972, + 0.0055111730471253395, + 0.003883958328515291, + 0.013270801864564419, + 0.006520080845803022, + 0.006170838605612516, + 0.02243700437247753, + 0.0021398465614765882, + 0.0018650743877515197, + 0.008182717487215996, + 0.009086493402719498, + 0.031385865062475204, + 0.012062383815646172, + 0.013654081150889397, + 0.05518689751625061, + 0.0020832789596170187, + 0.0024145194329321384 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.0017118098912760615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_12/index_tensor([1104, 1104])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_12/index_tensor([1104, 1104])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3476ba96f629c1c6afcf8169938ff84db04329b7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_12/index_tensor([1104, 1104])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot centers on the gnarled, dark gray branches of an old maple tree, silhouetted against the low, golden afternoon sunlight filtering through the forest canopy. Suddenly, a powerful, unseen gust of wind sweeps through the scene, causing the branches to whip violently. In a chaotic instant, dozens of vibrant red leaves are ripped from their stems, and the camera immediately tilts downward, perfectly tracking the ensuing cascade in mesmerizing slow motion. A mix of fiery red and deep scarlet, with papery edges and slightly curled, the leaves don't simply fall; they flee, tumble, twirl, and dance in swirls, their forms caught in the backlight like scattered pieces of luminous stained glass, silhouetted against the soft, out-of-focus autumnal tones of the deeper woods, creating a powerful and poetic atmosphere that captures both the raw power of nature and the melancholic beauty of decay.", + "seed": 10401, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_12/index_1104_seed_10401_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.012864084402099251, + "vlm_logprob": -0.012864084402099251, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.012864084402099251, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016173353418707848 + ], + "opposite_logprob_tokens": [ + -4.391173362731934 + ], + "logprob_reward": -0.016173353418707848, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016173353418707848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.331656813621521, + 0.06512065976858139, + 0.05486658215522766, + 0.20245881378650665, + 0.11562520265579224, + 0.044955089688301086, + 0.09193432331085205, + 0.03708287328481674, + 0.03566326946020126, + 0.020636428147554398 + ], + "credit_map": [ + 0.08476537466049194, + 0.01889808662235737, + 0.12439359724521637, + 0.020306644961237907, + 0.014754075556993484, + 0.018457962200045586, + 0.031443286687135696, + 0.01863773725926876, + 0.016501080244779587, + 0.004509556107223034, + 0.0058984034694731236, + 0.018928950652480125, + 0.003992758691310883, + 0.0023448513820767403, + 0.008856621570885181, + 0.004088432528078556, + 0.004949523136019707, + 0.004412709269672632, + 0.00794932059943676, + 0.007428311742842197, + 0.008835045620799065, + 0.005232286173850298, + 0.008788378909230232, + 0.0072709983214735985, + 0.013344774022698402, + 0.0064694699831306934, + 0.034208718687295914, + 0.015613039024174213, + 0.015277337282896042, + 0.0077391560189425945, + 0.07693672925233841, + 0.03286958113312721, + 0.014226952567696571, + 0.012002753093838692, + 0.018193092197179794, + 0.018726659938693047, + 0.016368702054023743, + 0.012497128918766975, + 0.012290041893720627, + 0.011319870129227638, + 0.004585534334182739, + 0.00395431462675333, + 0.005714477505534887, + 0.0059725153259932995, + 0.004300378262996674, + 0.005365277174860239, + 0.009109852835536003, + 0.005952735897153616, + 0.007500029169023037, + 0.004438927862793207, + 0.026326309889554977, + 0.00691457511857152, + 0.0058348774909973145, + 0.00611699465662241, + 0.006226515397429466, + 0.028576087206602097, + 0.004819776862859726, + 0.003498954465612769, + 0.009109851904213428, + 0.003318623872473836, + 0.0018723737448453903, + 0.0034792700316756964, + 0.0022241638507694006, + 0.008759855292737484, + 0.003478912403807044, + 0.003557750955224037, + 0.003571684705093503, + 0.003568468615412712, + 0.0039527942426502705, + 0.004874682519584894, + 0.00733557203784585, + 0.005323402117937803, + 0.0032392339780926704, + 0.002272645477205515, + 0.002206359291449189, + 0.002308128634467721, + 0.0016945466632023454, + 0.002594277961179614, + 0.0022875433787703514, + 0.004033691715449095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.016173353418707848 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022142434027045965 + ], + "opposite_logprob_tokens": [ + -6.127214431762695 + ], + "logprob_reward": -0.0022142434027045965, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022142434027045965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32575124502182007, + 0.02764703333377838, + 0.0699319913983345, + 0.06971769779920578, + 0.07611223310232162, + 0.022888151928782463, + 0.22160589694976807, + 0.07272864133119583, + 0.07696249336004257, + 0.03665463998913765 + ], + "credit_map": [ + 0.07995861023664474, + 0.020309999585151672, + 0.11895343661308289, + 0.019755590707063675, + 0.018072305247187614, + 0.018427401781082153, + 0.03067420795559883, + 0.019599679857492447, + 0.0066183083690702915, + 0.0021014094818383455, + 0.002575359772890806, + 0.0064610945992171764, + 0.002600101288408041, + 0.0013472067657858133, + 0.0031936189625412226, + 0.0027499340940266848, + 0.005861101672053337, + 0.013738003559410572, + 0.008340789005160332, + 0.005626635625958443, + 0.004648411180824041, + 0.003715852042660117, + 0.021122215315699577, + 0.006878983695060015, + 0.00575126800686121, + 0.0028874967247247696, + 0.007699344307184219, + 0.00752885639667511, + 0.0061006140895187855, + 0.004285227973014116, + 0.02438233606517315, + 0.011082552373409271, + 0.006295155733823776, + 0.005995043553411961, + 0.011728883720934391, + 0.012781462632119656, + 0.012907196767628193, + 0.009465504437685013, + 0.008613551035523415, + 0.008325430564582348, + 0.001874955021776259, + 0.002435196191072464, + 0.004427566193044186, + 0.002345059299841523, + 0.0027565909549593925, + 0.002300197957083583, + 0.0035571164917200804, + 0.0031914664432406425, + 0.01778370887041092, + 0.01133403554558754, + 0.06631293892860413, + 0.013089234009385109, + 0.01389703806489706, + 0.014843293465673923, + 0.013296869583427906, + 0.07104875892400742, + 0.007166093681007624, + 0.008774843998253345, + 0.01306847296655178, + 0.0069202035665512085, + 0.008707740344107151, + 0.010443621315062046, + 0.007806898560374975, + 0.009840761311352253, + 0.0073006474412977695, + 0.00830602366477251, + 0.009607047773897648, + 0.008529442362487316, + 0.008176224306225777, + 0.009615002200007439, + 0.01624334789812565, + 0.009184753522276878, + 0.005557245574891567, + 0.004458843730390072, + 0.0052483901381492615, + 0.004179751966148615, + 0.0037471202667802572, + 0.004311270546168089, + 0.004428696818649769, + 0.004723316989839077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.0022142434027045965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003197083715349436 + ], + "opposite_logprob_tokens": [ + -5.753197193145752 + ], + "logprob_reward": -0.003197083715349436, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003197083715349436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22058767080307007, + 0.07816523313522339, + 0.10204503685235977, + 0.07678356766700745, + 0.1127169206738472, + 0.037189897149801254, + 0.12481510639190674, + 0.07401632517576218, + 0.07708556205034256, + 0.09659460186958313 + ], + "credit_map": [ + 0.05032823234796524, + 0.014260994270443916, + 0.07656526565551758, + 0.014228817075490952, + 0.013958622701466084, + 0.012252927757799625, + 0.02116190642118454, + 0.017830923199653625, + 0.01511924434453249, + 0.005002469755709171, + 0.00850214995443821, + 0.021581554785370827, + 0.005250799935311079, + 0.003401519963517785, + 0.01333757396787405, + 0.005969934165477753, + 0.007300428580492735, + 0.021594896912574768, + 0.011784758418798447, + 0.010229034349322319, + 0.006523143965750933, + 0.006199495401233435, + 0.028378864750266075, + 0.010034426115453243, + 0.009257877245545387, + 0.0040225922130048275, + 0.014144916087388992, + 0.01564863882958889, + 0.007440346293151379, + 0.006859699729830027, + 0.010890726931393147, + 0.008518778719007969, + 0.012037181295454502, + 0.011533846147358418, + 0.01794310286641121, + 0.025720883160829544, + 0.012605463154613972, + 0.011908777058124542, + 0.009494124911725521, + 0.011473562568426132, + 0.003755784360691905, + 0.004051744472235441, + 0.004495848901569843, + 0.006595856510102749, + 0.004266486037522554, + 0.004514545667916536, + 0.003991961013525724, + 0.005517676472663879, + 0.014106138609349728, + 0.011071688495576382, + 0.025349078699946404, + 0.013869867660105228, + 0.014270259998738766, + 0.013382850214838982, + 0.010955950245261192, + 0.0218092892318964, + 0.008502544835209846, + 0.006492394953966141, + 0.014789573848247528, + 0.010767240077257156, + 0.005744012538343668, + 0.007687465287744999, + 0.009300868958234787, + 0.010732227936387062, + 0.008904926478862762, + 0.008519886061549187, + 0.007753766141831875, + 0.010577462613582611, + 0.007704099640250206, + 0.010843249037861824, + 0.012933548539876938, + 0.009848631918430328, + 0.013030098751187325, + 0.01163237914443016, + 0.014725695364177227, + 0.014854470267891884, + 0.0063773817382752895, + 0.010883313603699207, + 0.010414949618279934, + 0.01467631570994854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.003197083715349436 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636489786207676 + ], + "opposite_logprob_tokens": [ + -5.379636287689209 + ], + "logprob_reward": -0.004636489786207676, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636489786207676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37465375661849976, + 0.03920993208885193, + 0.10500948131084442, + 0.1306075006723404, + 0.12156691402196884, + 0.05416160076856613, + 0.05278247222304344, + 0.07773323357105255, + 0.022871971130371094, + 0.02140318974852562 + ], + "credit_map": [ + 0.09652968496084213, + 0.020640943199396133, + 0.1440780609846115, + 0.021221909672021866, + 0.014191791415214539, + 0.02092832699418068, + 0.0358264185488224, + 0.02123660035431385, + 0.004064095672219992, + 0.00239195441827178, + 0.0063669742085039616, + 0.004838894121348858, + 0.002703356556594372, + 0.001502497005276382, + 0.01464116107672453, + 0.0027009963523596525, + 0.008208146318793297, + 0.021217117086052895, + 0.01202835701406002, + 0.013968032784759998, + 0.0076732998713850975, + 0.006202745251357555, + 0.026627855375409126, + 0.009083922021090984, + 0.007307751104235649, + 0.004263635259121656, + 0.021225251257419586, + 0.008510890416800976, + 0.00947014894336462, + 0.004523462150245905, + 0.05014033615589142, + 0.025166010484099388, + 0.01584460958838463, + 0.013666216284036636, + 0.021036401391029358, + 0.026002677157521248, + 0.014270207844674587, + 0.009580014273524284, + 0.01041761226952076, + 0.01074916310608387, + 0.006117180921137333, + 0.00477177370339632, + 0.006670169532299042, + 0.0070510366931557655, + 0.005119788460433483, + 0.006092926487326622, + 0.011025414802134037, + 0.007313300855457783, + 0.005015670787543058, + 0.005447543226182461, + 0.009950386360287666, + 0.007948686368763447, + 0.0035871132276952267, + 0.005317113362252712, + 0.005389220081269741, + 0.010126731358468533, + 0.008951179683208466, + 0.008002725429832935, + 0.015788737684488297, + 0.012535564601421356, + 0.004349743481725454, + 0.0070161400362849236, + 0.006503983400762081, + 0.014585151337087154, + 0.002512948354706168, + 0.0033526213373988867, + 0.002651807852089405, + 0.0029189984779804945, + 0.0018097043503075838, + 0.0036018157843500376, + 0.002604601439088583, + 0.003419471438974142, + 0.003514522686600685, + 0.003487793728709221, + 0.0037496681325137615, + 0.002451673150062561, + 0.0016573626780882478, + 0.0025475183501839638, + 0.0013784895418211818, + 0.002616160549223423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.004636489786207676 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0380992516875267 + ], + "opposite_logprob_tokens": [ + -3.2880992889404297 + ], + "logprob_reward": -0.0380992516875267, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.0380992516875267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27280527353286743, + 0.08606381714344025, + 0.05953679606318474, + 0.10945729911327362, + 0.10148245841264725, + 0.05613110587000847, + 0.0793793573975563, + 0.07245239615440369, + 0.10434301197528839, + 0.05834854394197464 + ], + "credit_map": [ + 0.06916515529155731, + 0.014582514762878418, + 0.10468671470880508, + 0.01688980683684349, + 0.010659693740308285, + 0.015280921943485737, + 0.02662314847111702, + 0.014917287975549698, + 0.016681715846061707, + 0.004848959390074015, + 0.010361705906689167, + 0.022983888164162636, + 0.004062064923346043, + 0.003075482090935111, + 0.01957826316356659, + 0.004471731372177601, + 0.004110870882868767, + 0.011130711063742638, + 0.008959775790572166, + 0.009878994897007942, + 0.005476097576320171, + 0.003841797821223736, + 0.010271785780787468, + 0.005866756662726402, + 0.0110460314899683, + 0.005507688038051128, + 0.011119090020656586, + 0.022996552288532257, + 0.008808488957583904, + 0.007854072377085686, + 0.011166526935994625, + 0.030958838760852814, + 0.010741225443780422, + 0.01111811026930809, + 0.016351960599422455, + 0.02407493256032467, + 0.010402544401586056, + 0.00954816397279501, + 0.008567150682210922, + 0.010678364895284176, + 0.005843925289809704, + 0.004574484657496214, + 0.010921292938292027, + 0.009123378433287144, + 0.003976357169449329, + 0.006859437562525272, + 0.0061515923589468, + 0.00868063885718584, + 0.012795244343578815, + 0.008425739593803883, + 0.013460868038237095, + 0.008605845272541046, + 0.007929756306111813, + 0.010560867376625538, + 0.005969133228063583, + 0.011631898581981659, + 0.009923814795911312, + 0.00808941014111042, + 0.013273155316710472, + 0.0076774414628744125, + 0.006276332773268223, + 0.010465115308761597, + 0.007583872880786657, + 0.009163247421383858, + 0.00980981532484293, + 0.009679727256298065, + 0.010452089831233025, + 0.012029407545924187, + 0.013831619173288345, + 0.011709056794643402, + 0.02392517402768135, + 0.012906108982861042, + 0.008302179165184498, + 0.006916803307831287, + 0.008619779720902443, + 0.00886058621108532, + 0.006039025261998177, + 0.006234223488718271, + 0.006350905634462833, + 0.007025041617453098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.0380992516875267 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot centers on the gnarled, dark gray branches of an old maple tree, silhouetted against the low, golden afternoon sunlight filtering through the forest canopy. Suddenly, a powerful, unseen gust of wind sweeps through the scene, causing the branches to whip violently. In a chaotic instant, dozens of vibrant red leaves are ripped from their stems, and the camera immediately tilts downward, perfectly tracking the ensuing cascade in mesmerizing slow motion. A mix of fiery red and deep scarlet, with papery edges and slightly curled, the leaves don't simply fall; they flee, tumble, twirl, and dance in swirls, their forms caught in the backlight like scattered pieces of luminous stained glass, silhouetted against the soft, out-of-focus autumnal tones of the deeper woods, creating a powerful and poetic atmosphere that captures both the raw power of nature and the melancholic beauty of decay.", + "seed": 10402, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_12/index_1104_seed_10402_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0293698892230168, + "vlm_logprob": -0.0293698892230168, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.0293698892230168, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04831362143158913 + ], + "opposite_logprob_tokens": [ + -3.173313617706299 + ], + "logprob_reward": -0.04831362143158913, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04831362143158913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14182071387767792, + 0.03402767702937126, + 0.036396145820617676, + 0.13398151099681854, + 0.0851161852478981, + 0.3441427946090698, + 0.05658244341611862, + 0.05226115882396698, + 0.08818423748016357, + 0.027487121522426605 + ], + "credit_map": [ + 0.01398197840899229, + 0.01646197773516178, + 0.031365349888801575, + 0.011377612128853798, + 0.011538710445165634, + 0.016256961971521378, + 0.009827209636569023, + 0.03101092018187046, + 0.0030932766385376453, + 0.0035506421700119972, + 0.003292040666565299, + 0.004704716149717569, + 0.0034126066602766514, + 0.004297753795981407, + 0.0049936650320887566, + 0.006682977546006441, + 0.0057562063448131084, + 0.00599582027643919, + 0.0037247089203447104, + 0.003081756178289652, + 0.0038323767948895693, + 0.003641752293333411, + 0.006405549123883247, + 0.003957976587116718, + 0.011761062778532505, + 0.008675860241055489, + 0.01825818046927452, + 0.02615434117615223, + 0.013296263292431831, + 0.02339913882315159, + 0.017505235970020294, + 0.01493142917752266, + 0.005145592615008354, + 0.007487593684345484, + 0.012264279648661613, + 0.009874647483229637, + 0.01618616282939911, + 0.008792441338300705, + 0.015964165329933167, + 0.009401299990713596, + 0.029896151274442673, + 0.0811382308602333, + 0.08916424959897995, + 0.037417035549879074, + 0.03429492190480232, + 0.02283668704330921, + 0.022939201444387436, + 0.026456318795681, + 0.0046928683295845985, + 0.006440204102545977, + 0.011543847620487213, + 0.006812118925154209, + 0.0071549625135958195, + 0.005994390696287155, + 0.008981453254818916, + 0.004962598439306021, + 0.005510690156370401, + 0.010066181421279907, + 0.007590468041598797, + 0.007237404119223356, + 0.004503187723457813, + 0.003353722160682082, + 0.00875579658895731, + 0.005243709310889244, + 0.0066376724280416965, + 0.009752648882567883, + 0.007622922770678997, + 0.014212343841791153, + 0.010186776518821716, + 0.013632543385028839, + 0.019537901505827904, + 0.006601428147405386, + 0.0034554593730717897, + 0.0018336822977289557, + 0.003045538207516074, + 0.004013680852949619, + 0.004787771962583065, + 0.00332588329911232, + 0.0038858894258737564, + 0.003139216685667634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.04831362143158913 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001536618685349822 + ], + "opposite_logprob_tokens": [ + -6.501536846160889 + ], + "logprob_reward": -0.001536618685349822, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001536618685349822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055869944393634796, + 0.05805924907326698, + 0.04482070729136467, + 0.04658513143658638, + 0.11965145170688629, + 0.14560569822788239, + 0.0672416165471077, + 0.19508956372737885, + 0.13407905399799347, + 0.1329975724220276 + ], + "credit_map": [ + 0.0066924975253641605, + 0.005418492015451193, + 0.010410819202661514, + 0.007336183451116085, + 0.006888093892484903, + 0.008779491297900677, + 0.005044410936534405, + 0.00529995234683156, + 0.0064226617105305195, + 0.005785793531686068, + 0.004625434055924416, + 0.008178891614079475, + 0.005313694477081299, + 0.005840574856847525, + 0.012145777232944965, + 0.009746420197188854, + 0.005030997563153505, + 0.006029332056641579, + 0.005718626081943512, + 0.006324167363345623, + 0.004309895448386669, + 0.006096214987337589, + 0.006422462407499552, + 0.004889012314379215, + 0.0029038949869573116, + 0.0037524576764553785, + 0.008054176345467567, + 0.006917598191648722, + 0.003965928219258785, + 0.009768793359398842, + 0.005126813426613808, + 0.0060954708606004715, + 0.011396192014217377, + 0.018082886934280396, + 0.02047586254775524, + 0.01676984131336212, + 0.010330412536859512, + 0.009481878951191902, + 0.017867587506771088, + 0.015246790833771229, + 0.01353725790977478, + 0.02146373689174652, + 0.022130191326141357, + 0.02233806811273098, + 0.026928352192044258, + 0.013921868987381458, + 0.011648575775325298, + 0.013637653551995754, + 0.006270037963986397, + 0.008424402214586735, + 0.010457195341587067, + 0.011653563939034939, + 0.009104232303798199, + 0.006977022625505924, + 0.008089244365692139, + 0.0062659126706421375, + 0.00992859248071909, + 0.012607738375663757, + 0.042273733764886856, + 0.016814488917589188, + 0.012671764008700848, + 0.014536365866661072, + 0.05705750361084938, + 0.029199378564953804, + 0.01008689682930708, + 0.010069385170936584, + 0.010818461887538433, + 0.024950634688138962, + 0.016634326428174973, + 0.021524742245674133, + 0.030109915882349014, + 0.009884685277938843, + 0.0135004548355937, + 0.00999976322054863, + 0.0130861671641469, + 0.02167205885052681, + 0.018450632691383362, + 0.01843561977148056, + 0.022522013634443283, + 0.015330850146710873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.001536618685349822 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001948483637534082 + ], + "opposite_logprob_tokens": [ + -6.251948356628418 + ], + "logprob_reward": -0.001948483637534082, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001948483637534082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09709910303354263, + 0.0494065061211586, + 0.028607089072465897, + 0.06551012396812439, + 0.09388715773820877, + 0.12359806150197983, + 0.049848299473524094, + 0.34605273604393005, + 0.08460507541894913, + 0.06138581782579422 + ], + "credit_map": [ + 0.0111174201592803, + 0.011086801998317242, + 0.009524007327854633, + 0.010448715649545193, + 0.009992389939725399, + 0.0230945385992527, + 0.00891033560037613, + 0.012924893759191036, + 0.005331925582140684, + 0.0062076738104224205, + 0.0035001933574676514, + 0.007513778749853373, + 0.004534188192337751, + 0.004959456622600555, + 0.00943075306713581, + 0.007928539998829365, + 0.0026519806124269962, + 0.0035942704416811466, + 0.003208606503903866, + 0.004091080743819475, + 0.002460318151861429, + 0.005268624052405357, + 0.003965004347264767, + 0.0033672070130705833, + 0.006829071789979935, + 0.005624933633953333, + 0.006502178031951189, + 0.009988820180296898, + 0.006545165553689003, + 0.012665469199419022, + 0.007721364498138428, + 0.009633122943341732, + 0.008804470300674438, + 0.014278748072683811, + 0.01611780934035778, + 0.013733969070017338, + 0.00918850488960743, + 0.00662985909730196, + 0.012792875058948994, + 0.012340917252004147, + 0.010699174366891384, + 0.025843236595392227, + 0.02680004946887493, + 0.0162621159106493, + 0.015824226662516594, + 0.0077034602873027325, + 0.013152822852134705, + 0.007312982343137264, + 0.005233286879956722, + 0.0054303728975355625, + 0.005763707682490349, + 0.007204486057162285, + 0.006579811219125986, + 0.004928594455122948, + 0.006748448591679335, + 0.007959594950079918, + 0.026927437633275986, + 0.025690628215670586, + 0.09524612128734589, + 0.028853800147771835, + 0.02643739990890026, + 0.021084539592266083, + 0.10316133499145508, + 0.018651489168405533, + 0.00605690898373723, + 0.007168950513005257, + 0.0074695562943816185, + 0.010641597211360931, + 0.01091865636408329, + 0.01785329543054104, + 0.014645375311374664, + 0.009850743226706982, + 0.006522000301629305, + 0.0058470359072089195, + 0.006517503410577774, + 0.00610269233584404, + 0.014065545052289963, + 0.008913724683225155, + 0.007782690692692995, + 0.005634624511003494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.001948483637534082 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067780554294586 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067780554294586, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067780554294586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05923646315932274, + 0.022614726796746254, + 0.0327245332300663, + 0.023956814780831337, + 0.07112177461385727, + 0.3643348813056946, + 0.13694578409194946, + 0.12120386213064194, + 0.09340570122003555, + 0.07445544749498367 + ], + "credit_map": [ + 0.009848170913755894, + 0.006281271576881409, + 0.008664697408676147, + 0.0055739376693964005, + 0.006001575849950314, + 0.008443886414170265, + 0.004541240632534027, + 0.009881684556603432, + 0.002088947454467416, + 0.0034260451793670654, + 0.0020310725085437298, + 0.0034431186504662037, + 0.0022210017777979374, + 0.002490800339728594, + 0.004412220325320959, + 0.002501522656530142, + 0.0035048923455178738, + 0.004185816273093224, + 0.004662511870265007, + 0.005091738887131214, + 0.002424724167212844, + 0.004338642582297325, + 0.005054087843745947, + 0.0034621187951415777, + 0.0018533294787630439, + 0.0015329238958656788, + 0.0030962086748331785, + 0.004176810849457979, + 0.002173522487282753, + 0.004856886807829142, + 0.0028149718418717384, + 0.0034521622583270073, + 0.00420102383941412, + 0.00449260463938117, + 0.011689654551446438, + 0.011942765675485134, + 0.009875038638710976, + 0.0066371639259159565, + 0.013802150264382362, + 0.008481373079121113, + 0.03144412487745285, + 0.09134870022535324, + 0.08255040645599365, + 0.03280282020568848, + 0.035715483129024506, + 0.031630802899599075, + 0.022929605096578598, + 0.03591293841600418, + 0.011012285016477108, + 0.015952100977301598, + 0.02301020734012127, + 0.014976073987782001, + 0.01842198520898819, + 0.015265502035617828, + 0.02385910227894783, + 0.014448524452745914, + 0.007190975360572338, + 0.0069758095778524876, + 0.037290941923856735, + 0.008781634271144867, + 0.007377285975962877, + 0.006466826424002647, + 0.04071829095482826, + 0.006402099505066872, + 0.006854166742414236, + 0.006513944827020168, + 0.007433320861309767, + 0.019976209849119186, + 0.008919921703636646, + 0.01045234128832817, + 0.024950411170721054, + 0.0083053819835186, + 0.012248830869793892, + 0.007896745577454567, + 0.0051347906701266766, + 0.006967917084693909, + 0.014920477755367756, + 0.014102702960371971, + 0.007303991820663214, + 0.005879987496882677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.016067780554294586 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898294180631638 + ], + "opposite_logprob_tokens": [ + -2.5789828300476074 + ], + "logprob_reward": -0.07898294180631638, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898294180631638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0872533842921257, + 0.02764870785176754, + 0.03679366037249565, + 0.049525946378707886, + 0.10753042995929718, + 0.2019677311182022, + 0.1420619934797287, + 0.16320307552814484, + 0.1108168438076973, + 0.07319822907447815 + ], + "credit_map": [ + 0.012940553948283195, + 0.008425513282418251, + 0.016656575724482536, + 0.008623075671494007, + 0.00746788689866662, + 0.014467999339103699, + 0.008576939813792706, + 0.010094848461449146, + 0.003122262191027403, + 0.0040890867821872234, + 0.00289337825961411, + 0.002606183523312211, + 0.0032976088114082813, + 0.002731888322159648, + 0.0048692114651203156, + 0.004039087798446417, + 0.004457018803805113, + 0.005011661909520626, + 0.005304422229528427, + 0.004408593289554119, + 0.0034621679224073887, + 0.004568737465888262, + 0.005212855525314808, + 0.004368203692138195, + 0.0038563236594200134, + 0.0034846242051571608, + 0.00585028063505888, + 0.009991244412958622, + 0.0040915654972195625, + 0.006481761112809181, + 0.007682369090616703, + 0.008087778463959694, + 0.008292492479085922, + 0.015413603745400906, + 0.014120548963546753, + 0.018442008644342422, + 0.015325818210840225, + 0.007384470198303461, + 0.015504737384617329, + 0.013046751730144024, + 0.017968839034438133, + 0.032399412244558334, + 0.03625878691673279, + 0.03696781024336815, + 0.024970632046461105, + 0.012580830603837967, + 0.02352909743785858, + 0.017292337492108345, + 0.011425603181123734, + 0.015946343541145325, + 0.02297411859035492, + 0.014256108552217484, + 0.01863633655011654, + 0.015096587128937244, + 0.023953957483172417, + 0.019772933796048164, + 0.01807626709342003, + 0.026722023263573647, + 0.026959089562296867, + 0.019091825932264328, + 0.020207079127430916, + 0.01457778736948967, + 0.02678729221224785, + 0.010781710967421532, + 0.0066963182762265205, + 0.008296271786093712, + 0.009808552451431751, + 0.018961992114782333, + 0.011753823608160019, + 0.019061196595430374, + 0.023409588262438774, + 0.012829101644456387, + 0.008306424133479595, + 0.006669098976999521, + 0.00546182319521904, + 0.007663998752832413, + 0.013433573767542839, + 0.012909994460642338, + 0.010584336705505848, + 0.008168977685272694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.07898294180631638 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot centers on the gnarled, dark gray branches of an old maple tree, silhouetted against the low, golden afternoon sunlight filtering through the forest canopy. Suddenly, a powerful, unseen gust of wind sweeps through the scene, causing the branches to whip violently. In a chaotic instant, dozens of vibrant red leaves are ripped from their stems, and the camera immediately tilts downward, perfectly tracking the ensuing cascade in mesmerizing slow motion. A mix of fiery red and deep scarlet, with papery edges and slightly curled, the leaves don't simply fall; they flee, tumble, twirl, and dance in swirls, their forms caught in the backlight like scattered pieces of luminous stained glass, silhouetted against the soft, out-of-focus autumnal tones of the deeper woods, creating a powerful and poetic atmosphere that captures both the raw power of nature and the melancholic beauty of decay.", + "seed": 10403, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_12/index_1104_seed_10403_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.011507920059375464, + "vlm_logprob": -0.011507920059375464, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.011507920059375464, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020731192082166672 + ], + "opposite_logprob_tokens": [ + -4.145730972290039 + ], + "logprob_reward": -0.020731192082166672, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.020731192082166672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12067592889070511, + 0.05750848725438118, + 0.14324325323104858, + 0.10357557982206345, + 0.0764494463801384, + 0.030395416542887688, + 0.03089633770287037, + 0.13330455124378204, + 0.21713879704475403, + 0.08681228756904602 + ], + "credit_map": [ + 0.014357121661305428, + 0.015153547748923302, + 0.014121802523732185, + 0.010548212565481663, + 0.013245790265500546, + 0.013331630267202854, + 0.027129122987389565, + 0.012788685038685799, + 0.004933196119964123, + 0.005781732965260744, + 0.006624878849834204, + 0.003830562112852931, + 0.010205489583313465, + 0.006038089282810688, + 0.010660240426659584, + 0.009434293955564499, + 0.011243406683206558, + 0.008048184216022491, + 0.0315638966858387, + 0.011047718115150928, + 0.00893986877053976, + 0.01461965125054121, + 0.01906231790781021, + 0.03871819004416466, + 0.008256371133029461, + 0.008748408406972885, + 0.013021248392760754, + 0.026639210060238838, + 0.009625670500099659, + 0.01213712990283966, + 0.013162768445909023, + 0.01198475994169712, + 0.006883886642754078, + 0.006274116225540638, + 0.012680049054324627, + 0.013041146099567413, + 0.006606490816920996, + 0.010009892284870148, + 0.010559860616922379, + 0.010393993929028511, + 0.003645485732704401, + 0.004137012176215649, + 0.004801932722330093, + 0.004973409231752157, + 0.004166388418525457, + 0.0025043815840035677, + 0.003936813212931156, + 0.002229990204796195, + 0.002964763203635812, + 0.005282414611428976, + 0.005268210545182228, + 0.00539767649024725, + 0.002941109472885728, + 0.001966799609363079, + 0.005032374523580074, + 0.002042987383902073, + 0.01078795362263918, + 0.02042982541024685, + 0.018085939809679985, + 0.01972249150276184, + 0.014969727955758572, + 0.012374788522720337, + 0.020419443026185036, + 0.01651436649262905, + 0.018201645463705063, + 0.03416534885764122, + 0.015699390321969986, + 0.037142761051654816, + 0.03602006286382675, + 0.036632031202316284, + 0.016921313479542732, + 0.022356204688549042, + 0.009241299703717232, + 0.020268628373742104, + 0.006639542523771524, + 0.006991468369960785, + 0.019657757133245468, + 0.015862993896007538, + 0.004278526175767183, + 0.00387205951847136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.020731192082166672 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011986460303887725 + ], + "opposite_logprob_tokens": [ + -6.751198768615723 + ], + "logprob_reward": -0.0011986460303887725, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011986460303887725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0805695503950119, + 0.059405043721199036, + 0.07480978965759277, + 0.07192239165306091, + 0.026755070313811302, + 0.037402261048555374, + 0.030273258686065674, + 0.06535298377275467, + 0.07114313542842865, + 0.48236650228500366 + ], + "credit_map": [ + 0.011244981549680233, + 0.008656621910631657, + 0.012853041291236877, + 0.007691715843975544, + 0.008243243210017681, + 0.0069030458107590675, + 0.01466414425522089, + 0.010312763974070549, + 0.005024710204452276, + 0.0066937594674527645, + 0.004682785831391811, + 0.004202277399599552, + 0.012557584792375565, + 0.007593676447868347, + 0.01009547058492899, + 0.008554784581065178, + 0.006041512358933687, + 0.003920979797840118, + 0.018954772502183914, + 0.005085851531475782, + 0.0037956235464662313, + 0.007365351542830467, + 0.008644292131066322, + 0.021001411601901054, + 0.004123198799788952, + 0.004827321972697973, + 0.007875235751271248, + 0.012432632967829704, + 0.009231079369783401, + 0.010453937575221062, + 0.011987864039838314, + 0.010991130024194717, + 0.002651006216183305, + 0.0022117961198091507, + 0.003830095985904336, + 0.003179322462528944, + 0.0031293670181185007, + 0.0029758166056126356, + 0.003927224315702915, + 0.004850440658628941, + 0.003746419446542859, + 0.0045619686134159565, + 0.00669690640643239, + 0.004145096521824598, + 0.005939335562288761, + 0.004551460966467857, + 0.005057133734226227, + 0.0027039418928325176, + 0.0027707000263035297, + 0.0037769798655062914, + 0.004450798034667969, + 0.004199646878987551, + 0.0036586900241672993, + 0.0032999669201672077, + 0.0046983505599200726, + 0.003418128937482834, + 0.006003903690725565, + 0.008500722236931324, + 0.00637087132781744, + 0.015148227103054523, + 0.006445764563977718, + 0.008321340195834637, + 0.008254216983914375, + 0.006307943724095821, + 0.006234192289412022, + 0.010103884153068066, + 0.006485307589173317, + 0.010197299532592297, + 0.01045288983732462, + 0.008990217000246048, + 0.010305095463991165, + 0.008374256081879139, + 0.05093100294470787, + 0.1189405620098114, + 0.02658124454319477, + 0.02872893214225769, + 0.10566196590662003, + 0.09446294605731964, + 0.032873012125492096, + 0.02418683096766472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.0011986460303887725 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002201635157689452 + ], + "opposite_logprob_tokens": [ + -6.127201557159424 + ], + "logprob_reward": -0.002201635157689452, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002201635157689452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14516600966453552, + 0.03952237591147423, + 0.07385337352752686, + 0.054956357926130295, + 0.03658115863800049, + 0.042757797986269, + 0.05296468734741211, + 0.09591788053512573, + 0.1402437835931778, + 0.3180365264415741 + ], + "credit_map": [ + 0.01619396172463894, + 0.02120920829474926, + 0.024508433416485786, + 0.012563357129693031, + 0.01597783714532852, + 0.011176718398928642, + 0.022574404254555702, + 0.020962102338671684, + 0.003520496189594269, + 0.002703862963244319, + 0.006507827900350094, + 0.0034251934848725796, + 0.004295909311622381, + 0.003778395475819707, + 0.0075048464350402355, + 0.007785848807543516, + 0.006056418176740408, + 0.0072170523926615715, + 0.013486510142683983, + 0.0071145156398415565, + 0.004437674768269062, + 0.008398150093853474, + 0.01074066199362278, + 0.01640239916741848, + 0.004552630241960287, + 0.0040679797530174255, + 0.007955141365528107, + 0.00749190105125308, + 0.006859504617750645, + 0.008523298427462578, + 0.007715717423707247, + 0.007790188770741224, + 0.0032056053169071674, + 0.002780639100819826, + 0.005331842694431543, + 0.005665512289851904, + 0.005389987491071224, + 0.003365751588717103, + 0.005789603106677532, + 0.005052223335951567, + 0.004824470728635788, + 0.0058531914837658405, + 0.005977672059088945, + 0.0059533496387302876, + 0.007304733153432608, + 0.00452662818133831, + 0.004922698717564344, + 0.003395057050511241, + 0.00442699296399951, + 0.007189164869487286, + 0.0075268312357366085, + 0.009241356514394283, + 0.008151901885867119, + 0.004541185684502125, + 0.006473572459071875, + 0.00541368592530489, + 0.008691443130373955, + 0.01437915675342083, + 0.011128072626888752, + 0.015916980803012848, + 0.010039647109806538, + 0.011899233795702457, + 0.013118981383740902, + 0.010744375176727772, + 0.012620502151548862, + 0.022489307448267937, + 0.010118952952325344, + 0.017518777400255203, + 0.025629328563809395, + 0.02053721435368061, + 0.015400355681777, + 0.01592935249209404, + 0.03236786648631096, + 0.07255490869283676, + 0.018320035189390182, + 0.022556619718670845, + 0.07023978978395462, + 0.05936441570520401, + 0.022236837074160576, + 0.020396092906594276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.002201635157689452 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036140859592705965 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036140859592705965, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036140859592705965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0889682024717331, + 0.019880039617419243, + 0.06889449805021286, + 0.042169660329818726, + 0.06395897269248962, + 0.03560428321361542, + 0.02618837170302868, + 0.042656004428863525, + 0.11905357986688614, + 0.4926263988018036 + ], + "credit_map": [ + 0.014714706689119339, + 0.00867200642824173, + 0.011095833033323288, + 0.011716856621205807, + 0.005787708330899477, + 0.009099412709474564, + 0.015116496942937374, + 0.01276517566293478, + 0.00172170321457088, + 0.0013056860771030188, + 0.003916801419109106, + 0.0017222966998815536, + 0.0024493662640452385, + 0.00301484321244061, + 0.002696590032428503, + 0.0030527515336871147, + 0.005464920308440924, + 0.00465216813609004, + 0.014558343216776848, + 0.0071784150786697865, + 0.004128987900912762, + 0.007835419848561287, + 0.009123246185481548, + 0.01595299504697323, + 0.00339734205044806, + 0.003315509296953678, + 0.005585541017353535, + 0.004879304673522711, + 0.005588857922703028, + 0.005531651433557272, + 0.007417980115860701, + 0.006453473586589098, + 0.006709485314786434, + 0.005236869677901268, + 0.012565826997160912, + 0.010462229140102863, + 0.005793778691440821, + 0.007407199125736952, + 0.0115412138402462, + 0.00424237223342061, + 0.003245444968342781, + 0.004255115054547787, + 0.005826570093631744, + 0.004956439137458801, + 0.0054711150005459785, + 0.0037587613333016634, + 0.005003803409636021, + 0.003087033051997423, + 0.002280900487676263, + 0.004136369563639164, + 0.004015904851257801, + 0.0030140618328005075, + 0.003502494189888239, + 0.0019132349407300353, + 0.00504407100379467, + 0.0022813347168266773, + 0.0039042357821017504, + 0.008864200673997402, + 0.0036896211095154285, + 0.006276458501815796, + 0.006403582636266947, + 0.0037668440490961075, + 0.006986415479332209, + 0.0027646461967378855, + 0.010154732502996922, + 0.016809049993753433, + 0.012877145782113075, + 0.02123522013425827, + 0.017399828881025314, + 0.014096074737608433, + 0.015523984096944332, + 0.010957537218928337, + 0.05049702152609825, + 0.11779632419347763, + 0.029940588399767876, + 0.03098326176404953, + 0.10881167650222778, + 0.09345923364162445, + 0.03333960473537445, + 0.027798684313893318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0036140859592705965 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979404106736183 + ], + "opposite_logprob_tokens": [ + -3.5297939777374268 + ], + "logprob_reward": -0.02979404106736183, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979404106736183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08389294892549515, + 0.04405120387673378, + 0.04579543694853783, + 0.050752025097608566, + 0.08755137026309967, + 0.07583780586719513, + 0.04685543105006218, + 0.1694655418395996, + 0.19324490427970886, + 0.20255334675312042 + ], + "credit_map": [ + 0.00905547384172678, + 0.01588204875588417, + 0.008573045954108238, + 0.008889878168702126, + 0.0135354520753026, + 0.008314693346619606, + 0.011904450133442879, + 0.007737905718386173, + 0.00392715260386467, + 0.003598881885409355, + 0.009577256627380848, + 0.004022184759378433, + 0.0031513157300651073, + 0.005300784483551979, + 0.006725213024765253, + 0.0077484166249632835, + 0.003991184290498495, + 0.004099221434444189, + 0.008662094362080097, + 0.004774028901010752, + 0.003376019885763526, + 0.006810424849390984, + 0.005730212666094303, + 0.008352247066795826, + 0.005470083095133305, + 0.0066125718876719475, + 0.007321727462112904, + 0.0048247575759887695, + 0.005083523225039244, + 0.008816626854240894, + 0.00622731726616621, + 0.006395416799932718, + 0.010783599689602852, + 0.007116070482879877, + 0.01797887682914734, + 0.012303424067795277, + 0.008460462093353271, + 0.007429594174027443, + 0.01599808782339096, + 0.007481258828192949, + 0.008448710665106773, + 0.0189554151147604, + 0.014085640199482441, + 0.00793520174920559, + 0.009162584319710732, + 0.005229261238127947, + 0.008004605770111084, + 0.004016390535980463, + 0.004851201549172401, + 0.007673121988773346, + 0.006500765215605497, + 0.006418448407202959, + 0.00444779684767127, + 0.0046130758710205555, + 0.006236223503947258, + 0.006114800926297903, + 0.014184330590069294, + 0.026478108018636703, + 0.015340079553425312, + 0.02524726092815399, + 0.025226328521966934, + 0.013982790522277355, + 0.03789364919066429, + 0.011112996377050877, + 0.020775746554136276, + 0.026879623532295227, + 0.01676666922867298, + 0.02151641808450222, + 0.024465255439281464, + 0.033174481242895126, + 0.025235682725906372, + 0.0244310200214386, + 0.01843552105128765, + 0.04261372610926628, + 0.021882763132452965, + 0.017100758850574493, + 0.037417713552713394, + 0.03569705784320831, + 0.01803257130086422, + 0.011373236775398254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.02979404106736183 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot centers on the gnarled, dark gray branches of an old maple tree, silhouetted against the low, golden afternoon sunlight filtering through the forest canopy. Suddenly, a powerful, unseen gust of wind sweeps through the scene, causing the branches to whip violently. In a chaotic instant, dozens of vibrant red leaves are ripped from their stems, and the camera immediately tilts downward, perfectly tracking the ensuing cascade in mesmerizing slow motion. A mix of fiery red and deep scarlet, with papery edges and slightly curled, the leaves don't simply fall; they flee, tumble, twirl, and dance in swirls, their forms caught in the backlight like scattered pieces of luminous stained glass, silhouetted against the soft, out-of-focus autumnal tones of the deeper woods, creating a powerful and poetic atmosphere that captures both the raw power of nature and the melancholic beauty of decay.", + "seed": 10404, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_12/index_1104_seed_10404_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01850782735273242, + "vlm_logprob": -0.01850782735273242, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.01850782735273242, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026929708197712898 + ], + "opposite_logprob_tokens": [ + -3.7769296169281006 + ], + "logprob_reward": -0.026929708197712898, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.026929708197712898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06420327723026276, + 0.17445866763591766, + 0.1513935923576355, + 0.08599629253149033, + 0.07668167352676392, + 0.054309550672769547, + 0.11359411478042603, + 0.1838730126619339, + 0.042739905416965485, + 0.05274992808699608 + ], + "credit_map": [ + 0.0072284359484910965, + 0.004621230997145176, + 0.01137708779424429, + 0.007307255640625954, + 0.005814109928905964, + 0.007249488960951567, + 0.010283584706485271, + 0.0103220846503973, + 0.016735488548874855, + 0.012362943030893803, + 0.03910854458808899, + 0.024867096915841103, + 0.0181014034897089, + 0.017105378210544586, + 0.02699601650238037, + 0.01918179914355278, + 0.009528334252536297, + 0.007488562259823084, + 0.008221251890063286, + 0.007042068522423506, + 0.02325657568871975, + 0.04018021374940872, + 0.017842013388872147, + 0.03783459588885307, + 0.008768242783844471, + 0.010778481140732765, + 0.011743975803256035, + 0.014447477646172047, + 0.008197220973670483, + 0.007223543711006641, + 0.011606317944824696, + 0.01323104277253151, + 0.006193706300109625, + 0.004868333227932453, + 0.008270774967968464, + 0.011539174243807793, + 0.005437762476503849, + 0.00451952638104558, + 0.022874681279063225, + 0.012977723963558674, + 0.00353802228346467, + 0.006384979002177715, + 0.006731137167662382, + 0.0074867065995931625, + 0.005229613743722439, + 0.01001036912202835, + 0.007772959768772125, + 0.007155769970268011, + 0.006924700457602739, + 0.010640855878591537, + 0.019958846271038055, + 0.020106622949242592, + 0.015835076570510864, + 0.023064948618412018, + 0.010486520826816559, + 0.006576554384082556, + 0.015831492841243744, + 0.02353808842599392, + 0.014635010622441769, + 0.017414823174476624, + 0.03205738589167595, + 0.03239503130316734, + 0.031337540596723557, + 0.016663650050759315, + 0.004497229121625423, + 0.006384699605405331, + 0.00355400238186121, + 0.007210715673863888, + 0.00530754029750824, + 0.004651948343962431, + 0.005855944007635117, + 0.005277827847748995, + 0.007231310475617647, + 0.009614301845431328, + 0.00524080079048872, + 0.004955175332725048, + 0.008003235794603825, + 0.005393875762820244, + 0.007225661538541317, + 0.005085572134703398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 2 + ], + "selected_reward": -0.026929708197712898 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022161465603858232 + ], + "opposite_logprob_tokens": [ + -6.127216339111328 + ], + "logprob_reward": -0.0022161465603858232, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022161465603858232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06310269236564636, + 0.09730423986911774, + 0.03326399624347687, + 0.2189764529466629, + 0.0704563707113266, + 0.049193523824214935, + 0.08494598418474197, + 0.25079599022865295, + 0.05337066203355789, + 0.07859015464782715 + ], + "credit_map": [ + 0.009750881232321262, + 0.006994818802922964, + 0.006740848999470472, + 0.0067765917629003525, + 0.0050865476951003075, + 0.007444235496222973, + 0.00994846224784851, + 0.01036030799150467, + 0.010146650485694408, + 0.006560987327247858, + 0.02351405844092369, + 0.012071696110069752, + 0.011577162891626358, + 0.009559624828398228, + 0.013452060520648956, + 0.010421999730169773, + 0.0020774465519934893, + 0.0021789204329252243, + 0.0035265153273940086, + 0.003930192440748215, + 0.005064790602773428, + 0.005785161629319191, + 0.005673431791365147, + 0.005027538165450096, + 0.015173190273344517, + 0.011457745917141438, + 0.044188302010297775, + 0.050821688026189804, + 0.014872092753648758, + 0.01246559526771307, + 0.029675697907805443, + 0.04032212868332863, + 0.004192550666630268, + 0.004565665498375893, + 0.006519635207951069, + 0.012942267581820488, + 0.006451945286244154, + 0.005229610949754715, + 0.018997443839907646, + 0.011557254008948803, + 0.0029528974555432796, + 0.006056278012692928, + 0.005477519240230322, + 0.007470248732715845, + 0.004240168258547783, + 0.011170141398906708, + 0.0053164116106927395, + 0.0065098595805466175, + 0.005905951838940382, + 0.007867952808737755, + 0.017735889181494713, + 0.007351882755756378, + 0.0097257224842906, + 0.02266438491642475, + 0.007582385558634996, + 0.006111811846494675, + 0.02228330634534359, + 0.0343843437731266, + 0.024034632369875908, + 0.029953425750136375, + 0.04215356335043907, + 0.05501066893339157, + 0.02345852367579937, + 0.01951751485466957, + 0.004685843829065561, + 0.007090373896062374, + 0.005249553360044956, + 0.009732373058795929, + 0.007387684658169746, + 0.005461158696562052, + 0.006763611454516649, + 0.00700006540864706, + 0.007571590133011341, + 0.008505690842866898, + 0.006820788607001305, + 0.0077476645819842815, + 0.00768264289945364, + 0.007844029925763607, + 0.01959245838224888, + 0.012825285084545612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 1 + ], + "selected_reward": -0.0022161465603858232 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002202229807153344 + ], + "opposite_logprob_tokens": [ + -6.127202033996582 + ], + "logprob_reward": -0.002202229807153344, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002202229807153344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11053863912820816, + 0.09949645400047302, + 0.11600857228040695, + 0.13186919689178467, + 0.04861799255013466, + 0.09212931245565414, + 0.07601204514503479, + 0.18990804255008698, + 0.05246647447347641, + 0.08295325189828873 + ], + "credit_map": [ + 0.01644587144255638, + 0.012311025522649288, + 0.015536530874669552, + 0.010871809907257557, + 0.009024432860314846, + 0.012996679171919823, + 0.014996238984167576, + 0.018356051295995712, + 0.011451750993728638, + 0.008201832883059978, + 0.016208942979574203, + 0.013049490749835968, + 0.01205234881490469, + 0.011751271784305573, + 0.016706788912415504, + 0.010074038058519363, + 0.007601907476782799, + 0.005274447146803141, + 0.00876539945602417, + 0.007789871655404568, + 0.01754145883023739, + 0.025539929047226906, + 0.017827462404966354, + 0.025668110698461533, + 0.010035946033895016, + 0.008989698253571987, + 0.020396042615175247, + 0.01643001288175583, + 0.02131974883377552, + 0.011512087658047676, + 0.017949603497982025, + 0.025236060842871666, + 0.0036416572984308004, + 0.0033953229431062937, + 0.005927086807787418, + 0.008538921363651752, + 0.0036940702702850103, + 0.003396822838112712, + 0.010629142634570599, + 0.009394971653819084, + 0.004216192290186882, + 0.006559115368872881, + 0.01145472563803196, + 0.01221697498112917, + 0.010875153355300426, + 0.029390210285782814, + 0.006845918018370867, + 0.01057103369385004, + 0.005292021669447422, + 0.008197160437703133, + 0.01076521910727024, + 0.011454770341515541, + 0.010530451312661171, + 0.015478819608688354, + 0.006361714098602533, + 0.007931899279356003, + 0.014571568928658962, + 0.022005725651979446, + 0.016165031120181084, + 0.017009275034070015, + 0.031059324741363525, + 0.04090530797839165, + 0.027457622811198235, + 0.020734205842018127, + 0.0037802518345415592, + 0.0063417511992156506, + 0.004263498354703188, + 0.009166949428617954, + 0.007318011950701475, + 0.005019364412873983, + 0.009358176961541176, + 0.007218474522233009, + 0.00855813454836607, + 0.011240273714065552, + 0.007125398609787226, + 0.0073113529942929745, + 0.010163062252104282, + 0.007489826995879412, + 0.017685003578662872, + 0.013380206190049648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.002202229807153344 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535706162452698 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535706162452698, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535706162452698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04467059299349785, + 0.08742836862802505, + 0.16990071535110474, + 0.11582300812005997, + 0.07390882074832916, + 0.07559875398874283, + 0.1562601625919342, + 0.20299705862998962, + 0.05385179817676544, + 0.01956069841980934 + ], + "credit_map": [ + 0.00653374707326293, + 0.005089473444968462, + 0.006932008545845747, + 0.004080475308001041, + 0.0030593762639909983, + 0.00442779203876853, + 0.007914873771369457, + 0.006632843986153603, + 0.007338869851082563, + 0.008116531185805798, + 0.018989581614732742, + 0.011737274937331676, + 0.009673947468400002, + 0.008352510631084442, + 0.012834088876843452, + 0.010385571047663689, + 0.013361719436943531, + 0.008047609589993954, + 0.01638150028884411, + 0.01214023120701313, + 0.023976126685738564, + 0.03813701122999191, + 0.02213148958981037, + 0.035725027322769165, + 0.006173004396259785, + 0.007072821259498596, + 0.01655769534409046, + 0.025124119594693184, + 0.010682803578674793, + 0.006497371010482311, + 0.02403300814330578, + 0.019682183861732483, + 0.005792461801320314, + 0.007482744753360748, + 0.007077174261212349, + 0.01295409258455038, + 0.00778745673596859, + 0.004106379114091396, + 0.017348285764455795, + 0.011360223405063152, + 0.005466810427606106, + 0.009888263419270515, + 0.00794081948697567, + 0.010851815342903137, + 0.007862539030611515, + 0.012838712893426418, + 0.012009921483695507, + 0.008739873766899109, + 0.008922239765524864, + 0.013797678053379059, + 0.032087668776512146, + 0.025002121925354004, + 0.017896026372909546, + 0.03955262154340744, + 0.011804403737187386, + 0.007197401020675898, + 0.02041727490723133, + 0.024067796766757965, + 0.016478274017572403, + 0.018523162230849266, + 0.032591309398412704, + 0.033361926674842834, + 0.03771992027759552, + 0.019837399944663048, + 0.004222032148391008, + 0.0076871393248438835, + 0.004010489676147699, + 0.006984259933233261, + 0.006958998739719391, + 0.005642953794449568, + 0.00828190241008997, + 0.0100640207529068, + 0.0014767850516363978, + 0.0028395578265190125, + 0.0018122660694643855, + 0.0017844504909589887, + 0.001948817167431116, + 0.0025914020370692015, + 0.004037455189973116, + 0.0030699658673256636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.012535706162452698 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are gnarled, dark gray branches of an old maple tree visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do dozens of vibrant red leaves with papery edges and curled forms appear to be ripped from the branches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves tumble, twirl, and dance in swirls rather than simply falling straight down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera movement a tilt downward that tracks the cascade of leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the motion of the falling leaves depicted in slow motion?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048655346035957336 + ], + "opposite_logprob_tokens": [ + -3.0486552715301514 + ], + "logprob_reward": -0.048655346035957336, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048655346035957336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08297526091337204, + 0.0820806697010994, + 0.05862364172935486, + 0.08139213919639587, + 0.07125093042850494, + 0.11491188406944275, + 0.1337631344795227, + 0.19886073470115662, + 0.07044412940740585, + 0.10569745302200317 + ], + "credit_map": [ + 0.012986915186047554, + 0.01199603732675314, + 0.00895230658352375, + 0.007984508760273457, + 0.006764329504221678, + 0.009411963634192944, + 0.012256113812327385, + 0.012623093090951443, + 0.0067488932982087135, + 0.0070353043265640736, + 0.01079954020678997, + 0.015813404694199562, + 0.010321634821593761, + 0.00714813033118844, + 0.013232088647782803, + 0.010981673374772072, + 0.006055811885744333, + 0.003768058493733406, + 0.009573599323630333, + 0.008162235841155052, + 0.006308306474238634, + 0.0047328658401966095, + 0.012822858989238739, + 0.007199911866337061, + 0.007056014146655798, + 0.008468405343592167, + 0.008564838208258152, + 0.008074982091784477, + 0.009686127305030823, + 0.007934347726404667, + 0.019632112234830856, + 0.011975321918725967, + 0.0049766236916184425, + 0.005110299680382013, + 0.007560941390693188, + 0.012558305636048317, + 0.007641225587576628, + 0.004952251445502043, + 0.016462383791804314, + 0.011988908983767033, + 0.006446202751249075, + 0.008552294224500656, + 0.01322087086737156, + 0.018040701746940613, + 0.01139206811785698, + 0.024457203224301338, + 0.01673484779894352, + 0.01606770046055317, + 0.007826710119843483, + 0.01124721672385931, + 0.023964546620845795, + 0.017514633014798164, + 0.025333944708108902, + 0.024655569344758987, + 0.012516596354544163, + 0.010703918524086475, + 0.01629532128572464, + 0.021696168929338455, + 0.018290119245648384, + 0.018475038930773735, + 0.03898535296320915, + 0.041619978845119476, + 0.021273797377943993, + 0.022224977612495422, + 0.00376172480173409, + 0.007517548743635416, + 0.00513243256136775, + 0.011884399689733982, + 0.009553443640470505, + 0.008052241057157516, + 0.012957904487848282, + 0.011584440246224403, + 0.011401907540857792, + 0.0123789943754673, + 0.018055520951747894, + 0.016143495216965675, + 0.010475238785147667, + 0.010153823532164097, + 0.017105061560869217, + 0.009983418509364128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.048655346035957336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_13/index_tensor([5624, 5624])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_13/index_tensor([5624, 5624])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..36e1719e10c81b479f74194eff20bfe029e94e4f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_13/index_tensor([5624, 5624])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "visceral, first-person perspective shot violently lurches forward, plunging the viewer directly into the action, captured with the hyper-detailed, athletic style of a high-speed sports documentary. The camera rushes toward a dense, seemingly impenetrable wall of slender green Japanese bamboo. Bright, slightly diffused sunlight filters through the upper canopy, creating a dynamic effect of sharp highlights and deep shadows across the bamboo stems. As the moment of impact approaches, the shot shifts into extreme slow motion, revealing the complex physics of the collision with crystal clarity. Initial contact reveals the smooth, vibrant green bamboo stems bending unnaturally, their fibrous structure straining under immense force before they begin to shatter; the camera meticulously records the explosion of individual bamboo stems into a cloud of pale, woody fragments and a fine mist of internal moisture that catches the light. The forward momentum continues relentlessly, carving a path through the chaos as broken bamboo and torn, fluttering leaves whiz by, completely immersing the viewer in the raw, shocking force of the crash.", + "seed": 29385, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_13/index_5624_seed_29385_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04907463393174112, + "vlm_logprob": -0.04907463393174112, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.04907463393174112, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12510749697685242 + ], + "opposite_logprob_tokens": [ + -2.250107526779175 + ], + "logprob_reward": -0.12510749697685242, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12510749697685242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09706767648458481, + 0.02173660695552826, + 0.026554130017757416, + 0.26716840267181396, + 0.045380741357803345, + 0.14683039486408234, + 0.07275863736867905, + 0.19258557260036469, + 0.07820750027894974, + 0.0517103336751461 + ], + "credit_map": [ + 0.014256885275244713, + 0.0062094866298139095, + 0.012296536937355995, + 0.014259835705161095, + 0.008476263843476772, + 0.016702162101864815, + 0.015505785122513771, + 0.009360729716718197, + 0.0023464439436793327, + 0.0025205742567777634, + 0.0027293860912323, + 0.0032222019508481026, + 0.002235450316220522, + 0.00246486347168684, + 0.003730106633156538, + 0.002487578894942999, + 0.003451154101639986, + 0.0023730897810310125, + 0.0033742536325007677, + 0.003320785705000162, + 0.003966920543462038, + 0.0028582175727933645, + 0.0030442378483712673, + 0.004165471997112036, + 0.06988606601953506, + 0.023985762149095535, + 0.0231181550770998, + 0.024200445041060448, + 0.018010439351201057, + 0.017894133925437927, + 0.016947347670793533, + 0.07312605530023575, + 0.0024210659321397543, + 0.002283408073708415, + 0.005824516527354717, + 0.005076008848845959, + 0.00545021565631032, + 0.00573346484452486, + 0.012011283077299595, + 0.00658077746629715, + 0.017941391095519066, + 0.023364834487438202, + 0.02191569097340107, + 0.02441723830997944, + 0.012770180590450764, + 0.020278621464967728, + 0.012366724200546741, + 0.013775716535747051, + 0.008747827261686325, + 0.009563506580889225, + 0.01175131555646658, + 0.007193645462393761, + 0.008974178694188595, + 0.009202048182487488, + 0.00775469746440649, + 0.009571418166160583, + 0.02947455272078514, + 0.020592663437128067, + 0.029247770085930824, + 0.018091455101966858, + 0.03280717879533768, + 0.023996617645025253, + 0.018601523712277412, + 0.01977381482720375, + 0.009160093031823635, + 0.010022921487689018, + 0.017392151057720184, + 0.012689960189163685, + 0.0075996131636202335, + 0.008075844496488571, + 0.006818471010774374, + 0.006448452360928059, + 0.005229945294559002, + 0.005611454136669636, + 0.008724325336515903, + 0.006102410610765219, + 0.009227591566741467, + 0.005468863528221846, + 0.006031526252627373, + 0.005314220208674669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.12510749697685242 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020620493218302727 + ], + "opposite_logprob_tokens": [ + -3.895620584487915 + ], + "logprob_reward": -0.020620493218302727, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020620493218302727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08844191581010818, + 0.0692371279001236, + 0.040094856172800064, + 0.22405289113521576, + 0.062245577573776245, + 0.06995480507612228, + 0.06816962361335754, + 0.16803079843521118, + 0.18571946024894714, + 0.024052945896983147 + ], + "credit_map": [ + 0.014168918132781982, + 0.0075444504618644714, + 0.013839691877365112, + 0.007813691161572933, + 0.006632035132497549, + 0.014575197361409664, + 0.009184795431792736, + 0.014683141373097897, + 0.00683952309191227, + 0.007258928380906582, + 0.007692235987633467, + 0.013225456699728966, + 0.00644325464963913, + 0.00752847408875823, + 0.011719591915607452, + 0.008529667742550373, + 0.005288346204906702, + 0.005289448890835047, + 0.006144873797893524, + 0.0038866670802235603, + 0.006471047177910805, + 0.003942153416574001, + 0.004423878621309996, + 0.004648442380130291, + 0.05408337339758873, + 0.01946135237812996, + 0.01969497837126255, + 0.026374109089374542, + 0.014510159380733967, + 0.016357768326997757, + 0.01777617447078228, + 0.055794984102249146, + 0.0026293948758393526, + 0.0029144403524696827, + 0.010847147554159164, + 0.010137056931853294, + 0.008353304117918015, + 0.008387709967792034, + 0.009925195015966892, + 0.009051335044205189, + 0.004058856517076492, + 0.008843643590807915, + 0.009713339619338512, + 0.009831499308347702, + 0.0076128579676151276, + 0.01190410926938057, + 0.01188898365944624, + 0.006101522129029036, + 0.006914591882377863, + 0.006188237573951483, + 0.009460384026169777, + 0.0072960807010531425, + 0.010750367306172848, + 0.008453568443655968, + 0.010590012185275555, + 0.008516386151313782, + 0.03408665955066681, + 0.016149576753377914, + 0.03317783400416374, + 0.013766266405582428, + 0.017001217231154442, + 0.022969502955675125, + 0.010314780287444592, + 0.02056497149169445, + 0.02004346065223217, + 0.022847283631563187, + 0.04407733306288719, + 0.028659412637352943, + 0.016295839101076126, + 0.021567778661847115, + 0.014567763544619083, + 0.01766059920191765, + 0.0018645880045369267, + 0.0034819634165614843, + 0.0036806133575737476, + 0.002524095354601741, + 0.0030101074371486902, + 0.003664701944217086, + 0.002945643849670887, + 0.0028812363743782043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.020620493218302727 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897038757801056 + ], + "opposite_logprob_tokens": [ + -2.578970432281494 + ], + "logprob_reward": -0.07897038757801056, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897038757801056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07048541307449341, + 0.03735567629337311, + 0.22447797656059265, + 0.11854306608438492, + 0.0426027774810791, + 0.04447799548506737, + 0.07727006077766418, + 0.0964549258351326, + 0.24507121741771698, + 0.04326092451810837 + ], + "credit_map": [ + 0.008806182071566582, + 0.004283455666154623, + 0.009946906007826328, + 0.008092757314443588, + 0.0056939139030873775, + 0.014698711223900318, + 0.008693603798747063, + 0.010269875638186932, + 0.003905703779309988, + 0.004157594870775938, + 0.0037389404606074095, + 0.005450560711324215, + 0.004362616688013077, + 0.004226457327604294, + 0.0069981710985302925, + 0.004515627399086952, + 0.033532656729221344, + 0.024740038439631462, + 0.03648633882403374, + 0.029228268191218376, + 0.02555840276181698, + 0.021197030320763588, + 0.02307070791721344, + 0.030664514750242233, + 0.020601732656359673, + 0.01715986244380474, + 0.01190999522805214, + 0.01850764825940132, + 0.008988335728645325, + 0.011076195165514946, + 0.012658629566431046, + 0.017640655860304832, + 0.0017241798341274261, + 0.0020297772716730833, + 0.00635137315839529, + 0.00915833655744791, + 0.005115835927426815, + 0.005935347639024258, + 0.00797685794532299, + 0.004311061929911375, + 0.003974745515733957, + 0.008197248913347721, + 0.00615572789683938, + 0.006263693794608116, + 0.005246820393949747, + 0.0057523963041603565, + 0.0041974252089858055, + 0.0046899328008294106, + 0.009130608290433884, + 0.008229512721300125, + 0.011711747385561466, + 0.0075313332490623, + 0.009532302618026733, + 0.008979805745184422, + 0.011108511127531528, + 0.011046229861676693, + 0.008132980205118656, + 0.010779048316180706, + 0.012381973676383495, + 0.010071536526083946, + 0.01970513164997101, + 0.012352518737316132, + 0.014156234450638294, + 0.008875491097569466, + 0.03131795674562454, + 0.030012523755431175, + 0.050549931824207306, + 0.0370352528989315, + 0.021393124014139175, + 0.02819707989692688, + 0.01937413215637207, + 0.027191201224923134, + 0.003483496606349945, + 0.00495942635461688, + 0.00535944476723671, + 0.005935820750892162, + 0.006393023766577244, + 0.004648726899176836, + 0.006650140043348074, + 0.005830839276313782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.07897038757801056 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183324486017227 + ], + "opposite_logprob_tokens": [ + -4.01818323135376 + ], + "logprob_reward": -0.018183324486017227, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183324486017227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057788997888565063, + 0.0372878722846508, + 0.10017192363739014, + 0.13162609934806824, + 0.060810163617134094, + 0.06866170465946198, + 0.10085044056177139, + 0.06080159917473793, + 0.3335907757282257, + 0.048410482704639435 + ], + "credit_map": [ + 0.00832367967814207, + 0.004217346198856831, + 0.007600178476423025, + 0.007874231785535812, + 0.005843147169798613, + 0.009044446982443333, + 0.007651739288121462, + 0.007234226446598768, + 0.003930021543055773, + 0.0040530855767428875, + 0.004332836717367172, + 0.006127900909632444, + 0.004114191047847271, + 0.004248776938766241, + 0.0057461438700556755, + 0.004734913818538189, + 0.014993220567703247, + 0.014132690615952015, + 0.01644793339073658, + 0.011027641594409943, + 0.012703442014753819, + 0.009503443725407124, + 0.01023774966597557, + 0.011125802993774414, + 0.02623927779495716, + 0.013137992471456528, + 0.015553410165011883, + 0.015016810968518257, + 0.00877927616238594, + 0.00898575596511364, + 0.010686826892197132, + 0.03322674334049225, + 0.002996736206114292, + 0.003294676309451461, + 0.011183971539139748, + 0.007566660642623901, + 0.007652982138097286, + 0.009986711665987968, + 0.009184961207211018, + 0.008943461813032627, + 0.005838364828377962, + 0.006364074535667896, + 0.009610174223780632, + 0.010623198933899403, + 0.007523437961935997, + 0.009201613254845142, + 0.012995618395507336, + 0.006505216006189585, + 0.010978778824210167, + 0.01222471334040165, + 0.01504235714673996, + 0.009704910218715668, + 0.014464469626545906, + 0.012097550556063652, + 0.016304364427924156, + 0.010033302009105682, + 0.011715865693986416, + 0.004885711241513491, + 0.01141755934804678, + 0.0064567760564386845, + 0.004631293937563896, + 0.008665378205478191, + 0.006045260466635227, + 0.006983751431107521, + 0.041345056146383286, + 0.039693932980298996, + 0.07019361853599548, + 0.051814448088407516, + 0.0276725422590971, + 0.037138912826776505, + 0.03083391673862934, + 0.03489834815263748, + 0.004026568029075861, + 0.006264934781938791, + 0.004859196953475475, + 0.0040503740310668945, + 0.004105136264115572, + 0.007600703742355108, + 0.006263924762606621, + 0.011239646002650261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.018183324486017227 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002491467399522662 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.002491467399522662, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002491467399522662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13552609086036682, + 0.09424763917922974, + 0.16730843484401703, + 0.11805515736341476, + 0.08044645190238953, + 0.08864249289035797, + 0.049536898732185364, + 0.08748570084571838, + 0.1315726786851883, + 0.04717843234539032 + ], + "credit_map": [ + 0.01423435565084219, + 0.008936049416661263, + 0.018164504319429398, + 0.017356116324663162, + 0.022948041558265686, + 0.018885571509599686, + 0.019846079871058464, + 0.015155371278524399, + 0.006768438965082169, + 0.008612686768174171, + 0.013309257104992867, + 0.01571730338037014, + 0.011333384551107883, + 0.00965278223156929, + 0.017491880804300308, + 0.011361903510987759, + 0.024683313444256783, + 0.017879944294691086, + 0.022167401388287544, + 0.016975343227386475, + 0.02597089670598507, + 0.01764167658984661, + 0.01788324862718582, + 0.02410661242902279, + 0.021768396720290184, + 0.016272878274321556, + 0.009577048942446709, + 0.01954064890742302, + 0.008503847755491734, + 0.009418926201760769, + 0.010887534357607365, + 0.02208588272333145, + 0.003896631533280015, + 0.004235193599015474, + 0.01529170386493206, + 0.013424131087958813, + 0.012623359449207783, + 0.010708930902183056, + 0.010583044029772282, + 0.009683455340564251, + 0.005676772445440292, + 0.011537298560142517, + 0.01581263542175293, + 0.01378953643143177, + 0.008859386667609215, + 0.012628366239368916, + 0.012690793722867966, + 0.007647703867405653, + 0.007019827142357826, + 0.005899089854210615, + 0.009043589234352112, + 0.004179954994469881, + 0.006124721374362707, + 0.006730246357619762, + 0.006171157117933035, + 0.004368313122540712, + 0.00988063309341669, + 0.008122647181153297, + 0.013946987688541412, + 0.012624925933778286, + 0.007973561063408852, + 0.012210042215883732, + 0.008283132687211037, + 0.014443768188357353, + 0.014235742390155792, + 0.01602928154170513, + 0.029931992292404175, + 0.02077372558414936, + 0.011902638711035252, + 0.014810637570917606, + 0.010794113390147686, + 0.013094540685415268, + 0.005814462900161743, + 0.006355525925755501, + 0.007545425556600094, + 0.006151480600237846, + 0.005823277402669191, + 0.005341649986803532, + 0.00506962975487113, + 0.005076979752629995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.002491467399522662 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "visceral, first-person perspective shot violently lurches forward, plunging the viewer directly into the action, captured with the hyper-detailed, athletic style of a high-speed sports documentary. The camera rushes toward a dense, seemingly impenetrable wall of slender green Japanese bamboo. Bright, slightly diffused sunlight filters through the upper canopy, creating a dynamic effect of sharp highlights and deep shadows across the bamboo stems. As the moment of impact approaches, the shot shifts into extreme slow motion, revealing the complex physics of the collision with crystal clarity. Initial contact reveals the smooth, vibrant green bamboo stems bending unnaturally, their fibrous structure straining under immense force before they begin to shatter; the camera meticulously records the explosion of individual bamboo stems into a cloud of pale, woody fragments and a fine mist of internal moisture that catches the light. The forward momentum continues relentlessly, carving a path through the chaos as broken bamboo and torn, fluttering leaves whiz by, completely immersing the viewer in the raw, shocking force of the crash.", + "seed": 29386, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_13/index_5624_seed_29386_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06526872706599533, + "vlm_logprob": -0.06526872706599533, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.06526872706599533, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07984960824251175 + ], + "opposite_logprob_tokens": [ + -2.8298497200012207 + ], + "logprob_reward": -0.07984960824251175, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07984960824251175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23684653639793396, + 0.10661447048187256, + 0.0394449383020401, + 0.08731542527675629, + 0.04778832942247391, + 0.06052081286907196, + 0.0731426477432251, + 0.13994333148002625, + 0.1589498072862625, + 0.04943370446562767 + ], + "credit_map": [ + 0.017416339367628098, + 0.017353571951389313, + 0.020219195634126663, + 0.04459391161799431, + 0.023893125355243683, + 0.03979957848787308, + 0.029148058965802193, + 0.044422753155231476, + 0.01393132098019123, + 0.010708204470574856, + 0.01001722365617752, + 0.014838239178061485, + 0.010430436581373215, + 0.013726932927966118, + 0.01409521047025919, + 0.018866902217268944, + 0.0028054723516106606, + 0.004555936902761459, + 0.0032860003411769867, + 0.0048436252400279045, + 0.005378646310418844, + 0.005111808888614178, + 0.006258626934140921, + 0.007204821333289146, + 0.008436598815023899, + 0.007259146310389042, + 0.012637238018214703, + 0.009394113905727863, + 0.012611684389412403, + 0.016746439039707184, + 0.008934365585446358, + 0.01129583828151226, + 0.004916029516607523, + 0.003498590551316738, + 0.005726763978600502, + 0.0037061581388115883, + 0.00440000556409359, + 0.0074286931194365025, + 0.008602233603596687, + 0.009509853087365627, + 0.00614405469968915, + 0.0038879194762557745, + 0.008918666280806065, + 0.008088520728051662, + 0.005551110487431288, + 0.008757209405303001, + 0.009853658266365528, + 0.009319678880274296, + 0.008898126892745495, + 0.007137985900044441, + 0.010065063834190369, + 0.01069776713848114, + 0.006662104744464159, + 0.011419437825679779, + 0.007007823791354895, + 0.011254341341555119, + 0.03436627984046936, + 0.010019419714808464, + 0.011013992130756378, + 0.008389009162783623, + 0.00885876826941967, + 0.013732793740928173, + 0.013175204396247864, + 0.040387868881225586, + 0.012502877041697502, + 0.013431678526103497, + 0.02239232324063778, + 0.010504426434636116, + 0.023908404633402824, + 0.017477113753557205, + 0.01958085596561432, + 0.0391521193087101, + 0.0051270704716444016, + 0.007013737689703703, + 0.0035814642906188965, + 0.0035700127482414246, + 0.008805513381958008, + 0.011633810587227345, + 0.005513556767255068, + 0.004188535735011101 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.07984960824251175 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026394380256533623 + ], + "opposite_logprob_tokens": [ + -3.6513943672180176 + ], + "logprob_reward": -0.026394380256533623, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026394380256533623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17331404983997345, + 0.09648188948631287, + 0.051494043320417404, + 0.05282776057720184, + 0.16664648056030273, + 0.05949966982007027, + 0.06810925900936127, + 0.07093432545661926, + 0.1547902375459671, + 0.1059023067355156 + ], + "credit_map": [ + 0.01383613795042038, + 0.013275938108563423, + 0.011754505336284637, + 0.022652707993984222, + 0.030582444742321968, + 0.030178537592291832, + 0.017271438613533974, + 0.03376234322786331, + 0.010743826627731323, + 0.01172897033393383, + 0.007571263704448938, + 0.011262558400630951, + 0.016342751681804657, + 0.01493846345692873, + 0.013785245828330517, + 0.010108805261552334, + 0.0034445468336343765, + 0.005493300501257181, + 0.0052008493803441525, + 0.005939686205238104, + 0.007921530865132809, + 0.008340133354067802, + 0.007280619349330664, + 0.007873381488025188, + 0.004625645931810141, + 0.004365000408142805, + 0.009621393866837025, + 0.003840299788862467, + 0.0074121467769145966, + 0.007346716709434986, + 0.006422523409128189, + 0.009194032289087772, + 0.014479230158030987, + 0.010465818457305431, + 0.0222212765365839, + 0.012182485312223434, + 0.01925266534090042, + 0.03967604041099548, + 0.02061939798295498, + 0.02774956449866295, + 0.005892839748412371, + 0.0049691349267959595, + 0.009918758645653725, + 0.0060705081559717655, + 0.0071489885449409485, + 0.008804489858448505, + 0.007473994977772236, + 0.00922095775604248, + 0.007736708968877792, + 0.0050347172655165195, + 0.008469740860164165, + 0.00678031612187624, + 0.009444288909435272, + 0.009856613352894783, + 0.0069712442345917225, + 0.013815627433359623, + 0.010217709466814995, + 0.008812603540718555, + 0.006952287629246712, + 0.0060690101236104965, + 0.008029947988688946, + 0.009800178930163383, + 0.009771721437573433, + 0.011280861683189869, + 0.014509608037769794, + 0.013460888527333736, + 0.02246684394776821, + 0.011817548424005508, + 0.023246455937623978, + 0.014241890981793404, + 0.01867409609258175, + 0.03637290000915527, + 0.011303318664431572, + 0.018471375107765198, + 0.009715065360069275, + 0.00995124876499176, + 0.017896460369229317, + 0.014569985680282116, + 0.011176844127476215, + 0.012818013317883015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.026394380256533623 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015405148267746 + ], + "opposite_logprob_tokens": [ + -1.701540470123291 + ], + "logprob_reward": -0.2015405148267746, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015405148267746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1277940571308136, + 0.045580942183732986, + 0.08412182331085205, + 0.06367197632789612, + 0.26671987771987915, + 0.0706862136721611, + 0.10206201672554016, + 0.11207510530948639, + 0.0591292530298233, + 0.06815870851278305 + ], + "credit_map": [ + 0.0095971105620265, + 0.008066917769610882, + 0.009855806827545166, + 0.017083166167140007, + 0.01836431585252285, + 0.02214919961988926, + 0.014481681399047375, + 0.028195861726999283, + 0.00358604290522635, + 0.0038188539911061525, + 0.0043578664772212505, + 0.005306608974933624, + 0.006216354668140411, + 0.006349518429487944, + 0.006807475350797176, + 0.009138219989836216, + 0.005282524507492781, + 0.006543630268424749, + 0.007107260636985302, + 0.010265410877764225, + 0.010846407152712345, + 0.013318358920514584, + 0.020120183005928993, + 0.010638047009706497, + 0.006874486804008484, + 0.00599408894777298, + 0.008970336057245731, + 0.006223508156836033, + 0.007966901175677776, + 0.008448755368590355, + 0.010178876109421253, + 0.009015020914375782, + 0.02990259788930416, + 0.02613319270312786, + 0.03978338465094566, + 0.020524727180600166, + 0.029955601319670677, + 0.05206330865621567, + 0.03069019503891468, + 0.03766687214374542, + 0.009135630913078785, + 0.005156760104000568, + 0.010110353119671345, + 0.00795503705739975, + 0.007740418426692486, + 0.009136679582297802, + 0.010338451713323593, + 0.011112880893051624, + 0.010855844244360924, + 0.008400977589190006, + 0.013318621553480625, + 0.02106977440416813, + 0.007594054099172354, + 0.01237398386001587, + 0.011197051964700222, + 0.01725170761346817, + 0.029623394832015038, + 0.005345032084733248, + 0.010815189220011234, + 0.005355719476938248, + 0.004976199474185705, + 0.01094162929803133, + 0.009493614546954632, + 0.03552431985735893, + 0.0045948754996061325, + 0.008670800365507603, + 0.006838866509497166, + 0.005304741207510233, + 0.008387552574276924, + 0.008010544814169407, + 0.009908860549330711, + 0.0074130152352154255, + 0.009042480029165745, + 0.008779594674706459, + 0.005638152826577425, + 0.004156107548624277, + 0.011293184012174606, + 0.016542386263608932, + 0.006416895892471075, + 0.00628990214318037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.2015405148267746 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067545861005783 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.016067545861005783, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067545861005783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1463225781917572, + 0.046973396092653275, + 0.10738282650709152, + 0.07687652111053467, + 0.1146354079246521, + 0.0620390921831131, + 0.05498114973306656, + 0.077043317258358, + 0.12994977831840515, + 0.18379591405391693 + ], + "credit_map": [ + 0.014161262661218643, + 0.012306319549679756, + 0.008856399916112423, + 0.02086230181157589, + 0.026746511459350586, + 0.025444546714425087, + 0.016007745638489723, + 0.021937504410743713, + 0.00345910363830626, + 0.004224705975502729, + 0.0037282700650393963, + 0.006476937793195248, + 0.008397639729082584, + 0.007084236480295658, + 0.007066359277814627, + 0.006536145694553852, + 0.007561733014881611, + 0.012678583152592182, + 0.00852990709245205, + 0.012298648245632648, + 0.016131175681948662, + 0.01780964620411396, + 0.018712351098656654, + 0.01366078294813633, + 0.006450315471738577, + 0.00646398076787591, + 0.006554589606821537, + 0.010748754255473614, + 0.009848919697105885, + 0.01331900991499424, + 0.012122961692512035, + 0.011367990635335445, + 0.010426715947687626, + 0.009903755970299244, + 0.013020562939345837, + 0.010696395300328732, + 0.01522014755755663, + 0.017544668167829514, + 0.019907992333173752, + 0.017915166914463043, + 0.007984369061887264, + 0.003210071474313736, + 0.010487355291843414, + 0.008348353207111359, + 0.005509207956492901, + 0.006892741657793522, + 0.009007769636809826, + 0.010599221102893353, + 0.006315744947642088, + 0.003907954320311546, + 0.007703614886850119, + 0.009289002045989037, + 0.004931827075779438, + 0.006487328093498945, + 0.006881282664835453, + 0.009464393369853497, + 0.011291124857962132, + 0.006684156134724617, + 0.012284399941563606, + 0.007475723046809435, + 0.005787326954305172, + 0.009709116071462631, + 0.01276855543255806, + 0.011042912490665913, + 0.014420852065086365, + 0.02244570292532444, + 0.017252130433917046, + 0.011164389550685883, + 0.01662858948111534, + 0.013695811852812767, + 0.01423038262873888, + 0.020111922174692154, + 0.016268612816929817, + 0.018375953659415245, + 0.008456818759441376, + 0.008200081065297127, + 0.035386864095926285, + 0.06371542811393738, + 0.014083260670304298, + 0.019308896735310555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.016067545861005783 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024915861431509256 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.0024915861431509256, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024915861431509256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11321085691452026, + 0.09787681698799133, + 0.041488517075777054, + 0.06427057832479477, + 0.11119609326124191, + 0.1016625389456749, + 0.06511631608009338, + 0.045721106231212616, + 0.13106100261211395, + 0.22839616239070892 + ], + "credit_map": [ + 0.009732814505696297, + 0.010273817926645279, + 0.004571874625980854, + 0.015280213207006454, + 0.021858878433704376, + 0.017985347658395767, + 0.014500772580504417, + 0.019007138907909393, + 0.006743585225194693, + 0.008998293429613113, + 0.006427810061722994, + 0.01562691293656826, + 0.013624487444758415, + 0.013332678005099297, + 0.01638140343129635, + 0.01674165017902851, + 0.0035559264943003654, + 0.006788922939449549, + 0.0033329881262034178, + 0.004382132086902857, + 0.006380863022059202, + 0.006264328490942717, + 0.0041340421885252, + 0.006649313494563103, + 0.009423231706023216, + 0.004617991857230663, + 0.007383700925856829, + 0.006014641840010881, + 0.009556031785905361, + 0.014129241928458214, + 0.00536297308281064, + 0.007782772183418274, + 0.01135751698166132, + 0.007959924638271332, + 0.013546280562877655, + 0.009196998551487923, + 0.014030611142516136, + 0.020302826538681984, + 0.015649886801838875, + 0.019152047112584114, + 0.018236631527543068, + 0.007989750243723392, + 0.01728176884353161, + 0.011610365472733974, + 0.010844587348401546, + 0.013080621138215065, + 0.00961816031485796, + 0.01300064567476511, + 0.006455373018980026, + 0.005820274353027344, + 0.006881667301058769, + 0.01010982133448124, + 0.00810806080698967, + 0.007600876037031412, + 0.010639660991728306, + 0.009500585496425629, + 0.005044268444180489, + 0.00444416468963027, + 0.006027426104992628, + 0.004444403573870659, + 0.004945310764014721, + 0.004232512786984444, + 0.009735971689224243, + 0.0068470449186861515, + 0.011849450878798962, + 0.017198719084262848, + 0.013989140279591084, + 0.011694647371768951, + 0.02091873064637184, + 0.016034767031669617, + 0.016294879838824272, + 0.02308066561818123, + 0.023168202489614487, + 0.030169878154993057, + 0.012012687511742115, + 0.012112264521420002, + 0.04490630701184273, + 0.0663418099284172, + 0.016128528863191605, + 0.023556478321552277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.0024915861431509256 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "visceral, first-person perspective shot violently lurches forward, plunging the viewer directly into the action, captured with the hyper-detailed, athletic style of a high-speed sports documentary. The camera rushes toward a dense, seemingly impenetrable wall of slender green Japanese bamboo. Bright, slightly diffused sunlight filters through the upper canopy, creating a dynamic effect of sharp highlights and deep shadows across the bamboo stems. As the moment of impact approaches, the shot shifts into extreme slow motion, revealing the complex physics of the collision with crystal clarity. Initial contact reveals the smooth, vibrant green bamboo stems bending unnaturally, their fibrous structure straining under immense force before they begin to shatter; the camera meticulously records the explosion of individual bamboo stems into a cloud of pale, woody fragments and a fine mist of internal moisture that catches the light. The forward momentum continues relentlessly, carving a path through the chaos as broken bamboo and torn, fluttering leaves whiz by, completely immersing the viewer in the raw, shocking force of the crash.", + "seed": 29387, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_13/index_5624_seed_29387_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08018449996598065, + "vlm_logprob": -0.08018449996598065, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.08018449996598065, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0554681271314621 + ], + "opposite_logprob_tokens": [ + -3.1804680824279785 + ], + "logprob_reward": -0.0554681271314621, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0554681271314621, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05828026682138443, + 0.16864566504955292, + 0.15805684030056, + 0.055290043354034424, + 0.13055077195167542, + 0.07709141820669174, + 0.050617583096027374, + 0.12071262300014496, + 0.10043152421712875, + 0.0803232342004776 + ], + "credit_map": [ + 0.0061303796246647835, + 0.005636876914650202, + 0.004938787315040827, + 0.006066632457077503, + 0.005886424798518419, + 0.010511182248592377, + 0.0074516707099974155, + 0.011658313684165478, + 0.01508963480591774, + 0.023276854306459427, + 0.017783619463443756, + 0.021201469004154205, + 0.02154206484556198, + 0.023152122274041176, + 0.021325787529349327, + 0.0252741239964962, + 0.018994735553860664, + 0.0160510316491127, + 0.024869343265891075, + 0.02431483194231987, + 0.016713261604309082, + 0.018590368330478668, + 0.017707131803035736, + 0.02081613428890705, + 0.005769934970885515, + 0.007886089384555817, + 0.009176387451589108, + 0.005242795683443546, + 0.005216590128839016, + 0.011654875241219997, + 0.003486508736386895, + 0.00685686431825161, + 0.014479563571512699, + 0.019915664568543434, + 0.0171812791377306, + 0.021631939336657524, + 0.010359831154346466, + 0.018300095573067665, + 0.008488694205880165, + 0.02019369974732399, + 0.010812396183609962, + 0.012207779102027416, + 0.01016631256788969, + 0.012789134867489338, + 0.008185476996004581, + 0.008987442590296268, + 0.007732372730970383, + 0.006210502702742815, + 0.004447534680366516, + 0.00641622906550765, + 0.004640876315534115, + 0.006952029187232256, + 0.00662574078887701, + 0.00613029720261693, + 0.009172551333904266, + 0.006232327315956354, + 0.013771271333098412, + 0.020501870661973953, + 0.010617999359965324, + 0.008653518743813038, + 0.00885001476854086, + 0.010177155025303364, + 0.02110639587044716, + 0.027034396305680275, + 0.009956843219697475, + 0.0104000773280859, + 0.009623267687857151, + 0.015946311876177788, + 0.019343268126249313, + 0.010640675202012062, + 0.015981286764144897, + 0.008539794944226742, + 0.009949937462806702, + 0.010423003695905209, + 0.0101974718272686, + 0.010576983913779259, + 0.008321721106767654, + 0.009570898488163948, + 0.012559676542878151, + 0.008723536506295204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.0554681271314621 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029835810884833336 + ], + "opposite_logprob_tokens": [ + -3.5298357009887695 + ], + "logprob_reward": -0.029835810884833336, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029835810884833336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04900630563497543, + 0.11435136198997498, + 0.13757775723934174, + 0.05987030267715454, + 0.13823553919792175, + 0.11034409701824188, + 0.08050038665533066, + 0.14797963201999664, + 0.09144876897335052, + 0.07068584859371185 + ], + "credit_map": [ + 0.006173959467560053, + 0.0051468792371451855, + 0.004105810541659594, + 0.005862256977707148, + 0.007159512024372816, + 0.0065003433264791965, + 0.007832287810742855, + 0.006225254386663437, + 0.007468876428902149, + 0.008502804674208164, + 0.013604540377855301, + 0.012311050668358803, + 0.01797093078494072, + 0.02237892523407936, + 0.011308250948786736, + 0.02080598659813404, + 0.016532843932509422, + 0.014595460146665573, + 0.02233930304646492, + 0.02807406708598137, + 0.012867316603660583, + 0.013323587365448475, + 0.013671208173036575, + 0.01617397367954254, + 0.007305207196623087, + 0.01168056484311819, + 0.008730079978704453, + 0.006826454773545265, + 0.004662815947085619, + 0.009294168092310429, + 0.004151215311139822, + 0.007219797931611538, + 0.01164614874869585, + 0.01707225665450096, + 0.014109396375715733, + 0.021472368389368057, + 0.011461086571216583, + 0.022504417225718498, + 0.008633894845843315, + 0.031335968524217606, + 0.010882092639803886, + 0.012259822338819504, + 0.021333198994398117, + 0.015331953763961792, + 0.01347619853913784, + 0.01092094462364912, + 0.01359841600060463, + 0.012541467323899269, + 0.007784126792103052, + 0.016717780381441116, + 0.008289233781397343, + 0.008618933148682117, + 0.009834719821810722, + 0.010688708163797855, + 0.010049196891486645, + 0.008517688140273094, + 0.012525445781648159, + 0.017564499750733376, + 0.014654451981186867, + 0.02078351005911827, + 0.0187702439725399, + 0.012355253100395203, + 0.02590336836874485, + 0.02542285993695259, + 0.012188101187348366, + 0.014045817777514458, + 0.010173756629228592, + 0.012069744989275932, + 0.013380095362663269, + 0.008526506833732128, + 0.013473603874444962, + 0.007591138128191233, + 0.011313083581626415, + 0.008464865386486053, + 0.010183347389101982, + 0.00975237600505352, + 0.004965947940945625, + 0.007579166442155838, + 0.011321596801280975, + 0.007105470169335604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.029835810884833336 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520710825920105 + ], + "opposite_logprob_tokens": [ + -1.5020711421966553 + ], + "logprob_reward": -0.2520710825920105, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520710825920105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06928622722625732, + 0.14480812847614288, + 0.2743839621543884, + 0.05213988944888115, + 0.18269316852092743, + 0.05322365462779999, + 0.06853412836790085, + 0.07974544912576675, + 0.030631767585873604, + 0.04455369710922241 + ], + "credit_map": [ + 0.006770024541765451, + 0.007886761799454689, + 0.006101605016738176, + 0.007742066867649555, + 0.007893556728959084, + 0.010860872454941273, + 0.009663968347012997, + 0.012367370538413525, + 0.009832032024860382, + 0.022591691464185715, + 0.012718355283141136, + 0.017045047134160995, + 0.027573062106966972, + 0.0217550378292799, + 0.009018701501190662, + 0.024274202063679695, + 0.03038131631910801, + 0.04369469732046127, + 0.054808229207992554, + 0.03428357094526291, + 0.028628047555685043, + 0.022954193875193596, + 0.021964015439152718, + 0.037669871002435684, + 0.0061324592679739, + 0.009326702915132046, + 0.005108978599309921, + 0.005434297490864992, + 0.004124441184103489, + 0.008959193713963032, + 0.006116608623415232, + 0.006937200669199228, + 0.009906932711601257, + 0.01703278161585331, + 0.011899478733539581, + 0.02939632534980774, + 0.009992005303502083, + 0.03821943327784538, + 0.010891995392739773, + 0.05535421520471573, + 0.005608471110463142, + 0.002883080393075943, + 0.010711436159908772, + 0.005118364002555609, + 0.006901513785123825, + 0.005359221715480089, + 0.008481577038764954, + 0.00815998949110508, + 0.005729603115469217, + 0.01177830621600151, + 0.007100954651832581, + 0.008746200241148472, + 0.007461590692400932, + 0.009650317952036858, + 0.009504538960754871, + 0.008562608622014523, + 0.006606987677514553, + 0.011603341437876225, + 0.006014006678014994, + 0.00649599451571703, + 0.006605502683669329, + 0.00740954140201211, + 0.016330095008015633, + 0.018679967150092125, + 0.0027600647881627083, + 0.004298559855669737, + 0.0028980851639062166, + 0.004496345762163401, + 0.0048395199701189995, + 0.0030776970088481903, + 0.005061396863311529, + 0.0032000939827412367, + 0.004786514677107334, + 0.006283446215093136, + 0.00340969767421484, + 0.0067328778095543385, + 0.004239114001393318, + 0.006312762387096882, + 0.007801499217748642, + 0.004987781401723623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.2520710825920105 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062031928449869156 + ], + "opposite_logprob_tokens": [ + -2.8120319843292236 + ], + "logprob_reward": -0.062031928449869156, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062031928449869156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0733926072716713, + 0.08579367399215698, + 0.0867609903216362, + 0.04357726871967316, + 0.09342416375875473, + 0.06975461542606354, + 0.16674678027629852, + 0.2241571843624115, + 0.08342640101909637, + 0.07296621799468994 + ], + "credit_map": [ + 0.006917382124811411, + 0.004886814393103123, + 0.004658951424062252, + 0.007549501955509186, + 0.007161926478147507, + 0.016708258539438248, + 0.009483530186116695, + 0.016026249155402184, + 0.00518907280638814, + 0.011458470486104488, + 0.008594073355197906, + 0.010550073347985744, + 0.012146218679845333, + 0.014100976288318634, + 0.010926695540547371, + 0.0128280995413661, + 0.009922359138727188, + 0.012556645087897778, + 0.01122242771089077, + 0.014667394571006298, + 0.009084010496735573, + 0.011373649351298809, + 0.009098955430090427, + 0.00883555319160223, + 0.0047358241863548756, + 0.008505519479513168, + 0.005436192732304335, + 0.005292486399412155, + 0.004577108658850193, + 0.004724171012639999, + 0.004788346588611603, + 0.005517624318599701, + 0.01271106954663992, + 0.01735929399728775, + 0.00927825178951025, + 0.012248245999217033, + 0.006491456180810928, + 0.013187551870942116, + 0.007526491302996874, + 0.014621812850236893, + 0.008109678514301777, + 0.006116796284914017, + 0.014325402677059174, + 0.008125027641654015, + 0.00874913576990366, + 0.008176548406481743, + 0.006770421750843525, + 0.0093816127628088, + 0.01411444041877985, + 0.022771267220377922, + 0.012246986851096153, + 0.01877836510539055, + 0.027170071378350258, + 0.02943011187016964, + 0.024949433282017708, + 0.017286116257309914, + 0.025173358619213104, + 0.0410202220082283, + 0.023670131340622902, + 0.023779289796948433, + 0.02329588681459427, + 0.018888916820287704, + 0.028873665258288383, + 0.03945573419332504, + 0.010332931764423847, + 0.012303834781050682, + 0.00974371749907732, + 0.011214667931199074, + 0.011574912816286087, + 0.00821319967508316, + 0.012573864310979843, + 0.007469284813851118, + 0.008780957199633121, + 0.0070686000399291515, + 0.007978429086506367, + 0.012480066157877445, + 0.005786191672086716, + 0.00928330235183239, + 0.01392908114939928, + 0.007659594062715769 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.062031928449869156 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001515550771728158 + ], + "opposite_logprob_tokens": [ + -6.5015153884887695 + ], + "logprob_reward": -0.001515550771728158, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001515550771728158, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22085224092006683, + 0.09463932365179062, + 0.12413401156663895, + 0.051996178925037384, + 0.28119152784347534, + 0.0669582411646843, + 0.039004404097795486, + 0.03958076238632202, + 0.05860774964094162, + 0.02303551882505417 + ], + "credit_map": [ + 0.016545360907912254, + 0.014796579256653786, + 0.015087307430803776, + 0.027302321046590805, + 0.020123209804296494, + 0.03777880221605301, + 0.034516591578722, + 0.054702069610357285, + 0.005901863798499107, + 0.01817069947719574, + 0.006060849409550428, + 0.01122566219419241, + 0.018616218119859695, + 0.013648944906890392, + 0.010043617337942123, + 0.010971469804644585, + 0.016510624438524246, + 0.01983359269797802, + 0.017837204039096832, + 0.01904858835041523, + 0.013292912393808365, + 0.0119981300085783, + 0.012502443976700306, + 0.013110527768731117, + 0.004851764999330044, + 0.012003575451672077, + 0.006942173466086388, + 0.005077800713479519, + 0.004821885377168655, + 0.008493308909237385, + 0.004185268189758062, + 0.00562040600925684, + 0.02205902524292469, + 0.0285467691719532, + 0.021405218169093132, + 0.04091791808605194, + 0.023634450510144234, + 0.051346294581890106, + 0.01825784705579281, + 0.07502402365207672, + 0.005535800941288471, + 0.005902519915252924, + 0.008207646198570728, + 0.00992403831332922, + 0.008925740607082844, + 0.006711356807500124, + 0.012659074738621712, + 0.00909206923097372, + 0.0035064108669757843, + 0.006718703545629978, + 0.0032688325736671686, + 0.003037368645891547, + 0.0062424554489552975, + 0.005535285919904709, + 0.005646103527396917, + 0.0050492496229708195, + 0.004497861955314875, + 0.0036736319307237864, + 0.004182206466794014, + 0.0030761186499148607, + 0.004132393281906843, + 0.003480705665424466, + 0.008499897085130215, + 0.008037948980927467, + 0.006463023833930492, + 0.006064443849027157, + 0.005906309466809034, + 0.008512388914823532, + 0.010868284851312637, + 0.005710219964385033, + 0.009835761040449142, + 0.005247320514172316, + 0.0018224482191726565, + 0.004234975203871727, + 0.0020884531550109386, + 0.0032144079450517893, + 0.0026801496278494596, + 0.0026242753956466913, + 0.0036914010997861624, + 0.0026794092264026403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.001515550771728158 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "visceral, first-person perspective shot violently lurches forward, plunging the viewer directly into the action, captured with the hyper-detailed, athletic style of a high-speed sports documentary. The camera rushes toward a dense, seemingly impenetrable wall of slender green Japanese bamboo. Bright, slightly diffused sunlight filters through the upper canopy, creating a dynamic effect of sharp highlights and deep shadows across the bamboo stems. As the moment of impact approaches, the shot shifts into extreme slow motion, revealing the complex physics of the collision with crystal clarity. Initial contact reveals the smooth, vibrant green bamboo stems bending unnaturally, their fibrous structure straining under immense force before they begin to shatter; the camera meticulously records the explosion of individual bamboo stems into a cloud of pale, woody fragments and a fine mist of internal moisture that catches the light. The forward momentum continues relentlessly, carving a path through the chaos as broken bamboo and torn, fluttering leaves whiz by, completely immersing the viewer in the raw, shocking force of the crash.", + "seed": 29388, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_13/index_5624_seed_29388_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10662041688337923, + "vlm_logprob": -0.10662041688337923, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.10662041688337923, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09013703465461731 + ], + "opposite_logprob_tokens": [ + -2.715137004852295 + ], + "logprob_reward": -0.09013703465461731, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09013703465461731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1727123260498047, + 0.05088674649596214, + 0.05383586138486862, + 0.09667418897151947, + 0.035206783562898636, + 0.26421770453453064, + 0.16202668845653534, + 0.04091513529419899, + 0.08418725430965424, + 0.039337337017059326 + ], + "credit_map": [ + 0.012854082509875298, + 0.020347313955426216, + 0.04405584931373596, + 0.027693483978509903, + 0.009768890216946602, + 0.010436969809234142, + 0.016191262751817703, + 0.03136445954442024, + 0.002982281381264329, + 0.005594253074377775, + 0.006169091444462538, + 0.009813817217946053, + 0.006568781565874815, + 0.0061208102852106094, + 0.0070771570317447186, + 0.006560548208653927, + 0.005376596003770828, + 0.007722280919551849, + 0.0074856779538095, + 0.00984033104032278, + 0.004896324127912521, + 0.007373165339231491, + 0.005269151646643877, + 0.005872327368706465, + 0.008321838453412056, + 0.005918414331972599, + 0.024402478709816933, + 0.008090372197329998, + 0.007750945631414652, + 0.026746463030576706, + 0.007026225328445435, + 0.008417448028922081, + 0.007774500176310539, + 0.0026157440152019262, + 0.005995320156216621, + 0.003895449684932828, + 0.003382451133802533, + 0.0039014576468616724, + 0.003954688552767038, + 0.0036871694028377533, + 0.0206350926309824, + 0.015387291088700294, + 0.03013567440211773, + 0.04404844716191292, + 0.020720677450299263, + 0.030619921162724495, + 0.06843628734350204, + 0.03423430025577545, + 0.011550572700798512, + 0.01068679615855217, + 0.02481604740023613, + 0.022703683003783226, + 0.00996339786797762, + 0.033735521137714386, + 0.02325110323727131, + 0.025319552049040794, + 0.00495158089324832, + 0.004782593809068203, + 0.004799575544893742, + 0.003957534208893776, + 0.004653003998100758, + 0.005425495561212301, + 0.0044538043439388275, + 0.007891543209552765, + 0.008815031498670578, + 0.018373416736721992, + 0.010562681593000889, + 0.004533102735877037, + 0.020623715594410896, + 0.009071843698620796, + 0.00599212571978569, + 0.006215328350663185, + 0.0050249178893864155, + 0.008288687095046043, + 0.005903391633182764, + 0.003497311845421791, + 0.0069837081246078014, + 0.004200882278382778, + 0.002983333542943001, + 0.002455097623169422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.09013703465461731 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016114939004182816 + ], + "opposite_logprob_tokens": [ + -4.141114711761475 + ], + "logprob_reward": -0.016114939004182816, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016114939004182816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16521227359771729, + 0.09365495294332504, + 0.1318690925836563, + 0.05434533953666687, + 0.046591684222221375, + 0.05476617068052292, + 0.1059606745839119, + 0.09734413772821426, + 0.09824031591415405, + 0.1520153284072876 + ], + "credit_map": [ + 0.01490520965307951, + 0.01991083286702633, + 0.022812340408563614, + 0.0363088883459568, + 0.0144175561144948, + 0.017998509109020233, + 0.021151844412088394, + 0.0177070964127779, + 0.006805430632084608, + 0.010143019258975983, + 0.01702292449772358, + 0.010130204260349274, + 0.013111566193401814, + 0.01912708207964897, + 0.007472536992281675, + 0.009842193685472012, + 0.015714596956968307, + 0.02069905772805214, + 0.015630153939127922, + 0.024402495473623276, + 0.010595056228339672, + 0.01937817595899105, + 0.012519675306975842, + 0.012929894961416721, + 0.006172746419906616, + 0.00506179966032505, + 0.009698735550045967, + 0.008097955957055092, + 0.0036324248649179935, + 0.010447676293551922, + 0.004494824446737766, + 0.006739182397723198, + 0.00923633947968483, + 0.0030640834011137486, + 0.005598589777946472, + 0.004143810365349054, + 0.0036274930462241173, + 0.0067499722354114056, + 0.00985097698867321, + 0.004320420790463686, + 0.005501734092831612, + 0.00359300896525383, + 0.005983876530081034, + 0.00967459287494421, + 0.0034161554649472237, + 0.005623300559818745, + 0.013351213186979294, + 0.007622290402650833, + 0.010857225395739079, + 0.006659908220171928, + 0.014052418991923332, + 0.016201168298721313, + 0.0062936823815107346, + 0.020663706585764885, + 0.015892084687948227, + 0.01534048467874527, + 0.006481996271759272, + 0.009680503979325294, + 0.009390022605657578, + 0.011770190671086311, + 0.010369019582867622, + 0.01367152389138937, + 0.019830090925097466, + 0.01615079492330551, + 0.01202305406332016, + 0.011294394731521606, + 0.017925476655364037, + 0.007383299991488457, + 0.0156935453414917, + 0.015840277075767517, + 0.00797946099191904, + 0.010100815445184708, + 0.02149842120707035, + 0.02553083933889866, + 0.026366138830780983, + 0.014459890313446522, + 0.02537417970597744, + 0.017735477536916733, + 0.011560139246284962, + 0.009490241296589375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.016114939004182816 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870398998260498 + ], + "opposite_logprob_tokens": [ + -1.1370398998260498 + ], + "logprob_reward": -0.3870398998260498, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3870398998260498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20883294939994812, + 0.09212660789489746, + 0.07950825244188309, + 0.09187807887792587, + 0.03582383692264557, + 0.09188667684793472, + 0.1295209527015686, + 0.0827183797955513, + 0.05617969483137131, + 0.13152456283569336 + ], + "credit_map": [ + 0.019668929278850555, + 0.023884454742074013, + 0.04332074150443077, + 0.038787804543972015, + 0.0208517424762249, + 0.017885729670524597, + 0.019917448982596397, + 0.024516083300113678, + 0.00797528587281704, + 0.011152350343763828, + 0.011844038031995296, + 0.013504368253052235, + 0.012107321992516518, + 0.011991163715720177, + 0.011608795262873173, + 0.011943281628191471, + 0.005051691550761461, + 0.005608269479125738, + 0.009145703166723251, + 0.01766127720475197, + 0.006942728068679571, + 0.014148588292300701, + 0.009428595192730427, + 0.011521399021148682, + 0.007457138504832983, + 0.005343798082321882, + 0.02615223079919815, + 0.009194140322506428, + 0.005520396865904331, + 0.02477361634373665, + 0.004204711876809597, + 0.009232043288648129, + 0.00981785450130701, + 0.00364813394844532, + 0.004791364539414644, + 0.0030278211925178766, + 0.0032986029982566833, + 0.003961329348385334, + 0.003736963262781501, + 0.0035417689941823483, + 0.008433463051915169, + 0.007555122021585703, + 0.014372900128364563, + 0.01521431002765894, + 0.009233305230736732, + 0.013604207895696163, + 0.014997598715126514, + 0.0084757711738348, + 0.01093558594584465, + 0.008460396900773048, + 0.019495131447911263, + 0.01836567558348179, + 0.007290422450751066, + 0.026482779532670975, + 0.018234634771943092, + 0.020256323739886284, + 0.0063026538118720055, + 0.008933177217841148, + 0.008189680986106396, + 0.0114900479093194, + 0.008421584963798523, + 0.014702511951327324, + 0.011347721330821514, + 0.01333099976181984, + 0.009299165569245815, + 0.007712834049016237, + 0.013962152414023876, + 0.004427221138030291, + 0.005947520490735769, + 0.006120104808360338, + 0.004155402071774006, + 0.0045552910305559635, + 0.017294427379965782, + 0.017826298251748085, + 0.01970571279525757, + 0.01569215953350067, + 0.019638419151306152, + 0.016989024356007576, + 0.011930310167372227, + 0.012448208406567574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.3870398998260498 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809328377246857 + ], + "opposite_logprob_tokens": [ + -3.288093328475952 + ], + "logprob_reward": -0.03809328377246857, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809328377246857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28168442845344543, + 0.09604546427726746, + 0.10000569373369217, + 0.07165780663490295, + 0.02497154287993908, + 0.08025636523962021, + 0.04742644354701042, + 0.19975267350673676, + 0.04503348469734192, + 0.05316607281565666 + ], + "credit_map": [ + 0.01644442416727543, + 0.019275356084108353, + 0.0927383154630661, + 0.04239001125097275, + 0.014558480121195316, + 0.014428112655878067, + 0.011954005807638168, + 0.06989574432373047, + 0.011228247545659542, + 0.0114399129524827, + 0.01726567931473255, + 0.013280173763632774, + 0.010211460292339325, + 0.01132950372993946, + 0.010092318058013916, + 0.011198178865015507, + 0.008031672798097134, + 0.011585011146962643, + 0.00963797327131033, + 0.017015816643834114, + 0.010996496304869652, + 0.016720550134778023, + 0.01490488275885582, + 0.011113302782177925, + 0.005116431042551994, + 0.003456238890066743, + 0.02104176953434944, + 0.006788954604417086, + 0.00413561100140214, + 0.02303292602300644, + 0.004204739350825548, + 0.0038811410777270794, + 0.0045567709021270275, + 0.0033458999823778868, + 0.003910745959728956, + 0.0028789222706109285, + 0.0029074414633214474, + 0.002300343243405223, + 0.0023403808008879423, + 0.0027310384903103113, + 0.008502913638949394, + 0.004008191172033548, + 0.009990057907998562, + 0.014720985665917397, + 0.0051020607352256775, + 0.008598780259490013, + 0.018254553899168968, + 0.011078832671046257, + 0.003963368479162455, + 0.0033028218895196915, + 0.007602125406265259, + 0.0070718806236982346, + 0.0029702428728342056, + 0.00889734085649252, + 0.006704131606966257, + 0.006914537865668535, + 0.012730005197227001, + 0.021118517965078354, + 0.018315041437745094, + 0.022807065397500992, + 0.0200685765594244, + 0.02740403078496456, + 0.037276510149240494, + 0.040032919496297836, + 0.00506646279245615, + 0.008145324885845184, + 0.0067931716330349445, + 0.00394267775118351, + 0.006338780280202627, + 0.006731786299496889, + 0.0042643798515200615, + 0.0037509044632315636, + 0.008526813238859177, + 0.007576626725494862, + 0.0067077456042170525, + 0.006709594279527664, + 0.007624008227139711, + 0.006168881896883249, + 0.004386314190924168, + 0.005466095171868801 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.03809328377246857 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot filmed from a first-person perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera rush forward toward a dense wall of slender green Japanese bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video begin with a violent forward lurch before shifting to extreme slow motion at impact?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the bamboo stems shown bending unnaturally and shattering into pale woody fragments and a mist of moisture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by bright, slightly diffused sunlight creating sharp highlights and deep shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017169271595776081 + ], + "opposite_logprob_tokens": [ + -6.3767170906066895 + ], + "logprob_reward": -0.0017169271595776081, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017169271595776081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2622792422771454, + 0.0876086875796318, + 0.1554299294948578, + 0.07489809393882751, + 0.029194314032793045, + 0.061813339591026306, + 0.08077894896268845, + 0.09718242287635803, + 0.07999041676521301, + 0.07082462310791016 + ], + "credit_map": [ + 0.017951561138033867, + 0.034149169921875, + 0.04014338552951813, + 0.04500124603509903, + 0.028522267937660217, + 0.02634705789387226, + 0.03678010404109955, + 0.03338447958230972, + 0.005731509067118168, + 0.010187364183366299, + 0.014365581795573235, + 0.012151943519711494, + 0.012026892974972725, + 0.014394180849194527, + 0.010138948448002338, + 0.008612261153757572, + 0.018007982522249222, + 0.02528887428343296, + 0.017507590353488922, + 0.023894673213362694, + 0.016740499064326286, + 0.019473867490887642, + 0.019133267924189568, + 0.015383169054985046, + 0.006385493092238903, + 0.006258393172174692, + 0.012130174785852432, + 0.011139881797134876, + 0.009897194802761078, + 0.014799963682889938, + 0.00621697586029768, + 0.008070020005106926, + 0.007746731396764517, + 0.0028104581870138645, + 0.003841422265395522, + 0.002806272590532899, + 0.0027521110605448484, + 0.003422038396820426, + 0.003626585938036442, + 0.00218869443051517, + 0.007577837444841862, + 0.006018655840307474, + 0.007736672647297382, + 0.011461431160569191, + 0.0056604743003845215, + 0.008528471924364567, + 0.009110393933951855, + 0.005719405133277178, + 0.005900871939957142, + 0.005109064280986786, + 0.012723545543849468, + 0.01076419372111559, + 0.005134356673806906, + 0.01665733940899372, + 0.01154446229338646, + 0.012945112772285938, + 0.01034530159085989, + 0.009698386304080486, + 0.008979403413832188, + 0.010602651163935661, + 0.010580582544207573, + 0.01172979362308979, + 0.01680522784590721, + 0.01844107359647751, + 0.008536330424249172, + 0.009323203936219215, + 0.014626668766140938, + 0.007174139842391014, + 0.012572607025504112, + 0.012065200135111809, + 0.007344188634306192, + 0.008348079398274422, + 0.012861340306699276, + 0.0107529042288661, + 0.012759539298713207, + 0.008383127860724926, + 0.006521952338516712, + 0.008849423378705978, + 0.0052954968996346, + 0.0054008327424526215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0017169271595776081 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_14/index_tensor([5123, 5123])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_14/index_tensor([5123, 5123])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2368459884796a12716fe1a02aebb87ee6335bbf --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_14/index_tensor([5123, 5123])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle shot, the camera smoothly tilts downward, tracking a moving car as it prepares to make a smooth turn around a quaint Japanese street corner. It's a cream-white, vintage minivan, compact and exquisitely crafted with a meticulously maintained paint job, its warm sheen reflecting off the soft afternoon sun. Through the roof and windshield, an elderly woman can be clearly seen in the driver's seat. Her silver hair is meticulously pulled into a traditional high bun, secured with a simple wooden hairpin. She wears a dark blue silk kimono embellished with delicate white chrysanthemum patterns. Her face is kind and focused. Her hands are steady on the steering wheel, calmly and deliberately steering the car, gracefully tracing a smooth curve and precisely navigating into another narrow alley. The street where the vehicles are located is paved with irregular stone slabs, and on both sides are traditional Japanese wooden houses with dark tiled roofs and paper lattice windows. The warm sunlight casts the long shadows of the eaves and vehicles on the road, giving the entire picture a soft golden halo, creating a tranquil, nostalgic and aesthetically pleasing atmosphere.", + "seed": 29366, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_14/index_5123_seed_29366_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11016614958643914, + "vlm_logprob": -0.11016614958643914, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.11016614958643914, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42858782410621643 + ], + "opposite_logprob_tokens": [ + -1.178587794303894 + ], + "logprob_reward": -0.42858782410621643, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.42858782410621643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06829336285591125, + 0.1246388703584671, + 0.08205796033143997, + 0.04835953935980797, + 0.11129919439554214, + 0.06542297452688217, + 0.0648547038435936, + 0.3498408794403076, + 0.05075546354055405, + 0.03447704389691353 + ], + "credit_map": [ + 0.009040082804858685, + 0.007687709294259548, + 0.013477852568030357, + 0.012249570339918137, + 0.00739598274230957, + 0.004282298963516951, + 0.010114394128322601, + 0.004045471549034119, + 0.011150402016937733, + 0.02674693614244461, + 0.016175147145986557, + 0.020969005301594734, + 0.01238565519452095, + 0.013712409883737564, + 0.0102354371920228, + 0.013263877481222153, + 0.00947289913892746, + 0.00813408475369215, + 0.012166645377874374, + 0.013213230296969414, + 0.006857575383037329, + 0.009144986048340797, + 0.010573342442512512, + 0.012495206668972969, + 0.008493970148265362, + 0.004066768102347851, + 0.010015575215220451, + 0.0032853656448423862, + 0.0024779767263680696, + 0.00915490835905075, + 0.004211690276861191, + 0.0066532897762954235, + 0.02145693637430668, + 0.007539562880992889, + 0.01633881963789463, + 0.004757896531373262, + 0.0048993960954248905, + 0.020815692842006683, + 0.007976401597261429, + 0.027514494955539703, + 0.007712859660387039, + 0.008313514292240143, + 0.014249714091420174, + 0.00578303262591362, + 0.00572231225669384, + 0.006952381227165461, + 0.008016347885131836, + 0.008672814816236496, + 0.00753584410995245, + 0.00871280301362276, + 0.010324686765670776, + 0.006832404062151909, + 0.005090657621622086, + 0.011436667293310165, + 0.0062556094489991665, + 0.008666031062602997, + 0.024951230734586716, + 0.007663136348128319, + 0.13447929918766022, + 0.053869061172008514, + 0.010683024302124977, + 0.016865259036421776, + 0.015899658203125, + 0.08543024212121964, + 0.008382715284824371, + 0.007341877091675997, + 0.006909211631864309, + 0.006908099632710218, + 0.005932092200964689, + 0.004965537693351507, + 0.0051906355656683445, + 0.005125296767801046, + 0.0075101121328771114, + 0.005287372972816229, + 0.0037046431098133326, + 0.003668018849566579, + 0.004174973350018263, + 0.0030006244778633118, + 0.003556974930688739, + 0.003574325470253825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 4 + ], + "selected_reward": -0.42858782410621643 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020608583465218544 + ], + "opposite_logprob_tokens": [ + -3.89560866355896 + ], + "logprob_reward": -0.020608583465218544, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020608583465218544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04099249094724655, + 0.049652986228466034, + 0.014945796690881252, + 0.042713865637779236, + 0.031149502843618393, + 0.01830264739692211, + 0.03074781782925129, + 0.7073469758033752, + 0.04701138660311699, + 0.017136478796601295 + ], + "credit_map": [ + 0.005235857795923948, + 0.004876330960541964, + 0.006090895738452673, + 0.005076180212199688, + 0.004968790803104639, + 0.003246727865189314, + 0.006755896378308535, + 0.004741813987493515, + 0.006338920909911394, + 0.005393716506659985, + 0.00583264697343111, + 0.006520875729620457, + 0.004280753433704376, + 0.006109321024268866, + 0.00536069693043828, + 0.009816058911383152, + 0.0015558878658339381, + 0.0014180389698594809, + 0.0028058078605681658, + 0.0011863490799441934, + 0.001290906104259193, + 0.0021178072784096003, + 0.0013314560055732727, + 0.0032395441085100174, + 0.0058957659639418125, + 0.0029143390711396933, + 0.00801907293498516, + 0.0025251812767237425, + 0.002245175652205944, + 0.011372646316885948, + 0.003806962864473462, + 0.005934726446866989, + 0.004394324496388435, + 0.0029373634606599808, + 0.004906865768134594, + 0.0017750991974025965, + 0.0018022990552708507, + 0.004363036714494228, + 0.004276417661458254, + 0.006694100331515074, + 0.0018613925203680992, + 0.0023982562124729156, + 0.0034872195683419704, + 0.0020512505434453487, + 0.001374540152028203, + 0.002545864786952734, + 0.00225649937056005, + 0.0023276263382285833, + 0.004761298652738333, + 0.003310535801574588, + 0.005695023108273745, + 0.004162674769759178, + 0.0021253000013530254, + 0.003530769841745496, + 0.0030193033162504435, + 0.004142914433032274, + 0.04664327949285507, + 0.01302292849868536, + 0.2804431617259979, + 0.11473258584737778, + 0.019467514008283615, + 0.032072555273771286, + 0.030208434909582138, + 0.17075656354427338, + 0.006828129291534424, + 0.0064718821085989475, + 0.0054983897134661674, + 0.005601219367235899, + 0.006895925849676132, + 0.004989235196262598, + 0.0052896118722856045, + 0.00543699786067009, + 0.0029656144324690104, + 0.0028596767224371433, + 0.0022597527131438255, + 0.001999981701374054, + 0.002169791143387556, + 0.002003968693315983, + 0.001580011215992272, + 0.001297683804295957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 8 + ], + "selected_reward": -0.020608583465218544 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370789811015129 + ], + "opposite_logprob_tokens": [ + -3.408707857131958 + ], + "logprob_reward": -0.03370789811015129, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370789811015129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06957373768091202, + 0.03867015615105629, + 0.038481973111629486, + 0.025544874370098114, + 0.03971123322844505, + 0.015597551129758358, + 0.03522617369890213, + 0.5446322560310364, + 0.1455291211605072, + 0.047032929956912994 + ], + "credit_map": [ + 0.007911629043519497, + 0.008204386569559574, + 0.012559987604618073, + 0.011598973535001278, + 0.006168034393340349, + 0.005578682757914066, + 0.00976798590272665, + 0.007784050889313221, + 0.004900536499917507, + 0.00857685599476099, + 0.004974437411874533, + 0.006325920578092337, + 0.00379911414347589, + 0.003333895932883024, + 0.0031241104006767273, + 0.003635282162576914, + 0.005199545994400978, + 0.0037976971361786127, + 0.005598546005785465, + 0.005602772813290358, + 0.0026201712898910046, + 0.004345041234046221, + 0.0051805260591208935, + 0.006137668155133724, + 0.004437734372913837, + 0.003121463116258383, + 0.004784751683473587, + 0.0020811976864933968, + 0.0017430646112188697, + 0.003079652786254883, + 0.0033279438503086567, + 0.0029690645169466734, + 0.00624990975484252, + 0.002761497860774398, + 0.006904882378876209, + 0.0015237872721627355, + 0.0017501161200925708, + 0.010588918812572956, + 0.002340123988687992, + 0.007591990288347006, + 0.0012917923741042614, + 0.0014407711569219828, + 0.003857289208099246, + 0.0014319118345156312, + 0.0009313348564319313, + 0.0023247175849974155, + 0.0020826503168791533, + 0.0022370817605406046, + 0.003982711583375931, + 0.004759741947054863, + 0.0056474716402590275, + 0.004708545282483101, + 0.00264856219291687, + 0.005448582582175732, + 0.0039855437353253365, + 0.004045011475682259, + 0.0348261222243309, + 0.010317600332200527, + 0.21708662807941437, + 0.08843080699443817, + 0.01522806566208601, + 0.02426767535507679, + 0.023129228502511978, + 0.13134609162807465, + 0.02063942328095436, + 0.025756092742085457, + 0.01423900481313467, + 0.01652391441166401, + 0.023285645991563797, + 0.015480194240808487, + 0.01578967086970806, + 0.013815167360007763, + 0.0075456527993083, + 0.007898269221186638, + 0.007913559675216675, + 0.005046747159212828, + 0.007636437192559242, + 0.00306256883777678, + 0.004504402168095112, + 0.0034252896439284086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.03370789811015129 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029834074899554253 + ], + "opposite_logprob_tokens": [ + -3.529834032058716 + ], + "logprob_reward": -0.029834074899554253, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029834074899554253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028100157156586647, + 0.026577159762382507, + 0.022600146010518074, + 0.012684506364166737, + 0.02789868414402008, + 0.009923571720719337, + 0.017351893708109856, + 0.6776344180107117, + 0.09160727262496948, + 0.0856221541762352 + ], + "credit_map": [ + 0.0033242579083889723, + 0.003047279780730605, + 0.0037007341161370277, + 0.004733119625598192, + 0.003827630076557398, + 0.002865189453586936, + 0.0043298895470798016, + 0.0022720620036125183, + 0.003633580170571804, + 0.0037078221794217825, + 0.0029704810585826635, + 0.003668717108666897, + 0.002745859557762742, + 0.00329931965097785, + 0.002649034606292844, + 0.0039023507852107286, + 0.0026955667417496443, + 0.002268174197524786, + 0.0037284977734088898, + 0.003092524129897356, + 0.0017826950643211603, + 0.0019573115278035402, + 0.002623317064717412, + 0.0044520641677081585, + 0.001815033727325499, + 0.001009916770271957, + 0.003247389104217291, + 0.0006590220145881176, + 0.0005659195012412965, + 0.0026650778017938137, + 0.0011233717668801546, + 0.001598777249455452, + 0.0045497179962694645, + 0.001992542063817382, + 0.003189300885424018, + 0.0014650528319180012, + 0.0014255645219236612, + 0.004612805787473917, + 0.0032673170790076256, + 0.007396386936306953, + 0.0008478653035126626, + 0.000982443569228053, + 0.00162651261780411, + 0.0006527365767396986, + 0.0007886342355050147, + 0.002522594528272748, + 0.0010620965622365475, + 0.0014406908303499222, + 0.0030917145777493715, + 0.0019930494017899036, + 0.00240827864035964, + 0.002206536242738366, + 0.0014642988098785281, + 0.00244684936478734, + 0.0012359459651634097, + 0.0025052239652723074, + 0.04432867094874382, + 0.011133039370179176, + 0.26936644315719604, + 0.10884518921375275, + 0.018940379843115807, + 0.030691074207425117, + 0.02883295901119709, + 0.16549676656723022, + 0.011923532001674175, + 0.011876420117914677, + 0.010100863873958588, + 0.016387339681386948, + 0.013556377030909061, + 0.008128493092954159, + 0.009554363787174225, + 0.010079893283545971, + 0.014252899214625359, + 0.014043185859918594, + 0.013865463435649872, + 0.009051654487848282, + 0.011771593242883682, + 0.008286559954285622, + 0.00776403583586216, + 0.006586769595742226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.029834074899554253 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038092367351055145 + ], + "opposite_logprob_tokens": [ + -3.2880923748016357 + ], + "logprob_reward": -0.038092367351055145, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038092367351055145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13623549044132233, + 0.14443838596343994, + 0.06228649243712425, + 0.05499909445643425, + 0.07899399101734161, + 0.10545529425144196, + 0.03489996865391731, + 0.21133846044540405, + 0.11808796972036362, + 0.053264912217855453 + ], + "credit_map": [ + 0.013173714280128479, + 0.013919147662818432, + 0.021339938044548035, + 0.02351735346019268, + 0.014384118840098381, + 0.013052962720394135, + 0.017200808972120285, + 0.01964743249118328, + 0.01801307499408722, + 0.029006510972976685, + 0.016733648255467415, + 0.024460522457957268, + 0.014630027115345001, + 0.015187270939350128, + 0.013625295832753181, + 0.012782031670212746, + 0.008559104055166245, + 0.004988582339137793, + 0.00983163621276617, + 0.0084611177444458, + 0.0045143975876271725, + 0.008677472360432148, + 0.007663217838853598, + 0.009590962901711464, + 0.007518022786825895, + 0.003592907451093197, + 0.011984906159341335, + 0.0025525924284011126, + 0.0023372014984488487, + 0.016211941838264465, + 0.003309402149170637, + 0.007492113392800093, + 0.010763912461698055, + 0.006552880629897118, + 0.017822645604610443, + 0.003667352953925729, + 0.003474762197583914, + 0.014148294925689697, + 0.006131772417575121, + 0.016432365402579308, + 0.00998626183718443, + 0.009577332064509392, + 0.017380699515342712, + 0.005652269814163446, + 0.006898035295307636, + 0.03154916688799858, + 0.007469289004802704, + 0.016942238435149193, + 0.006003790535032749, + 0.0034214570187032223, + 0.00450840825214982, + 0.0032367052044719458, + 0.002143912948668003, + 0.007841037586331367, + 0.002562718465924263, + 0.005181935150176287, + 0.01428203471004963, + 0.003868251573294401, + 0.08250211179256439, + 0.03360355272889137, + 0.006233524531126022, + 0.010204985737800598, + 0.009244806133210659, + 0.05139917880296707, + 0.019252628087997437, + 0.013615484349429607, + 0.011614267714321613, + 0.010805312544107437, + 0.02302716299891472, + 0.01109748613089323, + 0.016039930284023285, + 0.012635687366127968, + 0.008114258758723736, + 0.009610735811293125, + 0.0072675347328186035, + 0.006375494878739119, + 0.008292839862406254, + 0.005491177551448345, + 0.0041167656891047955, + 0.003996099345386028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.038092367351055145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle shot, the camera smoothly tilts downward, tracking a moving car as it prepares to make a smooth turn around a quaint Japanese street corner. It's a cream-white, vintage minivan, compact and exquisitely crafted with a meticulously maintained paint job, its warm sheen reflecting off the soft afternoon sun. Through the roof and windshield, an elderly woman can be clearly seen in the driver's seat. Her silver hair is meticulously pulled into a traditional high bun, secured with a simple wooden hairpin. She wears a dark blue silk kimono embellished with delicate white chrysanthemum patterns. Her face is kind and focused. Her hands are steady on the steering wheel, calmly and deliberately steering the car, gracefully tracing a smooth curve and precisely navigating into another narrow alley. The street where the vehicles are located is paved with irregular stone slabs, and on both sides are traditional Japanese wooden houses with dark tiled roofs and paper lattice windows. The warm sunlight casts the long shadows of the eaves and vehicles on the road, giving the entire picture a soft golden halo, creating a tranquil, nostalgic and aesthetically pleasing atmosphere.", + "seed": 29367, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_14/index_5123_seed_29367_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.23881572633981704, + "vlm_logprob": -0.23881572633981704, + "vlm_yes_no_margin": 1.925, + "vlm_token_credit": -0.23881572633981704, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7265512347221375 + ], + "opposite_logprob_tokens": [ + -0.7265512347221375 + ], + "logprob_reward": -0.7265512347221375, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7265512347221375, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3563457131385803, + 0.03179546073079109, + 0.00928368978202343, + 0.024793358519673347, + 0.011106371879577637, + 0.03181411325931549, + 0.14902055263519287, + 0.017658034339547157, + 0.327041357755661, + 0.04114134982228279 + ], + "credit_map": [ + 0.020520837977528572, + 0.15506120026111603, + 0.006735978648066521, + 0.08962129801511765, + 0.05329571291804314, + 0.010196579620242119, + 0.013135072775185108, + 0.007779011037200689, + 0.0023398592602461576, + 0.002543320180848241, + 0.00228292727842927, + 0.008649474941194057, + 0.0016151944873854518, + 0.0017134391237050295, + 0.010085150599479675, + 0.002566094044595957, + 0.0014029120793566108, + 0.0009553206036798656, + 0.00115069595631212, + 0.0012598198372870684, + 0.0009786187438294291, + 0.0013055024901404977, + 0.0014316400047391653, + 0.0007991801830939949, + 0.0024037661496549845, + 0.00285398680716753, + 0.00273673003539443, + 0.004101595841348171, + 0.002900083549320698, + 0.0029551649931818247, + 0.0037551322020590305, + 0.003086900105699897, + 0.001326539902947843, + 0.001205633976496756, + 0.0012366381706669927, + 0.001365344855003059, + 0.0015475077088922262, + 0.0012684277025982738, + 0.002104552462697029, + 0.0010517268674448133, + 0.002689850050956011, + 0.0015832236967980862, + 0.008614631369709969, + 0.003922986797988415, + 0.002088095061480999, + 0.0022764031309634447, + 0.00813649408519268, + 0.002502431394532323, + 0.008016610518097878, + 0.006291832774877548, + 0.022563548758625984, + 0.007199197541922331, + 0.005419135093688965, + 0.05954141914844513, + 0.032184913754463196, + 0.007803888525813818, + 0.001386352232657373, + 0.0016073613660410047, + 0.004133742768317461, + 0.0022160483058542013, + 0.0013666431186720729, + 0.0020910294260829687, + 0.003293792949989438, + 0.0015630649868398905, + 0.021493053063750267, + 0.009065022692084312, + 0.11362535506486893, + 0.015572787262499332, + 0.021386969834566116, + 0.029608994722366333, + 0.0903448760509491, + 0.025944296270608902, + 0.0029635548125952482, + 0.0021850168704986572, + 0.012775624170899391, + 0.00242342846468091, + 0.004829895682632923, + 0.0032151800114661455, + 0.003050166182219982, + 0.009698483161628246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.7265512347221375 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332417666912079 + ], + "opposite_logprob_tokens": [ + -3.7733242511749268 + ], + "logprob_reward": -0.02332417666912079, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332417666912079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5234761834144592, + 0.0206881333142519, + 0.022382264956831932, + 0.016754424199461937, + 0.012880571186542511, + 0.06575693935155869, + 0.04178091883659363, + 0.01322498731315136, + 0.236247718334198, + 0.04680780693888664 + ], + "credit_map": [ + 0.030577192083001137, + 0.2233418971300125, + 0.01106566283851862, + 0.14470353722572327, + 0.06737760454416275, + 0.014672387391328812, + 0.020580805838108063, + 0.011157117784023285, + 0.0015508041251450777, + 0.0031532219145447016, + 0.0018192741554230452, + 0.0034943840000778437, + 0.0018969690427184105, + 0.0015626485692337155, + 0.005279062781482935, + 0.001931772567331791, + 0.0016810077941045165, + 0.0031201879028230906, + 0.0028756314422935247, + 0.003005400998517871, + 0.002852080622687936, + 0.0033512853551656008, + 0.0030706413090229034, + 0.0024260308127850294, + 0.001759659149684012, + 0.0016246647574007511, + 0.0020082842092961073, + 0.001577891525812447, + 0.002703377278521657, + 0.0020886543206870556, + 0.0022992833983153105, + 0.0026926102582365274, + 0.0016942806541919708, + 0.002119466196745634, + 0.0021223484072834253, + 0.001221360289491713, + 0.001132841338403523, + 0.0010443299543112516, + 0.0024605290964245796, + 0.0010854152496904135, + 0.00553800817579031, + 0.004180404357612133, + 0.012541212141513824, + 0.00838138721883297, + 0.004310861229896545, + 0.005336829461157322, + 0.01682037115097046, + 0.00864786934107542, + 0.004346426576375961, + 0.0024146686773747206, + 0.0043875230476260185, + 0.002423649886623025, + 0.0016618247609585524, + 0.006383075378835201, + 0.008032144978642464, + 0.012131609953939915, + 0.0013705422170460224, + 0.0014216122217476368, + 0.0014790582936257124, + 0.0016590681625530124, + 0.0026379521004855633, + 0.0017351864371448755, + 0.0015530773671343923, + 0.001368491561152041, + 0.015995288267731667, + 0.00663570174947381, + 0.08143265545368195, + 0.011165923438966274, + 0.01584448106586933, + 0.021191416308283806, + 0.06546667963266373, + 0.018515603616833687, + 0.0032213348895311356, + 0.0016225436702370644, + 0.016181036829948425, + 0.0023815901950001717, + 0.005286339204758406, + 0.0029560187831521034, + 0.0029037671629339457, + 0.012255178764462471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.02332417666912079 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25218531489372253 + ], + "opposite_logprob_tokens": [ + -1.502185344696045 + ], + "logprob_reward": -0.25218531489372253, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25218531489372253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10081306844949722, + 0.1243639662861824, + 0.026410553604364395, + 0.0965605303645134, + 0.03116649203002453, + 0.07465247064828873, + 0.23019182682037354, + 0.0782189667224884, + 0.06091376394033432, + 0.1767083704471588 + ], + "credit_map": [ + 0.007832389324903488, + 0.02706034667789936, + 0.00693971524015069, + 0.010590174235403538, + 0.02963758073747158, + 0.005924401339143515, + 0.0055875396355986595, + 0.007240919396281242, + 0.0051507591269910336, + 0.006644396111369133, + 0.0069205258041620255, + 0.04630633816123009, + 0.005131465848535299, + 0.004285729490220547, + 0.042661819607019424, + 0.007262927945703268, + 0.00392052112147212, + 0.002503705210983753, + 0.004317954182624817, + 0.0030873159412294626, + 0.0026226399932056665, + 0.0028769748751074076, + 0.0039596883580088615, + 0.0031217532232403755, + 0.009793905541300774, + 0.007989908568561077, + 0.012652709148824215, + 0.013304414227604866, + 0.013151192106306553, + 0.011622613295912743, + 0.012847558595240116, + 0.015198228880763054, + 0.004027361981570721, + 0.0026238271966576576, + 0.008106602355837822, + 0.003070780774578452, + 0.003218067577108741, + 0.002876805141568184, + 0.004238148685544729, + 0.0030048969201743603, + 0.007551922928541899, + 0.004701200872659683, + 0.019462106749415398, + 0.008693349547684193, + 0.0047685024328529835, + 0.005721996072679758, + 0.017393726855516434, + 0.00635967031121254, + 0.012512789107859135, + 0.01051065232604742, + 0.031148968264460564, + 0.01303141750395298, + 0.009341985918581486, + 0.08785173296928406, + 0.0539773628115654, + 0.011816916055977345, + 0.005925594363361597, + 0.00599793903529644, + 0.02206471562385559, + 0.00664530461654067, + 0.00746449688449502, + 0.009661665186285973, + 0.015421164222061634, + 0.005038085393607616, + 0.004344905726611614, + 0.0021950802765786648, + 0.01848352886736393, + 0.003183316206559539, + 0.004409782122820616, + 0.005778853315860033, + 0.0163450688123703, + 0.006173226982355118, + 0.011075906455516815, + 0.003962918184697628, + 0.06670990586280823, + 0.0075494651682674885, + 0.01808083802461624, + 0.009448951110243797, + 0.009043029509484768, + 0.05083734542131424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 1 + ], + "selected_reward": -0.25218531489372253 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07906809449195862 + ], + "opposite_logprob_tokens": [ + -2.579068183898926 + ], + "logprob_reward": -0.07906809449195862, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07906809449195862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2667384147644043, + 0.11859887838363647, + 0.047686878591775894, + 0.10116299986839294, + 0.01940922625362873, + 0.06009265035390854, + 0.0682249516248703, + 0.029627513140439987, + 0.06189284846186638, + 0.2265656441450119 + ], + "credit_map": [ + 0.018698226660490036, + 0.08112563192844391, + 0.013612241484224796, + 0.09434471279382706, + 0.019882449880242348, + 0.012544458732008934, + 0.015697922557592392, + 0.010832753963768482, + 0.00558746512979269, + 0.00782366655766964, + 0.0097526665776968, + 0.04405803233385086, + 0.004468432627618313, + 0.0046120136976242065, + 0.03492605686187744, + 0.0073705511167645454, + 0.0038876980543136597, + 0.005981089547276497, + 0.006833682302385569, + 0.00663842773064971, + 0.006286080460995436, + 0.006016711238771677, + 0.006057784892618656, + 0.005985405761748552, + 0.009516783058643341, + 0.009923926554620266, + 0.012666614726185799, + 0.013573363423347473, + 0.012716683559119701, + 0.013922430574893951, + 0.013388303108513355, + 0.015454908832907677, + 0.002204736229032278, + 0.0017621546285226941, + 0.0020262827165424824, + 0.0020618895068764687, + 0.003370103193446994, + 0.0021009340416640043, + 0.003291662083938718, + 0.002591464202851057, + 0.009901691228151321, + 0.004179816693067551, + 0.013479157350957394, + 0.006699732504785061, + 0.003674997016787529, + 0.005591167137026787, + 0.009176903404295444, + 0.007389183156192303, + 0.0061241392977535725, + 0.0034866430796682835, + 0.012100464664399624, + 0.003342095762491226, + 0.0024685084354132414, + 0.018103696405887604, + 0.007142894901335239, + 0.01545651163905859, + 0.0022425747010856867, + 0.0034480371978133917, + 0.004250950179994106, + 0.002803765470162034, + 0.008207525126636028, + 0.004117764998227358, + 0.002626907778903842, + 0.0019299871055409312, + 0.004460725001990795, + 0.002397804521024227, + 0.020102819427847862, + 0.0038186225574463606, + 0.005587549414485693, + 0.00685778446495533, + 0.013958667404949665, + 0.004708878695964813, + 0.014289837330579758, + 0.005484932567924261, + 0.08511604368686676, + 0.00980036798864603, + 0.022714080289006233, + 0.012533984147012234, + 0.01152846310287714, + 0.06509793549776077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.07906809449195862 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294981092214584 + ], + "opposite_logprob_tokens": [ + -2.237949848175049 + ], + "logprob_reward": -0.11294981092214584, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294981092214584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22490033507347107, + 0.17684397101402283, + 0.040680840611457825, + 0.10173793137073517, + 0.0143219493329525, + 0.05452686548233032, + 0.24863453209400177, + 0.057011768221855164, + 0.048203952610492706, + 0.033137813210487366 + ], + "credit_map": [ + 0.015783097594976425, + 0.08676068484783173, + 0.012477193959057331, + 0.046634264290332794, + 0.03940606489777565, + 0.006616930011659861, + 0.00941830687224865, + 0.00780380517244339, + 0.01141926646232605, + 0.015145585872232914, + 0.01149636972695589, + 0.04558195918798447, + 0.012234874069690704, + 0.011950782500207424, + 0.05407033488154411, + 0.014944798313081264, + 0.004190841224044561, + 0.006811202969402075, + 0.004998123273253441, + 0.005882889498025179, + 0.0029008418787270784, + 0.004229895304888487, + 0.007061223965138197, + 0.0046058231964707375, + 0.01087014377117157, + 0.018791615962982178, + 0.012134725227952003, + 0.011169261299073696, + 0.01274807658046484, + 0.009854587726294994, + 0.011954246088862419, + 0.014215284027159214, + 0.001748846611008048, + 0.001563186990097165, + 0.0016479409532621503, + 0.0016345129115507007, + 0.002265798393636942, + 0.0016259244875982404, + 0.0021647007670253515, + 0.001671039848588407, + 0.010986139066517353, + 0.0032086740247905254, + 0.004772154614329338, + 0.003557598916813731, + 0.00425155833363533, + 0.006175671238452196, + 0.00960790365934372, + 0.011967170052230358, + 0.017362987622618675, + 0.012561437673866749, + 0.01758170686662197, + 0.015353229828178883, + 0.010070551186800003, + 0.06969095766544342, + 0.06093761324882507, + 0.04507605731487274, + 0.0060508050955832005, + 0.005485618952661753, + 0.006756284739822149, + 0.006119315512478352, + 0.010384757071733475, + 0.007271768990904093, + 0.00835286546498537, + 0.0065903556533157825, + 0.0038654012605547905, + 0.002132842782884836, + 0.014042622409760952, + 0.0036193025298416615, + 0.005511389579623938, + 0.005631499458104372, + 0.008786709047853947, + 0.004614188335835934, + 0.002019230742007494, + 0.0011840815423056483, + 0.011417500674724579, + 0.0017813824815675616, + 0.0037562160287052393, + 0.0019885911606252193, + 0.0028185832779854536, + 0.00817223358899355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.11294981092214584 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle shot, the camera smoothly tilts downward, tracking a moving car as it prepares to make a smooth turn around a quaint Japanese street corner. It's a cream-white, vintage minivan, compact and exquisitely crafted with a meticulously maintained paint job, its warm sheen reflecting off the soft afternoon sun. Through the roof and windshield, an elderly woman can be clearly seen in the driver's seat. Her silver hair is meticulously pulled into a traditional high bun, secured with a simple wooden hairpin. She wears a dark blue silk kimono embellished with delicate white chrysanthemum patterns. Her face is kind and focused. Her hands are steady on the steering wheel, calmly and deliberately steering the car, gracefully tracing a smooth curve and precisely navigating into another narrow alley. The street where the vehicles are located is paved with irregular stone slabs, and on both sides are traditional Japanese wooden houses with dark tiled roofs and paper lattice windows. The warm sunlight casts the long shadows of the eaves and vehicles on the road, giving the entire picture a soft golden halo, creating a tranquil, nostalgic and aesthetically pleasing atmosphere.", + "seed": 29368, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_14/index_5123_seed_29368_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07371697667986155, + "vlm_logprob": -0.07371697667986155, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.07371697667986155, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17893558740615845 + ], + "opposite_logprob_tokens": [ + -2.0539355278015137 + ], + "logprob_reward": -0.17893558740615845, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.17893558740615845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08766437321901321, + 0.05364302918314934, + 0.15430425107479095, + 0.33944207429885864, + 0.08897967636585236, + 0.1502576619386673, + 0.04054351523518562, + 0.031774766743183136, + 0.031087709590792656, + 0.02230299450457096 + ], + "credit_map": [ + 0.010200661607086658, + 0.006303281523287296, + 0.007578038144856691, + 0.012671458534896374, + 0.008568755351006985, + 0.018215997144579887, + 0.011733614839613438, + 0.012392555363476276, + 0.006590442731976509, + 0.004170150496065617, + 0.007406620308756828, + 0.006851182784885168, + 0.008886544033885002, + 0.006201528944075108, + 0.006259496323764324, + 0.007277058437466621, + 0.015008688904345036, + 0.013112392276525497, + 0.016585860401391983, + 0.01777050457894802, + 0.024698399007320404, + 0.024964094161987305, + 0.0208760853856802, + 0.021288208663463593, + 0.03137429431080818, + 0.030915910378098488, + 0.02456481382250786, + 0.0786389410495758, + 0.03891070932149887, + 0.07488255947828293, + 0.02721364237368107, + 0.03294117748737335, + 0.008456182666122913, + 0.008328579366207123, + 0.013606617227196693, + 0.0132417818531394, + 0.009047836996614933, + 0.01061023585498333, + 0.016758589074015617, + 0.008929850533604622, + 0.0075112334452569485, + 0.00635979138314724, + 0.006943574175238609, + 0.046060603111982346, + 0.007764158770442009, + 0.013007602654397488, + 0.009580794721841812, + 0.053029902279376984, + 0.003976898733526468, + 0.0029933371115475893, + 0.006210368126630783, + 0.00805147085338831, + 0.0030100622680038214, + 0.006153623573482037, + 0.004544788971543312, + 0.005602962337434292, + 0.0033523801248520613, + 0.0028271309565752745, + 0.00392611650750041, + 0.003986545372754335, + 0.005084722302854061, + 0.004406422842293978, + 0.004946531727910042, + 0.003244915744289756, + 0.0035494884941726923, + 0.0025069215334951878, + 0.003621392883360386, + 0.0029205717146396637, + 0.003595861839130521, + 0.005153517238795757, + 0.005986175499856472, + 0.00375378062017262, + 0.0017317170277237892, + 0.0069128465838730335, + 0.0010761716403067112, + 0.0016726823523640633, + 0.0017123478464782238, + 0.006246793549507856, + 0.001358572393655777, + 0.0015918620629236102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 5 + ], + "selected_reward": -0.17893558740615845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009805961512029171 + ], + "opposite_logprob_tokens": [ + -4.634806156158447 + ], + "logprob_reward": -0.009805961512029171, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009805961512029171, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12175385653972626, + 0.11714662611484528, + 0.1065337061882019, + 0.2026231735944748, + 0.1090114414691925, + 0.08436569571495056, + 0.07046010345220566, + 0.040105387568473816, + 0.04857255145907402, + 0.09942744672298431 + ], + "credit_map": [ + 0.008556663990020752, + 0.010279442183673382, + 0.009133955463767052, + 0.014521745033562183, + 0.009360717609524727, + 0.027288561686873436, + 0.029691744595766068, + 0.012921029701828957, + 0.013353647664189339, + 0.01555372029542923, + 0.010219537653028965, + 0.013514071702957153, + 0.012898474000394344, + 0.016830435022711754, + 0.01616453006863594, + 0.01861220970749855, + 0.01035325974225998, + 0.008497084490954876, + 0.013236652128398418, + 0.01614759862422943, + 0.013647320680320263, + 0.01868245378136635, + 0.015099448151886463, + 0.010869884863495827, + 0.017394591122865677, + 0.011476445943117142, + 0.012771782465279102, + 0.055412717163562775, + 0.013810585252940655, + 0.06115395575761795, + 0.015778006985783577, + 0.014825083315372467, + 0.010376263409852982, + 0.009133122861385345, + 0.015886198729276657, + 0.016409602016210556, + 0.016055911779403687, + 0.014808526262640953, + 0.014471424743533134, + 0.011870394460856915, + 0.006365766283124685, + 0.006446226499974728, + 0.008266612887382507, + 0.026957914233207703, + 0.006558768916875124, + 0.016889378428459167, + 0.006852169055491686, + 0.006028864067047834, + 0.006409747060388327, + 0.009238659404218197, + 0.008596664294600487, + 0.01376357115805149, + 0.008080048486590385, + 0.00821614172309637, + 0.00716347387060523, + 0.008991796523332596, + 0.0034507354721426964, + 0.003639835864305496, + 0.004703969229012728, + 0.005637625232338905, + 0.00667866226285696, + 0.006122917402535677, + 0.005280859302729368, + 0.00459078187122941, + 0.004184043034911156, + 0.003781968727707863, + 0.006033784709870815, + 0.006760707125067711, + 0.00499294139444828, + 0.010751660913228989, + 0.006654159165918827, + 0.00541328638792038, + 0.006076092831790447, + 0.0062173111364245415, + 0.006528057623654604, + 0.02335330843925476, + 0.010941757820546627, + 0.028081852942705154, + 0.009516339749097824, + 0.00871272198855877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.009805961512029171 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296886950731277 + ], + "opposite_logprob_tokens": [ + -2.237968921661377 + ], + "logprob_reward": -0.11296886950731277, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296886950731277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13958843052387238, + 0.07817571610212326, + 0.15172886848449707, + 0.059767432510852814, + 0.04702204465866089, + 0.08137696981430054, + 0.05959216505289078, + 0.03398221358656883, + 0.2418636977672577, + 0.10690252482891083 + ], + "credit_map": [ + 0.014308111742138863, + 0.01459596585482359, + 0.012834467925131321, + 0.016623910516500473, + 0.016280528157949448, + 0.027527742087841034, + 0.01728610135614872, + 0.02013159729540348, + 0.00885279756039381, + 0.007890614680945873, + 0.0059835887514054775, + 0.007681960240006447, + 0.015211284160614014, + 0.012744227424263954, + 0.011667272076010704, + 0.008143954910337925, + 0.011922375299036503, + 0.0099171232432127, + 0.014981476590037346, + 0.030293243005871773, + 0.01457983162254095, + 0.027636736631393433, + 0.025813259184360504, + 0.01658480614423752, + 0.0074791694059967995, + 0.005350480321794748, + 0.006524207536131144, + 0.007508550304919481, + 0.007751769386231899, + 0.008969496004283428, + 0.00848916545510292, + 0.007694584783166647, + 0.004331047181040049, + 0.003698634449392557, + 0.004764339420944452, + 0.0064492919482290745, + 0.004457204136997461, + 0.01223352923989296, + 0.006003915797919035, + 0.0050840824842453, + 0.006668949034065008, + 0.004866565577685833, + 0.005457625258713961, + 0.024895964190363884, + 0.0052108303643763065, + 0.02023090422153473, + 0.006587309762835503, + 0.007458817213773727, + 0.008736307732760906, + 0.006967382039874792, + 0.008204727433621883, + 0.0069085522554814816, + 0.005569856148213148, + 0.008088449947535992, + 0.00835738331079483, + 0.00675950525328517, + 0.0033046642784029245, + 0.0028497641906142235, + 0.0025543367955833673, + 0.007250303868204355, + 0.004627097398042679, + 0.006295345723628998, + 0.0037283410783857107, + 0.003372358391061425, + 0.016006050631403923, + 0.01706046424806118, + 0.022003425285220146, + 0.02966398373246193, + 0.030911892652511597, + 0.05303402990102768, + 0.04708901420235634, + 0.026094812899827957, + 0.0057386551052331924, + 0.008998987264931202, + 0.006263685878366232, + 0.023860039189457893, + 0.008493474684655666, + 0.03892252594232559, + 0.006903208326548338, + 0.007721944712102413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.11296886950731277 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869406670331955 + ], + "opposite_logprob_tokens": [ + -3.048694133758545 + ], + "logprob_reward": -0.04869406670331955, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869406670331955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3068241775035858, + 0.12269511818885803, + 0.09714008867740631, + 0.11043798178434372, + 0.04820246994495392, + 0.0749317854642868, + 0.05673622339963913, + 0.03330576419830322, + 0.09698070585727692, + 0.052745722234249115 + ], + "credit_map": [ + 0.027023285627365112, + 0.028223270550370216, + 0.025240138173103333, + 0.0439629927277565, + 0.02619103714823723, + 0.051411278545856476, + 0.07552407681941986, + 0.029248110949993134, + 0.00919399131089449, + 0.015437202528119087, + 0.009662440977990627, + 0.016160748898983, + 0.016578376293182373, + 0.01757042668759823, + 0.020247047767043114, + 0.017844879999756813, + 0.010481485165655613, + 0.009567687287926674, + 0.009161075577139854, + 0.013395474292337894, + 0.015656596049666405, + 0.015617667697370052, + 0.012924961745738983, + 0.010335140861570835, + 0.009257479570806026, + 0.006119412835687399, + 0.00812077522277832, + 0.029008831828832626, + 0.006217068061232567, + 0.036730922758579254, + 0.009025043807923794, + 0.005958452820777893, + 0.004460741300135851, + 0.0038859103806316853, + 0.0054415324702858925, + 0.010194223374128342, + 0.004703040700405836, + 0.00623419601470232, + 0.008801278658211231, + 0.00448154890909791, + 0.005422291811555624, + 0.004639382939785719, + 0.007895023562014103, + 0.019782155752182007, + 0.0059601678512990475, + 0.018619762733578682, + 0.00764616159722209, + 0.004966835491359234, + 0.005585293285548687, + 0.0076277246698737144, + 0.006199112627655268, + 0.006014841143041849, + 0.007524699438363314, + 0.007774734869599342, + 0.0075591010972857475, + 0.008450711145997047, + 0.0038514109328389168, + 0.002823858056217432, + 0.0031022008042782545, + 0.005395341198891401, + 0.004268583841621876, + 0.004661260638386011, + 0.00583875784650445, + 0.0033643515780568123, + 0.006656075362116098, + 0.00686399033293128, + 0.00863662175834179, + 0.01038992591202259, + 0.01233583502471447, + 0.02260645665228367, + 0.01784999668598175, + 0.011641808785498142, + 0.00417875312268734, + 0.009280719794332981, + 0.0028670423198491335, + 0.005758375860750675, + 0.004078625235706568, + 0.01701272837817669, + 0.004478715360164642, + 0.0050907619297504425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.04869406670331955 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180398270487785 + ], + "opposite_logprob_tokens": [ + -4.0181803703308105 + ], + "logprob_reward": -0.018180398270487785, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180398270487785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14403128623962402, + 0.09025335311889648, + 0.10628988593816757, + 0.062030404806137085, + 0.14334701001644135, + 0.15558172762393951, + 0.12335827946662903, + 0.08434993028640747, + 0.03826961666345596, + 0.05248852074146271 + ], + "credit_map": [ + 0.01527844276279211, + 0.014224132522940636, + 0.01140731479972601, + 0.011287163943052292, + 0.012807969003915787, + 0.03537558764219284, + 0.026425838470458984, + 0.017224831506609917, + 0.00740883918479085, + 0.017176860943436623, + 0.004717475734651089, + 0.007816988974809647, + 0.009505200199782848, + 0.0139713604003191, + 0.008564691990613937, + 0.021091928705573082, + 0.007605742663145065, + 0.008197866380214691, + 0.010405711829662323, + 0.015431562438607216, + 0.01567346602678299, + 0.01791282370686531, + 0.016238126903772354, + 0.014824592508375645, + 0.007554042153060436, + 0.005333945155143738, + 0.00488943001255393, + 0.01294540986418724, + 0.007784867659211159, + 0.011393454857170582, + 0.005300256423652172, + 0.006828999612480402, + 0.013402833603322506, + 0.012322881259024143, + 0.01989704743027687, + 0.018297772854566574, + 0.018106579780578613, + 0.01942308247089386, + 0.026905430480837822, + 0.014991376549005508, + 0.008300918154418468, + 0.007314620073884726, + 0.007207488175481558, + 0.056360144168138504, + 0.005813282448798418, + 0.01917959377169609, + 0.009057722985744476, + 0.04234795272350311, + 0.010417800396680832, + 0.012909598648548126, + 0.01353888213634491, + 0.01722068153321743, + 0.013951984234154224, + 0.02126269042491913, + 0.018524277955293655, + 0.015532364137470722, + 0.007836766541004181, + 0.00689662666991353, + 0.006732159294188023, + 0.01013000588864088, + 0.010946521535515785, + 0.018679393455386162, + 0.009779345244169235, + 0.01334910374134779, + 0.0029677983839064837, + 0.003279665717855096, + 0.004081190098077059, + 0.004060075618326664, + 0.004885676316916943, + 0.006907923612743616, + 0.007185941096395254, + 0.0049013481475412846, + 0.00421035336330533, + 0.008385153487324715, + 0.004042952787131071, + 0.004853660240769386, + 0.005421370733529329, + 0.014975521713495255, + 0.004559372551739216, + 0.00604013679549098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.018180398270487785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle shot, the camera smoothly tilts downward, tracking a moving car as it prepares to make a smooth turn around a quaint Japanese street corner. It's a cream-white, vintage minivan, compact and exquisitely crafted with a meticulously maintained paint job, its warm sheen reflecting off the soft afternoon sun. Through the roof and windshield, an elderly woman can be clearly seen in the driver's seat. Her silver hair is meticulously pulled into a traditional high bun, secured with a simple wooden hairpin. She wears a dark blue silk kimono embellished with delicate white chrysanthemum patterns. Her face is kind and focused. Her hands are steady on the steering wheel, calmly and deliberately steering the car, gracefully tracing a smooth curve and precisely navigating into another narrow alley. The street where the vehicles are located is paved with irregular stone slabs, and on both sides are traditional Japanese wooden houses with dark tiled roofs and paper lattice windows. The warm sunlight casts the long shadows of the eaves and vehicles on the road, giving the entire picture a soft golden halo, creating a tranquil, nostalgic and aesthetically pleasing atmosphere.", + "seed": 29369, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_14/index_5123_seed_29369_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1392065592110157, + "vlm_logprob": -0.1392065592110157, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.1392065592110157, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4650939702987671 + ], + "opposite_logprob_tokens": [ + -1.090093970298767 + ], + "logprob_reward": -0.4650939702987671, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4650939702987671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09936397522687912, + 0.11524570733308792, + 0.15177534520626068, + 0.07658087462186813, + 0.08652374148368835, + 0.15069077908992767, + 0.07865656167268753, + 0.06138763576745987, + 0.13202795386314392, + 0.04774738475680351 + ], + "credit_map": [ + 0.011219925247132778, + 0.011397307738661766, + 0.010690771043300629, + 0.0105813667178154, + 0.012698793783783913, + 0.012451976537704468, + 0.010976244695484638, + 0.01934758760035038, + 0.010880145244300365, + 0.013230754062533379, + 0.010735840536653996, + 0.027461247518658638, + 0.007916848175227642, + 0.011843685060739517, + 0.01397407241165638, + 0.019203105941414833, + 0.011848374269902706, + 0.012338551692664623, + 0.005570076406002045, + 0.01391294039785862, + 0.05412941426038742, + 0.011311867274343967, + 0.006843725685030222, + 0.035820405930280685, + 0.010141439735889435, + 0.01041233167052269, + 0.008468285202980042, + 0.008167923428118229, + 0.019537266343832016, + 0.0064401160925626755, + 0.005654691252857447, + 0.007758823689073324, + 0.013964437879621983, + 0.017144666984677315, + 0.0062140123918652534, + 0.008189214393496513, + 0.011869833804666996, + 0.01204235665500164, + 0.007238728925585747, + 0.009860491380095482, + 0.009356752969324589, + 0.007680254057049751, + 0.005133498460054398, + 0.04952624812722206, + 0.056267328560352325, + 0.012083995155990124, + 0.00622631935402751, + 0.0044163903221488, + 0.006586262956261635, + 0.00625250767916441, + 0.0041680773720145226, + 0.016441363841295242, + 0.023007037118077278, + 0.005344551056623459, + 0.011527975089848042, + 0.005328786093741655, + 0.006854910869151354, + 0.006943310145288706, + 0.005726505536586046, + 0.01235860213637352, + 0.0054202633909881115, + 0.006057801656424999, + 0.012799384072422981, + 0.005226858425885439, + 0.013270637020468712, + 0.013571708463132381, + 0.018489226698875427, + 0.01526717096567154, + 0.01889575459063053, + 0.012135379947721958, + 0.026167355477809906, + 0.014230716973543167, + 0.004822415299713612, + 0.004244224168360233, + 0.006500819232314825, + 0.007207355462014675, + 0.004538615234196186, + 0.009725570678710938, + 0.0039916993118822575, + 0.006716681644320488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 8 + ], + "selected_reward": -0.4650939702987671 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020620260387659073 + ], + "opposite_logprob_tokens": [ + -3.895620346069336 + ], + "logprob_reward": -0.020620260387659073, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020620260387659073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1452876627445221, + 0.08243776857852936, + 0.16787123680114746, + 0.07919943332672119, + 0.04027928039431572, + 0.17752236127853394, + 0.045366451144218445, + 0.14329683780670166, + 0.10533316433429718, + 0.013405807316303253 + ], + "credit_map": [ + 0.012620946392416954, + 0.01563398912549019, + 0.022682949900627136, + 0.02351737581193447, + 0.018066564574837685, + 0.01570776104927063, + 0.017817765474319458, + 0.019240308552980423, + 0.007922778837382793, + 0.010850546881556511, + 0.010277478955686092, + 0.012531071901321411, + 0.010407964698970318, + 0.009850672446191311, + 0.010148448869585991, + 0.010448808781802654, + 0.011872845701873302, + 0.013083375990390778, + 0.00566191878169775, + 0.00917565543204546, + 0.06441044062376022, + 0.012071113102138042, + 0.008655465207993984, + 0.042940426617860794, + 0.007897548377513885, + 0.012580251321196556, + 0.007723547983914614, + 0.01225578784942627, + 0.010171659290790558, + 0.0068392157554626465, + 0.013542722910642624, + 0.0081887012347579, + 0.005031182896345854, + 0.007412020117044449, + 0.0027430285699665546, + 0.0046867914497852325, + 0.0058753942139446735, + 0.005245610605925322, + 0.0053571611642837524, + 0.003928090445697308, + 0.010566993616521358, + 0.011433923617005348, + 0.0075195664539933205, + 0.05604587867856026, + 0.0630444809794426, + 0.015493559651076794, + 0.007109658792614937, + 0.006308295298367739, + 0.005691182799637318, + 0.006573131773620844, + 0.0067422171123325825, + 0.004121045581996441, + 0.005149297881871462, + 0.003567434847354889, + 0.009443812072277069, + 0.004078331403434277, + 0.012430483475327492, + 0.014856740832328796, + 0.013556837104260921, + 0.024949660524725914, + 0.016223153099417686, + 0.018754059448838234, + 0.026807203888893127, + 0.01571868546307087, + 0.00969772133976221, + 0.0091711962595582, + 0.009651909582316875, + 0.010435638949275017, + 0.019438689574599266, + 0.010955394245684147, + 0.021511804312467575, + 0.014470808207988739, + 0.0013174640480428934, + 0.001453522127121687, + 0.002183809643611312, + 0.001196285244077444, + 0.0019238563254475594, + 0.0019716168753802776, + 0.0013284272281453013, + 0.0020308252424001694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.020620260387659073 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297482997179031 + ], + "opposite_logprob_tokens": [ + -2.2379748821258545 + ], + "logprob_reward": -0.11297482997179031, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297482997179031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09326880425214767, + 0.07285534590482712, + 0.051886171102523804, + 0.09942843019962311, + 0.03518584370613098, + 0.30826810002326965, + 0.06409724801778793, + 0.057962849736213684, + 0.1639643907546997, + 0.053082771599292755 + ], + "credit_map": [ + 0.00826835073530674, + 0.006068650633096695, + 0.01879405789077282, + 0.012337169609963894, + 0.012365661561489105, + 0.010461158119142056, + 0.013223814778029919, + 0.011749938130378723, + 0.007853160612285137, + 0.009886052459478378, + 0.006700581405311823, + 0.01326733361929655, + 0.00912307109683752, + 0.008058172650635242, + 0.008292642422020435, + 0.009674330241978168, + 0.0032979489769786596, + 0.004311685916036367, + 0.002217900939285755, + 0.0035665403120219707, + 0.01770976558327675, + 0.004429406486451626, + 0.0026540523394942284, + 0.013698870316147804, + 0.01510019600391388, + 0.011268412694334984, + 0.012867330573499203, + 0.015136584639549255, + 0.016916759312152863, + 0.007297439035028219, + 0.010278847999870777, + 0.010562858544290066, + 0.00404131505638361, + 0.005295435898005962, + 0.003926195669919252, + 0.004374624229967594, + 0.0044601489789783955, + 0.004396805074065924, + 0.0045240516774356365, + 0.004167267121374607, + 0.019942712038755417, + 0.01857847347855568, + 0.011565961875021458, + 0.09505771100521088, + 0.11197233200073242, + 0.029348429292440414, + 0.011977537535130978, + 0.009824930690228939, + 0.007501615211367607, + 0.008410160429775715, + 0.01088450662791729, + 0.007867798209190369, + 0.007875119335949421, + 0.0057810526341199875, + 0.010274004191160202, + 0.005502997897565365, + 0.006135638803243637, + 0.005077616777271032, + 0.004217240493744612, + 0.005355192814022303, + 0.005641436204314232, + 0.011780754663050175, + 0.00858310330659151, + 0.011171868070960045, + 0.015851644799113274, + 0.014588721096515656, + 0.015061781741678715, + 0.016360590234398842, + 0.03716132417321205, + 0.018346069380640984, + 0.025683464482426643, + 0.020910784602165222, + 0.005034424364566803, + 0.004287436604499817, + 0.006683222483843565, + 0.0038037006743252277, + 0.007271592039614916, + 0.00950677040964365, + 0.011983836069703102, + 0.0045117903500795364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.11297482997179031 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869383946061134 + ], + "opposite_logprob_tokens": [ + -3.048693895339966 + ], + "logprob_reward": -0.04869383946061134, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869383946061134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06872717291116714, + 0.04788808524608612, + 0.019351240247488022, + 0.0662747174501419, + 0.01636834628880024, + 0.54293292760849, + 0.04588525369763374, + 0.06073745712637901, + 0.09195292741060257, + 0.03988182917237282 + ], + "credit_map": [ + 0.006419327575713396, + 0.006211338564753532, + 0.011948127299547195, + 0.01618327759206295, + 0.0056129381991922855, + 0.007705407682806253, + 0.007472184952348471, + 0.007174584548920393, + 0.005150218494236469, + 0.006985895801335573, + 0.003991138655692339, + 0.008448810316622257, + 0.005624487996101379, + 0.004380633123219013, + 0.00506387185305357, + 0.008243032731115818, + 0.0018621019553393126, + 0.0023778032045811415, + 0.001934983883984387, + 0.0025068917311728, + 0.002671971684321761, + 0.0033181849867105484, + 0.0022153244353830814, + 0.0024639791809022427, + 0.006485209800302982, + 0.012379588559269905, + 0.005897859577089548, + 0.008717844262719154, + 0.008954840712249279, + 0.00458437018096447, + 0.012213469482958317, + 0.007041549310088158, + 0.0027627001982182264, + 0.0033355194609612226, + 0.0012168858665972948, + 0.0016864833887666464, + 0.002232067519798875, + 0.0023343288339674473, + 0.00141230970621109, + 0.0013880525948479772, + 0.03214816376566887, + 0.033504437655210495, + 0.02220284752547741, + 0.17289848625659943, + 0.19529449939727783, + 0.046987734735012054, + 0.02145802043378353, + 0.018438804894685745, + 0.0034640931989997625, + 0.0033732939045876265, + 0.004978840239346027, + 0.009614178910851479, + 0.012762876227498055, + 0.002588198985904455, + 0.006421337369829416, + 0.0026824367232620716, + 0.004939757287502289, + 0.005521683022379875, + 0.0043525369837880135, + 0.009879054501652718, + 0.006430316716432571, + 0.009275669232010841, + 0.013316328637301922, + 0.007022124715149403, + 0.00817063543945551, + 0.008042502216994762, + 0.007785750553011894, + 0.00963285006582737, + 0.012797541916370392, + 0.00964709185063839, + 0.01754015125334263, + 0.01833641156554222, + 0.0033281431533396244, + 0.003472225973382592, + 0.0032410959247499704, + 0.0042821625247597694, + 0.008963670581579208, + 0.007677275687456131, + 0.004461122211068869, + 0.004456136841326952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.04869383946061134 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle and tilting downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a cream-white vintage minivan with a meticulously maintained paint job?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elderly woman with silver hair in a traditional high bun visible driving the vehicle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the driver wearing a dark blue silk kimono with white chrysanthemum patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a Japanese street corner with irregular stone slabs and traditional wooden houses?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864989593625069 + ], + "opposite_logprob_tokens": [ + -3.048649787902832 + ], + "logprob_reward": -0.04864989593625069, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864989593625069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2288568615913391, + 0.04059692844748497, + 0.0793677344918251, + 0.12750403583049774, + 0.06075765937566757, + 0.24090202152729034, + 0.09929709136486053, + 0.044195227324962616, + 0.03386928141117096, + 0.04465320333838463 + ], + "credit_map": [ + 0.023180341348052025, + 0.026610827073454857, + 0.031045541167259216, + 0.030587198212742805, + 0.02822585217654705, + 0.021292945370078087, + 0.02821102924644947, + 0.0397031232714653, + 0.0038180560804903507, + 0.002719502430409193, + 0.00426595751196146, + 0.0068137142807245255, + 0.006315466016530991, + 0.003984570503234863, + 0.00441866647452116, + 0.008260990492999554, + 0.006880523636937141, + 0.015425875782966614, + 0.005705398041754961, + 0.006045839283615351, + 0.013529323972761631, + 0.01774616353213787, + 0.005205747205764055, + 0.008828854188323021, + 0.018507640808820724, + 0.02346814051270485, + 0.01899581402540207, + 0.016661835834383965, + 0.01901230216026306, + 0.01136102806776762, + 0.007886849343776703, + 0.011610417626798153, + 0.00855104997754097, + 0.009638485498726368, + 0.005172901786863804, + 0.006258911918848753, + 0.0075616901740431786, + 0.007245352957397699, + 0.010880311019718647, + 0.005448951385915279, + 0.014073352329432964, + 0.012216977775096893, + 0.009295601397752762, + 0.0804695188999176, + 0.08784738928079605, + 0.017992863431572914, + 0.010999958962202072, + 0.008006337098777294, + 0.006260716821998358, + 0.007401642389595509, + 0.0074103763327002525, + 0.030195439234375954, + 0.034045711159706116, + 0.004843176808208227, + 0.006215586327016354, + 0.002924437401816249, + 0.004373946692794561, + 0.003627584781497717, + 0.0051270220428705215, + 0.00657147541642189, + 0.005336000584065914, + 0.0054700011387467384, + 0.0077644167467951775, + 0.005924778524786234, + 0.0029448620043694973, + 0.002854786580428481, + 0.0031339158304035664, + 0.004828430712223053, + 0.005329940002411604, + 0.003551406553015113, + 0.003891837317496538, + 0.007334098219871521, + 0.004780764691531658, + 0.0035920836962759495, + 0.007032894063740969, + 0.004977464210242033, + 0.006056732032448053, + 0.005355465691536665, + 0.005305692553520203, + 0.007552101742476225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.04864989593625069 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_15/index_tensor([801, 801])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_15/index_tensor([801, 801])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..faca5f1ab23d3e940099f729c74dd76338c99a55 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_15/index_tensor([801, 801])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fixed, top-down, bird's-eye view immerses the viewer in a classic pixel art world, its visuals reminiscent of the tense atmosphere of a vintage video game. At the center of the frame, a simple, square life raft, constructed from bright orange blocks, struggles amidst a vast ocean of dark blue pixels. Massive waves, procedurally composed of alternating dark navy and slightly lighter cyan blocks, create a dynamic sense of depth, their crests dotted with striking foam composed of pure white pixels. Rather than a smooth, fluid motion, the life raft is violently tossed up and down in this digital turmoil, jerking and jerking with each wave, pushed to the corners of the frame and pulled back toward the center. The scene utilizes a flat, shadowless lighting scheme and a limited, saturated color palette. The strong color contrast and blocky dynamics create a tense atmosphere of isolation and life-or-death struggle, perfectly recreating the unique aesthetic and pressure of survival games from the 8-bit era.", + "seed": 16661, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_15/index_801_seed_16661_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07071596058085561, + "vlm_logprob": -0.07071596058085561, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.07071596058085561, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28725337982177734 + ], + "opposite_logprob_tokens": [ + -1.4122533798217773 + ], + "logprob_reward": -0.28725337982177734, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28725337982177734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12169890850782394, + 0.13816848397254944, + 0.08460213243961334, + 0.04486611485481262, + 0.073394775390625, + 0.04473352059721947, + 0.29901057481765747, + 0.05365218594670296, + 0.05109105631709099, + 0.08878228068351746 + ], + "credit_map": [ + 0.015460495837032795, + 0.01570701040327549, + 0.011900640092790127, + 0.01570095121860504, + 0.02334360033273697, + 0.010109024122357368, + 0.01765129528939724, + 0.011825893074274063, + 0.01085758302360773, + 0.044943057000637054, + 0.005630032159388065, + 0.01048178132623434, + 0.009294688701629639, + 0.006789257749915123, + 0.007448253221809864, + 0.04272382706403732, + 0.011410738341510296, + 0.014156045392155647, + 0.006270392332226038, + 0.006955837830901146, + 0.01500299759209156, + 0.006785227917134762, + 0.012534240260720253, + 0.011486654169857502, + 0.0030690422281622887, + 0.0065668183378875256, + 0.0048052906058728695, + 0.00502903014421463, + 0.00888129323720932, + 0.005461698863655329, + 0.005185141693800688, + 0.005867800209671259, + 0.00931472796946764, + 0.007692388258874416, + 0.010697104036808014, + 0.008760149590671062, + 0.011507732793688774, + 0.009296130388975143, + 0.005683154799044132, + 0.010443384759128094, + 0.0024643144570291042, + 0.0044166152365505695, + 0.00618354557082057, + 0.00500353891402483, + 0.009701347909867764, + 0.005590385757386684, + 0.004931735806167126, + 0.006442037411034107, + 0.041081853210926056, + 0.020308854058384895, + 0.06132538989186287, + 0.018581148236989975, + 0.029537012800574303, + 0.08590148389339447, + 0.01093877013772726, + 0.031336039304733276, + 0.003965028095990419, + 0.006283735390752554, + 0.00845499150454998, + 0.0057560354471206665, + 0.012494211085140705, + 0.00727506447583437, + 0.00412228936329484, + 0.005300828255712986, + 0.004905095789581537, + 0.005929320584982634, + 0.003919143229722977, + 0.008086332120001316, + 0.00751894898712635, + 0.007903399877250195, + 0.008640565909445286, + 0.00418825214728713, + 0.004320359323173761, + 0.032359592616558075, + 0.0030803768895566463, + 0.004752911161631346, + 0.006785191595554352, + 0.030251577496528625, + 0.0033851535990834236, + 0.0038471214938908815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.28725337982177734 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038128167390823364 + ], + "opposite_logprob_tokens": [ + -3.288128137588501 + ], + "logprob_reward": -0.038128167390823364, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038128167390823364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3402930796146393, + 0.11596240103244781, + 0.04704274982213974, + 0.0715223103761673, + 0.0802244022488594, + 0.05041880160570145, + 0.16335858404636383, + 0.04407127574086189, + 0.03356356546282768, + 0.05354287475347519 + ], + "credit_map": [ + 0.04035943001508713, + 0.0405307337641716, + 0.03650889918208122, + 0.05009550601243973, + 0.0562344454228878, + 0.03029460646212101, + 0.04982055723667145, + 0.036448884755373, + 0.008760610595345497, + 0.032574161887168884, + 0.005796298384666443, + 0.01610937900841236, + 0.007997400127351284, + 0.004873465280979872, + 0.007858755066990852, + 0.03199232369661331, + 0.004871852695941925, + 0.005533813964575529, + 0.0032722370233386755, + 0.007437658030539751, + 0.007664500270038843, + 0.0032818655017763376, + 0.008410619571805, + 0.006570200435817242, + 0.0053649116307497025, + 0.009308608248829842, + 0.0050470163114368916, + 0.011371816508471966, + 0.013290503062307835, + 0.005006373394280672, + 0.013413551263511181, + 0.008719531819224358, + 0.006501088384538889, + 0.009794841520488262, + 0.009762526489794254, + 0.014969146810472012, + 0.010871640406548977, + 0.005570674780756235, + 0.013272248208522797, + 0.009482234716415405, + 0.0032186342868953943, + 0.003635522909462452, + 0.007965445518493652, + 0.005615300498902798, + 0.015253017656505108, + 0.006314410362392664, + 0.005755799822509289, + 0.002660674974322319, + 0.02026290073990822, + 0.01561836339533329, + 0.017675751820206642, + 0.019125670194625854, + 0.026833118870854378, + 0.029007121920585632, + 0.020852837711572647, + 0.013982822187244892, + 0.003407259937375784, + 0.005169943440705538, + 0.006597194354981184, + 0.004724040627479553, + 0.010062193498015404, + 0.006592694669961929, + 0.003262400161474943, + 0.0042555504478514194, + 0.002355435164645314, + 0.003267069114372134, + 0.004144035745412111, + 0.005056107882410288, + 0.004633291158825159, + 0.00492491340264678, + 0.007176963612437248, + 0.0020057461224496365, + 0.0025076023302972317, + 0.01897498406469822, + 0.001912852399982512, + 0.0030308447312563658, + 0.003750507952645421, + 0.01824973337352276, + 0.0032917470671236515, + 0.001824603765271604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.038128167390823364 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093601368367672 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.004093601368367672, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093601368367672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16880804300308228, + 0.26422932744026184, + 0.04779297858476639, + 0.08809495717287064, + 0.05400855839252472, + 0.040158648043870926, + 0.19841672480106354, + 0.03198253735899925, + 0.021566061303019524, + 0.08494208753108978 + ], + "credit_map": [ + 0.020243791863322258, + 0.02474820986390114, + 0.014934828504920006, + 0.020222952589392662, + 0.028519315645098686, + 0.012140580452978611, + 0.022136932238936424, + 0.025861423462629318, + 0.01717323437333107, + 0.09160961210727692, + 0.008169217966496944, + 0.024981942027807236, + 0.014517330564558506, + 0.012505467049777508, + 0.012380007654428482, + 0.0828925222158432, + 0.007028726860880852, + 0.009522581472992897, + 0.004118342883884907, + 0.004128200002014637, + 0.008298060856759548, + 0.004858818370848894, + 0.0036634602583944798, + 0.006174786947667599, + 0.008338390849530697, + 0.014390405267477036, + 0.007230239454656839, + 0.008318085223436356, + 0.018122227862477303, + 0.010313428938388824, + 0.00827998761087656, + 0.013102193363010883, + 0.006846773438155651, + 0.0070932358503341675, + 0.006781465373933315, + 0.004816548433154821, + 0.009251657873392105, + 0.005709952209144831, + 0.00384781276807189, + 0.009661110118031502, + 0.002020568121224642, + 0.003595008049160242, + 0.006012126803398132, + 0.004232070874422789, + 0.009701213799417019, + 0.003568527288734913, + 0.005562442820519209, + 0.005466691683977842, + 0.015616604126989841, + 0.010068146511912346, + 0.04175736755132675, + 0.01853344216942787, + 0.020752692595124245, + 0.05000346153974533, + 0.024372490122914314, + 0.017312537878751755, + 0.002374225528910756, + 0.0032522568944841623, + 0.004503208678215742, + 0.003648622427135706, + 0.006637943908572197, + 0.004408820997923613, + 0.0029335778672248125, + 0.004223882686346769, + 0.0013171510072425008, + 0.0029378158506006002, + 0.0028371030930429697, + 0.0022456541191786528, + 0.0030548498034477234, + 0.003504382912069559, + 0.002420986769720912, + 0.003248116932809353, + 0.004053667653352022, + 0.03106841631233692, + 0.0020457005593925714, + 0.004105694591999054, + 0.005933745298534632, + 0.03208113834261894, + 0.002709324238821864, + 0.002944404259324074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.004093601368367672 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329447865486145 + ], + "opposite_logprob_tokens": [ + -3.773294448852539 + ], + "logprob_reward": -0.02329447865486145, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329447865486145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12726077437400818, + 0.17711853981018066, + 0.04145277664065361, + 0.10706491768360138, + 0.06302527338266373, + 0.07442013174295425, + 0.30834364891052246, + 0.01725165545940399, + 0.04158351570367813, + 0.0424787737429142 + ], + "credit_map": [ + 0.017175450921058655, + 0.016461586579680443, + 0.012116559781134129, + 0.016666041687130928, + 0.019499054178595543, + 0.015331840142607689, + 0.01663840003311634, + 0.013371855951845646, + 0.011787453666329384, + 0.045794855803251266, + 0.012099451385438442, + 0.03011617250740528, + 0.01341270562261343, + 0.009680033661425114, + 0.011411115527153015, + 0.04281673580408096, + 0.005548021290451288, + 0.006325280759483576, + 0.0035241669975221157, + 0.006302410736680031, + 0.0041519529186189175, + 0.004489172715693712, + 0.006552122533321381, + 0.004559648688882589, + 0.00904683768749237, + 0.016769615933299065, + 0.006721207406371832, + 0.011227918788790703, + 0.02564237266778946, + 0.008884995244443417, + 0.01588878594338894, + 0.012883187271654606, + 0.005420922767370939, + 0.010125862434506416, + 0.009349962696433067, + 0.0077421520836651325, + 0.008720485493540764, + 0.005739911459386349, + 0.007829410023987293, + 0.008096573874354362, + 0.005312558729201555, + 0.007293837144970894, + 0.010403754189610481, + 0.007239221129566431, + 0.014077865518629551, + 0.008847184479236603, + 0.00713301869109273, + 0.014112688601016998, + 0.02827356569468975, + 0.016574906185269356, + 0.0659705325961113, + 0.02280113846063614, + 0.03642481565475464, + 0.08111860603094101, + 0.030034776777029037, + 0.027145294472575188, + 0.0012073371326550841, + 0.0013327018823474646, + 0.0038962282706052065, + 0.0016900948248803616, + 0.002862613182514906, + 0.003642427735030651, + 0.0012738930527120829, + 0.0013463598443195224, + 0.0032165718730539083, + 0.0037454008124768734, + 0.006428942549973726, + 0.006069935858249664, + 0.005580060184001923, + 0.0072265383787453175, + 0.00729604996740818, + 0.0020200139842927456, + 0.0032431951258331537, + 0.015779471024870872, + 0.0019359663128852844, + 0.0034613164607435465, + 0.004445247817784548, + 0.009512466378509998, + 0.0021412940695881844, + 0.001959814690053463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.02329447865486145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008101756684482098 + ], + "opposite_logprob_tokens": [ + -7.125810146331787 + ], + "logprob_reward": -0.0008101756684482098, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008101756684482098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23844827711582184, + 0.07881910353899002, + 0.068984754383564, + 0.06626244634389877, + 0.07618693262338638, + 0.04606391116976738, + 0.246717169880867, + 0.04754079878330231, + 0.05891001224517822, + 0.07206661254167557 + ], + "credit_map": [ + 0.03130722790956497, + 0.029668811708688736, + 0.02394687570631504, + 0.02943028323352337, + 0.04070238396525383, + 0.019707340747117996, + 0.03418435901403427, + 0.02950100786983967, + 0.007297108415514231, + 0.02295854687690735, + 0.006129261571913958, + 0.007252572104334831, + 0.00558583065867424, + 0.003979178611189127, + 0.004499789793044329, + 0.021116817370057106, + 0.010119467973709106, + 0.011564699001610279, + 0.0057081785053014755, + 0.005890061147511005, + 0.011750245466828346, + 0.005460275802761316, + 0.008680262602865696, + 0.009811567142605782, + 0.006220942363142967, + 0.009089624509215355, + 0.0065685962326824665, + 0.008263539522886276, + 0.015343141742050648, + 0.005311235785484314, + 0.009011771529912949, + 0.0064535909332334995, + 0.008531346917152405, + 0.00874046329408884, + 0.01046464778482914, + 0.009838792495429516, + 0.01182143110781908, + 0.008172495290637016, + 0.006683314684778452, + 0.011934446170926094, + 0.002346326131373644, + 0.004666255321353674, + 0.007401301525533199, + 0.0054693324491381645, + 0.011461809277534485, + 0.004127912223339081, + 0.005623007193207741, + 0.004967967048287392, + 0.014800912700593472, + 0.014027389697730541, + 0.05044344440102577, + 0.027559000998735428, + 0.029459450393915176, + 0.055425360798835754, + 0.03717857226729393, + 0.01782304234802723, + 0.0035181876737624407, + 0.004905728157609701, + 0.00728763360530138, + 0.005421478766947985, + 0.00954635813832283, + 0.006459204480051994, + 0.005081465933471918, + 0.005320738069713116, + 0.0047101727686822414, + 0.006103305611759424, + 0.004743881057947874, + 0.009832791052758694, + 0.008262407034635544, + 0.009012691676616669, + 0.011917833238840103, + 0.00432693213224411, + 0.004011699929833412, + 0.01858457364141941, + 0.003665757831186056, + 0.006079109385609627, + 0.0065102106891572475, + 0.024994000792503357, + 0.003651815000921488, + 0.00456944340839982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0008101756684482098 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed, top-down, bird's-eye view immerses the viewer in a classic pixel art world, its visuals reminiscent of the tense atmosphere of a vintage video game. At the center of the frame, a simple, square life raft, constructed from bright orange blocks, struggles amidst a vast ocean of dark blue pixels. Massive waves, procedurally composed of alternating dark navy and slightly lighter cyan blocks, create a dynamic sense of depth, their crests dotted with striking foam composed of pure white pixels. Rather than a smooth, fluid motion, the life raft is violently tossed up and down in this digital turmoil, jerking and jerking with each wave, pushed to the corners of the frame and pulled back toward the center. The scene utilizes a flat, shadowless lighting scheme and a limited, saturated color palette. The strong color contrast and blocky dynamics create a tense atmosphere of isolation and life-or-death struggle, perfectly recreating the unique aesthetic and pressure of survival games from the 8-bit era.", + "seed": 16662, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_15/index_801_seed_16662_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.008804717904422433, + "vlm_logprob": -0.008804717904422433, + "vlm_yes_no_margin": 5.525, + "vlm_token_credit": -0.008804717904422433, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02648574858903885 + ], + "opposite_logprob_tokens": [ + -3.7764856815338135 + ], + "logprob_reward": -0.02648574858903885, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02648574858903885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20357394218444824, + 0.05349680036306381, + 0.08947106450796127, + 0.12210183590650558, + 0.04487046226859093, + 0.06752313673496246, + 0.08239854872226715, + 0.0885484591126442, + 0.1340518295764923, + 0.11396395415067673 + ], + "credit_map": [ + 0.01462548691779375, + 0.012678087688982487, + 0.03176365792751312, + 0.017228731885552406, + 0.043892309069633484, + 0.03591553866863251, + 0.024908337742090225, + 0.022561799734830856, + 0.004969432950019836, + 0.014542599208652973, + 0.00536440359428525, + 0.005455725360661745, + 0.0083969971165061, + 0.005442312452942133, + 0.005041820462793112, + 0.0042835115455091, + 0.019282778725028038, + 0.008477064780890942, + 0.007609558291733265, + 0.006729431916028261, + 0.01205639448016882, + 0.019236082211136818, + 0.008796428330242634, + 0.007283318322151899, + 0.013865197077393532, + 0.010738333687186241, + 0.012837549671530724, + 0.022572683170437813, + 0.016262097284197807, + 0.016440140083432198, + 0.01658538542687893, + 0.012800449505448341, + 0.0046628862619400024, + 0.003973677754402161, + 0.007551058195531368, + 0.005804875865578651, + 0.006725686602294445, + 0.004031654447317123, + 0.007521164137870073, + 0.004599461331963539, + 0.0101141557097435, + 0.008020828478038311, + 0.006850958801805973, + 0.007863220758736134, + 0.006798028480261564, + 0.011938614770770073, + 0.008241623640060425, + 0.007695705164223909, + 0.004359258804470301, + 0.005483837798237801, + 0.025265861302614212, + 0.003700678003951907, + 0.027007397264242172, + 0.005163978319615126, + 0.008737169206142426, + 0.002680369885638356, + 0.007946216501295567, + 0.0128384530544281, + 0.016712604090571404, + 0.006619608029723167, + 0.01222862210124731, + 0.008777705952525139, + 0.006422404199838638, + 0.01700284145772457, + 0.03742482140660286, + 0.010123912245035172, + 0.013160605914890766, + 0.009233731776475906, + 0.016597844660282135, + 0.009160495363175869, + 0.008665272034704685, + 0.02968515083193779, + 0.014005502685904503, + 0.016170231625437737, + 0.012751687318086624, + 0.02626274898648262, + 0.012145280838012695, + 0.011956322938203812, + 0.011569978669285774, + 0.00910219643265009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.02648574858903885 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028285279404371977 + ], + "opposite_logprob_tokens": [ + -5.877828598022461 + ], + "logprob_reward": -0.0028285279404371977, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028285279404371977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20453321933746338, + 0.059010546654462814, + 0.10113231837749481, + 0.147197887301445, + 0.027697322890162468, + 0.09034452587366104, + 0.08966401219367981, + 0.06690043956041336, + 0.14932870864868164, + 0.06419109553098679 + ], + "credit_map": [ + 0.01896582916378975, + 0.01930120401084423, + 0.047410305589437485, + 0.015221317298710346, + 0.037394121289253235, + 0.028276877477765083, + 0.021141309291124344, + 0.016822239384055138, + 0.005932277534157038, + 0.009636548347771168, + 0.0071560959331691265, + 0.007492543198168278, + 0.01151183433830738, + 0.005736282095313072, + 0.00676958030089736, + 0.0047753797844052315, + 0.025838198140263557, + 0.009500962682068348, + 0.008371546864509583, + 0.006846519187092781, + 0.012567116878926754, + 0.020807011052966118, + 0.00932263769209385, + 0.007878306321799755, + 0.0158648993819952, + 0.011462725698947906, + 0.017267640680074692, + 0.021318605169653893, + 0.016985543072223663, + 0.02097221091389656, + 0.030576786026358604, + 0.012749463319778442, + 0.0027549697551876307, + 0.0022144827526062727, + 0.004877564497292042, + 0.004190155304968357, + 0.004461904522031546, + 0.0019557990599423647, + 0.004991845693439245, + 0.0022506017703562975, + 0.012186731211841106, + 0.018975697457790375, + 0.009704888798296452, + 0.007452463265508413, + 0.00816480815410614, + 0.020220061764121056, + 0.007185946684330702, + 0.006453919690102339, + 0.004672167357057333, + 0.005225832108408213, + 0.02971736341714859, + 0.0038850605487823486, + 0.03096332959830761, + 0.005544197279959917, + 0.0067350734025239944, + 0.0029209854546934366, + 0.006181873846799135, + 0.007957728579640388, + 0.008898619562387466, + 0.005202509928494692, + 0.016408752650022507, + 0.005643446929752827, + 0.007588185835629702, + 0.009019319899380207, + 0.037630707025527954, + 0.011952158063650131, + 0.009903344325721264, + 0.011929843574762344, + 0.02333085983991623, + 0.011711865663528442, + 0.009541001170873642, + 0.033328909426927567, + 0.005022404249757528, + 0.005918012000620365, + 0.007032722234725952, + 0.012217316776514053, + 0.019158778712153435, + 0.004181486088782549, + 0.005561069585382938, + 0.0050993082113564014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.0028285279404371977 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193756565451622 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.003193756565451622, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193756565451622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26718926429748535, + 0.07938481122255325, + 0.10725536942481995, + 0.1784742921590805, + 0.05857352912425995, + 0.10989399999380112, + 0.06657157093286514, + 0.07960889488458633, + 0.04146797955036163, + 0.011580240912735462 + ], + "credit_map": [ + 0.030941592529416084, + 0.016038773581385612, + 0.023132817819714546, + 0.022671980783343315, + 0.032804928719997406, + 0.07316289097070694, + 0.020936310291290283, + 0.04749998822808266, + 0.008521903306245804, + 0.021148134022951126, + 0.008546315133571625, + 0.008921467699110508, + 0.010600108653306961, + 0.008057220838963985, + 0.0069278148002922535, + 0.006661851890385151, + 0.02300439402461052, + 0.0139978202059865, + 0.009015707299113274, + 0.0066620949655771255, + 0.009558339603245258, + 0.031014326959848404, + 0.0067038387060165405, + 0.007298852317035198, + 0.018211722373962402, + 0.016075745224952698, + 0.018011217936873436, + 0.03691060096025467, + 0.02025735192000866, + 0.025260498747229576, + 0.029232384636998177, + 0.01451479084789753, + 0.007038521580398083, + 0.005187870003283024, + 0.008612260222434998, + 0.008324947208166122, + 0.0075788055546581745, + 0.0039374916814267635, + 0.011531365104019642, + 0.00636227335780859, + 0.014281390234827995, + 0.022908836603164673, + 0.014283706434071064, + 0.014940311200916767, + 0.011799268424510956, + 0.013687868602573872, + 0.010694229044020176, + 0.007298402953892946, + 0.004465174395591021, + 0.005880242213606834, + 0.021910125389695168, + 0.0033977669663727283, + 0.01716158725321293, + 0.005657223518937826, + 0.005245591513812542, + 0.002853858983144164, + 0.008573896251618862, + 0.020317982882261276, + 0.010818427428603172, + 0.008930896408855915, + 0.0068800197914242744, + 0.008172472938895226, + 0.0056180935353040695, + 0.010297111235558987, + 0.003684296738356352, + 0.003754963865503669, + 0.0049173892475664616, + 0.003504486521705985, + 0.014617236331105232, + 0.004309191834181547, + 0.0036332940217107534, + 0.0030471228528767824, + 0.0012082788161933422, + 0.001273346715606749, + 0.0012316425563767552, + 0.00176896620541811, + 0.0025934495497494936, + 0.0012695088516920805, + 0.001027477439492941, + 0.0012075714766979218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.003193756565451622 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078749783337116 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078749783337116, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078749783337116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17310044169425964, + 0.08293796330690384, + 0.22545529901981354, + 0.20063255727291107, + 0.060147713869810104, + 0.031759459525346756, + 0.03254038840532303, + 0.03253987804055214, + 0.14940467476844788, + 0.011481691151857376 + ], + "credit_map": [ + 0.02131257764995098, + 0.011537350714206696, + 0.020346544682979584, + 0.008748157881200314, + 0.015918215736746788, + 0.047532252967357635, + 0.015533638186752796, + 0.03217170760035515, + 0.008629221469163895, + 0.01982841268181801, + 0.00945583451539278, + 0.008337120525538921, + 0.013725594617426395, + 0.008696370758116245, + 0.00808270089328289, + 0.006182692013680935, + 0.07052204012870789, + 0.016170047223567963, + 0.0133931003510952, + 0.010358527302742004, + 0.011446711607277393, + 0.080622099339962, + 0.005988307762891054, + 0.016954436898231506, + 0.02243862673640251, + 0.02399626187980175, + 0.025754742324352264, + 0.031120378524065018, + 0.021198129281401634, + 0.022972648963332176, + 0.03138041868805885, + 0.02177133597433567, + 0.006562026217579842, + 0.006006548646837473, + 0.008812070824205875, + 0.009777737781405449, + 0.006490127649158239, + 0.00523383030667901, + 0.011347385123372078, + 0.005917983129620552, + 0.0044921329244971275, + 0.004648276139050722, + 0.0034587038680911064, + 0.0027080425061285496, + 0.00596088171005249, + 0.004018610808998346, + 0.004247595090419054, + 0.0022252153139561415, + 0.0026639425195753574, + 0.0036743758246302605, + 0.007614270318299532, + 0.0026534516364336014, + 0.005224877502769232, + 0.003474441124126315, + 0.004669941030442715, + 0.0025650840252637863, + 0.0027185233775526285, + 0.004499819129705429, + 0.004810031037777662, + 0.00254170224070549, + 0.005020110867917538, + 0.0028365254402160645, + 0.002487911144271493, + 0.007625250145792961, + 0.042518250644207, + 0.011165323667228222, + 0.009392845444381237, + 0.013954229652881622, + 0.024945911020040512, + 0.01071763876825571, + 0.009231037460267544, + 0.027479426935315132, + 0.0012901111040264368, + 0.0011102232383564115, + 0.0013109780848026276, + 0.0017193076200783253, + 0.002766100922599435, + 0.0013325785985216498, + 0.0009468001662753522, + 0.0010055904276669025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.011078749783337116 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0004368066438473761 + ], + "opposite_logprob_tokens": [ + -7.750436782836914 + ], + "logprob_reward": -0.0004368066438473761, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0004368066438473761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2324172556400299, + 0.14628168940544128, + 0.17973096668720245, + 0.1581069529056549, + 0.0610920786857605, + 0.05473523959517479, + 0.053054168820381165, + 0.03252718970179558, + 0.04909542575478554, + 0.03295900300145149 + ], + "credit_map": [ + 0.03017081506550312, + 0.01707303151488304, + 0.051804687827825546, + 0.015683935955166817, + 0.0175221785902977, + 0.03637934476137161, + 0.01846431754529476, + 0.045318976044654846, + 0.017451459541916847, + 0.03376789018511772, + 0.013402889482676983, + 0.015840869396924973, + 0.023765871301293373, + 0.014880849048495293, + 0.014601248316466808, + 0.012570612132549286, + 0.05329794064164162, + 0.014979187399148941, + 0.01084629725664854, + 0.007787726819515228, + 0.011781451292335987, + 0.06021467596292496, + 0.00773668335750699, + 0.01308701653033495, + 0.017150167375802994, + 0.014066251926124096, + 0.020644763484597206, + 0.029773889109492302, + 0.017491960898041725, + 0.019406765699386597, + 0.024366946890950203, + 0.015206226147711277, + 0.006128272507339716, + 0.006134392693638802, + 0.009566478431224823, + 0.008017105981707573, + 0.010621615685522556, + 0.004759481642395258, + 0.01099252700805664, + 0.004872215446084738, + 0.004679950885474682, + 0.007249629124999046, + 0.005060171242803335, + 0.010527939535677433, + 0.0075361160561442375, + 0.008102587424218655, + 0.007218648679554462, + 0.0043602087534964085, + 0.004345978144556284, + 0.006394415628165007, + 0.009357846342027187, + 0.004389598499983549, + 0.007748871576040983, + 0.005931098014116287, + 0.010779507458209991, + 0.004106858745217323, + 0.003089342499151826, + 0.006168261636048555, + 0.004202006384730339, + 0.0037794956006109715, + 0.004203120246529579, + 0.0029627846088260412, + 0.0021933557000011206, + 0.005928829312324524, + 0.006353986449539661, + 0.004342956468462944, + 0.0051910970360040665, + 0.0048688966780900955, + 0.016211850568652153, + 0.004755106288939714, + 0.003936808090656996, + 0.003434728365391493, + 0.005181706976145506, + 0.0035235595423728228, + 0.004111332353204489, + 0.0073165581561625, + 0.00317125185392797, + 0.003685689065605402, + 0.0032223237212747335, + 0.002746584126725793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.0004368066438473761 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed, top-down, bird's-eye view immerses the viewer in a classic pixel art world, its visuals reminiscent of the tense atmosphere of a vintage video game. At the center of the frame, a simple, square life raft, constructed from bright orange blocks, struggles amidst a vast ocean of dark blue pixels. Massive waves, procedurally composed of alternating dark navy and slightly lighter cyan blocks, create a dynamic sense of depth, their crests dotted with striking foam composed of pure white pixels. Rather than a smooth, fluid motion, the life raft is violently tossed up and down in this digital turmoil, jerking and jerking with each wave, pushed to the corners of the frame and pulled back toward the center. The scene utilizes a flat, shadowless lighting scheme and a limited, saturated color palette. The strong color contrast and blocky dynamics create a tense atmosphere of isolation and life-or-death struggle, perfectly recreating the unique aesthetic and pressure of survival games from the 8-bit era.", + "seed": 16663, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_15/index_801_seed_16663_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01120861847884953, + "vlm_logprob": -0.01120861847884953, + "vlm_yes_no_margin": 5.4, + "vlm_token_credit": -0.01120861847884953, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020711224526166916 + ], + "opposite_logprob_tokens": [ + -4.145711421966553 + ], + "logprob_reward": -0.020711224526166916, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.020711224526166916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44241732358932495, + 0.05116571858525276, + 0.1480240821838379, + 0.08592207729816437, + 0.06348703056573868, + 0.03348448500037193, + 0.06521479785442352, + 0.009033530950546265, + 0.03668889403343201, + 0.06456209719181061 + ], + "credit_map": [ + 0.03054274059832096, + 0.02318723127245903, + 0.10666323453187943, + 0.01607484742999077, + 0.014400653541088104, + 0.1812080442905426, + 0.010568791069090366, + 0.05977172031998634, + 0.005791720002889633, + 0.006171730812638998, + 0.008599208667874336, + 0.004568222910165787, + 0.003950851038098335, + 0.007196754217147827, + 0.007185195107012987, + 0.007702031638473272, + 0.0251627117395401, + 0.011178785935044289, + 0.041286543011665344, + 0.008805133402347565, + 0.010558457113802433, + 0.02528364397585392, + 0.00905575044453144, + 0.016693035140633583, + 0.02557152323424816, + 0.008411618880927563, + 0.011393914930522442, + 0.004308430012315512, + 0.003813203191384673, + 0.017637114971876144, + 0.00309983198530972, + 0.011686425656080246, + 0.008280017413198948, + 0.005941023584455252, + 0.012811552733182907, + 0.004970211535692215, + 0.0064362697303295135, + 0.011611820198595524, + 0.00497811334207654, + 0.008458012714982033, + 0.006243994925171137, + 0.00313968094997108, + 0.006451902445405722, + 0.0021109164226800203, + 0.00287057738751173, + 0.00408119335770607, + 0.0017169031780213118, + 0.006869311910122633, + 0.015236109495162964, + 0.006066238507628441, + 0.01235069241374731, + 0.004288618452847004, + 0.004379167687147856, + 0.010933603160083294, + 0.002551475539803505, + 0.009408879093825817, + 0.0016989634605124593, + 0.000873517943546176, + 0.0013718002010136843, + 0.0005620230222120881, + 0.0008215304696932435, + 0.0016523277154192328, + 0.0006177322356961668, + 0.0014356342144310474, + 0.004788549151271582, + 0.0029233540408313274, + 0.012591943144798279, + 0.001620049704797566, + 0.0016901874914765358, + 0.003656738670542836, + 0.0017740349285304546, + 0.007644027937203646, + 0.013372792862355709, + 0.005655430257320404, + 0.009144073352217674, + 0.002613431541249156, + 0.0032832822762429714, + 0.019114984199404716, + 0.002194660482928157, + 0.00918344035744667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.020711224526166916 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015259062638506293 + ], + "opposite_logprob_tokens": [ + -6.50152587890625 + ], + "logprob_reward": -0.0015259062638506293, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015259062638506293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3530817925930023, + 0.08870846033096313, + 0.16330905258655548, + 0.0701623409986496, + 0.16304655373096466, + 0.017488328740000725, + 0.02532312087714672, + 0.014633671380579472, + 0.03936952352523804, + 0.06487719714641571 + ], + "credit_map": [ + 0.02000817470252514, + 0.015444007702171803, + 0.10566864162683487, + 0.013499902561306953, + 0.010472980327904224, + 0.04580918699502945, + 0.012258085422217846, + 0.12992079555988312, + 0.009809572249650955, + 0.010505123063921928, + 0.014278918504714966, + 0.007692965678870678, + 0.01051601953804493, + 0.00930357351899147, + 0.01389621663838625, + 0.012706068344414234, + 0.024837056174874306, + 0.012725229375064373, + 0.038736555725336075, + 0.008479622192680836, + 0.013083034195005894, + 0.030090350657701492, + 0.0173063725233078, + 0.018050823360681534, + 0.014638676308095455, + 0.008065415546298027, + 0.006536789238452911, + 0.005378178786486387, + 0.00673689367249608, + 0.014124698005616665, + 0.005570258945226669, + 0.009111428633332253, + 0.027168327942490578, + 0.016124142333865166, + 0.020255016162991524, + 0.01271319855004549, + 0.0223092008382082, + 0.031408071517944336, + 0.012645265087485313, + 0.020423341542482376, + 0.0024466942995786667, + 0.0020643288735300303, + 0.002853924874216318, + 0.001618206617422402, + 0.0021695129107683897, + 0.0018254142487421632, + 0.0014233567053452134, + 0.0030868886969983578, + 0.005313624627888203, + 0.0020447110291570425, + 0.0037138184998184443, + 0.0014056441141292453, + 0.0016484599327668548, + 0.0047996677458286285, + 0.0012081469176337123, + 0.005189047195017338, + 0.002740113530308008, + 0.0011902624974027276, + 0.002174197696149349, + 0.0010627952869981527, + 0.0019374338444322348, + 0.002528982236981392, + 0.0012662516674026847, + 0.001733632991090417, + 0.004025815520435572, + 0.0022031422704458237, + 0.015222093090415001, + 0.001608146820217371, + 0.0029657240957021713, + 0.0024929577484726906, + 0.0016833191039040685, + 0.009168327786028385, + 0.018076276406645775, + 0.005711756646633148, + 0.005682666320353746, + 0.003720237873494625, + 0.003908962476998568, + 0.01869676634669304, + 0.004612131044268608, + 0.004468394909054041 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.0015259062638506293 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031931623816490173 + ], + "opposite_logprob_tokens": [ + -5.753193378448486 + ], + "logprob_reward": -0.0031931623816490173, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031931623816490173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15657790005207062, + 0.11710216104984283, + 0.16459926962852478, + 0.06080153211951256, + 0.179783433675766, + 0.12999428808689117, + 0.057754792273044586, + 0.023352187126874924, + 0.04785139486193657, + 0.0621829628944397 + ], + "credit_map": [ + 0.015241120010614395, + 0.011247959919273853, + 0.01997433416545391, + 0.006941651925444603, + 0.009556269273161888, + 0.04226507246494293, + 0.010108045302331448, + 0.04124346375465393, + 0.011682389304041862, + 0.014190573245286942, + 0.016658274456858635, + 0.011808163486421108, + 0.012182146310806274, + 0.0169999822974205, + 0.018164267763495445, + 0.01541637908667326, + 0.02203340269625187, + 0.0144463200122118, + 0.022255655378103256, + 0.009287552908062935, + 0.013739807531237602, + 0.031637802720069885, + 0.019872793927788734, + 0.03132596239447594, + 0.013680371455848217, + 0.006383028347045183, + 0.006850186735391617, + 0.003799185622483492, + 0.00458357622846961, + 0.014586144126951694, + 0.003874504705891013, + 0.007044542580842972, + 0.03657640889286995, + 0.015903225168585777, + 0.04778994619846344, + 0.009707343764603138, + 0.012971378862857819, + 0.027612891048192978, + 0.0077594066970050335, + 0.021462857723236084, + 0.02470676600933075, + 0.011300627142190933, + 0.02749202586710453, + 0.009006408974528313, + 0.007759901694953442, + 0.01588059589266777, + 0.00614528264850378, + 0.02770269103348255, + 0.011568493209779263, + 0.005070564802736044, + 0.008660201914608479, + 0.0035831639543175697, + 0.0039920206181705, + 0.010792713612318039, + 0.0035299400333315134, + 0.010557703673839569, + 0.004174658562988043, + 0.0016663726419210434, + 0.003796955803409219, + 0.0012801667908206582, + 0.0019478702452033758, + 0.005039334297180176, + 0.0016163698164746165, + 0.003830461297184229, + 0.007627840619534254, + 0.005267064552754164, + 0.00748808216303587, + 0.0030767186544835567, + 0.003751721465960145, + 0.0073764268308877945, + 0.003434561425819993, + 0.00982898660004139, + 0.012158505618572235, + 0.008265797980129719, + 0.011931668035686016, + 0.0030425661243498325, + 0.0031588806305080652, + 0.012358610518276691, + 0.0024031957145780325, + 0.008863748051226139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.0031931623816490173 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029801908880472183 + ], + "opposite_logprob_tokens": [ + -3.529801845550537 + ], + "logprob_reward": -0.029801908880472183, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029801908880472183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5472223162651062, + 0.08295554667711258, + 0.08295994997024536, + 0.04142158105969429, + 0.05372696742415428, + 0.035251684486866, + 0.035401616245508194, + 0.012177146971225739, + 0.07355058193206787, + 0.03533262386918068 + ], + "credit_map": [ + 0.03778047487139702, + 0.028587237000465393, + 0.16688190400600433, + 0.01860252022743225, + 0.016668589785695076, + 0.22313068807125092, + 0.010120846331119537, + 0.04545007273554802, + 0.009993382729589939, + 0.008154005743563175, + 0.013210896402597427, + 0.008384824730455875, + 0.008720412850379944, + 0.00878904853016138, + 0.014534336514770985, + 0.011168639175593853, + 0.011395879089832306, + 0.007752754259854555, + 0.01554196234792471, + 0.005785736255347729, + 0.009119236841797829, + 0.011949336156249046, + 0.01089517679065466, + 0.010519864037632942, + 0.009081411175429821, + 0.005503548309206963, + 0.004350923001766205, + 0.0028453152626752853, + 0.004861065186560154, + 0.006568122189491987, + 0.004414367023855448, + 0.00379682844504714, + 0.008593224920332432, + 0.005647988524287939, + 0.008964866399765015, + 0.0041862232610583305, + 0.005726178176701069, + 0.010115907527506351, + 0.0034764192532747984, + 0.007016158662736416, + 0.0071389335207641125, + 0.00317281112074852, + 0.0053832922130823135, + 0.0027699307538568974, + 0.0023052238393574953, + 0.004174075089395046, + 0.0015581909101456404, + 0.008749225176870823, + 0.006209087558090687, + 0.0030497706029564142, + 0.00648400979116559, + 0.0032563505228608847, + 0.0031187720596790314, + 0.005130348261445761, + 0.0028859018348157406, + 0.005267378408461809, + 0.0019655111245810986, + 0.001027974532917142, + 0.0021145872306078672, + 0.0010362429311499, + 0.0015018472913652658, + 0.0017905577551573515, + 0.0010530565632507205, + 0.0016873696586117148, + 0.011850809678435326, + 0.007649173494428396, + 0.020842794328927994, + 0.004974041599780321, + 0.005675225052982569, + 0.0071030231192708015, + 0.004772332962602377, + 0.010683181695640087, + 0.009108252823352814, + 0.0021545516792684793, + 0.004037945996969938, + 0.0013693510554730892, + 0.0021057561971247196, + 0.010745834559202194, + 0.001604472636245191, + 0.004206460434943438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.029801908880472183 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008108903421089053 + ], + "opposite_logprob_tokens": [ + -7.1258111000061035 + ], + "logprob_reward": -0.0008108903421089053, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008108903421089053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24050019681453705, + 0.06480579823255539, + 0.1891281008720398, + 0.04388836398720741, + 0.06182865425944328, + 0.06464161723852158, + 0.05150126665830612, + 0.015219505876302719, + 0.10535957664251328, + 0.1631269007921219 + ], + "credit_map": [ + 0.025657737627625465, + 0.014372671954333782, + 0.07403603941202164, + 0.008618888445198536, + 0.010218953713774681, + 0.05522444099187851, + 0.010678602382540703, + 0.041692860424518585, + 0.007317712530493736, + 0.0070491270162165165, + 0.00868785660713911, + 0.006322997156530619, + 0.007431515958160162, + 0.007948682643473148, + 0.011604569852352142, + 0.008443332277238369, + 0.033836718648672104, + 0.013790186494588852, + 0.050045743584632874, + 0.01166203711181879, + 0.012744504027068615, + 0.03046952188014984, + 0.016929911449551582, + 0.019649475812911987, + 0.007290360983461142, + 0.005892802961170673, + 0.004530900623649359, + 0.00332639436237514, + 0.004511594772338867, + 0.00979031901806593, + 0.003393382765352726, + 0.005152606405317783, + 0.014736827462911606, + 0.005230305250734091, + 0.01316859107464552, + 0.0028642246033996344, + 0.004077186342328787, + 0.01008134800940752, + 0.0034707807935774326, + 0.008199392817914486, + 0.012199698947370052, + 0.006015240214765072, + 0.013350835070014, + 0.004228812642395496, + 0.0036847747396677732, + 0.008230275474488735, + 0.0032142335548996925, + 0.013717749156057835, + 0.010514491237699986, + 0.004935446660965681, + 0.009480823762714863, + 0.003365457756444812, + 0.003500764723867178, + 0.008349291980266571, + 0.003691536607220769, + 0.007663457188755274, + 0.0023839131463319063, + 0.0013818989973515272, + 0.0025160289369523525, + 0.0010585231939330697, + 0.0015497541753575206, + 0.002529514255002141, + 0.0013011707924306393, + 0.002498702146112919, + 0.012660124339163303, + 0.008129659108817577, + 0.0364614799618721, + 0.004813336301594973, + 0.00857146643102169, + 0.009086692705750465, + 0.004678684286773205, + 0.020958134904503822, + 0.025751257315278053, + 0.014879830181598663, + 0.028049279004335403, + 0.007200210355222225, + 0.009881540201604366, + 0.039529070258140564, + 0.008021720685064793, + 0.02981400303542614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.0008108903421089053 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed, top-down, bird's-eye view immerses the viewer in a classic pixel art world, its visuals reminiscent of the tense atmosphere of a vintage video game. At the center of the frame, a simple, square life raft, constructed from bright orange blocks, struggles amidst a vast ocean of dark blue pixels. Massive waves, procedurally composed of alternating dark navy and slightly lighter cyan blocks, create a dynamic sense of depth, their crests dotted with striking foam composed of pure white pixels. Rather than a smooth, fluid motion, the life raft is violently tossed up and down in this digital turmoil, jerking and jerking with each wave, pushed to the corners of the frame and pulled back toward the center. The scene utilizes a flat, shadowless lighting scheme and a limited, saturated color palette. The strong color contrast and blocky dynamics create a tense atmosphere of isolation and life-or-death struggle, perfectly recreating the unique aesthetic and pressure of survival games from the 8-bit era.", + "seed": 16664, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_15/index_801_seed_16664_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01731888351496309, + "vlm_logprob": -0.01731888351496309, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.01731888351496309, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030980058014392853 + ], + "opposite_logprob_tokens": [ + -3.655980110168457 + ], + "logprob_reward": -0.030980058014392853, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.030980058014392853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28976303339004517, + 0.26505202054977417, + 0.18619702756404877, + 0.020498257130384445, + 0.02496679686009884, + 0.02579653449356556, + 0.05910699814558029, + 0.034504130482673645, + 0.07718095928430557, + 0.01693420298397541 + ], + "credit_map": [ + 0.0750393196940422, + 0.022763175889849663, + 0.06600244343280792, + 0.025286266580224037, + 0.022819314152002335, + 0.03477730229496956, + 0.013339133933186531, + 0.029736099764704704, + 0.067220039665699, + 0.019584819674491882, + 0.04546208679676056, + 0.019458193331956863, + 0.02148224040865898, + 0.033799055963754654, + 0.017266299575567245, + 0.04077928140759468, + 0.03274969756603241, + 0.01166506391018629, + 0.03446897491812706, + 0.009403026662766933, + 0.015371869318187237, + 0.026898019015789032, + 0.010313862934708595, + 0.04532650113105774, + 0.003549630520865321, + 0.0016039017355069518, + 0.0037352563813328743, + 0.0017820770153775811, + 0.0019054977456107736, + 0.0029860343784093857, + 0.002013328019529581, + 0.0029225307516753674, + 0.002818034728989005, + 0.00203731469810009, + 0.004887305200099945, + 0.0021100137382745743, + 0.0022377485875040293, + 0.0033036451786756516, + 0.0021465786267071962, + 0.005426155403256416, + 0.002497483976185322, + 0.002085575135424733, + 0.002218969166278839, + 0.0031419002916663885, + 0.005173127632588148, + 0.0035198451951146126, + 0.0039872219786047935, + 0.0031724118161946535, + 0.006970912683755159, + 0.005683458875864744, + 0.006514399778097868, + 0.004474413115531206, + 0.009270347654819489, + 0.0120787862688303, + 0.00789264589548111, + 0.00622203666716814, + 0.005815921351313591, + 0.002384649822488427, + 0.006626698188483715, + 0.0021571707911789417, + 0.0035359621979296207, + 0.00469641899690032, + 0.0026320049073547125, + 0.006655306089669466, + 0.007620200514793396, + 0.00423702085390687, + 0.024568406865000725, + 0.004469814244657755, + 0.003409219440072775, + 0.0093691386282444, + 0.003724303562194109, + 0.019782859832048416, + 0.001846542116254568, + 0.0015500551089644432, + 0.002645050175487995, + 0.0010584035189822316, + 0.0020622671581804752, + 0.0027859967667609453, + 0.0014123610453680158, + 0.0035735268611460924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.030980058014392853 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002834114944562316 + ], + "opposite_logprob_tokens": [ + -5.877834320068359 + ], + "logprob_reward": -0.002834114944562316, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002834114944562316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14278919994831085, + 0.39841121435165405, + 0.15711699426174164, + 0.019573643803596497, + 0.06355614215135574, + 0.054276127368211746, + 0.02830062434077263, + 0.01960090734064579, + 0.05328645929694176, + 0.06308872997760773 + ], + "credit_map": [ + 0.03818438574671745, + 0.009526682086288929, + 0.019818531349301338, + 0.011097578331828117, + 0.011358940973877907, + 0.019004182890057564, + 0.01232758816331625, + 0.021471308544278145, + 0.09208113700151443, + 0.030577486380934715, + 0.04815121740102768, + 0.038996320217847824, + 0.03694339469075203, + 0.04949845373630524, + 0.03449751064181328, + 0.06766565144062042, + 0.026941264048218727, + 0.009111148305237293, + 0.0335213802754879, + 0.00782142486423254, + 0.013159330002963543, + 0.02186456136405468, + 0.009081021882593632, + 0.03561684861779213, + 0.002738446230068803, + 0.0017079712124541402, + 0.0027787929866462946, + 0.0021740747615695, + 0.002252465346828103, + 0.0024514535907655954, + 0.0024668509140610695, + 0.003003589576110244, + 0.008683604188263416, + 0.004949883092194796, + 0.008144799619913101, + 0.005282580852508545, + 0.009982449002563953, + 0.008360828272998333, + 0.006747931241989136, + 0.011404065415263176, + 0.006483002100139856, + 0.0034983085934072733, + 0.006688822992146015, + 0.004948439076542854, + 0.007133919280022383, + 0.007733721751719713, + 0.006643782835453749, + 0.011146132834255695, + 0.0035185401793569326, + 0.002260445151478052, + 0.004829127341508865, + 0.002693179063498974, + 0.0026140990667045116, + 0.004532776772975922, + 0.003672392340376973, + 0.004180065356194973, + 0.0031156346667557955, + 0.001291349995881319, + 0.004049686715006828, + 0.002108838874846697, + 0.0018331115134060383, + 0.0027269700076431036, + 0.0022861638572067022, + 0.0021891524083912373, + 0.010576640255749226, + 0.004484496079385281, + 0.005550164729356766, + 0.006179760675877333, + 0.008270522579550743, + 0.006412578281015158, + 0.006639279890805483, + 0.005173016805201769, + 0.008101072162389755, + 0.004177768249064684, + 0.0058312127366662025, + 0.004863611422479153, + 0.005494754295796156, + 0.01739134080708027, + 0.00528579531237483, + 0.011943177320063114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.002834114944562316 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00319268717430532 + ], + "opposite_logprob_tokens": [ + -5.753192901611328 + ], + "logprob_reward": -0.00319268717430532, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00319268717430532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17990639805793762, + 0.11634103953838348, + 0.21555525064468384, + 0.0565289743244648, + 0.06904611736536026, + 0.060188859701156616, + 0.030118068680167198, + 0.11001445353031158, + 0.10534334182739258, + 0.05695755034685135 + ], + "credit_map": [ + 0.02848731353878975, + 0.010927191004157066, + 0.03228956460952759, + 0.011492786929011345, + 0.019938189536333084, + 0.0339675098657608, + 0.0116448774933815, + 0.031158961355686188, + 0.029244672507047653, + 0.008776159957051277, + 0.019200097769498825, + 0.009751991368830204, + 0.011830626055598259, + 0.014817139133810997, + 0.007370563689619303, + 0.015349800698459148, + 0.055754806846380234, + 0.01448262669146061, + 0.033794794231653214, + 0.016153555363416672, + 0.01876833476126194, + 0.03483199328184128, + 0.012529582716524601, + 0.02923954278230667, + 0.009492803364992142, + 0.0039079817943274975, + 0.01290096528828144, + 0.004692973103374243, + 0.004790592473000288, + 0.007571271155029535, + 0.004056266043335199, + 0.009116116911172867, + 0.01115320436656475, + 0.007160454522818327, + 0.011834450997412205, + 0.0048400890082120895, + 0.007526830304414034, + 0.010294423438608646, + 0.004880562890321016, + 0.011356102302670479, + 0.009671149775385857, + 0.005850496701896191, + 0.008707646280527115, + 0.004438372794538736, + 0.0076601202599704266, + 0.010465686209499836, + 0.005186049267649651, + 0.008209331892430782, + 0.004249298945069313, + 0.0018284727120772004, + 0.0052336580120027065, + 0.002308368682861328, + 0.0029541882686316967, + 0.005864736624062061, + 0.002673177048563957, + 0.005006169900298119, + 0.019994324073195457, + 0.008663454093039036, + 0.02200060524046421, + 0.008053484372794628, + 0.009114311076700687, + 0.015437021851539612, + 0.008986766450107098, + 0.017764484509825706, + 0.009956819005310535, + 0.005741034634411335, + 0.03220393881201744, + 0.005967673379927874, + 0.005283525213599205, + 0.013050301931798458, + 0.0055489144288003445, + 0.02759113721549511, + 0.007149832788854837, + 0.004233361221849918, + 0.007229441311210394, + 0.004634765908122063, + 0.004797894041985273, + 0.012006668373942375, + 0.0037822260055691004, + 0.01312335766851902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.00319268717430532 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486685186624527 + ], + "opposite_logprob_tokens": [ + -3.048668622970581 + ], + "logprob_reward": -0.0486685186624527, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486685186624527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30711817741394043, + 0.07712677866220474, + 0.16401436924934387, + 0.0425400547683239, + 0.06376679986715317, + 0.03656228631734848, + 0.0570572204887867, + 0.13255004584789276, + 0.04145406186580658, + 0.07781020551919937 + ], + "credit_map": [ + 0.07701535522937775, + 0.020237833261489868, + 0.07282997667789459, + 0.022795243188738823, + 0.015595083124935627, + 0.04264160245656967, + 0.014937766827642918, + 0.04106533154845238, + 0.01695752888917923, + 0.007286696694791317, + 0.010892615653574467, + 0.006051613483577967, + 0.010636689141392708, + 0.010128052905201912, + 0.005898086819797754, + 0.009275492280721664, + 0.01680232398211956, + 0.010993831790983677, + 0.03830774128437042, + 0.013835765421390533, + 0.013150027953088284, + 0.023085257038474083, + 0.011883199214935303, + 0.03595621883869171, + 0.006376524921506643, + 0.0042023551650345325, + 0.005549782421439886, + 0.004549643490463495, + 0.0056771887466311455, + 0.005513468757271767, + 0.005066943820565939, + 0.00560414232313633, + 0.006944333203136921, + 0.005832912400364876, + 0.010110163129866123, + 0.005528345238417387, + 0.007468730211257935, + 0.00757910031825304, + 0.005864005070179701, + 0.014439214952290058, + 0.004248577170073986, + 0.0031173396855592728, + 0.004462383221834898, + 0.0031300270929932594, + 0.006412144750356674, + 0.004799636080861092, + 0.005089405924081802, + 0.005302770994603634, + 0.00665033794939518, + 0.003159098792821169, + 0.010294611565768719, + 0.002992141293361783, + 0.0033856809604912996, + 0.0138242831453681, + 0.002968363231047988, + 0.013782703317701817, + 0.021390995010733604, + 0.008646786212921143, + 0.035222992300987244, + 0.010839194990694523, + 0.01387855689972639, + 0.019621150568127632, + 0.009481609798967838, + 0.013468767516314983, + 0.00410429947078228, + 0.003105706302449107, + 0.009744579903781414, + 0.00307902111671865, + 0.005000937730073929, + 0.004509642254561186, + 0.004299205727875233, + 0.007610669359564781, + 0.01021285355091095, + 0.005609655752778053, + 0.012225559912621975, + 0.007821540348231792, + 0.00646605622023344, + 0.011067187413573265, + 0.005968564655631781, + 0.01843879371881485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0486685186624527 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle a fixed top-down bird's-eye view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central object a square life raft constructed from bright orange blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the waves composed of alternating dark navy and cyan blocks with white pixel foam?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the life raft moving with a jerking up-and-down motion rather than smooth fluid motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene use a flat, shadowless lighting scheme with a limited saturated color palette?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009190387791022658 + ], + "opposite_logprob_tokens": [ + -7.000918865203857 + ], + "logprob_reward": -0.0009190387791022658, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009190387791022658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27425214648246765, + 0.09389488399028778, + 0.08273854106664658, + 0.030444806441664696, + 0.04802650585770607, + 0.10434328019618988, + 0.16749629378318787, + 0.08836270868778229, + 0.039784833788871765, + 0.07065603882074356 + ], + "credit_map": [ + 0.0769386813044548, + 0.01785970665514469, + 0.06329695135354996, + 0.02071356028318405, + 0.015615304931998253, + 0.035076599568128586, + 0.011149381287395954, + 0.033601950854063034, + 0.019931819289922714, + 0.007435898296535015, + 0.016683606430888176, + 0.007429706864058971, + 0.009865070693194866, + 0.01255240011960268, + 0.007087325677275658, + 0.012909051962196827, + 0.011985208839178085, + 0.006394318770617247, + 0.022587891668081284, + 0.005437551531940699, + 0.00634717708453536, + 0.012693280354142189, + 0.004238918889313936, + 0.013054191134870052, + 0.005013646092265844, + 0.0025636267382651567, + 0.006052662618458271, + 0.0020992362406104803, + 0.0038183515425771475, + 0.0043697599321603775, + 0.002245508134365082, + 0.004282013047486544, + 0.007920837961137295, + 0.004328991752117872, + 0.008671182207763195, + 0.0031362720765173435, + 0.005009816028177738, + 0.005825402215123177, + 0.0033847433514893055, + 0.009749259799718857, + 0.013298152945935726, + 0.006061655003577471, + 0.01755957305431366, + 0.005965167190879583, + 0.00915940199047327, + 0.015013201162219048, + 0.007186940871179104, + 0.030099183320999146, + 0.018944403156638145, + 0.009598109871149063, + 0.026893025264143944, + 0.007612684741616249, + 0.01256620790809393, + 0.043802324682474136, + 0.009412235580384731, + 0.03866730257868767, + 0.015868907794356346, + 0.00592795992270112, + 0.01504779327660799, + 0.005072330124676228, + 0.009564164094626904, + 0.012977059930562973, + 0.007421459536999464, + 0.016483034938573837, + 0.0059589967131614685, + 0.003213381627574563, + 0.010575422085821629, + 0.0029248546343296766, + 0.0036867940798401833, + 0.004250743892043829, + 0.003325588069856167, + 0.005849050357937813, + 0.010610820725560188, + 0.0050333421677351, + 0.007616826798766851, + 0.0029465288389474154, + 0.004072615411132574, + 0.022193415090441704, + 0.002752391155809164, + 0.015430103987455368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0009190387791022658 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_2/index_tensor([3596, 3596])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_2/index_tensor([3596, 3596])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..25234db7352a0a5013b87962bdc9545d6df2c5a8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_2/index_tensor([3596, 3596])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly zooming in, slightly overhead, medium shot focuses on a warm, round wooden table where three young friends sit intimately around a steaming hot pot. In the center, a vibrant red oil-based broth simmers vigorously, bubbling with dried chilies and Sichuan peppercorns. A milky, clear broth in the other half gently steams. Rich, white steam rises under the warm yellow light of a chandelier, creating a soft halo. A woman with long, curly brown hair and wearing a loose, off-white knit sweater, her face a healthy, steamy flush, smiles cheerfully. Across from her, a man wearing black-framed glasses and a dark blue hoodie gestures and excitedly talks, drawing laughter from his friends. As the camera slowly approaches, a woman carefully picks up a crisp, red and white marbled beef roll with a pair of long wooden chopsticks and steadily and calmly dips it into the boiling red oil. The moment the meat touches the broth, a string of fine bubbles rises around it, and the color of the fresh meat quickly darkens. The table is laden with a variety of fresh ingredients—emerald green vegetables, tender white tofu, and exquisitely presented meats. The entire scene is filled with the laughter of friends gathering together and the warm and cozy atmosphere of life, creating an atmosphere of supreme comfort and intimacy.", + "seed": 23274, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_2/index_3596_seed_23274_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.019475467689335346, + "vlm_logprob": -0.019475467689335346, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.019475467689335346, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022835038602352142 + ], + "opposite_logprob_tokens": [ + -4.7728352546691895 + ], + "logprob_reward": -0.022835038602352142, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.022835038602352142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11949892342090607, + 0.16363762319087982, + 0.032170217484235764, + 0.03515719622373581, + 0.04401298984885216, + 0.3693419098854065, + 0.10400323569774628, + 0.03486461192369461, + 0.051528200507164, + 0.045785095542669296 + ], + "credit_map": [ + 0.018768787384033203, + 0.02281331829726696, + 0.014767667278647423, + 0.019206535071134567, + 0.013708756305277348, + 0.014568164013326168, + 0.008736787363886833, + 0.006928912363946438, + 0.020461512729525566, + 0.016171569004654884, + 0.009052685461938381, + 0.00874325167387724, + 0.053484633564949036, + 0.0225369930267334, + 0.011908787302672863, + 0.021278202533721924, + 0.006689960137009621, + 0.0036744247190654278, + 0.002931104274466634, + 0.003826092230156064, + 0.004085245076566935, + 0.004253945779055357, + 0.0019719130359590054, + 0.0047375354915857315, + 0.006771830376237631, + 0.0039380015805363655, + 0.003271459136158228, + 0.002173319924622774, + 0.0033889259211719036, + 0.0035017607733607292, + 0.006132866255939007, + 0.005979033187031746, + 0.00449296785518527, + 0.005850040353834629, + 0.002819726476445794, + 0.004386499058455229, + 0.00621021119877696, + 0.010156756266951561, + 0.006031831260770559, + 0.0040649548172950745, + 0.02996932342648506, + 0.11724267154932022, + 0.014133881777524948, + 0.02011316828429699, + 0.019937876611948013, + 0.019251467660069466, + 0.12592530250549316, + 0.022768229246139526, + 0.015743253752589226, + 0.011188204400241375, + 0.009717525914311409, + 0.015731019899249077, + 0.017697550356388092, + 0.013928942382335663, + 0.013578226789832115, + 0.006418509408831596, + 0.0036703518126159906, + 0.003965095151215792, + 0.0034428657963871956, + 0.006025942973792553, + 0.005097231827676296, + 0.0044394852593541145, + 0.004643153864890337, + 0.0035804843064397573, + 0.003578023286536336, + 0.005677186883985996, + 0.005551532842218876, + 0.006091699935495853, + 0.01084030419588089, + 0.008246364071965218, + 0.007218260783702135, + 0.0043248278088867664, + 0.004902695771306753, + 0.006581397727131844, + 0.005866250954568386, + 0.0060884831473231316, + 0.0057367668487131596, + 0.006046036724001169, + 0.006797661539167166, + 0.003765807254239917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.022835038602352142 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022143623791635036 + ], + "opposite_logprob_tokens": [ + -6.127214431762695 + ], + "logprob_reward": -0.0022143623791635036, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022143623791635036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09747612476348877, + 0.1990896314382553, + 0.0627003163099289, + 0.020270109176635742, + 0.07183153182268143, + 0.17325779795646667, + 0.2261604219675064, + 0.04407680407166481, + 0.06808778643608093, + 0.03704950958490372 + ], + "credit_map": [ + 0.009337417781352997, + 0.01643110252916813, + 0.01587481051683426, + 0.012816517613828182, + 0.01357007585465908, + 0.012290861457586288, + 0.00873769074678421, + 0.008417638018727303, + 0.05758247524499893, + 0.018315570428967476, + 0.019545186311006546, + 0.013721482828259468, + 0.029133394360542297, + 0.02046169340610504, + 0.024354180321097374, + 0.015975678339600563, + 0.010094950906932354, + 0.007952994666993618, + 0.003953723236918449, + 0.022207725793123245, + 0.003848334075883031, + 0.004774495027959347, + 0.006350034382194281, + 0.0035180547274649143, + 0.004001301247626543, + 0.0024699014611542225, + 0.002103053964674473, + 0.002204574877396226, + 0.0021763003896921873, + 0.002659787889569998, + 0.001643694587983191, + 0.0030114948749542236, + 0.007250968832522631, + 0.008412860333919525, + 0.006455136463046074, + 0.01002416294068098, + 0.006445013917982578, + 0.009142925031483173, + 0.01783277466893196, + 0.006267689634114504, + 0.03819267451763153, + 0.06346224248409271, + 0.009057532995939255, + 0.00946059450507164, + 0.01242841873317957, + 0.00870826281607151, + 0.025778362527489662, + 0.00616971030831337, + 0.02000412344932556, + 0.028865208849310875, + 0.028816601261496544, + 0.04200589656829834, + 0.025424426421523094, + 0.027667364105582237, + 0.03619208186864853, + 0.017184726893901825, + 0.005342637654393911, + 0.004128209315240383, + 0.005128163378685713, + 0.008226447738707066, + 0.00481655728071928, + 0.004506878554821014, + 0.007143851835280657, + 0.004784060176461935, + 0.0070786564610898495, + 0.007750899996608496, + 0.00870775431394577, + 0.015219509601593018, + 0.007410014048218727, + 0.006889385636895895, + 0.010578341782093048, + 0.004453225061297417, + 0.00442827632650733, + 0.007495338562875986, + 0.003561563789844513, + 0.004641679115593433, + 0.004961407743394375, + 0.003656033193692565, + 0.005420143250375986, + 0.0028850689996033907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.0022143623791635036 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048645809292793274 + ], + "opposite_logprob_tokens": [ + -3.0486457347869873 + ], + "logprob_reward": -0.048645809292793274, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048645809292793274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2489510476589203, + 0.11807630211114883, + 0.053289592266082764, + 0.014514775015413761, + 0.038535717874765396, + 0.314365029335022, + 0.06974897533655167, + 0.01940178871154785, + 0.048020314425230026, + 0.07509654015302658 + ], + "credit_map": [ + 0.0239492729306221, + 0.04517683759331703, + 0.019759666174650192, + 0.034062378108501434, + 0.045801714062690735, + 0.0300719253718853, + 0.027727274224162102, + 0.022401954978704453, + 0.010047455318272114, + 0.011027179658412933, + 0.007848073728382587, + 0.008255159482359886, + 0.042793016880750656, + 0.015234011225402355, + 0.010399437509477139, + 0.012471968308091164, + 0.007553727366030216, + 0.006422334350645542, + 0.003203770611435175, + 0.019566427916288376, + 0.003528821049258113, + 0.0034228162840008736, + 0.005652332212775946, + 0.003939364105463028, + 0.002746304962784052, + 0.0023791168350726366, + 0.0013679356779903173, + 0.0016408762894570827, + 0.0016719700070098042, + 0.0017943689599633217, + 0.0013919757911935449, + 0.0015222267247736454, + 0.003921398892998695, + 0.005468045361340046, + 0.0025757162366062403, + 0.0038040343206375837, + 0.00428803451359272, + 0.005148726049810648, + 0.010379839688539505, + 0.0029499216470867395, + 0.024417059496045113, + 0.10029486566781998, + 0.01166833657771349, + 0.01749592088162899, + 0.01696222648024559, + 0.016606807708740234, + 0.108230359852314, + 0.018689433112740517, + 0.0065724230371415615, + 0.010813623666763306, + 0.00787021592259407, + 0.013174150139093399, + 0.007378553971648216, + 0.00765431672334671, + 0.011212939396500587, + 0.005072754342108965, + 0.0022674475330859423, + 0.0018282082164660096, + 0.0015786343719810247, + 0.003177385311573744, + 0.003118152031674981, + 0.0033619471359997988, + 0.001914791064336896, + 0.002155221998691559, + 0.004284529481083155, + 0.00736670196056366, + 0.0036327787674963474, + 0.00983800645917654, + 0.006740313954651356, + 0.005368315149098635, + 0.007373080123215914, + 0.003416588297113776, + 0.011830107308924198, + 0.015085781924426556, + 0.005249698180705309, + 0.009523716755211353, + 0.008547810837626457, + 0.008869776502251625, + 0.009365668520331383, + 0.006623989436775446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.048645809292793274 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605992205440998 + ], + "opposite_logprob_tokens": [ + -4.141059875488281 + ], + "logprob_reward": -0.01605992205440998, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605992205440998, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12402013689279556, + 0.1913956254720688, + 0.03027934767305851, + 0.025771956890821457, + 0.0393514409661293, + 0.35568103194236755, + 0.05560286343097687, + 0.039899926632642746, + 0.06731091439723969, + 0.07068677246570587 + ], + "credit_map": [ + 0.013032278046011925, + 0.027403270825743675, + 0.008786393329501152, + 0.017335332930088043, + 0.02281392551958561, + 0.013350875116884708, + 0.010308588854968548, + 0.010989479720592499, + 0.01967662200331688, + 0.014274819754064083, + 0.009465297684073448, + 0.011119414120912552, + 0.07602465152740479, + 0.026231475174427032, + 0.01508728414773941, + 0.01951606757938862, + 0.006759638898074627, + 0.0024802819825708866, + 0.002469210186973214, + 0.0074277035892009735, + 0.0027911076322197914, + 0.003470450406894088, + 0.003167290473356843, + 0.0017136697424575686, + 0.007926286198198795, + 0.0029129041358828545, + 0.0024648529943078756, + 0.0016654119826853275, + 0.0023943244013935328, + 0.003384172450751066, + 0.001693426864221692, + 0.003330579726025462, + 0.0038047246634960175, + 0.005722117144614458, + 0.0042194765992462635, + 0.005666727665811777, + 0.0043426561169326305, + 0.006127669010311365, + 0.006227812264114618, + 0.0032402602955698967, + 0.03265422582626343, + 0.1137828528881073, + 0.013385951519012451, + 0.019262703135609627, + 0.018511276692152023, + 0.01817195490002632, + 0.11812803894281387, + 0.021784063428640366, + 0.008115675300359726, + 0.005769992712885141, + 0.005619939882308245, + 0.006819828879088163, + 0.008559091947972775, + 0.00846684817224741, + 0.005632681306451559, + 0.006618806626647711, + 0.0037466532085090876, + 0.004226844757795334, + 0.004570849239826202, + 0.008294973522424698, + 0.0044398317113518715, + 0.004033002536743879, + 0.0064967479556798935, + 0.004091027658432722, + 0.006022482179105282, + 0.009796354919672012, + 0.004946230910718441, + 0.011351332068443298, + 0.012259052135050297, + 0.009234240278601646, + 0.008759510703384876, + 0.00494172191247344, + 0.009475396946072578, + 0.013828754425048828, + 0.006959277670830488, + 0.008061191998422146, + 0.009144103154540062, + 0.006754850037395954, + 0.010856392793357372, + 0.005606807768344879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.01605992205440998 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762220611795783 + ], + "opposite_logprob_tokens": [ + -4.882622241973877 + ], + "logprob_reward": -0.00762220611795783, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762220611795783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19295647740364075, + 0.6116703748703003, + 0.0196373388171196, + 0.03296482935547829, + 0.03024977445602417, + 0.05295591801404953, + 0.01718878000974655, + 0.009403668344020844, + 0.013800942339003086, + 0.01917191594839096 + ], + "credit_map": [ + 0.017616096884012222, + 0.03511146456003189, + 0.028428632766008377, + 0.02807953767478466, + 0.025787191465497017, + 0.01600664108991623, + 0.027189528569579124, + 0.014737377874553204, + 0.13959859311580658, + 0.07090923935174942, + 0.0645274892449379, + 0.054763808846473694, + 0.07192657887935638, + 0.07124342769384384, + 0.08296278864145279, + 0.05573837831616402, + 0.00447741849347949, + 0.0018846724415197968, + 0.0015779174864292145, + 0.004401771817356348, + 0.0020251276437193155, + 0.0024206540547311306, + 0.001806222484447062, + 0.0010435518343001604, + 0.004319510422646999, + 0.005453416146337986, + 0.0021875023376196623, + 0.004478614777326584, + 0.003870750777423382, + 0.002693628892302513, + 0.006872483994811773, + 0.003088919911533594, + 0.0031116458121687174, + 0.004723922815173864, + 0.0035184361040592194, + 0.003464702283963561, + 0.0032653529196977615, + 0.005688520614057779, + 0.0038319772575050592, + 0.002645213156938553, + 0.01398425456136465, + 0.013159958645701408, + 0.002928669797256589, + 0.003316108603030443, + 0.004461381584405899, + 0.003312391461804509, + 0.00976833887398243, + 0.002024809131398797, + 0.001963886199519038, + 0.0023915134370326996, + 0.0016104733804240823, + 0.0029800713527947664, + 0.0017276874277740717, + 0.0017098913667723536, + 0.0033405849244445562, + 0.001464670174755156, + 0.0008205931517295539, + 0.0008825822733342648, + 0.001097255852073431, + 0.0015414775116369128, + 0.0015073207905516028, + 0.001285573118366301, + 0.0014253535773605108, + 0.0008435114286839962, + 0.0013738296693190932, + 0.0013633514754474163, + 0.0016939201159402728, + 0.0016657066298648715, + 0.0023825643584132195, + 0.0021049194037914276, + 0.001988454954698682, + 0.001228194683790207, + 0.0025106293614953756, + 0.004313173703849316, + 0.0016223221318796277, + 0.0021838604006916285, + 0.0028105247765779495, + 0.0019552765879780054, + 0.0022812450770288706, + 0.0014948822790756822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.00762220611795783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly zooming in, slightly overhead, medium shot focuses on a warm, round wooden table where three young friends sit intimately around a steaming hot pot. In the center, a vibrant red oil-based broth simmers vigorously, bubbling with dried chilies and Sichuan peppercorns. A milky, clear broth in the other half gently steams. Rich, white steam rises under the warm yellow light of a chandelier, creating a soft halo. A woman with long, curly brown hair and wearing a loose, off-white knit sweater, her face a healthy, steamy flush, smiles cheerfully. Across from her, a man wearing black-framed glasses and a dark blue hoodie gestures and excitedly talks, drawing laughter from his friends. As the camera slowly approaches, a woman carefully picks up a crisp, red and white marbled beef roll with a pair of long wooden chopsticks and steadily and calmly dips it into the boiling red oil. The moment the meat touches the broth, a string of fine bubbles rises around it, and the color of the fresh meat quickly darkens. The table is laden with a variety of fresh ingredients—emerald green vegetables, tender white tofu, and exquisitely presented meats. The entire scene is filled with the laughter of friends gathering together and the warm and cozy atmosphere of life, creating an atmosphere of supreme comfort and intimacy.", + "seed": 23275, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_2/index_3596_seed_23275_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025905710365623235, + "vlm_logprob": -0.025905710365623235, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.025905710365623235, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0197315476834774 + ], + "opposite_logprob_tokens": [ + -4.519731521606445 + ], + "logprob_reward": -0.0197315476834774, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0197315476834774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1693662852048874, + 0.09022549539804459, + 0.05465656891465187, + 0.12205882370471954, + 0.07186190038919449, + 0.05002693459391594, + 0.04506780952215195, + 0.10411692410707474, + 0.13938571512699127, + 0.1532335877418518 + ], + "credit_map": [ + 0.014413855038583279, + 0.013843613676726818, + 0.014306374825537205, + 0.029655680060386658, + 0.02109833061695099, + 0.021928060799837112, + 0.017233366146683693, + 0.036886993795633316, + 0.010830234736204147, + 0.009005993604660034, + 0.006295421626418829, + 0.011776171624660492, + 0.010609345510601997, + 0.016834983602166176, + 0.009070264175534248, + 0.015803074464201927, + 0.006142230238765478, + 0.007369489409029484, + 0.00446490291506052, + 0.006978632416576147, + 0.010869484394788742, + 0.007428786251693964, + 0.006923269014805555, + 0.00447977427393198, + 0.009236613288521767, + 0.01390825118869543, + 0.011733192950487137, + 0.02361438423395157, + 0.02861670032143593, + 0.011920562945306301, + 0.01099774707108736, + 0.012031376361846924, + 0.0074745831079781055, + 0.01362254936248064, + 0.006066274829208851, + 0.008378532715141773, + 0.016578862443566322, + 0.007836594246327877, + 0.007476629223674536, + 0.004427872132509947, + 0.0037232907488942146, + 0.005566582549363375, + 0.008403879590332508, + 0.007936757989227772, + 0.0063193547539412975, + 0.003799577709287405, + 0.00840464886277914, + 0.0058728414587676525, + 0.0060788108967244625, + 0.005303398706018925, + 0.008318840526044369, + 0.004985465668141842, + 0.008177447132766247, + 0.0034809941425919533, + 0.004162443336099386, + 0.004560407716780901, + 0.027872323989868164, + 0.012229330837726593, + 0.00852870475500822, + 0.014222461730241776, + 0.009878941811621189, + 0.013695470057427883, + 0.007316829636693001, + 0.010372853837907314, + 0.010765510611236095, + 0.032698675990104675, + 0.01121275033801794, + 0.031967755407094955, + 0.01528187096118927, + 0.013980011455714703, + 0.013651560992002487, + 0.00982757844030857, + 0.03413904830813408, + 0.02166895568370819, + 0.014679000712931156, + 0.01915145479142666, + 0.013404931873083115, + 0.02200496755540371, + 0.013076310977339745, + 0.015108922496438026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0197315476834774 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022168601863086224 + ], + "opposite_logprob_tokens": [ + -6.127216815948486 + ], + "logprob_reward": -0.0022168601863086224, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022168601863086224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07186393439769745, + 0.11522264778614044, + 0.03756439685821533, + 0.13525593280792236, + 0.04249629005789757, + 0.03146519884467125, + 0.025737008079886436, + 0.04117085039615631, + 0.1600470095872879, + 0.3391767144203186 + ], + "credit_map": [ + 0.007579265162348747, + 0.007994381710886955, + 0.005958793684840202, + 0.010184669867157936, + 0.010110076516866684, + 0.01153649017214775, + 0.008444137871265411, + 0.010056122206151485, + 0.009357835166156292, + 0.009367084130644798, + 0.006091054528951645, + 0.011221867986023426, + 0.009320005774497986, + 0.027834605425596237, + 0.009944272227585316, + 0.032085925340652466, + 0.005052364431321621, + 0.005000629927963018, + 0.0033962081652134657, + 0.004073827061802149, + 0.005909455008804798, + 0.0063966307789087296, + 0.0045687551610171795, + 0.0031665258575230837, + 0.010310674086213112, + 0.0140659986063838, + 0.010993234813213348, + 0.027101071551442146, + 0.033347245305776596, + 0.013069125823676586, + 0.012974089942872524, + 0.013394495472311974, + 0.003995007369667292, + 0.007871896028518677, + 0.004919837228953838, + 0.005482620559632778, + 0.007644255645573139, + 0.004129298962652683, + 0.004628156777471304, + 0.003825215622782707, + 0.002586750080808997, + 0.003540505887940526, + 0.005888359155505896, + 0.0034151810687035322, + 0.005185303278267384, + 0.0022663376294076443, + 0.0045834993943572044, + 0.003999264445155859, + 0.0033055932726711035, + 0.003159801010042429, + 0.0048255156725645065, + 0.0030213219579309225, + 0.0023897262290120125, + 0.0034303779248148203, + 0.0031342555303126574, + 0.002470415784046054, + 0.00445631705224514, + 0.005662273149937391, + 0.00531661557033658, + 0.008719578385353088, + 0.004140839446336031, + 0.004402888473123312, + 0.004774448927491903, + 0.0036978882271796465, + 0.00999821349978447, + 0.05646057799458504, + 0.010713649913668633, + 0.03261745721101761, + 0.021631501615047455, + 0.007605841848999262, + 0.012600868940353394, + 0.008418895304203033, + 0.07026834040880203, + 0.027565013617277145, + 0.039314765483140945, + 0.048094287514686584, + 0.022999711334705353, + 0.04490206390619278, + 0.05168193578720093, + 0.034350596368312836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.0022168601863086224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898029685020447 + ], + "opposite_logprob_tokens": [ + -2.5789802074432373 + ], + "logprob_reward": -0.07898029685020447, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898029685020447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2804037034511566, + 0.10904627293348312, + 0.10132050514221191, + 0.059181567281484604, + 0.04705619066953659, + 0.04238348454236984, + 0.03418358415365219, + 0.056113675236701965, + 0.12948709726333618, + 0.14082390069961548 + ], + "credit_map": [ + 0.025032319128513336, + 0.03127892687916756, + 0.019187236204743385, + 0.025463860481977463, + 0.050709135830402374, + 0.06651225686073303, + 0.021875720471143723, + 0.0403442420065403, + 0.008546097204089165, + 0.010021956637501717, + 0.008233395405113697, + 0.018544161692261696, + 0.012491810135543346, + 0.01719292439520359, + 0.018252551555633545, + 0.015763379633426666, + 0.010845780372619629, + 0.010661324486136436, + 0.0073617203161120415, + 0.01499304361641407, + 0.021895483136177063, + 0.014048337936401367, + 0.01359286904335022, + 0.00792195089161396, + 0.004267100244760513, + 0.006099278572946787, + 0.004740631673485041, + 0.013448210433125496, + 0.011521082371473312, + 0.00480927899479866, + 0.007591204717755318, + 0.006704779341816902, + 0.004204707685858011, + 0.008203055709600449, + 0.004530375357717276, + 0.006483488716185093, + 0.007758231367915869, + 0.0053456127643585205, + 0.005741352215409279, + 0.004789367318153381, + 0.006166380364447832, + 0.006442591082304716, + 0.006012321449816227, + 0.0037408138159662485, + 0.0071871899999678135, + 0.004149833228439093, + 0.003946638200432062, + 0.004737717565149069, + 0.004938534926623106, + 0.003910133149474859, + 0.004241248592734337, + 0.003421847242861986, + 0.006450823042541742, + 0.005179687403142452, + 0.003065514611080289, + 0.0029757965821772814, + 0.01044449396431446, + 0.009592987596988678, + 0.007390324492007494, + 0.007424443494528532, + 0.006805733311921358, + 0.004460962954908609, + 0.0051100668497383595, + 0.004884661640971899, + 0.00951325986534357, + 0.04524359107017517, + 0.008269758895039558, + 0.013585399836301804, + 0.028378257527947426, + 0.009065271355211735, + 0.007105386815965176, + 0.008326178416609764, + 0.029882313683629036, + 0.012972955591976643, + 0.015613879077136517, + 0.0158864613622427, + 0.009877575561404228, + 0.021213814616203308, + 0.022190600633621216, + 0.013186299242079258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.07898029685020447 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534411624073982 + ], + "opposite_logprob_tokens": [ + -4.3875346183776855 + ], + "logprob_reward": -0.012534411624073982, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534411624073982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11300910264253616, + 0.07499115914106369, + 0.03187399357557297, + 0.15643659234046936, + 0.03168390691280365, + 0.18906056880950928, + 0.060005106031894684, + 0.1075119823217392, + 0.11986112594604492, + 0.11556646227836609 + ], + "credit_map": [ + 0.010487259365618229, + 0.011038287542760372, + 0.007806714624166489, + 0.013699139468371868, + 0.018087072297930717, + 0.024624129757285118, + 0.011489289812743664, + 0.015777213498950005, + 0.0054499246180057526, + 0.005726421717554331, + 0.005218593869358301, + 0.010607963427901268, + 0.006588413845747709, + 0.014261524192988873, + 0.00948941707611084, + 0.017648909240961075, + 0.0037674270570278168, + 0.004076347686350346, + 0.002726717619225383, + 0.004605709575116634, + 0.006901134736835957, + 0.004023286513984203, + 0.0034387961495667696, + 0.0023345788940787315, + 0.012252278625965118, + 0.014162730425596237, + 0.01223030686378479, + 0.0372413732111454, + 0.03166290372610092, + 0.017606502398848534, + 0.016146376729011536, + 0.015134110115468502, + 0.0029266958590596914, + 0.005890298634767532, + 0.0034823198802769184, + 0.0038475512992590666, + 0.005200375337153673, + 0.0033983022440224886, + 0.004193189088255167, + 0.0027451757341623306, + 0.016605110839009285, + 0.021288251504302025, + 0.02566869929432869, + 0.025380324572324753, + 0.035185981541872025, + 0.01593063771724701, + 0.024480409920215607, + 0.02452118508517742, + 0.005307057406753302, + 0.01035913173109293, + 0.006748113315552473, + 0.007272429298609495, + 0.01064205076545477, + 0.008450944907963276, + 0.005708715412765741, + 0.005516668315976858, + 0.024730604141950607, + 0.01674470119178295, + 0.008592075668275356, + 0.014374818652868271, + 0.009998660534620285, + 0.01569291763007641, + 0.006930070463567972, + 0.010448139160871506, + 0.011242806911468506, + 0.04293317347764969, + 0.00618335697799921, + 0.01528310589492321, + 0.023570777848362923, + 0.006187232211232185, + 0.007442767266184092, + 0.007017902564257383, + 0.02692989446222782, + 0.01001927349716425, + 0.015200207941234112, + 0.013647147454321384, + 0.008847377263009548, + 0.017292719334363937, + 0.012824580073356628, + 0.010805271565914154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.012534411624073982 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065435484051704 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.016065435484051704, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065435484051704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08925558626651764, + 0.3890911638736725, + 0.03404621779918671, + 0.22710084915161133, + 0.06075612083077431, + 0.013134823180735111, + 0.020476045086979866, + 0.03997097164392471, + 0.06992177665233612, + 0.056246399879455566 + ], + "credit_map": [ + 0.010575667023658752, + 0.006889231503009796, + 0.005211282521486282, + 0.011430836282670498, + 0.017364786937832832, + 0.015733135864138603, + 0.009320457465946674, + 0.012730199843645096, + 0.03842978551983833, + 0.02938954532146454, + 0.016043879091739655, + 0.022059034556150436, + 0.037215471267700195, + 0.10984264314174652, + 0.019955478608608246, + 0.11615535616874695, + 0.006685527041554451, + 0.004041555803269148, + 0.0016933126607909799, + 0.003184311091899872, + 0.0074985092505812645, + 0.004506545141339302, + 0.0034090792760252953, + 0.0030273841693997383, + 0.016965411603450775, + 0.023805085569620132, + 0.017141422256827354, + 0.0463024266064167, + 0.05687302350997925, + 0.02205800823867321, + 0.02157827652990818, + 0.022377226501703262, + 0.006092254538089037, + 0.010825038887560368, + 0.005116526503115892, + 0.00713868485763669, + 0.012941124849021435, + 0.007290612440556288, + 0.00665754871442914, + 0.004694338422268629, + 0.001010824111290276, + 0.0014263701159507036, + 0.0012235388858243823, + 0.001666884170845151, + 0.002392580034211278, + 0.001434824662283063, + 0.001470857416279614, + 0.002508945995941758, + 0.0018430401105433702, + 0.0019137487979605794, + 0.0026917282957583666, + 0.002435836475342512, + 0.004442363046109676, + 0.0032308034133166075, + 0.0016187855508178473, + 0.002299739746376872, + 0.010175530798733234, + 0.004358022008091211, + 0.0042085968889296055, + 0.005266088992357254, + 0.004401765298098326, + 0.004918295424431562, + 0.002589165698736906, + 0.004053512588143349, + 0.008006511256098747, + 0.021128928288817406, + 0.003915863577276468, + 0.005804555956274271, + 0.015267931856215, + 0.00697048706933856, + 0.0041398461908102036, + 0.0046876599080860615, + 0.010912032797932625, + 0.0050971717573702335, + 0.007417008746415377, + 0.006849511526525021, + 0.004848169162869453, + 0.007491649128496647, + 0.008190395310521126, + 0.005440467037260532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.016065435484051704 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly zooming in, slightly overhead, medium shot focuses on a warm, round wooden table where three young friends sit intimately around a steaming hot pot. In the center, a vibrant red oil-based broth simmers vigorously, bubbling with dried chilies and Sichuan peppercorns. A milky, clear broth in the other half gently steams. Rich, white steam rises under the warm yellow light of a chandelier, creating a soft halo. A woman with long, curly brown hair and wearing a loose, off-white knit sweater, her face a healthy, steamy flush, smiles cheerfully. Across from her, a man wearing black-framed glasses and a dark blue hoodie gestures and excitedly talks, drawing laughter from his friends. As the camera slowly approaches, a woman carefully picks up a crisp, red and white marbled beef roll with a pair of long wooden chopsticks and steadily and calmly dips it into the boiling red oil. The moment the meat touches the broth, a string of fine bubbles rises around it, and the color of the fresh meat quickly darkens. The table is laden with a variety of fresh ingredients—emerald green vegetables, tender white tofu, and exquisitely presented meats. The entire scene is filled with the laughter of friends gathering together and the warm and cozy atmosphere of life, creating an atmosphere of supreme comfort and intimacy.", + "seed": 23276, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_2/index_3596_seed_23276_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02259177519008517, + "vlm_logprob": -0.02259177519008517, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.02259177519008517, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03216974809765816 + ], + "opposite_logprob_tokens": [ + -4.157169818878174 + ], + "logprob_reward": -0.03216974809765816, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.03216974809765816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07562364637851715, + 0.030524441972374916, + 0.40345677733421326, + 0.03750649094581604, + 0.0412529855966568, + 0.11082223802804947, + 0.14512145519256592, + 0.025835344567894936, + 0.07829514145851135, + 0.051561467349529266 + ], + "credit_map": [ + 0.005704144481569529, + 0.00721747986972332, + 0.0076329209841787815, + 0.008568421006202698, + 0.021857691928744316, + 0.010545967146754265, + 0.009020315483212471, + 0.005076712928712368, + 0.0038921074010431767, + 0.003760314779356122, + 0.00413168128579855, + 0.00479091377928853, + 0.0045245056971907616, + 0.0034163130912929773, + 0.0029087658040225506, + 0.0030998431611806154, + 0.016881903633475304, + 0.012370540760457516, + 0.14705148339271545, + 0.016018832102417946, + 0.15391089022159576, + 0.027265390381217003, + 0.013225367292761803, + 0.016732383519411087, + 0.003119588829576969, + 0.005536837037652731, + 0.004199162591248751, + 0.0028228159062564373, + 0.008138707838952541, + 0.005850072950124741, + 0.005124124698340893, + 0.00271518318913877, + 0.006843230687081814, + 0.0037185654509812593, + 0.004636771511286497, + 0.0034556088503450155, + 0.007606951519846916, + 0.0049066729843616486, + 0.0054823774844408035, + 0.004602810833603144, + 0.027917979285120964, + 0.00469356682151556, + 0.008981416933238506, + 0.006981253158301115, + 0.01137599628418684, + 0.03648265823721886, + 0.008001548238098621, + 0.006387836299836636, + 0.014153634198009968, + 0.009429470635950565, + 0.01875966414809227, + 0.03309573233127594, + 0.019085457548499107, + 0.010240620002150536, + 0.02451745793223381, + 0.01583942025899887, + 0.0027225958183407784, + 0.0029426070395857096, + 0.0028418898582458496, + 0.0048053814098238945, + 0.0030124606564641, + 0.002456334186717868, + 0.0027108704671263695, + 0.0043432097882032394, + 0.007275440730154514, + 0.0051721371710300446, + 0.0249316543340683, + 0.008716287091374397, + 0.0051144822500646114, + 0.005756620783358812, + 0.006778365932404995, + 0.014550167135894299, + 0.004587304778397083, + 0.005001832731068134, + 0.006569996941834688, + 0.006752646062523127, + 0.005956300999969244, + 0.014097060076892376, + 0.0037226637359708548, + 0.004873666912317276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.03216974809765816 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00527632562443614 + ], + "opposite_logprob_tokens": [ + -5.255276203155518 + ], + "logprob_reward": -0.00527632562443614, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00527632562443614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031263839453458786, + 0.01757574826478958, + 0.2560121417045593, + 0.02871634066104889, + 0.025313233956694603, + 0.16849786043167114, + 0.10570330172777176, + 0.028475668281316757, + 0.3019651174545288, + 0.036476753652095795 + ], + "credit_map": [ + 0.0027430050540715456, + 0.003285621525719762, + 0.0025287852622568607, + 0.0034129128325730562, + 0.0062807463109493256, + 0.005302408244460821, + 0.005336008034646511, + 0.002374354749917984, + 0.002216741442680359, + 0.0020032664760947227, + 0.0018299143994227052, + 0.001666969619691372, + 0.0029909852892160416, + 0.002295470330864191, + 0.002117277355864644, + 0.0024551230017095804, + 0.012259876355528831, + 0.008914443664252758, + 0.09122712910175323, + 0.010042704641819, + 0.09866945445537567, + 0.017683569341897964, + 0.008230810984969139, + 0.008984167128801346, + 0.003589809639379382, + 0.004224421922117472, + 0.0038954932242631912, + 0.0016775232506915927, + 0.007053531240671873, + 0.0038527792785316706, + 0.002848101779818535, + 0.0015746813733130693, + 0.0028666702564805746, + 0.002160419011488557, + 0.004123149439692497, + 0.003154504345729947, + 0.002199979266151786, + 0.0028139385394752026, + 0.0047042532823979855, + 0.003290320048108697, + 0.04845115914940834, + 0.00761769525706768, + 0.012733463197946548, + 0.007722090929746628, + 0.01569799892604351, + 0.058092162013053894, + 0.009898658841848373, + 0.008284608833491802, + 0.009363709017634392, + 0.006532192695885897, + 0.013384862802922726, + 0.02370074950158596, + 0.016710415482521057, + 0.007905622012913227, + 0.017409637570381165, + 0.010696113109588623, + 0.003266168525442481, + 0.0030673430301249027, + 0.004189784172922373, + 0.004100772086530924, + 0.0027130316011607647, + 0.002591862343251705, + 0.0038289029616862535, + 0.004717804957181215, + 0.03184221312403679, + 0.01459289900958538, + 0.12153229862451553, + 0.02199430949985981, + 0.013630824163556099, + 0.016797775402665138, + 0.014960620552301407, + 0.06661418080329895, + 0.002859600819647312, + 0.008332493714988232, + 0.00400583678856492, + 0.006143988110125065, + 0.004930923227220774, + 0.0036511605139821768, + 0.003390825819224119, + 0.0031619274523109198 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.00527632562443614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810085728764534 + ], + "opposite_logprob_tokens": [ + -3.2881009578704834 + ], + "logprob_reward": -0.03810085728764534, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810085728764534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08696985244750977, + 0.06347668170928955, + 0.5791575908660889, + 0.018747597932815552, + 0.010608727112412453, + 0.029898686334490776, + 0.13985022902488708, + 0.04295359551906586, + 0.01587560586631298, + 0.012461411766707897 + ], + "credit_map": [ + 0.009473144076764584, + 0.007402009330689907, + 0.008437611162662506, + 0.008590683341026306, + 0.01678975112736225, + 0.01757151447236538, + 0.01241231244057417, + 0.006292824633419514, + 0.010119330137968063, + 0.004364484455436468, + 0.006522944197058678, + 0.0049281339161098, + 0.010145318694412708, + 0.007194075267761946, + 0.006786681711673737, + 0.013415717519819736, + 0.024132536724209785, + 0.015863284468650818, + 0.2183363139629364, + 0.025205928832292557, + 0.20936666429042816, + 0.03574046120047569, + 0.01886221021413803, + 0.03165019303560257, + 0.0018278311472386122, + 0.0017954648938030005, + 0.0011876847129315138, + 0.001489777467213571, + 0.0039858585223555565, + 0.0024345179554075003, + 0.004864365793764591, + 0.0011620963923633099, + 0.0013952431036159396, + 0.0009132368140853941, + 0.0011159551795572042, + 0.0014777189353480935, + 0.0013760639121755958, + 0.0012226159451529384, + 0.0020113929640501738, + 0.001096499734558165, + 0.007845205254852772, + 0.0017647743225097656, + 0.003798333229497075, + 0.0018694645259529352, + 0.002662045182660222, + 0.00873600784689188, + 0.0015259102219715714, + 0.001696947729215026, + 0.013865312561392784, + 0.008051221258938313, + 0.017907749861478806, + 0.03165159747004509, + 0.018408726900815964, + 0.010118760168552399, + 0.02360331080853939, + 0.016243554651737213, + 0.004593818448483944, + 0.005365588236600161, + 0.004936700686812401, + 0.0041139195673167706, + 0.005189551040530205, + 0.004175752401351929, + 0.0035124621354043484, + 0.011065798811614513, + 0.001537951873615384, + 0.0014260574243962765, + 0.004736734554171562, + 0.0016285823658108711, + 0.0014694276032969356, + 0.0008280526963062584, + 0.0014100372791290283, + 0.002838761080056429, + 0.002547701122239232, + 0.0011409888975322247, + 0.0016494408482685685, + 0.0018131053075194359, + 0.0012267476413398981, + 0.002095828065648675, + 0.0008442719117738307, + 0.0011433283798396587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.03810085728764534 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071322485804558 + ], + "opposite_logprob_tokens": [ + -4.51107120513916 + ], + "logprob_reward": -0.011071322485804558, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071322485804558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03335500508546829, + 0.017990147694945335, + 0.011502159759402275, + 0.009229746647179127, + 0.054307155311107635, + 0.08460326492786407, + 0.7174944281578064, + 0.008960474282503128, + 0.04990966245532036, + 0.012647930532693863 + ], + "credit_map": [ + 0.0032546238508075476, + 0.004072289448231459, + 0.004762463737279177, + 0.0037778730038553476, + 0.008709384128451347, + 0.003329699160531163, + 0.003669778583571315, + 0.0017788949189707637, + 0.002877382794395089, + 0.0020269553642719984, + 0.0026857759803533554, + 0.0015896615805104375, + 0.002825099742040038, + 0.0016059556510299444, + 0.0019183135591447353, + 0.0024610029067844152, + 0.001846078783273697, + 0.001115080201998353, + 0.001419372158125043, + 0.00101698306389153, + 0.0016461912309750915, + 0.0016069021075963974, + 0.0010964494431391358, + 0.0017551032360643148, + 0.001124444417655468, + 0.0011425359407439828, + 0.001703915768302977, + 0.0008450833847746253, + 0.000985293765552342, + 0.0011195959523320198, + 0.0016869802493602037, + 0.0006218976341187954, + 0.012703864835202694, + 0.00478987256065011, + 0.005540295969694853, + 0.004348314367234707, + 0.0052281212992966175, + 0.00782664492726326, + 0.009408559650182724, + 0.004461484961211681, + 0.01907353475689888, + 0.0058815861120820045, + 0.010305576026439667, + 0.006860287860035896, + 0.007585036568343639, + 0.021068468689918518, + 0.007835890166461468, + 0.005992883816361427, + 0.06511779129505157, + 0.04091190546751022, + 0.09103010594844818, + 0.1619308441877365, + 0.1108463704586029, + 0.05372055247426033, + 0.11983776837587357, + 0.0740991085767746, + 0.0011074681533500552, + 0.000966304331086576, + 0.0009290823363699019, + 0.0011624798644334078, + 0.0010999691439792514, + 0.0010041932109743357, + 0.0010918129701167345, + 0.0015991637483239174, + 0.005113551858812571, + 0.0033706072717905045, + 0.018911922350525856, + 0.003204713575541973, + 0.002996642142534256, + 0.003075510961934924, + 0.003134883474558592, + 0.010101832449436188, + 0.0011892722686752677, + 0.0014181127771735191, + 0.0019217270892113447, + 0.0022208953741937876, + 0.0014857831411063671, + 0.002405076287686825, + 0.0011645011836662889, + 0.0008425631676800549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.011071322485804558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026340622454881668 + ], + "opposite_logprob_tokens": [ + -3.6513407230377197 + ], + "logprob_reward": -0.026340622454881668, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026340622454881668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06019891798496246, + 0.025983959436416626, + 0.32559865713119507, + 0.01749490387737751, + 0.03990683704614639, + 0.14687226712703705, + 0.1530686318874359, + 0.01392817310988903, + 0.18435636162757874, + 0.03259136527776718 + ], + "credit_map": [ + 0.005396946333348751, + 0.003984321374446154, + 0.006862173788249493, + 0.006881142035126686, + 0.009663608856499195, + 0.007978934794664383, + 0.012937941588461399, + 0.006493841763585806, + 0.003912918269634247, + 0.002081130398437381, + 0.004268110264092684, + 0.0026775230653584003, + 0.0037376529071480036, + 0.002840847708284855, + 0.0024982518516480923, + 0.003967521246522665, + 0.012957731261849403, + 0.01069727260619402, + 0.12046726047992706, + 0.013346095569431782, + 0.12199480831623077, + 0.021141894161701202, + 0.010699928738176823, + 0.014293629676103592, + 0.003275661962106824, + 0.0019313804805278778, + 0.0021519733127206564, + 0.0016720748972147703, + 0.0020924019627273083, + 0.0024797040969133377, + 0.0022425511851906776, + 0.0016491531860083342, + 0.0048103174194693565, + 0.003153224941343069, + 0.004424665588885546, + 0.005617461632937193, + 0.004080228973180056, + 0.004486465826630592, + 0.010203089565038681, + 0.003131381468847394, + 0.034963205456733704, + 0.00972118228673935, + 0.012637664563953876, + 0.009701171889901161, + 0.01467163860797882, + 0.04448080435395241, + 0.013145700097084045, + 0.007550879381597042, + 0.01443265751004219, + 0.0091091962531209, + 0.019681284204125404, + 0.035763777792453766, + 0.02196062169969082, + 0.010775018483400345, + 0.026057898998260498, + 0.015288173221051693, + 0.002108854241669178, + 0.0013282689033076167, + 0.0021896460093557835, + 0.0015462947776541114, + 0.0015635864110663533, + 0.0012904260074719787, + 0.0013133782194927335, + 0.002587716793641448, + 0.019374065101146698, + 0.007925828918814659, + 0.07671311497688293, + 0.010542782954871655, + 0.007695070002228022, + 0.01031462848186493, + 0.007789382711052895, + 0.04400146007537842, + 0.0034995435271412134, + 0.0024898203555494547, + 0.007549729663878679, + 0.0021877626422792673, + 0.00528810964897275, + 0.0058724042028188705, + 0.0024115126579999924, + 0.00329247722402215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.026340622454881668 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly zooming in, slightly overhead, medium shot focuses on a warm, round wooden table where three young friends sit intimately around a steaming hot pot. In the center, a vibrant red oil-based broth simmers vigorously, bubbling with dried chilies and Sichuan peppercorns. A milky, clear broth in the other half gently steams. Rich, white steam rises under the warm yellow light of a chandelier, creating a soft halo. A woman with long, curly brown hair and wearing a loose, off-white knit sweater, her face a healthy, steamy flush, smiles cheerfully. Across from her, a man wearing black-framed glasses and a dark blue hoodie gestures and excitedly talks, drawing laughter from his friends. As the camera slowly approaches, a woman carefully picks up a crisp, red and white marbled beef roll with a pair of long wooden chopsticks and steadily and calmly dips it into the boiling red oil. The moment the meat touches the broth, a string of fine bubbles rises around it, and the color of the fresh meat quickly darkens. The table is laden with a variety of fresh ingredients—emerald green vegetables, tender white tofu, and exquisitely presented meats. The entire scene is filled with the laughter of friends gathering together and the warm and cozy atmosphere of life, creating an atmosphere of supreme comfort and intimacy.", + "seed": 23277, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_2/index_3596_seed_23277_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.026431931694969536, + "vlm_logprob": -0.026431931694969536, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.026431931694969536, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03664174675941467 + ], + "opposite_logprob_tokens": [ + -4.161641597747803 + ], + "logprob_reward": -0.03664174675941467, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.03664174675941467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16151362657546997, + 0.07829957455396652, + 0.04372590407729149, + 0.11470097303390503, + 0.08546943962574005, + 0.23807673156261444, + 0.09176377952098846, + 0.038023803383111954, + 0.07144124060869217, + 0.0769849494099617 + ], + "credit_map": [ + 0.015791069716215134, + 0.012296031229197979, + 0.013676906935870647, + 0.017492497339844704, + 0.028829332441091537, + 0.01408936083316803, + 0.04126114770770073, + 0.018077276647090912, + 0.007282326463609934, + 0.01003007311373949, + 0.010047373361885548, + 0.009755002334713936, + 0.011788575910031796, + 0.0056021567434072495, + 0.01466280221939087, + 0.009131251834332943, + 0.003639820497483015, + 0.0050941351801157, + 0.0034598263446241617, + 0.004958766978234053, + 0.006443301681429148, + 0.004141887184232473, + 0.011977237649261951, + 0.004010925069451332, + 0.010343295522034168, + 0.03675694763660431, + 0.005869310349225998, + 0.007573586888611317, + 0.03355531394481659, + 0.006303178612142801, + 0.00874497089534998, + 0.005554363131523132, + 0.007550051901489496, + 0.008461927063763142, + 0.010426691733300686, + 0.007399839349091053, + 0.013378352858126163, + 0.01276961900293827, + 0.010970029979944229, + 0.014512920752167702, + 0.06963174790143967, + 0.04149830713868141, + 0.019316554069519043, + 0.023605715483427048, + 0.028340257704257965, + 0.017251137644052505, + 0.019452422857284546, + 0.01898057572543621, + 0.00854333397001028, + 0.021183179691433907, + 0.005523835774511099, + 0.006392371840775013, + 0.027942094951868057, + 0.0071776085533201694, + 0.009948336519300938, + 0.005053014028817415, + 0.008306412026286125, + 0.005903882905840874, + 0.004728102125227451, + 0.0036885614972561598, + 0.005813601892441511, + 0.0028677652589976788, + 0.0038969407323747873, + 0.0028185376431792974, + 0.007500761188566685, + 0.007550735492259264, + 0.0072379810735583305, + 0.010158219374716282, + 0.011157999746501446, + 0.009463128633797169, + 0.012408608570694923, + 0.0059638069942593575, + 0.011971961706876755, + 0.011971482075750828, + 0.005347676575183868, + 0.0076513211242854595, + 0.013077572919428349, + 0.009143177419900894, + 0.00872241985052824, + 0.009099338203668594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.03664174675941467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032164526637643576 + ], + "opposite_logprob_tokens": [ + -5.75321626663208 + ], + "logprob_reward": -0.0032164526637643576, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032164526637643576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1332845538854599, + 0.1303674280643463, + 0.07082843035459518, + 0.05356232449412346, + 0.03902215138077736, + 0.06514614820480347, + 0.04207624867558479, + 0.03118305839598179, + 0.09478632360696793, + 0.33974334597587585 + ], + "credit_map": [ + 0.010574142448604107, + 0.014070946723222733, + 0.01082614902406931, + 0.014011829160153866, + 0.015430323779582977, + 0.011362925171852112, + 0.03542793542146683, + 0.021580301225185394, + 0.012098629958927631, + 0.012586784549057484, + 0.01804828643798828, + 0.018642595037817955, + 0.017443375661969185, + 0.010429459623992443, + 0.030036834999918938, + 0.011081459000706673, + 0.006092609837651253, + 0.008256005123257637, + 0.005943967029452324, + 0.00974102970212698, + 0.007654278073459864, + 0.00550107704475522, + 0.01911327801644802, + 0.008526184596121311, + 0.005325938109308481, + 0.015178674831986427, + 0.003098202170804143, + 0.005970296449959278, + 0.013131140731275082, + 0.003713487181812525, + 0.004690167959779501, + 0.002454417757689953, + 0.0064114052802324295, + 0.004793399944901466, + 0.003881720593199134, + 0.0030256779864430428, + 0.005799666978418827, + 0.004432414658367634, + 0.006944787222892046, + 0.0037330775521695614, + 0.017759375274181366, + 0.009738428518176079, + 0.005047934129834175, + 0.0073891268111765385, + 0.008678855374455452, + 0.00646752305328846, + 0.004990275949239731, + 0.0050746300257742405, + 0.005422168876975775, + 0.007234393153339624, + 0.0025329815689474344, + 0.00421450799331069, + 0.010954082012176514, + 0.003601094475015998, + 0.0043461620807647705, + 0.003770858747884631, + 0.005235624965280294, + 0.006128416862338781, + 0.0029799712356179953, + 0.003544574836269021, + 0.004807611461728811, + 0.003330698236823082, + 0.002982216188684106, + 0.0021739474032074213, + 0.01019431184977293, + 0.011234590783715248, + 0.007414625957608223, + 0.014151488430798054, + 0.020016783848404884, + 0.008009771816432476, + 0.014366320334374905, + 0.009398436173796654, + 0.04363066703081131, + 0.042410459369421005, + 0.028560668230056763, + 0.032289352267980576, + 0.04591275751590729, + 0.03246413543820381, + 0.05766403675079346, + 0.05681126192212105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.0032164526637643576 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204671785235405 + ], + "opposite_logprob_tokens": [ + -2.812046766281128 + ], + "logprob_reward": -0.06204671785235405, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204671785235405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21625441312789917, + 0.14511917531490326, + 0.07542136311531067, + 0.15295027196407318, + 0.05940984934568405, + 0.12633852660655975, + 0.05347447097301483, + 0.030561769381165504, + 0.0868600457906723, + 0.05361009016633034 + ], + "credit_map": [ + 0.02128472924232483, + 0.02299576811492443, + 0.022156480699777603, + 0.017800623551011086, + 0.02883005701005459, + 0.02393043413758278, + 0.028152016922831535, + 0.05110432207584381, + 0.01879066787660122, + 0.02074410207569599, + 0.01626351848244667, + 0.01276322826743126, + 0.029378505423665047, + 0.013720471411943436, + 0.015625787898898125, + 0.01783289574086666, + 0.0074460203759372234, + 0.007438719738274813, + 0.00828342605382204, + 0.010246374644339085, + 0.010279590263962746, + 0.00649168761447072, + 0.015932351350784302, + 0.00930318795144558, + 0.013759802095592022, + 0.048406533896923065, + 0.008562566712498665, + 0.01676676981151104, + 0.03240662068128586, + 0.01481029111891985, + 0.012003669515252113, + 0.006234000902622938, + 0.012158215045928955, + 0.011019482277333736, + 0.005446004215627909, + 0.004318034276366234, + 0.010006031021475792, + 0.007364416029304266, + 0.0044484566897153854, + 0.004649213049560785, + 0.03171814605593681, + 0.01894058845937252, + 0.008711064234375954, + 0.010259980335831642, + 0.02396836318075657, + 0.013980873860418797, + 0.008516591042280197, + 0.01024292130023241, + 0.005471688695251942, + 0.012929435819387436, + 0.002769396873190999, + 0.003921537194401026, + 0.018147755414247513, + 0.0035366856027394533, + 0.0037483712658286095, + 0.002949601737782359, + 0.004553241189569235, + 0.004420043434947729, + 0.003090995829552412, + 0.004188387654721737, + 0.004106166772544384, + 0.0026832374278455973, + 0.005075203254818916, + 0.002444495679810643, + 0.010395891033113003, + 0.01289840042591095, + 0.009496280923485756, + 0.009504751302301884, + 0.01447151880711317, + 0.015646401792764664, + 0.0067214504815638065, + 0.007725355215370655, + 0.010070874355733395, + 0.006673307158052921, + 0.006539235357195139, + 0.004973634146153927, + 0.008182108402252197, + 0.003157753963023424, + 0.0093908179551363, + 0.004622360225766897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.06204671785235405 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189071021974087 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.014189071021974087, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189071021974087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16105684638023376, + 0.09938626736402512, + 0.04570198431611061, + 0.09595363587141037, + 0.0593680664896965, + 0.18795454502105713, + 0.09765513241291046, + 0.05461110547184944, + 0.07109961658716202, + 0.12721280753612518 + ], + "credit_map": [ + 0.01686544343829155, + 0.0154342669993639, + 0.015445969998836517, + 0.013801568187773228, + 0.019787734374403954, + 0.01482664979994297, + 0.03241593763232231, + 0.03247926011681557, + 0.013233215548098087, + 0.0152931222692132, + 0.006815513130277395, + 0.007051850203424692, + 0.028464967384934425, + 0.007451427634805441, + 0.0076978737488389015, + 0.013378291390836239, + 0.0055624074302613735, + 0.005910779349505901, + 0.004268810618668795, + 0.006227415055036545, + 0.00993162952363491, + 0.003823754843324423, + 0.005462799221277237, + 0.004514383152127266, + 0.007369332481175661, + 0.028410449624061584, + 0.006794675253331661, + 0.00945302750915289, + 0.025866994634270668, + 0.005033859051764011, + 0.007440210320055485, + 0.00558508513495326, + 0.008020554669201374, + 0.007951400242745876, + 0.008430778980255127, + 0.005354462657123804, + 0.00872036349028349, + 0.009896193630993366, + 0.006458689924329519, + 0.004535618703812361, + 0.03729557618498802, + 0.028823386877775192, + 0.012924554757773876, + 0.016129933297634125, + 0.03883606195449829, + 0.021988948807120323, + 0.017118778079748154, + 0.01483729388564825, + 0.010693782940506935, + 0.02063179761171341, + 0.005832490045577288, + 0.007121950853615999, + 0.02745506353676319, + 0.006567242555320263, + 0.010466577485203743, + 0.008886227384209633, + 0.008075091987848282, + 0.009458694607019424, + 0.004698215518146753, + 0.007325595710426569, + 0.007977333851158619, + 0.005001198500394821, + 0.00609477236866951, + 0.005980199668556452, + 0.009212116710841656, + 0.011858243495225906, + 0.006294149439781904, + 0.007734610233455896, + 0.012073582969605923, + 0.007527545094490051, + 0.008846057578921318, + 0.007553310599178076, + 0.013537529855966568, + 0.018766479566693306, + 0.015558134764432907, + 0.011941525153815746, + 0.014798523858189583, + 0.015508128330111504, + 0.02180573344230652, + 0.015296743251383305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.014189071021974087 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a medium shot that slowly zooms in from a slightly overhead angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a warm, round wooden table with a steaming hot pot containing a vibrant red broth with chilies and peppercorns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are three young friends visible sitting around the table, including a woman with long curly brown hair in an off-white sweater and a man in a dark blue hoodie with black-framed glasses?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a woman use long wooden chopsticks to dip a red and white marbled beef roll into the boiling broth, causing bubbles to rise and the meat to darken?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm yellow light from a chandelier, creating a soft halo of steam and a cozy atmosphere?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065670177340508 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.016065670177340508, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065670177340508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27456626296043396, + 0.14089368283748627, + 0.08913709968328476, + 0.0493713840842247, + 0.057159435003995895, + 0.08327078819274902, + 0.059559039771556854, + 0.060649674385786057, + 0.08740204572677612, + 0.09799057245254517 + ], + "credit_map": [ + 0.0242410060018301, + 0.03272407874464989, + 0.023027148097753525, + 0.022962534800171852, + 0.03333812579512596, + 0.031196482479572296, + 0.04342905059456825, + 0.06364785134792328, + 0.017049362882971764, + 0.0160871259868145, + 0.013324575498700142, + 0.014880993403494358, + 0.027492493391036987, + 0.013003614731132984, + 0.020838923752307892, + 0.018216606229543686, + 0.012854075990617275, + 0.01464566309005022, + 0.010501213371753693, + 0.011280132457613945, + 0.013737006112933159, + 0.006164299324154854, + 0.012122083455324173, + 0.007832634262740612, + 0.005364108830690384, + 0.014156712219119072, + 0.002457684138789773, + 0.005145105998963118, + 0.009244379587471485, + 0.005290934816002846, + 0.0035894725006073713, + 0.004122990183532238, + 0.006977013777941465, + 0.00640420988202095, + 0.00530755240470171, + 0.005816539749503136, + 0.012806286104023457, + 0.0071979123167693615, + 0.007659600581973791, + 0.004990320187062025, + 0.022915469482541084, + 0.012387526221573353, + 0.006297231186181307, + 0.009050901047885418, + 0.011908854357898235, + 0.005990427453070879, + 0.008213277906179428, + 0.006507096812129021, + 0.006691259797662497, + 0.011749772354960442, + 0.0033621625043451786, + 0.004931917414069176, + 0.02029550075531006, + 0.0036307908594608307, + 0.0048784115351736546, + 0.004019224084913731, + 0.012053783982992172, + 0.010427609086036682, + 0.005002896301448345, + 0.007460777647793293, + 0.010108551010489464, + 0.00568403908982873, + 0.005669363774359226, + 0.004242652095854282, + 0.017145343124866486, + 0.009774661622941494, + 0.012510916218161583, + 0.013462194241583347, + 0.010583317838609219, + 0.006966941989958286, + 0.008896760642528534, + 0.008061902597546577, + 0.01432292815297842, + 0.011967882513999939, + 0.007162932306528091, + 0.00990052055567503, + 0.012661455199122429, + 0.010020853020250797, + 0.014923692680895329, + 0.01703030802309513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.016065670177340508 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_3/index_tensor([3187, 3187])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_3/index_tensor([3187, 3187])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e8b91588b20db9aa07c45d4071b303fb9a215d63 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_3/index_tensor([3187, 3187])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level wide-angle lens creates a tranquil scene, capturing a picturesque barn painted a warm, classic red hue, its weathered wood panels revealing deep grain, the paint gently fading. The barn sits perfectly amidst lush, rolling meadows, vibrant with life and swaying gently in the breeze. Soft, golden sunlight filters through large, fluffy white clouds floating above a pale blue sky, casting a gentle glow on the weathered timbers and creating long, soft shadows across the field. Vibrant wildflowers in purple, yellow, and white dot the meadow, adding a touch of natural color. A rustic split-section wood fence, its wood aged to a silvery gray, lines the perimeter of the property, its form creating a leading line that draws the viewer's eye to the distant, hazy treeline on the horizon. The entire scene is imbued with a tranquil and idyllic atmosphere, capturing the serene, timeless essence of rural beauty.", + "seed": 3086, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_3/index_3187_seed_3086_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08301459718495607, + "vlm_logprob": -0.08301459718495607, + "vlm_yes_no_margin": 3.449999952316284, + "vlm_token_credit": -0.08301459718495607, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007210778072476387 + ], + "opposite_logprob_tokens": [ + -5.257210731506348 + ], + "logprob_reward": -0.007210778072476387, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.007210778072476387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10918164998292923, + 0.08614379167556763, + 0.3671776056289673, + 0.060793232172727585, + 0.05351554602384567, + 0.05216939374804497, + 0.17062623798847198, + 0.04969685152173042, + 0.036531008780002594, + 0.014164716005325317 + ], + "credit_map": [ + 0.012576594017446041, + 0.01377708651125431, + 0.010508110746741295, + 0.015503397211432457, + 0.010553696192800999, + 0.012132002972066402, + 0.016992323100566864, + 0.017138434574007988, + 0.013755274936556816, + 0.01596452295780182, + 0.008045886643230915, + 0.008492857217788696, + 0.010004710406064987, + 0.00865446962416172, + 0.011726075783371925, + 0.009499997831881046, + 0.162342831492424, + 0.15632987022399902, + 0.006472204811871052, + 0.010201865807175636, + 0.00990671943873167, + 0.005885356571525335, + 0.007497617043554783, + 0.008541126735508442, + 0.021327786147594452, + 0.02011614851653576, + 0.002760767238214612, + 0.0034888298250734806, + 0.0033467442262917757, + 0.0033731248695403337, + 0.0023204993922263384, + 0.004059334751218557, + 0.01432864274829626, + 0.0136652747169137, + 0.003282710909843445, + 0.0030193147249519825, + 0.0059980168007314205, + 0.005171662196516991, + 0.002890464384108782, + 0.005159459542483091, + 0.01035307813435793, + 0.009937469847500324, + 0.002732503227889538, + 0.00383169949054718, + 0.007591061759740114, + 0.005935213528573513, + 0.0051566436886787415, + 0.006631727330386639, + 0.0701616108417511, + 0.06721708923578262, + 0.004471407737582922, + 0.0049378108233213425, + 0.007456300780177116, + 0.004987670108675957, + 0.004854785744100809, + 0.006539560854434967, + 0.0074721877463161945, + 0.007000639569014311, + 0.003893256885930896, + 0.005353054963052273, + 0.00633277790620923, + 0.007546011824160814, + 0.004388078115880489, + 0.007710839621722698, + 0.01164398342370987, + 0.010950770229101181, + 0.0024646043311804533, + 0.002044311026111245, + 0.00196828693151474, + 0.0022397725842893124, + 0.0020659633446484804, + 0.003153317142277956, + 0.0023748453240841627, + 0.0020788409747183323, + 0.001387152005918324, + 0.0014537337701767683, + 0.0016819483134895563, + 0.0015771595062687993, + 0.0017642966704443097, + 0.0018467407207936049 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.007210778072476387 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31346678733825684 + ], + "opposite_logprob_tokens": [ + -1.3134667873382568 + ], + "logprob_reward": -0.31346678733825684, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31346678733825684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08606851100921631, + 0.04158620908856392, + 0.5318357944488525, + 0.0599142387509346, + 0.016301000490784645, + 0.018606187775731087, + 0.153373584151268, + 0.07002189755439758, + 0.01023522112518549, + 0.012057402171194553 + ], + "credit_map": [ + 0.012431096285581589, + 0.014613481238484383, + 0.01265845075249672, + 0.015509982593357563, + 0.0066429018042981625, + 0.006606241222470999, + 0.007118237670511007, + 0.010488114319741726, + 0.006619150284677744, + 0.007453485857695341, + 0.0036346439737826586, + 0.003954335115849972, + 0.004770343657582998, + 0.003974507562816143, + 0.006330407690256834, + 0.004849337041378021, + 0.2363470047712326, + 0.22725069522857666, + 0.008815530687570572, + 0.014297615736722946, + 0.013705580495297909, + 0.007583879865705967, + 0.011123724281787872, + 0.012711742892861366, + 0.02615346945822239, + 0.024632848799228668, + 0.0012702615931630135, + 0.001897851238027215, + 0.002113507827743888, + 0.001177113619633019, + 0.0010805130004882812, + 0.0015886693727225065, + 0.003928094170987606, + 0.0038186467718333006, + 0.0012189557310193777, + 0.0018227803520858288, + 0.0013461046619340777, + 0.0012980093015357852, + 0.0015138351591303945, + 0.001354573993012309, + 0.0046991691924631596, + 0.004493251908570528, + 0.0009343439596705139, + 0.0009507223730906844, + 0.002121900673955679, + 0.00179008929990232, + 0.0014438384678214788, + 0.0021728714928030968, + 0.05776911601424217, + 0.05523389205336571, + 0.004088008310645819, + 0.004529920872300863, + 0.00945650227367878, + 0.008507844060659409, + 0.005128478165715933, + 0.008659826591610909, + 0.0092393159866333, + 0.008610601536929607, + 0.010574308224022388, + 0.013571478426456451, + 0.006257480476051569, + 0.005191758740693331, + 0.00848996452987194, + 0.008086984045803547, + 0.0015853767981752753, + 0.001556545146740973, + 0.001178799313493073, + 0.001090730307623744, + 0.001153362332843244, + 0.0010734982788562775, + 0.0010060679633170366, + 0.001590841100551188, + 0.001891900086775422, + 0.00182236242108047, + 0.001824313192628324, + 0.001956493826583028, + 0.0015507673379033804, + 0.000874209450557828, + 0.001145371119491756, + 0.0009919853182509542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.31346678733825684 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189188368618488 + ], + "opposite_logprob_tokens": [ + -4.26418924331665 + ], + "logprob_reward": -0.014189188368618488, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189188368618488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09124217182397842, + 0.1633596122264862, + 0.34817391633987427, + 0.10266024619340897, + 0.06515815854072571, + 0.024479107931256294, + 0.12715616822242737, + 0.046982042491436005, + 0.016827845945954323, + 0.013960679993033409 + ], + "credit_map": [ + 0.012042248621582985, + 0.012681241147220135, + 0.012011275626718998, + 0.014777850359678268, + 0.00643461849540472, + 0.007144985254853964, + 0.010825488716363907, + 0.015324468724429607, + 0.025164926424622536, + 0.029129046946763992, + 0.014194847084581852, + 0.016593704000115395, + 0.01832623966038227, + 0.01717926189303398, + 0.02377970516681671, + 0.018991883844137192, + 0.1474021077156067, + 0.14108459651470184, + 0.007479927968233824, + 0.01276656799018383, + 0.010440325364470482, + 0.0072999675758183, + 0.010385673493146896, + 0.011314785107970238, + 0.017352599650621414, + 0.016552086919546127, + 0.015020027756690979, + 0.018706783652305603, + 0.006935717537999153, + 0.00699398759752512, + 0.009909085929393768, + 0.011189958080649376, + 0.01756448484957218, + 0.01658886857330799, + 0.005639375187456608, + 0.006403437349945307, + 0.003930855076760054, + 0.004773114342242479, + 0.005042582750320435, + 0.005215443670749664, + 0.0034453889820724726, + 0.0034484094940125942, + 0.003173933597281575, + 0.002890645992010832, + 0.0030968880746513605, + 0.002814555075019598, + 0.0029608779586851597, + 0.0026484099216759205, + 0.04639226943254471, + 0.04456394165754318, + 0.0034592284355312586, + 0.0035023551899939775, + 0.008861340582370758, + 0.00794741790741682, + 0.004610077012330294, + 0.007819552905857563, + 0.010508901439607143, + 0.01021997444331646, + 0.004603005014359951, + 0.005718272645026445, + 0.0031628806609660387, + 0.003985908813774586, + 0.003921502269804478, + 0.004861601162701845, + 0.0037469537928700447, + 0.00350717781111598, + 0.0015057172859087586, + 0.001628242083825171, + 0.0015431642532348633, + 0.0012974232668057084, + 0.0012446371838450432, + 0.0023545322474092245, + 0.0023174728266894817, + 0.0021833053324371576, + 0.0011826878180727363, + 0.0016832697438076138, + 0.0017310611438006163, + 0.0014148318441584706, + 0.0014808338601142168, + 0.0019672184716910124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.014189188368618488 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818203739821911 + ], + "opposite_logprob_tokens": [ + -4.018181800842285 + ], + "logprob_reward": -0.01818203739821911, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818203739821911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03848104923963547, + 0.03815750032663345, + 0.1251184642314911, + 0.058282557874917984, + 0.06260231137275696, + 0.10710916668176651, + 0.4280247092247009, + 0.12065141648054123, + 0.015071634203195572, + 0.006501271855086088 + ], + "credit_map": [ + 0.005434487946331501, + 0.006207773461937904, + 0.0038836870808154345, + 0.0042048352770507336, + 0.002783628646284342, + 0.003782683052122593, + 0.005291968118399382, + 0.006891984958201647, + 0.0057082828134298325, + 0.0067747836001217365, + 0.0034736786037683487, + 0.003895306494086981, + 0.004076093900948763, + 0.003855035873129964, + 0.006088286638259888, + 0.004286034032702446, + 0.05627714842557907, + 0.05421280115842819, + 0.0019061177736148238, + 0.0026174006052315235, + 0.003041589166969061, + 0.0015365068102255464, + 0.0022849051747471094, + 0.003241983475163579, + 0.021349532529711723, + 0.020143626257777214, + 0.0019901415798813105, + 0.0029432643204927444, + 0.0034766471944749355, + 0.0029795970767736435, + 0.0020422039087861776, + 0.003357544308528304, + 0.023604055866599083, + 0.022691750898957253, + 0.0018122780602425337, + 0.0017349055269733071, + 0.0038906908594071865, + 0.003603726625442505, + 0.002230516169220209, + 0.003034386783838272, + 0.040224526077508926, + 0.03898526728153229, + 0.0031823941972106695, + 0.0029691553208976984, + 0.006397933699190617, + 0.005244733300060034, + 0.004609014373272657, + 0.0054961382411420345, + 0.16626092791557312, + 0.15913191437721252, + 0.009749648161232471, + 0.010699166916310787, + 0.025838255882263184, + 0.021342750638723373, + 0.013188461773097515, + 0.021813584491610527, + 0.03422043099999428, + 0.032551635056734085, + 0.007619254291057587, + 0.009313623420894146, + 0.007776462938636541, + 0.01077768299728632, + 0.0072804526425898075, + 0.011111877858638763, + 0.003273752983659506, + 0.003208153648301959, + 0.0012385098962113261, + 0.0012955707497894764, + 0.0010746457846835256, + 0.001726912334561348, + 0.001229151152074337, + 0.002024936955422163, + 0.0011310982517898083, + 0.0010857582092285156, + 0.0006237052148208022, + 0.0005900099640712142, + 0.0010828366503119469, + 0.000691402645315975, + 0.0004936131299473345, + 0.0008028477313928306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.01818203739821911 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202419474720955 + ], + "opposite_logprob_tokens": [ + -2.8120241165161133 + ], + "logprob_reward": -0.06202419474720955, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202419474720955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11442188918590546, + 0.030355796217918396, + 0.1050897017121315, + 0.12290941923856735, + 0.05097989737987518, + 0.10495849698781967, + 0.34739676117897034, + 0.03685443103313446, + 0.0654989629983902, + 0.021534619852900505 + ], + "credit_map": [ + 0.0165549349039793, + 0.017203640192747116, + 0.016346320509910583, + 0.016660869121551514, + 0.01013596449047327, + 0.010061888955533504, + 0.011237289756536484, + 0.016220971941947937, + 0.004689084831625223, + 0.005083182360976934, + 0.002441384131088853, + 0.003749541938304901, + 0.0023879946675151587, + 0.0033442366402596235, + 0.0037421933375298977, + 0.004918178543448448, + 0.04570680856704712, + 0.04412486404180527, + 0.0017532441997900605, + 0.0022833349648863077, + 0.003346460172906518, + 0.0021337117068469524, + 0.0026326843071728945, + 0.0031085917726159096, + 0.04731561243534088, + 0.0443718358874321, + 0.006404633168131113, + 0.008402346633374691, + 0.004408726003021002, + 0.0031800922006368637, + 0.0037191740702837706, + 0.005107009783387184, + 0.015755869448184967, + 0.014966951683163643, + 0.003942741546779871, + 0.004574119113385677, + 0.0028060749173164368, + 0.0027431068010628223, + 0.0024688895791769028, + 0.0037221426609903574, + 0.042421914637088776, + 0.04039622098207474, + 0.004318424500524998, + 0.004171267151832581, + 0.0035340969916433096, + 0.0027514975517988205, + 0.0040236422792077065, + 0.0033414410427212715, + 0.14658695459365845, + 0.13928799331188202, + 0.01186507847160101, + 0.011378765106201172, + 0.012409613467752934, + 0.006049741990864277, + 0.010514041408896446, + 0.009304570034146309, + 0.0105659831315279, + 0.010162374936044216, + 0.0027475703973323107, + 0.0038025309331715107, + 0.002440101932734251, + 0.0019194860942661762, + 0.0025379082653671503, + 0.0026784762740135193, + 0.024084841832518578, + 0.02283385396003723, + 0.0037831359077244997, + 0.003612465923652053, + 0.0028611691668629646, + 0.0024552817922085524, + 0.0026757644955068827, + 0.0031924548093229532, + 0.0035084497649222612, + 0.0035016245674341917, + 0.002569682663306594, + 0.0013815523125231266, + 0.001887753140181303, + 0.0023571490310132504, + 0.00325472978875041, + 0.0030736797489225864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.06202419474720955 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle lens creates a tranquil scene, capturing a picturesque barn painted a warm, classic red hue, its weathered wood panels revealing deep grain, the paint gently fading. The barn sits perfectly amidst lush, rolling meadows, vibrant with life and swaying gently in the breeze. Soft, golden sunlight filters through large, fluffy white clouds floating above a pale blue sky, casting a gentle glow on the weathered timbers and creating long, soft shadows across the field. Vibrant wildflowers in purple, yellow, and white dot the meadow, adding a touch of natural color. A rustic split-section wood fence, its wood aged to a silvery gray, lines the perimeter of the property, its form creating a leading line that draws the viewer's eye to the distant, hazy treeline on the horizon. The entire scene is imbued with a tranquil and idyllic atmosphere, capturing the serene, timeless essence of rural beauty.", + "seed": 3087, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_3/index_3187_seed_3087_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03507325891405344, + "vlm_logprob": -0.03507325891405344, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.03507325891405344, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006818361580371857 + ], + "opposite_logprob_tokens": [ + -5.7568182945251465 + ], + "logprob_reward": -0.006818361580371857, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.006818361580371857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15680578351020813, + 0.044058214873075485, + 0.061788223683834076, + 0.06129055842757225, + 0.09376456588506699, + 0.056410662829875946, + 0.06320397555828094, + 0.13905110955238342, + 0.18842799961566925, + 0.1351989209651947 + ], + "credit_map": [ + 0.01657336764037609, + 0.012301472947001457, + 0.02097901701927185, + 0.020036708563566208, + 0.012319042347371578, + 0.02079659514129162, + 0.018608825281262398, + 0.035190753638744354, + 0.006301796995103359, + 0.0035094486083835363, + 0.0058540771715343, + 0.0059190294705331326, + 0.003330182284116745, + 0.006172248627990484, + 0.006363263353705406, + 0.006608165800571442, + 0.007283659651875496, + 0.003922265954315662, + 0.009542868472635746, + 0.0061401440761983395, + 0.004994676914066076, + 0.0129129933193326, + 0.006798652932047844, + 0.010192959569394588, + 0.005780385807156563, + 0.005174587480723858, + 0.008289302699267864, + 0.008858168497681618, + 0.005857130978256464, + 0.007323114201426506, + 0.0077143143862485886, + 0.0122935576364398, + 0.00981996301561594, + 0.005619320087134838, + 0.018954219296574593, + 0.015717463567852974, + 0.006439974531531334, + 0.010785583406686783, + 0.01162679772824049, + 0.014801240526139736, + 0.006021014880388975, + 0.004698993172496557, + 0.005701834335923195, + 0.008169927634298801, + 0.006577589549124241, + 0.00962870940566063, + 0.007264136336743832, + 0.008348457515239716, + 0.010793796740472317, + 0.0038620843552052975, + 0.004759666975587606, + 0.004805759992450476, + 0.008943993598222733, + 0.01766907423734665, + 0.004799924790859222, + 0.007569668814539909, + 0.01735319383442402, + 0.011682625859975815, + 0.01664040796458721, + 0.013351608999073505, + 0.010410823859274387, + 0.01939966343343258, + 0.020576242357492447, + 0.029636552557349205, + 0.029024893417954445, + 0.012956936843693256, + 0.013650897890329361, + 0.024265257641673088, + 0.018458811566233635, + 0.04640049859881401, + 0.021114302799105644, + 0.022556396201252937, + 0.014331781305372715, + 0.008157026953995228, + 0.019272729754447937, + 0.023356454446911812, + 0.011518828570842743, + 0.02341024950146675, + 0.016662446781992912, + 0.018489407375454903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.006818361580371857 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029812553897500038 + ], + "opposite_logprob_tokens": [ + -3.5298125743865967 + ], + "logprob_reward": -0.029812553897500038, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029812553897500038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10809534043073654, + 0.13748778402805328, + 0.2321118414402008, + 0.12122540175914764, + 0.06865888088941574, + 0.03410807251930237, + 0.037326354533433914, + 0.1041077971458435, + 0.10344154387712479, + 0.053437042981386185 + ], + "credit_map": [ + 0.014726927503943443, + 0.007444961927831173, + 0.011237027123570442, + 0.010726875625550747, + 0.009227853268384933, + 0.023647673428058624, + 0.013410511426627636, + 0.017673511058092117, + 0.011709814891219139, + 0.00677165063098073, + 0.018128059804439545, + 0.0164694394916296, + 0.009197069332003593, + 0.012858464382588863, + 0.02306281216442585, + 0.03929048404097557, + 0.028594013303518295, + 0.013737955130636692, + 0.028282159939408302, + 0.028774185106158257, + 0.01878584548830986, + 0.0416865237057209, + 0.03883194550871849, + 0.033419206738471985, + 0.01795712299644947, + 0.008578117936849594, + 0.011645427905023098, + 0.009974916465580463, + 0.01269443891942501, + 0.028178784996271133, + 0.01483981218189001, + 0.01735677942633629, + 0.007975118234753609, + 0.004595842678099871, + 0.014408689923584461, + 0.01163109578192234, + 0.005052265245467424, + 0.008010933175683022, + 0.0074068959802389145, + 0.009578037075698376, + 0.00526439631357789, + 0.002581134671345353, + 0.0028484815265983343, + 0.0026839270722121, + 0.0036563663743436337, + 0.008635160513222218, + 0.0036968018393963575, + 0.004741807468235493, + 0.004213754553347826, + 0.0036607219371944666, + 0.003875410882756114, + 0.0055105616338551044, + 0.0036885859444737434, + 0.005546214524656534, + 0.004638794809579849, + 0.006192308850586414, + 0.016514241695404053, + 0.007128484547138214, + 0.006495953071862459, + 0.007755773141980171, + 0.010593365877866745, + 0.02804182842373848, + 0.011756841093301773, + 0.01582130789756775, + 0.01447173859924078, + 0.007638226263225079, + 0.008996003307402134, + 0.01483532227575779, + 0.00974676851183176, + 0.021092703565955162, + 0.012961646541953087, + 0.013699132017791271, + 0.007864839397370815, + 0.003012032015249133, + 0.004307080525904894, + 0.006468906532973051, + 0.005693742539733648, + 0.0138401472941041, + 0.004947157111018896, + 0.007303137332201004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.029812553897500038 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778928942978382 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009778928942978382, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778928942978382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16786065697669983, + 0.20819702744483948, + 0.028557654470205307, + 0.043603308498859406, + 0.05381465703248978, + 0.0377841591835022, + 0.06065806746482849, + 0.10103411972522736, + 0.19963310658931732, + 0.09885731339454651 + ], + "credit_map": [ + 0.019131207838654518, + 0.010124842636287212, + 0.017736464738845825, + 0.021195128560066223, + 0.010500135831534863, + 0.02401134930551052, + 0.018785882741212845, + 0.04637560620903969, + 0.021226542070508003, + 0.015580195933580399, + 0.027340427041053772, + 0.02750801295042038, + 0.015807367861270905, + 0.01980874314904213, + 0.03155379742383957, + 0.04937192425131798, + 0.003370968159288168, + 0.0022927867248654366, + 0.00341574614867568, + 0.002385280327871442, + 0.002810124075040221, + 0.005173210985958576, + 0.0043378253467381, + 0.004771708976477385, + 0.004575976170599461, + 0.004316103644669056, + 0.00854212511330843, + 0.00460770633071661, + 0.0035416167229413986, + 0.004747665487229824, + 0.005189803894609213, + 0.008082306012511253, + 0.007015598472207785, + 0.0037469398230314255, + 0.009929061867296696, + 0.008133954368531704, + 0.004408051259815693, + 0.008075720630586147, + 0.00621217442676425, + 0.006293144542723894, + 0.006722817663103342, + 0.002386445179581642, + 0.0020772761199623346, + 0.0023099135141819715, + 0.004353974014520645, + 0.012366208247840405, + 0.00308958045206964, + 0.004477938637137413, + 0.008320429362356663, + 0.00633501959964633, + 0.004863271955400705, + 0.006950064096599817, + 0.008099038153886795, + 0.013873804360628128, + 0.005941945593804121, + 0.006274482235312462, + 0.012792633846402168, + 0.00805335957556963, + 0.009397630579769611, + 0.009953579865396023, + 0.009491907432675362, + 0.01828008145093918, + 0.014062208123505116, + 0.019002700224518776, + 0.030625302344560623, + 0.012342157773673534, + 0.016215821728110313, + 0.02756618708372116, + 0.01984601467847824, + 0.046896327286958694, + 0.020990917459130287, + 0.02515033818781376, + 0.012181833386421204, + 0.007185616996139288, + 0.014772141352295876, + 0.01659690961241722, + 0.009710529819130898, + 0.015539295971393585, + 0.01065524946898222, + 0.01221571397036314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.009778928942978382 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606062613427639 + ], + "opposite_logprob_tokens": [ + -4.141060829162598 + ], + "logprob_reward": -0.01606062613427639, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606062613427639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0712457150220871, + 0.12773752212524414, + 0.05053691565990448, + 0.06979860365390778, + 0.28797078132629395, + 0.022090978920459747, + 0.04501756280660629, + 0.10262344032526016, + 0.18988439440727234, + 0.03309418261051178 + ], + "credit_map": [ + 0.009449807927012444, + 0.006257711909711361, + 0.007750998716801405, + 0.009588908404111862, + 0.00468857865780592, + 0.01127255242317915, + 0.008034858852624893, + 0.014202292077243328, + 0.011642449535429478, + 0.00697712367400527, + 0.016318030655384064, + 0.013245207257568836, + 0.009260688908398151, + 0.011957946233451366, + 0.01983693242073059, + 0.038499124348163605, + 0.005253802053630352, + 0.003614254994317889, + 0.009966908022761345, + 0.006006432231515646, + 0.0035542589612305164, + 0.008234741166234016, + 0.005506602581590414, + 0.008399910293519497, + 0.008212740533053875, + 0.007733539212495089, + 0.011144629679620266, + 0.009260226041078568, + 0.00744579778984189, + 0.010824447497725487, + 0.005990476813167334, + 0.009186739102005959, + 0.030321283265948296, + 0.01663338579237461, + 0.050852470099925995, + 0.04596021771430969, + 0.020432185381650925, + 0.04746899753808975, + 0.03484192118048668, + 0.04146026074886322, + 0.0025517407339066267, + 0.0019611353054642677, + 0.0020212791860103607, + 0.0023658692371100187, + 0.0017210845835506916, + 0.004594992380589247, + 0.003535305615514517, + 0.0033395695500075817, + 0.005920196417719126, + 0.004109556321054697, + 0.003384934738278389, + 0.004972176160663366, + 0.004010320641100407, + 0.0086359279230237, + 0.0056707183830440044, + 0.008313726633787155, + 0.014309822581708431, + 0.008877253159880638, + 0.008513827808201313, + 0.008306389674544334, + 0.010455741547048092, + 0.0226447656750679, + 0.012302746064960957, + 0.017212877050042152, + 0.029337484389543533, + 0.012378538027405739, + 0.01510861236602068, + 0.02322671189904213, + 0.0203056912869215, + 0.046740639954805374, + 0.018464574590325356, + 0.02432212606072426, + 0.004117592703551054, + 0.002413651440292597, + 0.004525772295892239, + 0.004892989061772823, + 0.002261877292767167, + 0.005813328083604574, + 0.0038844975642859936, + 0.005184469278901815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.01606062613427639 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11289582401514053 + ], + "opposite_logprob_tokens": [ + -2.2378957271575928 + ], + "logprob_reward": -0.11289582401514053, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11289582401514053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1039721891283989, + 0.03799428418278694, + 0.06855375319719315, + 0.16809406876564026, + 0.05993509665131569, + 0.02120821364223957, + 0.071177639067173, + 0.04086696356534958, + 0.3792347013950348, + 0.04896309971809387 + ], + "credit_map": [ + 0.011961088515818119, + 0.00822402723133564, + 0.00973349530249834, + 0.020687289535999298, + 0.008363251574337482, + 0.012359561398625374, + 0.012852014973759651, + 0.019791454076766968, + 0.0033927238546311855, + 0.004591242875903845, + 0.005326494108885527, + 0.006850938778370619, + 0.0030959027353674173, + 0.0029207130428403616, + 0.00668285833671689, + 0.005133409984409809, + 0.008095292374491692, + 0.005857101175934076, + 0.008637658320367336, + 0.009062607772648335, + 0.0060806721448898315, + 0.008625838905572891, + 0.012473615817725658, + 0.009720956906676292, + 0.0177772119641304, + 0.021279871463775635, + 0.03192546218633652, + 0.02640707604587078, + 0.020186908543109894, + 0.013967405073344707, + 0.018316084519028664, + 0.018234027549624443, + 0.006912356708198786, + 0.003866348182782531, + 0.008825818076729774, + 0.00742186838760972, + 0.005035134498029947, + 0.01198581699281931, + 0.007060580886900425, + 0.008827168494462967, + 0.003165532136335969, + 0.0019811061210930347, + 0.0021308138966560364, + 0.002700410783290863, + 0.0024208766408264637, + 0.0031794686801731586, + 0.0031709761824458838, + 0.002459026640281081, + 0.012341848574578762, + 0.0035076383501291275, + 0.003829055465757847, + 0.004721892066299915, + 0.008966933004558086, + 0.022386102005839348, + 0.00607388885691762, + 0.00935027189552784, + 0.006114303600043058, + 0.0021069429349154234, + 0.003260534256696701, + 0.0035052523016929626, + 0.004115824121981859, + 0.01058364100754261, + 0.004824397154152393, + 0.006356062367558479, + 0.06230911612510681, + 0.020332926884293556, + 0.031858641654253006, + 0.04349428415298462, + 0.03938867151737213, + 0.10028790682554245, + 0.03476701304316521, + 0.04679609090089798, + 0.007593509275466204, + 0.0020188861526548862, + 0.0035778935998678207, + 0.005328974220901728, + 0.004890953656286001, + 0.0126383351162076, + 0.004431392066180706, + 0.00848314817994833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.11289582401514053 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle lens creates a tranquil scene, capturing a picturesque barn painted a warm, classic red hue, its weathered wood panels revealing deep grain, the paint gently fading. The barn sits perfectly amidst lush, rolling meadows, vibrant with life and swaying gently in the breeze. Soft, golden sunlight filters through large, fluffy white clouds floating above a pale blue sky, casting a gentle glow on the weathered timbers and creating long, soft shadows across the field. Vibrant wildflowers in purple, yellow, and white dot the meadow, adding a touch of natural color. A rustic split-section wood fence, its wood aged to a silvery gray, lines the perimeter of the property, its form creating a leading line that draws the viewer's eye to the distant, hazy treeline on the horizon. The entire scene is imbued with a tranquil and idyllic atmosphere, capturing the serene, timeless essence of rural beauty.", + "seed": 3088, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_3/index_3187_seed_3088_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.029914145823568106, + "vlm_logprob": -0.029914145823568106, + "vlm_yes_no_margin": 3.975, + "vlm_token_credit": -0.029914145823568106, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004541202913969755 + ], + "opposite_logprob_tokens": [ + -5.754541397094727 + ], + "logprob_reward": -0.004541202913969755, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.004541202913969755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42236700654029846, + 0.07897073030471802, + 0.05731825903058052, + 0.045269276946783066, + 0.06368998438119888, + 0.06531375646591187, + 0.050003305077552795, + 0.08191540092229843, + 0.04950135201215744, + 0.08565089106559753 + ], + "credit_map": [ + 0.08346808701753616, + 0.06794656813144684, + 0.057454053312540054, + 0.03995158523321152, + 0.03299286216497421, + 0.05508474260568619, + 0.035981688648462296, + 0.049487438052892685, + 0.013716030865907669, + 0.01212400197982788, + 0.009316930547356606, + 0.010311723686754704, + 0.0073114680126309395, + 0.00995706208050251, + 0.00841598492115736, + 0.007817537523806095, + 0.008905955590307713, + 0.0075777508318424225, + 0.006795587949454784, + 0.005339009687304497, + 0.00472809374332428, + 0.010675402358174324, + 0.00462270388379693, + 0.008673757314682007, + 0.009490618482232094, + 0.007582041434943676, + 0.005001181736588478, + 0.003609300358220935, + 0.003564480459317565, + 0.0082291504368186, + 0.00340681872330606, + 0.004385688807815313, + 0.009045803919434547, + 0.009624245576560497, + 0.007365555502474308, + 0.0090391356498003, + 0.007238901685923338, + 0.008420074358582497, + 0.00678488751873374, + 0.006171387620270252, + 0.010353985242545605, + 0.011267312802374363, + 0.009037662297487259, + 0.007394927553832531, + 0.006967740599066019, + 0.006711130030453205, + 0.008412705734372139, + 0.005168300122022629, + 0.015500507317483425, + 0.01665588654577732, + 0.0032115094363689423, + 0.002461463212966919, + 0.002114039845764637, + 0.004576494917273521, + 0.003369783516973257, + 0.0021136256400495768, + 0.012910687364637852, + 0.014441346749663353, + 0.008238974027335644, + 0.008428980596363544, + 0.00832461193203926, + 0.01029595173895359, + 0.01322296354919672, + 0.006051893811672926, + 0.0077870190143585205, + 0.006485721096396446, + 0.005893527530133724, + 0.005126317963004112, + 0.006185705773532391, + 0.007442954462021589, + 0.006826678290963173, + 0.003753429278731346, + 0.012394082732498646, + 0.01676301658153534, + 0.00979503896087408, + 0.007117795292288065, + 0.006716250907629728, + 0.016652900725603104, + 0.00995978806167841, + 0.00625202152878046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.004541202913969755 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370801359415054 + ], + "opposite_logprob_tokens": [ + -3.408708095550537 + ], + "logprob_reward": -0.03370801359415054, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370801359415054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2772224545478821, + 0.11216972023248672, + 0.08729161322116852, + 0.029404129832983017, + 0.055355384945869446, + 0.04974447190761566, + 0.04088233783841133, + 0.20942945778369904, + 0.03595711663365364, + 0.10254335403442383 + ], + "credit_map": [ + 0.05545416474342346, + 0.053567513823509216, + 0.04299527034163475, + 0.02330523543059826, + 0.017127294093370438, + 0.031689487397670746, + 0.02319241315126419, + 0.029891077429056168, + 0.023267166689038277, + 0.02354736067354679, + 0.01902421936392784, + 0.010165577754378319, + 0.005912965163588524, + 0.012833972461521626, + 0.008334522135555744, + 0.009083935990929604, + 0.016056135296821594, + 0.015727439895272255, + 0.014471616595983505, + 0.00859121885150671, + 0.006613411940634251, + 0.011924554593861103, + 0.006313379388302565, + 0.007593858055770397, + 0.008200599811971188, + 0.007336803246289492, + 0.0032397122122347355, + 0.002187818055972457, + 0.001498099765740335, + 0.0027063004672527313, + 0.0016596631612628698, + 0.0025751325301826, + 0.011444058269262314, + 0.01101565733551979, + 0.008879708126187325, + 0.005467370618134737, + 0.002859402447938919, + 0.007193286903202534, + 0.0033190695103257895, + 0.005176832899451256, + 0.009159193374216557, + 0.008646995760500431, + 0.00881008617579937, + 0.004592448938637972, + 0.003950358368456364, + 0.006760854739695787, + 0.0036626579239964485, + 0.004161877557635307, + 0.00845465436577797, + 0.008857701905071735, + 0.004436134360730648, + 0.0027885299641638994, + 0.0024100756272673607, + 0.007138230837881565, + 0.002816338324919343, + 0.00398067245259881, + 0.07979010045528412, + 0.07742656767368317, + 0.010773993097245693, + 0.007302679121494293, + 0.006130749825388193, + 0.012489917688071728, + 0.00795045681297779, + 0.0075649782083928585, + 0.007147488184273243, + 0.0073753842152655125, + 0.004311643075197935, + 0.0027591485995799303, + 0.0021462368313223124, + 0.006480673793703318, + 0.0031803231686353683, + 0.002556220628321171, + 0.017583781853318214, + 0.0199724193662405, + 0.019680695608258247, + 0.011129560880362988, + 0.004460903815925121, + 0.012414127588272095, + 0.008122400380671024, + 0.00917946919798851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.03370801359415054 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817992888391018 + ], + "opposite_logprob_tokens": [ + -4.018179893493652 + ], + "logprob_reward": -0.01817992888391018, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817992888391018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3102879822254181, + 0.06601838767528534, + 0.10282614827156067, + 0.11444521695375443, + 0.12250292301177979, + 0.08005238324403763, + 0.056090228259563446, + 0.1007552519440651, + 0.022857563570141792, + 0.02416389063000679 + ], + "credit_map": [ + 0.07051526010036469, + 0.0611303448677063, + 0.04163740202784538, + 0.028138218447566032, + 0.027513673529028893, + 0.034655097872018814, + 0.02805470861494541, + 0.018643269315361977, + 0.013031276874244213, + 0.011628893204033375, + 0.007231997326016426, + 0.007053788751363754, + 0.004575128201395273, + 0.011748777702450752, + 0.005017977673560381, + 0.0057305446825921535, + 0.02401166968047619, + 0.024387307465076447, + 0.010158093646168709, + 0.00869304221123457, + 0.006891997531056404, + 0.012746189720928669, + 0.008798030205070972, + 0.007139808963984251, + 0.030618863180279732, + 0.029349707067012787, + 0.013038165867328644, + 0.009896098636090755, + 0.00585524458438158, + 0.00915226899087429, + 0.007239812519401312, + 0.009295043535530567, + 0.018840068951249123, + 0.015011188574135303, + 0.02515716664493084, + 0.015524782240390778, + 0.009037250652909279, + 0.018783144652843475, + 0.00930347666144371, + 0.010845848359167576, + 0.01448410376906395, + 0.015281222760677338, + 0.015925228595733643, + 0.00902392528951168, + 0.0046682655811309814, + 0.009916736744344234, + 0.005290130618959665, + 0.00546276243403554, + 0.020220225676894188, + 0.020790638402104378, + 0.0024321789387613535, + 0.0017305260989814997, + 0.0023374136071652174, + 0.0042674727737903595, + 0.0019370859954506159, + 0.002374685136601329, + 0.023419931530952454, + 0.02505076304078102, + 0.012124436907470226, + 0.007632086984813213, + 0.0065422384068369865, + 0.009964033029973507, + 0.009386532008647919, + 0.006635218858718872, + 0.003964630886912346, + 0.003793062176555395, + 0.0029611459467560053, + 0.0021244084928184748, + 0.0017922354163601995, + 0.0034857881255447865, + 0.0026529179885983467, + 0.002083372324705124, + 0.003923051059246063, + 0.003132751677185297, + 0.005803207866847515, + 0.003054512431845069, + 0.0012519396841526031, + 0.0021719252690672874, + 0.0025104822125285864, + 0.002316017635166645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.01817992888391018 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188600704073906 + ], + "opposite_logprob_tokens": [ + -4.264188766479492 + ], + "logprob_reward": -0.014188600704073906, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188600704073906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29952162504196167, + 0.037163253873586655, + 0.03733324259519577, + 0.039607640355825424, + 0.07475559413433075, + 0.10196579247713089, + 0.038875382393598557, + 0.3144272565841675, + 0.03908749297261238, + 0.01726275123655796 + ], + "credit_map": [ + 0.07831529527902603, + 0.06582208722829819, + 0.03739701956510544, + 0.016633415594697, + 0.024585319682955742, + 0.03173007071018219, + 0.023945670574903488, + 0.021092725917696953, + 0.008122632279992104, + 0.005687331780791283, + 0.005141885951161385, + 0.0033226849045604467, + 0.0022199219092726707, + 0.005531454924494028, + 0.0027729051653295755, + 0.004364429973065853, + 0.00875372625887394, + 0.008119308389723301, + 0.0043290709145367146, + 0.0021476515103131533, + 0.002110914560034871, + 0.006752596702426672, + 0.0018531217938289046, + 0.0032668448984622955, + 0.007431057747453451, + 0.006194346584379673, + 0.008094077929854393, + 0.00440030824393034, + 0.002210348378866911, + 0.004739562049508095, + 0.003260966157540679, + 0.0032769711688160896, + 0.010303682647645473, + 0.00992998480796814, + 0.01350005716085434, + 0.007782532833516598, + 0.004798971116542816, + 0.013508746400475502, + 0.005985027179121971, + 0.008946588262915611, + 0.026214130222797394, + 0.02524707280099392, + 0.014298437163233757, + 0.008122248575091362, + 0.004858840722590685, + 0.009785402566194534, + 0.00707180704921484, + 0.006367843598127365, + 0.011288555338978767, + 0.011955381371080875, + 0.004018120467662811, + 0.002185797318816185, + 0.0019144064281135798, + 0.0032124922145158052, + 0.0020263080950826406, + 0.0022743188310414553, + 0.06648638099431992, + 0.07574339956045151, + 0.03636043146252632, + 0.023269090801477432, + 0.018183084204792976, + 0.04670468345284462, + 0.029243644326925278, + 0.018436508253216743, + 0.007381788920611143, + 0.009094204753637314, + 0.003873166861012578, + 0.0028143078088760376, + 0.002400533528998494, + 0.006988884415477514, + 0.0029422221705317497, + 0.0035923814866691828, + 0.0026432282757014036, + 0.002931259572505951, + 0.00397334061563015, + 0.0017464865231886506, + 0.0010572513565421104, + 0.0024041172582656145, + 0.0014805368846282363, + 0.0010265312157571316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.014188600704073906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895298302173615 + ], + "opposite_logprob_tokens": [ + -2.5789530277252197 + ], + "logprob_reward": -0.07895298302173615, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895298302173615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.215118408203125, + 0.06575474888086319, + 0.10452250391244888, + 0.03663087263703346, + 0.0356033518910408, + 0.05522990599274635, + 0.11859937012195587, + 0.11354835331439972, + 0.20017538964748383, + 0.054816994816064835 + ], + "credit_map": [ + 0.06111445277929306, + 0.04710708558559418, + 0.020288990810513496, + 0.01290983147919178, + 0.017553824931383133, + 0.0223048347979784, + 0.015445034950971603, + 0.018394382670521736, + 0.016609307378530502, + 0.015201834961771965, + 0.006849763449281454, + 0.006183874327689409, + 0.004721979610621929, + 0.006024541798979044, + 0.005384375341236591, + 0.004779082722961903, + 0.030929770320653915, + 0.03145357221364975, + 0.00601972034201026, + 0.005710730794817209, + 0.005219932645559311, + 0.014706170186400414, + 0.00453777564689517, + 0.005944833159446716, + 0.006364487577229738, + 0.006414622999727726, + 0.0038460211362689734, + 0.0030655497685074806, + 0.003703000722452998, + 0.007554422132670879, + 0.0026532167103141546, + 0.003029556479305029, + 0.0057312967255711555, + 0.005590254906564951, + 0.004931026604026556, + 0.00394562678411603, + 0.0035109154414385557, + 0.005986727774143219, + 0.003092473605647683, + 0.0028150316793471575, + 0.015031839720904827, + 0.01496325433254242, + 0.005492459051311016, + 0.0032778207678347826, + 0.003871652763336897, + 0.004689501598477364, + 0.004293433390557766, + 0.0036099462304264307, + 0.04551837965846062, + 0.045482534915208817, + 0.00631294772028923, + 0.0035517506767064333, + 0.003399667562916875, + 0.005498063284903765, + 0.003963800147175789, + 0.004872230812907219, + 0.033861562609672546, + 0.03501388058066368, + 0.007761987391859293, + 0.0058113704435527325, + 0.004794251639395952, + 0.012964114546775818, + 0.005838473327457905, + 0.007502725347876549, + 0.027609463781118393, + 0.03369968384504318, + 0.026896363124251366, + 0.017599547281861305, + 0.0115382494404912, + 0.0476677380502224, + 0.019349820911884308, + 0.01581454463303089, + 0.008790782652795315, + 0.010690107010304928, + 0.008313266560435295, + 0.0046697575598955154, + 0.00318508711643517, + 0.01058038417249918, + 0.00419502891600132, + 0.004392585251480341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.07895298302173615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level wide-angle lens creates a tranquil scene, capturing a picturesque barn painted a warm, classic red hue, its weathered wood panels revealing deep grain, the paint gently fading. The barn sits perfectly amidst lush, rolling meadows, vibrant with life and swaying gently in the breeze. Soft, golden sunlight filters through large, fluffy white clouds floating above a pale blue sky, casting a gentle glow on the weathered timbers and creating long, soft shadows across the field. Vibrant wildflowers in purple, yellow, and white dot the meadow, adding a touch of natural color. A rustic split-section wood fence, its wood aged to a silvery gray, lines the perimeter of the property, its form creating a leading line that draws the viewer's eye to the distant, hazy treeline on the horizon. The entire scene is imbued with a tranquil and idyllic atmosphere, capturing the serene, timeless essence of rural beauty.", + "seed": 3089, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_3/index_3187_seed_3089_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.025050224084407092, + "vlm_logprob": -0.025050224084407092, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.025050224084407092, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004009542521089315 + ], + "opposite_logprob_tokens": [ + -6.00400972366333 + ], + "logprob_reward": -0.004009542521089315, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.004009542521089315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17427395284175873, + 0.1847842037677765, + 0.1311333328485489, + 0.05725528672337532, + 0.12323905527591705, + 0.02888663299381733, + 0.10759422928094864, + 0.07229400426149368, + 0.051999498158693314, + 0.0685398057103157 + ], + "credit_map": [ + 0.016677362844347954, + 0.016677459701895714, + 0.029388295486569405, + 0.027918677777051926, + 0.01749304309487343, + 0.017503149807453156, + 0.023821722716093063, + 0.024794243276119232, + 0.01692354679107666, + 0.019520321860909462, + 0.038214311003685, + 0.027274927124381065, + 0.015661053359508514, + 0.013809758238494396, + 0.024497831240296364, + 0.028882455080747604, + 0.01293348427861929, + 0.014686905778944492, + 0.026633385568857193, + 0.019117530435323715, + 0.010232619941234589, + 0.011333387345075607, + 0.016670463606715202, + 0.019525552168488503, + 0.006093234289437532, + 0.008049641735851765, + 0.0092897554859519, + 0.009133663028478622, + 0.005740222055464983, + 0.005001210141927004, + 0.007325487211346626, + 0.006622073706239462, + 0.012149598449468613, + 0.012348338030278683, + 0.025972630828619003, + 0.01756669208407402, + 0.011253182776272297, + 0.010595819912850857, + 0.016084104776382446, + 0.017268698662519455, + 0.00303838262334466, + 0.0033368465956300497, + 0.0036276397295296192, + 0.00400429917499423, + 0.0036764112301170826, + 0.004563732072710991, + 0.0032394069712609053, + 0.0033999159932136536, + 0.012157213874161243, + 0.01072684396058321, + 0.02009759284555912, + 0.011607383377850056, + 0.01374016609042883, + 0.007792117539793253, + 0.014370527118444443, + 0.01710238866508007, + 0.008310236036777496, + 0.009733524173498154, + 0.010773408226668835, + 0.011057814583182335, + 0.007913977839052677, + 0.006616006605327129, + 0.008575670421123505, + 0.009313362650573254, + 0.005901047959923744, + 0.004735932219773531, + 0.009957093745470047, + 0.00516878068447113, + 0.004854370374232531, + 0.00585024943575263, + 0.006414579227566719, + 0.009117445908486843, + 0.007018709089607, + 0.0090443454682827, + 0.010370918549597263, + 0.010852920822799206, + 0.006803153548389673, + 0.007168399635702372, + 0.008748677559196949, + 0.008532678708434105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.004009542521089315 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043073561042547226 + ], + "opposite_logprob_tokens": [ + -3.1680736541748047 + ], + "logprob_reward": -0.043073561042547226, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043073561042547226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3426686227321625, + 0.16572730243206024, + 0.09220538288354874, + 0.02211257815361023, + 0.05694578215479851, + 0.031967077404260635, + 0.06832297146320343, + 0.05045078322291374, + 0.07290181517601013, + 0.09669767320156097 + ], + "credit_map": [ + 0.03698233515024185, + 0.02806684374809265, + 0.06180877238512039, + 0.03970850631594658, + 0.030332308262586594, + 0.044726401567459106, + 0.0430917963385582, + 0.057951658964157104, + 0.017200665548443794, + 0.012910629622638226, + 0.038385991007089615, + 0.015891194343566895, + 0.012116575613617897, + 0.015672648325562477, + 0.02111312933266163, + 0.03243647515773773, + 0.010234599933028221, + 0.008885659277439117, + 0.018377358093857765, + 0.010121287778019905, + 0.009415724314749241, + 0.008600120432674885, + 0.012172036804258823, + 0.01439859438687563, + 0.0029713839758187532, + 0.0022047695238143206, + 0.003113101702183485, + 0.001969975419342518, + 0.0031418907456099987, + 0.0032822568900883198, + 0.0023142434656620026, + 0.00311495503410697, + 0.006163828540593386, + 0.006403421517461538, + 0.010857310146093369, + 0.006249822210520506, + 0.0067714606411755085, + 0.004818996414542198, + 0.006654631812125444, + 0.009026311337947845, + 0.003486853325739503, + 0.0032072702888399363, + 0.005148197989910841, + 0.004303708206862211, + 0.0034563106019049883, + 0.0037574705202132463, + 0.003625655546784401, + 0.004981611389666796, + 0.007669136859476566, + 0.008478390984237194, + 0.00881842989474535, + 0.007002442143857479, + 0.009834949858486652, + 0.00592090142890811, + 0.008846059441566467, + 0.0117526575922966, + 0.0049782972782850266, + 0.005765438545495272, + 0.009804923087358475, + 0.007365751545876265, + 0.003958316054195166, + 0.00466175377368927, + 0.006454110611230135, + 0.007462189998477697, + 0.007661896757781506, + 0.005611206870526075, + 0.017061788588762283, + 0.0074579427018761635, + 0.008083382621407509, + 0.004768284969031811, + 0.00963143166154623, + 0.01262588333338499, + 0.009800107218325138, + 0.008345368318259716, + 0.01949724368751049, + 0.011333483271300793, + 0.0105786407366395, + 0.008044794201850891, + 0.012754503637552261, + 0.016343535855412483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.043073561042547226 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535353191196918 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535353191196918, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535353191196918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12380470335483551, + 0.17596673965454102, + 0.07491844892501831, + 0.17860466241836548, + 0.10310167074203491, + 0.049088433384895325, + 0.12049344927072525, + 0.09118857979774475, + 0.018146878108382225, + 0.06468643993139267 + ], + "credit_map": [ + 0.014964569360017776, + 0.01185127068310976, + 0.017955653369426727, + 0.016044380143284798, + 0.013583352789282799, + 0.01638132333755493, + 0.015871351584792137, + 0.017152804881334305, + 0.01915920525789261, + 0.014974385499954224, + 0.03732922673225403, + 0.020340770483016968, + 0.014239673502743244, + 0.017053602263331413, + 0.022290818393230438, + 0.030579060316085815, + 0.010837530717253685, + 0.007201831787824631, + 0.007922022603452206, + 0.008146627806127071, + 0.011165165342390537, + 0.011317960917949677, + 0.00829619076102972, + 0.010031115263700485, + 0.023128073662519455, + 0.018610935658216476, + 0.02744802087545395, + 0.015696361660957336, + 0.028658369556069374, + 0.022432582452893257, + 0.017344044521450996, + 0.025286275893449783, + 0.015297416597604752, + 0.009936920367181301, + 0.008142353966832161, + 0.009593949653208256, + 0.01703113317489624, + 0.017033129930496216, + 0.011319530196487904, + 0.014747238717973232, + 0.006772586144506931, + 0.0054649957455694675, + 0.0051199509762227535, + 0.005740760825574398, + 0.009165509603917599, + 0.007243412081152201, + 0.00488325348123908, + 0.004697964061051607, + 0.016810646280646324, + 0.016641076654195786, + 0.012086468748748302, + 0.013230721466243267, + 0.01993039809167385, + 0.012580927461385727, + 0.01401672512292862, + 0.015196481719613075, + 0.013645149767398834, + 0.012762274593114853, + 0.010250760242342949, + 0.009404403157532215, + 0.016647281125187874, + 0.011306976899504662, + 0.008128867484629154, + 0.009042867459356785, + 0.0027376075740903616, + 0.002503738971427083, + 0.0014293475542217493, + 0.0023773368448019028, + 0.0029325950890779495, + 0.002655457006767392, + 0.0017114243237301707, + 0.0017993710935115814, + 0.006596859078854322, + 0.004756694193929434, + 0.013100248761475086, + 0.007245150860399008, + 0.005462362430989742, + 0.0064190709963440895, + 0.00867449026554823, + 0.012431565672159195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.012535353191196918 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036118291318416595 + ], + "opposite_logprob_tokens": [ + -5.628612041473389 + ], + "logprob_reward": -0.0036118291318416595, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036118291318416595, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16611292958259583, + 0.09583315253257751, + 0.08188950270414352, + 0.09935358166694641, + 0.05786019563674927, + 0.13780327141284943, + 0.16369716823101044, + 0.05947837978601456, + 0.0646277666091919, + 0.07334405928850174 + ], + "credit_map": [ + 0.022495519369840622, + 0.012670987285673618, + 0.025372061878442764, + 0.01774601638317108, + 0.016063282266259193, + 0.027682088315486908, + 0.019072042778134346, + 0.025010937824845314, + 0.012877458706498146, + 0.009729050099849701, + 0.015450523234903812, + 0.008612235076725483, + 0.015198752284049988, + 0.013141999952495098, + 0.009981919080018997, + 0.010841209441423416, + 0.010910309851169586, + 0.013428314588963985, + 0.006577313411980867, + 0.012904134579002857, + 0.011993045918643475, + 0.009067458100616932, + 0.008450856432318687, + 0.00855807401239872, + 0.014059113338589668, + 0.015465956181287766, + 0.009128591977059841, + 0.01241999864578247, + 0.014640706591308117, + 0.011608390137553215, + 0.011299293488264084, + 0.010731532238423824, + 0.006349553819745779, + 0.005769873969256878, + 0.011226233094930649, + 0.0056635732762515545, + 0.006542158778756857, + 0.00477253645658493, + 0.00680736918002367, + 0.01072890218347311, + 0.018297068774700165, + 0.010882739908993244, + 0.019364455714821815, + 0.015221447683870792, + 0.014044861309230328, + 0.02315795235335827, + 0.016377463936805725, + 0.020457275211811066, + 0.02023603953421116, + 0.01978764869272709, + 0.022775648161768913, + 0.015041448175907135, + 0.024054503068327904, + 0.013923104852437973, + 0.020684240385890007, + 0.027194548398256302, + 0.006284103263169527, + 0.005376336630433798, + 0.01181778497993946, + 0.007832569070160389, + 0.007453250233083963, + 0.0058959233574569225, + 0.006921135354787111, + 0.007897275499999523, + 0.007741041015833616, + 0.007333445828408003, + 0.010323396883904934, + 0.007711593061685562, + 0.005468025803565979, + 0.008174518123269081, + 0.00718421908095479, + 0.010691527277231216, + 0.0075500174425542355, + 0.005349862389266491, + 0.015877963975071907, + 0.007648339029401541, + 0.009170523844659328, + 0.00584730040282011, + 0.008936826139688492, + 0.012963222339749336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.0036118291318416595 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a barn painted a warm, classic red hue with weathered wood panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the barn situated in lush, rolling meadows containing vibrant wildflowers in purple, yellow, and white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a rustic split-section wood fence aged to a silvery gray visible lining the property perimeter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, golden sunlight filtering through large, fluffy white clouds in a pale blue sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, eye-level wide-angle view?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062020834535360336 + ], + "opposite_logprob_tokens": [ + -2.812020778656006 + ], + "logprob_reward": -0.062020834535360336, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062020834535360336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13435731828212738, + 0.12738129496574402, + 0.1960124522447586, + 0.13685862720012665, + 0.054206639528274536, + 0.04187646508216858, + 0.109787717461586, + 0.07760464400053024, + 0.08234897255897522, + 0.03956594318151474 + ], + "credit_map": [ + 0.018916713073849678, + 0.010004379786550999, + 0.013701464980840683, + 0.013656050898134708, + 0.02048894576728344, + 0.026827620342373848, + 0.018240975216031075, + 0.012521157041192055, + 0.014513680711388588, + 0.014059370383620262, + 0.025859758257865906, + 0.012565403245389462, + 0.01358044147491455, + 0.012973016127943993, + 0.012511339038610458, + 0.021318279206752777, + 0.024325532838702202, + 0.02760852314531803, + 0.03178999572992325, + 0.020757019519805908, + 0.02612452208995819, + 0.019541455432772636, + 0.018239926546812057, + 0.02762546017765999, + 0.01623697578907013, + 0.014795452356338501, + 0.030146272853016853, + 0.012739880941808224, + 0.017264869064092636, + 0.012380245141685009, + 0.014473932795226574, + 0.0188209880143404, + 0.005672760307788849, + 0.004261037800461054, + 0.011277757585048676, + 0.006163048092275858, + 0.006225850433111191, + 0.004642859101295471, + 0.006870016921311617, + 0.009093308821320534, + 0.005953847896307707, + 0.0038331043906509876, + 0.006466015242040157, + 0.0037690605968236923, + 0.005935939960181713, + 0.004796039313077927, + 0.00566846551373601, + 0.0054539903067052364, + 0.012338319793343544, + 0.013015612959861755, + 0.015834808349609375, + 0.013409034349024296, + 0.013876768760383129, + 0.009640682488679886, + 0.013264356181025505, + 0.018408119678497314, + 0.009329953230917454, + 0.012174860574305058, + 0.008116161450743675, + 0.01039030309766531, + 0.009751002304255962, + 0.007944640703499317, + 0.007771293632686138, + 0.01212642528116703, + 0.009010592475533485, + 0.011672264896333218, + 0.013120290823280811, + 0.013384252786636353, + 0.008143717423081398, + 0.007645638659596443, + 0.010797453112900257, + 0.008574750274419785, + 0.004060856532305479, + 0.0033990126103162766, + 0.008458084426820278, + 0.0044122193939983845, + 0.004041779786348343, + 0.0029835482127964497, + 0.005114675499498844, + 0.007095762994140387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.062020834535360336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_4/index_tensor([1493, 1493])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_4/index_tensor([1493, 1493])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9e9db1286b939f0f14c9c731ece2763734073fed --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_4/index_tensor([1493, 1493])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic, extreme slow-motion shot, shot from a low, upward angle, focuses on a Ka-52 Alligator helicopter, hovering silently in the gloomy night sky. Below it looms a Gothic city shrouded in darkness, its silhouettes of countless steepled churches and sinister gargoyles looming in the thin moonlight. The helicopter's gunmetal hull is studded with sharp rivets, intricate panel lines, and angular weapon pylons. Its signature contra-rotating rotors remain frozen in the air, reflecting the faint, cold glow of the city below. Suddenly, an invisible, destructive force erupts from its core. Rather than a violent explosion, the helicopter's fuselage is gracefully yet brutally ripped apart in an eerie silence. The metal skin slowly tears, twists, and peels away like paper, revealing the complex mechanical skeleton beneath. Electrical wires drift helplessly like severed nerves, hydraulic lines emit tiny droplets of oil that solidify in slow motion, and delicate structural beams bend and break inch by inch, as if fragile bones are crumbling under irresistible pressure. Countless tiny metal fragments and screws float in the air, and each internal short circuit releases a brief, bright spark, momentarily illuminating the disintegrating mechanical wonder. The entire scene, rendered in cold tones, presents a surreal aesthetic of technological destruction, imbued with apocalyptic grandeur and tranquility.", + "seed": 5511, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_4/index_1493_seed_5511_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02904856512323022, + "vlm_logprob": -0.02904856512323022, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.02904856512323022, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11785286664962769 + ], + "opposite_logprob_tokens": [ + -2.3678529262542725 + ], + "logprob_reward": -0.11785286664962769, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11785286664962769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043192315846681595, + 0.8026360869407654, + 0.015411016531288624, + 0.0113436970859766, + 0.013768463395535946, + 0.016906805336475372, + 0.0044161430560052395, + 0.019505884498357773, + 0.018895002081990242, + 0.053924597799777985 + ], + "credit_map": [ + 0.00404566852375865, + 0.004779717419296503, + 0.0048240115866065025, + 0.005066724494099617, + 0.0052491603419184685, + 0.006535240914672613, + 0.005965101066976786, + 0.006726696155965328, + 0.09237997233867645, + 0.02203954942524433, + 0.016848169267177582, + 0.03924600034952164, + 0.027405202388763428, + 0.19437409937381744, + 0.3341323137283325, + 0.07621081173419952, + 0.0018305464182049036, + 0.0009496480342932045, + 0.0014369726413860917, + 0.0015520867891609669, + 0.0015918646240606904, + 0.003982914611697197, + 0.0016535837203264236, + 0.0024134006816893816, + 0.001498050638474524, + 0.0011535733938217163, + 0.0009826463647186756, + 0.0017023071413859725, + 0.001356899505481124, + 0.001244245213456452, + 0.0018035275861620903, + 0.0016024478245526552, + 0.0020655544940382242, + 0.001439784187823534, + 0.0012124734930694103, + 0.001405986025929451, + 0.002171325497329235, + 0.001418171334080398, + 0.002045515924692154, + 0.002009653951972723, + 0.0019431712571531534, + 0.0012601225171238184, + 0.0016798075521364808, + 0.0022226152941584587, + 0.003322148695588112, + 0.001961352536454797, + 0.0026167440228164196, + 0.00190084392670542, + 0.0007057567127048969, + 0.0005457833758555353, + 0.0006731514004059136, + 0.0006968057714402676, + 0.00046527408994734287, + 0.0005774754099547863, + 0.00040747359162196517, + 0.00034442346077412367, + 0.0020928436424583197, + 0.0016810678644105792, + 0.003018478397279978, + 0.004242320545017719, + 0.0020962166599929333, + 0.002582025481387973, + 0.0025646211579442024, + 0.001228311681188643, + 0.0017442406387999654, + 0.0015513519756495953, + 0.0019739442504942417, + 0.0031560808420181274, + 0.0019474055152386427, + 0.003394193947315216, + 0.0025172436144202948, + 0.0026105442084372044, + 0.01575641892850399, + 0.002867515664547682, + 0.002890197094529867, + 0.005042072385549545, + 0.014656043611466885, + 0.0037896649446338415, + 0.007097302004694939, + 0.001825388171710074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.11785286664962769 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009821425192058086 + ], + "opposite_logprob_tokens": [ + -4.63482141494751 + ], + "logprob_reward": -0.009821425192058086, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009821425192058086, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09046387672424316, + 0.6466198563575745, + 0.032020509243011475, + 0.030953451991081238, + 0.05381174385547638, + 0.016128767281770706, + 0.015853190794587135, + 0.04925266280770302, + 0.03637531027197838, + 0.02852058969438076 + ], + "credit_map": [ + 0.007690771017223597, + 0.010837049223482609, + 0.010204127989709377, + 0.009411122649908066, + 0.014783087186515331, + 0.010962162166833878, + 0.012548644095659256, + 0.014026915654540062, + 0.07254412770271301, + 0.017778640612959862, + 0.014553523622453213, + 0.031479187309741974, + 0.022644709795713425, + 0.15803393721580505, + 0.2683432400226593, + 0.06124251335859299, + 0.004066606983542442, + 0.001698794192634523, + 0.0025271051563322544, + 0.0020258771255612373, + 0.003394288709387183, + 0.009788109920918941, + 0.0034858211874961853, + 0.005033908877521753, + 0.00402861787006259, + 0.0012346695875748992, + 0.0071522812359035015, + 0.0028772582300007343, + 0.002465982921421528, + 0.0067099761217832565, + 0.0033738259226083755, + 0.003110840916633606, + 0.008533532731235027, + 0.0051232618279755116, + 0.004225905053317547, + 0.004558410029858351, + 0.00791971292346716, + 0.005603370722383261, + 0.007018734235316515, + 0.010828820988535881, + 0.0022048968821763992, + 0.001142411958426237, + 0.001752562471665442, + 0.0015750073362141848, + 0.0026643499732017517, + 0.0018979982705786824, + 0.0024837206583470106, + 0.0024078222922980785, + 0.0028785662725567818, + 0.0015137442387640476, + 0.0017196439439430833, + 0.00171481107827276, + 0.0031845869962126017, + 0.00219017849303782, + 0.0014027287252247334, + 0.0012489320943132043, + 0.005987195763736963, + 0.004244767129421234, + 0.006334483157843351, + 0.01126155350357294, + 0.006213934626430273, + 0.0064401328563690186, + 0.005490733776241541, + 0.0032798643223941326, + 0.00291431020013988, + 0.0028826254419982433, + 0.0029133360367268324, + 0.006366770714521408, + 0.004428420681506395, + 0.010214990004897118, + 0.003078467445448041, + 0.0035763902124017477, + 0.005968054290860891, + 0.0029647983610630035, + 0.0016594502376392484, + 0.002832459518685937, + 0.007056732662022114, + 0.0022811407689005136, + 0.0036608679220080376, + 0.0020970881450921297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.009821425192058086 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012542887590825558 + ], + "opposite_logprob_tokens": [ + -4.387542724609375 + ], + "logprob_reward": -0.012542887590825558, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012542887590825558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09353238344192505, + 0.6910310983657837, + 0.023890729993581772, + 0.028756033629179, + 0.014961888082325459, + 0.029615409672260284, + 0.005384054500609636, + 0.06745947897434235, + 0.01973295584321022, + 0.0256359726190567 + ], + "credit_map": [ + 0.01114956010133028, + 0.007952585816383362, + 0.011295408941805363, + 0.013808200135827065, + 0.006835755426436663, + 0.013676481321454048, + 0.01541112270206213, + 0.01340327225625515, + 0.08154867589473724, + 0.019639374688267708, + 0.017416425049304962, + 0.03500119596719742, + 0.024469798430800438, + 0.16077086329460144, + 0.285325825214386, + 0.06685895472764969, + 0.0033293135929852724, + 0.0015897463308647275, + 0.003157340921461582, + 0.0022174546029418707, + 0.0022905163932591677, + 0.00558982789516449, + 0.002517316723242402, + 0.0031992143485695124, + 0.004328175913542509, + 0.0012937594437971711, + 0.006498384289443493, + 0.0026630174834281206, + 0.0018171928822994232, + 0.0064418138936161995, + 0.0025685669388622046, + 0.0031451229006052017, + 0.0026605722960084677, + 0.0017191640799865127, + 0.0011329324916005135, + 0.0016877650050446391, + 0.002547046635299921, + 0.001455203746445477, + 0.0015589564573019743, + 0.002200246788561344, + 0.004333960823714733, + 0.002086605178192258, + 0.002582666464149952, + 0.0030337884090840816, + 0.0069402907975018024, + 0.003383569186553359, + 0.0036380216479301453, + 0.0036165083292871714, + 0.0009381809504702687, + 0.00044978069490753114, + 0.0004250514321029186, + 0.0005240883911028504, + 0.0013374232221394777, + 0.000772631901782006, + 0.00036868167808279395, + 0.000568216375540942, + 0.008102262392640114, + 0.005569200962781906, + 0.012210024520754814, + 0.013833492994308472, + 0.007028493098914623, + 0.009287451393902302, + 0.00709297601133585, + 0.004335578065365553, + 0.001773123862221837, + 0.0014561157440766692, + 0.0020426760893315077, + 0.0035620355047285557, + 0.0021294401958584785, + 0.004514995962381363, + 0.0019916840828955173, + 0.0022628840524703264, + 0.006816194858402014, + 0.0017074906500056386, + 0.0019670070614665747, + 0.002607210772112012, + 0.005317992530763149, + 0.0018406795570626855, + 0.003966328222304583, + 0.001413068501278758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.012542887590825558 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028234163764864206 + ], + "opposite_logprob_tokens": [ + -5.877823352813721 + ], + "logprob_reward": -0.0028234163764864206, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028234163764864206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06942221522331238, + 0.5378808379173279, + 0.04047197103500366, + 0.03305458277463913, + 0.030824508517980576, + 0.05537329614162445, + 0.02055649273097515, + 0.038562387228012085, + 0.06100626289844513, + 0.11284748464822769 + ], + "credit_map": [ + 0.007156485691666603, + 0.006706310901790857, + 0.009263752959668636, + 0.010756970383226871, + 0.009008004330098629, + 0.009263377636671066, + 0.007614050526171923, + 0.009653255343437195, + 0.06270831823348999, + 0.014606562443077564, + 0.012666848488152027, + 0.026279354467988014, + 0.019953889772295952, + 0.12811699509620667, + 0.22111858427524567, + 0.05243032053112984, + 0.005874050781130791, + 0.002628220710903406, + 0.005345548037439585, + 0.0030693255830556154, + 0.0042191739194095135, + 0.009903092868626118, + 0.003972724545747042, + 0.005459834821522236, + 0.004186431411653757, + 0.0034219424705952406, + 0.004348078742623329, + 0.003959095571190119, + 0.00403589429333806, + 0.004002477042376995, + 0.005398788955062628, + 0.0037018759176135063, + 0.004984007682651281, + 0.0028404563199728727, + 0.0036521474830806255, + 0.003127784002572298, + 0.004724536556750536, + 0.004683984909206629, + 0.002801918424665928, + 0.004009673837572336, + 0.005928979255259037, + 0.003322542179375887, + 0.00831395760178566, + 0.00696679949760437, + 0.008506412617862225, + 0.006358160171657801, + 0.00985876377671957, + 0.006117676850408316, + 0.0029650493524968624, + 0.0020933065097779036, + 0.002807245124131441, + 0.0026833047159016132, + 0.0025670176837593317, + 0.0023261324968189, + 0.0024475441314280033, + 0.0026668920181691647, + 0.004468151833862066, + 0.0028146449476480484, + 0.004406238440424204, + 0.004085093270987272, + 0.005949027370661497, + 0.005742458160966635, + 0.007916625589132309, + 0.0031801501754671335, + 0.006627316121011972, + 0.0047076246701180935, + 0.008202403783798218, + 0.008914466015994549, + 0.006729308050125837, + 0.011776501312851906, + 0.007560690399259329, + 0.006487952079623938, + 0.02779565565288067, + 0.016453437507152557, + 0.007714482955634594, + 0.012407484464347363, + 0.016248781234025955, + 0.00749780610203743, + 0.014529125764966011, + 0.010200702585279942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.0028234163764864206 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002202229807153344 + ], + "opposite_logprob_tokens": [ + -6.127202033996582 + ], + "logprob_reward": -0.002202229807153344, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002202229807153344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08276781439781189, + 0.10504066199064255, + 0.02943607047200203, + 0.02672027237713337, + 0.11459101736545563, + 0.0748986080288887, + 0.02640555426478386, + 0.383777916431427, + 0.04800986126065254, + 0.10835225880146027 + ], + "credit_map": [ + 0.014619172550737858, + 0.008000153116881847, + 0.008595239371061325, + 0.01695553958415985, + 0.0062307496555149555, + 0.009764361195266247, + 0.010086745023727417, + 0.00851584691554308, + 0.015078159980475903, + 0.004981836304068565, + 0.009365024045109749, + 0.008280055597424507, + 0.009358922950923443, + 0.026947341859340668, + 0.020994683727622032, + 0.010034636594355106, + 0.003704434959217906, + 0.0034208442084491253, + 0.002466580830514431, + 0.004335425328463316, + 0.0038562093395739794, + 0.004104620777070522, + 0.004343806765973568, + 0.0032041489612311125, + 0.003953027538955212, + 0.0019721759017556906, + 0.0035040387883782387, + 0.003693781793117523, + 0.0025110719725489616, + 0.003534722840413451, + 0.004337741993367672, + 0.0032137108501046896, + 0.019569339230656624, + 0.011697480455040932, + 0.009253657422959805, + 0.012217666022479534, + 0.019102705642580986, + 0.01373907458037138, + 0.013014711439609528, + 0.015996381640434265, + 0.01068512350320816, + 0.006169966887682676, + 0.007672981359064579, + 0.007988055236637592, + 0.016594210639595985, + 0.008852217346429825, + 0.007538688834756613, + 0.009397360496222973, + 0.0037258751690387726, + 0.0029898902866989374, + 0.00299456506036222, + 0.002882653381675482, + 0.006279094144701958, + 0.003373820800334215, + 0.001957203494384885, + 0.0022024521604180336, + 0.039750054478645325, + 0.03003043122589588, + 0.07450538873672485, + 0.0772339478135109, + 0.03801725432276726, + 0.06260769069194794, + 0.040952183306217194, + 0.02068096399307251, + 0.005718102212995291, + 0.0038124846760183573, + 0.0069997357204556465, + 0.00547695392742753, + 0.005527012515813112, + 0.008200030773878098, + 0.005727808456867933, + 0.0065477341413497925, + 0.016629066318273544, + 0.019413979724049568, + 0.01144267525523901, + 0.012653953395783901, + 0.013461295515298843, + 0.010545175522565842, + 0.011817824095487595, + 0.012388288974761963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 9 + ], + "selected_reward": -0.002202229807153344 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, extreme slow-motion shot, shot from a low, upward angle, focuses on a Ka-52 Alligator helicopter, hovering silently in the gloomy night sky. Below it looms a Gothic city shrouded in darkness, its silhouettes of countless steepled churches and sinister gargoyles looming in the thin moonlight. The helicopter's gunmetal hull is studded with sharp rivets, intricate panel lines, and angular weapon pylons. Its signature contra-rotating rotors remain frozen in the air, reflecting the faint, cold glow of the city below. Suddenly, an invisible, destructive force erupts from its core. Rather than a violent explosion, the helicopter's fuselage is gracefully yet brutally ripped apart in an eerie silence. The metal skin slowly tears, twists, and peels away like paper, revealing the complex mechanical skeleton beneath. Electrical wires drift helplessly like severed nerves, hydraulic lines emit tiny droplets of oil that solidify in slow motion, and delicate structural beams bend and break inch by inch, as if fragile bones are crumbling under irresistible pressure. Countless tiny metal fragments and screws float in the air, and each internal short circuit releases a brief, bright spark, momentarily illuminating the disintegrating mechanical wonder. The entire scene, rendered in cold tones, presents a surreal aesthetic of technological destruction, imbued with apocalyptic grandeur and tranquility.", + "seed": 5512, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_4/index_1493_seed_5512_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.026532732974737884, + "vlm_logprob": -0.026532732974737884, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.026532732974737884, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11401837319135666 + ], + "opposite_logprob_tokens": [ + -2.364018440246582 + ], + "logprob_reward": -0.11401837319135666, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11401837319135666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10096413642168045, + 0.2225496917963028, + 0.056909553706645966, + 0.12215316295623779, + 0.0806376188993454, + 0.14628131687641144, + 0.07585135847330093, + 0.030869198963046074, + 0.08067221939563751, + 0.0831117033958435 + ], + "credit_map": [ + 0.013454485684633255, + 0.010963392443954945, + 0.01507885754108429, + 0.00829846691340208, + 0.013211434707045555, + 0.008332706987857819, + 0.01703839935362339, + 0.014586393721401691, + 0.0266265906393528, + 0.023481052368879318, + 0.02643333561718464, + 0.01916722021996975, + 0.02542070485651493, + 0.03946778550744057, + 0.03406210243701935, + 0.02789090946316719, + 0.006162173580378294, + 0.006673477124422789, + 0.00777352973818779, + 0.00606150645762682, + 0.006003204267472029, + 0.006183851044625044, + 0.009204556234180927, + 0.008847257122397423, + 0.017709432169795036, + 0.013366230763494968, + 0.01893269456923008, + 0.01578458398580551, + 0.008942906744778156, + 0.01181416492909193, + 0.01910450868308544, + 0.016498642042279243, + 0.009242297150194645, + 0.011988110840320587, + 0.013165990822017193, + 0.011933519504964352, + 0.009802166372537613, + 0.006787600927054882, + 0.008590159937739372, + 0.009127771481871605, + 0.01666547730565071, + 0.014439658261835575, + 0.020740797743201256, + 0.01641894318163395, + 0.017188670113682747, + 0.023606278002262115, + 0.013597537763416767, + 0.023623960092663765, + 0.0076837362721562386, + 0.008341417647898197, + 0.011834150180220604, + 0.010896311141550541, + 0.005448025651276112, + 0.01042119413614273, + 0.010444940067827702, + 0.010781582444906235, + 0.0038363803178071976, + 0.004766132682561874, + 0.004299581050872803, + 0.0032609396148473024, + 0.0025089432019740343, + 0.00272379070520401, + 0.0030720396898686886, + 0.006401393096894026, + 0.009281723760068417, + 0.008827844634652138, + 0.012839620932936668, + 0.010760667733848095, + 0.008782614953815937, + 0.010810771025717258, + 0.009465803392231464, + 0.009903177618980408, + 0.010221265256404877, + 0.00939398817718029, + 0.009210660122334957, + 0.013213014230132103, + 0.012056969106197357, + 0.009288218803703785, + 0.00816288497298956, + 0.011564692482352257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 3 + ], + "selected_reward": -0.11401837319135666 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0053076306357979774 + ], + "opposite_logprob_tokens": [ + -5.255307674407959 + ], + "logprob_reward": -0.0053076306357979774, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0053076306357979774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16710707545280457, + 0.15762004256248474, + 0.07250364124774933, + 0.10010182112455368, + 0.10202108323574066, + 0.14183537662029266, + 0.044846951961517334, + 0.05329529568552971, + 0.05452464893460274, + 0.1061440259218216 + ], + "credit_map": [ + 0.02796756662428379, + 0.01763511635363102, + 0.030720384791493416, + 0.013178923167288303, + 0.011381066404283047, + 0.017344331368803978, + 0.02500065043568611, + 0.023879047483205795, + 0.026370909065008163, + 0.01105337031185627, + 0.026839761063456535, + 0.018738437443971634, + 0.007360031362622976, + 0.015725864097476006, + 0.030371755361557007, + 0.02115992270410061, + 0.009638212621212006, + 0.007120540831238031, + 0.013563689775764942, + 0.00815525557845831, + 0.006102023646235466, + 0.006390134338289499, + 0.011839153245091438, + 0.009694638662040234, + 0.012071319855749607, + 0.009627816267311573, + 0.018169762566685677, + 0.013851379975676537, + 0.007707972079515457, + 0.008606554009020329, + 0.014929215423762798, + 0.015137811191380024, + 0.011590832844376564, + 0.014835343696177006, + 0.01200901996344328, + 0.012983505614101887, + 0.011627855710685253, + 0.009507564827799797, + 0.015109915286302567, + 0.01435705367475748, + 0.019030224531888962, + 0.012971427291631699, + 0.018307656049728394, + 0.01541763637214899, + 0.01512108463793993, + 0.020163696259260178, + 0.01698884554207325, + 0.0238348301500082, + 0.005072284955531359, + 0.004514914937317371, + 0.005755115300416946, + 0.00706531573086977, + 0.0034268288873136044, + 0.005334277171641588, + 0.005747810937464237, + 0.00793040357530117, + 0.0062756845727562904, + 0.007420725654810667, + 0.008233859203755856, + 0.004662449937313795, + 0.006109306588768959, + 0.006671465467661619, + 0.0037021527532488108, + 0.01021965779364109, + 0.005563804879784584, + 0.004991357214748859, + 0.006423654034733772, + 0.007786224130541086, + 0.007470764219760895, + 0.007113604806363583, + 0.00572214787825942, + 0.00945309642702341, + 0.016357505694031715, + 0.010236676782369614, + 0.014424772933125496, + 0.015892991796135902, + 0.018270477652549744, + 0.010765177197754383, + 0.009057920426130295, + 0.011138522066175938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0053076306357979774 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864158570766449 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.00864158570766449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864158570766449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15435270965099335, + 0.14894084632396698, + 0.08417369425296783, + 0.279224157333374, + 0.07070460170507431, + 0.07181765884160995, + 0.036230966448783875, + 0.04020671918988228, + 0.02797151729464531, + 0.08637712150812149 + ], + "credit_map": [ + 0.018707014620304108, + 0.016879292204976082, + 0.018207885324954987, + 0.014380727894604206, + 0.018704108893871307, + 0.021826662123203278, + 0.02187267877161503, + 0.023774337023496628, + 0.02747337520122528, + 0.01415182650089264, + 0.01798788644373417, + 0.013384466990828514, + 0.010303515940904617, + 0.020195066928863525, + 0.02634117379784584, + 0.019103536382317543, + 0.00973562616854906, + 0.00977720320224762, + 0.01234909612685442, + 0.01268702745437622, + 0.008263405412435532, + 0.007121379021555185, + 0.011130565777420998, + 0.01310939434915781, + 0.04068014398217201, + 0.02511267177760601, + 0.05566767603158951, + 0.04253429174423218, + 0.014615119434893131, + 0.01950627937912941, + 0.04074743762612343, + 0.040360525250434875, + 0.008919837884604931, + 0.00678524374961853, + 0.010776270180940628, + 0.012500269338488579, + 0.007528604939579964, + 0.0055867754854261875, + 0.006247303914278746, + 0.01236029714345932, + 0.00948810949921608, + 0.01223584171384573, + 0.010133223608136177, + 0.008533218875527382, + 0.007326431106775999, + 0.00619670981541276, + 0.007342218887060881, + 0.010561903938651085, + 0.004116443917155266, + 0.005026328843086958, + 0.00391749944537878, + 0.005063472781330347, + 0.0032981110271066427, + 0.004573139362037182, + 0.003671108977869153, + 0.00656486302614212, + 0.0059226881712675095, + 0.007593641057610512, + 0.004958894103765488, + 0.004487341269850731, + 0.004152868874371052, + 0.0035741443280130625, + 0.0044718049466609955, + 0.005045337602496147, + 0.0032417678739875555, + 0.0034278768580406904, + 0.004160292446613312, + 0.003630401799455285, + 0.003017289098352194, + 0.003577942494302988, + 0.002782670548185706, + 0.004133276641368866, + 0.010196281597018242, + 0.009630633518099785, + 0.00799797847867012, + 0.011091587133705616, + 0.01936511881649494, + 0.008549855090677738, + 0.007618832401931286, + 0.01192683819681406 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.00864158570766449 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024937265552580357 + ], + "opposite_logprob_tokens": [ + -6.002493858337402 + ], + "logprob_reward": -0.0024937265552580357, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024937265552580357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08649804443120956, + 0.13882170617580414, + 0.04813085123896599, + 0.03932340443134308, + 0.1656491756439209, + 0.08641288429498672, + 0.10200385749340057, + 0.036548785865306854, + 0.13309182226657867, + 0.1635194718837738 + ], + "credit_map": [ + 0.008418946526944637, + 0.00936797820031643, + 0.009216135367751122, + 0.007408875972032547, + 0.011200697161257267, + 0.009927765466272831, + 0.01614176295697689, + 0.014815891161561012, + 0.022405069321393967, + 0.011467722244560719, + 0.028417840600013733, + 0.01607581228017807, + 0.008380383253097534, + 0.013814842328429222, + 0.021168000996112823, + 0.017092041671276093, + 0.0055092740803956985, + 0.005151926539838314, + 0.007644545752555132, + 0.005036232527345419, + 0.005181563086807728, + 0.0044770510867238045, + 0.007551931776106358, + 0.007578327786177397, + 0.005338545423001051, + 0.0036718917544931173, + 0.008340951055288315, + 0.003458151128143072, + 0.0026381397619843483, + 0.0036765148397535086, + 0.006683762650936842, + 0.005515448283404112, + 0.021133437752723694, + 0.028619926422834396, + 0.01696638949215412, + 0.020044775679707527, + 0.01722806505858898, + 0.017374295741319656, + 0.020265856757760048, + 0.024016432464122772, + 0.009504664689302444, + 0.013405026867985725, + 0.017364470288157463, + 0.009829174727201462, + 0.0067927357740700245, + 0.007955312728881836, + 0.008477939292788506, + 0.013083563186228275, + 0.014197569340467453, + 0.009709229692816734, + 0.017459804192185402, + 0.015090757980942726, + 0.00811973586678505, + 0.009347797371447086, + 0.013877230696380138, + 0.014201732352375984, + 0.006280866451561451, + 0.009469733573496342, + 0.004013017285615206, + 0.0030487109906971455, + 0.004840061534196138, + 0.003764087101444602, + 0.002039328683167696, + 0.0030929783824831247, + 0.01665765419602394, + 0.01059306226670742, + 0.021728338673710823, + 0.021524876356124878, + 0.014438322745263577, + 0.01771775633096695, + 0.010102979838848114, + 0.02032882533967495, + 0.024878354743123055, + 0.01698152720928192, + 0.023595208302140236, + 0.030095873400568962, + 0.018839174881577492, + 0.021439936012029648, + 0.014713444747030735, + 0.01297596376389265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.0024937265552580357 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022023487836122513 + ], + "opposite_logprob_tokens": [ + -6.12720251083374 + ], + "logprob_reward": -0.0022023487836122513, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022023487836122513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07062577456235886, + 0.14111535251140594, + 0.05139782652258873, + 0.08090876787900925, + 0.14823156595230103, + 0.21766462922096252, + 0.05583658069372177, + 0.043684303760528564, + 0.03926927596330643, + 0.1512659192085266 + ], + "credit_map": [ + 0.010604741051793098, + 0.007033518049865961, + 0.011486328206956387, + 0.006623698864132166, + 0.008388745598495007, + 0.006111994851380587, + 0.01152680441737175, + 0.008849943988025188, + 0.0166790634393692, + 0.01117622945457697, + 0.022510100156068802, + 0.014048747718334198, + 0.01139424741268158, + 0.022031530737876892, + 0.02734380215406418, + 0.01593162678182125, + 0.006071932148188353, + 0.005085816606879234, + 0.00806249026209116, + 0.0059105767868459225, + 0.0050141713581979275, + 0.004238322377204895, + 0.008556877262890339, + 0.008457641117274761, + 0.008376502431929111, + 0.008883302100002766, + 0.01190909557044506, + 0.01480664312839508, + 0.00830975454300642, + 0.008691505528986454, + 0.01203012466430664, + 0.007901841774582863, + 0.01584548130631447, + 0.019406108185648918, + 0.01771371439099312, + 0.01924545131623745, + 0.020392227917909622, + 0.01844785176217556, + 0.019479339942336082, + 0.017701392993330956, + 0.025106413289904594, + 0.018830088898539543, + 0.03173446282744408, + 0.024823514744639397, + 0.022652244195342064, + 0.0251290425658226, + 0.02980491891503334, + 0.039583951234817505, + 0.006991246249526739, + 0.00787417497485876, + 0.0062640756368637085, + 0.006933549884706736, + 0.004971545655280352, + 0.008784518577158451, + 0.006445621140301228, + 0.007571847178041935, + 0.006498923059552908, + 0.007920251227915287, + 0.005127534735947847, + 0.004768313840031624, + 0.004708786029368639, + 0.0032833830919116735, + 0.005842429585754871, + 0.005534682888537645, + 0.004485198762267828, + 0.004271881654858589, + 0.00490995030850172, + 0.004769941326230764, + 0.006769061554223299, + 0.004880904220044613, + 0.003880412084981799, + 0.0053019276820123196, + 0.01695230044424534, + 0.01502926740795374, + 0.015703311190009117, + 0.023660628125071526, + 0.02696862258017063, + 0.016195476055145264, + 0.012941602617502213, + 0.023814713582396507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 4 + ], + "selected_reward": -0.0022023487836122513 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, extreme slow-motion shot, shot from a low, upward angle, focuses on a Ka-52 Alligator helicopter, hovering silently in the gloomy night sky. Below it looms a Gothic city shrouded in darkness, its silhouettes of countless steepled churches and sinister gargoyles looming in the thin moonlight. The helicopter's gunmetal hull is studded with sharp rivets, intricate panel lines, and angular weapon pylons. Its signature contra-rotating rotors remain frozen in the air, reflecting the faint, cold glow of the city below. Suddenly, an invisible, destructive force erupts from its core. Rather than a violent explosion, the helicopter's fuselage is gracefully yet brutally ripped apart in an eerie silence. The metal skin slowly tears, twists, and peels away like paper, revealing the complex mechanical skeleton beneath. Electrical wires drift helplessly like severed nerves, hydraulic lines emit tiny droplets of oil that solidify in slow motion, and delicate structural beams bend and break inch by inch, as if fragile bones are crumbling under irresistible pressure. Countless tiny metal fragments and screws float in the air, and each internal short circuit releases a brief, bright spark, momentarily illuminating the disintegrating mechanical wonder. The entire scene, rendered in cold tones, presents a surreal aesthetic of technological destruction, imbued with apocalyptic grandeur and tranquility.", + "seed": 5513, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_4/index_1493_seed_5513_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.031382403313182296, + "vlm_logprob": -0.031382403313182296, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.031382403313182296, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1265897899866104 + ], + "opposite_logprob_tokens": [ + -2.251589775085449 + ], + "logprob_reward": -0.1265897899866104, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1265897899866104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10487183928489685, + 0.04284358024597168, + 0.03550361841917038, + 0.032936424016952515, + 0.04500441253185272, + 0.026332449167966843, + 0.02052273415029049, + 0.05507607012987137, + 0.33256426453590393, + 0.3043445944786072 + ], + "credit_map": [ + 0.00973527692258358, + 0.005580475553870201, + 0.01829742267727852, + 0.008694141171872616, + 0.007823268882930279, + 0.012886444106698036, + 0.025553209707140923, + 0.016301598399877548, + 0.010665143840014935, + 0.002869907533749938, + 0.0019257423700764775, + 0.0034308191388845444, + 0.0022295627277344465, + 0.006204103119671345, + 0.005379069596529007, + 0.010139230638742447, + 0.007079613860696554, + 0.003512831637635827, + 0.003887687809765339, + 0.003970857709646225, + 0.0025823479518294334, + 0.007221287116408348, + 0.0036074791569262743, + 0.0036415138747543097, + 0.007402805611491203, + 0.0027853259816765785, + 0.0036357822827994823, + 0.004018109291791916, + 0.0026909890584647655, + 0.0037517924793064594, + 0.005194618366658688, + 0.0034570014104247093, + 0.0074033234268426895, + 0.005304559133946896, + 0.0048846048302948475, + 0.003557290881872177, + 0.003959251567721367, + 0.008010412566363811, + 0.005923563614487648, + 0.005961406975984573, + 0.005280749872326851, + 0.0023918699007481337, + 0.0037545396480709314, + 0.002693043788895011, + 0.0023077221121639013, + 0.0033078498672693968, + 0.0029286760836839676, + 0.003667998593300581, + 0.0030623949132859707, + 0.0019739314448088408, + 0.004099764861166477, + 0.002158327028155327, + 0.003168220166116953, + 0.002057023346424103, + 0.0022663294803351164, + 0.0017367430264130235, + 0.004360460210591555, + 0.011583034880459309, + 0.0051733204163610935, + 0.015767784789204597, + 0.006634831894189119, + 0.0028967501129955053, + 0.006480248644948006, + 0.002179637085646391, + 0.016795838251709938, + 0.06998321413993835, + 0.012079983949661255, + 0.06937523186206818, + 0.13017307221889496, + 0.013460827991366386, + 0.012478409335017204, + 0.00821769330650568, + 0.017519168555736542, + 0.035918429493904114, + 0.027062201872467995, + 0.04109903797507286, + 0.11804252862930298, + 0.03519359976053238, + 0.016885705292224884, + 0.012623944319784641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.1265897899866104 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011116711422801018 + ], + "opposite_logprob_tokens": [ + -4.5111165046691895 + ], + "logprob_reward": -0.011116711422801018, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011116711422801018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17272049188613892, + 0.18143025040626526, + 0.10066471993923187, + 0.04996778815984726, + 0.08419810980558395, + 0.017283964902162552, + 0.05168537423014641, + 0.0260455384850502, + 0.11183727532625198, + 0.2041664868593216 + ], + "credit_map": [ + 0.01850597932934761, + 0.009972687810659409, + 0.0275084488093853, + 0.01113240234553814, + 0.012937500141561031, + 0.029804131016135216, + 0.03717818856239319, + 0.025681164115667343, + 0.052191462367773056, + 0.00703733554109931, + 0.007002742495387793, + 0.012495540082454681, + 0.005732355639338493, + 0.02788957767188549, + 0.02071320451796055, + 0.048368051648139954, + 0.020122963935136795, + 0.004757999908179045, + 0.01167517714202404, + 0.007686922792345285, + 0.004510344937443733, + 0.020278537645936012, + 0.0130482641980052, + 0.018584515899419785, + 0.011597983539104462, + 0.006030168384313583, + 0.004529664758592844, + 0.005176305305212736, + 0.004631407093256712, + 0.006247028708457947, + 0.007409173995256424, + 0.004346057772636414, + 0.013543456792831421, + 0.01222617644816637, + 0.011002670973539352, + 0.008493859320878983, + 0.005768835544586182, + 0.00962914526462555, + 0.012963791377842426, + 0.010570185258984566, + 0.0025258478708565235, + 0.0024971337988972664, + 0.002315493766218424, + 0.00166930234991014, + 0.0016010445542633533, + 0.002007757080718875, + 0.0025611896999180317, + 0.0021061962470412254, + 0.003163291607052088, + 0.002350342459976673, + 0.0036460324190557003, + 0.01775580272078514, + 0.017514768987894058, + 0.0024934415705502033, + 0.002196887042373419, + 0.0025648092851042747, + 0.003185426816344261, + 0.003886790480464697, + 0.004180310759693384, + 0.005571978632360697, + 0.0038460770156234503, + 0.001350941602140665, + 0.0025938735343515873, + 0.0014301425544545054, + 0.006053257267922163, + 0.01643506810069084, + 0.004256114363670349, + 0.03125174716114998, + 0.04137474298477173, + 0.0044970340095460415, + 0.00491762813180685, + 0.003051700070500374, + 0.014097683131694794, + 0.02897450141608715, + 0.011809744872152805, + 0.02501210756599903, + 0.07470910996198654, + 0.023816825821995735, + 0.0125819630920887, + 0.013164559379220009 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.011116711422801018 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075171530246735 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.016075171530246735, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075171530246735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2140021026134491, + 0.02536821737885475, + 0.10351041704416275, + 0.015634840354323387, + 0.12101684510707855, + 0.03053511306643486, + 0.0322677344083786, + 0.018680039793252945, + 0.21158701181411743, + 0.2273975908756256 + ], + "credit_map": [ + 0.025542017072439194, + 0.012312531471252441, + 0.04129120707511902, + 0.016460224986076355, + 0.016297614201903343, + 0.026659827679395676, + 0.043394941836595535, + 0.03204375505447388, + 0.005238373763859272, + 0.0014064953429624438, + 0.002783441450446844, + 0.0017950634937733412, + 0.0015214405721053481, + 0.005555992946028709, + 0.0029518466908484697, + 0.004115565679967403, + 0.021814826875925064, + 0.00469220383092761, + 0.014413001015782356, + 0.007507104426622391, + 0.003958373796194792, + 0.024106690660119057, + 0.010607766918838024, + 0.01641046442091465, + 0.002828399883583188, + 0.0015781486872583628, + 0.0015972454566508532, + 0.0015556903090327978, + 0.0013496128376573324, + 0.001714262180030346, + 0.003160386113449931, + 0.0018510970985516906, + 0.02049860544502735, + 0.01672067493200302, + 0.01729600317776203, + 0.010983194224536419, + 0.008345920592546463, + 0.01504750456660986, + 0.014637958258390427, + 0.017486998811364174, + 0.007144520059227943, + 0.0034508819226175547, + 0.004232021514326334, + 0.0026711514219641685, + 0.003122973022982478, + 0.0027843944262713194, + 0.003213826334103942, + 0.003915349021553993, + 0.0027772884350270033, + 0.0018641329370439053, + 0.0034071269910782576, + 0.01036540511995554, + 0.008853583596646786, + 0.0017082683043554425, + 0.001677564694546163, + 0.0016143684042617679, + 0.0026482404209673405, + 0.0033907094039022923, + 0.003913543652743101, + 0.0028909826651215553, + 0.001693874248303473, + 0.001311163417994976, + 0.0016690390184521675, + 0.0011624874314293265, + 0.01110185869038105, + 0.035960603505373, + 0.0076294271275401115, + 0.05514021962881088, + 0.07974246144294739, + 0.00838204100728035, + 0.008397741243243217, + 0.005232683848589659, + 0.013060481287539005, + 0.027609141543507576, + 0.018750935792922974, + 0.031147336587309837, + 0.08860769122838974, + 0.02531479112803936, + 0.012772602960467339, + 0.010134636424481869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.016075171530246735 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019452712731435895 + ], + "opposite_logprob_tokens": [ + -6.251945495605469 + ], + "logprob_reward": -0.0019452712731435895, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019452712731435895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020588161423802376, + 0.012618916109204292, + 0.04703711345791817, + 0.009365258738398552, + 0.04840739816427231, + 0.027243996039032936, + 0.04328117147088051, + 0.025920480489730835, + 0.4444650113582611, + 0.3210725486278534 + ], + "credit_map": [ + 0.0018661236390471458, + 0.0011121295392513275, + 0.0014510670443996787, + 0.002535433042794466, + 0.0015295386547222733, + 0.00314681651070714, + 0.005510902963578701, + 0.0034361479338258505, + 0.001421110238879919, + 0.0008421168313361704, + 0.0013693798100575805, + 0.0015256087062880397, + 0.0009734531049616635, + 0.0023531182669103146, + 0.00234279315918684, + 0.0017913351766765118, + 0.007792861666530371, + 0.0019883359782397747, + 0.007189116440713406, + 0.0033362878020852804, + 0.0020119105465710163, + 0.007895191200077534, + 0.006845233030617237, + 0.009978175163269043, + 0.0017236460698768497, + 0.0006953179254196584, + 0.0011768032563850284, + 0.001049477024935186, + 0.001373755862005055, + 0.0010409599635750055, + 0.0010995117481797934, + 0.0012057857820764184, + 0.00931513961404562, + 0.006055298261344433, + 0.005206068977713585, + 0.0038142420817166567, + 0.003623014548793435, + 0.007184097543358803, + 0.006840706802904606, + 0.00636883033439517, + 0.003676675958558917, + 0.003961102105677128, + 0.004812182858586311, + 0.0029717939905822277, + 0.002286877017468214, + 0.00256374035961926, + 0.00298431352712214, + 0.003987308591604233, + 0.0032913703471422195, + 0.002415422350168228, + 0.0030159223824739456, + 0.014488030225038528, + 0.013844282366335392, + 0.0023480765521526337, + 0.0018235007300972939, + 0.002054563257843256, + 0.0022205347195267677, + 0.005534064024686813, + 0.0023901083040982485, + 0.007304775062948465, + 0.0032706004567444324, + 0.0017716229194775224, + 0.002434168942272663, + 0.000994603498838842, + 0.02237699367105961, + 0.09383227676153183, + 0.016221653670072556, + 0.0881844162940979, + 0.17781314253807068, + 0.018070673570036888, + 0.017040129750967026, + 0.010925697162747383, + 0.01970745623111725, + 0.03754323720932007, + 0.030677011236548424, + 0.042758893221616745, + 0.12164998054504395, + 0.03738784417510033, + 0.0179545059800148, + 0.01339359488338232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.0019452712731435895 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011850723531097174 + ], + "opposite_logprob_tokens": [ + -6.751184940338135 + ], + "logprob_reward": -0.0011850723531097174, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011850723531097174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052602458745241165, + 0.17224039137363434, + 0.05736041069030762, + 0.027817798778414726, + 0.043405238538980484, + 0.08074145019054413, + 0.05225995555520058, + 0.06727191060781479, + 0.29911696910858154, + 0.14718331396579742 + ], + "credit_map": [ + 0.004653004463762045, + 0.0030551471281796694, + 0.006478236522525549, + 0.0045531452633440495, + 0.003606122452765703, + 0.008197169750928879, + 0.014086948707699776, + 0.007972689345479012, + 0.04978020861744881, + 0.006475840229541063, + 0.00697375787422061, + 0.012905420735478401, + 0.0053188977763056755, + 0.025574058294296265, + 0.0188139695674181, + 0.04639825224876404, + 0.013985431753098965, + 0.0024295991752296686, + 0.006836968939751387, + 0.004070923663675785, + 0.0025126789696514606, + 0.013397221453487873, + 0.007294444367289543, + 0.006833148188889027, + 0.005203379783779383, + 0.0025098451878875494, + 0.003010770073160529, + 0.0021075974218547344, + 0.002801942639052868, + 0.003059166483581066, + 0.00553814135491848, + 0.0035869600251317024, + 0.007084643002599478, + 0.004557873122394085, + 0.0049605644308030605, + 0.003915074747055769, + 0.003519196528941393, + 0.0051253074780106544, + 0.008176431991159916, + 0.006066151428967714, + 0.013360756449401379, + 0.007216491270810366, + 0.009077378548681736, + 0.00698948884382844, + 0.009917334653437138, + 0.010230852290987968, + 0.01466289907693863, + 0.009286263026297092, + 0.0063027506694197655, + 0.003599356161430478, + 0.008393066935241222, + 0.009787991642951965, + 0.013944882899522781, + 0.003452002303674817, + 0.0028907216619700193, + 0.0038891886360943317, + 0.008593276143074036, + 0.01298926305025816, + 0.0120150838047266, + 0.0093440106138587, + 0.00851413980126381, + 0.005011756904423237, + 0.006101030856370926, + 0.004703352693468332, + 0.015778714790940285, + 0.07549285143613815, + 0.011022567749023438, + 0.048734553158283234, + 0.11797545850276947, + 0.012170744128525257, + 0.010648376308381557, + 0.007293732836842537, + 0.00794721208512783, + 0.01735128089785576, + 0.010953866876661777, + 0.01958129182457924, + 0.055644441395998, + 0.020378176122903824, + 0.009218412451446056, + 0.006108652334660292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.0011850723531097174 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic, extreme slow-motion shot, shot from a low, upward angle, focuses on a Ka-52 Alligator helicopter, hovering silently in the gloomy night sky. Below it looms a Gothic city shrouded in darkness, its silhouettes of countless steepled churches and sinister gargoyles looming in the thin moonlight. The helicopter's gunmetal hull is studded with sharp rivets, intricate panel lines, and angular weapon pylons. Its signature contra-rotating rotors remain frozen in the air, reflecting the faint, cold glow of the city below. Suddenly, an invisible, destructive force erupts from its core. Rather than a violent explosion, the helicopter's fuselage is gracefully yet brutally ripped apart in an eerie silence. The metal skin slowly tears, twists, and peels away like paper, revealing the complex mechanical skeleton beneath. Electrical wires drift helplessly like severed nerves, hydraulic lines emit tiny droplets of oil that solidify in slow motion, and delicate structural beams bend and break inch by inch, as if fragile bones are crumbling under irresistible pressure. Countless tiny metal fragments and screws float in the air, and each internal short circuit releases a brief, bright spark, momentarily illuminating the disintegrating mechanical wonder. The entire scene, rendered in cold tones, presents a surreal aesthetic of technological destruction, imbued with apocalyptic grandeur and tranquility.", + "seed": 5514, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_4/index_1493_seed_5514_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.026737773884087802, + "vlm_logprob": -0.026737773884087802, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.026737773884087802, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10974439978599548 + ], + "opposite_logprob_tokens": [ + -2.3597443103790283 + ], + "logprob_reward": -0.10974439978599548, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10974439978599548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15585117042064667, + 0.11831490695476532, + 0.07361922413110733, + 0.10582410544157028, + 0.32598838210105896, + 0.035432275384664536, + 0.08511555939912796, + 0.05339369550347328, + 0.016825372353196144, + 0.029635265469551086 + ], + "credit_map": [ + 0.012059022672474384, + 0.010883406735956669, + 0.03821802884340286, + 0.023572275415062904, + 0.009543518535792828, + 0.021108927205204964, + 0.020401546731591225, + 0.0200644563883543, + 0.020547306165099144, + 0.010561345145106316, + 0.014454401098191738, + 0.01443251222372055, + 0.009321993216872215, + 0.022106442600488663, + 0.014283028431236744, + 0.012607892043888569, + 0.008121300488710403, + 0.006530840881168842, + 0.009992814622819424, + 0.008753837086260319, + 0.007421308197081089, + 0.009442758746445179, + 0.011981693096458912, + 0.011374677531421185, + 0.010060557164251804, + 0.00604651216417551, + 0.016586942598223686, + 0.00991248618811369, + 0.008770570158958435, + 0.014823921024799347, + 0.011907210573554039, + 0.02771591581404209, + 0.06749410182237625, + 0.012203584425151348, + 0.0365278385579586, + 0.008983812294900417, + 0.014045189134776592, + 0.12657912075519562, + 0.01403907872736454, + 0.04611566290259361, + 0.004506963770836592, + 0.005200129002332687, + 0.005261600483208895, + 0.0031696928199380636, + 0.0052301338873803616, + 0.004769747145473957, + 0.003251258749514818, + 0.004042751155793667, + 0.007095220498740673, + 0.008109917864203453, + 0.006498869974166155, + 0.006683202926069498, + 0.009710792452096939, + 0.012977492064237595, + 0.01305797602981329, + 0.02098209224641323, + 0.007249475456774235, + 0.005828521214425564, + 0.008947644382715225, + 0.006768739316612482, + 0.005908196326345205, + 0.008911519311368465, + 0.004376572091132402, + 0.005403031129390001, + 0.001961975358426571, + 0.002236725762486458, + 0.002111272420734167, + 0.0017051526810973883, + 0.0019562554080039263, + 0.0020507711451500654, + 0.0012150354450568557, + 0.00358818587847054, + 0.003482215804979205, + 0.0049186041578650475, + 0.0033073413651436567, + 0.004334562923759222, + 0.004051862750202417, + 0.003271907800808549, + 0.003166376380249858, + 0.003102394752204418 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.10974439978599548 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005294349975883961 + ], + "opposite_logprob_tokens": [ + -5.255294322967529 + ], + "logprob_reward": -0.005294349975883961, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005294349975883961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20335949957370758, + 0.12267802655696869, + 0.074707992374897, + 0.08379542827606201, + 0.07022329419851303, + 0.07622162997722626, + 0.2983931601047516, + 0.029166877269744873, + 0.026178063824772835, + 0.015276048332452774 + ], + "credit_map": [ + 0.020515913143754005, + 0.01843586377799511, + 0.04495276138186455, + 0.027583245187997818, + 0.010683164931833744, + 0.036987435072660446, + 0.02423347346484661, + 0.019967632368206978, + 0.010952897369861603, + 0.004781860858201981, + 0.0348619669675827, + 0.013918102718889713, + 0.005225122440606356, + 0.021596526727080345, + 0.01887749694287777, + 0.012464053928852081, + 0.006443314719945192, + 0.0035980690736323595, + 0.011679287068545818, + 0.00695837615057826, + 0.005181680433452129, + 0.009770297445356846, + 0.014260402880609035, + 0.01681656204164028, + 0.01200039777904749, + 0.008050380274653435, + 0.013646217063069344, + 0.00593380955979228, + 0.006273246370255947, + 0.012341908179223537, + 0.008294990286231041, + 0.017254481092095375, + 0.007899451069533825, + 0.006177480332553387, + 0.010276421904563904, + 0.004847438540309668, + 0.00510190986096859, + 0.01576521061360836, + 0.008416852913796902, + 0.011738527566194534, + 0.013858676888048649, + 0.008658756501972675, + 0.009483358822762966, + 0.0057930052280426025, + 0.00952952727675438, + 0.012910804711282253, + 0.007216453552246094, + 0.008771047927439213, + 0.05129702389240265, + 0.025091249495744705, + 0.032177139073610306, + 0.023365860804915428, + 0.02897454798221588, + 0.037228479981422424, + 0.06329119205474854, + 0.0369676798582077, + 0.0037614379543811083, + 0.003912209998816252, + 0.004685197025537491, + 0.003891485743224621, + 0.002490587532520294, + 0.004903014749288559, + 0.0025560325011610985, + 0.002966910135000944, + 0.0038675053510814905, + 0.004256410989910364, + 0.002347230911254883, + 0.0029070815071463585, + 0.00360941793769598, + 0.0037042947951704264, + 0.0019598067738115788, + 0.0035263136960566044, + 0.00207315431907773, + 0.002307961694896221, + 0.00194948457647115, + 0.002309926552698016, + 0.002198349218815565, + 0.00155687949154526, + 0.0014040347887203097, + 0.0014762572245672345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.005294349975883961 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541121803224087 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541121803224087, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541121803224087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15941987931728363, + 0.29417935013771057, + 0.12434568256139755, + 0.050464630126953125, + 0.13694599270820618, + 0.052327677607536316, + 0.06174078211188316, + 0.04121260344982147, + 0.044652897864580154, + 0.034710463136434555 + ], + "credit_map": [ + 0.014413450844585896, + 0.016086578369140625, + 0.03012842684984207, + 0.02419360727071762, + 0.01694633439183235, + 0.016097962856292725, + 0.020145894959568977, + 0.02140762470662594, + 0.027336591854691505, + 0.012625714763998985, + 0.07247435301542282, + 0.03392600268125534, + 0.014056244865059853, + 0.048046745359897614, + 0.04838327318429947, + 0.037330444902181625, + 0.010078378953039646, + 0.006910914555191994, + 0.02617488242685795, + 0.011489816009998322, + 0.007403559982776642, + 0.019158395007252693, + 0.018460499122738838, + 0.02466924488544464, + 0.0064618769101798534, + 0.006020043045282364, + 0.007498532999306917, + 0.004099028650671244, + 0.004436404909938574, + 0.006761278957128525, + 0.005217444617301226, + 0.009970022365450859, + 0.026635512709617615, + 0.011815965175628662, + 0.015865463763475418, + 0.006587985437363386, + 0.007281012833118439, + 0.04069140926003456, + 0.007844940759241581, + 0.020223699510097504, + 0.004897423088550568, + 0.005570461042225361, + 0.008526639081537724, + 0.0038180924020707607, + 0.0057025267742574215, + 0.008040135726332664, + 0.007326281163841486, + 0.008446121588349342, + 0.010304133407771587, + 0.006587048526853323, + 0.009405293501913548, + 0.005421095062047243, + 0.004744015634059906, + 0.0062904031947255135, + 0.01120370626449585, + 0.007785087451338768, + 0.005595669616013765, + 0.00437317555770278, + 0.0073354048654437065, + 0.005983609706163406, + 0.0043877805583179, + 0.006421566475182772, + 0.003937749192118645, + 0.00317764631472528, + 0.005997197236865759, + 0.006183836609125137, + 0.0037531487178057432, + 0.00461712758988142, + 0.0051790704019367695, + 0.006532532628625631, + 0.0036556627601385117, + 0.00873432494699955, + 0.004265621770173311, + 0.005575054790824652, + 0.0034907932858914137, + 0.0064319586381316185, + 0.005096083506941795, + 0.00312586291693151, + 0.0032020038925111294, + 0.0035230829380452633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.012541121803224087 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024915861431509256 + ], + "opposite_logprob_tokens": [ + -6.002491474151611 + ], + "logprob_reward": -0.0024915861431509256, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024915861431509256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22129441797733307, + 0.06375014781951904, + 0.05473995953798294, + 0.03913063555955887, + 0.15418089926242828, + 0.10279477387666702, + 0.13350431621074677, + 0.06259442865848541, + 0.05237184837460518, + 0.11563859134912491 + ], + "credit_map": [ + 0.016463354229927063, + 0.014078306034207344, + 0.05363224074244499, + 0.03380396217107773, + 0.009690632112324238, + 0.033062778413295746, + 0.029182281345129013, + 0.03138086199760437, + 0.009149654768407345, + 0.005987279117107391, + 0.007731269579380751, + 0.007824555039405823, + 0.004930566065013409, + 0.010712497867643833, + 0.007983703166246414, + 0.009430626407265663, + 0.006392719689756632, + 0.005102206487208605, + 0.008278335444629192, + 0.005743895657360554, + 0.004229404963552952, + 0.007908237166702747, + 0.00887641403824091, + 0.008208743296563625, + 0.006435994524508715, + 0.0032451325096189976, + 0.006038497667759657, + 0.004192229360342026, + 0.002828715369105339, + 0.0035574049688875675, + 0.004786431789398193, + 0.008046231232583523, + 0.024568060413002968, + 0.008406488224864006, + 0.020726609975099564, + 0.005738421343266964, + 0.007463767658919096, + 0.05408570170402527, + 0.01019368413835764, + 0.02299816533923149, + 0.011892469599843025, + 0.011090540327131748, + 0.012705002911388874, + 0.0071497573517262936, + 0.011592796072363853, + 0.01874614879488945, + 0.012091823853552341, + 0.017526231706142426, + 0.01229465939104557, + 0.01047549769282341, + 0.010752965696156025, + 0.008585200645029545, + 0.0155209731310606, + 0.018106937408447266, + 0.021175971254706383, + 0.036592110991477966, + 0.007469581440091133, + 0.005188771057873964, + 0.009408073499798775, + 0.004905663430690765, + 0.006900687236338854, + 0.006778772920370102, + 0.005874726455658674, + 0.01606815494596958, + 0.0045725638046860695, + 0.006712025031447411, + 0.0060532656498253345, + 0.007201934698969126, + 0.007436162792146206, + 0.006811237893998623, + 0.004093864466995001, + 0.009490792639553547, + 0.015047647058963776, + 0.013489266857504845, + 0.015985609963536263, + 0.020481795072555542, + 0.02041960321366787, + 0.00867533776909113, + 0.010178669355809689, + 0.011360662057995796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0024915861431509256 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Ka-52 Alligator helicopter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene shot from a low, upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a Gothic city with steepled churches and gargoyles under a gloomy night sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the helicopter's fuselage tear, twist, and peel away like paper to reveal a mechanical skeleton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the entire scene rendered in cold tones with floating metal fragments and brief bright sparks?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036174117121845484 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.0036174117121845484, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036174117121845484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13329102098941803, + 0.08018244802951813, + 0.09493899345397949, + 0.03398784250020981, + 0.14922410249710083, + 0.08153517544269562, + 0.15173006057739258, + 0.07319775223731995, + 0.10966451466083527, + 0.0922480896115303 + ], + "credit_map": [ + 0.011538843624293804, + 0.009517067112028599, + 0.027155650779604912, + 0.01802903786301613, + 0.008692391216754913, + 0.026979146525263786, + 0.016670461744070053, + 0.014708425849676132, + 0.01202512625604868, + 0.004836510866880417, + 0.011737865395843983, + 0.009529569186270237, + 0.005611697677522898, + 0.015896936878561974, + 0.010931420139968395, + 0.009613320231437683, + 0.007455328479409218, + 0.004829185549169779, + 0.015121412463486195, + 0.009689862839877605, + 0.00549016147851944, + 0.014829369261860847, + 0.017821017652750015, + 0.01970265433192253, + 0.004382592625916004, + 0.002662067534402013, + 0.005617836490273476, + 0.0033566621132194996, + 0.0024855134543031454, + 0.0036914076190441847, + 0.0035387943498790264, + 0.00825296901166439, + 0.02291703410446644, + 0.00746059650555253, + 0.02030065283179283, + 0.005082007497549057, + 0.006991446949541569, + 0.05805542320013046, + 0.007547913584858179, + 0.02086903341114521, + 0.008916412480175495, + 0.011274975724518299, + 0.008905768394470215, + 0.007676581386476755, + 0.009933779947459698, + 0.014424607157707214, + 0.008947709575295448, + 0.011455335654318333, + 0.01192532666027546, + 0.012223774567246437, + 0.010386296547949314, + 0.012161705642938614, + 0.015951795503497124, + 0.02503717504441738, + 0.024516763165593147, + 0.039527226239442825, + 0.010291341692209244, + 0.007608283311128616, + 0.01496576052159071, + 0.010680115781724453, + 0.006300457753241062, + 0.009970278479158878, + 0.0069518303498625755, + 0.0064296843484044075, + 0.020964812487363815, + 0.01556651946157217, + 0.014731016010046005, + 0.013123951852321625, + 0.013304406777024269, + 0.010418672114610672, + 0.007957148365676403, + 0.013597989454865456, + 0.011946061626076698, + 0.009724787436425686, + 0.010472230613231659, + 0.014164742082357407, + 0.01248859241604805, + 0.010503925383090973, + 0.01196863129734993, + 0.010979121550917625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.0036174117121845484 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_5/index_tensor([366, 366])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_5/index_tensor([366, 366])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..374099bbe3bbb5b66d653912dcbd829bf8668677 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_5/index_tensor([366, 366])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a long, low-angle shot, the camera moving slowly and steadily around a heavy wooden table in a dimly lit office, creating a sense of quiet observation. The entire scene is rendered in a high-quality vintage style, defined by a dark color scheme and a pervasive matte texture that absorbs light. Under the spotlight of a low-hanging desk lamp, a teenage South Asian girl with long, black hair is completely absorbed in her work. She wears thin, black-rimmed glasses perched on her nose and a simple charcoal-gray cotton T-shirt. She leans forward, her posture intensely focused, yet a subtle, playful curiosity animates her expression. Her hands move with care as she assembles a small, intricate mechanical object; her fingers steadily and deliberately place a small component with a pair of tweezers, then gently press it into place. The surrounding office is shrouded in deep shadow, and the warm, isolated lighting creates a stark contrast, highlighting the matte finish of the wooden table, the soft fabric of her shirt, and the focused intensity of her creative process, creating an atmosphere of intimacy and tranquility.", + "seed": 29059, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_5/index_366_seed_29059_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06115241427905858, + "vlm_logprob": -0.06115241427905858, + "vlm_yes_no_margin": 3.449999976158142, + "vlm_token_credit": -0.06115241427905858, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03736329823732376 + ], + "opposite_logprob_tokens": [ + -3.787363290786743 + ], + "logprob_reward": -0.03736329823732376, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.03736329823732376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2090257853269577, + 0.09544593095779419, + 0.0970090702176094, + 0.11928808689117432, + 0.09454076737165451, + 0.16925466060638428, + 0.03783843666315079, + 0.03771091252565384, + 0.09616249054670334, + 0.04372384771704674 + ], + "credit_map": [ + 0.03893427550792694, + 0.026040710508823395, + 0.020845763385295868, + 0.025846248492598534, + 0.0259273499250412, + 0.0302698016166687, + 0.021550429984927177, + 0.019611207768321037, + 0.014816767536103725, + 0.016917472705245018, + 0.006640047766268253, + 0.0073699806816875935, + 0.018119364976882935, + 0.012090791016817093, + 0.008957882411777973, + 0.01053362712264061, + 0.015358932316303253, + 0.015683073550462723, + 0.008097894489765167, + 0.007723832502961159, + 0.015475750900804996, + 0.015413138084113598, + 0.007636315189301968, + 0.011620131321251392, + 0.023488884791731834, + 0.019978147000074387, + 0.009260933846235275, + 0.010225050151348114, + 0.016910068690776825, + 0.019001927226781845, + 0.008115410804748535, + 0.012307656928896904, + 0.008956562727689743, + 0.01846952922642231, + 0.011068051680922508, + 0.007312642876058817, + 0.015488512814044952, + 0.012808780185878277, + 0.006469303276389837, + 0.013967389240860939, + 0.02411087602376938, + 0.02892870083451271, + 0.016018405556678772, + 0.011126217432320118, + 0.021559616550803185, + 0.023614339530467987, + 0.0163047444075346, + 0.027591761201620102, + 0.006315792445093393, + 0.005976364482194185, + 0.004465965088456869, + 0.0025100144557654858, + 0.005897762253880501, + 0.0058236378245055676, + 0.002809997648000717, + 0.0040389057248830795, + 0.004331052768975496, + 0.004271040204912424, + 0.006924574263393879, + 0.002967627253383398, + 0.005096535664051771, + 0.005589071195572615, + 0.0023441086523234844, + 0.006186903920024633, + 0.009487778879702091, + 0.020898550748825073, + 0.00742038618773222, + 0.006800514180213213, + 0.01991124451160431, + 0.015826158225536346, + 0.005503583699464798, + 0.010314270853996277, + 0.004194019827991724, + 0.010445961728692055, + 0.0028869996313005686, + 0.0026190823409706354, + 0.009758333675563335, + 0.006491849198937416, + 0.0022418482694774866, + 0.005085754208266735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.03736329823732376 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033730603754520416 + ], + "opposite_logprob_tokens": [ + -3.4087305068969727 + ], + "logprob_reward": -0.033730603754520416, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033730603754520416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23284122347831726, + 0.08070849627256393, + 0.09474129229784012, + 0.21091514825820923, + 0.02531225048005581, + 0.047763243317604065, + 0.02405359223484993, + 0.06143205985426903, + 0.162079319357872, + 0.06015335023403168 + ], + "credit_map": [ + 0.03125838190317154, + 0.03485757112503052, + 0.019721155986189842, + 0.021315429359674454, + 0.061482854187488556, + 0.028644265606999397, + 0.017072923481464386, + 0.01848863624036312, + 0.012591023929417133, + 0.012067736126482487, + 0.006178208161145449, + 0.00817728042602539, + 0.012471586465835571, + 0.010990153066813946, + 0.009557151235640049, + 0.00867535825818777, + 0.014634333550930023, + 0.008098789490759373, + 0.01093135867267847, + 0.009477070532739162, + 0.015982430428266525, + 0.01681070774793625, + 0.010173375718295574, + 0.008633228950202465, + 0.03548260033130646, + 0.03174206614494324, + 0.0189479049295187, + 0.017464328557252884, + 0.02765931934118271, + 0.042746126651763916, + 0.017046328634023666, + 0.019826458767056465, + 0.0029446985572576523, + 0.005150903016328812, + 0.0022897692397236824, + 0.002157108159735799, + 0.004670219961553812, + 0.002940261969342828, + 0.002079019322991371, + 0.003080269554629922, + 0.005582493729889393, + 0.00917678140103817, + 0.005457214545458555, + 0.003573949448764324, + 0.007079364266246557, + 0.007192094344645739, + 0.0030526472255587578, + 0.006648697890341282, + 0.003382632974535227, + 0.003372661769390106, + 0.0026740883477032185, + 0.0019314076052978635, + 0.004123583901673555, + 0.0035552838817238808, + 0.002066579880192876, + 0.0029473535250872374, + 0.0045461468398571014, + 0.00905675906687975, + 0.006161265075206757, + 0.005512629169970751, + 0.011325896717607975, + 0.008943501859903336, + 0.0037630132865160704, + 0.012122847139835358, + 0.01673807017505169, + 0.03335874155163765, + 0.0123945577070117, + 0.012154479511082172, + 0.035306159406900406, + 0.02527432143688202, + 0.009247914887964725, + 0.017605070024728775, + 0.005595312919467688, + 0.012339849025011063, + 0.004324357025325298, + 0.004272907506674528, + 0.01398241613060236, + 0.008925403468310833, + 0.0028182296082377434, + 0.00789487361907959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.033730603754520416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14283713698387146 + ], + "opposite_logprob_tokens": [ + -2.0178370475769043 + ], + "logprob_reward": -0.14283713698387146, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14283713698387146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08455056697130203, + 0.10679000616073608, + 0.08747989684343338, + 0.5521603226661682, + 0.04864491894841194, + 0.016388658434152603, + 0.024440154433250427, + 0.016448764130473137, + 0.026284631341695786, + 0.03681201487779617 + ], + "credit_map": [ + 0.015059123747050762, + 0.011835996061563492, + 0.010674642398953438, + 0.006775010842829943, + 0.012188568711280823, + 0.011475306004285812, + 0.008391707204282284, + 0.008150219917297363, + 0.014726188033819199, + 0.01701047271490097, + 0.007996669970452785, + 0.007541178725659847, + 0.019467530772089958, + 0.0183546282351017, + 0.009413434192538261, + 0.012279899790883064, + 0.013485134579241276, + 0.010983756743371487, + 0.011867339722812176, + 0.010539699345827103, + 0.01273066271096468, + 0.009494705125689507, + 0.009647766128182411, + 0.00873083807528019, + 0.08691643923521042, + 0.04797889664769173, + 0.06221839785575867, + 0.050966937094926834, + 0.08081185817718506, + 0.11981265991926193, + 0.05226725712418556, + 0.05118793621659279, + 0.005465180613100529, + 0.00838476326316595, + 0.004693060647696257, + 0.0036061068531125784, + 0.0067648268304765224, + 0.008302413858473301, + 0.0035593383945524693, + 0.007869231514632702, + 0.0015296287601813674, + 0.002850546734407544, + 0.0021197779569774866, + 0.0014046194264665246, + 0.0029299105517566204, + 0.00222229422070086, + 0.0015996102010831237, + 0.0017322730273008347, + 0.0027515566907823086, + 0.0037036582361906767, + 0.0021739224903285503, + 0.0020835513714700937, + 0.0051675341092050076, + 0.0038809587713330984, + 0.0017785109812393785, + 0.002900464227423072, + 0.0021872692741453648, + 0.0024046944454312325, + 0.0020298524759709835, + 0.0012888044584542513, + 0.0019831014797091484, + 0.002601229352876544, + 0.001324211247265339, + 0.002629602327942848, + 0.0024460703134536743, + 0.005080174654722214, + 0.002205764874815941, + 0.0020084388088434935, + 0.004885629750788212, + 0.004022085107862949, + 0.0023832018487155437, + 0.0032532697077840567, + 0.003529665991663933, + 0.007386027369648218, + 0.0029281238093972206, + 0.0033208089880645275, + 0.005570495035499334, + 0.005796070676296949, + 0.00339678511954844, + 0.004884041380137205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 2 + ], + "selected_reward": -0.14283713698387146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028234163764864206 + ], + "opposite_logprob_tokens": [ + -5.877823352813721 + ], + "logprob_reward": -0.0028234163764864206, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028234163764864206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1285138577222824, + 0.10676232725381851, + 0.13586856424808502, + 0.2761412560939789, + 0.05636050924658775, + 0.1203075647354126, + 0.06178664416074753, + 0.028541356325149536, + 0.05013919994235039, + 0.03557869419455528 + ], + "credit_map": [ + 0.016866929829120636, + 0.014092962257564068, + 0.013379961252212524, + 0.018262647092342377, + 0.018489990383386612, + 0.01757284440100193, + 0.013688703998923302, + 0.016159817576408386, + 0.017051005735993385, + 0.013423771597445011, + 0.016684291884303093, + 0.013169799000024796, + 0.012145948596298695, + 0.017249366268515587, + 0.00791651476174593, + 0.00912163220345974, + 0.022698983550071716, + 0.019108407199382782, + 0.013631336390972137, + 0.011324862018227577, + 0.02419924922287464, + 0.01937907189130783, + 0.012231805361807346, + 0.01329484861344099, + 0.03891010582447052, + 0.020840199664235115, + 0.03145371004939079, + 0.0246252603828907, + 0.03743397444486618, + 0.06638582795858383, + 0.028418000787496567, + 0.028074169531464577, + 0.006902455352246761, + 0.00964425690472126, + 0.005864151753485203, + 0.005525569897145033, + 0.010111860930919647, + 0.0065789250656962395, + 0.004066879395395517, + 0.007666409946978092, + 0.014950796030461788, + 0.019473575055599213, + 0.013765513896942139, + 0.009772546589374542, + 0.014419909566640854, + 0.017223753035068512, + 0.011515124700963497, + 0.019186336547136307, + 0.00615043705329299, + 0.010685086250305176, + 0.005471305921673775, + 0.004820497706532478, + 0.009538288228213787, + 0.01003800518810749, + 0.004308251664042473, + 0.010774776339530945, + 0.002962912665680051, + 0.0045935302041471004, + 0.003487739944830537, + 0.003027950180694461, + 0.00396231934428215, + 0.004337496589869261, + 0.002336198929697275, + 0.003833209630101919, + 0.004267754033207893, + 0.009750607423484325, + 0.004431440494954586, + 0.004230744205415249, + 0.008221354335546494, + 0.00896026473492384, + 0.003661117749288678, + 0.006615915801376104, + 0.003962449263781309, + 0.005326625891029835, + 0.0037454834673553705, + 0.0033544732723385096, + 0.005964187905192375, + 0.004879258573055267, + 0.0022522436920553446, + 0.006093975622206926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0028234163764864206 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900761604309082 + ], + "opposite_logprob_tokens": [ + -2.464007616043091 + ], + "logprob_reward": -0.08900761604309082, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900761604309082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12135608494281769, + 0.09331759810447693, + 0.13788633048534393, + 0.15858207643032074, + 0.05270339921116829, + 0.07282577455043793, + 0.03116130642592907, + 0.09667667746543884, + 0.07419316470623016, + 0.16129758954048157 + ], + "credit_map": [ + 0.01731647551059723, + 0.014895012602210045, + 0.014085846021771431, + 0.01128347311168909, + 0.01829678937792778, + 0.023004047572612762, + 0.012389034032821655, + 0.010085406713187695, + 0.015192447230219841, + 0.017310667783021927, + 0.00925415102392435, + 0.008949190378189087, + 0.016191668808460236, + 0.013049603439867496, + 0.006589162163436413, + 0.006780700758099556, + 0.021330485120415688, + 0.015281962230801582, + 0.01928846165537834, + 0.01601189374923706, + 0.020629407837986946, + 0.01604798622429371, + 0.013170958496630192, + 0.01612517610192299, + 0.0225653275847435, + 0.012884831987321377, + 0.01895890384912491, + 0.01200313214212656, + 0.02141599915921688, + 0.036057665944099426, + 0.016497347503900528, + 0.018198860809206963, + 0.006731225177645683, + 0.010337610729038715, + 0.006093658972531557, + 0.005268047098070383, + 0.008221333846449852, + 0.007424245122820139, + 0.0035528994631022215, + 0.005074380896985531, + 0.009189489297568798, + 0.01570870913565159, + 0.005860034842044115, + 0.004541398957371712, + 0.01340442057698965, + 0.011161050759255886, + 0.0063334982842206955, + 0.006627173628658056, + 0.0029921308159828186, + 0.005605521611869335, + 0.00329133914783597, + 0.0022364184260368347, + 0.005533865187317133, + 0.00524131441488862, + 0.0021430819761008024, + 0.004117636941373348, + 0.01043616607785225, + 0.014015673659741879, + 0.01341172493994236, + 0.009788412600755692, + 0.013156943023204803, + 0.01468893326818943, + 0.008507989346981049, + 0.012670833617448807, + 0.00706665450707078, + 0.013127310201525688, + 0.008538385853171349, + 0.005767210852354765, + 0.010605261661112309, + 0.012828350067138672, + 0.005415475461632013, + 0.010844510979950428, + 0.01753726787865162, + 0.032192062586545944, + 0.013528987765312195, + 0.015052368864417076, + 0.023525677621364594, + 0.024496853351593018, + 0.015922602266073227, + 0.019041771069169044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 2 + ], + "selected_reward": -0.08900761604309082 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a long, low-angle shot, the camera moving slowly and steadily around a heavy wooden table in a dimly lit office, creating a sense of quiet observation. The entire scene is rendered in a high-quality vintage style, defined by a dark color scheme and a pervasive matte texture that absorbs light. Under the spotlight of a low-hanging desk lamp, a teenage South Asian girl with long, black hair is completely absorbed in her work. She wears thin, black-rimmed glasses perched on her nose and a simple charcoal-gray cotton T-shirt. She leans forward, her posture intensely focused, yet a subtle, playful curiosity animates her expression. Her hands move with care as she assembles a small, intricate mechanical object; her fingers steadily and deliberately place a small component with a pair of tweezers, then gently press it into place. The surrounding office is shrouded in deep shadow, and the warm, isolated lighting creates a stark contrast, highlighting the matte finish of the wooden table, the soft fabric of her shirt, and the focused intensity of her creative process, creating an atmosphere of intimacy and tranquility.", + "seed": 29060, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_5/index_366_seed_29060_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07781931711360812, + "vlm_logprob": -0.07781931711360812, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.07781931711360812, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034674327820539474 + ], + "opposite_logprob_tokens": [ + -3.9096744060516357 + ], + "logprob_reward": -0.034674327820539474, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.034674327820539474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2369261533021927, + 0.10416650772094727, + 0.10035763680934906, + 0.08359505236148834, + 0.09289912134408951, + 0.1688690334558487, + 0.08437823504209518, + 0.05493951588869095, + 0.04522180184721947, + 0.028646957129240036 + ], + "credit_map": [ + 0.025855418294668198, + 0.03259909152984619, + 0.044848691672086716, + 0.02664853446185589, + 0.03796936571598053, + 0.037242840975522995, + 0.01541003305464983, + 0.01635216549038887, + 0.01171841099858284, + 0.011911813169717789, + 0.016607588157057762, + 0.01652134582400322, + 0.015287689864635468, + 0.01797771453857422, + 0.007245059125125408, + 0.006896894425153732, + 0.011384716257452965, + 0.011830088682472706, + 0.0173776987940073, + 0.008719337172806263, + 0.018796758726239204, + 0.016558807343244553, + 0.006871617864817381, + 0.008818610571324825, + 0.008316226303577423, + 0.010404913686215878, + 0.0117650693282485, + 0.005645230878144503, + 0.02259577438235283, + 0.013417419977486134, + 0.006775678135454655, + 0.0046747419983148575, + 0.009826987981796265, + 0.014366284012794495, + 0.011275476776063442, + 0.007795579731464386, + 0.019224409013986588, + 0.012754973955452442, + 0.008561982773244381, + 0.009093426167964935, + 0.015548434108495712, + 0.028310921043157578, + 0.021984681487083435, + 0.013658064417541027, + 0.0400763563811779, + 0.023679768666625023, + 0.01108139380812645, + 0.01452940609306097, + 0.009893500246107578, + 0.01112340483814478, + 0.006671532057225704, + 0.00673440657556057, + 0.018553009256720543, + 0.010980900377035141, + 0.010024133138358593, + 0.010397357866168022, + 0.005992793943732977, + 0.009089725092053413, + 0.006967420224100351, + 0.004007967654615641, + 0.009823966771364212, + 0.006446691695600748, + 0.004848259035497904, + 0.007762689609080553, + 0.004565659444779158, + 0.010040485300123692, + 0.004162331577390432, + 0.0034829487558454275, + 0.0072235483676195145, + 0.006532997824251652, + 0.0041404240764677525, + 0.005073408596217632, + 0.0028267186135053635, + 0.0035643393639475107, + 0.0025376940611749887, + 0.0027584275230765343, + 0.005403192713856697, + 0.0037214988842606544, + 0.0045866393484175205, + 0.0032484482508152723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.034674327820539474 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026366282254457474 + ], + "opposite_logprob_tokens": [ + -3.6513662338256836 + ], + "logprob_reward": -0.026366282254457474, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026366282254457474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17048606276512146, + 0.0846431627869606, + 0.13266433775424957, + 0.08250845968723297, + 0.19259287416934967, + 0.052420008927583694, + 0.03809491544961929, + 0.12152321636676788, + 0.04903152585029602, + 0.07603549212217331 + ], + "credit_map": [ + 0.014534799382090569, + 0.025184039026498795, + 0.021589497104287148, + 0.024908702820539474, + 0.03063618764281273, + 0.025402667000889778, + 0.016645386815071106, + 0.011584773659706116, + 0.008115990087389946, + 0.006163453683257103, + 0.01592184789478779, + 0.008210254833102226, + 0.018756449222564697, + 0.014974536374211311, + 0.005164142232388258, + 0.0073364851996302605, + 0.013445913791656494, + 0.012878353707492352, + 0.022582320496439934, + 0.009588072076439857, + 0.033047519624233246, + 0.022560831159353256, + 0.008465174585580826, + 0.010096145793795586, + 0.006608112715184689, + 0.006937363184988499, + 0.020540621131658554, + 0.003687829477712512, + 0.014226109720766544, + 0.022829344496130943, + 0.0037879126612097025, + 0.0038911672309041023, + 0.015619762241840363, + 0.021262459456920624, + 0.039741918444633484, + 0.012686552479863167, + 0.03997965529561043, + 0.03969098627567291, + 0.014353555627167225, + 0.00925796665251255, + 0.005197644233703613, + 0.009143192321062088, + 0.005952546838670969, + 0.003331112675368786, + 0.014778067357838154, + 0.0075955018401145935, + 0.003216991201043129, + 0.0032049445435404778, + 0.004192235879600048, + 0.00646802457049489, + 0.002534709870815277, + 0.0024092283565551043, + 0.008858284913003445, + 0.005436329636722803, + 0.004252313636243343, + 0.0039437878876924515, + 0.011636517941951752, + 0.02576150931417942, + 0.014913949184119701, + 0.009948033839464188, + 0.02218952216207981, + 0.018322689458727837, + 0.00879334844648838, + 0.009957629255950451, + 0.005853055045008659, + 0.013175987638533115, + 0.005131442099809647, + 0.003943212796002626, + 0.007517458405345678, + 0.00773973111063242, + 0.002543049631640315, + 0.0031275846995413303, + 0.008757596835494041, + 0.018872007727622986, + 0.006146846804767847, + 0.004996723495423794, + 0.014067714102566242, + 0.012190358713269234, + 0.003643675008788705, + 0.007360569201409817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.026366282254457474 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031787306070328 + ], + "opposite_logprob_tokens": [ + -2.35031795501709 + ], + "logprob_reward": -0.10031787306070328, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031787306070328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1471233069896698, + 0.21536485850811005, + 0.24438142776489258, + 0.05559801682829857, + 0.1566925197839737, + 0.03188823163509369, + 0.028796685859560966, + 0.031660184264183044, + 0.03491482883691788, + 0.0535799041390419 + ], + "credit_map": [ + 0.0192089956253767, + 0.0170351043343544, + 0.01932344026863575, + 0.013997744768857956, + 0.033765606582164764, + 0.018564308062195778, + 0.014738441444933414, + 0.010489660315215588, + 0.018699152395129204, + 0.017333531752228737, + 0.03623776510357857, + 0.023987384513020515, + 0.028633834794163704, + 0.03665706142783165, + 0.02041465789079666, + 0.03340147063136101, + 0.01618756726384163, + 0.014257727190852165, + 0.07001432031393051, + 0.010508517734706402, + 0.03254127502441406, + 0.06766945123672485, + 0.019466664642095566, + 0.013735906220972538, + 0.004636730998754501, + 0.00550750270485878, + 0.012679866515100002, + 0.0030058857519179583, + 0.008515719324350357, + 0.013318121433258057, + 0.0037173533346503973, + 0.004216834902763367, + 0.013370760716497898, + 0.015042233280837536, + 0.03429074212908745, + 0.009982599876821041, + 0.029722275212407112, + 0.03518878296017647, + 0.011177299544215202, + 0.00791783258318901, + 0.0039381911046803, + 0.004862415604293346, + 0.003592823399230838, + 0.0025686677545309067, + 0.00718876114115119, + 0.0045354110188782215, + 0.002585403621196747, + 0.002616559388116002, + 0.002991662360727787, + 0.0057132663205266, + 0.0034365314058959484, + 0.002520937705412507, + 0.004637932404875755, + 0.004067406989634037, + 0.002436678158119321, + 0.0029922695830464363, + 0.003405270166695118, + 0.005961108487099409, + 0.0034475233405828476, + 0.0022943136282265186, + 0.0070623005740344524, + 0.004393207374960184, + 0.0023074173368513584, + 0.002789041493088007, + 0.003761210246011615, + 0.004594695754349232, + 0.003883630270138383, + 0.0024781206157058477, + 0.004918408580124378, + 0.004844390321522951, + 0.0037055877037346363, + 0.006728785578161478, + 0.006588818971067667, + 0.0060174427926540375, + 0.005716417916119099, + 0.0040504783391952515, + 0.008442815393209457, + 0.00707587692886591, + 0.0051430300809443, + 0.010545019060373306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.10031787306070328 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200921531766653 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.002200921531766653, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200921531766653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21430224180221558, + 0.12324253469705582, + 0.14134131371974945, + 0.12407343834638596, + 0.06233564391732216, + 0.04248368740081787, + 0.05395348742604256, + 0.08367928862571716, + 0.1078650951385498, + 0.04672330990433693 + ], + "credit_map": [ + 0.022488821297883987, + 0.03725682944059372, + 0.03144192323088646, + 0.015200565569102764, + 0.04015889763832092, + 0.029915105551481247, + 0.01705183833837509, + 0.02078823558986187, + 0.013835969381034374, + 0.01277663093060255, + 0.02036011591553688, + 0.01585349626839161, + 0.019692037254571915, + 0.016603482887148857, + 0.012172556482255459, + 0.01194823905825615, + 0.015580382198095322, + 0.012523815967142582, + 0.02576851285994053, + 0.010212170891463757, + 0.026561589911580086, + 0.02457263134419918, + 0.012209477834403515, + 0.013912728056311607, + 0.009211885742843151, + 0.010609454475343227, + 0.02974621020257473, + 0.007215396035462618, + 0.023769060149788857, + 0.029844673350453377, + 0.007972449995577335, + 0.00570430513471365, + 0.006604430265724659, + 0.007590234279632568, + 0.011461718939244747, + 0.00346567970700562, + 0.016077041625976562, + 0.010253236629068851, + 0.0034825457260012627, + 0.0034007506910711527, + 0.0049653728492558, + 0.006500189192593098, + 0.005886252503842115, + 0.0037931599654257298, + 0.008085829205811024, + 0.005704597104340792, + 0.0039597381837666035, + 0.0035885449033230543, + 0.005530951544642448, + 0.011112285777926445, + 0.006046769674867392, + 0.004710683599114418, + 0.010056672617793083, + 0.007036966271698475, + 0.004771889187395573, + 0.004687263630330563, + 0.011593780480325222, + 0.016208723187446594, + 0.008119223639369011, + 0.0061135063879191875, + 0.015004746615886688, + 0.01245272345840931, + 0.005363807547837496, + 0.008822768926620483, + 0.01468846295028925, + 0.02751525677740574, + 0.011031482368707657, + 0.007871775887906551, + 0.01542748510837555, + 0.01634354144334793, + 0.006979865953326225, + 0.008007220923900604, + 0.005797712132334709, + 0.011184639297425747, + 0.003016703762114048, + 0.002738366136327386, + 0.008899612352252007, + 0.006944052875041962, + 0.0031796784605830908, + 0.00496253976598382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.002200921531766653 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22553718090057373 + ], + "opposite_logprob_tokens": [ + -1.6005371809005737 + ], + "logprob_reward": -0.22553718090057373, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22553718090057373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1392631232738495, + 0.10567443817853928, + 0.14523984491825104, + 0.12479529529809952, + 0.1053827553987503, + 0.0360625796020031, + 0.08941468596458435, + 0.08448541164398193, + 0.09571923315525055, + 0.07396269589662552 + ], + "credit_map": [ + 0.015759184956550598, + 0.01867382973432541, + 0.02315986528992653, + 0.01754607819020748, + 0.02162049524486065, + 0.01970105804502964, + 0.011072050780057907, + 0.011730553582310677, + 0.013829614967107773, + 0.024743124842643738, + 0.015257303602993488, + 0.010322785936295986, + 0.015068290755152702, + 0.014911492355167866, + 0.0066842264495790005, + 0.004857603460550308, + 0.009933602064847946, + 0.014285712502896786, + 0.0455716997385025, + 0.0061883567832410336, + 0.014341375790536404, + 0.04561734199523926, + 0.005078539717942476, + 0.004223207477480173, + 0.011213980615139008, + 0.012707182206213474, + 0.025780757889151573, + 0.008985966444015503, + 0.02731916680932045, + 0.02442197874188423, + 0.008197270333766937, + 0.006168987136334181, + 0.008558880537748337, + 0.0104354377835989, + 0.022369200363755226, + 0.007658816408365965, + 0.01917983777821064, + 0.022607915103435516, + 0.008524334989488125, + 0.0060483370907604694, + 0.0036587456706911325, + 0.005856723990291357, + 0.0042526256293058395, + 0.003568302607163787, + 0.008349679410457611, + 0.004696345422416925, + 0.0025760303251445293, + 0.003104125615209341, + 0.008614892140030861, + 0.018909655511379242, + 0.010066338814795017, + 0.007170663680881262, + 0.018570829182863235, + 0.011747186072170734, + 0.006828309968113899, + 0.0075068119913339615, + 0.01111904438585043, + 0.022256139665842056, + 0.007426904514431953, + 0.0061794607900083065, + 0.013302377425134182, + 0.013871252536773682, + 0.004483204334974289, + 0.005847022868692875, + 0.012132542207837105, + 0.02717125415802002, + 0.008466552011668682, + 0.007185730617493391, + 0.013601972721517086, + 0.015197432599961758, + 0.005750076845288277, + 0.00621366873383522, + 0.009144168347120285, + 0.01023942045867443, + 0.006254301872104406, + 0.005013945512473583, + 0.013381369411945343, + 0.009698157198727131, + 0.006027060095220804, + 0.014204270206391811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.22553718090057373 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a long, low-angle shot, the camera moving slowly and steadily around a heavy wooden table in a dimly lit office, creating a sense of quiet observation. The entire scene is rendered in a high-quality vintage style, defined by a dark color scheme and a pervasive matte texture that absorbs light. Under the spotlight of a low-hanging desk lamp, a teenage South Asian girl with long, black hair is completely absorbed in her work. She wears thin, black-rimmed glasses perched on her nose and a simple charcoal-gray cotton T-shirt. She leans forward, her posture intensely focused, yet a subtle, playful curiosity animates her expression. Her hands move with care as she assembles a small, intricate mechanical object; her fingers steadily and deliberately place a small component with a pair of tweezers, then gently press it into place. The surrounding office is shrouded in deep shadow, and the warm, isolated lighting creates a stark contrast, highlighting the matte finish of the wooden table, the soft fabric of her shirt, and the focused intensity of her creative process, creating an atmosphere of intimacy and tranquility.", + "seed": 29061, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_5/index_366_seed_29061_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07215621345676482, + "vlm_logprob": -0.07215621345676482, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.07215621345676482, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06891395151615143 + ], + "opposite_logprob_tokens": [ + -3.1939139366149902 + ], + "logprob_reward": -0.06891395151615143, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06891395151615143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01629343442618847, + 0.011051586829125881, + 0.7648652195930481, + 0.01873840019106865, + 0.14300371706485748, + 0.007516106590628624, + 0.008625875227153301, + 0.00687534362077713, + 0.0059592947363853455, + 0.017071083188056946 + ], + "credit_map": [ + 0.001707069342955947, + 0.002515916246920824, + 0.0020413629245013, + 0.00211335183121264, + 0.002485880395397544, + 0.0019627693109214306, + 0.0013699322007596493, + 0.0020971510093659163, + 0.0013504887465387583, + 0.0015430452767759562, + 0.0016149948351085186, + 0.0010776116978377104, + 0.0015297859208658338, + 0.0017563350265845656, + 0.0008603267488069832, + 0.0013189992168918252, + 0.02042190544307232, + 0.017580091953277588, + 0.004634341225028038, + 0.0045328582637012005, + 0.3549899160861969, + 0.3481251299381256, + 0.005881285294890404, + 0.00869970116764307, + 0.001710383454337716, + 0.0015606213128194213, + 0.001041311421431601, + 0.0006938673905096948, + 0.006005288567394018, + 0.005188454873859882, + 0.0006476507405750453, + 0.0018908227793872356, + 0.003907662816345692, + 0.003780363593250513, + 0.0016280763084068894, + 0.0013859502505511045, + 0.0654386654496193, + 0.06343032419681549, + 0.0012145701330155134, + 0.00221810070797801, + 0.0007552728638984263, + 0.0010896569583564997, + 0.0008025309653021395, + 0.0004432846908457577, + 0.0015532649122178555, + 0.0014889456797391176, + 0.0004046793037559837, + 0.000978470896370709, + 0.0012372516794130206, + 0.0011378929484635592, + 0.001170051866210997, + 0.0005926536396145821, + 0.0019020618638023734, + 0.0012129415990784764, + 0.0006379918195307255, + 0.0007350297528319061, + 0.000726242084056139, + 0.0010768553474918008, + 0.001409117947332561, + 0.0006000316934660077, + 0.001256598625332117, + 0.0007582853431813419, + 0.0004319525323808193, + 0.0006162605131976306, + 0.0005222017643973231, + 0.0010990723967552185, + 0.0007379173766821623, + 0.0006910117226652801, + 0.0012177921598777175, + 0.0008025499992072582, + 0.0004158449301030487, + 0.00047290450311265886, + 0.0012467792257666588, + 0.0030020177364349365, + 0.0043705906718969345, + 0.0011435180203989148, + 0.002147813094779849, + 0.0037433309480547905, + 0.0006170243141241372, + 0.0008000095258466899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.06891395151615143 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486975833773613 + ], + "opposite_logprob_tokens": [ + -3.0486974716186523 + ], + "logprob_reward": -0.0486975833773613, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486975833773613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056559573858976364, + 0.011944816447794437, + 0.7962744832038879, + 0.006585296709090471, + 0.022369492799043655, + 0.024657081812620163, + 0.03370252624154091, + 0.020310046151280403, + 0.004776806104928255, + 0.022819839417934418 + ], + "credit_map": [ + 0.006452648434787989, + 0.007252049632370472, + 0.005131347570568323, + 0.0045784469693899155, + 0.01266630832105875, + 0.011151513084769249, + 0.00585918128490448, + 0.0034680792596191168, + 0.0011734835570678115, + 0.0018493473762646317, + 0.0015138235175982118, + 0.0011229743249714375, + 0.002189844148233533, + 0.002540712943300605, + 0.000956595700699836, + 0.0005980347050353885, + 0.021050678566098213, + 0.01829293742775917, + 0.004888023715466261, + 0.0048150960355997086, + 0.3698503077030182, + 0.3625072240829468, + 0.006015262100845575, + 0.00885497871786356, + 0.000669286644551903, + 0.0007341498276218772, + 0.00045261040213517845, + 0.000306758563965559, + 0.0019553122110664845, + 0.001744298730045557, + 0.0002599891449790448, + 0.0004628915339708328, + 0.0013016702141612768, + 0.0018397151725366712, + 0.0011539836414158344, + 0.0008778577321209013, + 0.007733454927802086, + 0.007486483082175255, + 0.0009084572084248066, + 0.001067870995029807, + 0.003153319237753749, + 0.0039146775379776955, + 0.0028966849204152822, + 0.001510279718786478, + 0.005339365918189287, + 0.005374765954911709, + 0.0012464801548048854, + 0.001221507671289146, + 0.007224707398563623, + 0.00455664424225688, + 0.007683692965656519, + 0.0018565789796411991, + 0.0048683276399970055, + 0.00401899591088295, + 0.0016346722841262817, + 0.0018589041428640485, + 0.0021485956385731697, + 0.0030359947122633457, + 0.0036420247051864862, + 0.0017460489179939032, + 0.003674133913591504, + 0.0031321232672780752, + 0.001015246962197125, + 0.0019158783834427595, + 0.000368894194252789, + 0.0008221089956350625, + 0.00044242703006602824, + 0.000438960938481614, + 0.0009645435493439436, + 0.0009708763100206852, + 0.0003444456378929317, + 0.00042454921640455723, + 0.0012291016755625606, + 0.0020709754899144173, + 0.007902010343968868, + 0.0011819684877991676, + 0.0015084878541529179, + 0.006271762773394585, + 0.001097168424166739, + 0.0015583642525598407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.0486975833773613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270654797554016 + ], + "opposite_logprob_tokens": [ + -2.127065420150757 + ], + "logprob_reward": -0.1270654797554016, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270654797554016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021012291312217712, + 0.0022200800012797117, + 0.9369610548019409, + 0.0010680416598916054, + 0.014702754095196724, + 0.0029431493021547794, + 0.0030316098127514124, + 0.003361588343977928, + 0.002164968755096197, + 0.012534509412944317 + ], + "credit_map": [ + 0.0023317798040807247, + 0.0020715217106044292, + 0.002128317952156067, + 0.0017690581735223532, + 0.004534244537353516, + 0.003651439445093274, + 0.0024680623319000006, + 0.002057865960523486, + 0.00030282093212008476, + 0.0003002795856446028, + 0.0003259749209973961, + 0.00017173985543195158, + 0.0003087646618951112, + 0.00028495080186985433, + 0.0001900701317936182, + 0.0003354790969751775, + 0.025367368012666702, + 0.02114749699831009, + 0.005685596261173487, + 0.005599719937890768, + 0.43570736050605774, + 0.425437331199646, + 0.007158334832638502, + 0.010857737623155117, + 9.895517723634839e-05, + 0.00011127930338261649, + 0.00010425507934996858, + 8.761446952121332e-05, + 0.0002249414537800476, + 0.0002071186463581398, + 6.955207936698571e-05, + 0.00016432539268862456, + 0.00048453189083375037, + 0.000546504685189575, + 0.00033884961158037186, + 0.0003303846751805395, + 0.006108640693128109, + 0.005930320359766483, + 0.00035927645512856543, + 0.0006042442983016372, + 0.00019722805882338434, + 0.00029278264264576137, + 0.00016002113989088684, + 0.00011683793127303943, + 0.0008036814979277551, + 0.000834673410281539, + 0.0001711782970232889, + 0.0003667459823191166, + 0.0004437465104274452, + 0.00032170480699278414, + 0.0004774882399942726, + 0.00024285029212478548, + 0.0005426882416941226, + 0.0004299669526517391, + 0.0002684364444576204, + 0.00030472796061076224, + 0.0003508915542624891, + 0.0005523311556316912, + 0.0006181479548104107, + 0.00031155982287600636, + 0.0005504395812749863, + 0.00042033710633404553, + 0.00023568210599478334, + 0.00032219872809946537, + 0.0001600163523107767, + 0.0004914321471005678, + 0.00017465942073613405, + 0.0001436149759683758, + 0.00027563521871343255, + 0.0006007215706631541, + 0.00019070184498559684, + 0.00012818712275475264, + 0.0009253230527974665, + 0.0012512158136814833, + 0.0030650924891233444, + 0.0009914470138028264, + 0.0015244917012751102, + 0.0026881115045398474, + 0.0009432330261915922, + 0.0011455941712483764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.1270654797554016 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031998169142752886 + ], + "opposite_logprob_tokens": [ + -5.753200054168701 + ], + "logprob_reward": -0.0031998169142752886, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031998169142752886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0446062870323658, + 0.01567857898771763, + 0.7107048630714417, + 0.02085103467106819, + 0.12329647690057755, + 0.00812557153403759, + 0.010018564760684967, + 0.008794273249804974, + 0.014582079835236073, + 0.043342240154743195 + ], + "credit_map": [ + 0.00550572806969285, + 0.00539050716906786, + 0.004292833618819714, + 0.00468438072130084, + 0.00825479719787836, + 0.0067645348608493805, + 0.006110349204391241, + 0.003603156888857484, + 0.0016992000164464116, + 0.0015383947174996138, + 0.002113903872668743, + 0.0009531349060125649, + 0.0033267776016145945, + 0.0033755896147340536, + 0.0008379979990422726, + 0.0018335802014917135, + 0.019310228526592255, + 0.01621541567146778, + 0.004360492806881666, + 0.004230794496834278, + 0.3301461637020111, + 0.3227631747722626, + 0.0054418062791228294, + 0.008236766792833805, + 0.0020346108358353376, + 0.002361469203606248, + 0.0013143668184056878, + 0.0011471096659079194, + 0.005174911580979824, + 0.00474111782386899, + 0.0011049744207412004, + 0.0029724736232310534, + 0.0044923643581569195, + 0.004218237940222025, + 0.001675145816989243, + 0.0015191315906122327, + 0.05554996058344841, + 0.0519128181040287, + 0.0017663706094026566, + 0.0021624548826366663, + 0.0008175377151928842, + 0.0007178412051871419, + 0.0007499777129851282, + 0.00030890843481756747, + 0.002117582131177187, + 0.002365294611081481, + 0.0005699570174328983, + 0.0004784722113981843, + 0.001952284132130444, + 0.0016681242268532515, + 0.002274218015372753, + 0.0007641112897545099, + 0.0011795879108831286, + 0.0011077184462919831, + 0.0005065473960712552, + 0.0005659732851199806, + 0.001266252831555903, + 0.001391847268678248, + 0.0011407239362597466, + 0.0005869946326129138, + 0.0014924004208296537, + 0.0011280217440798879, + 0.0008343334775418043, + 0.0009536999277770519, + 0.001186475739814341, + 0.00299730128608644, + 0.0014082265552133322, + 0.0014396023470908403, + 0.002556560328230262, + 0.00317400717176497, + 0.0007183477282524109, + 0.001101558213122189, + 0.002691147616133094, + 0.009129198268055916, + 0.009884028695523739, + 0.0031747114844620228, + 0.006140824407339096, + 0.009919379837810993, + 0.00111798953730613, + 0.0012849582126364112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.0031998169142752886 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290423572063446 + ], + "opposite_logprob_tokens": [ + -2.2379043102264404 + ], + "logprob_reward": -0.11290423572063446, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290423572063446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029950806871056557, + 0.0076448735781013966, + 0.872519314289093, + 0.006270820274949074, + 0.02678212895989418, + 0.008213687688112259, + 0.010173642076551914, + 0.006036463193595409, + 0.010085276328027248, + 0.022323042154312134 + ], + "credit_map": [ + 0.003840223653241992, + 0.003578630741685629, + 0.003026938997209072, + 0.002762711141258478, + 0.006034893449395895, + 0.004726371727883816, + 0.0035170018672943115, + 0.0024640329647809267, + 0.0009490094380453229, + 0.0010981030063703656, + 0.0012214801972731948, + 0.0008282713824883103, + 0.0011052732588723302, + 0.001162399654276669, + 0.0006578924949280918, + 0.0006224446697160602, + 0.023804781958460808, + 0.01994159445166588, + 0.0052607920952141285, + 0.005127831362187862, + 0.4052950143814087, + 0.3964180648326874, + 0.006686353590339422, + 0.009984847158193588, + 0.001148848095908761, + 0.0008527532918378711, + 0.0010151041205972433, + 0.0004208807658869773, + 0.000916209421120584, + 0.001057580579072237, + 0.0002838079526554793, + 0.0005756359896622598, + 0.0011856523342430592, + 0.0010172334732487798, + 0.0006977265584282577, + 0.0003955110441893339, + 0.01109064370393753, + 0.011186208575963974, + 0.0003969718236476183, + 0.0008121823193505406, + 0.0013724713353440166, + 0.0011103200959041715, + 0.0009859474375844002, + 0.0006832409417256713, + 0.0012903740862384439, + 0.0012411142233759165, + 0.0005204631015658379, + 0.0010097568156197667, + 0.002179148141294718, + 0.0010234785731881857, + 0.0020804053638130426, + 0.0005184691399335861, + 0.001751252100802958, + 0.001329369144514203, + 0.00042040421976707876, + 0.0008711148984730244, + 0.0007062104414217174, + 0.0009022572776302695, + 0.001187603804282844, + 0.0006081411265768111, + 0.0009479053551331162, + 0.0006582186324521899, + 0.000300692452583462, + 0.0007254338124766946, + 0.0007872935966588557, + 0.0014801754150539637, + 0.0016388865187764168, + 0.0013032510178163648, + 0.0018105785129591823, + 0.0014302808558568358, + 0.0006410030764527619, + 0.000993807683698833, + 0.0011388264829292893, + 0.0029167220927774906, + 0.007666459772735834, + 0.0010551388841122389, + 0.0016448915703222156, + 0.006471979897469282, + 0.0006487026112154126, + 0.0007803209009580314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.11290423572063446 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a long, low-angle shot, the camera moving slowly and steadily around a heavy wooden table in a dimly lit office, creating a sense of quiet observation. The entire scene is rendered in a high-quality vintage style, defined by a dark color scheme and a pervasive matte texture that absorbs light. Under the spotlight of a low-hanging desk lamp, a teenage South Asian girl with long, black hair is completely absorbed in her work. She wears thin, black-rimmed glasses perched on her nose and a simple charcoal-gray cotton T-shirt. She leans forward, her posture intensely focused, yet a subtle, playful curiosity animates her expression. Her hands move with care as she assembles a small, intricate mechanical object; her fingers steadily and deliberately place a small component with a pair of tweezers, then gently press it into place. The surrounding office is shrouded in deep shadow, and the warm, isolated lighting creates a stark contrast, highlighting the matte finish of the wooden table, the soft fabric of her shirt, and the focused intensity of her creative process, creating an atmosphere of intimacy and tranquility.", + "seed": 29062, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_5/index_366_seed_29062_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02349138886202127, + "vlm_logprob": -0.02349138886202127, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.02349138886202127, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01630976051092148 + ], + "opposite_logprob_tokens": [ + -5.26630973815918 + ], + "logprob_reward": -0.01630976051092148, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.01630976051092148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1355360448360443, + 0.03562642261385918, + 0.1135864406824112, + 0.06232018768787384, + 0.054386161267757416, + 0.14062798023223877, + 0.17751814424991608, + 0.16666008532047272, + 0.04312847927212715, + 0.07061008363962173 + ], + "credit_map": [ + 0.018633974716067314, + 0.02186228334903717, + 0.010865071788430214, + 0.013532456010580063, + 0.019242683425545692, + 0.016068177297711372, + 0.014104656875133514, + 0.021226730197668076, + 0.004695785231888294, + 0.006192751228809357, + 0.0027088304050266743, + 0.0026097476948052645, + 0.004352687392383814, + 0.005180994980037212, + 0.0023864114191383123, + 0.007499213796108961, + 0.01525404118001461, + 0.0183375533670187, + 0.007028487976640463, + 0.010876530781388283, + 0.016504820436239243, + 0.020754586905241013, + 0.01072592195123434, + 0.014104499481618404, + 0.006029949523508549, + 0.006546737626194954, + 0.0028690388426184654, + 0.0036341154482215643, + 0.004741137847304344, + 0.017916522920131683, + 0.004064581822603941, + 0.01651810109615326, + 0.007961239665746689, + 0.0070191784761846066, + 0.0030818188097327948, + 0.005277086980640888, + 0.006123221945017576, + 0.01054330077022314, + 0.004585173446685076, + 0.009795138612389565, + 0.016276909038424492, + 0.021661194041371346, + 0.012705735862255096, + 0.013117944821715355, + 0.027910800650715828, + 0.01817803457379341, + 0.015524490736424923, + 0.015252865850925446, + 0.026399392634630203, + 0.017398420721292496, + 0.016842063516378403, + 0.012506096623837948, + 0.023745611310005188, + 0.032211944460868835, + 0.01531242672353983, + 0.033102188259363174, + 0.03741499409079552, + 0.022492730990052223, + 0.018797418102622032, + 0.014717712067067623, + 0.013585975393652916, + 0.023948462679982185, + 0.014829090796411037, + 0.02087370678782463, + 0.00864244531840086, + 0.004488639999181032, + 0.004351725801825523, + 0.00463871518149972, + 0.0055664279498159885, + 0.006664529908448458, + 0.003751373616978526, + 0.005024620797485113, + 0.009577135555446148, + 0.01071739848703146, + 0.00844988226890564, + 0.008205101825296879, + 0.008211719803512096, + 0.010562463663518429, + 0.006761297583580017, + 0.008125086314976215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.01630976051092148 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026361288502812386 + ], + "opposite_logprob_tokens": [ + -3.6513612270355225 + ], + "logprob_reward": -0.026361288502812386, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026361288502812386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16387370228767395, + 0.14516329765319824, + 0.12200798839330673, + 0.13019360601902008, + 0.11760740727186203, + 0.10876604914665222, + 0.04625142365694046, + 0.09145575016736984, + 0.03676143288612366, + 0.03791935369372368 + ], + "credit_map": [ + 0.02005956321954727, + 0.032013844698667526, + 0.013258659280836582, + 0.014179954305291176, + 0.020639928057789803, + 0.021564599126577377, + 0.0226645078510046, + 0.01949264295399189, + 0.018616585060954094, + 0.01972738839685917, + 0.008462943136692047, + 0.014599636197090149, + 0.017849743366241455, + 0.02800525538623333, + 0.012669987045228481, + 0.025231748819351196, + 0.015882564708590508, + 0.020050324499607086, + 0.009533396922051907, + 0.009743276052176952, + 0.01663873717188835, + 0.022782839834690094, + 0.009816101752221584, + 0.017560744658112526, + 0.008900845423340797, + 0.01241003256291151, + 0.004390060435980558, + 0.007715131621807814, + 0.010141386650502682, + 0.04250982776284218, + 0.006096064578741789, + 0.03803025931119919, + 0.01668090559542179, + 0.010026850737631321, + 0.004928853828459978, + 0.005842875689268112, + 0.007778209634125233, + 0.03371737524867058, + 0.005927827674895525, + 0.03270450979471207, + 0.008606364019215107, + 0.0139118991792202, + 0.010996436700224876, + 0.013378430157899857, + 0.02290440909564495, + 0.01492579746991396, + 0.012322358787059784, + 0.01172035001218319, + 0.006360940635204315, + 0.00781513936817646, + 0.0030411805491894484, + 0.0036057946272194386, + 0.009288576431572437, + 0.006774955429136753, + 0.003906089346855879, + 0.005458742380142212, + 0.01290079765021801, + 0.014090563170611858, + 0.007320870645344257, + 0.0063185119070112705, + 0.019379332661628723, + 0.014727627858519554, + 0.006760571151971817, + 0.00995747558772564, + 0.007938970811665058, + 0.00483887130394578, + 0.00227502197958529, + 0.0018846424063667655, + 0.006178447511047125, + 0.0059088957495987415, + 0.0026254775002598763, + 0.0051111020147800446, + 0.0036754910834133625, + 0.007309452164918184, + 0.004061892628669739, + 0.003084313590079546, + 0.0065008290112018585, + 0.005417271051555872, + 0.0037737912498414516, + 0.0040963138453662395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.026361288502812386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043059974908828735 + ], + "opposite_logprob_tokens": [ + -3.168060064315796 + ], + "logprob_reward": -0.043059974908828735, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043059974908828735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11746833473443985, + 0.07544641941785812, + 0.0684833899140358, + 0.14551898837089539, + 0.10558845847845078, + 0.11880042403936386, + 0.12125899642705917, + 0.14847294986248016, + 0.04614671319723129, + 0.052815333008766174 + ], + "credit_map": [ + 0.016460511833429337, + 0.01932188868522644, + 0.010418417863547802, + 0.011693756096065044, + 0.012334789149463177, + 0.012484381906688213, + 0.01722848229110241, + 0.017526108771562576, + 0.008413786999881268, + 0.013276451267302036, + 0.00733594223856926, + 0.010775532573461533, + 0.008778467774391174, + 0.010795500129461288, + 0.0070610553957521915, + 0.009009677916765213, + 0.007184037007391453, + 0.008032849058508873, + 0.011112382635474205, + 0.009203308261930943, + 0.00744394026696682, + 0.012008527293801308, + 0.006325585767626762, + 0.00717275682836771, + 0.016338035464286804, + 0.012367557734251022, + 0.014031329192221165, + 0.019511209800839424, + 0.012578717432916164, + 0.034070126712322235, + 0.010767481289803982, + 0.02585453726351261, + 0.015905480831861496, + 0.008055735379457474, + 0.004161444958299398, + 0.0051109520718455315, + 0.007295308168977499, + 0.03088325820863247, + 0.005212935619056225, + 0.028963344171643257, + 0.0108798174187541, + 0.016120729967951775, + 0.01134263351559639, + 0.011890961788594723, + 0.025701364502310753, + 0.01651158183813095, + 0.012992200441658497, + 0.013361128978431225, + 0.01688973605632782, + 0.011877580545842648, + 0.014339019544422626, + 0.009204051457345486, + 0.018662981688976288, + 0.019743258133530617, + 0.011529863812029362, + 0.019012507051229477, + 0.016146300360560417, + 0.015953388065099716, + 0.011316164396703243, + 0.01330217719078064, + 0.02431805618107319, + 0.025375239551067352, + 0.016363991424441338, + 0.025697635486721992, + 0.009504014626145363, + 0.007532916963100433, + 0.003087770426645875, + 0.0039952779188752174, + 0.003766515990719199, + 0.007354310248047113, + 0.003965363837778568, + 0.006940542720258236, + 0.005092357285320759, + 0.009993024170398712, + 0.007384432014077902, + 0.004055451601743698, + 0.0074949623085558414, + 0.007058329414576292, + 0.006114628631621599, + 0.005622144788503647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.043059974908828735 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019421778852120042 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.0019421778852120042, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019421778852120042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08783943206071854, + 0.02568269520998001, + 0.17648252844810486, + 0.06737317889928818, + 0.3827027976512909, + 0.07360837608575821, + 0.08224742114543915, + 0.04768236726522446, + 0.030512789264321327, + 0.025868380442261696 + ], + "credit_map": [ + 0.010358093306422234, + 0.01656845584511757, + 0.00522188562899828, + 0.007403879426419735, + 0.01599639467895031, + 0.012636338360607624, + 0.008686504326760769, + 0.010967887938022614, + 0.003171381540596485, + 0.00405883090570569, + 0.0016941609792411327, + 0.0026974908541888, + 0.004513742867857218, + 0.004259198904037476, + 0.002168327569961548, + 0.003119564149528742, + 0.022558167576789856, + 0.02962176874279976, + 0.0113044623285532, + 0.01823686808347702, + 0.027066724374890327, + 0.031581979244947433, + 0.015336071141064167, + 0.02077651023864746, + 0.007315254770219326, + 0.005305867176502943, + 0.0034034610725939274, + 0.003888143692165613, + 0.004775006324052811, + 0.02139238826930523, + 0.0038383007049560547, + 0.01745476759970188, + 0.0523739792406559, + 0.022696318104863167, + 0.013619941659271717, + 0.014663802459836006, + 0.021092351526021957, + 0.12353657931089401, + 0.015868550166487694, + 0.11885128915309906, + 0.006402857601642609, + 0.00923699326813221, + 0.007557399570941925, + 0.007988831028342247, + 0.014932296238839626, + 0.010603840462863445, + 0.008386311121284962, + 0.008499850519001484, + 0.015913430601358414, + 0.00881212018430233, + 0.007598047610372305, + 0.005621602293103933, + 0.007346548140048981, + 0.015190396457910538, + 0.006969359703361988, + 0.014795922674238682, + 0.008040555752813816, + 0.006460319273173809, + 0.00513702305033803, + 0.004868726711720228, + 0.006690625566989183, + 0.00701708672568202, + 0.003990126773715019, + 0.0054779076017439365, + 0.006938639096915722, + 0.004826658871024847, + 0.0018531129462644458, + 0.0021720766089856625, + 0.004275523591786623, + 0.004038304090499878, + 0.0027901718858629465, + 0.0036183057818561792, + 0.002417539479210973, + 0.0040274206548929214, + 0.003058751579374075, + 0.0037228153087198734, + 0.0035296280402690172, + 0.0031081277411431074, + 0.002863447181880474, + 0.0031406518537551165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.0019421778852120042 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit office with a heavy wooden table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a teenage South Asian girl with long black hair wearing black-rimmed glasses and a charcoal-gray cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the girl using tweezers to assemble a small, intricate mechanical object?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the visual style defined by a dark color scheme, matte texture, and warm isolated lighting from a low-hanging desk lamp?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera move slowly and steadily around the table in a long, low-angle shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029783742502331734 + ], + "opposite_logprob_tokens": [ + -3.5297837257385254 + ], + "logprob_reward": -0.029783742502331734, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029783742502331734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06847786903381348, + 0.03748972341418266, + 0.08941856771707535, + 0.0560469925403595, + 0.21598689258098602, + 0.22256989777088165, + 0.08492069691419601, + 0.15612885355949402, + 0.03370710834860802, + 0.03525344654917717 + ], + "credit_map": [ + 0.010683353058993816, + 0.015477698296308517, + 0.00484837219119072, + 0.005274449475109577, + 0.010189760476350784, + 0.006519502028822899, + 0.007069580722600222, + 0.00841514952480793, + 0.00505090830847621, + 0.007241753861308098, + 0.001920703798532486, + 0.0035276985727250576, + 0.006105391774326563, + 0.005949113052338362, + 0.0029536690562963486, + 0.004740484524518251, + 0.011307980865240097, + 0.012167423963546753, + 0.0048931255005300045, + 0.008020242676138878, + 0.01364106498658657, + 0.018778201192617416, + 0.007758969906717539, + 0.012851548381149769, + 0.005660457536578178, + 0.003859652439132333, + 0.0022820273879915476, + 0.0028395322151482105, + 0.006570328958332539, + 0.017905205488204956, + 0.002661806298419833, + 0.01426797453314066, + 0.04640563204884529, + 0.01501975767314434, + 0.006774288602173328, + 0.008797327056527138, + 0.012551471590995789, + 0.06184448301792145, + 0.00609892513602972, + 0.058494988828897476, + 0.01834719441831112, + 0.0306368637830019, + 0.02064898982644081, + 0.026161054149270058, + 0.04564699903130531, + 0.03257886692881584, + 0.025669924914836884, + 0.02287997677922249, + 0.017022842541337013, + 0.007767370902001858, + 0.008523157797753811, + 0.006319798994809389, + 0.008779790252447128, + 0.014973217621445656, + 0.007518471218645573, + 0.014016037806868553, + 0.021547328680753708, + 0.02354464866220951, + 0.013795272447168827, + 0.010827172547578812, + 0.029392721131443977, + 0.0248622614890337, + 0.013990635983645916, + 0.01816880702972412, + 0.005442339926958084, + 0.0045760576613247395, + 0.0028367252089083195, + 0.002949932822957635, + 0.006508365273475647, + 0.005000061821192503, + 0.0030541387386620045, + 0.003339482471346855, + 0.004914891440421343, + 0.006740274839103222, + 0.004052725154906511, + 0.0025495330337435007, + 0.005201856140047312, + 0.00511120306327939, + 0.003385411575436592, + 0.003297547111287713 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 7 + ], + "selected_reward": -0.029783742502331734 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_6/index_tensor([8206, 8206])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_6/index_tensor([8206, 8206])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bb1446acdd1e2b9d357b661b379d6cdd62ca5ed3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_6/index_tensor([8206, 8206])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Rendered in a warm and captivating style inspired by Studio Ghibli, this close-up shot tilts smoothly upward, tracking a whimsical hand-drawn menu as it gracefully rises through the air, propelled by an unseen magical force. The menu itself is made of thick, cream-colored parchment paper, with slightly crinkled and soft edges. The camera maintains sharp focus on the menu's intricate artwork, showcasing a vibrant array of cakes and sweets rendered in delicate, vibrant watercolors with delicate ink outlines: a puffy strawberry shortcake with sparkling red berries, a miniature fruit tart with a lustrous glaze, and a plump, cream-filled chocolate pastry. As the menu floats calmly upward with a subtle, almost imperceptible sway, the sunlit bakery interior behind it is cast in a soft, dreamy blur. Through this beautiful bokeh, one can see the warm wood textures of the shelves filled with round bread and the soft glow of the glass display cases. Golden sunlight streams in through a large window, illuminating the dust particles dancing in the air and bathing the entire scene in a warm, honeyed glow, creating an atmosphere of purity, tenderness, charm, and comfort.", + "seed": 19900, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_6/index_8206_seed_19900_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0592738788574934, + "vlm_logprob": -0.0592738788574934, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.0592738788574934, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054061487317085266 + ], + "opposite_logprob_tokens": [ + -3.1790614128112793 + ], + "logprob_reward": -0.054061487317085266, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.054061487317085266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0690612643957138, + 0.14506827294826508, + 0.029208101332187653, + 0.041638851165771484, + 0.023273231461644173, + 0.30338385701179504, + 0.0896177887916565, + 0.10551100969314575, + 0.16264791786670685, + 0.03058968111872673 + ], + "credit_map": [ + 0.012012745253741741, + 0.007822362706065178, + 0.008621281012892723, + 0.006802109070122242, + 0.008643685840070248, + 0.007600327953696251, + 0.009652834385633469, + 0.007905918173491955, + 0.004826484713703394, + 0.003127086441963911, + 0.0053801569156348705, + 0.004502632189542055, + 0.057176023721694946, + 0.02219502255320549, + 0.04145481437444687, + 0.006406053435057402, + 0.002328481525182724, + 0.0026783980429172516, + 0.002470464212819934, + 0.0022777405101805925, + 0.003972491715103388, + 0.003733268938958645, + 0.004311420954763889, + 0.007435835897922516, + 0.002468753606081009, + 0.003004970494657755, + 0.0036797018256038427, + 0.0033064489252865314, + 0.00444717425853014, + 0.006600538734346628, + 0.013748797588050365, + 0.004382462240755558, + 0.002359190257266164, + 0.002954703290015459, + 0.003489798866212368, + 0.002569863572716713, + 0.0027581932954490185, + 0.0024169895332306623, + 0.0036181325558573008, + 0.0031063593924045563, + 0.01957997865974903, + 0.008729337714612484, + 0.01586165837943554, + 0.00653349794447422, + 0.0169293861836195, + 0.061323750764131546, + 0.1374431848526001, + 0.03698305785655975, + 0.009558507241308689, + 0.006122879683971405, + 0.005365659948438406, + 0.004750996828079224, + 0.004497061017900705, + 0.019706469029188156, + 0.010005354881286621, + 0.029610855504870415, + 0.011618256568908691, + 0.00971139594912529, + 0.008061634376645088, + 0.01033297274261713, + 0.008665169589221478, + 0.023897934705018997, + 0.007239084225147963, + 0.025984562933444977, + 0.023812206462025642, + 0.009483643807470798, + 0.008002502843737602, + 0.006380376871675253, + 0.006677121855318546, + 0.06669370830059052, + 0.005191760137677193, + 0.03640659153461456, + 0.004908300936222076, + 0.005055270157754421, + 0.003479134291410446, + 0.0029327874071896076, + 0.0024729070719331503, + 0.003966436721384525, + 0.0027591318357735872, + 0.005015710834413767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.054061487317085266 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004337757825851 + ], + "opposite_logprob_tokens": [ + -2.6950433254241943 + ], + "logprob_reward": -0.07004337757825851, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004337757825851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04629512131214142, + 0.08932340890169144, + 0.043954379856586456, + 0.04606233537197113, + 0.014700453728437424, + 0.24417836964130402, + 0.22809095680713654, + 0.041802778840065, + 0.181032195687294, + 0.0645599439740181 + ], + "credit_map": [ + 0.0066696335561573505, + 0.006837998051196337, + 0.004914058838039637, + 0.005866168532520533, + 0.005826878361403942, + 0.004880722612142563, + 0.005711246747523546, + 0.005588422995060682, + 0.0033904474694281816, + 0.0032465150579810143, + 0.0048116957768797874, + 0.006150030065327883, + 0.030363602563738823, + 0.011812747456133366, + 0.024488398805260658, + 0.005059977062046528, + 0.004034479148685932, + 0.0021376130171120167, + 0.0030706115067005157, + 0.0024027947802096605, + 0.008256988599896431, + 0.006820307578891516, + 0.006717761047184467, + 0.010513829998672009, + 0.0021490019280463457, + 0.002504967851564288, + 0.0032939871307462454, + 0.0033933587837964296, + 0.004970788024365902, + 0.007745014503598213, + 0.01714375801384449, + 0.004861463792622089, + 0.0014371526194736362, + 0.001514763804152608, + 0.0015076036797836423, + 0.0010996165219694376, + 0.0015692835440859199, + 0.001678219996392727, + 0.003573508933186531, + 0.0023203068412840366, + 0.01554795354604721, + 0.008620942942798138, + 0.012883467599749565, + 0.008411800488829613, + 0.014003688469529152, + 0.04894477128982544, + 0.10623973608016968, + 0.029526017606258392, + 0.011398142203688622, + 0.008720161393284798, + 0.008091107942163944, + 0.007489204406738281, + 0.008906343951821327, + 0.07406766712665558, + 0.01682502217590809, + 0.09259330481290817, + 0.0055200872011482716, + 0.008038494735956192, + 0.003676473395898938, + 0.004954865202307701, + 0.004617191385477781, + 0.005861555226147175, + 0.003555361879989505, + 0.0055787526071071625, + 0.040330495685338974, + 0.01706383191049099, + 0.013086866587400436, + 0.013080896809697151, + 0.00833116378635168, + 0.04972531646490097, + 0.01089843176305294, + 0.02851521037518978, + 0.014085572212934494, + 0.0061040716245770454, + 0.009324360638856888, + 0.0023823934607207775, + 0.0033784504048526287, + 0.008056407794356346, + 0.005736369173973799, + 0.015492326579988003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.07004337757825851 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048674196004867554 + ], + "opposite_logprob_tokens": [ + -3.0486741065979004 + ], + "logprob_reward": -0.048674196004867554, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048674196004867554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08907853066921234, + 0.053053371608257294, + 0.016435734927654266, + 0.025692246854305267, + 0.016381975263357162, + 0.3910895586013794, + 0.1795208752155304, + 0.05384853482246399, + 0.08530241250991821, + 0.08959675580263138 + ], + "credit_map": [ + 0.015878761187195778, + 0.009591837413609028, + 0.010733462870121002, + 0.008892130106687546, + 0.011098925024271011, + 0.007836873643100262, + 0.015126802958548069, + 0.009919733740389347, + 0.003898079739883542, + 0.002891180571168661, + 0.004178910981863737, + 0.004344291519373655, + 0.016178805381059647, + 0.00622806791216135, + 0.012280871160328388, + 0.0030531648080796003, + 0.0012152749113738537, + 0.0013561726082116365, + 0.001691385987214744, + 0.001339634065516293, + 0.0023264936171472073, + 0.0021765420679003, + 0.0024656762834638357, + 0.0038645556196570396, + 0.0016349826473742723, + 0.002106924308463931, + 0.0019538006745278835, + 0.002167398575693369, + 0.002760503441095352, + 0.005191153846681118, + 0.007365907076746225, + 0.0025115753524005413, + 0.0014657540014013648, + 0.002664938336238265, + 0.002259286819025874, + 0.002483747201040387, + 0.001868931227363646, + 0.00148831179831177, + 0.0021417669486254454, + 0.0020092392805963755, + 0.024341817945241928, + 0.01111673191189766, + 0.01842804253101349, + 0.008857287466526031, + 0.02144394814968109, + 0.08246849477291107, + 0.17806924879550934, + 0.04636398330330849, + 0.03457782417535782, + 0.014059836976230145, + 0.01196309830993414, + 0.011679156683385372, + 0.00995586346834898, + 0.023459551855921745, + 0.01891781948506832, + 0.054907720535993576, + 0.006485672667622566, + 0.005310093052685261, + 0.005235945340245962, + 0.005576516035944223, + 0.004785788245499134, + 0.010907948017120361, + 0.005195881240069866, + 0.010350695811212063, + 0.01567542739212513, + 0.005427937489002943, + 0.004427216947078705, + 0.002169085666537285, + 0.003470917697995901, + 0.03431348130106926, + 0.003163235727697611, + 0.01665511354804039, + 0.015038891695439816, + 0.013906360603868961, + 0.014937958680093288, + 0.006411431357264519, + 0.007671489380300045, + 0.010885505937039852, + 0.007196882274001837, + 0.013548239134252071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.048674196004867554 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329273335635662 + ], + "opposite_logprob_tokens": [ + -3.7732927799224854 + ], + "logprob_reward": -0.02329273335635662, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329273335635662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09277571737766266, + 0.18428762257099152, + 0.09072120487689972, + 0.031283341348171234, + 0.035303615033626556, + 0.05485554784536362, + 0.0799281969666481, + 0.15948234498500824, + 0.234872505068779, + 0.03648986294865608 + ], + "credit_map": [ + 0.006498711183667183, + 0.004884737078100443, + 0.004787565674632788, + 0.006578688509762287, + 0.014365576207637787, + 0.023678677156567574, + 0.017121655866503716, + 0.014860108494758606, + 0.006128143519163132, + 0.003969639074057341, + 0.006139971315860748, + 0.005306502804160118, + 0.07257339358329773, + 0.029533352702856064, + 0.051591090857982635, + 0.009045516140758991, + 0.005924318451434374, + 0.005526358727365732, + 0.00386382220312953, + 0.004073685500770807, + 0.01554866787046194, + 0.014564483426511288, + 0.015181981958448887, + 0.02603788860142231, + 0.0022564763203263283, + 0.001387040945701301, + 0.0019389057997614145, + 0.00223109545186162, + 0.0029598139226436615, + 0.005916614551097155, + 0.008627133443951607, + 0.005966261029243469, + 0.003270380198955536, + 0.0023307267110794783, + 0.0030572989489883184, + 0.0020325374789536, + 0.003922935575246811, + 0.004467670805752277, + 0.010598192922770977, + 0.005623874720185995, + 0.003186512039974332, + 0.002546005416661501, + 0.0031884126365184784, + 0.002042100764811039, + 0.003365797223523259, + 0.011909962631762028, + 0.02273101918399334, + 0.005885737482458353, + 0.006156157236546278, + 0.005463925190269947, + 0.0037552572321146727, + 0.005585141014307737, + 0.00423408392816782, + 0.019213004037737846, + 0.009177177213132381, + 0.02634344808757305, + 0.015563287772238255, + 0.015343858860433102, + 0.015813274309039116, + 0.015576926060020924, + 0.014671268872916698, + 0.03657807782292366, + 0.010964134708046913, + 0.03497150540351868, + 0.03501734510064125, + 0.011578488163650036, + 0.011142613366246223, + 0.004556918982416391, + 0.008993173949420452, + 0.10038110613822937, + 0.007311634719371796, + 0.05589122325181961, + 0.005939094349741936, + 0.005919214338064194, + 0.005670452024787664, + 0.0028049363754689693, + 0.0033678761683404446, + 0.004657394252717495, + 0.002910719718784094, + 0.005220179446041584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.02329273335635662 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029760003089905 + ], + "opposite_logprob_tokens": [ + -2.350297689437866 + ], + "logprob_reward": -0.10029760003089905, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029760003089905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0629553571343422, + 0.22617308795452118, + 0.055626872926950455, + 0.02508729137480259, + 0.016613438725471497, + 0.4146661162376404, + 0.055867210030555725, + 0.0666636973619461, + 0.03297180309891701, + 0.04337513819336891 + ], + "credit_map": [ + 0.004760848358273506, + 0.004581040237098932, + 0.004332229495048523, + 0.005628758575767279, + 0.010162072256207466, + 0.012122627347707748, + 0.011167258955538273, + 0.01020052284002304, + 0.0059760273434221745, + 0.004193716682493687, + 0.007864957675337791, + 0.005994624923914671, + 0.09153986722230911, + 0.037934549152851105, + 0.06121731549501419, + 0.011452031321823597, + 0.003502361476421356, + 0.005484568886458874, + 0.002846105257049203, + 0.002549223368987441, + 0.009796354919672012, + 0.008416935801506042, + 0.00823398306965828, + 0.014797341078519821, + 0.0022537419572472572, + 0.0012896012049168348, + 0.0017922475235536695, + 0.0015265541151165962, + 0.005823499523103237, + 0.0028334450908005238, + 0.00625260965898633, + 0.0033155903220176697, + 0.0018668848788365722, + 0.0016792342066764832, + 0.001914355205371976, + 0.0015036106342449784, + 0.0020262382458895445, + 0.0017245238414034247, + 0.0031646715942770243, + 0.0027339213993400335, + 0.026136232540011406, + 0.012513506226241589, + 0.02151723951101303, + 0.009631011635065079, + 0.022670086473226547, + 0.08538433909416199, + 0.18751120567321777, + 0.04930248484015465, + 0.014937271364033222, + 0.0037190858274698257, + 0.0045897443778812885, + 0.003307152772322297, + 0.003488454269245267, + 0.008806961588561535, + 0.007547219283878803, + 0.009471320547163486, + 0.007190614007413387, + 0.006430078763514757, + 0.005572965834289789, + 0.00681596202775836, + 0.006498406175523996, + 0.014648860320448875, + 0.006462350487709045, + 0.01304446067661047, + 0.006541497074067593, + 0.0037046344950795174, + 0.0023512740153819323, + 0.00399386789649725, + 0.002540676621720195, + 0.003876927075907588, + 0.0024923463352024555, + 0.007470578886568546, + 0.007805097382515669, + 0.005324509460479021, + 0.007393607869744301, + 0.0031040399335324764, + 0.0030601713806390762, + 0.005054814275354147, + 0.0037309869658201933, + 0.007901913486421108 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.10029760003089905 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in a warm and captivating style inspired by Studio Ghibli, this close-up shot tilts smoothly upward, tracking a whimsical hand-drawn menu as it gracefully rises through the air, propelled by an unseen magical force. The menu itself is made of thick, cream-colored parchment paper, with slightly crinkled and soft edges. The camera maintains sharp focus on the menu's intricate artwork, showcasing a vibrant array of cakes and sweets rendered in delicate, vibrant watercolors with delicate ink outlines: a puffy strawberry shortcake with sparkling red berries, a miniature fruit tart with a lustrous glaze, and a plump, cream-filled chocolate pastry. As the menu floats calmly upward with a subtle, almost imperceptible sway, the sunlit bakery interior behind it is cast in a soft, dreamy blur. Through this beautiful bokeh, one can see the warm wood textures of the shelves filled with round bread and the soft glow of the glass display cases. Golden sunlight streams in through a large window, illuminating the dust particles dancing in the air and bathing the entire scene in a warm, honeyed glow, creating an atmosphere of purity, tenderness, charm, and comfort.", + "seed": 19901, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_6/index_8206_seed_19901_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13239974472671748, + "vlm_logprob": -0.13239974472671748, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.13239974472671748, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03654922917485237 + ], + "opposite_logprob_tokens": [ + -3.5365493297576904 + ], + "logprob_reward": -0.03654922917485237, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03654922917485237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06511982530355453, + 0.035461802035570145, + 0.12535592913627625, + 0.04410737380385399, + 0.25407296419143677, + 0.23258917033672333, + 0.06723877787590027, + 0.048614196479320526, + 0.04114731401205063, + 0.08629265427589417 + ], + "credit_map": [ + 0.007231065537780523, + 0.011157076805830002, + 0.008104612119495869, + 0.004512231331318617, + 0.00914747640490532, + 0.009117447771131992, + 0.005313929636031389, + 0.010535991750657558, + 0.003903807606548071, + 0.004570807795971632, + 0.002745269099250436, + 0.0032566413283348083, + 0.0052842795848846436, + 0.006406216882169247, + 0.0026429370045661926, + 0.006651848554611206, + 0.022804060950875282, + 0.016141390427947044, + 0.012636610306799412, + 0.01302263606339693, + 0.019216371700167656, + 0.012254173867404461, + 0.015977568924427032, + 0.013303129002451897, + 0.006076867692172527, + 0.004049092531204224, + 0.006471433211117983, + 0.00453731045126915, + 0.005977557972073555, + 0.006843345705419779, + 0.00470410380512476, + 0.005447668489068747, + 0.06493637710809708, + 0.01682138442993164, + 0.02863648161292076, + 0.015549561008810997, + 0.022169142961502075, + 0.05572865158319473, + 0.01629950851202011, + 0.03393187001347542, + 0.015639575198292732, + 0.013537365943193436, + 0.01205440703779459, + 0.01034500915557146, + 0.013566119596362114, + 0.08014599233865738, + 0.01186763122677803, + 0.07543307542800903, + 0.008085397072136402, + 0.008359361439943314, + 0.010052124969661236, + 0.00844730343669653, + 0.007632720284163952, + 0.00821998342871666, + 0.006919375620782375, + 0.00952251348644495, + 0.005157324485480785, + 0.009446576237678528, + 0.0041960314847528934, + 0.00779747823253274, + 0.005122446920722723, + 0.005185295362025499, + 0.005734873004257679, + 0.00597417401149869, + 0.005323535762727261, + 0.0054848031140863895, + 0.006339940708130598, + 0.0036280015483498573, + 0.003862703451886773, + 0.004474993795156479, + 0.005337238311767578, + 0.006696098484098911, + 0.010326376184821129, + 0.01208532601594925, + 0.009040103293955326, + 0.010106037370860577, + 0.009069113992154598, + 0.011068717576563358, + 0.009807866998016834, + 0.014789114706218243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.03654922917485237 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07005004584789276 + ], + "opposite_logprob_tokens": [ + -2.695050001144409 + ], + "logprob_reward": -0.07005004584789276, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07005004584789276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08257982134819031, + 0.07208435237407684, + 0.06445936113595963, + 0.04315419867634773, + 0.34103941917419434, + 0.14756721258163452, + 0.09220848232507706, + 0.026278095319867134, + 0.08366240561008453, + 0.04696669057011604 + ], + "credit_map": [ + 0.010551097802817822, + 0.012183560989797115, + 0.008679509162902832, + 0.008483746089041233, + 0.010296243242919445, + 0.013819345273077488, + 0.00633246311917901, + 0.012233851477503777, + 0.00876164622604847, + 0.010305373929440975, + 0.005932483356446028, + 0.006314413156360388, + 0.009338367730379105, + 0.009199385531246662, + 0.0065748197957873344, + 0.015657858923077583, + 0.0060168965719640255, + 0.005296669434756041, + 0.011438606306910515, + 0.006790889427065849, + 0.006195194087922573, + 0.007801437750458717, + 0.008099811151623726, + 0.012819861993193626, + 0.0049072736874222755, + 0.004624298773705959, + 0.005227978341281414, + 0.006400845479220152, + 0.006748597137629986, + 0.00528535945340991, + 0.005429015029221773, + 0.004530833102762699, + 0.08750145882368088, + 0.025958625599741936, + 0.03602943196892738, + 0.022794662043452263, + 0.030003603547811508, + 0.08050395548343658, + 0.0209652166813612, + 0.037282466888427734, + 0.011395311914384365, + 0.006403373088687658, + 0.010241379030048847, + 0.005525105632841587, + 0.006907518487423658, + 0.0514768548309803, + 0.008294148370623589, + 0.047323524951934814, + 0.01179124228656292, + 0.009361855685710907, + 0.012288287281990051, + 0.010540108196437359, + 0.009385713376104832, + 0.01386287808418274, + 0.009928153827786446, + 0.015050248242914677, + 0.003484196262434125, + 0.004808510188013315, + 0.002821999602019787, + 0.0033915170934051275, + 0.002530183643102646, + 0.0028795383404940367, + 0.002221723087131977, + 0.004140425007790327, + 0.009163481183350086, + 0.013676472939550877, + 0.011237222701311111, + 0.008747364394366741, + 0.006589970085769892, + 0.009689677506685257, + 0.011173630133271217, + 0.013384592719376087, + 0.006702784448862076, + 0.006755253300070763, + 0.005545102525502443, + 0.005004026927053928, + 0.005314210895448923, + 0.005128822289407253, + 0.006204470060765743, + 0.006312016397714615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.07005004584789276 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700526013970375 + ], + "opposite_logprob_tokens": [ + -2.6950526237487793 + ], + "logprob_reward": -0.0700526013970375, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700526013970375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14394362270832062, + 0.04167087376117706, + 0.2378809005022049, + 0.032668158411979675, + 0.08034396171569824, + 0.21684956550598145, + 0.05911492928862572, + 0.04848215728998184, + 0.04502571374177933, + 0.09402010589838028 + ], + "credit_map": [ + 0.016494089737534523, + 0.011403009295463562, + 0.013889117166399956, + 0.013274392113089561, + 0.010135827586054802, + 0.03581623733043671, + 0.010974728502333164, + 0.031956225633621216, + 0.007019905839115381, + 0.0035111408215016127, + 0.003184705041348934, + 0.0034411048982292414, + 0.002763387281447649, + 0.005013672169297934, + 0.0031685258727520704, + 0.013568433001637459, + 0.042609572410583496, + 0.029533376917243004, + 0.025010092183947563, + 0.025926103815436363, + 0.03819357603788376, + 0.02403506077826023, + 0.02847365476191044, + 0.02409948967397213, + 0.0036314381286501884, + 0.0038757806178182364, + 0.005784638691693544, + 0.0031984359957277775, + 0.005530522204935551, + 0.003945834003388882, + 0.003416231367737055, + 0.003285278333351016, + 0.016537241637706757, + 0.007016052026301622, + 0.00988942850381136, + 0.005327528342604637, + 0.0084364153444767, + 0.019523832947015762, + 0.004219941329210997, + 0.009393525309860706, + 0.05078360065817833, + 0.013108623214066029, + 0.013243983499705791, + 0.009246675297617912, + 0.012478828430175781, + 0.07346268743276596, + 0.015729546546936035, + 0.028795622289180756, + 0.011869366280734539, + 0.004972732160240412, + 0.005285050719976425, + 0.004628192167729139, + 0.003986644558608532, + 0.007766684517264366, + 0.004388776607811451, + 0.016217483207583427, + 0.005177871789783239, + 0.008631359785795212, + 0.0040421262383461, + 0.00806568842381239, + 0.005857065320014954, + 0.004897605162113905, + 0.005957663059234619, + 0.005852782167494297, + 0.005621201824396849, + 0.008217989467084408, + 0.006562991067767143, + 0.006408549379557371, + 0.004696451127529144, + 0.00366410193964839, + 0.0037461102474480867, + 0.0061083207838237286, + 0.011555631645023823, + 0.011833281256258488, + 0.015266010537743568, + 0.009414811618626118, + 0.010481405071914196, + 0.010754534974694252, + 0.010654964484274387, + 0.01405948493629694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.0700526013970375 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077099479734898 + ], + "opposite_logprob_tokens": [ + -4.511076927185059 + ], + "logprob_reward": -0.011077099479734898, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077099479734898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16739721596240997, + 0.0413014218211174, + 0.1667247712612152, + 0.04381047561764717, + 0.1585562825202942, + 0.17745503783226013, + 0.04631033539772034, + 0.05261708050966263, + 0.058009788393974304, + 0.08781754970550537 + ], + "credit_map": [ + 0.028584858402609825, + 0.01576734147965908, + 0.019537243992090225, + 0.01628076657652855, + 0.013421268202364445, + 0.034246914088726044, + 0.010940025560557842, + 0.028618818148970604, + 0.006177044939249754, + 0.0046437275595963, + 0.004093336872756481, + 0.003765065222978592, + 0.0034974950831383467, + 0.005798658821731806, + 0.003147866576910019, + 0.010178228840231895, + 0.02215915359556675, + 0.018645934760570526, + 0.02029472589492798, + 0.01944490149617195, + 0.0200981292873621, + 0.01595616713166237, + 0.02697109431028366, + 0.02315467596054077, + 0.005983813665807247, + 0.004530654288828373, + 0.007208764087408781, + 0.004031663294881582, + 0.005864265374839306, + 0.006231390871107578, + 0.00468982569873333, + 0.005270102992653847, + 0.03568685054779053, + 0.011795338243246078, + 0.0244736485183239, + 0.010501766577363014, + 0.013106527738273144, + 0.02555587701499462, + 0.012100825086236, + 0.025335462763905525, + 0.04559784755110741, + 0.01055111363530159, + 0.012428510002791882, + 0.005731805693358183, + 0.008401195518672466, + 0.05863591283559799, + 0.013580485247075558, + 0.02252817153930664, + 0.005607428029179573, + 0.004154000896960497, + 0.009544682689011097, + 0.003782987594604492, + 0.0035777208395302296, + 0.007417318876832724, + 0.005404955241829157, + 0.00682124262675643, + 0.0048792604357004166, + 0.00666538393124938, + 0.005824523977935314, + 0.005801644641906023, + 0.005239186808466911, + 0.008754066191613674, + 0.0076584648340940475, + 0.007794554345309734, + 0.006714653689414263, + 0.008514565415680408, + 0.009335167706012726, + 0.007734550163149834, + 0.005738220177590847, + 0.0049643502570688725, + 0.007986257784068584, + 0.0070220245979726315, + 0.013240469619631767, + 0.007487130351364613, + 0.027059029787778854, + 0.006592846475541592, + 0.0054625594057142735, + 0.008422485552728176, + 0.0068275947123765945, + 0.012725443579256535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.011077099479734898 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742697477340698 + ], + "opposite_logprob_tokens": [ + -0.9742697477340698 + ], + "logprob_reward": -0.4742697477340698, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742697477340698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06857549399137497, + 0.03571592643857002, + 0.06625320762395859, + 0.031415458768606186, + 0.1579713225364685, + 0.29532554745674133, + 0.06484110653400421, + 0.05038604885339737, + 0.11775992065668106, + 0.11175587773323059 + ], + "credit_map": [ + 0.009608020074665546, + 0.0067384205758571625, + 0.006067866925150156, + 0.007488956209272146, + 0.008403613232076168, + 0.012181708589196205, + 0.00549575500190258, + 0.012591155245900154, + 0.005251499358564615, + 0.005258527118712664, + 0.0031399806030094624, + 0.0027913041412830353, + 0.004665642976760864, + 0.006109403911978006, + 0.0029775691218674183, + 0.005522000137716532, + 0.011578897945582867, + 0.008670707233250141, + 0.006348790600895882, + 0.00770762050524354, + 0.009915286675095558, + 0.00605527451261878, + 0.008259376510977745, + 0.00771725969389081, + 0.004149504471570253, + 0.00442501762881875, + 0.0029869028367102146, + 0.0038958541117608547, + 0.005865172483026981, + 0.0036350958980619907, + 0.0031781713478267193, + 0.003279742319136858, + 0.03571441397070885, + 0.012834031134843826, + 0.021716531366109848, + 0.01250730361789465, + 0.015989311039447784, + 0.025790881365537643, + 0.011243244633078575, + 0.0221756212413311, + 0.06680414080619812, + 0.01526773814111948, + 0.016757529228925705, + 0.008915744721889496, + 0.013182644732296467, + 0.10808037221431732, + 0.01989542320370674, + 0.04642197862267494, + 0.011270738206803799, + 0.004766139667481184, + 0.007845627143979073, + 0.00570726627483964, + 0.005073356907814741, + 0.008595927618443966, + 0.0071319229900836945, + 0.014450132846832275, + 0.006017395295202732, + 0.0069571430794894695, + 0.006688221823424101, + 0.008051034063100815, + 0.004117338918149471, + 0.006267223507165909, + 0.0050302292220294476, + 0.0072574676014482975, + 0.01465661358088255, + 0.022702636197209358, + 0.019775724038481712, + 0.012742714956402779, + 0.011952164582908154, + 0.008595959283411503, + 0.013919252902269363, + 0.013414868153631687, + 0.015670541673898697, + 0.019564460963010788, + 0.018353961408138275, + 0.010455125942826271, + 0.012033500708639622, + 0.009653511457145214, + 0.01112519670277834, + 0.014899586327373981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.4742697477340698 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in a warm and captivating style inspired by Studio Ghibli, this close-up shot tilts smoothly upward, tracking a whimsical hand-drawn menu as it gracefully rises through the air, propelled by an unseen magical force. The menu itself is made of thick, cream-colored parchment paper, with slightly crinkled and soft edges. The camera maintains sharp focus on the menu's intricate artwork, showcasing a vibrant array of cakes and sweets rendered in delicate, vibrant watercolors with delicate ink outlines: a puffy strawberry shortcake with sparkling red berries, a miniature fruit tart with a lustrous glaze, and a plump, cream-filled chocolate pastry. As the menu floats calmly upward with a subtle, almost imperceptible sway, the sunlit bakery interior behind it is cast in a soft, dreamy blur. Through this beautiful bokeh, one can see the warm wood textures of the shelves filled with round bread and the soft glow of the glass display cases. Golden sunlight streams in through a large window, illuminating the dust particles dancing in the air and bathing the entire scene in a warm, honeyed glow, creating an atmosphere of purity, tenderness, charm, and comfort.", + "seed": 19902, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_6/index_8206_seed_19902_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.045027464255690576, + "vlm_logprob": -0.045027464255690576, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.045027464255690576, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05457329377532005 + ], + "opposite_logprob_tokens": [ + -3.0545732975006104 + ], + "logprob_reward": -0.05457329377532005, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05457329377532005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1686875969171524, + 0.294929176568985, + 0.08253354579210281, + 0.08486416935920715, + 0.042679086327552795, + 0.06713958084583282, + 0.11694619804620743, + 0.09692808240652084, + 0.015578893013298512, + 0.029713694006204605 + ], + "credit_map": [ + 0.010458866134285927, + 0.020320488139986992, + 0.01482362113893032, + 0.011510137468576431, + 0.017859263345599174, + 0.026044005528092384, + 0.022616315633058548, + 0.04505489394068718, + 0.03183263912796974, + 0.04550637677311897, + 0.03887025639414787, + 0.026042310521006584, + 0.045614778995513916, + 0.03465849533677101, + 0.03161250427365303, + 0.04079180955886841, + 0.009034907445311546, + 0.009039836935698986, + 0.009024256840348244, + 0.011460836045444012, + 0.014604100957512856, + 0.008317367173731327, + 0.009824085049331188, + 0.011228152550756931, + 0.006576908752322197, + 0.007492731790989637, + 0.009134219959378242, + 0.012710633687675, + 0.011567842215299606, + 0.013394210487604141, + 0.007910940796136856, + 0.01607668586075306, + 0.003964561503380537, + 0.011100693605840206, + 0.003037312999367714, + 0.006142713129520416, + 0.004726255778223276, + 0.004562100861221552, + 0.0040632858872413635, + 0.005082158371806145, + 0.008117679506540298, + 0.005414370913058519, + 0.005319682415574789, + 0.006260182708501816, + 0.008524644188582897, + 0.017681462690234184, + 0.006524822674691677, + 0.009296737611293793, + 0.014554748311638832, + 0.015439908020198345, + 0.014844701625406742, + 0.011963682249188423, + 0.017067233100533485, + 0.013609238900244236, + 0.009154648520052433, + 0.020312033593654633, + 0.011328241787850857, + 0.012865506112575531, + 0.008330758661031723, + 0.010152817703783512, + 0.0162857286632061, + 0.01505891140550375, + 0.009229772724211216, + 0.013676348142325878, + 0.0014408916467800736, + 0.0016504938248544931, + 0.001559075084514916, + 0.0017881925450637937, + 0.001344982534646988, + 0.0017595404060557485, + 0.002857730956748128, + 0.003177986480295658, + 0.0032913796603679657, + 0.004218850284814835, + 0.003306447993963957, + 0.0036943324375897646, + 0.005552199669182301, + 0.0030231636483222246, + 0.003262964077293873, + 0.0033643555361777544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.05457329377532005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003026455640793 + ], + "opposite_logprob_tokens": [ + -2.6950302124023438 + ], + "logprob_reward": -0.07003026455640793, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003026455640793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1922028809785843, + 0.1557576060295105, + 0.05383427068591118, + 0.12082323431968689, + 0.057977817952632904, + 0.13048118352890015, + 0.10715319961309433, + 0.04708951711654663, + 0.0954652950167656, + 0.03921499475836754 + ], + "credit_map": [ + 0.013185501098632812, + 0.019904864951968193, + 0.021923327818512917, + 0.017268525436520576, + 0.019752467051148415, + 0.02897552214562893, + 0.021938221529126167, + 0.04925443232059479, + 0.0164367463439703, + 0.02165200002491474, + 0.0207053292542696, + 0.013318481855094433, + 0.02600286528468132, + 0.01774379424750805, + 0.019718607887625694, + 0.02017976902425289, + 0.0065346648916602135, + 0.005629478953778744, + 0.004778795875608921, + 0.006823628209531307, + 0.008127081207931042, + 0.005749490112066269, + 0.007208306808024645, + 0.008982818573713303, + 0.010645073838531971, + 0.012347668409347534, + 0.017209524288773537, + 0.01728428155183792, + 0.017542028799653053, + 0.017891958355903625, + 0.011727789416909218, + 0.016174906864762306, + 0.00852543581277132, + 0.008204733021557331, + 0.00445352541282773, + 0.010348640382289886, + 0.0058785974979400635, + 0.006362054962664843, + 0.006392877083271742, + 0.007811950519680977, + 0.014866167679429054, + 0.012039914727210999, + 0.008136478252708912, + 0.011601623147726059, + 0.022739680483937263, + 0.03144904598593712, + 0.014720986597239971, + 0.014927282929420471, + 0.014150449074804783, + 0.009964694269001484, + 0.018475443124771118, + 0.014285474084317684, + 0.011961504817008972, + 0.012142266146838665, + 0.010219409130513668, + 0.015953946858644485, + 0.005947530269622803, + 0.005463627632707357, + 0.006201604381203651, + 0.004717424511909485, + 0.0052109938114881516, + 0.005500034429132938, + 0.007552430033683777, + 0.006495870184153318, + 0.00830135028809309, + 0.00883630570024252, + 0.008765497244894505, + 0.013958469964563847, + 0.009408608078956604, + 0.010138334706425667, + 0.0213859211653471, + 0.014670804142951965, + 0.004576869308948517, + 0.004388000350445509, + 0.004645560402423143, + 0.003053088206797838, + 0.0074291224591434, + 0.004888440016657114, + 0.005290348082780838, + 0.00494356220588088 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.07003026455640793 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819257251918316 + ], + "opposite_logprob_tokens": [ + -4.018192768096924 + ], + "logprob_reward": -0.01819257251918316, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819257251918316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14345267415046692, + 0.06103753298521042, + 0.08258761465549469, + 0.07450541108846664, + 0.14226621389389038, + 0.11879022419452667, + 0.15902648866176605, + 0.12818662822246552, + 0.02971111610531807, + 0.06043609604239464 + ], + "credit_map": [ + 0.008403376676142216, + 0.010726519860327244, + 0.014261603355407715, + 0.009682347066700459, + 0.01307015772908926, + 0.021689390763640404, + 0.01982978917658329, + 0.04578948765993118, + 0.006579575128853321, + 0.006974899210035801, + 0.007024184335023165, + 0.006792806088924408, + 0.009630589745938778, + 0.00708159152418375, + 0.008330567739903927, + 0.00862332247197628, + 0.00776985939592123, + 0.010804134421050549, + 0.006799300201237202, + 0.013234629295766354, + 0.015754375606775284, + 0.008869498036801815, + 0.010190188884735107, + 0.009165626950562, + 0.007574760355055332, + 0.006789089646190405, + 0.006364202126860619, + 0.015025726519525051, + 0.009877786971628666, + 0.011295507661998272, + 0.0076382155530154705, + 0.009940131567418575, + 0.013147419318556786, + 0.03394647315144539, + 0.011236381717026234, + 0.019984692335128784, + 0.016783401370048523, + 0.017592977732419968, + 0.0116917435079813, + 0.017883136868476868, + 0.013954193331301212, + 0.011554433032870293, + 0.007565595209598541, + 0.00850463192909956, + 0.026822404935956, + 0.026464218273758888, + 0.010825414210557938, + 0.013099349103868008, + 0.019168365746736526, + 0.01727304235100746, + 0.01867004483938217, + 0.014614402316510677, + 0.02709125727415085, + 0.020679449662566185, + 0.011395934969186783, + 0.030133992433547974, + 0.014664122834801674, + 0.01656852848827839, + 0.011078710667788982, + 0.015080629847943783, + 0.020879607647657394, + 0.01983312889933586, + 0.013097966089844704, + 0.016983939334750175, + 0.0027878647670149803, + 0.002311734016984701, + 0.0027360611129552126, + 0.004559597000479698, + 0.002710800152271986, + 0.003250895533710718, + 0.0038747426588088274, + 0.007479421328753233, + 0.006592476740479469, + 0.007334744557738304, + 0.006694566924124956, + 0.006082963198423386, + 0.013474611565470695, + 0.006549738347530365, + 0.0074427989311516285, + 0.006264201831072569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.01819257251918316 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369844704866409 + ], + "opposite_logprob_tokens": [ + -3.408698558807373 + ], + "logprob_reward": -0.03369844704866409, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369844704866409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2678107023239136, + 0.3959540128707886, + 0.03267797455191612, + 0.048862095922231674, + 0.029558559879660606, + 0.04849540814757347, + 0.07471120357513428, + 0.0629388839006424, + 0.017885562032461166, + 0.02110554836690426 + ], + "credit_map": [ + 0.02968614362180233, + 0.02877083420753479, + 0.027854936197400093, + 0.01893717609345913, + 0.03409825637936592, + 0.03524690121412277, + 0.03502844646573067, + 0.058188002556562424, + 0.04075578600168228, + 0.06006569787859917, + 0.04687201976776123, + 0.027690693736076355, + 0.07123973220586777, + 0.04818638414144516, + 0.04230818897485733, + 0.05883551761507988, + 0.003077243221923709, + 0.0065107750706374645, + 0.0033990375231951475, + 0.004435889422893524, + 0.004295549355447292, + 0.004507280420511961, + 0.0038248393684625626, + 0.0026273573748767376, + 0.004204961005598307, + 0.004363618325442076, + 0.0049665034748613834, + 0.008123367093503475, + 0.006160513032227755, + 0.007627865299582481, + 0.0048502772115170956, + 0.008564988151192665, + 0.003989594057202339, + 0.0044852131977677345, + 0.0019212457118555903, + 0.0038666187319904566, + 0.00411189952865243, + 0.004410475958138704, + 0.0027653269935399294, + 0.004008186049759388, + 0.005514299962669611, + 0.003933243453502655, + 0.00264207785949111, + 0.0039922017604112625, + 0.006177439354360104, + 0.014133593067526817, + 0.004016022197902203, + 0.00808653049170971, + 0.007723280228674412, + 0.007824664004147053, + 0.007608816958963871, + 0.009818069636821747, + 0.010269587859511375, + 0.01092139445245266, + 0.00787491723895073, + 0.012670477852225304, + 0.008483625017106533, + 0.006429542787373066, + 0.006806892342865467, + 0.007096163462847471, + 0.007177354302257299, + 0.007981739938259125, + 0.006859827786684036, + 0.012103739194571972, + 0.0020364646334201097, + 0.0022861012257635593, + 0.0020809711422771215, + 0.002539311069995165, + 0.0016513840528205037, + 0.0017867718124762177, + 0.0016966596012935042, + 0.003807897213846445, + 0.002547321142628789, + 0.0034172297455370426, + 0.0021265309769660234, + 0.0022976428736001253, + 0.0035039621870964766, + 0.0023438730277121067, + 0.0022038158494979143, + 0.0026651727966964245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.03369844704866409 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864274337887764 + ], + "opposite_logprob_tokens": [ + -3.048642635345459 + ], + "logprob_reward": -0.04864274337887764, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864274337887764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09548849612474442, + 0.2635948359966278, + 0.07076400518417358, + 0.03859126195311546, + 0.06082770228385925, + 0.21241846680641174, + 0.17546245455741882, + 0.019195474684238434, + 0.03173479065299034, + 0.031922537833452225 + ], + "credit_map": [ + 0.007802589330822229, + 0.014628654345870018, + 0.011390658095479012, + 0.008218611590564251, + 0.01243346743285656, + 0.012856047600507736, + 0.010448910295963287, + 0.017709556967020035, + 0.02614723891019821, + 0.044099412858486176, + 0.03347405046224594, + 0.02330179326236248, + 0.035740870982408524, + 0.031223248690366745, + 0.02867395058274269, + 0.04093426093459129, + 0.007019848097115755, + 0.010893397033214569, + 0.00844812486320734, + 0.009765823371708393, + 0.009985222481191158, + 0.00834785494953394, + 0.008809924125671387, + 0.007493805605918169, + 0.005157827399671078, + 0.0033477460965514183, + 0.003741631517186761, + 0.004837432410567999, + 0.006139599252492189, + 0.005554047878831625, + 0.0037501719780266285, + 0.006062806583940983, + 0.007378842681646347, + 0.006993589922785759, + 0.005509199574589729, + 0.007055278867483139, + 0.006330093834549189, + 0.010788489133119583, + 0.006130227353423834, + 0.010641979984939098, + 0.024376463145017624, + 0.01930418610572815, + 0.011404675431549549, + 0.016999302431941032, + 0.03610359504818916, + 0.054922375828027725, + 0.023523420095443726, + 0.02578446827828884, + 0.022687500342726707, + 0.01818704418838024, + 0.0247596874833107, + 0.027017742395401, + 0.018461236730217934, + 0.02073950693011284, + 0.0172392800450325, + 0.026370462030172348, + 0.0024805262219160795, + 0.0019215878564864397, + 0.0033416468650102615, + 0.002085397718474269, + 0.0020023726392537355, + 0.0022053360007703304, + 0.0031821569427847862, + 0.0019764502067118883, + 0.003060822607949376, + 0.003650257596746087, + 0.0029353112913668156, + 0.0057772425934672356, + 0.003587972605600953, + 0.00330051826313138, + 0.0062394398264586926, + 0.003183226101100445, + 0.003782027866691351, + 0.0034439510200172663, + 0.004738601855933666, + 0.0028037226293236017, + 0.0073819877579808235, + 0.0033217414747923613, + 0.0034278789535164833, + 0.0030226288363337517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.04864274337887764 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Rendered in a warm and captivating style inspired by Studio Ghibli, this close-up shot tilts smoothly upward, tracking a whimsical hand-drawn menu as it gracefully rises through the air, propelled by an unseen magical force. The menu itself is made of thick, cream-colored parchment paper, with slightly crinkled and soft edges. The camera maintains sharp focus on the menu's intricate artwork, showcasing a vibrant array of cakes and sweets rendered in delicate, vibrant watercolors with delicate ink outlines: a puffy strawberry shortcake with sparkling red berries, a miniature fruit tart with a lustrous glaze, and a plump, cream-filled chocolate pastry. As the menu floats calmly upward with a subtle, almost imperceptible sway, the sunlit bakery interior behind it is cast in a soft, dreamy blur. Through this beautiful bokeh, one can see the warm wood textures of the shelves filled with round bread and the soft glow of the glass display cases. Golden sunlight streams in through a large window, illuminating the dust particles dancing in the air and bathing the entire scene in a warm, honeyed glow, creating an atmosphere of purity, tenderness, charm, and comfort.", + "seed": 19903, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_6/index_8206_seed_19903_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06961774118244649, + "vlm_logprob": -0.06961774118244649, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.06961774118244649, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.040486015379428864 + ], + "opposite_logprob_tokens": [ + -3.6654860973358154 + ], + "logprob_reward": -0.040486015379428864, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.040486015379428864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3206993341445923, + 0.10933339595794678, + 0.18954698741436005, + 0.18851186335086823, + 0.057144541293382645, + 0.026370659470558167, + 0.042122986167669296, + 0.029398977756500244, + 0.012998614460229874, + 0.02387263812124729 + ], + "credit_map": [ + 0.020823951810598373, + 0.009251562878489494, + 0.12561650574207306, + 0.01539748627692461, + 0.01159709133207798, + 0.1094546690583229, + 0.010870061814785004, + 0.017688024789094925, + 0.01265717763453722, + 0.009769910015165806, + 0.018642399460077286, + 0.014489192515611649, + 0.009845949709415436, + 0.014313967898488045, + 0.012094620615243912, + 0.017520183697342873, + 0.015614344738423824, + 0.009003179147839546, + 0.054170142859220505, + 0.013869067654013634, + 0.009865806438028812, + 0.01557631604373455, + 0.017465073615312576, + 0.053983062505722046, + 0.029894376173615456, + 0.01224896963685751, + 0.04294727370142937, + 0.01592223532497883, + 0.013006854802370071, + 0.02517252415418625, + 0.018980687484145164, + 0.030338965356349945, + 0.012669906951487064, + 0.00488872267305851, + 0.0063839140348136425, + 0.004910661838948727, + 0.005137655418366194, + 0.006197585724294186, + 0.00580599345266819, + 0.011150103993713856, + 0.007919197902083397, + 0.0013291737996041775, + 0.0022563759703189135, + 0.0013154197949916124, + 0.001803147722966969, + 0.0020947540178894997, + 0.0025676225777715445, + 0.007084968034178019, + 0.0046238889917731285, + 0.003909970633685589, + 0.0061890557408332825, + 0.005688190460205078, + 0.004032899625599384, + 0.00487595796585083, + 0.006304140668362379, + 0.006498882547020912, + 0.0027288668788969517, + 0.004018161445856094, + 0.0032679762225598097, + 0.003298870986327529, + 0.0037429509684443474, + 0.00438927486538887, + 0.004097272641956806, + 0.0038556058425456285, + 0.0015752618201076984, + 0.0016515884781256318, + 0.0019875725265592337, + 0.0013632421614602208, + 0.0015204735100269318, + 0.0019445393700152636, + 0.0012377556413412094, + 0.0017181809525936842, + 0.0024059375282377005, + 0.0019737391266971827, + 0.0024449594784528017, + 0.003963941242545843, + 0.0015772399492561817, + 0.002882043831050396, + 0.0032968681771308184, + 0.005327912047505379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.040486015379428864 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032790154218674 + ], + "opposite_logprob_tokens": [ + -2.350327968597412 + ], + "logprob_reward": -0.10032790154218674, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032790154218674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16057643294334412, + 0.08149942755699158, + 0.21565558016300201, + 0.16046801209449768, + 0.06659073382616043, + 0.08485381305217743, + 0.029537014663219452, + 0.13804768025875092, + 0.01705998368561268, + 0.04571128636598587 + ], + "credit_map": [ + 0.010703257285058498, + 0.005719941109418869, + 0.06299293041229248, + 0.019116031005978584, + 0.005985843949019909, + 0.03800654411315918, + 0.007517228834331036, + 0.010534672066569328, + 0.008290743455290794, + 0.008460796438157558, + 0.01734072156250477, + 0.007333359681069851, + 0.007900937460362911, + 0.011210167780518532, + 0.012529188767075539, + 0.00843350775539875, + 0.017739631235599518, + 0.007747802417725325, + 0.07901431620121002, + 0.01032944954931736, + 0.007201055530458689, + 0.015785347670316696, + 0.008452143520116806, + 0.069385826587677, + 0.022052442654967308, + 0.01047130674123764, + 0.030861899256706238, + 0.013497945852577686, + 0.012524548918008804, + 0.026495810598134995, + 0.01833028346300125, + 0.026233766227960587, + 0.013736997731029987, + 0.004926522262394428, + 0.006870800629258156, + 0.004639435559511185, + 0.007882148027420044, + 0.00717162573710084, + 0.007773543708026409, + 0.013589665293693542, + 0.030402543023228645, + 0.0044326819479465485, + 0.007315721362829208, + 0.0041725654155015945, + 0.0033832977060228586, + 0.005127498880028725, + 0.005357923451811075, + 0.024661587551236153, + 0.0037554865702986717, + 0.003447633469477296, + 0.002741272794082761, + 0.00319654680788517, + 0.0032172491773962975, + 0.0049169911071658134, + 0.003633915912359953, + 0.004627919755876064, + 0.012359005399048328, + 0.015746640041470528, + 0.017350593581795692, + 0.017476005479693413, + 0.016551539301872253, + 0.015498175285756588, + 0.028295399621129036, + 0.014770331792533398, + 0.0017597959376871586, + 0.0030432450585067272, + 0.0018227886175736785, + 0.002038374077528715, + 0.002063231309875846, + 0.0024446314200758934, + 0.0022397313732653856, + 0.001648187986575067, + 0.004005045164376497, + 0.006373800337314606, + 0.004500274546444416, + 0.005342254415154457, + 0.003850865876302123, + 0.004702075384557247, + 0.005759254563599825, + 0.011177721433341503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.10032790154218674 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026357922703027725 + ], + "opposite_logprob_tokens": [ + -3.651357889175415 + ], + "logprob_reward": -0.026357922703027725, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026357922703027725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6618750691413879, + 0.05793021246790886, + 0.16464832425117493, + 0.034672386944293976, + 0.031104257330298424, + 0.017113395035266876, + 0.00941709615290165, + 0.007734267972409725, + 0.007448049262166023, + 0.008056947961449623 + ], + "credit_map": [ + 0.0329836830496788, + 0.013983286917209625, + 0.27813148498535156, + 0.07549422979354858, + 0.02130461111664772, + 0.19883576035499573, + 0.012843945063650608, + 0.028298066928982735, + 0.004115892108529806, + 0.0028426351491361856, + 0.018075380474328995, + 0.004883312154561281, + 0.0033776559866964817, + 0.0047606211155653, + 0.004512495826929808, + 0.015362215228378773, + 0.012653288431465626, + 0.0041984631679952145, + 0.06223238632082939, + 0.007604728452861309, + 0.00379235134460032, + 0.010837670415639877, + 0.006038710940629244, + 0.05729071423411369, + 0.005431103985756636, + 0.004503929056227207, + 0.0056643676944077015, + 0.0037711530458182096, + 0.003043262753635645, + 0.003744021523743868, + 0.0038065791595727205, + 0.004707970190793276, + 0.008522399701178074, + 0.0023796334862709045, + 0.0022895873989909887, + 0.002561543369665742, + 0.0019891569390892982, + 0.002512723207473755, + 0.0022180418018251657, + 0.008631172589957714, + 0.003668468678370118, + 0.0013122677337378263, + 0.001802863902412355, + 0.001143726403824985, + 0.001426716218702495, + 0.0021146147046238184, + 0.0023802758660167456, + 0.0032644609455019236, + 0.0009298466611653566, + 0.0010847552912309766, + 0.001001182827167213, + 0.0016204501735046506, + 0.001046879100613296, + 0.0010425260988995433, + 0.0014948246534913778, + 0.0011966313468292356, + 0.0007818281301297247, + 0.0005769736599177122, + 0.0008633163524791598, + 0.0008321602945216, + 0.0007950449944473803, + 0.0014098068932071328, + 0.0010397202568128705, + 0.0014354169834405184, + 0.0008585291216149926, + 0.0007597057265229523, + 0.0013344385661184788, + 0.0008275171858258545, + 0.0007568772416561842, + 0.0010416572913527489, + 0.0009893376845866442, + 0.0008799862116575241, + 0.0006337188533507288, + 0.0009265763219445944, + 0.0008476785151287913, + 0.0009373213979415596, + 0.0006575667066499591, + 0.0009286956628784537, + 0.0010896299500018358, + 0.0020357610192149878 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.026357922703027725 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020584410056471825 + ], + "opposite_logprob_tokens": [ + -3.8955843448638916 + ], + "logprob_reward": -0.020584410056471825, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020584410056471825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18226376175880432, + 0.21188487112522125, + 0.21520249545574188, + 0.03645852953195572, + 0.11881604790687561, + 0.034550223499536514, + 0.02912786789238453, + 0.08792965859174728, + 0.04176262393593788, + 0.04200390726327896 + ], + "credit_map": [ + 0.009891589172184467, + 0.005456642713397741, + 0.05597993731498718, + 0.018416615203022957, + 0.006805843207985163, + 0.0699288472533226, + 0.0055303070694208145, + 0.010253965854644775, + 0.015403284691274166, + 0.00686732167378068, + 0.0776643231511116, + 0.012369757518172264, + 0.010227381251752377, + 0.01986185647547245, + 0.009307026863098145, + 0.06018391624093056, + 0.016120940446853638, + 0.005440901033580303, + 0.07761018723249435, + 0.008808049373328686, + 0.007473199628293514, + 0.017422666773200035, + 0.00652672303840518, + 0.07579982280731201, + 0.0048227692022919655, + 0.003916926681995392, + 0.005638435482978821, + 0.0036500762216746807, + 0.003020535921677947, + 0.0051255133002996445, + 0.004293604288250208, + 0.005990668665617704, + 0.03216050937771797, + 0.006869570352137089, + 0.007676043082028627, + 0.0074043115600943565, + 0.010367950424551964, + 0.012226337566971779, + 0.007455648388713598, + 0.03465568274259567, + 0.004543454386293888, + 0.0020115473307669163, + 0.005414480809122324, + 0.0028711196500808, + 0.003182344138622284, + 0.004644602071493864, + 0.007582786493003368, + 0.004299888852983713, + 0.003549951361492276, + 0.0019886146765202284, + 0.003626074641942978, + 0.00287067168392241, + 0.003007463412359357, + 0.0035680467262864113, + 0.00556914834305644, + 0.004947897512465715, + 0.0073088305070996284, + 0.009587448090314865, + 0.009556940756738186, + 0.010210791602730751, + 0.009694154374301434, + 0.014203011058270931, + 0.015778539702296257, + 0.011589949950575829, + 0.004716773983091116, + 0.0026011376176029444, + 0.006910081487149, + 0.003226690459996462, + 0.005163604859262705, + 0.008467591367661953, + 0.004715371411293745, + 0.0059613729827106, + 0.0029171514324843884, + 0.004165549762547016, + 0.004343394190073013, + 0.007363845594227314, + 0.0031554060988128185, + 0.004817830864340067, + 0.007740194909274578, + 0.007500535808503628 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.020584410056471825 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a hand-drawn menu made of thick, cream-colored parchment paper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the menu float upward through the air with a subtle, almost imperceptible sway?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the menu display specific items including a strawberry shortcake, a fruit tart, and a chocolate pastry?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a sunlit bakery interior with blurred shelves of bread and glass display cases?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth upward tilt while maintaining sharp focus on the menu?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16033245623111725 + ], + "opposite_logprob_tokens": [ + -1.910332441329956 + ], + "logprob_reward": -0.16033245623111725, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16033245623111725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2608165144920349, + 0.03333013877272606, + 0.27147677540779114, + 0.04729585349559784, + 0.06398194283246994, + 0.11122636497020721, + 0.09199003130197525, + 0.057193271815776825, + 0.017876338213682175, + 0.04481274634599686 + ], + "credit_map": [ + 0.014457856304943562, + 0.0057280720211565495, + 0.1080460473895073, + 0.019870471209287643, + 0.008857160806655884, + 0.08480227738618851, + 0.005940258037298918, + 0.013114354573190212, + 0.0044149667955935, + 0.004125568084418774, + 0.005464974790811539, + 0.004052473697811365, + 0.004995985422283411, + 0.0037469572853296995, + 0.003420739434659481, + 0.0031084760557860136, + 0.018572738394141197, + 0.01003185659646988, + 0.09529861062765121, + 0.015472586266696453, + 0.009424706920981407, + 0.0181672815233469, + 0.01585766114294529, + 0.08865132182836533, + 0.0060244472697377205, + 0.0034200435038655996, + 0.014457947574555874, + 0.0044197035022079945, + 0.0025345676112920046, + 0.005666617304086685, + 0.0041481987573206425, + 0.006624323315918446, + 0.015966229140758514, + 0.004290549084544182, + 0.006418992765247822, + 0.00416242890059948, + 0.003527793800458312, + 0.008526939898729324, + 0.005442872643470764, + 0.015646133571863174, + 0.04081577807664871, + 0.003366349032148719, + 0.011592722497880459, + 0.004190560430288315, + 0.004307913593947887, + 0.006924856454133987, + 0.00550717581063509, + 0.03452100604772568, + 0.010569659993052483, + 0.008118260651826859, + 0.013238906860351562, + 0.012711719609797001, + 0.007719373796135187, + 0.012725278735160828, + 0.012978657148778439, + 0.013928175903856754, + 0.005228235851973295, + 0.0046719638630747795, + 0.006566890049725771, + 0.006237685214728117, + 0.005266560707241297, + 0.0072175851091742516, + 0.01285113487392664, + 0.009153217077255249, + 0.001563770929351449, + 0.002070132177323103, + 0.002117894124239683, + 0.002019706415012479, + 0.002603733679279685, + 0.0033091362565755844, + 0.0015670321881771088, + 0.00262493290938437, + 0.005065510980784893, + 0.004408600274473429, + 0.004573872312903404, + 0.00670586945489049, + 0.0035663535818457603, + 0.005518818274140358, + 0.006528056692332029, + 0.00844566524028778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.16033245623111725 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_7/index_tensor([8329, 8329])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_7/index_tensor([8329, 8329])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9e365d9446295ddfc9ed639781630ecba36a1c91 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_7/index_tensor([8329, 8329])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle macro shot immerses the viewer in a Giger-esque biomechanical landscape, initially focusing on the sickly, bone-white, waxy, bell-shaped flowers of a mutant hosta. Faint, translucent textures etched into the petals tremble with an unnaturally high frequency, creating a subtle, unsettling blur at the edges. Its stem, far from organic, seamlessly merges with a blackened, ribbed metal tube that takes on the texture of pitted iron and an oily sheen. From the grotesque seam where flesh meets metal, a thick, viscous, iridescent black liquid meticulously gushes forth in heavy droplets, oozing and dripping downward with syrupy slowness, tracing the icy contours of the metal. The camera then begins a slow, steady pan back, revealing the alien plant's full, monstrous form. This retreat unveils a vast, surreal setting, a perpetual twilight sky, the ground a mosaic of interlocking plates resembling a massive, carved rib cage. The plant proves to be just one specimen within a dense, nightmarish forest of similar biomechanical horrors, their skeletal forms silhouetted against a dim, hazy horizon. The scene is lit by a single, cold, distant light source, casting long, deep shadows that accentuate the eerie fusion of organic and mechanical textures, creating an oppressive and deeply unsettling atmosphere of profound, alien dread.", + "seed": 24158, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_7/index_8329_seed_24158_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.025162750110030174, + "vlm_logprob": -0.025162750110030174, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.025162750110030174, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007746417541056871 + ], + "opposite_logprob_tokens": [ + -5.25774621963501 + ], + "logprob_reward": -0.007746417541056871, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.007746417541056871, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3405362367630005, + 0.061784036457538605, + 0.05811111256480217, + 0.06911832094192505, + 0.09085885435342789, + 0.03843776509165764, + 0.14135552942752838, + 0.07112983614206314, + 0.08619672060012817, + 0.04247155413031578 + ], + "credit_map": [ + 0.030504707247018814, + 0.07393482327461243, + 0.021209483966231346, + 0.04956445470452309, + 0.06270140409469604, + 0.04298463091254234, + 0.02886463701725006, + 0.03077213279902935, + 0.007269984111189842, + 0.010588588193058968, + 0.00671800272539258, + 0.007172289304435253, + 0.010185005143284798, + 0.006433461792767048, + 0.007474601734429598, + 0.005942108109593391, + 0.005279540549963713, + 0.008706925436854362, + 0.006671446841210127, + 0.007254027761518955, + 0.007106292061507702, + 0.007441251538693905, + 0.010011865757405758, + 0.005639772396534681, + 0.009559477679431438, + 0.0073799854144454, + 0.0101081021130085, + 0.008054099045693874, + 0.008610893040895462, + 0.00966955628246069, + 0.009486840106546879, + 0.006249373313039541, + 0.0102543318644166, + 0.009124594740569592, + 0.014745024032890797, + 0.0214390829205513, + 0.009684250690042973, + 0.009491367265582085, + 0.00705511961132288, + 0.00906508881598711, + 0.004370845854282379, + 0.005513511132448912, + 0.006666446570307016, + 0.006629811134189367, + 0.0039051661733537912, + 0.0039749485440552235, + 0.003975002560764551, + 0.0034020368475466967, + 0.017826741561293602, + 0.01781904511153698, + 0.05234667658805847, + 0.015064573846757412, + 0.010842937044799328, + 0.012250498868525028, + 0.006377565208822489, + 0.008827490732073784, + 0.01596223935484886, + 0.006719817407429218, + 0.017342086881399155, + 0.005995078478008509, + 0.006071712821722031, + 0.006351917050778866, + 0.007212081924080849, + 0.005474905017763376, + 0.013470107689499855, + 0.012638111598789692, + 0.012424924410879612, + 0.012867527082562447, + 0.010343283414840698, + 0.009140083566308022, + 0.006833390332758427, + 0.008479299955070019, + 0.006910307332873344, + 0.005126318894326687, + 0.009236350655555725, + 0.004777975380420685, + 0.004382777493447065, + 0.003025353653356433, + 0.005408291704952717, + 0.0036041822750121355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.007746417541056871 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900860905647278 + ], + "opposite_logprob_tokens": [ + -2.5790085792541504 + ], + "logprob_reward": -0.07900860905647278, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900860905647278, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29259586334228516, + 0.06421441584825516, + 0.08138009905815125, + 0.07114551216363907, + 0.02469748444855213, + 0.0549645759165287, + 0.11465686559677124, + 0.16403959691524506, + 0.09376019984483719, + 0.03854534775018692 + ], + "credit_map": [ + 0.02546551078557968, + 0.07534420490264893, + 0.016639232635498047, + 0.03978424146771431, + 0.05875055119395256, + 0.03126712143421173, + 0.023157773539423943, + 0.022187232971191406, + 0.0097482455894351, + 0.0064267055131495, + 0.006172345485538244, + 0.008107341825962067, + 0.006453061942011118, + 0.009194040670990944, + 0.01158127561211586, + 0.006531400140374899, + 0.006213080137968063, + 0.0066291410475969315, + 0.0051069436594843864, + 0.011904128827154636, + 0.0073433551006019115, + 0.009349789470434189, + 0.026579352095723152, + 0.00825431663542986, + 0.006974902935326099, + 0.008772776462137699, + 0.008487533777952194, + 0.009623358026146889, + 0.010964175686240196, + 0.009950571693480015, + 0.010460796765983105, + 0.005911401938647032, + 0.0027963370084762573, + 0.0024492060765624046, + 0.003938665147870779, + 0.004436127375811338, + 0.002737127710133791, + 0.0035679605789482594, + 0.0025985930114984512, + 0.0021734684705734253, + 0.00821711216121912, + 0.008322625420987606, + 0.007074684835970402, + 0.007814827375113964, + 0.007831948809325695, + 0.0048100706189870834, + 0.00524293864145875, + 0.005650368984788656, + 0.019546611234545708, + 0.011273074895143509, + 0.02207857184112072, + 0.011126731522381306, + 0.02113277092576027, + 0.010438734665513039, + 0.010316123254597187, + 0.00874424446374178, + 0.03380497545003891, + 0.015382647514343262, + 0.04217660799622536, + 0.01742267981171608, + 0.01711498387157917, + 0.012678300961852074, + 0.014189986512064934, + 0.011269417591392994, + 0.013986673206090927, + 0.01284879446029663, + 0.013379727490246296, + 0.01774713024497032, + 0.01230639312416315, + 0.008827966637909412, + 0.007234342861920595, + 0.007429175078868866, + 0.00756562314927578, + 0.005885586608201265, + 0.005007980391383171, + 0.0049359784461557865, + 0.005000961944460869, + 0.003117823973298073, + 0.003167858114466071, + 0.003863536985591054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.07900860905647278 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947037134319544 + ], + "opposite_logprob_tokens": [ + -5.130947113037109 + ], + "logprob_reward": -0.005947037134319544, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947037134319544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3474237024784088, + 0.10625223070383072, + 0.08901169151067734, + 0.05419046804308891, + 0.06566287577152252, + 0.02533966489136219, + 0.12158730626106262, + 0.04527447000145912, + 0.10861237347126007, + 0.03664521500468254 + ], + "credit_map": [ + 0.029187660664319992, + 0.08015327155590057, + 0.0222339928150177, + 0.04823900759220123, + 0.06798826903104782, + 0.04062381759285927, + 0.02857242152094841, + 0.030425263568758965, + 0.011007881723344326, + 0.01516791246831417, + 0.011661591939628124, + 0.014105966314673424, + 0.015942057594656944, + 0.01288765948265791, + 0.01634008064866066, + 0.009139074012637138, + 0.01047282014042139, + 0.010776054114103317, + 0.008365301415324211, + 0.011436141096055508, + 0.008245767094194889, + 0.013875609263777733, + 0.01716529205441475, + 0.008674707263708115, + 0.005431921686977148, + 0.005215912591665983, + 0.006645057816058397, + 0.006819816306233406, + 0.007092407438904047, + 0.004505798686295748, + 0.013461549766361713, + 0.005018005613237619, + 0.006471872795373201, + 0.007411155849695206, + 0.009459075517952442, + 0.013733289204537868, + 0.007189691998064518, + 0.007543974090367556, + 0.005911058280616999, + 0.007942757569253445, + 0.0024459483101963997, + 0.003608129220083356, + 0.003367022145539522, + 0.0032540042884647846, + 0.005458473227918148, + 0.002592238364741206, + 0.0025410291273146868, + 0.002072819508612156, + 0.019353223964571953, + 0.009776361286640167, + 0.037926413118839264, + 0.01085925754159689, + 0.017339864745736122, + 0.009854773059487343, + 0.008190354332327843, + 0.008287067525088787, + 0.007554061710834503, + 0.007449995260685682, + 0.008770243264734745, + 0.0053164102137088776, + 0.006005913484841585, + 0.0033154585398733616, + 0.00369782536290586, + 0.003164564725011587, + 0.01661359891295433, + 0.015511871315538883, + 0.016278082504868507, + 0.01394401304423809, + 0.013317054137587547, + 0.01331842690706253, + 0.00905713438987732, + 0.01057219598442316, + 0.006599448621273041, + 0.007575531955808401, + 0.005372642073780298, + 0.0051890225149691105, + 0.0037014095578342676, + 0.002135148737579584, + 0.003345210338011384, + 0.002726802136749029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.005947037134319544 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533116154372692 + ], + "opposite_logprob_tokens": [ + -4.387533187866211 + ], + "logprob_reward": -0.012533116154372692, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533116154372692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12241043150424957, + 0.1440112590789795, + 0.22596025466918945, + 0.0377408005297184, + 0.0629960373044014, + 0.03637075051665306, + 0.08454421162605286, + 0.07716348022222519, + 0.10588619858026505, + 0.10291662812232971 + ], + "credit_map": [ + 0.009430873207747936, + 0.034207433462142944, + 0.008166108280420303, + 0.014994186349213123, + 0.028675558045506477, + 0.008660531602799892, + 0.007980303838849068, + 0.010295435786247253, + 0.021463964134454727, + 0.011628418229520321, + 0.018021978437900543, + 0.018323317170143127, + 0.015028647147119045, + 0.02087845839560032, + 0.022684471681714058, + 0.015981998294591904, + 0.025496438145637512, + 0.02297040820121765, + 0.020027540624141693, + 0.030776692554354668, + 0.019607743248343468, + 0.03936101868748665, + 0.04282631352543831, + 0.024894103407859802, + 0.004099576268345118, + 0.0033140918239951134, + 0.00464166235178709, + 0.004452670458704233, + 0.004144524689763784, + 0.003579414216801524, + 0.009546061046421528, + 0.003962801303714514, + 0.00819111056625843, + 0.006610166747123003, + 0.01052454300224781, + 0.010094432160258293, + 0.007703269831836224, + 0.006379599682986736, + 0.005296086426824331, + 0.008196831680834293, + 0.0057250033132731915, + 0.004509038757532835, + 0.0051575652323663235, + 0.004301918670535088, + 0.004759281408041716, + 0.0044582136906683445, + 0.0029504538979381323, + 0.004509276710450649, + 0.014229502528905869, + 0.006254150532186031, + 0.021690629422664642, + 0.0077146561816334724, + 0.014704408124089241, + 0.007218292914330959, + 0.006596126593649387, + 0.0061364490538835526, + 0.012425145134329796, + 0.009378032758831978, + 0.017601685598492622, + 0.005992772988975048, + 0.011627995409071445, + 0.006717262323945761, + 0.007284405641257763, + 0.006136181298643351, + 0.0158638134598732, + 0.015979938209056854, + 0.013253373093903065, + 0.0166652612388134, + 0.012486658990383148, + 0.012816743925213814, + 0.009039374999701977, + 0.009781032800674438, + 0.019359495490789413, + 0.02074550651013851, + 0.012666761875152588, + 0.012344694696366787, + 0.011803790926933289, + 0.007830585353076458, + 0.008561640977859497, + 0.009604149498045444 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.012533116154372692 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020578570663928986 + ], + "opposite_logprob_tokens": [ + -3.895578622817993 + ], + "logprob_reward": -0.020578570663928986, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020578570663928986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28667840361595154, + 0.07128266245126724, + 0.08866941928863525, + 0.13732293248176575, + 0.07715719193220139, + 0.053408410400152206, + 0.09589448571205139, + 0.061151374131441116, + 0.035274915397167206, + 0.0931602269411087 + ], + "credit_map": [ + 0.028028523549437523, + 0.04358961805701256, + 0.01854703389108181, + 0.041076432913541794, + 0.04511730372905731, + 0.05052810534834862, + 0.02290305122733116, + 0.03688833862543106, + 0.0068519688211381435, + 0.008949859999120235, + 0.010619009844958782, + 0.010776272974908352, + 0.010387260466814041, + 0.008144414983689785, + 0.008998560719192028, + 0.006555311381816864, + 0.006369059439748526, + 0.008537786081433296, + 0.007061783690005541, + 0.013760807923972607, + 0.006394727621227503, + 0.012316848151385784, + 0.024188317358493805, + 0.010040088556706905, + 0.013368702493607998, + 0.015292981639504433, + 0.01850949414074421, + 0.016114871948957443, + 0.014777573756873608, + 0.023142721503973007, + 0.02252774126827717, + 0.01358884572982788, + 0.010118881240487099, + 0.007119550369679928, + 0.014066215604543686, + 0.01505951676517725, + 0.007891993969678879, + 0.008946840651333332, + 0.005868340842425823, + 0.008085845038294792, + 0.010582970455288887, + 0.006068721413612366, + 0.00901168305426836, + 0.005735564511269331, + 0.004584727808833122, + 0.004993216600269079, + 0.003941988106817007, + 0.008489538915455341, + 0.01895783469080925, + 0.008136379532516003, + 0.026705225929617882, + 0.008296565152704716, + 0.012968400493264198, + 0.007434210740029812, + 0.007064553443342447, + 0.006331309210509062, + 0.013417810201644897, + 0.004934880416840315, + 0.012356958352029324, + 0.006627105642110109, + 0.010903965681791306, + 0.004057547077536583, + 0.004926302470266819, + 0.003926800098270178, + 0.007196242921054363, + 0.004863875452429056, + 0.005324686877429485, + 0.005973040591925383, + 0.003789400914683938, + 0.003508985275402665, + 0.002526482567191124, + 0.0020922012627124786, + 0.014526166021823883, + 0.015354123897850513, + 0.019124075770378113, + 0.015270554460585117, + 0.007479202933609486, + 0.005180134437978268, + 0.009733375161886215, + 0.00649259751662612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.020578570663928986 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle macro shot immerses the viewer in a Giger-esque biomechanical landscape, initially focusing on the sickly, bone-white, waxy, bell-shaped flowers of a mutant hosta. Faint, translucent textures etched into the petals tremble with an unnaturally high frequency, creating a subtle, unsettling blur at the edges. Its stem, far from organic, seamlessly merges with a blackened, ribbed metal tube that takes on the texture of pitted iron and an oily sheen. From the grotesque seam where flesh meets metal, a thick, viscous, iridescent black liquid meticulously gushes forth in heavy droplets, oozing and dripping downward with syrupy slowness, tracing the icy contours of the metal. The camera then begins a slow, steady pan back, revealing the alien plant's full, monstrous form. This retreat unveils a vast, surreal setting, a perpetual twilight sky, the ground a mosaic of interlocking plates resembling a massive, carved rib cage. The plant proves to be just one specimen within a dense, nightmarish forest of similar biomechanical horrors, their skeletal forms silhouetted against a dim, hazy horizon. The scene is lit by a single, cold, distant light source, casting long, deep shadows that accentuate the eerie fusion of organic and mechanical textures, creating an oppressive and deeply unsettling atmosphere of profound, alien dread.", + "seed": 24159, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_7/index_8329_seed_24159_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03500917097553611, + "vlm_logprob": -0.03500917097553611, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.03500917097553611, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005431418772786856 + ], + "opposite_logprob_tokens": [ + -5.505431652069092 + ], + "logprob_reward": -0.005431418772786856, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005431418772786856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5408822894096375, + 0.04915846139192581, + 0.0553201325237751, + 0.07419067621231079, + 0.056292515248060226, + 0.03324669972062111, + 0.049543190747499466, + 0.05023399367928505, + 0.05166785046458244, + 0.039464183151721954 + ], + "credit_map": [ + 0.03632272034883499, + 0.029202233999967575, + 0.04106820002198219, + 0.056901149451732635, + 0.025650447234511375, + 0.1404847949743271, + 0.043282631784677505, + 0.16797010600566864, + 0.0052877310663461685, + 0.0064906170591712, + 0.006244555581361055, + 0.004604370333254337, + 0.00582619896158576, + 0.007285721600055695, + 0.006498894654214382, + 0.0069203744642436504, + 0.00896324310451746, + 0.005498428363353014, + 0.00804095808416605, + 0.004126598127186298, + 0.004695269279181957, + 0.007576164323836565, + 0.003695202525705099, + 0.01272426825016737, + 0.016558846458792686, + 0.006791022606194019, + 0.008440096862614155, + 0.004141698125749826, + 0.004621386993676424, + 0.016756882891058922, + 0.00523412274196744, + 0.011646619066596031, + 0.0069260671734809875, + 0.006813052110373974, + 0.004704542923718691, + 0.00553296459838748, + 0.00936750415712595, + 0.008255509659647942, + 0.008032447658479214, + 0.006660424172878265, + 0.0019393209367990494, + 0.0037044882774353027, + 0.0022587471175938845, + 0.0066841403022408485, + 0.003874232294037938, + 0.005506632383912802, + 0.004719768185168505, + 0.004559369292110205, + 0.0039036262314766645, + 0.009789235889911652, + 0.0046764155849814415, + 0.007593899499624968, + 0.007176067214459181, + 0.006057498045265675, + 0.004868722055107355, + 0.005477724131196737, + 0.0042429473251104355, + 0.005651002749800682, + 0.00501667894423008, + 0.006429190281778574, + 0.007389830891042948, + 0.009362980723381042, + 0.005543943028897047, + 0.006597417872399092, + 0.0042355055920779705, + 0.007091446314007044, + 0.0047483425587415695, + 0.01170269213616848, + 0.006200340576469898, + 0.006561641581356525, + 0.005894760135561228, + 0.005233116447925568, + 0.0038644929882138968, + 0.004083692096173763, + 0.0037818700075149536, + 0.005070074461400509, + 0.006656845565885305, + 0.005448855925351381, + 0.005019082687795162, + 0.005539271514862776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.005431418772786856 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270582377910614 + ], + "opposite_logprob_tokens": [ + -2.127058267593384 + ], + "logprob_reward": -0.1270582377910614, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270582377910614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26130548119544983, + 0.10641584545373917, + 0.11652646958827972, + 0.12289150804281235, + 0.05438661947846413, + 0.0775584876537323, + 0.04020468890666962, + 0.07753000408411026, + 0.10121642798185349, + 0.041964393109083176 + ], + "credit_map": [ + 0.023332679644227028, + 0.026286136358976364, + 0.031663525849580765, + 0.03504011780023575, + 0.017439842224121094, + 0.044509366154670715, + 0.0288136824965477, + 0.054220154881477356, + 0.015067629516124725, + 0.012412797659635544, + 0.012716837227344513, + 0.010065274313092232, + 0.016907446086406708, + 0.012219307944178581, + 0.014274373650550842, + 0.012752173468470573, + 0.017858384177088737, + 0.009901520796120167, + 0.014618352986872196, + 0.010923166759312153, + 0.017135044559836388, + 0.014190382324159145, + 0.0075665744952857494, + 0.024333041161298752, + 0.025092488154768944, + 0.012489274144172668, + 0.011794542893767357, + 0.007816251367330551, + 0.009771096520125866, + 0.02856219932436943, + 0.012928737327456474, + 0.014436918310821056, + 0.004969533067196608, + 0.007621145341545343, + 0.00419261259958148, + 0.006236402317881584, + 0.008919045329093933, + 0.0062519884668290615, + 0.007647024467587471, + 0.008548865094780922, + 0.006100822705775499, + 0.007785022724419832, + 0.004856826737523079, + 0.015138855203986168, + 0.007277116645127535, + 0.011356968432664871, + 0.011267120949923992, + 0.013775748200714588, + 0.00334269180893898, + 0.005186775233596563, + 0.004807960242033005, + 0.003512507537379861, + 0.00649538729339838, + 0.006833595689386129, + 0.004878971725702286, + 0.0051468005403876305, + 0.006869781296700239, + 0.008620986714959145, + 0.011602109298110008, + 0.006621969863772392, + 0.009390383027493954, + 0.01325483899563551, + 0.010558806359767914, + 0.010611126199364662, + 0.007987862452864647, + 0.009794730693101883, + 0.012458994053304195, + 0.01687784679234028, + 0.016520768404006958, + 0.014423834159970284, + 0.0103891771286726, + 0.01276321616023779, + 0.00372995319776237, + 0.004428780637681484, + 0.003761140862479806, + 0.005446088965982199, + 0.007054544519633055, + 0.005864748731255531, + 0.004913695622235537, + 0.006765443831682205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.1270582377910614 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036139669828116894 + ], + "opposite_logprob_tokens": [ + -5.6286139488220215 + ], + "logprob_reward": -0.0036139669828116894, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036139669828116894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4407120943069458, + 0.053647350519895554, + 0.10354533791542053, + 0.05066455528140068, + 0.06268727779388428, + 0.061431389302015305, + 0.0711534395813942, + 0.05410400778055191, + 0.05347921699285507, + 0.04857533052563667 + ], + "credit_map": [ + 0.023202117532491684, + 0.017150087282061577, + 0.026977136731147766, + 0.030449166893959045, + 0.01587093621492386, + 0.14671175181865692, + 0.02639719285070896, + 0.1539537012577057, + 0.005640102084726095, + 0.006190762855112553, + 0.0077479747124016285, + 0.004334418568760157, + 0.007068099454045296, + 0.006716503296047449, + 0.0068246894516050816, + 0.009124796837568283, + 0.012117253616452217, + 0.015775533393025398, + 0.014734813943505287, + 0.010195515118539333, + 0.014154372736811638, + 0.011522258631885052, + 0.006641418673098087, + 0.018404170870780945, + 0.008426682092249393, + 0.007282480597496033, + 0.006067479029297829, + 0.0036317617632448673, + 0.004514822270721197, + 0.006599403917789459, + 0.004196049179881811, + 0.009945877827703953, + 0.005966328550130129, + 0.00852982234209776, + 0.005062699317932129, + 0.010258061811327934, + 0.009524966590106487, + 0.007515059784054756, + 0.008614178746938705, + 0.0072161611169576645, + 0.0048840721137821674, + 0.007863950915634632, + 0.003976297099143267, + 0.01161067746579647, + 0.006621085572987795, + 0.00880921259522438, + 0.008947838097810745, + 0.00871825497597456, + 0.006045597605407238, + 0.01249856036156416, + 0.007217995822429657, + 0.009677249006927013, + 0.011508139781653881, + 0.008475345559418201, + 0.007094384636729956, + 0.008636172860860825, + 0.0038292815443128347, + 0.004266171250492334, + 0.008275340311229229, + 0.007219926454126835, + 0.00711377477273345, + 0.010259129106998444, + 0.005886378698050976, + 0.007254006806761026, + 0.004609243478626013, + 0.006142533849924803, + 0.0053227748721838, + 0.009623652324080467, + 0.008397698402404785, + 0.007881133817136288, + 0.006351367104798555, + 0.005150811746716499, + 0.0044985245913267136, + 0.004743233323097229, + 0.005498313810676336, + 0.005877879448235035, + 0.007390874437987804, + 0.005834512412548065, + 0.006715565454214811, + 0.008016426116228104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.0036139669828116894 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033691875636577606 + ], + "opposite_logprob_tokens": [ + -3.408691883087158 + ], + "logprob_reward": -0.033691875636577606, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033691875636577606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.357585608959198, + 0.06438450515270233, + 0.09193125367164612, + 0.044963862746953964, + 0.0984373465180397, + 0.06962531805038452, + 0.09935499727725983, + 0.09466061741113663, + 0.04383622482419014, + 0.03522023186087608 + ], + "credit_map": [ + 0.01794413849711418, + 0.011134142056107521, + 0.021369531750679016, + 0.01666804403066635, + 0.010358475148677826, + 0.1361662894487381, + 0.015007910318672657, + 0.12893711030483246, + 0.005898594856262207, + 0.008370401337742805, + 0.007848245091736317, + 0.006345232482999563, + 0.009172051213681698, + 0.009402194991707802, + 0.009969238191843033, + 0.007378553040325642, + 0.010951072908937931, + 0.012501221150159836, + 0.014784242026507854, + 0.009308901615440845, + 0.012597969733178616, + 0.009952783584594727, + 0.005359902512282133, + 0.01647515781223774, + 0.008466954343020916, + 0.004698248580098152, + 0.005707400385290384, + 0.003360756440088153, + 0.003280542092397809, + 0.006204014178365469, + 0.00460710097104311, + 0.008638844825327396, + 0.008916497230529785, + 0.013556492514908314, + 0.007832269184291363, + 0.013670428656041622, + 0.014227297157049179, + 0.011565828695893288, + 0.017580365762114525, + 0.011088159866631031, + 0.006262237671762705, + 0.009034448303282261, + 0.006020278669893742, + 0.012414298951625824, + 0.009301687590777874, + 0.008661486208438873, + 0.008192592300474644, + 0.009738290682435036, + 0.006691824644804001, + 0.02361488901078701, + 0.008162089623510838, + 0.018541330471634865, + 0.012430169619619846, + 0.008302804082632065, + 0.010723954997956753, + 0.010887942276895046, + 0.005698659457266331, + 0.008136151358485222, + 0.008450810797512531, + 0.012783302925527096, + 0.015416804701089859, + 0.01504739560186863, + 0.01631881296634674, + 0.012808683328330517, + 0.004026519600301981, + 0.0044709909707307816, + 0.005746849812567234, + 0.007409817539155483, + 0.0061960783787071705, + 0.006156271789222956, + 0.005019122734665871, + 0.004810575395822525, + 0.0036850348114967346, + 0.0036280888598412275, + 0.003570798086002469, + 0.005505845416337252, + 0.005142289213836193, + 0.004658496472984552, + 0.004710061941295862, + 0.004319613799452782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.033691875636577606 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052503556944429874 + ], + "opposite_logprob_tokens": [ + -5.255250453948975 + ], + "logprob_reward": -0.0052503556944429874, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052503556944429874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37461304664611816, + 0.07782867550849915, + 0.059729285538196564, + 0.07633815705776215, + 0.08414587378501892, + 0.07977673411369324, + 0.0737740620970726, + 0.05204961076378822, + 0.053965140134096146, + 0.06777949631214142 + ], + "credit_map": [ + 0.019392982125282288, + 0.01075867097824812, + 0.022929290309548378, + 0.019781840965151787, + 0.00872416328638792, + 0.14379990100860596, + 0.016396090388298035, + 0.13283008337020874, + 0.0073969061486423016, + 0.009301726706326008, + 0.010942176915705204, + 0.009555119089782238, + 0.011032722890377045, + 0.00886034406721592, + 0.011639134027063847, + 0.009100537747144699, + 0.010819694958627224, + 0.006389956921339035, + 0.009717818349599838, + 0.004714634735137224, + 0.007902556098997593, + 0.005288837011903524, + 0.003632898209616542, + 0.01126288902014494, + 0.01064947061240673, + 0.015034355223178864, + 0.00883093848824501, + 0.006519399583339691, + 0.010141124948859215, + 0.008520171046257019, + 0.006476591806858778, + 0.010166102088987827, + 0.0072921025566756725, + 0.011969241313636303, + 0.007354839704930782, + 0.011366138234734535, + 0.013226012699306011, + 0.010559683665633202, + 0.011204445734620094, + 0.011173403821885586, + 0.007063950877636671, + 0.010978827252984047, + 0.005310918670147657, + 0.014001843519508839, + 0.007502328604459763, + 0.012375397607684135, + 0.011658779345452785, + 0.01088467612862587, + 0.007040613796561956, + 0.007930289022624493, + 0.0067592803388834, + 0.008820620365440845, + 0.01038131583482027, + 0.014280877076089382, + 0.009311622940003872, + 0.009249433875083923, + 0.0039640129543840885, + 0.005339835304766893, + 0.00537176663056016, + 0.0056836167350411415, + 0.007183055859059095, + 0.009699804708361626, + 0.00772301247343421, + 0.007084501441568136, + 0.0058697983622550964, + 0.005522437859326601, + 0.004738286137580872, + 0.006832003127783537, + 0.01129131019115448, + 0.008067812770605087, + 0.00668514147400856, + 0.004958339035511017, + 0.005611511413007975, + 0.006415033712983131, + 0.008088897913694382, + 0.010223216377198696, + 0.013379978016018867, + 0.006463614758104086, + 0.009927164763212204, + 0.007670068182051182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0052503556944429874 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle macro shot immerses the viewer in a Giger-esque biomechanical landscape, initially focusing on the sickly, bone-white, waxy, bell-shaped flowers of a mutant hosta. Faint, translucent textures etched into the petals tremble with an unnaturally high frequency, creating a subtle, unsettling blur at the edges. Its stem, far from organic, seamlessly merges with a blackened, ribbed metal tube that takes on the texture of pitted iron and an oily sheen. From the grotesque seam where flesh meets metal, a thick, viscous, iridescent black liquid meticulously gushes forth in heavy droplets, oozing and dripping downward with syrupy slowness, tracing the icy contours of the metal. The camera then begins a slow, steady pan back, revealing the alien plant's full, monstrous form. This retreat unveils a vast, surreal setting, a perpetual twilight sky, the ground a mosaic of interlocking plates resembling a massive, carved rib cage. The plant proves to be just one specimen within a dense, nightmarish forest of similar biomechanical horrors, their skeletal forms silhouetted against a dim, hazy horizon. The scene is lit by a single, cold, distant light source, casting long, deep shadows that accentuate the eerie fusion of organic and mechanical textures, creating an oppressive and deeply unsettling atmosphere of profound, alien dread.", + "seed": 24160, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_7/index_8329_seed_24160_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05430313814431429, + "vlm_logprob": -0.05430313814431429, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.05430313814431429, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0304553359746933 + ], + "opposite_logprob_tokens": [ + -3.6554553508758545 + ], + "logprob_reward": -0.0304553359746933, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0304553359746933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22139130532741547, + 0.08474421501159668, + 0.1463032364845276, + 0.04022493585944176, + 0.06723498553037643, + 0.06730159372091293, + 0.11188913136720657, + 0.05496535822749138, + 0.15239107608795166, + 0.05355420336127281 + ], + "credit_map": [ + 0.01983749121427536, + 0.01715780608355999, + 0.020702611654996872, + 0.024834178388118744, + 0.04067499563097954, + 0.03399952873587608, + 0.036515332758426666, + 0.027669351547956467, + 0.0072457618080079556, + 0.00753448624163866, + 0.008606159128248692, + 0.006608177442103624, + 0.0081480098888278, + 0.015258446335792542, + 0.013146478682756424, + 0.01819668896496296, + 0.010606310330331326, + 0.017505569383502007, + 0.016343723982572556, + 0.015157456509768963, + 0.02454371005296707, + 0.01830526441335678, + 0.02011958509683609, + 0.023721611127257347, + 0.0031508866231888533, + 0.004247997421771288, + 0.004803667310625315, + 0.004725460894405842, + 0.008322544395923615, + 0.005235522985458374, + 0.005456145852804184, + 0.004282712936401367, + 0.005058112554252148, + 0.005795540753751993, + 0.010089018382132053, + 0.009123616851866245, + 0.012157448567450047, + 0.00927872583270073, + 0.008316428400576115, + 0.007416092325001955, + 0.005177176557481289, + 0.007597656454890966, + 0.0084841912612319, + 0.011281662620604038, + 0.01725120097398758, + 0.00712173618376255, + 0.005567260552197695, + 0.004820703994482756, + 0.007274376228451729, + 0.011516764760017395, + 0.011795871891081333, + 0.01363297738134861, + 0.016768626868724823, + 0.014259830117225647, + 0.01105797104537487, + 0.025582706555724144, + 0.00491070281714201, + 0.004795527085661888, + 0.004405069164931774, + 0.007881724275648594, + 0.006989645306020975, + 0.006298031657934189, + 0.007366836536675692, + 0.012317824177443981, + 0.010349447838962078, + 0.01635441742837429, + 0.015293657779693604, + 0.021787691861391068, + 0.0294045377522707, + 0.01570896804332733, + 0.027270901948213577, + 0.016221454367041588, + 0.006257275119423866, + 0.0056693474762141705, + 0.0075922394171357155, + 0.007862838916480541, + 0.005289134569466114, + 0.004948130808770657, + 0.006072556599974632, + 0.00986267626285553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.0304553359746933 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20163649320602417 + ], + "opposite_logprob_tokens": [ + -1.701636552810669 + ], + "logprob_reward": -0.20163649320602417, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20163649320602417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24246464669704437, + 0.07832784205675125, + 0.162992462515831, + 0.04837527871131897, + 0.030825914815068245, + 0.09431637078523636, + 0.13162802159786224, + 0.06797972321510315, + 0.044918593019247055, + 0.0981711745262146 + ], + "credit_map": [ + 0.02112496830523014, + 0.0190338846296072, + 0.02540132775902748, + 0.03096817620098591, + 0.042213212698698044, + 0.0368054024875164, + 0.037737905979156494, + 0.029179781675338745, + 0.006569207645952702, + 0.00795475859194994, + 0.007923060096800327, + 0.011745613068342209, + 0.0074517433531582355, + 0.012593695893883705, + 0.010061416774988174, + 0.014028353616595268, + 0.019531987607479095, + 0.015442597679793835, + 0.018202822655439377, + 0.017614902928471565, + 0.021905874833464622, + 0.022750340402126312, + 0.021991336718201637, + 0.025552598759531975, + 0.00417714286595583, + 0.008258967660367489, + 0.006301913410425186, + 0.006363883148878813, + 0.00622034864500165, + 0.00638780789449811, + 0.005072467960417271, + 0.005592744331806898, + 0.0019529247656464577, + 0.003313013119623065, + 0.004123678430914879, + 0.0038302901666611433, + 0.005622676573693752, + 0.0040452019311487675, + 0.004271849058568478, + 0.003666279371827841, + 0.006717759650200605, + 0.009928615763783455, + 0.01171072293072939, + 0.0159224271774292, + 0.023676617071032524, + 0.010044988244771957, + 0.009088094346225262, + 0.007227146532386541, + 0.008235003799200058, + 0.012661214917898178, + 0.012566638179123402, + 0.022113747894763947, + 0.018350599333643913, + 0.011045648716390133, + 0.013387629762291908, + 0.03326752409338951, + 0.006063039880245924, + 0.006656428799033165, + 0.005728451069444418, + 0.00979052484035492, + 0.008439320139586926, + 0.007645111996680498, + 0.009011480957269669, + 0.014645369723439217, + 0.003140633460134268, + 0.00457492470741272, + 0.0042487503960728645, + 0.0064877960830926895, + 0.010155930183827877, + 0.004174672998487949, + 0.007308132015168667, + 0.00482775317505002, + 0.011356492526829243, + 0.010547509416937828, + 0.014494611881673336, + 0.011915319599211216, + 0.011422878131270409, + 0.009227704256772995, + 0.011905205436050892, + 0.01730145514011383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.20163649320602417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640880040824413 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.004640880040824413, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640880040824413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09321638196706772, + 0.13556382060050964, + 0.10122385621070862, + 0.10222925245761871, + 0.05508425086736679, + 0.13781847059726715, + 0.1696816384792328, + 0.04467223957180977, + 0.07544887065887451, + 0.08506124466657639 + ], + "credit_map": [ + 0.00803756806999445, + 0.007068063132464886, + 0.009924616664648056, + 0.015296139754354954, + 0.014905589632689953, + 0.01328533235937357, + 0.012158039957284927, + 0.01254102773964405, + 0.012214787304401398, + 0.011143972165882587, + 0.013153020292520523, + 0.01551352720707655, + 0.013509590178728104, + 0.023128215223550797, + 0.01924576237797737, + 0.02765493467450142, + 0.008310463279485703, + 0.013565420173108578, + 0.011821822263300419, + 0.007256580051034689, + 0.014861609786748886, + 0.01512094121426344, + 0.010868504643440247, + 0.01941850781440735, + 0.0071715437807142735, + 0.009363293647766113, + 0.010295386426150799, + 0.013274810276925564, + 0.012136358767747879, + 0.016570862382650375, + 0.015815097838640213, + 0.017601890489459038, + 0.0037518050521612167, + 0.004617448430508375, + 0.007191481068730354, + 0.00629389425739646, + 0.009848945774137974, + 0.00798349641263485, + 0.008181548677384853, + 0.007215625140815973, + 0.01043596863746643, + 0.015042086131870747, + 0.014886698685586452, + 0.01788165792822838, + 0.03132788464426994, + 0.015501215122640133, + 0.018939625471830368, + 0.01380332000553608, + 0.011766360141336918, + 0.020830487832427025, + 0.021295610815286636, + 0.020887713879346848, + 0.025859851390123367, + 0.024186737835407257, + 0.018001431599259377, + 0.026853423565626144, + 0.0039764223620295525, + 0.004102693870663643, + 0.0047935787588357925, + 0.006759378127753735, + 0.005115775857120752, + 0.004873182624578476, + 0.005932152271270752, + 0.009119050577282906, + 0.005007940344512463, + 0.009416180662810802, + 0.007811453193426132, + 0.010467487387359142, + 0.012869427911937237, + 0.007715617306530476, + 0.010621828958392143, + 0.011538926512002945, + 0.010792098939418793, + 0.009152173064649105, + 0.013114715926349163, + 0.01056278869509697, + 0.009315215982496738, + 0.008067714050412178, + 0.009051351808011532, + 0.015005175955593586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.004640880040824413 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195770025253296 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195770025253296, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195770025253296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09701679646968842, + 0.192781001329422, + 0.0770643800497055, + 0.08009914308786392, + 0.07240337878465652, + 0.13434657454490662, + 0.12408949434757233, + 0.06481762230396271, + 0.10821425914764404, + 0.049167364835739136 + ], + "credit_map": [ + 0.007735473569482565, + 0.007668375037610531, + 0.007953363470733166, + 0.01230641920119524, + 0.014577179215848446, + 0.015609010122716427, + 0.016669752076268196, + 0.014497223310172558, + 0.016485651955008507, + 0.013968795537948608, + 0.01883046142756939, + 0.023358087986707687, + 0.016849718987941742, + 0.032857298851013184, + 0.02667679823935032, + 0.04375418275594711, + 0.00588576402515173, + 0.009140372276306152, + 0.008433200418949127, + 0.006167924962937832, + 0.01110883243381977, + 0.011725052259862423, + 0.00955009926110506, + 0.01505313627421856, + 0.008724147453904152, + 0.007304623257368803, + 0.009291133843362331, + 0.008250132203102112, + 0.008027900941669941, + 0.010167358443140984, + 0.01538680586963892, + 0.012947043403983116, + 0.004596103448420763, + 0.005843770690262318, + 0.0077649555169045925, + 0.009443145245313644, + 0.011732039973139763, + 0.010212438181042671, + 0.009840833023190498, + 0.012970094569027424, + 0.011459293775260448, + 0.015367750078439713, + 0.016880927607417107, + 0.018131515011191368, + 0.03748007491230965, + 0.013775955885648727, + 0.011654311791062355, + 0.009596752934157848, + 0.007543992251157761, + 0.00761261535808444, + 0.007419710047543049, + 0.02768995240330696, + 0.0069754477590322495, + 0.014316800981760025, + 0.009090385399758816, + 0.043440595269203186, + 0.005832647904753685, + 0.00663755415007472, + 0.006282440386712551, + 0.008789756335318089, + 0.010738528333604336, + 0.0073753283359110355, + 0.0076446630991995335, + 0.01151670329272747, + 0.00770907336845994, + 0.011852477677166462, + 0.008710095658898354, + 0.01700141467154026, + 0.030224688351154327, + 0.010313878767192364, + 0.015837732702493668, + 0.006564894691109657, + 0.006365589331835508, + 0.0052536409348249435, + 0.007946833036839962, + 0.006117684300988913, + 0.00540761835873127, + 0.0046636685729026794, + 0.005542923230677843, + 0.007869405671954155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 6 + ], + "selected_reward": -0.014195770025253296 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020587211474776268 + ], + "opposite_logprob_tokens": [ + -3.895587205886841 + ], + "logprob_reward": -0.020587211474776268, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020587211474776268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060151293873786926, + 0.13209591805934906, + 0.12049680203199387, + 0.06638279557228088, + 0.024267923086881638, + 0.09550967067480087, + 0.2802130877971649, + 0.07841642946004868, + 0.027940403670072556, + 0.11452563107013702 + ], + "credit_map": [ + 0.004047631751745939, + 0.005113390740007162, + 0.004954125732183456, + 0.010153837502002716, + 0.009657260961830616, + 0.009311883710324764, + 0.010254417546093464, + 0.006658756174147129, + 0.01050659641623497, + 0.008477727882564068, + 0.01306049432605505, + 0.013366063125431538, + 0.010708656162023544, + 0.024702658876776695, + 0.0189820509403944, + 0.032291676849126816, + 0.00869044754654169, + 0.01379257719963789, + 0.012625322677195072, + 0.013831614516675472, + 0.020949479192495346, + 0.01556572038680315, + 0.01730232499539852, + 0.017739327624440193, + 0.005146815907210112, + 0.0076465243473649025, + 0.006577587686479092, + 0.00814120564609766, + 0.007201892323791981, + 0.009628024883568287, + 0.010215713642537594, + 0.011825041845440865, + 0.002930945483967662, + 0.002436150098219514, + 0.003514868440106511, + 0.00236689206212759, + 0.0034657421056181192, + 0.0030367588624358177, + 0.002916922327131033, + 0.0035996469669044018, + 0.0065452479757368565, + 0.0101132458075881, + 0.011170714162290096, + 0.013731294311583042, + 0.01675737276673317, + 0.010599114000797272, + 0.015730973333120346, + 0.010861708782613277, + 0.0130236791446805, + 0.014442062936723232, + 0.016771722584962845, + 0.046391911804676056, + 0.015231438912451267, + 0.058357685804367065, + 0.015521078370511532, + 0.10047353059053421, + 0.006198636721819639, + 0.005597649607807398, + 0.00611210847273469, + 0.011922118254005909, + 0.007683145813643932, + 0.010523473843932152, + 0.010377765633165836, + 0.02000153250992298, + 0.002278842031955719, + 0.0026773656718432903, + 0.0023587027098983526, + 0.0037105621304363012, + 0.0071705300360918045, + 0.002890856470912695, + 0.004030589014291763, + 0.0028229583986103535, + 0.014414660632610321, + 0.010647247545421124, + 0.018330009654164314, + 0.013178248889744282, + 0.011577010154724121, + 0.010323015041649342, + 0.013255701400339603, + 0.02279973402619362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.020587211474776268 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle macro shot immerses the viewer in a Giger-esque biomechanical landscape, initially focusing on the sickly, bone-white, waxy, bell-shaped flowers of a mutant hosta. Faint, translucent textures etched into the petals tremble with an unnaturally high frequency, creating a subtle, unsettling blur at the edges. Its stem, far from organic, seamlessly merges with a blackened, ribbed metal tube that takes on the texture of pitted iron and an oily sheen. From the grotesque seam where flesh meets metal, a thick, viscous, iridescent black liquid meticulously gushes forth in heavy droplets, oozing and dripping downward with syrupy slowness, tracing the icy contours of the metal. The camera then begins a slow, steady pan back, revealing the alien plant's full, monstrous form. This retreat unveils a vast, surreal setting, a perpetual twilight sky, the ground a mosaic of interlocking plates resembling a massive, carved rib cage. The plant proves to be just one specimen within a dense, nightmarish forest of similar biomechanical horrors, their skeletal forms silhouetted against a dim, hazy horizon. The scene is lit by a single, cold, distant light source, casting long, deep shadows that accentuate the eerie fusion of organic and mechanical textures, creating an oppressive and deeply unsettling atmosphere of profound, alien dread.", + "seed": 24161, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_7/index_8329_seed_24161_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0517265610396862, + "vlm_logprob": -0.0517265610396862, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.0517265610396862, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030442269518971443 + ], + "opposite_logprob_tokens": [ + -3.655442237854004 + ], + "logprob_reward": -0.030442269518971443, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.030442269518971443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31991609930992126, + 0.08604162186384201, + 0.09757009893655777, + 0.028282806277275085, + 0.042253296822309494, + 0.038183990865945816, + 0.10307356715202332, + 0.05182491987943649, + 0.040308877825737, + 0.19254475831985474 + ], + "credit_map": [ + 0.0355457179248333, + 0.025289157405495644, + 0.07327688485383987, + 0.04739149287343025, + 0.0323743037879467, + 0.02314598672091961, + 0.04582850635051727, + 0.037064049392938614, + 0.009485077112913132, + 0.006697372067719698, + 0.011082952842116356, + 0.016084162518382072, + 0.007930456660687923, + 0.005803104490041733, + 0.02084021270275116, + 0.008118280209600925, + 0.006831466220319271, + 0.013697854243218899, + 0.012974820099771023, + 0.015316970646381378, + 0.012186006642878056, + 0.005398275796324015, + 0.023884423077106476, + 0.007280286401510239, + 0.0032986807636916637, + 0.0029882024973630905, + 0.005663127638399601, + 0.004728103522211313, + 0.0027153922710567713, + 0.0027148944791406393, + 0.003782762447372079, + 0.0023916426580399275, + 0.0043517523445189, + 0.005038014613091946, + 0.004214538726955652, + 0.004719045478850603, + 0.00513747101649642, + 0.0075702304020524025, + 0.006676090881228447, + 0.004546153824776411, + 0.0037331562489271164, + 0.004148305393755436, + 0.005744182970374823, + 0.004599936306476593, + 0.003984376788139343, + 0.006048162002116442, + 0.004005540162324905, + 0.005920328665524721, + 0.010391958989202976, + 0.01248998660594225, + 0.013535356149077415, + 0.017339594662189484, + 0.01392553560435772, + 0.009864290244877338, + 0.012373500503599644, + 0.01315334439277649, + 0.004608495626598597, + 0.004235443659126759, + 0.007434698287397623, + 0.005908130668103695, + 0.007860714569687843, + 0.00973849929869175, + 0.0055449786596000195, + 0.006493960507214069, + 0.0033083336893469095, + 0.004700115416198969, + 0.005280875600874424, + 0.0043598562479019165, + 0.004010817501693964, + 0.004364873748272657, + 0.0037826523184776306, + 0.01050135400146246, + 0.01643456518650055, + 0.028207262977957726, + 0.022237472236156464, + 0.030098438262939453, + 0.03088696114718914, + 0.0171044934540987, + 0.018329735845327377, + 0.02924584038555622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.030442269518971443 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16042080521583557 + ], + "opposite_logprob_tokens": [ + -1.9104207754135132 + ], + "logprob_reward": -0.16042080521583557, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16042080521583557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1785760521888733, + 0.07888912409543991, + 0.06191342696547508, + 0.056417543441057205, + 0.095554880797863, + 0.04886074364185333, + 0.0525517463684082, + 0.04170459881424904, + 0.06442692130804062, + 0.3211049735546112 + ], + "credit_map": [ + 0.016513321548700333, + 0.014952069148421288, + 0.03684011474251747, + 0.028859129175543785, + 0.019715143367648125, + 0.011981822550296783, + 0.026979275047779083, + 0.02273518033325672, + 0.011753097176551819, + 0.007206130772829056, + 0.010642876848578453, + 0.014774300158023834, + 0.008284146897494793, + 0.006316975690424442, + 0.012305209413170815, + 0.007606389932334423, + 0.005201667547225952, + 0.007528130896389484, + 0.007893534377217293, + 0.008448147214949131, + 0.009231504052877426, + 0.004428978078067303, + 0.014368262141942978, + 0.004813208244740963, + 0.005925585050135851, + 0.006333843804895878, + 0.00942554883658886, + 0.011232648976147175, + 0.005713961087167263, + 0.005187662318348885, + 0.006860082969069481, + 0.005738215986639261, + 0.022924035787582397, + 0.008421213366091251, + 0.011727576144039631, + 0.010159183293581009, + 0.006884540896862745, + 0.009469677694141865, + 0.010178286582231522, + 0.01579037681221962, + 0.004535994492471218, + 0.006975851953029633, + 0.005312909837812185, + 0.008818065747618675, + 0.005037473049014807, + 0.007793341297656298, + 0.004576613195240498, + 0.005810498725622892, + 0.006403747946023941, + 0.0065675657242536545, + 0.007423833943903446, + 0.0071244193241000175, + 0.006319474894553423, + 0.006386459339410067, + 0.006404716521501541, + 0.005921528674662113, + 0.003906984813511372, + 0.0035777436569333076, + 0.006076409015804529, + 0.004634303506463766, + 0.00613179849460721, + 0.006706480868160725, + 0.005009854212403297, + 0.005661028437316418, + 0.005813759285956621, + 0.006956914439797401, + 0.013957373797893524, + 0.007122865878045559, + 0.00676892651244998, + 0.010320969857275486, + 0.005625090096145868, + 0.00786102470010519, + 0.019840072840452194, + 0.02037949487566948, + 0.027281902730464935, + 0.01813824102282524, + 0.09963756799697876, + 0.025866597890853882, + 0.01568642072379589, + 0.09427468478679657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.16042080521583557 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640285581350327 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640285581350327, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640285581350327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18154707551002502, + 0.06285295635461807, + 0.10206333547830582, + 0.03940250724554062, + 0.06179823726415634, + 0.06194350868463516, + 0.11934632062911987, + 0.07485086470842361, + 0.05626516416668892, + 0.23993004858493805 + ], + "credit_map": [ + 0.016955973580479622, + 0.01748618297278881, + 0.037759967148303986, + 0.026784729212522507, + 0.019594833254814148, + 0.012280753813683987, + 0.028197649866342545, + 0.022486986592411995, + 0.009154802188277245, + 0.00562660489231348, + 0.009296286851167679, + 0.008080788888037205, + 0.005754157435148954, + 0.00790201686322689, + 0.010960386134684086, + 0.006077917758375406, + 0.0067214579321444035, + 0.011489964090287685, + 0.01361107174307108, + 0.01791209541261196, + 0.01283371914178133, + 0.007192331366240978, + 0.023387815803289413, + 0.008914876729249954, + 0.004752620123326778, + 0.00315466639585793, + 0.009398147463798523, + 0.008537814021110535, + 0.002821052446961403, + 0.004042841028422117, + 0.0036538452841341496, + 0.0030415209475904703, + 0.005809412803500891, + 0.008585666306316853, + 0.005047795828431845, + 0.006263521034270525, + 0.007310125511139631, + 0.009540785104036331, + 0.010973100550472736, + 0.00826783012598753, + 0.004944893531501293, + 0.013074164278805256, + 0.004725564271211624, + 0.011226403526961803, + 0.005435266997665167, + 0.009715243242681026, + 0.00521025201305747, + 0.007611721754074097, + 0.014427521266043186, + 0.015689903870224953, + 0.013392123393714428, + 0.020506953820586205, + 0.015977246686816216, + 0.010679042898118496, + 0.013960331678390503, + 0.014713195152580738, + 0.007489297538995743, + 0.005985509138554335, + 0.007852981798350811, + 0.005588390398770571, + 0.018538113683462143, + 0.007603990845382214, + 0.006045157555490732, + 0.015747422352433205, + 0.004829646088182926, + 0.004809134639799595, + 0.007078647613525391, + 0.011707084253430367, + 0.008388746529817581, + 0.00468128127977252, + 0.004324088804423809, + 0.010446535423398018, + 0.014582096599042416, + 0.015211338177323341, + 0.019908087328076363, + 0.016252851113677025, + 0.07126495242118835, + 0.018589086830615997, + 0.013225401751697063, + 0.07089624553918839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.008640285581350327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305826127529144 + ], + "opposite_logprob_tokens": [ + -3.168058156967163 + ], + "logprob_reward": -0.04305826127529144, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305826127529144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.142934650182724, + 0.14492370188236237, + 0.06183759868144989, + 0.07932128012180328, + 0.12017188221216202, + 0.09260125458240509, + 0.05832310765981674, + 0.04657783359289169, + 0.15113137662410736, + 0.10217732191085815 + ], + "credit_map": [ + 0.010831148363649845, + 0.015274353325366974, + 0.025336498394608498, + 0.01971515454351902, + 0.014237780123949051, + 0.011994613334536552, + 0.025208856910467148, + 0.020336246117949486, + 0.011415932327508926, + 0.00963112898170948, + 0.011956525966525078, + 0.021611452102661133, + 0.016126787289977074, + 0.031696002930402756, + 0.017797140404582024, + 0.024688737466931343, + 0.005506998393684626, + 0.007559238467365503, + 0.007358667906373739, + 0.011173103004693985, + 0.0075474688783288, + 0.00421015452593565, + 0.012702967040240765, + 0.005778994411230087, + 0.008262557908892632, + 0.007321523502469063, + 0.015010037459433079, + 0.012824632227420807, + 0.005698941648006439, + 0.010035662911832333, + 0.012263377197086811, + 0.007904544472694397, + 0.02387103997170925, + 0.014318069443106651, + 0.012343604117631912, + 0.01345749106258154, + 0.009714551270008087, + 0.013394467532634735, + 0.0178519319742918, + 0.015220723114907742, + 0.011087499558925629, + 0.01055356115102768, + 0.013838253915309906, + 0.010099870152771473, + 0.011476285755634308, + 0.011141488328576088, + 0.010994342155754566, + 0.013409952633082867, + 0.005487978924065828, + 0.0052610523998737335, + 0.0077355834655463696, + 0.007943175733089447, + 0.007859470322728157, + 0.007805665023624897, + 0.008864335715770721, + 0.007365846540778875, + 0.003595151240006089, + 0.004534988198429346, + 0.0053702471777796745, + 0.005507063586264849, + 0.008501733653247356, + 0.008642004802823067, + 0.004479076247662306, + 0.005947568453848362, + 0.009794630110263824, + 0.010650559328496456, + 0.020800234749913216, + 0.02466966025531292, + 0.025775548070669174, + 0.0193269532173872, + 0.015984129160642624, + 0.024129673838615417, + 0.009310921654105186, + 0.011460498906672001, + 0.012241368182003498, + 0.01834760047495365, + 0.013589606620371342, + 0.009038898162543774, + 0.01136202085763216, + 0.016826406121253967 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.04305826127529144 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle macro view focusing on sickly, bone-white, waxy, bell-shaped flowers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the flower petals show faint, translucent textures etched into them that tremble with high frequency?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the flower stem seamlessly merge with a blackened, ribbed metal tube having a pitted iron texture and oily sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a thick, viscous, iridescent black liquid gushing from the seam between flesh and metal and dripping downward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow, steady pan back to reveal a vast setting with a ground resembling a massive, carved rib cage?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607118360698223 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.01607118360698223, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607118360698223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09643782675266266, + 0.08321014791727066, + 0.07492037862539291, + 0.0751592367887497, + 0.09004112333059311, + 0.05760042741894722, + 0.13672524690628052, + 0.041371043771505356, + 0.1410522162914276, + 0.20348234474658966 + ], + "credit_map": [ + 0.00941671896725893, + 0.00817457027733326, + 0.02002660557627678, + 0.014472969807684422, + 0.006785073783248663, + 0.005892227403819561, + 0.018961986526846886, + 0.012707688845694065, + 0.01592964120209217, + 0.008521215058863163, + 0.008554267697036266, + 0.010046938434243202, + 0.007447735872119665, + 0.015259483829140663, + 0.008894314058125019, + 0.008556547574698925, + 0.005224731285125017, + 0.008566586300730705, + 0.011358930729329586, + 0.011669907718896866, + 0.009679477661848068, + 0.0037753612268716097, + 0.018112104386091232, + 0.00653328001499176, + 0.006695033982396126, + 0.008005939424037933, + 0.01631208136677742, + 0.014986897818744183, + 0.0073149665258824825, + 0.0053688823245465755, + 0.008633828721940517, + 0.007841606624424458, + 0.010320616886019707, + 0.012212431989610195, + 0.010332872159779072, + 0.011572854593396187, + 0.010251744650304317, + 0.013735808432102203, + 0.01237828191369772, + 0.009236516430974007, + 0.005772728007286787, + 0.012924443930387497, + 0.004640767816454172, + 0.009759431704878807, + 0.005152033641934395, + 0.007276184391230345, + 0.0045693074353039265, + 0.007505530025810003, + 0.011270429939031601, + 0.02127736434340477, + 0.01725745014846325, + 0.025369228795170784, + 0.014114807359874249, + 0.013528114184737206, + 0.01694609224796295, + 0.016961771994829178, + 0.0036646323278546333, + 0.0043241530656814575, + 0.0063104634173214436, + 0.00383459753356874, + 0.006102025508880615, + 0.008561435155570507, + 0.00421922979876399, + 0.004354510456323624, + 0.010046711191534996, + 0.009124759584665298, + 0.017701372504234314, + 0.022686054930090904, + 0.018153369426727295, + 0.02275179699063301, + 0.011484313756227493, + 0.0291038416326046, + 0.01172154676169157, + 0.010830334387719631, + 0.01828235760331154, + 0.011751024052500725, + 0.0627998635172844, + 0.017278587445616722, + 0.010214071720838547, + 0.06060457602143288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.01607118360698223 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_8/index_tensor([3924, 3924])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_8/index_tensor([3924, 3924])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b7964a36ae65cb5466fbfbb0993d4aea37a06a6a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_8/index_tensor([3924, 3924])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot follows a wild boar, the camera gently shaking with its bumpy gait, creating a tense and chaotic sense of pursuit. The massive boar sprints forward through the dense jungle at a reckless, clumsy, and furious pace, its coarse, dark brown mane matted with chunks of damp, dark mud and scattered dead leaves. Two thick, curved ivory tusks jut fiercely from either side of its muddy muzzle, its small, black eyes wide with utter terror. Dappled sunlight filters through the dense canopy, casting shifting patches of light and shadow across its moving form and the forest floor. Its heavy hooves splatter mud and debris on the slippery, decaying leaf-strewn ground, while its powerful shoulders and head shove aside low-hanging branches and brush that block its path. The whole picture is full of primitive dynamics and a tense atmosphere of escape. The main colors are composed of dark green and earthy brown, creating a damp and wild feeling.", + "seed": 10525, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_8/index_3924_seed_10525_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04042515372857451, + "vlm_logprob": -0.04042515372857451, + "vlm_yes_no_margin": 3.899999976158142, + "vlm_token_credit": -0.04042515372857451, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024487411603331566 + ], + "opposite_logprob_tokens": [ + -3.7744874954223633 + ], + "logprob_reward": -0.024487411603331566, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.024487411603331566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1417190432548523, + 0.06553208827972412, + 0.07251693308353424, + 0.11778073012828827, + 0.0541255846619606, + 0.13060754537582397, + 0.03714165836572647, + 0.06530389189720154, + 0.05322788283228874, + 0.26204466819763184 + ], + "credit_map": [ + 0.01905949041247368, + 0.021274805068969727, + 0.01395486667752266, + 0.02180839329957962, + 0.0199801716953516, + 0.01870035007596016, + 0.009895752184092999, + 0.017045224085450172, + 0.008098195306956768, + 0.012325110845267773, + 0.005306866019964218, + 0.005826449487358332, + 0.008777542971074581, + 0.01535390317440033, + 0.004374274518340826, + 0.005469748750329018, + 0.007523233070969582, + 0.007950641214847565, + 0.005874182563275099, + 0.007013930939137936, + 0.011197767220437527, + 0.008535297587513924, + 0.014268900267779827, + 0.010152986273169518, + 0.018829435110092163, + 0.024492064490914345, + 0.018494093790650368, + 0.01922699622809887, + 0.012721175327897072, + 0.008243477903306484, + 0.008421814069151878, + 0.007351670414209366, + 0.006130700930953026, + 0.008380296640098095, + 0.004748316016048193, + 0.007372891530394554, + 0.006095655262470245, + 0.009380408562719822, + 0.005840233992785215, + 0.006177080795168877, + 0.010343054309487343, + 0.01007455587387085, + 0.012085618451237679, + 0.02769509330391884, + 0.015107234008610249, + 0.02701135165989399, + 0.016617778688669205, + 0.011672860942780972, + 0.003708174452185631, + 0.004614610690623522, + 0.0027656774036586285, + 0.004817903507500887, + 0.005570025648921728, + 0.005080081056803465, + 0.00603993097320199, + 0.0045452541671693325, + 0.012018593028187752, + 0.004958035424351692, + 0.012241095304489136, + 0.008990931324660778, + 0.009085032157599926, + 0.004932330921292305, + 0.007662850897759199, + 0.005415030289441347, + 0.004139892291277647, + 0.005321579519659281, + 0.00371082522906363, + 0.005451006814837456, + 0.012373602949082851, + 0.01095611322671175, + 0.007075830828398466, + 0.00419903127476573, + 0.015146455727517605, + 0.02392277680337429, + 0.02157607674598694, + 0.03406665474176407, + 0.03593471273779869, + 0.064326710999012, + 0.042621199041604996, + 0.02445005625486374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.024487411603331566 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018202288076281548 + ], + "opposite_logprob_tokens": [ + -4.018202304840088 + ], + "logprob_reward": -0.018202288076281548, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018202288076281548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1321382373571396, + 0.1232035905122757, + 0.04690488800406456, + 0.12328264862298965, + 0.09134487807750702, + 0.07147860527038574, + 0.040459513664245605, + 0.06423715502023697, + 0.07601214200258255, + 0.2309383749961853 + ], + "credit_map": [ + 0.01663910038769245, + 0.01868426613509655, + 0.010668418370187283, + 0.026121629402041435, + 0.013444237411022186, + 0.020995939150452614, + 0.013087164610624313, + 0.012497473508119583, + 0.018155859783291817, + 0.025254186242818832, + 0.01622701808810234, + 0.013452005572617054, + 0.021984215825796127, + 0.012443672865629196, + 0.00669419951736927, + 0.008992433547973633, + 0.006341404281556606, + 0.006296343170106411, + 0.006566681899130344, + 0.004152029752731323, + 0.005917532369494438, + 0.004305824637413025, + 0.00771040515974164, + 0.005614668596535921, + 0.01874079741537571, + 0.029547015205025673, + 0.017230359837412834, + 0.01612177863717079, + 0.015561128966510296, + 0.011491937562823296, + 0.006682977080345154, + 0.007906662300229073, + 0.011892215348780155, + 0.011167839169502258, + 0.009216617792844772, + 0.011364713311195374, + 0.00965135544538498, + 0.014568286016583443, + 0.012404825538396835, + 0.011079027317464352, + 0.008669043891131878, + 0.01061695721000433, + 0.006470080930739641, + 0.008507946506142616, + 0.009596320800483227, + 0.006589790806174278, + 0.00796928908675909, + 0.013059171847999096, + 0.003386343829333782, + 0.0047814082354307175, + 0.004630456678569317, + 0.004345706198364496, + 0.006794567685574293, + 0.0047541395761072636, + 0.007199393119663, + 0.004567496478557587, + 0.013013857416808605, + 0.005969068966805935, + 0.007644480559974909, + 0.009475216269493103, + 0.011379935778677464, + 0.004130483604967594, + 0.008235746063292027, + 0.004388360772281885, + 0.005648417863994837, + 0.007383130490779877, + 0.009268106892704964, + 0.006209036335349083, + 0.017060013487935066, + 0.015302862040698528, + 0.00997097697108984, + 0.005169598385691643, + 0.015576719306409359, + 0.02751108631491661, + 0.01975497230887413, + 0.028942544013261795, + 0.0316409096121788, + 0.05604073777794838, + 0.03237864002585411, + 0.019092747941613197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.018202288076281548 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1427968442440033 + ], + "opposite_logprob_tokens": [ + -2.017796754837036 + ], + "logprob_reward": -0.1427968442440033, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1427968442440033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10882992297410965, + 0.140666201710701, + 0.16820591688156128, + 0.0328204408288002, + 0.048443764448165894, + 0.04690780118107796, + 0.07850252836942673, + 0.07900727540254593, + 0.06594834476709366, + 0.23066778481006622 + ], + "credit_map": [ + 0.013855756260454655, + 0.014284322038292885, + 0.009525485336780548, + 0.011433136649429798, + 0.013999122194945812, + 0.020121004432439804, + 0.01174606941640377, + 0.013865025714039803, + 0.015469897538423538, + 0.0407392717897892, + 0.007519394624978304, + 0.010191812179982662, + 0.02800578624010086, + 0.01216913852840662, + 0.013132349587976933, + 0.013438543304800987, + 0.017177393659949303, + 0.020023763179779053, + 0.014973008073866367, + 0.013555769808590412, + 0.023732388392090797, + 0.022089842706918716, + 0.03010335937142372, + 0.02655038610100746, + 0.004061548504978418, + 0.005889049265533686, + 0.0056564779952168465, + 0.004910270683467388, + 0.003603806719183922, + 0.003877077018842101, + 0.0019218220841139555, + 0.0029003871604800224, + 0.0073039778508245945, + 0.006698390934616327, + 0.003923843149095774, + 0.007460984401404858, + 0.005853492300957441, + 0.006360882427543402, + 0.006407433655112982, + 0.004434758331626654, + 0.004679023753851652, + 0.004698529839515686, + 0.0055953324772417545, + 0.00942196324467659, + 0.004024815279990435, + 0.008078785613179207, + 0.004189638886600733, + 0.006219713017344475, + 0.009702973999083042, + 0.009785125963389874, + 0.01192691270262003, + 0.00544026168063283, + 0.010197548195719719, + 0.012480603531002998, + 0.01068513561040163, + 0.008283966220915318, + 0.016502248123288155, + 0.006822189316153526, + 0.014728065580129623, + 0.011405151337385178, + 0.006937437690794468, + 0.0053670695051550865, + 0.010198273696005344, + 0.007046838756650686, + 0.00545540964230895, + 0.010198447853326797, + 0.008584362454712391, + 0.00698950095102191, + 0.012746084481477737, + 0.008480988442897797, + 0.010018477216362953, + 0.0034750772174447775, + 0.015158177353441715, + 0.024297021329402924, + 0.021679386496543884, + 0.031671930104494095, + 0.030761877074837685, + 0.04060832038521767, + 0.043462835252285004, + 0.023028235882520676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.1427968442440033 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0125422989949584 + ], + "opposite_logprob_tokens": [ + -4.387542247772217 + ], + "logprob_reward": -0.0125422989949584, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0125422989949584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11224190890789032, + 0.11558327078819275, + 0.11046864837408066, + 0.07906124740839005, + 0.10219085216522217, + 0.05924783647060394, + 0.07664413005113602, + 0.06578565388917923, + 0.15895894169807434, + 0.11981754750013351 + ], + "credit_map": [ + 0.015077788382768631, + 0.01440292689949274, + 0.009643711149692535, + 0.018120063468813896, + 0.011373095214366913, + 0.01818668656051159, + 0.009209154173731804, + 0.016228482127189636, + 0.010995505377650261, + 0.0348486453294754, + 0.004922227468341589, + 0.007967200130224228, + 0.031290385872125626, + 0.008289650082588196, + 0.009291096590459347, + 0.007978551089763641, + 0.011114359833300114, + 0.012858572416007519, + 0.016447965055704117, + 0.009811429306864738, + 0.012486249208450317, + 0.015564042143523693, + 0.01703161559998989, + 0.015154408290982246, + 0.013375606387853622, + 0.020013196393847466, + 0.00941650290042162, + 0.007434138096868992, + 0.01066728588193655, + 0.007017479278147221, + 0.00546483276411891, + 0.005672199185937643, + 0.014676893129944801, + 0.016057180240750313, + 0.009765584021806717, + 0.016639402136206627, + 0.014919955283403397, + 0.008593329228460789, + 0.01300677191466093, + 0.00853173527866602, + 0.006510872859507799, + 0.005484418477863073, + 0.004544728901237249, + 0.015545145608484745, + 0.004578414373099804, + 0.010013493709266186, + 0.004547651391476393, + 0.008023113012313843, + 0.008851844817399979, + 0.010560124181210995, + 0.006448626518249512, + 0.0050107152201235294, + 0.012168888933956623, + 0.01905229687690735, + 0.008763578720390797, + 0.0057880571112036705, + 0.006493335124105215, + 0.007956091314554214, + 0.006358844228088856, + 0.012144859880208969, + 0.008859455585479736, + 0.010159728117287159, + 0.008233862929046154, + 0.005579476710408926, + 0.01644718460738659, + 0.026741934940218925, + 0.01660195179283619, + 0.017330430448055267, + 0.026023557409644127, + 0.024194767698645592, + 0.01969202607870102, + 0.011927084997296333, + 0.0076548196375370026, + 0.013731757178902626, + 0.011824551969766617, + 0.017763987183570862, + 0.014458570629358292, + 0.026203911751508713, + 0.018060335889458656, + 0.010119621641933918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.0125422989949584 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096925724297762 + ], + "opposite_logprob_tokens": [ + -5.504096984863281 + ], + "logprob_reward": -0.004096925724297762, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096925724297762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3978694975376129, + 0.04203987866640091, + 0.1713280975818634, + 0.04849192127585411, + 0.05804764851927757, + 0.05718422681093216, + 0.03448657691478729, + 0.059814561158418655, + 0.06629329919815063, + 0.06444428861141205 + ], + "credit_map": [ + 0.050552066415548325, + 0.04869425669312477, + 0.06382731348276138, + 0.06104838103055954, + 0.04797833785414696, + 0.043991025537252426, + 0.03717255964875221, + 0.044605545699596405, + 0.004909418523311615, + 0.007130276877433062, + 0.0031775294337421656, + 0.0035842338111251593, + 0.006012621335685253, + 0.005972292274236679, + 0.004959560465067625, + 0.006293949205428362, + 0.02045542560517788, + 0.018213242292404175, + 0.015533367171883583, + 0.011906281113624573, + 0.02269015647470951, + 0.028297599405050278, + 0.027085572481155396, + 0.027146467939019203, + 0.004816824104636908, + 0.0063794804736971855, + 0.00533876521512866, + 0.005002152640372515, + 0.0052837589755654335, + 0.01116164494305849, + 0.004595902282744646, + 0.00591339822858572, + 0.007745709735900164, + 0.01028345339000225, + 0.006084188353270292, + 0.008829469792544842, + 0.007496291771531105, + 0.006667186971753836, + 0.005617313086986542, + 0.00532403914257884, + 0.005459378939121962, + 0.004155081696808338, + 0.006383317522704601, + 0.013079291209578514, + 0.005978309083729982, + 0.01186178158968687, + 0.004440078046172857, + 0.005826995242387056, + 0.0029248446226119995, + 0.004008736461400986, + 0.0025303130969405174, + 0.00318085472099483, + 0.0067828018218278885, + 0.00716965040192008, + 0.004607881885021925, + 0.003281497163698077, + 0.011470054276287556, + 0.006233168300241232, + 0.009677795693278313, + 0.00824957899749279, + 0.005761224776506424, + 0.005581412930041552, + 0.0071447137743234634, + 0.005696614272892475, + 0.007408243138343096, + 0.008778606541454792, + 0.011130541563034058, + 0.009671274572610855, + 0.007791528943926096, + 0.006567896343767643, + 0.009102755226194859, + 0.005842456594109535, + 0.004378408193588257, + 0.006322778761386871, + 0.005847497843205929, + 0.008653664961457253, + 0.011148885823786259, + 0.011267904192209244, + 0.010026047006249428, + 0.006799103692173958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.004096925724297762 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a wild boar, the camera gently shaking with its bumpy gait, creating a tense and chaotic sense of pursuit. The massive boar sprints forward through the dense jungle at a reckless, clumsy, and furious pace, its coarse, dark brown mane matted with chunks of damp, dark mud and scattered dead leaves. Two thick, curved ivory tusks jut fiercely from either side of its muddy muzzle, its small, black eyes wide with utter terror. Dappled sunlight filters through the dense canopy, casting shifting patches of light and shadow across its moving form and the forest floor. Its heavy hooves splatter mud and debris on the slippery, decaying leaf-strewn ground, while its powerful shoulders and head shove aside low-hanging branches and brush that block its path. The whole picture is full of primitive dynamics and a tense atmosphere of escape. The main colors are composed of dark green and earthy brown, creating a damp and wild feeling.", + "seed": 10526, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_8/index_3924_seed_10526_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.025732660479843617, + "vlm_logprob": -0.025732660479843617, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.025732660479843617, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013735185377299786 + ], + "opposite_logprob_tokens": [ + -4.388735294342041 + ], + "logprob_reward": -0.013735185377299786, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.013735185377299786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1951402723789215, + 0.10816281288862228, + 0.07410041242837906, + 0.19127307832241058, + 0.07163641601800919, + 0.03731180727481842, + 0.045554086565971375, + 0.05819052457809448, + 0.09232984483242035, + 0.12630072236061096 + ], + "credit_map": [ + 0.02184380404651165, + 0.04259645938873291, + 0.024214569479227066, + 0.030409449711441994, + 0.023968437686562538, + 0.01956113986670971, + 0.01733425445854664, + 0.015212144702672958, + 0.011535254307091236, + 0.02480018325150013, + 0.00912062730640173, + 0.019116785377264023, + 0.010838939808309078, + 0.01782223768532276, + 0.005920341704040766, + 0.00900843646377325, + 0.005649538245052099, + 0.007270997855812311, + 0.005294677335768938, + 0.005587609019130468, + 0.0057092285715043545, + 0.014976928941905499, + 0.009871851652860641, + 0.019739581272006035, + 0.02657533437013626, + 0.023646056652069092, + 0.01769893616437912, + 0.00912593211978674, + 0.04764421284198761, + 0.022985801100730896, + 0.018798261880874634, + 0.024798555299639702, + 0.006158787291496992, + 0.014622928574681282, + 0.004450043197721243, + 0.00744260661303997, + 0.013982944190502167, + 0.007511673029512167, + 0.007266473025083542, + 0.010200956836342812, + 0.0066857608035206795, + 0.003884240286424756, + 0.003368105972185731, + 0.004766907077282667, + 0.003979100845754147, + 0.0038755740970373154, + 0.005219051614403725, + 0.005533067975193262, + 0.007471134420484304, + 0.0033066673204302788, + 0.007037146482616663, + 0.006140867248177528, + 0.0029757185839116573, + 0.009130407124757767, + 0.004659424535930157, + 0.004832718521356583, + 0.006524518132209778, + 0.008815906941890717, + 0.00782004278153181, + 0.008510705083608627, + 0.005746215116232634, + 0.007199566811323166, + 0.00825569685548544, + 0.005317875184118748, + 0.016929009929299355, + 0.00925502646714449, + 0.009453320875763893, + 0.014915366657078266, + 0.007573308423161507, + 0.0065226140432059765, + 0.015780776739120483, + 0.011900423094630241, + 0.0125764524564147, + 0.008552785962820053, + 0.0363459512591362, + 0.00758037343621254, + 0.00719437375664711, + 0.006314000580459833, + 0.012151059694588184, + 0.03558572381734848 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.013735185377299786 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596086978912354 + ], + "opposite_logprob_tokens": [ + -3.8955960273742676 + ], + "logprob_reward": -0.020596086978912354, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596086978912354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1078047901391983, + 0.10579312592744827, + 0.20379380881786346, + 0.1776721179485321, + 0.030547423288226128, + 0.02954769693315029, + 0.04366878420114517, + 0.13250669836997986, + 0.08950157463550568, + 0.07916408032178879 + ], + "credit_map": [ + 0.012966273352503777, + 0.021759264171123505, + 0.013407386839389801, + 0.013339408673346043, + 0.007825693115592003, + 0.010130496695637703, + 0.015659987926483154, + 0.012716270051896572, + 0.012337942607700825, + 0.015138044953346252, + 0.009962093085050583, + 0.016741564497351646, + 0.019742297008633614, + 0.01701757125556469, + 0.00867738202214241, + 0.006176214199513197, + 0.016588781028985977, + 0.013904492370784283, + 0.021636351943016052, + 0.014880220405757427, + 0.016727378591895103, + 0.03236917033791542, + 0.03595282509922981, + 0.051734551787376404, + 0.024336932227015495, + 0.02032436430454254, + 0.015674619004130363, + 0.008834400214254856, + 0.04714076593518257, + 0.020811671391129494, + 0.016934849321842194, + 0.023614490404725075, + 0.003239582758396864, + 0.003484186949208379, + 0.0024224354419857264, + 0.003432664554566145, + 0.00705175893381238, + 0.003812467446550727, + 0.003973690327256918, + 0.003130633383989334, + 0.0034801282454282045, + 0.0024888559710234404, + 0.0020382332149893045, + 0.0038496647030115128, + 0.0030135144479572773, + 0.0022238946985453367, + 0.007983648218214512, + 0.004469755105674267, + 0.006059571169316769, + 0.0043504498898983, + 0.005063303746283054, + 0.00478760851547122, + 0.0031939002219587564, + 0.011080567725002766, + 0.003531302325427532, + 0.005602075252681971, + 0.011739100329577923, + 0.021092407405376434, + 0.015707259997725487, + 0.018532438203692436, + 0.011859574355185032, + 0.01579339988529682, + 0.02006971463561058, + 0.017712784931063652, + 0.01013005431741476, + 0.011787372641265392, + 0.011314873583614826, + 0.012598246335983276, + 0.007254051510244608, + 0.009646846912801266, + 0.014590393751859665, + 0.012179719284176826, + 0.008871876634657383, + 0.007731970865279436, + 0.020396167412400246, + 0.0055215186439454556, + 0.007997644133865833, + 0.005075301509350538, + 0.0075643835589289665, + 0.016005204990506172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.020596086978912354 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897435128688812 + ], + "opposite_logprob_tokens": [ + -2.5789742469787598 + ], + "logprob_reward": -0.07897435128688812, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897435128688812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0867888480424881, + 0.09351633489131927, + 0.10734757035970688, + 0.4889478385448456, + 0.031008297577500343, + 0.0263529010117054, + 0.021318454295396805, + 0.06063871830701828, + 0.04174097627401352, + 0.042340055108070374 + ], + "credit_map": [ + 0.010206494480371475, + 0.01711972802877426, + 0.008845100179314613, + 0.009920951910316944, + 0.011791589669883251, + 0.007933305576443672, + 0.011103721335530281, + 0.009867968037724495, + 0.011300432495772839, + 0.019399071112275124, + 0.0074887508526444435, + 0.01591419242322445, + 0.013650095090270042, + 0.01616714708507061, + 0.004970288369804621, + 0.0046263630501925945, + 0.008161098696291447, + 0.0075870417058467865, + 0.007777273654937744, + 0.006151714362204075, + 0.006195992697030306, + 0.022439729422330856, + 0.016291501000523567, + 0.03274322673678398, + 0.07294367253780365, + 0.06137467920780182, + 0.04678488150238991, + 0.02380272001028061, + 0.1125975176692009, + 0.06085200607776642, + 0.04650023207068443, + 0.06409212946891785, + 0.002303660847246647, + 0.0063422927632927895, + 0.0017697194125503302, + 0.003221228253096342, + 0.005914832465350628, + 0.0033962191082537174, + 0.0038123242557048798, + 0.0042480239644646645, + 0.006395517382770777, + 0.0024506018962711096, + 0.002080096397548914, + 0.003081914270296693, + 0.003195105353370309, + 0.001891923719085753, + 0.0034756516106426716, + 0.003782094456255436, + 0.0021580206230282784, + 0.0020579129923135042, + 0.0020253516267985106, + 0.002654528943821788, + 0.0014580823481082916, + 0.00524931401014328, + 0.00289919157512486, + 0.0028160535730421543, + 0.008021967485547066, + 0.0060560149140655994, + 0.008506317622959614, + 0.011253389529883862, + 0.004774945788085461, + 0.0073355697095394135, + 0.008649572730064392, + 0.006040951702743769, + 0.006459744181483984, + 0.00402802461758256, + 0.005309939384460449, + 0.007250277791172266, + 0.0034130148123949766, + 0.004024633206427097, + 0.005526587832719088, + 0.005728759802877903, + 0.005615330766886473, + 0.0032030800357460976, + 0.010066287592053413, + 0.003829801455140114, + 0.0024133711121976376, + 0.002063204301521182, + 0.004756493493914604, + 0.010392486117780209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 1 + ], + "selected_reward": -0.07897435128688812 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537471950054169 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537471950054169, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537471950054169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10646293312311172, + 0.11914218962192535, + 0.06159503385424614, + 0.18687006831169128, + 0.05126316472887993, + 0.06596363335847855, + 0.0358235202729702, + 0.08765066415071487, + 0.06606650352478027, + 0.21916228532791138 + ], + "credit_map": [ + 0.013750162906944752, + 0.019141029566526413, + 0.013291539624333382, + 0.014353807084262371, + 0.01067784707993269, + 0.01088719628751278, + 0.012956355698406696, + 0.01140498649328947, + 0.011240724474191666, + 0.01951715722680092, + 0.009583448991179466, + 0.0252170879393816, + 0.015377260744571686, + 0.020346077159047127, + 0.006923987064510584, + 0.010936446487903595, + 0.0077254981733858585, + 0.007919042371213436, + 0.007569293957203627, + 0.0062896558083593845, + 0.00715889036655426, + 0.009046253748238087, + 0.007799128070473671, + 0.008087273687124252, + 0.015094264410436153, + 0.015541940927505493, + 0.01618247479200363, + 0.014230195432901382, + 0.052205294370651245, + 0.03009839542210102, + 0.01927088387310505, + 0.024246614426374435, + 0.0051902662962675095, + 0.006239221431314945, + 0.004152878187596798, + 0.006118627730756998, + 0.010415884666144848, + 0.0065240259282290936, + 0.007997401058673859, + 0.004624855704605579, + 0.008066813461482525, + 0.005853277165442705, + 0.008350122720003128, + 0.011182966642081738, + 0.005714945029467344, + 0.005303000565618277, + 0.01311628706753254, + 0.008376218378543854, + 0.005022732075303793, + 0.0029319196473807096, + 0.005135887768119574, + 0.0042928894981741905, + 0.0028862047474831343, + 0.004974657204002142, + 0.006450679153203964, + 0.004128552973270416, + 0.013051625341176987, + 0.01042146235704422, + 0.010590024292469025, + 0.01205489132553339, + 0.007707476150244474, + 0.009565535932779312, + 0.015641896054148674, + 0.00861775130033493, + 0.01141392719000578, + 0.00586262671276927, + 0.006710614077746868, + 0.008352921344339848, + 0.004730085842311382, + 0.0070858788676559925, + 0.010325809009373188, + 0.011584642343223095, + 0.0228127408772707, + 0.01692821830511093, + 0.06015825271606445, + 0.013946819119155407, + 0.022629475221037865, + 0.013176601380109787, + 0.02020033821463585, + 0.049309827387332916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.012537471950054169 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820206806063652 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.002820206806063652, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820206806063652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27841803431510925, + 0.09066690504550934, + 0.08802325278520584, + 0.08061768114566803, + 0.034162748605012894, + 0.047839127480983734, + 0.05808231607079506, + 0.1256290227174759, + 0.06821932643651962, + 0.12834160029888153 + ], + "credit_map": [ + 0.028760861605405807, + 0.03990444168448448, + 0.03893674910068512, + 0.03266137093305588, + 0.030285267159342766, + 0.027065424248576164, + 0.042542681097984314, + 0.03826124593615532, + 0.008634114637970924, + 0.013375505805015564, + 0.010519216768443584, + 0.012164508923888206, + 0.00989212654531002, + 0.011692720465362072, + 0.0076800016686320305, + 0.01670871488749981, + 0.00886642374098301, + 0.01312493160367012, + 0.011205831542611122, + 0.005713955033570528, + 0.016373613849282265, + 0.012846479192376137, + 0.005901639815419912, + 0.01399038266390562, + 0.007024566642940044, + 0.007049350533634424, + 0.0056147403083741665, + 0.006238161586225033, + 0.024987593293190002, + 0.010163669474422932, + 0.007501876913011074, + 0.012037730775773525, + 0.004415481351315975, + 0.006371011026203632, + 0.002328439150005579, + 0.0028367042541503906, + 0.004174416419118643, + 0.004643763415515423, + 0.005048801656812429, + 0.004344130866229534, + 0.005686044692993164, + 0.004261988215148449, + 0.005562333390116692, + 0.006326564587652683, + 0.005673895590007305, + 0.0046999710611999035, + 0.009439815767109394, + 0.006188519764691591, + 0.004656159318983555, + 0.005716080777347088, + 0.004838523920625448, + 0.007855587638914585, + 0.00471497280523181, + 0.016536077484488487, + 0.006028153467923403, + 0.007736758794635534, + 0.01617356948554516, + 0.01227081473916769, + 0.020804917439818382, + 0.02513311244547367, + 0.008339909836649895, + 0.015930358320474625, + 0.01550016924738884, + 0.011476175859570503, + 0.008440623991191387, + 0.007959336042404175, + 0.006784271448850632, + 0.01363189984112978, + 0.007713329046964645, + 0.005904295016080141, + 0.00866063591092825, + 0.009124945849180222, + 0.012623175047338009, + 0.011568508110940456, + 0.02868315577507019, + 0.011774004437029362, + 0.013149753212928772, + 0.011805730871856213, + 0.017106885090470314, + 0.021630389615893364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.002820206806063652 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a wild boar, the camera gently shaking with its bumpy gait, creating a tense and chaotic sense of pursuit. The massive boar sprints forward through the dense jungle at a reckless, clumsy, and furious pace, its coarse, dark brown mane matted with chunks of damp, dark mud and scattered dead leaves. Two thick, curved ivory tusks jut fiercely from either side of its muddy muzzle, its small, black eyes wide with utter terror. Dappled sunlight filters through the dense canopy, casting shifting patches of light and shadow across its moving form and the forest floor. Its heavy hooves splatter mud and debris on the slippery, decaying leaf-strewn ground, while its powerful shoulders and head shove aside low-hanging branches and brush that block its path. The whole picture is full of primitive dynamics and a tense atmosphere of escape. The main colors are composed of dark green and earthy brown, creating a damp and wild feeling.", + "seed": 10527, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_8/index_3924_seed_10527_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027521588653326035, + "vlm_logprob": -0.027521588653326035, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.027521588653326035, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023095306009054184 + ], + "opposite_logprob_tokens": [ + -3.8980953693389893 + ], + "logprob_reward": -0.023095306009054184, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.023095306009054184, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21487835049629211, + 0.08325055241584778, + 0.060189150273799896, + 0.12809668481349945, + 0.09383843094110489, + 0.07101664692163467, + 0.18619577586650848, + 0.05818350985646248, + 0.048373982310295105, + 0.05597696453332901 + ], + "credit_map": [ + 0.022778784856200218, + 0.02815113216638565, + 0.024238986894488335, + 0.03732771798968315, + 0.029721705242991447, + 0.02193419449031353, + 0.026003627106547356, + 0.024722209200263023, + 0.006812970619648695, + 0.010741189122200012, + 0.009441952221095562, + 0.01166129857301712, + 0.010943148285150528, + 0.009981264360249043, + 0.013085165992379189, + 0.01058356836438179, + 0.0072033400647342205, + 0.007104307878762484, + 0.00876912847161293, + 0.012785028666257858, + 0.0056233517825603485, + 0.007713022641837597, + 0.005793324206024408, + 0.00519764656201005, + 0.012694635428488255, + 0.017083944752812386, + 0.01061057299375534, + 0.02020539715886116, + 0.018190203234553337, + 0.014758097939193249, + 0.017237018793821335, + 0.01731681264936924, + 0.006812635809183121, + 0.011487440206110477, + 0.01711476594209671, + 0.016036974266171455, + 0.01105216983705759, + 0.010121126659214497, + 0.011573020368814468, + 0.009640300646424294, + 0.005533386487513781, + 0.011179206892848015, + 0.0062142545357346535, + 0.005119557026773691, + 0.010065252892673016, + 0.011414634995162487, + 0.010798369534313679, + 0.010691985487937927, + 0.022965220734477043, + 0.028051942586898804, + 0.02048345096409321, + 0.023824134841561317, + 0.022426513954997063, + 0.02243245206773281, + 0.02597799152135849, + 0.020034076645970345, + 0.006518466863781214, + 0.008490352891385555, + 0.0077693527564406395, + 0.00769353611394763, + 0.00705817062407732, + 0.007826766930520535, + 0.007261775899678469, + 0.005565086845308542, + 0.004734749440103769, + 0.007306382060050964, + 0.006878592539578676, + 0.006966036278754473, + 0.008804908022284508, + 0.005401823204010725, + 0.004770193714648485, + 0.0035112982150167227, + 0.007903221063315868, + 0.009659984149038792, + 0.007737245410680771, + 0.005122282076627016, + 0.007281850092113018, + 0.005281354300677776, + 0.006280907429754734, + 0.00671011907979846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.023095306009054184 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308223113417625 + ], + "opposite_logprob_tokens": [ + -3.773308277130127 + ], + "logprob_reward": -0.023308223113417625, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308223113417625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16989202797412872, + 0.09486132115125656, + 0.06522917002439499, + 0.16098955273628235, + 0.13319551944732666, + 0.06814555078744888, + 0.06403359770774841, + 0.06995785236358643, + 0.0722096711397171, + 0.10148579627275467 + ], + "credit_map": [ + 0.016603469848632812, + 0.02128165028989315, + 0.03559708222746849, + 0.02165084145963192, + 0.018444756045937538, + 0.016557464376091957, + 0.017287813127040863, + 0.022468943148851395, + 0.006738102529197931, + 0.011647449806332588, + 0.010739308781921864, + 0.016746504232287407, + 0.014139449223876, + 0.011778496205806732, + 0.010854744352400303, + 0.012217250652611256, + 0.004531677812337875, + 0.008204269222915173, + 0.005397377535700798, + 0.017459748312830925, + 0.011266447603702545, + 0.006908458191901445, + 0.005695741157978773, + 0.005765441805124283, + 0.018725354224443436, + 0.021997995674610138, + 0.021643640473484993, + 0.019367700442671776, + 0.022500935941934586, + 0.018364911898970604, + 0.018605254590511322, + 0.019783733412623405, + 0.010887865908443928, + 0.01696622371673584, + 0.020561158657073975, + 0.014772878028452396, + 0.018518496304750443, + 0.02173786610364914, + 0.016286838799715042, + 0.013464180752635002, + 0.004884554073214531, + 0.013141410425305367, + 0.005712166894227266, + 0.0066408258862793446, + 0.00782804936170578, + 0.009373040869832039, + 0.01236764620989561, + 0.008197847753763199, + 0.008886301890015602, + 0.008993136696517467, + 0.00594252347946167, + 0.0067618535831570625, + 0.009177064523100853, + 0.007532482501119375, + 0.008839321322739124, + 0.007900902070105076, + 0.011652507819235325, + 0.00788977462798357, + 0.012371190823614597, + 0.006480023264884949, + 0.007220024708658457, + 0.011928205378353596, + 0.006801914423704147, + 0.005614199675619602, + 0.006887662224471569, + 0.013589861802756786, + 0.01095843780785799, + 0.010546725243330002, + 0.009969103150069714, + 0.006233071908354759, + 0.008454078808426857, + 0.0055707283318042755, + 0.009795360267162323, + 0.014850569888949394, + 0.01478035468608141, + 0.011178423650562763, + 0.024354971945285797, + 0.010794063098728657, + 0.008238712325692177, + 0.0074933297000825405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.023308223113417625 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001993060112 + ], + "opposite_logprob_tokens": [ + -2.6950199604034424 + ], + "logprob_reward": -0.07001993060112, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001993060112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17178094387054443, + 0.0715157613158226, + 0.06788051128387451, + 0.2352234125137329, + 0.08864398300647736, + 0.048953693360090256, + 0.08100482821464539, + 0.10433986037969589, + 0.07296647876501083, + 0.05769054964184761 + ], + "credit_map": [ + 0.01820945367217064, + 0.021168099716305733, + 0.016490530222654343, + 0.027799125760793686, + 0.02901093102991581, + 0.020204326137900352, + 0.02024873159825802, + 0.018649762496352196, + 0.005398713517934084, + 0.00697983568534255, + 0.009634355083107948, + 0.012317328713834286, + 0.011438668705523014, + 0.010760936886072159, + 0.009035165421664715, + 0.005950751714408398, + 0.006832113955169916, + 0.008049482479691505, + 0.007884801365435123, + 0.01144537702202797, + 0.010478739626705647, + 0.00962129794061184, + 0.008101086132228374, + 0.005467615090310574, + 0.01940903812646866, + 0.038700345903635025, + 0.020370541140437126, + 0.03562870994210243, + 0.033322256058454514, + 0.025168877094984055, + 0.03616943955421448, + 0.026454204693436623, + 0.005747323855757713, + 0.009107671678066254, + 0.006435109302401543, + 0.013517368584871292, + 0.015624230727553368, + 0.0168712567538023, + 0.012319682165980339, + 0.009021341800689697, + 0.005244718864560127, + 0.01234948355704546, + 0.0029985469300299883, + 0.005655629560351372, + 0.005620949435979128, + 0.004842002410441637, + 0.0041458564810454845, + 0.008096504956483841, + 0.009342106990516186, + 0.008642229251563549, + 0.006387181580066681, + 0.0074889641255140305, + 0.011360579170286655, + 0.011418682523071766, + 0.013764605857431889, + 0.012600479647517204, + 0.007853865623474121, + 0.014395849779248238, + 0.008535378612577915, + 0.012706086039543152, + 0.01595260761678219, + 0.014054225757718086, + 0.021692674607038498, + 0.009149172343313694, + 0.005452694371342659, + 0.013834604062139988, + 0.010285856202244759, + 0.008343104273080826, + 0.013704794459044933, + 0.011043280363082886, + 0.005389302968978882, + 0.004912838339805603, + 0.005181091837584972, + 0.010115502402186394, + 0.008143537677824497, + 0.007171399425715208, + 0.010219162330031395, + 0.005569890141487122, + 0.007239192258566618, + 0.004050774499773979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.07001993060112 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540415860712528 + ], + "opposite_logprob_tokens": [ + -4.387540340423584 + ], + "logprob_reward": -0.012540415860712528, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540415860712528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15291409194469452, + 0.07719922065734863, + 0.033291470259428024, + 0.08370472490787506, + 0.08898379653692245, + 0.03953651338815689, + 0.11306364834308624, + 0.12043822556734085, + 0.17614267766475677, + 0.11472560465335846 + ], + "credit_map": [ + 0.017921239137649536, + 0.019136281684041023, + 0.01533794216811657, + 0.023201752454042435, + 0.024625184014439583, + 0.016649313271045685, + 0.018066363409161568, + 0.017976028844714165, + 0.005665519740432501, + 0.009875642135739326, + 0.007484141271561384, + 0.016162483021616936, + 0.01105530932545662, + 0.010052787140011787, + 0.007956892251968384, + 0.008946443907916546, + 0.005686628166586161, + 0.0038110699970275164, + 0.004127254243940115, + 0.004403718281537294, + 0.004275172483175993, + 0.0039029403124004602, + 0.003314114874228835, + 0.0037705700378865004, + 0.011631019413471222, + 0.008822917006909847, + 0.008959616534411907, + 0.008710619993507862, + 0.012586806900799274, + 0.009566710330545902, + 0.011365287005901337, + 0.01206174585968256, + 0.01036897860467434, + 0.00924187246710062, + 0.012822207063436508, + 0.01942974142730236, + 0.0066774757578969, + 0.014009540900588036, + 0.010115403681993484, + 0.006318580824881792, + 0.0035674336832016706, + 0.006006443407386541, + 0.0027200172189623117, + 0.0055701411329209805, + 0.004288710653781891, + 0.005311226937919855, + 0.005702631548047066, + 0.006369905546307564, + 0.011850048787891865, + 0.018551284447312355, + 0.011439002119004726, + 0.012171231210231781, + 0.010771074332296848, + 0.01392569113522768, + 0.016256073489785194, + 0.018099237233400345, + 0.01805782876908779, + 0.015430586412549019, + 0.020044730976223946, + 0.012278957292437553, + 0.01659565605223179, + 0.01230094488710165, + 0.015567784197628498, + 0.010161741636693478, + 0.012636951170861721, + 0.02316427417099476, + 0.02215713821351528, + 0.024983607232570648, + 0.028286956250667572, + 0.011639510281383991, + 0.039730627089738846, + 0.01354362815618515, + 0.01106198225170374, + 0.0177001990377903, + 0.015491646714508533, + 0.013045836240053177, + 0.02923707850277424, + 0.010183199308812618, + 0.008641424588859081, + 0.009364240802824497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.012540415860712528 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864406768232584 + ], + "opposite_logprob_tokens": [ + -4.758644104003906 + ], + "logprob_reward": -0.00864406768232584, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864406768232584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20163588225841522, + 0.06912975758314133, + 0.045655228197574615, + 0.06737609207630157, + 0.09839075803756714, + 0.04828494042158127, + 0.09623758494853973, + 0.12167531251907349, + 0.055770810693502426, + 0.1958436518907547 + ], + "credit_map": [ + 0.02826014533638954, + 0.024139609187841415, + 0.040097035467624664, + 0.02545066364109516, + 0.017169833183288574, + 0.025598114356398582, + 0.01451890729367733, + 0.02640155889093876, + 0.006537223234772682, + 0.010583659633994102, + 0.01003063004463911, + 0.011173439212143421, + 0.007110963109880686, + 0.0074296388775110245, + 0.006374489050358534, + 0.00988970696926117, + 0.00399864511564374, + 0.0058610388077795506, + 0.005485699977725744, + 0.007500366773456335, + 0.007701388094574213, + 0.006430626381188631, + 0.004305578768253326, + 0.004371876362711191, + 0.007700161077082157, + 0.008452633395791054, + 0.011107871308922768, + 0.0055722882971167564, + 0.010015291161835194, + 0.007898254320025444, + 0.006395052187144756, + 0.010234539397060871, + 0.01083506178110838, + 0.010289470665156841, + 0.016869179904460907, + 0.023211084306240082, + 0.00822523981332779, + 0.01044501643627882, + 0.012384938076138496, + 0.0061307684518396854, + 0.003668667282909155, + 0.009533140808343887, + 0.0034705845173448324, + 0.0082459207624197, + 0.005997715052217245, + 0.005396982654929161, + 0.006367729045450687, + 0.005604199133813381, + 0.014129919931292534, + 0.012403283268213272, + 0.010896416381001472, + 0.011307401582598686, + 0.008746241219341755, + 0.01272971834987402, + 0.011759700253605843, + 0.014264898374676704, + 0.011996110901236534, + 0.02306346781551838, + 0.01445864886045456, + 0.012358386069536209, + 0.012219848111271858, + 0.017500776797533035, + 0.017683323472738266, + 0.012394743040204048, + 0.005019202828407288, + 0.007218549959361553, + 0.010718104429543018, + 0.008196944370865822, + 0.008621670305728912, + 0.004830375779420137, + 0.00445999950170517, + 0.006705957930535078, + 0.022106751799583435, + 0.03095852956175804, + 0.025111138820648193, + 0.0191721823066473, + 0.034972529858350754, + 0.02051965706050396, + 0.02182530052959919, + 0.021177547052502632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.00864406768232584 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot follows a wild boar, the camera gently shaking with its bumpy gait, creating a tense and chaotic sense of pursuit. The massive boar sprints forward through the dense jungle at a reckless, clumsy, and furious pace, its coarse, dark brown mane matted with chunks of damp, dark mud and scattered dead leaves. Two thick, curved ivory tusks jut fiercely from either side of its muddy muzzle, its small, black eyes wide with utter terror. Dappled sunlight filters through the dense canopy, casting shifting patches of light and shadow across its moving form and the forest floor. Its heavy hooves splatter mud and debris on the slippery, decaying leaf-strewn ground, while its powerful shoulders and head shove aside low-hanging branches and brush that block its path. The whole picture is full of primitive dynamics and a tense atmosphere of escape. The main colors are composed of dark green and earthy brown, creating a damp and wild feeling.", + "seed": 10528, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_8/index_3924_seed_10528_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.026804901752620937, + "vlm_logprob": -0.026804901752620937, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.026804901752620937, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03543170541524887 + ], + "opposite_logprob_tokens": [ + -3.4104316234588623 + ], + "logprob_reward": -0.03543170541524887, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03543170541524887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24393987655639648, + 0.06478117406368256, + 0.06032377481460571, + 0.10132120549678802, + 0.17150361835956573, + 0.06933127343654633, + 0.03189921751618385, + 0.05221956968307495, + 0.15758101642131805, + 0.047099266201257706 + ], + "credit_map": [ + 0.017219284549355507, + 0.03690922260284424, + 0.02537520043551922, + 0.07522709667682648, + 0.01676275208592415, + 0.018432220444083214, + 0.019318152219057083, + 0.03469593822956085, + 0.005306223873049021, + 0.01079532876610756, + 0.006678718142211437, + 0.013951057568192482, + 0.006735155824571848, + 0.007147172931581736, + 0.006302559282630682, + 0.007864952087402344, + 0.0038337227888405323, + 0.009013091214001179, + 0.007135889958590269, + 0.006074948236346245, + 0.012478041462600231, + 0.006449171341955662, + 0.00855585653334856, + 0.006783054210245609, + 0.01070333644747734, + 0.013953397050499916, + 0.010997331701219082, + 0.009435571730136871, + 0.02057611010968685, + 0.012924116104841232, + 0.013397584669291973, + 0.009333748370409012, + 0.01967564970254898, + 0.01974230632185936, + 0.025651078671216965, + 0.01739264279603958, + 0.02413296326994896, + 0.024843791499733925, + 0.023835374042391777, + 0.016229795292019844, + 0.016297757625579834, + 0.004669191315770149, + 0.024423394352197647, + 0.00616845116019249, + 0.003589791478589177, + 0.005321901757270098, + 0.005071769468486309, + 0.0037890090607106686, + 0.008916251361370087, + 0.0031267742160707712, + 0.0036040055565536022, + 0.0024956855922937393, + 0.006839536130428314, + 0.0024736816994845867, + 0.002567521994933486, + 0.0018757591024041176, + 0.006291795521974564, + 0.011048021726310253, + 0.005528193432837725, + 0.005343638826161623, + 0.010900222696363926, + 0.004624121822416782, + 0.0048743500374257565, + 0.003609222359955311, + 0.015122116543352604, + 0.016774699091911316, + 0.02177054062485695, + 0.01270689070224762, + 0.01819736324250698, + 0.03985347971320152, + 0.01424818579107523, + 0.018907735124230385, + 0.0036571957170963287, + 0.004864047281444073, + 0.002915391232818365, + 0.003574526170268655, + 0.0151328444480896, + 0.008037771098315716, + 0.004279321059584618, + 0.004638169892132282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.03543170541524887 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812139853835106 + ], + "opposite_logprob_tokens": [ + -3.288121461868286 + ], + "logprob_reward": -0.03812139853835106, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812139853835106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11376409232616425, + 0.0820033997297287, + 0.07203478366136551, + 0.043304335325956345, + 0.05963362753391266, + 0.3007589876651764, + 0.0656185895204544, + 0.057203296571969986, + 0.12357740104198456, + 0.08210150897502899 + ], + "credit_map": [ + 0.010177251882851124, + 0.019708488136529922, + 0.013566911220550537, + 0.02603960782289505, + 0.008132223971188068, + 0.009274673648178577, + 0.011879167519509792, + 0.01498576533049345, + 0.005505637731403112, + 0.01388124655932188, + 0.006506915204226971, + 0.01402834989130497, + 0.010950714349746704, + 0.00918582733720541, + 0.012233112938702106, + 0.009711594320833683, + 0.0054923053830862045, + 0.011958407238125801, + 0.00772496173158288, + 0.007714058738201857, + 0.010481255128979683, + 0.010656967759132385, + 0.00981592945754528, + 0.008190897293388844, + 0.0038189238402992487, + 0.007494182325899601, + 0.003273407695814967, + 0.004423432983458042, + 0.010218597948551178, + 0.0058175139129161835, + 0.004726124927401543, + 0.0035321535542607307, + 0.009760086424648762, + 0.00734128849580884, + 0.008246599696576595, + 0.006838035304099321, + 0.006823711097240448, + 0.006409245077520609, + 0.008432739414274693, + 0.005781922955065966, + 0.09518983215093613, + 0.008961085230112076, + 0.1298462450504303, + 0.022841013967990875, + 0.005882637109607458, + 0.01246492750942707, + 0.011767633259296417, + 0.013805638067424297, + 0.014561989344656467, + 0.006347747053951025, + 0.017640233039855957, + 0.005012496840208769, + 0.008646207861602306, + 0.004277256783097982, + 0.0051599121652543545, + 0.00397274736315012, + 0.0058272466994822025, + 0.006265715695917606, + 0.005838686600327492, + 0.008512461557984352, + 0.007878675125539303, + 0.010428345762193203, + 0.007733939215540886, + 0.004718226380646229, + 0.009439258836209774, + 0.01516428217291832, + 0.016566740348935127, + 0.012696231715381145, + 0.017935538664460182, + 0.025953464210033417, + 0.013188722543418407, + 0.012633158825337887, + 0.006794344633817673, + 0.008358171209692955, + 0.004626226611435413, + 0.005707456264644861, + 0.027797462418675423, + 0.014540943317115307, + 0.008131149224936962, + 0.00614574970677495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.03812139853835106 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043056320399045944 + ], + "opposite_logprob_tokens": [ + -3.1680562496185303 + ], + "logprob_reward": -0.043056320399045944, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043056320399045944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1525588482618332, + 0.0971185564994812, + 0.0741717591881752, + 0.09721148014068604, + 0.11276323348283768, + 0.18470382690429688, + 0.06649722903966904, + 0.06964097172021866, + 0.0584617480635643, + 0.08687232434749603 + ], + "credit_map": [ + 0.018301092088222504, + 0.02555864304304123, + 0.01876380480825901, + 0.021219568327069283, + 0.011312306858599186, + 0.022956619039177895, + 0.013482209295034409, + 0.0209646075963974, + 0.009437529370188713, + 0.020517488941550255, + 0.00919370073825121, + 0.014759044162929058, + 0.009504174813628197, + 0.010466781444847584, + 0.01131246518343687, + 0.01192737277597189, + 0.007700462359935045, + 0.011747553013265133, + 0.007576579228043556, + 0.009277250617742538, + 0.010498653165996075, + 0.010177534073591232, + 0.009826620109379292, + 0.007367110811173916, + 0.00959939043968916, + 0.012818275019526482, + 0.010333853773772717, + 0.008978589437901974, + 0.019638564437627792, + 0.012692675925791264, + 0.01391488965600729, + 0.009235243313014507, + 0.01305517926812172, + 0.01846044696867466, + 0.014026499353349209, + 0.015940655022859573, + 0.015111311338841915, + 0.01274033635854721, + 0.014098785817623138, + 0.0093300212174654, + 0.04283875972032547, + 0.010959711857140064, + 0.05815930664539337, + 0.01990712806582451, + 0.009060952812433243, + 0.01544735487550497, + 0.0162688959389925, + 0.012061710469424725, + 0.017815638333559036, + 0.006658026482909918, + 0.005236551631242037, + 0.004965831991285086, + 0.016904713585972786, + 0.005761280655860901, + 0.00522062461823225, + 0.003934564534574747, + 0.00832489226013422, + 0.012091518379747868, + 0.008536163717508316, + 0.009160119108855724, + 0.010004004463553429, + 0.006782176438719034, + 0.0096743144094944, + 0.0050677829422056675, + 0.0042466758750379086, + 0.0065666018053889275, + 0.005904914811253548, + 0.005172275006771088, + 0.007813829928636551, + 0.01630071923136711, + 0.005426483694463968, + 0.007030246779322624, + 0.005891317501664162, + 0.007981076836585999, + 0.005903522949665785, + 0.004848862066864967, + 0.027751442044973373, + 0.016826607286930084, + 0.009121057577431202, + 0.008548432029783726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.043056320399045944 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783650748431683 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783650748431683, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783650748431683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1431828886270523, + 0.052316077053546906, + 0.05961746349930763, + 0.06014354154467583, + 0.06788937747478485, + 0.2698926627635956, + 0.05482802540063858, + 0.10322780162096024, + 0.07559195160865784, + 0.11331024765968323 + ], + "credit_map": [ + 0.014586195349693298, + 0.02360397018492222, + 0.014657498337328434, + 0.03209923952817917, + 0.009559748694300652, + 0.016288666054606438, + 0.014082899317145348, + 0.018304655328392982, + 0.005307596642524004, + 0.009098202921450138, + 0.006829528603702784, + 0.006031179800629616, + 0.0038064399268478155, + 0.007147074677050114, + 0.006894147489219904, + 0.0072019039653241634, + 0.005426893476396799, + 0.009745089337229729, + 0.00815523136407137, + 0.0081643620505929, + 0.0062879049219191074, + 0.005424061324447393, + 0.009441844187676907, + 0.006972074508666992, + 0.006859638728201389, + 0.007724081631749868, + 0.0073730600997805595, + 0.0060980007983744144, + 0.01016385480761528, + 0.007869241759181023, + 0.007951440289616585, + 0.006104219239205122, + 0.00892704539000988, + 0.00824262760579586, + 0.010788596235215664, + 0.007715210784226656, + 0.008630331605672836, + 0.008194138295948505, + 0.008558811619877815, + 0.006832613609731197, + 0.08262830972671509, + 0.008927132934331894, + 0.11363893747329712, + 0.020259801298379898, + 0.005624186247587204, + 0.01319872960448265, + 0.012296895496547222, + 0.013318671844899654, + 0.012188121676445007, + 0.00433076499029994, + 0.009921562857925892, + 0.0054513500072062016, + 0.011912564747035503, + 0.00408407486975193, + 0.0031735296361148357, + 0.0037660514935851097, + 0.011051115579903126, + 0.01759910024702549, + 0.013067408464848995, + 0.012041675858199596, + 0.016911888495087624, + 0.012782637029886246, + 0.0126581322401762, + 0.00711583998054266, + 0.0071042063646018505, + 0.0070610130205750465, + 0.011085236445069313, + 0.0067270235158503056, + 0.008214463479816914, + 0.019730186089873314, + 0.006765131372958422, + 0.008904687128961086, + 0.009007871150970459, + 0.012348825111985207, + 0.015357141382992268, + 0.007883948273956776, + 0.03110717236995697, + 0.017256928607821465, + 0.010205788537859917, + 0.010142560116946697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.009783650748431683 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wild boar with coarse dark brown fur, two thick curved ivory tusks, and small black eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the boar covered in damp dark mud and scattered dead leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the boar sprinting forward through a dense jungle, shoving aside low-hanging branches and brush?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera use a low-angle tracking shot that gently shakes to mimic the boar's bumpy gait?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by dappled sunlight filtering through the canopy, creating shifting patches of light and shadow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631433662027121 + ], + "opposite_logprob_tokens": [ + -4.882631301879883 + ], + "logprob_reward": -0.007631433662027121, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631433662027121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15423616766929626, + 0.0877300351858139, + 0.06228017061948776, + 0.07540437579154968, + 0.07859113812446594, + 0.1533586084842682, + 0.19980044662952423, + 0.052353035658597946, + 0.1054472029209137, + 0.030798805877566338 + ], + "credit_map": [ + 0.0146066565066576, + 0.019029049202799797, + 0.020381858572363853, + 0.014836348593235016, + 0.010553115978837013, + 0.023809215053915977, + 0.029147066175937653, + 0.021872850134968758, + 0.011940575204789639, + 0.010801922529935837, + 0.012118869461119175, + 0.011657169088721275, + 0.007298135664314032, + 0.013009010814130306, + 0.009517396800220013, + 0.011386943981051445, + 0.006667603738605976, + 0.010502330958843231, + 0.00927992258220911, + 0.005898723378777504, + 0.006833393592387438, + 0.007417160551995039, + 0.009027210995554924, + 0.006653822958469391, + 0.008729908615350723, + 0.012606349773705006, + 0.0069457124918699265, + 0.0067810495384037495, + 0.015301674604415894, + 0.00967530719935894, + 0.0075346482917666435, + 0.00782972015440464, + 0.010679925791919231, + 0.00862190779298544, + 0.012171733193099499, + 0.00828227587044239, + 0.010842345654964447, + 0.009948152117431164, + 0.010539643466472626, + 0.007505148183554411, + 0.03559265285730362, + 0.011764354072511196, + 0.046424854546785355, + 0.017912492156028748, + 0.0084006879478693, + 0.014569710940122604, + 0.009583146311342716, + 0.009110704064369202, + 0.05510159209370613, + 0.019635476171970367, + 0.0273290928453207, + 0.014471837319433689, + 0.04261279106140137, + 0.014690041542053223, + 0.015255237929522991, + 0.010704363696277142, + 0.0066296132281422615, + 0.009447695687413216, + 0.0049814204685389996, + 0.007650254759937525, + 0.006814775988459587, + 0.00679408572614193, + 0.0062689450569450855, + 0.003766243578866124, + 0.008622602559626102, + 0.011540563777089119, + 0.01717885583639145, + 0.00857919454574585, + 0.01386539451777935, + 0.023970533162355423, + 0.009985726326704025, + 0.011704332195222378, + 0.0032202936708927155, + 0.0025695303920656443, + 0.0027052981313318014, + 0.0033356002531945705, + 0.007449001539498568, + 0.004297057632356882, + 0.004745976999402046, + 0.002476046560332179 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.007631433662027121 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_9/index_tensor([5875, 5875])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_9/index_tensor([5875, 5875])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cc416756738092ae2efe4fb79d57be1aa8465673 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_9/index_tensor([5875, 5875])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures the scene in a brightly lit modern kitchen. A man with slightly tousled brown hair casually walks toward a set of lower cabinets. He wears a red and white checkered, short-sleeved button-down shirt and contrasting teal checkered shorts, creating a relaxed, homely look. He pauses before a white, rickety cabinet and reaches down deliberately, his fingers grasping the small, brushed metal knob. He smoothly pulls the door open, revealing shelves filled with a messy yet organized pantry: an assortment of unlabeled glass jars, colorful tins, and plain cardboard boxes. Holding a simple glass jar with a plain metal lid in his right hand, he takes a step closer, bending slightly at his knees and waist, and carefully reaches into the cabinet. Steadily and accurately, he places the jar on the bottom shelf, nestled between a stack of jars and a box. Once his fingers release the jar, he straightens and begins to rotate his body to the left. As he begins to turn away from the open cabinet, his movements are fluid and unhurried, completing a simple chore. The scene is bathed in warm, soft light from overhead fixtures, creating a comfortable and tranquil atmosphere.", + "seed": 11771, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_9/index_5875_seed_11771_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09658260568976403, + "vlm_logprob": -0.09658260568976403, + "vlm_yes_no_margin": 2.8999999523162843, + "vlm_token_credit": -0.09658260568976403, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28803688287734985 + ], + "opposite_logprob_tokens": [ + -1.413036823272705 + ], + "logprob_reward": -0.28803688287734985, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28803688287734985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21611535549163818, + 0.04545843228697777, + 0.039579764008522034, + 0.042150817811489105, + 0.07658649981021881, + 0.20944486558437347, + 0.14964041113853455, + 0.07388943433761597, + 0.09670181572437286, + 0.05043265223503113 + ], + "credit_map": [ + 0.016742443665862083, + 0.026803329586982727, + 0.026571357622742653, + 0.035344064235687256, + 0.031059743836522102, + 0.02065790444612503, + 0.04246200993657112, + 0.016474509611725807, + 0.0051264893263578415, + 0.006614371202886105, + 0.0052692038007080555, + 0.005406912416219711, + 0.004146865103393793, + 0.006722417660057545, + 0.005877967458218336, + 0.006294206716120243, + 0.003515522228553891, + 0.0028308765031397343, + 0.003980109468102455, + 0.0038640082348138094, + 0.007154411636292934, + 0.004105773288756609, + 0.004881176166236401, + 0.009247889742255211, + 0.007080955896526575, + 0.009163152426481247, + 0.006477234419435263, + 0.003959714435040951, + 0.0035814051516354084, + 0.0035658921115100384, + 0.0038762495387345552, + 0.0044462126679718494, + 0.009182753972709179, + 0.007874393835663795, + 0.01580854319036007, + 0.011182421818375587, + 0.006686259992420673, + 0.01163491141051054, + 0.0076762051321566105, + 0.006541009526699781, + 0.025196339935064316, + 0.017726602032780647, + 0.015827253460884094, + 0.014091961085796356, + 0.025523751974105835, + 0.014146380126476288, + 0.0636727437376976, + 0.033259838819503784, + 0.02614506706595421, + 0.01962432451546192, + 0.021861504763364792, + 0.013707766309380531, + 0.01865423657000065, + 0.016490980982780457, + 0.01439798902720213, + 0.018758540973067284, + 0.006815822329372168, + 0.008357218466699123, + 0.00781895313411951, + 0.008073192089796066, + 0.008117902092635632, + 0.006913746241480112, + 0.013154039159417152, + 0.014638558030128479, + 0.008219428360462189, + 0.013643382117152214, + 0.008092688396573067, + 0.011699912138283253, + 0.020550794899463654, + 0.009566417895257473, + 0.01174637209624052, + 0.01318281888961792, + 0.0061570159159600735, + 0.004853876773267984, + 0.005830930545926094, + 0.004421295132488012, + 0.008551375940442085, + 0.011405005119740963, + 0.0048098317347466946, + 0.004403324332088232 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.28803688287734985 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018211066722869873 + ], + "opposite_logprob_tokens": [ + -4.0182108879089355 + ], + "logprob_reward": -0.018211066722869873, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018211066722869873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22512954473495483, + 0.06676401197910309, + 0.03814677894115448, + 0.059244006872177124, + 0.06703843921422958, + 0.11450248956680298, + 0.14867667853832245, + 0.10455348342657089, + 0.05275799334049225, + 0.12318655103445053 + ], + "credit_map": [ + 0.01309980172663927, + 0.021319987252354622, + 0.02949702739715576, + 0.04643768072128296, + 0.0238860622048378, + 0.020997095853090286, + 0.03649289906024933, + 0.03339898958802223, + 0.005813912488520145, + 0.006189746782183647, + 0.005722351837903261, + 0.006341677159070969, + 0.009062152355909348, + 0.007517235819250345, + 0.011875001713633537, + 0.01424193661659956, + 0.0032920432277023792, + 0.0023281474132090807, + 0.005731719546020031, + 0.005752998869866133, + 0.007028615567833185, + 0.0032544229179620743, + 0.0062250904738903046, + 0.004533741157501936, + 0.008134719915688038, + 0.00769949983805418, + 0.008186906576156616, + 0.008155692368745804, + 0.00842391699552536, + 0.005076522473245859, + 0.007801433093845844, + 0.0057653156109154224, + 0.006607488263398409, + 0.007016950286924839, + 0.005376983899623156, + 0.004537370055913925, + 0.018335731700062752, + 0.010099944658577442, + 0.007694185245782137, + 0.007369783706963062, + 0.013157294131815434, + 0.011289659887552261, + 0.009382755495607853, + 0.006050635129213333, + 0.021606406196951866, + 0.023179057985544205, + 0.015918511897325516, + 0.013918173499405384, + 0.01284002885222435, + 0.020622611045837402, + 0.011651741340756416, + 0.009652223438024521, + 0.04347711428999901, + 0.014866284094750881, + 0.020028719678521156, + 0.015537962317466736, + 0.008436678908765316, + 0.012237885035574436, + 0.0113973468542099, + 0.008630286902189255, + 0.013881450518965721, + 0.009479367174208164, + 0.02502295933663845, + 0.015467516146600246, + 0.005775680299848318, + 0.008115665055811405, + 0.005612964276224375, + 0.00652353884652257, + 0.007280787453055382, + 0.00904727540910244, + 0.004728502128273249, + 0.005673579405993223, + 0.018954908475279808, + 0.012891235761344433, + 0.00830914732068777, + 0.01035388559103012, + 0.018917446956038475, + 0.037547409534454346, + 0.009444176219403744, + 0.006768341176211834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.018211066722869873 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026336906477808952 + ], + "opposite_logprob_tokens": [ + -3.651336908340454 + ], + "logprob_reward": -0.026336906477808952, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026336906477808952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1803375631570816, + 0.05808888375759125, + 0.09603758901357651, + 0.08565665036439896, + 0.1042184978723526, + 0.18193183839321136, + 0.06673972308635712, + 0.03168535977602005, + 0.08006826043128967, + 0.11523561924695969 + ], + "credit_map": [ + 0.017537835985422134, + 0.02664824016392231, + 0.024222958832979202, + 0.030062319710850716, + 0.02116798236966133, + 0.016062162816524506, + 0.023920999839901924, + 0.020715055987238884, + 0.008359267376363277, + 0.009126012213528156, + 0.0077246795408427715, + 0.006730877328664064, + 0.005855316761881113, + 0.008568533696234226, + 0.005910187028348446, + 0.005814009346067905, + 0.009607666172087193, + 0.0064308540895581245, + 0.007564275059849024, + 0.012904470786452293, + 0.018763842061161995, + 0.00975142139941454, + 0.01098542008548975, + 0.020029639825224876, + 0.01923407055437565, + 0.011067110113799572, + 0.02174987643957138, + 0.008127675391733646, + 0.006825045682489872, + 0.00533643364906311, + 0.007941444404423237, + 0.005374981556087732, + 0.0077710277400910854, + 0.005741480737924576, + 0.020307505503296852, + 0.007619177456945181, + 0.03141847997903824, + 0.01400213222950697, + 0.009199848398566246, + 0.008158847689628601, + 0.018785947933793068, + 0.019767649471759796, + 0.013122416101396084, + 0.01006823219358921, + 0.020980823785066605, + 0.011270728893578053, + 0.056482456624507904, + 0.03145357593894005, + 0.008344375528395176, + 0.008151393383741379, + 0.008038883097469807, + 0.005025099962949753, + 0.01515929028391838, + 0.006178160663694143, + 0.007903228513896465, + 0.007939282804727554, + 0.003068875055760145, + 0.00476430170238018, + 0.004389834124594927, + 0.005499206017702818, + 0.004597257822751999, + 0.0028393035754561424, + 0.003689427161589265, + 0.0028371536172926426, + 0.008365126326680183, + 0.010950363241136074, + 0.008124844171106815, + 0.010845771990716457, + 0.015389016829431057, + 0.007382313720881939, + 0.008868135511875153, + 0.010142680257558823, + 0.014458483085036278, + 0.011937067843973637, + 0.006317437160760164, + 0.004492015112191439, + 0.02589949406683445, + 0.04371514171361923, + 0.004010501317679882, + 0.004405464977025986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.026336906477808952 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270398646593094 + ], + "opposite_logprob_tokens": [ + -2.127039909362793 + ], + "logprob_reward": -0.1270398646593094, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270398646593094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.234537735581398, + 0.036055684089660645, + 0.03683294728398323, + 0.12151797115802765, + 0.14074274897575378, + 0.14431461691856384, + 0.05693897232413292, + 0.06265261769294739, + 0.04269932955503464, + 0.12370727211236954 + ], + "credit_map": [ + 0.01025653351098299, + 0.024558277800679207, + 0.03201454505324364, + 0.045294154435396194, + 0.02494923397898674, + 0.020714683458209038, + 0.04587636888027191, + 0.03087395429611206, + 0.004575000610202551, + 0.005098563153296709, + 0.004682617262005806, + 0.003517165780067444, + 0.004458579700440168, + 0.0056974138133227825, + 0.0033645187504589558, + 0.004661827813833952, + 0.003491403302177787, + 0.002802525646984577, + 0.004779766779392958, + 0.006032281555235386, + 0.007233212236315012, + 0.0033385350834578276, + 0.004758467897772789, + 0.004396759904921055, + 0.03404079005122185, + 0.008697150275111198, + 0.03621393069624901, + 0.00980861485004425, + 0.008845030330121517, + 0.009663678705692291, + 0.007898333482444286, + 0.006350454408675432, + 0.013508264906704426, + 0.007048434112221003, + 0.0341835543513298, + 0.015643281862139702, + 0.033862289041280746, + 0.01728730835020542, + 0.010180000215768814, + 0.00902964174747467, + 0.01367840263992548, + 0.0145656643435359, + 0.01639268919825554, + 0.010384142398834229, + 0.01635095663368702, + 0.014463264495134354, + 0.034565556794404984, + 0.02391395904123783, + 0.004525832831859589, + 0.005315148271620274, + 0.004163844510912895, + 0.0035902790259569883, + 0.015707459300756454, + 0.007496509235352278, + 0.007291089743375778, + 0.008848819881677628, + 0.0038045363035053015, + 0.004641932435333729, + 0.004731256980448961, + 0.0055958726443350315, + 0.01014269981533289, + 0.00703270873054862, + 0.013252206146717072, + 0.0134514095261693, + 0.0038419964257627726, + 0.0075608049519360065, + 0.004554372746497393, + 0.005566225852817297, + 0.006547928787767887, + 0.005414061713963747, + 0.0031602743547409773, + 0.006053667049854994, + 0.017155593261122704, + 0.014173481613397598, + 0.008280175738036633, + 0.006152186542749405, + 0.02485317550599575, + 0.04092803969979286, + 0.005536586046218872, + 0.006628044415265322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.1270398646593094 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288307711482048 + ], + "opposite_logprob_tokens": [ + -3.7732882499694824 + ], + "logprob_reward": -0.023288307711482048, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288307711482048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4310445785522461, + 0.039972975850105286, + 0.05519651249051094, + 0.03532028943300247, + 0.03755011782050133, + 0.09346567094326019, + 0.05575333163142204, + 0.10148049145936966, + 0.02646183967590332, + 0.12375422567129135 + ], + "credit_map": [ + 0.04908835515379906, + 0.07370130717754364, + 0.061747223138809204, + 0.0564345046877861, + 0.051944103091955185, + 0.033812057226896286, + 0.05416911467909813, + 0.0501478910446167, + 0.006142352242022753, + 0.005380243994295597, + 0.004163878969848156, + 0.005905752070248127, + 0.004461390431970358, + 0.004186008591204882, + 0.00451459176838398, + 0.00521875498816371, + 0.0052630226127803326, + 0.0034978745970875025, + 0.004782871343195438, + 0.009195350110530853, + 0.00749018881469965, + 0.005593093577772379, + 0.005865071434527636, + 0.013509036973118782, + 0.006226298864930868, + 0.003521811915561557, + 0.0061731780879199505, + 0.0037383651360869408, + 0.004821468610316515, + 0.002978983800858259, + 0.00529756024479866, + 0.0025626211427152157, + 0.00418475829064846, + 0.005359383299946785, + 0.00522145489230752, + 0.002446182770654559, + 0.004308722447603941, + 0.006275553721934557, + 0.004244926851242781, + 0.005509132519364357, + 0.007931498810648918, + 0.011068171821534634, + 0.006507959216833115, + 0.006096226628869772, + 0.010967056266963482, + 0.007468591444194317, + 0.02683219686150551, + 0.016593966633081436, + 0.008650965057313442, + 0.0069505441933870316, + 0.008318552747368813, + 0.004852148238569498, + 0.009368136525154114, + 0.005480736959725618, + 0.006416696589440107, + 0.005715548060834408, + 0.008731422014534473, + 0.01136311236768961, + 0.012269626371562481, + 0.015290207229554653, + 0.013338711112737656, + 0.008708968758583069, + 0.015810493379831314, + 0.015967946499586105, + 0.0026250071823596954, + 0.0032599407713860273, + 0.002745836740359664, + 0.0035502505488693714, + 0.0045515443198382854, + 0.003244372783228755, + 0.002349533373489976, + 0.004135353025048971, + 0.01385553926229477, + 0.01437295414507389, + 0.01060958206653595, + 0.0072401431389153, + 0.024654243141412735, + 0.034794148057699203, + 0.007083591539412737, + 0.011144012212753296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.023288307711482048 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures the scene in a brightly lit modern kitchen. A man with slightly tousled brown hair casually walks toward a set of lower cabinets. He wears a red and white checkered, short-sleeved button-down shirt and contrasting teal checkered shorts, creating a relaxed, homely look. He pauses before a white, rickety cabinet and reaches down deliberately, his fingers grasping the small, brushed metal knob. He smoothly pulls the door open, revealing shelves filled with a messy yet organized pantry: an assortment of unlabeled glass jars, colorful tins, and plain cardboard boxes. Holding a simple glass jar with a plain metal lid in his right hand, he takes a step closer, bending slightly at his knees and waist, and carefully reaches into the cabinet. Steadily and accurately, he places the jar on the bottom shelf, nestled between a stack of jars and a box. Once his fingers release the jar, he straightens and begins to rotate his body to the left. As he begins to turn away from the open cabinet, his movements are fluid and unhurried, completing a simple chore. The scene is bathed in warm, soft light from overhead fixtures, creating a comfortable and tranquil atmosphere.", + "seed": 11772, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_9/index_5875_seed_11772_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09733625780791044, + "vlm_logprob": -0.09733625780791044, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.09733625780791044, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2251104712486267 + ], + "opposite_logprob_tokens": [ + -1.7251105308532715 + ], + "logprob_reward": -0.2251104712486267, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2251104712486267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21093891561031342, + 0.056672364473342896, + 0.09291477501392365, + 0.0982864573597908, + 0.045911334455013275, + 0.060711704194545746, + 0.04210395738482475, + 0.12686055898666382, + 0.17848122119903564, + 0.08711879700422287 + ], + "credit_map": [ + 0.03302968665957451, + 0.03334823250770569, + 0.027778906747698784, + 0.02060094103217125, + 0.03336300328373909, + 0.022462768480181694, + 0.027797246351838112, + 0.012558099813759327, + 0.005645396187901497, + 0.015215694904327393, + 0.008236045017838478, + 0.005098304245620966, + 0.007328896317631006, + 0.0053613740019500256, + 0.005772275384515524, + 0.004014375153928995, + 0.013497857376933098, + 0.017832400277256966, + 0.01567542739212513, + 0.012521696276962757, + 0.010961745865643024, + 0.009540854021906853, + 0.006647857371717691, + 0.006236930377781391, + 0.013020928017795086, + 0.01335671916604042, + 0.01740959845483303, + 0.0168981421738863, + 0.011378854513168335, + 0.0071159061044454575, + 0.011621438898146152, + 0.007484858855605125, + 0.006304865702986717, + 0.006430016830563545, + 0.008337742649018764, + 0.00458017410710454, + 0.007212531752884388, + 0.004299866501241922, + 0.005311441142112017, + 0.00343469250947237, + 0.010279808193445206, + 0.00711104879155755, + 0.010752780362963676, + 0.008171665482223034, + 0.00443094689399004, + 0.007603561971336603, + 0.00544923497363925, + 0.006912652403116226, + 0.004665141925215721, + 0.006765719968825579, + 0.004315512254834175, + 0.005672826897352934, + 0.006719894241541624, + 0.005093242973089218, + 0.005287745967507362, + 0.0035838645417243242, + 0.017978716641664505, + 0.01695924438536167, + 0.019973430782556534, + 0.012012246996164322, + 0.02433880791068077, + 0.013830199837684631, + 0.011537924408912659, + 0.0102299889549613, + 0.023179633542895317, + 0.026626404374837875, + 0.03126917779445648, + 0.024784935638308525, + 0.02447757124900818, + 0.018917877227067947, + 0.012837148271501064, + 0.016388466581702232, + 0.009104953147470951, + 0.010511159896850586, + 0.013704382814466953, + 0.014878761023283005, + 0.010199310258030891, + 0.008649995550513268, + 0.012760930694639683, + 0.007309292908757925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.2251104712486267 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008666994981467724 + ], + "opposite_logprob_tokens": [ + -4.7586669921875 + ], + "logprob_reward": -0.008666994981467724, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008666994981467724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15516957640647888, + 0.10123489052057266, + 0.11011290550231934, + 0.1304939240217209, + 0.10362675040960312, + 0.10417284816503525, + 0.05489557608962059, + 0.09260416030883789, + 0.06873244792222977, + 0.07895689457654953 + ], + "credit_map": [ + 0.021213261410593987, + 0.022060170769691467, + 0.021210622042417526, + 0.011252104304730892, + 0.025478770956397057, + 0.01647312194108963, + 0.023586710914969444, + 0.013894818723201752, + 0.011388123966753483, + 0.018854539841413498, + 0.012814858928322792, + 0.006692624185234308, + 0.013374991714954376, + 0.016427190974354744, + 0.015059156343340874, + 0.006623406894505024, + 0.011023221537470818, + 0.012851673178374767, + 0.010888825170695782, + 0.006546101067215204, + 0.02291836403310299, + 0.024956153705716133, + 0.012949180789291859, + 0.007979393936693668, + 0.020573431625962257, + 0.01832440122961998, + 0.023935630917549133, + 0.02490437589585781, + 0.012229569256305695, + 0.00859683845192194, + 0.012458985671401024, + 0.009470708668231964, + 0.013872288167476654, + 0.013173680752515793, + 0.013609545305371284, + 0.009847387671470642, + 0.01632339507341385, + 0.011676693335175514, + 0.01500720251351595, + 0.010116564109921455, + 0.018313080072402954, + 0.011698934249579906, + 0.02504579722881317, + 0.0137495007365942, + 0.009026066400110722, + 0.009355749934911728, + 0.009171017445623875, + 0.00781270395964384, + 0.00549653684720397, + 0.008616013452410698, + 0.004920713137835264, + 0.006813675165176392, + 0.0109030120074749, + 0.005237256176769733, + 0.008798989467322826, + 0.00410938635468483, + 0.015629269182682037, + 0.016774989664554596, + 0.01282298844307661, + 0.00815892219543457, + 0.01612132042646408, + 0.00816390011459589, + 0.008303102105855942, + 0.006629678886383772, + 0.00821855291724205, + 0.00823177956044674, + 0.01831037551164627, + 0.008858246728777885, + 0.007593653164803982, + 0.005990472622215748, + 0.005165185313671827, + 0.006364183034747839, + 0.009096077643334866, + 0.009092234075069427, + 0.007642125245183706, + 0.008631445467472076, + 0.01584889180958271, + 0.010552641935646534, + 0.01026811171323061, + 0.007825374603271484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.008666994981467724 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779755026102066 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779755026102066, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779755026102066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0608750581741333, + 0.04479752480983734, + 0.05437600612640381, + 0.16793908178806305, + 0.06456227600574493, + 0.12126185745000839, + 0.062011536210775375, + 0.20365647971630096, + 0.1425113081932068, + 0.07800885289907455 + ], + "credit_map": [ + 0.009182272478938103, + 0.007754133082926273, + 0.00848890095949173, + 0.004065516404807568, + 0.011647140607237816, + 0.006338442210108042, + 0.00847604125738144, + 0.004922613501548767, + 0.005114800296723843, + 0.010103107430040836, + 0.0062719592824578285, + 0.00543934665620327, + 0.004816476721316576, + 0.004617622122168541, + 0.006062781438231468, + 0.0023714378476142883, + 0.0071636284701526165, + 0.010118997655808926, + 0.007839495316147804, + 0.006124609615653753, + 0.007028317078948021, + 0.006355880293995142, + 0.005994951352477074, + 0.0037501289043575525, + 0.033722102642059326, + 0.024074558168649673, + 0.04067954421043396, + 0.020797722041606903, + 0.017497681081295013, + 0.013309307396411896, + 0.008742264471948147, + 0.009115917608141899, + 0.010056691244244576, + 0.010528062470257282, + 0.01268208771944046, + 0.005936512257903814, + 0.006841491442173719, + 0.005749838892370462, + 0.00829343032091856, + 0.004474165849387646, + 0.0203566811978817, + 0.015231026336550713, + 0.02009003236889839, + 0.02866072952747345, + 0.009212344884872437, + 0.007825604639947414, + 0.012212221510708332, + 0.007673218846321106, + 0.006132006645202637, + 0.009181885048747063, + 0.005773458164185286, + 0.009207332506775856, + 0.009938383474946022, + 0.007482586428523064, + 0.008231757208704948, + 0.006064129993319511, + 0.028276195749640465, + 0.04504035785794258, + 0.028657304123044014, + 0.02172066643834114, + 0.023721840232610703, + 0.026441967114806175, + 0.0159026850014925, + 0.013895473442971706, + 0.014613146893680096, + 0.01948506385087967, + 0.02114003337919712, + 0.02475919760763645, + 0.016884570941329002, + 0.01837419904768467, + 0.01700262911617756, + 0.01025247573852539, + 0.009542602114379406, + 0.008950643241405487, + 0.010069250129163265, + 0.011420818977057934, + 0.01334376074373722, + 0.0069143385626375675, + 0.01179729774594307, + 0.005970143247395754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.009779755026102066 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280490577220917 + ], + "opposite_logprob_tokens": [ + -2.0178048610687256 + ], + "logprob_reward": -0.14280490577220917, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14280490577220917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052206531167030334, + 0.09302709251642227, + 0.03620850667357445, + 0.12786416709423065, + 0.19586026668548584, + 0.19244849681854248, + 0.0810781940817833, + 0.06993399560451508, + 0.07687738537788391, + 0.07449532300233841 + ], + "credit_map": [ + 0.00628990214318037, + 0.006075748708099127, + 0.006849068682640791, + 0.0057002282701432705, + 0.008118118159472942, + 0.006511134095489979, + 0.008188378065824509, + 0.004473954439163208, + 0.012598639354109764, + 0.026117589324712753, + 0.015889326110482216, + 0.014148151502013206, + 0.006513264030218124, + 0.005753949284553528, + 0.007737561594694853, + 0.004268611315637827, + 0.005448875483125448, + 0.007502265740185976, + 0.005633923225104809, + 0.0036328483838588, + 0.004749227315187454, + 0.004205402452498674, + 0.00257924385368824, + 0.002456719521433115, + 0.01847083307802677, + 0.021741904318332672, + 0.026986904442310333, + 0.014522593468427658, + 0.012346248142421246, + 0.011554088443517685, + 0.009085839614272118, + 0.013155749067664146, + 0.027271201834082603, + 0.03382932394742966, + 0.03369937464594841, + 0.0266578309237957, + 0.026189707219600677, + 0.01479915902018547, + 0.02362692914903164, + 0.009786725975573063, + 0.029890095815062523, + 0.019657792523503304, + 0.025578675791621208, + 0.0551295131444931, + 0.015212835744023323, + 0.016224680468440056, + 0.019596418365836143, + 0.011158490553498268, + 0.008443991653621197, + 0.011584398336708546, + 0.007067844737321138, + 0.011164376512169838, + 0.01481187716126442, + 0.0061913831159472466, + 0.012915327213704586, + 0.008898993022739887, + 0.013586007989943027, + 0.011285287328064442, + 0.008447786793112755, + 0.00785429310053587, + 0.009465724229812622, + 0.007409713231027126, + 0.006560118868947029, + 0.005325060337781906, + 0.010034679435193539, + 0.010792585089802742, + 0.0159747451543808, + 0.009713764302432537, + 0.009783956222236156, + 0.008090930990874767, + 0.006894949823617935, + 0.005591777618974447, + 0.008856149390339851, + 0.00787362176924944, + 0.007102575618773699, + 0.01186364982277155, + 0.011399250477552414, + 0.011130549013614655, + 0.008833705447614193, + 0.0074358233250677586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.14280490577220917 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031916201114655 + ], + "opposite_logprob_tokens": [ + -2.3503191471099854 + ], + "logprob_reward": -0.10031916201114655, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031916201114655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2122001349925995, + 0.10531625151634216, + 0.11690495908260345, + 0.06366508454084396, + 0.03790685907006264, + 0.07403429597616196, + 0.06973113119602203, + 0.09904276579618454, + 0.12010041624307632, + 0.10109809786081314 + ], + "credit_map": [ + 0.04708903655409813, + 0.02432972751557827, + 0.04074517637491226, + 0.01505606435239315, + 0.017677538096904755, + 0.024635963141918182, + 0.03215720131993294, + 0.010509422980248928, + 0.011248440481722355, + 0.03430311009287834, + 0.016762403771281242, + 0.013471332378685474, + 0.00922488421201706, + 0.007389864418655634, + 0.007290195673704147, + 0.0056260195560753345, + 0.01644040085375309, + 0.02485515922307968, + 0.01910519227385521, + 0.018331699073314667, + 0.011223571375012398, + 0.012283124029636383, + 0.008492433466017246, + 0.006173377390950918, + 0.008393215015530586, + 0.008033630438148975, + 0.0119395786896348, + 0.007065173238515854, + 0.0067708189599215984, + 0.005640040151774883, + 0.009194280952215195, + 0.006628348026424646, + 0.006743456237018108, + 0.004350797738879919, + 0.00598166324198246, + 0.003811896312981844, + 0.004867920186370611, + 0.004401656799018383, + 0.004856156650930643, + 0.002893314464017749, + 0.012327004224061966, + 0.009087641723453999, + 0.01743590459227562, + 0.018637413159012794, + 0.00505836820229888, + 0.003885789541527629, + 0.004903737921267748, + 0.002698441967368126, + 0.007762979716062546, + 0.010489517822861671, + 0.0066427248530089855, + 0.009946615435183048, + 0.012338305823504925, + 0.007879083976149559, + 0.009237994439899921, + 0.005433907266706228, + 0.014354083687067032, + 0.019126448780298233, + 0.01710718497633934, + 0.01200823299586773, + 0.01385966595262289, + 0.0074117607437074184, + 0.009313761256635189, + 0.005861625075340271, + 0.016302114352583885, + 0.015870729461312294, + 0.01586294360458851, + 0.012695148587226868, + 0.015868406742811203, + 0.014695599675178528, + 0.010836108587682247, + 0.017969366163015366, + 0.009769695810973644, + 0.011267793364822865, + 0.010163648054003716, + 0.019930049777030945, + 0.01890525594353676, + 0.008220549672842026, + 0.015492651611566544, + 0.007348456420004368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.10031916201114655 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures the scene in a brightly lit modern kitchen. A man with slightly tousled brown hair casually walks toward a set of lower cabinets. He wears a red and white checkered, short-sleeved button-down shirt and contrasting teal checkered shorts, creating a relaxed, homely look. He pauses before a white, rickety cabinet and reaches down deliberately, his fingers grasping the small, brushed metal knob. He smoothly pulls the door open, revealing shelves filled with a messy yet organized pantry: an assortment of unlabeled glass jars, colorful tins, and plain cardboard boxes. Holding a simple glass jar with a plain metal lid in his right hand, he takes a step closer, bending slightly at his knees and waist, and carefully reaches into the cabinet. Steadily and accurately, he places the jar on the bottom shelf, nestled between a stack of jars and a box. Once his fingers release the jar, he straightens and begins to rotate his body to the left. As he begins to turn away from the open cabinet, his movements are fluid and unhurried, completing a simple chore. The scene is bathed in warm, soft light from overhead fixtures, creating a comfortable and tranquil atmosphere.", + "seed": 11773, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_9/index_5875_seed_11773_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09014625335112214, + "vlm_logprob": -0.09014625335112214, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.09014625335112214, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.24882107973098755 + ], + "opposite_logprob_tokens": [ + -1.6238210201263428 + ], + "logprob_reward": -0.24882107973098755, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.24882107973098755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15032440423965454, + 0.0478469654917717, + 0.04886287450790405, + 0.030279379338026047, + 0.06552527844905853, + 0.07402937114238739, + 0.09906386584043503, + 0.11029012501239777, + 0.28144896030426025, + 0.09232880920171738 + ], + "credit_map": [ + 0.013452418148517609, + 0.022903205826878548, + 0.02062767930328846, + 0.028766589239239693, + 0.018959136679768562, + 0.022189240902662277, + 0.011706208810210228, + 0.011719925329089165, + 0.005650227889418602, + 0.006634227465838194, + 0.008179265074431896, + 0.009486586786806583, + 0.0057335118763148785, + 0.006049341522157192, + 0.003365712473168969, + 0.0027480905409902334, + 0.0071022105403244495, + 0.0068052988499403, + 0.009361226111650467, + 0.007927555590867996, + 0.006053784396499395, + 0.004490412771701813, + 0.004055613186210394, + 0.003066771198064089, + 0.0033793740440160036, + 0.0049372646026313305, + 0.004430324304848909, + 0.003745090216398239, + 0.005115837790071964, + 0.0024461792781949043, + 0.004275398328900337, + 0.0019499115878716111, + 0.006795292720198631, + 0.008074324578046799, + 0.010062295012176037, + 0.009753664955496788, + 0.0077871051616966724, + 0.009456903673708439, + 0.0074532534927129745, + 0.00614243745803833, + 0.004302255809307098, + 0.006885075941681862, + 0.00871642492711544, + 0.01840643398463726, + 0.008135318756103516, + 0.006222454831004143, + 0.016817621886730194, + 0.004543785471469164, + 0.007967622950673103, + 0.013073829934000969, + 0.00934478361159563, + 0.0114157535135746, + 0.015327413566410542, + 0.011077731847763062, + 0.01730429194867611, + 0.013552443124353886, + 0.007354261819273233, + 0.014004380442202091, + 0.007229266222566366, + 0.014210768975317478, + 0.01384829357266426, + 0.01573321595788002, + 0.02197437919676304, + 0.015935556963086128, + 0.019163405522704124, + 0.021331191062927246, + 0.0708390474319458, + 0.03811141103506088, + 0.02487342804670334, + 0.03274234011769295, + 0.05891745537519455, + 0.015470682643353939, + 0.008927591145038605, + 0.009251058101654053, + 0.018240205943584442, + 0.014972544275224209, + 0.014992390759289265, + 0.013284837827086449, + 0.007197892758995295, + 0.005462289322167635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.24882107973098755 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026382768526673317 + ], + "opposite_logprob_tokens": [ + -3.6513826847076416 + ], + "logprob_reward": -0.026382768526673317, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026382768526673317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11374179273843765, + 0.07029585540294647, + 0.10907074809074402, + 0.17068707942962646, + 0.1869911104440689, + 0.05063490942120552, + 0.031971514225006104, + 0.06910783797502518, + 0.12199785560369492, + 0.07550130039453506 + ], + "credit_map": [ + 0.016120795160531998, + 0.014088993892073631, + 0.015091555193066597, + 0.017631318420171738, + 0.01454225368797779, + 0.014826567843556404, + 0.00970381498336792, + 0.011736501939594746, + 0.006143782287836075, + 0.008110562339425087, + 0.010871363803744316, + 0.00831815879791975, + 0.009250988252460957, + 0.007597220595926046, + 0.011809860356152058, + 0.00819392316043377, + 0.009095434099435806, + 0.01507012639194727, + 0.010118218138813972, + 0.015663523226976395, + 0.016870010644197464, + 0.021502016112208366, + 0.013403382152318954, + 0.007348043378442526, + 0.018173947930336, + 0.01919584535062313, + 0.029015110805630684, + 0.020765159279108047, + 0.021222466602921486, + 0.01776127517223358, + 0.024372247979044914, + 0.020181048661470413, + 0.019420932978391647, + 0.016839666292071342, + 0.024746762588620186, + 0.03226320073008537, + 0.019276540726423264, + 0.026538655161857605, + 0.02225276082754135, + 0.025652602314949036, + 0.005166452843695879, + 0.005419621244072914, + 0.006914798635989428, + 0.005057343281805515, + 0.005975129082798958, + 0.008486859500408173, + 0.005533262621611357, + 0.008081447333097458, + 0.0020723806228488684, + 0.0027094376273453236, + 0.0033979243598878384, + 0.0070136431604623795, + 0.0020280040334910154, + 0.002617812482640147, + 0.007534322794526815, + 0.004597993567585945, + 0.004986132960766554, + 0.009704804979264736, + 0.004751990549266338, + 0.010079816915094852, + 0.008586457930505276, + 0.01014677807688713, + 0.012472430244088173, + 0.008379428647458553, + 0.008574141189455986, + 0.009412234649062157, + 0.030451035127043724, + 0.016972409561276436, + 0.011581611819565296, + 0.013520436361432076, + 0.024517033249139786, + 0.0069689666852355, + 0.005592611618340015, + 0.005917987320572138, + 0.010386906564235687, + 0.008672986179590225, + 0.010231174528598785, + 0.010908196680247784, + 0.010346874594688416, + 0.013444573618471622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.026382768526673317 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630369160324335 + ], + "opposite_logprob_tokens": [ + -4.882630348205566 + ], + "logprob_reward": -0.007630369160324335, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630369160324335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10359025001525879, + 0.04652213677763939, + 0.14643718302249908, + 0.12553510069847107, + 0.16827116906642914, + 0.0796353742480278, + 0.12309936434030533, + 0.05576487258076668, + 0.07089829444885254, + 0.08024623990058899 + ], + "credit_map": [ + 0.010728158056735992, + 0.013453440740704536, + 0.015285543166100979, + 0.019276877865195274, + 0.01172769907861948, + 0.011685077100992203, + 0.007123601622879505, + 0.01430985052138567, + 0.004357592668384314, + 0.0072359261102974415, + 0.009866033680737019, + 0.00842340849339962, + 0.003898515133187175, + 0.0070027513429522514, + 0.003766354639083147, + 0.0019715565722435713, + 0.022890083491802216, + 0.021296370774507523, + 0.03463743254542351, + 0.026029467582702637, + 0.01626681536436081, + 0.009999158792197704, + 0.008721799589693546, + 0.006596054881811142, + 0.019820084795355797, + 0.01849995367228985, + 0.024014746770262718, + 0.013919278047978878, + 0.01821134053170681, + 0.010693715885281563, + 0.014450055547058582, + 0.005925935227423906, + 0.02193458005785942, + 0.030133601278066635, + 0.019352348521351814, + 0.03684926778078079, + 0.016312042251229286, + 0.01509382389485836, + 0.01732070930302143, + 0.01127479039132595, + 0.009455508552491665, + 0.01036188006401062, + 0.01286403276026249, + 0.01218795869499445, + 0.00781591422855854, + 0.007687730249017477, + 0.013343132100999355, + 0.005919213406741619, + 0.01005165558308363, + 0.010322739370167255, + 0.01521387044340372, + 0.019070418551564217, + 0.01489598210901022, + 0.012465503998100758, + 0.027439149096608162, + 0.01364004798233509, + 0.004437230993062258, + 0.007173854857683182, + 0.005932708736509085, + 0.01139161642640829, + 0.005905241705477238, + 0.007186621427536011, + 0.00820824783295393, + 0.005529353395104408, + 0.005948399659246206, + 0.010643230751156807, + 0.011154127307236195, + 0.010986397042870522, + 0.008552193641662598, + 0.007979029789566994, + 0.01051370706409216, + 0.005121208261698484, + 0.007997742854058743, + 0.011276311241090298, + 0.009881362318992615, + 0.014250966720283031, + 0.009659664705395699, + 0.009515859186649323, + 0.010675589554011822, + 0.0069887456484138966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.007630369160324335 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293192207813263 + ], + "opposite_logprob_tokens": [ + -2.237931966781616 + ], + "logprob_reward": -0.11293192207813263, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293192207813263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10473792999982834, + 0.059968557208776474, + 0.045613761991262436, + 0.08672385662794113, + 0.15813376009464264, + 0.23063363134860992, + 0.08022196590900421, + 0.07954216003417969, + 0.11536119133234024, + 0.03906312584877014 + ], + "credit_map": [ + 0.010147662833333015, + 0.023183070123195648, + 0.013938738964498043, + 0.028689589351415634, + 0.009424007497727871, + 0.00608815299347043, + 0.007979744113981724, + 0.005286954343318939, + 0.006180956494063139, + 0.011220728978514671, + 0.014286624267697334, + 0.007648022845387459, + 0.005038264207541943, + 0.006163913756608963, + 0.005333693232387304, + 0.004096352029591799, + 0.006515819113701582, + 0.0060624005272984505, + 0.007869015447795391, + 0.009292633272707462, + 0.003970082849264145, + 0.003601173171773553, + 0.004191296175122261, + 0.004111338406801224, + 0.01140043418854475, + 0.013567310757935047, + 0.01286718063056469, + 0.010314024984836578, + 0.014806091785430908, + 0.006674527190625668, + 0.011240986175835133, + 0.00585330231115222, + 0.02340753935277462, + 0.018861930817365646, + 0.020437339320778847, + 0.03121238946914673, + 0.021714558824896812, + 0.0181496050208807, + 0.013220040127635002, + 0.011130351573228836, + 0.027565594762563705, + 0.03150709345936775, + 0.040446456521749496, + 0.027630778029561043, + 0.031312018632888794, + 0.02384897693991661, + 0.030980294570326805, + 0.017342394217848778, + 0.007702216971665621, + 0.009618464857339859, + 0.010647879913449287, + 0.010875171050429344, + 0.007656728383153677, + 0.009409522637724876, + 0.014773514121770859, + 0.009538467973470688, + 0.006136870477348566, + 0.009629555977880955, + 0.0060773142613470554, + 0.013437656685709953, + 0.008851963095366955, + 0.011010381393134594, + 0.009190321899950504, + 0.015208102762699127, + 0.006794429384171963, + 0.00790251512080431, + 0.024429062381386757, + 0.013672901317477226, + 0.012567054480314255, + 0.013878563418984413, + 0.02979855425655842, + 0.006318117491900921, + 0.0023193913511931896, + 0.002616168698295951, + 0.0028758456464856863, + 0.0034452099353075027, + 0.006306071765720844, + 0.007342254277318716, + 0.0057175797410309315, + 0.008440601639449596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.11293192207813263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054965127259492874 + ], + "opposite_logprob_tokens": [ + -2.929965019226074 + ], + "logprob_reward": -0.054965127259492874, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054965127259492874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.182185098528862, + 0.04577955603599548, + 0.03787703439593315, + 0.07896489650011063, + 0.05238219350576401, + 0.08266877382993698, + 0.1452467441558838, + 0.18556860089302063, + 0.09259071946144104, + 0.09673634171485901 + ], + "credit_map": [ + 0.01966700702905655, + 0.03918648883700371, + 0.026777975261211395, + 0.047548260539770126, + 0.01511120330542326, + 0.015996035188436508, + 0.0077624013647437096, + 0.01013573631644249, + 0.0033264048397541046, + 0.008636299520730972, + 0.008290356956422329, + 0.008654795587062836, + 0.004269175231456757, + 0.007065175101161003, + 0.0033722971566021442, + 0.002165056299418211, + 0.003921941388398409, + 0.005330604501068592, + 0.0050344751216471195, + 0.0065862867049872875, + 0.0063610319048166275, + 0.004554950166493654, + 0.0029700868763029575, + 0.003117663087323308, + 0.009622293524444103, + 0.013623835518956184, + 0.010554498061537743, + 0.008636475540697575, + 0.014006863348186016, + 0.007682465482503176, + 0.009983138181269169, + 0.004855336621403694, + 0.007131044752895832, + 0.006757244002074003, + 0.00624062167480588, + 0.009238213300704956, + 0.00665108859539032, + 0.006961517035961151, + 0.004597305785864592, + 0.004805165808647871, + 0.004936836194247007, + 0.005853394977748394, + 0.01475851982831955, + 0.014843007549643517, + 0.011199587024748325, + 0.00865537766367197, + 0.016409220173954964, + 0.006012842524796724, + 0.013513981364667416, + 0.019132912158966064, + 0.014166014268994331, + 0.022547800093889236, + 0.015554564073681831, + 0.011585528030991554, + 0.03386944159865379, + 0.014876523055136204, + 0.013381212018430233, + 0.019650649279356003, + 0.011961041018366814, + 0.024760153144598007, + 0.024022355675697327, + 0.027870088815689087, + 0.03759674355387688, + 0.02632637694478035, + 0.0057685039937496185, + 0.006430185399949551, + 0.022631697356700897, + 0.01121668890118599, + 0.009022724814713001, + 0.010871033184230328, + 0.02207874320447445, + 0.004571164958178997, + 0.008478128351271152, + 0.008523313328623772, + 0.018275676295161247, + 0.0116216279566288, + 0.022004399448633194, + 0.012349105440080166, + 0.009760398417711258, + 0.005723700858652592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.054965127259492874 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures the scene in a brightly lit modern kitchen. A man with slightly tousled brown hair casually walks toward a set of lower cabinets. He wears a red and white checkered, short-sleeved button-down shirt and contrasting teal checkered shorts, creating a relaxed, homely look. He pauses before a white, rickety cabinet and reaches down deliberately, his fingers grasping the small, brushed metal knob. He smoothly pulls the door open, revealing shelves filled with a messy yet organized pantry: an assortment of unlabeled glass jars, colorful tins, and plain cardboard boxes. Holding a simple glass jar with a plain metal lid in his right hand, he takes a step closer, bending slightly at his knees and waist, and carefully reaches into the cabinet. Steadily and accurately, he places the jar on the bottom shelf, nestled between a stack of jars and a box. Once his fingers release the jar, he straightens and begins to rotate his body to the left. As he begins to turn away from the open cabinet, his movements are fluid and unhurried, completing a simple chore. The scene is bathed in warm, soft light from overhead fixtures, creating a comfortable and tranquil atmosphere.", + "seed": 11774, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000136/group_9/index_5875_seed_11774_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1884199120104313, + "vlm_logprob": -0.1884199120104313, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.1884199120104313, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7694486975669861 + ], + "opposite_logprob_tokens": [ + -0.6444486975669861 + ], + "logprob_reward": -0.7694486975669861, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7694486975669861, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22853058576583862, + 0.17831814289093018, + 0.031004944816231728, + 0.08896981924772263, + 0.10370346158742905, + 0.0912342295050621, + 0.03658919408917427, + 0.06057019904255867, + 0.05977729335427284, + 0.12130216509103775 + ], + "credit_map": [ + 0.011908435262739658, + 0.017378121614456177, + 0.014154156669974327, + 0.01895308494567871, + 0.04707547277212143, + 0.04102940484881401, + 0.04759450629353523, + 0.03043738566339016, + 0.011911696754395962, + 0.016860222443938255, + 0.010500467382371426, + 0.01225297711789608, + 0.02322283200919628, + 0.03973933309316635, + 0.02876516431570053, + 0.035065434873104095, + 0.004886910319328308, + 0.005088912323117256, + 0.0034570482093840837, + 0.0028432561084628105, + 0.002891130279749632, + 0.00488200644031167, + 0.003641158575192094, + 0.0033145202323794365, + 0.010400949977338314, + 0.03540728986263275, + 0.014417176134884357, + 0.008802548982203007, + 0.003979901783168316, + 0.004476313944905996, + 0.0033637124579399824, + 0.008121918886899948, + 0.0138644864782691, + 0.01250026561319828, + 0.010864054784178734, + 0.01510394737124443, + 0.013212229125201702, + 0.01310465857386589, + 0.013770968653261662, + 0.011282836087048054, + 0.009548849426209927, + 0.016046393662691116, + 0.0090335039421916, + 0.012590530328452587, + 0.01071122009307146, + 0.011013383977115154, + 0.01179303415119648, + 0.010497297160327435, + 0.0034523906651884317, + 0.0074214027263224125, + 0.00225046300329268, + 0.006849280092865229, + 0.004075912293046713, + 0.0036197323352098465, + 0.003620800795033574, + 0.005299207754433155, + 0.00565783865749836, + 0.013080203905701637, + 0.004321780521422625, + 0.008143764920532703, + 0.006520002614706755, + 0.007009911350905895, + 0.007655471563339233, + 0.00818121712654829, + 0.005565720144659281, + 0.006441903300583363, + 0.013256847858428955, + 0.004181596916168928, + 0.004747793544083834, + 0.0055184378288686275, + 0.004980255849659443, + 0.015084734186530113, + 0.01589357666671276, + 0.020596209913492203, + 0.019135333597660065, + 0.016746725887060165, + 0.01671416684985161, + 0.009397493675351143, + 0.005926224868744612, + 0.01689242012798786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.7694486975669861 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012577380053699017 + ], + "opposite_logprob_tokens": [ + -4.387577533721924 + ], + "logprob_reward": -0.012577380053699017, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012577380053699017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05138507857918739, + 0.17982643842697144, + 0.044370707124471664, + 0.22931897640228271, + 0.09085718542337418, + 0.06713554263114929, + 0.09483662992715836, + 0.11780578643083572, + 0.05565124377608299, + 0.06881239265203476 + ], + "credit_map": [ + 0.004105913918465376, + 0.00501372991129756, + 0.004378656391054392, + 0.00463853869587183, + 0.010530748404562473, + 0.007850038819015026, + 0.0060943011194467545, + 0.008773146197199821, + 0.029855135828256607, + 0.01683523692190647, + 0.031739741563797, + 0.0273202583193779, + 0.012252714484930038, + 0.020617278292775154, + 0.01917305588722229, + 0.022033018991351128, + 0.003969252575188875, + 0.00348685123026371, + 0.003254795679822564, + 0.010959801264107227, + 0.004318068269640207, + 0.005733068101108074, + 0.006228456739336252, + 0.006420412100851536, + 0.02731127291917801, + 0.08825857937335968, + 0.026704512536525726, + 0.027708454057574272, + 0.010109782218933105, + 0.011320400983095169, + 0.01043515745550394, + 0.027470815926790237, + 0.013800540007650852, + 0.009810395538806915, + 0.010652883909642696, + 0.02179250493645668, + 0.00791305210441351, + 0.008385930210351944, + 0.009978653863072395, + 0.008523222990334034, + 0.008452179841697216, + 0.010177825577557087, + 0.010388735681772232, + 0.014866932295262814, + 0.006465432234108448, + 0.004626655951142311, + 0.006477250251919031, + 0.005680527072399855, + 0.008769537322223186, + 0.019979329779744148, + 0.00994845014065504, + 0.01655568741261959, + 0.008365313522517681, + 0.008271623402833939, + 0.008908278308808804, + 0.014038411900401115, + 0.010094022378325462, + 0.025434089824557304, + 0.0100320465862751, + 0.02133343182504177, + 0.012678270228207111, + 0.009937705472111702, + 0.01408290769904852, + 0.01421330962330103, + 0.004522696603089571, + 0.005422656890004873, + 0.01011388935148716, + 0.006645622197538614, + 0.004214284010231495, + 0.0052382261492311954, + 0.005526603665202856, + 0.01396726444363594, + 0.00563619052991271, + 0.011393468827009201, + 0.011073834262788296, + 0.016971079632639885, + 0.010422509163618088, + 0.005037318915128708, + 0.004978129640221596, + 0.0032998572569340467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 7 + ], + "selected_reward": -0.012577380053699017 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074387468397617 + ], + "opposite_logprob_tokens": [ + -4.511074542999268 + ], + "logprob_reward": -0.011074387468397617, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074387468397617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09080842137336731, + 0.08464157581329346, + 0.08061469346284866, + 0.17849671840667725, + 0.11734168976545334, + 0.11754397302865982, + 0.04046124219894409, + 0.11186940968036652, + 0.06982334703207016, + 0.1083989292383194 + ], + "credit_map": [ + 0.006136642303317785, + 0.008564354851841927, + 0.008433780632913113, + 0.007750967517495155, + 0.016649475321173668, + 0.013785832561552525, + 0.013587052933871746, + 0.015900323167443275, + 0.005437473300844431, + 0.011228062212467194, + 0.004868852440267801, + 0.010178152471780777, + 0.013776532374322414, + 0.012120191007852554, + 0.012227233499288559, + 0.014805082231760025, + 0.006834234111011028, + 0.006746202707290649, + 0.005311240442097187, + 0.012585942633450031, + 0.009878188371658325, + 0.010982627049088478, + 0.013706903904676437, + 0.014569351449608803, + 0.014953737147152424, + 0.06675952672958374, + 0.04086723551154137, + 0.016060180962085724, + 0.00789211131632328, + 0.008591621182858944, + 0.005770446732640266, + 0.017601847648620605, + 0.017986075952649117, + 0.012958689592778683, + 0.013517232611775398, + 0.0253774281591177, + 0.011221333406865597, + 0.01187900174409151, + 0.013746925629675388, + 0.01065500546246767, + 0.012771292589604855, + 0.01820860616862774, + 0.018790945410728455, + 0.021067703142762184, + 0.014637050218880177, + 0.010122262872755527, + 0.010671150870621204, + 0.01127496175467968, + 0.002368162153288722, + 0.006942007690668106, + 0.002639330690726638, + 0.005992056801915169, + 0.007132027763873339, + 0.004189122933894396, + 0.00597034627571702, + 0.005228186957538128, + 0.009493792429566383, + 0.0240404661744833, + 0.008351156488060951, + 0.01514936052262783, + 0.012362383306026459, + 0.01136505976319313, + 0.014035689644515514, + 0.017071498557925224, + 0.007460483815521002, + 0.010335065424442291, + 0.006972193252295256, + 0.007906787097454071, + 0.01018755603581667, + 0.006967529188841581, + 0.012272592633962631, + 0.007721147499978542, + 0.0083081666380167, + 0.019260521978139877, + 0.009911240078508854, + 0.012187322601675987, + 0.02603190392255783, + 0.010598678141832352, + 0.007890337146818638, + 0.014210757799446583 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.011074387468397617 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0700037032365799 + ], + "opposite_logprob_tokens": [ + -2.6950037479400635 + ], + "logprob_reward": -0.0700037032365799, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0700037032365799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054763104766607285, + 0.03786652907729149, + 0.027319353073835373, + 0.18364891409873962, + 0.08177252113819122, + 0.1961655467748642, + 0.05459527671337128, + 0.09812509268522263, + 0.19606249034404755, + 0.06968116760253906 + ], + "credit_map": [ + 0.0028174170292913914, + 0.007547753397375345, + 0.003793196752667427, + 0.0060648531652987, + 0.008882053196430206, + 0.007740152068436146, + 0.0089980848133564, + 0.00891959574073553, + 0.0023861953523010015, + 0.006494517903774977, + 0.002202422358095646, + 0.004847789648920298, + 0.00609212089329958, + 0.005790235474705696, + 0.004775123205035925, + 0.005278128664940596, + 0.0029264274053275585, + 0.0018886409234255552, + 0.0031558312475681305, + 0.0033679974731057882, + 0.0028603789396584034, + 0.0032864147797226906, + 0.004833401646465063, + 0.005000260192900896, + 0.014550743624567986, + 0.07323826849460602, + 0.04141274467110634, + 0.014762241393327713, + 0.008070508949458599, + 0.007182911969721317, + 0.004470493644475937, + 0.019961019977927208, + 0.013826136477291584, + 0.009307542815804482, + 0.009630433283746243, + 0.021183473989367485, + 0.00740171130746603, + 0.008130835369229317, + 0.006764337886124849, + 0.005528053268790245, + 0.019236812368035316, + 0.030478425323963165, + 0.022284843027591705, + 0.04303118586540222, + 0.018275495618581772, + 0.01908605918288231, + 0.02267579361796379, + 0.021096937358379364, + 0.004148660693317652, + 0.009833562187850475, + 0.0044902837835252285, + 0.008075268007814884, + 0.0068310704082250595, + 0.007183455396443605, + 0.006600652821362019, + 0.007432322949171066, + 0.009709049947559834, + 0.02246820740401745, + 0.008126343600451946, + 0.013349941000342369, + 0.009382661432027817, + 0.009150223806500435, + 0.010806165635585785, + 0.015132502652704716, + 0.017762204632163048, + 0.020441457629203796, + 0.040993914008140564, + 0.012454905547201633, + 0.020106632262468338, + 0.015559372492134571, + 0.021694747731089592, + 0.04704924672842026, + 0.006862244103103876, + 0.00884734746068716, + 0.009900441393256187, + 0.0214613638818264, + 0.007492739707231522, + 0.00489769596606493, + 0.00442082853987813, + 0.005798507947474718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.0700037032365799 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a brightly lit modern kitchen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a red and white checkered short-sleeved button-down shirt and teal checkered shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man open a white rickety cabinet to reveal shelves with glass jars, tins, and cardboard boxes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the man place a glass jar with a plain metal lid onto the bottom shelf of the cabinet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in warm, soft light from overhead fixtures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899539172649384 + ], + "opposite_logprob_tokens": [ + -2.5789954662323 + ], + "logprob_reward": -0.07899539172649384, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899539172649384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11911361664533615, + 0.04024052247405052, + 0.05960594490170479, + 0.3719666302204132, + 0.10933033376932144, + 0.0792040303349495, + 0.031363073736429214, + 0.03936424478888512, + 0.03262999653816223, + 0.11718156933784485 + ], + "credit_map": [ + 0.005927469115704298, + 0.010774720460176468, + 0.01126052625477314, + 0.014616888016462326, + 0.018680624663829803, + 0.01628803461790085, + 0.021460700780153275, + 0.02010466530919075, + 0.002891272772103548, + 0.004625516943633556, + 0.002268170937895775, + 0.003675608430057764, + 0.003905327059328556, + 0.006659250240772963, + 0.006541043519973755, + 0.009674333967268467, + 0.0065909787081182, + 0.0045770625583827496, + 0.006231978535652161, + 0.00868113711476326, + 0.007259194739162922, + 0.007224756292998791, + 0.009339369833469391, + 0.009701461531221867, + 0.027821635827422142, + 0.15357211232185364, + 0.07527537643909454, + 0.030264731496572495, + 0.016105225309729576, + 0.014164146035909653, + 0.0096546346321702, + 0.04510876536369324, + 0.015949958935379982, + 0.013392298482358456, + 0.012941443361341953, + 0.02586219273507595, + 0.011486291885375977, + 0.011044599115848541, + 0.01042669452726841, + 0.008226861245930195, + 0.008747496642172337, + 0.013263853266835213, + 0.008634272962808609, + 0.012378126382827759, + 0.00889044813811779, + 0.0071661085821688175, + 0.011208112351596355, + 0.008915610611438751, + 0.0024385282304137945, + 0.0044928742572665215, + 0.0022518483456224203, + 0.004061118699610233, + 0.0048974184319376945, + 0.0052590263076126575, + 0.004146145191043615, + 0.0038161135744303465, + 0.0036573675461113453, + 0.007358585484325886, + 0.003160002874210477, + 0.005482088308781385, + 0.004913961049169302, + 0.004247418139129877, + 0.0036569531075656414, + 0.006887868512421846, + 0.0035342355258762836, + 0.003101508365944028, + 0.007646105717867613, + 0.0028657300863415003, + 0.002762412652373314, + 0.002938917838037014, + 0.002583106281235814, + 0.0071979789063334465, + 0.010918070562183857, + 0.030199915170669556, + 0.015194311738014221, + 0.025231026113033295, + 0.015661275014281273, + 0.007735362742096186, + 0.007343467324972153, + 0.004898143466562033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.07899539172649384 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_0/index_tensor([2462, 2462])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_0/index_tensor([2462, 2462])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..69f13fc812423284e1663dbec9e13f67e6ff6dae --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_0/index_tensor([2462, 2462])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up, captured in a fluid, orbiting motion, focuses on an ancient copper lock cylinder suspended in a pure black void. Cast from bronze, its surface is mottled with age-worn patina and minute scratches. A mysterious blue glow permeates its inner workings, imbuing the cold metal with a breath of life. As the camera slowly pans, the cylinder begins to disintegrate with a grace that defies the laws of physics. Countless delicate gears, springs, pins, and levers float and spread silently outward, like stars imbued with life, forming a dazzling mechanical nebula in the darkness. Each tiny component rotates independently, reflecting the faint glow of its core. Then, as the camera continues to track, the scattered components, as if drawn by an invisible force, begin to converge steadily and deliberately towards the center along a precise trajectory. Gears mesh seamlessly, pins slide precisely into their holes, and finally, perfectly reassemble into a complete lock cylinder. The glow within becomes introverted and profound. The whole process, set against a deep background, creates a mysterious atmosphere of focus, tranquility and full of fantasy mechanical aesthetics.", + "seed": 16563, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_0/index_2462_seed_16563_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09683660222217441, + "vlm_logprob": -0.09683660222217441, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.09683660222217441, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022528549656271935 + ], + "opposite_logprob_tokens": [ + -3.897528648376465 + ], + "logprob_reward": -0.022528549656271935, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.022528549656271935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3005028963088989, + 0.13871000707149506, + 0.11268545687198639, + 0.12051525712013245, + 0.08241382241249084, + 0.06000332161784172, + 0.03332415595650673, + 0.02313506044447422, + 0.09306401759386063, + 0.03564601019024849 + ], + "credit_map": [ + 0.026435066014528275, + 0.04045375809073448, + 0.02401108480989933, + 0.024139704182744026, + 0.03027927130460739, + 0.06875587999820709, + 0.022032542154192924, + 0.06439557671546936, + 0.009073141030967236, + 0.018151504918932915, + 0.0109552638605237, + 0.023418353870511055, + 0.01850060746073723, + 0.02730542980134487, + 0.007719625718891621, + 0.0235860925167799, + 0.007911325432360172, + 0.012204567901790142, + 0.008698069490492344, + 0.010383502580225468, + 0.016381895169615746, + 0.0218606349080801, + 0.010913745500147343, + 0.024331722408533096, + 0.015661869198083878, + 0.018577130511403084, + 0.008034247905015945, + 0.009206031449139118, + 0.016339153051376343, + 0.018619338050484657, + 0.011379923671483994, + 0.02269757352769375, + 0.012237606570124626, + 0.013310105539858341, + 0.008400698192417622, + 0.005975988693535328, + 0.00877225399017334, + 0.00995319988578558, + 0.009771663695573807, + 0.013992313295602798, + 0.006754257716238499, + 0.006627317983657122, + 0.006472459528595209, + 0.0065410262905061245, + 0.004735189955681562, + 0.008588728494942188, + 0.011624258011579514, + 0.0086600910872221, + 0.005049863830208778, + 0.005566976964473724, + 0.004216772969812155, + 0.00295220036059618, + 0.0026766262017190456, + 0.0034666042774915695, + 0.00463090417906642, + 0.0047642081044614315, + 0.002987462328746915, + 0.0036418279632925987, + 0.003216349985450506, + 0.002715486101806164, + 0.0020524882711470127, + 0.0017379831988364458, + 0.0034871166571974754, + 0.003296349197626114, + 0.015675431117415428, + 0.008606324903666973, + 0.018397318199276924, + 0.008786664344370365, + 0.009457462467253208, + 0.008859273977577686, + 0.011320545338094234, + 0.011961004696786404, + 0.005194539669901133, + 0.004182341508567333, + 0.005574936978518963, + 0.00315381633117795, + 0.004359049256891012, + 0.0033698389306664467, + 0.004333407152444124, + 0.0054780845530331135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.022528549656271935 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03373325243592262 + ], + "opposite_logprob_tokens": [ + -3.408733367919922 + ], + "logprob_reward": -0.03373325243592262, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03373325243592262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18492437899112701, + 0.17581726610660553, + 0.12641285359859467, + 0.09138385206460953, + 0.1116911992430687, + 0.1158248633146286, + 0.04331900179386139, + 0.021085545420646667, + 0.08814705908298492, + 0.04139399528503418 + ], + "credit_map": [ + 0.01613682322204113, + 0.027962494641542435, + 0.019977180287241936, + 0.014974781312048435, + 0.02709324099123478, + 0.030676763504743576, + 0.020998945459723473, + 0.027104148641228676, + 0.010025577619671822, + 0.020985523238778114, + 0.013557469472289085, + 0.029126456007361412, + 0.027737900614738464, + 0.03365229070186615, + 0.01046526338905096, + 0.03026679717004299, + 0.009127875789999962, + 0.012965714558959007, + 0.010963760316371918, + 0.008606921881437302, + 0.013597783632576466, + 0.02927137352526188, + 0.008376491256058216, + 0.03350292518734932, + 0.012245755642652512, + 0.014990542083978653, + 0.006644001230597496, + 0.006404040846973658, + 0.011643607169389725, + 0.014089897274971008, + 0.008119119331240654, + 0.017246892675757408, + 0.01775076426565647, + 0.021114878356456757, + 0.010220952332019806, + 0.009565379470586777, + 0.010155374184250832, + 0.013565806671977043, + 0.011391589418053627, + 0.01792646013200283, + 0.010166825726628304, + 0.012819602154195309, + 0.010908578522503376, + 0.014326722361147404, + 0.008243583142757416, + 0.01965351216495037, + 0.022863443940877914, + 0.016842592507600784, + 0.006422603502869606, + 0.007216483820229769, + 0.004953189752995968, + 0.004259803332388401, + 0.0037604111712425947, + 0.00451985327526927, + 0.005687217693775892, + 0.006499439477920532, + 0.002487229648977518, + 0.0026720762252807617, + 0.0026051225140690804, + 0.0030630731489509344, + 0.002188329817727208, + 0.0019344117026776075, + 0.0032249477226287127, + 0.002910355105996132, + 0.016302790492773056, + 0.00812290795147419, + 0.01728961616754532, + 0.008229369297623634, + 0.008370497263967991, + 0.007701720576733351, + 0.010995321907103062, + 0.011134826578199863, + 0.00425077835097909, + 0.0050092763267457485, + 0.0058452473022043705, + 0.005058237351477146, + 0.005013652145862579, + 0.005290791857987642, + 0.006028888281434774, + 0.004897125065326691 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03373325243592262 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015220972709357738 + ], + "opposite_logprob_tokens": [ + -6.501522064208984 + ], + "logprob_reward": -0.0015220972709357738, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015220972709357738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22464331984519958, + 0.0850365161895752, + 0.1115814596414566, + 0.12190838158130646, + 0.13895659148693085, + 0.09854646027088165, + 0.07234562188386917, + 0.030987059697508812, + 0.06680630147457123, + 0.049188293516635895 + ], + "credit_map": [ + 0.018915388733148575, + 0.031780268996953964, + 0.015819642692804337, + 0.01675628498196602, + 0.021315814927220345, + 0.053623586893081665, + 0.01461966335773468, + 0.051812656223773956, + 0.006293754559010267, + 0.011291975155472755, + 0.006149630062282085, + 0.012011905200779438, + 0.010419642552733421, + 0.017519675195217133, + 0.004761764779686928, + 0.016588162630796432, + 0.008683602325618267, + 0.008035252802073956, + 0.010135079734027386, + 0.008252049796283245, + 0.011302376165986061, + 0.024588827043771744, + 0.011818638071417809, + 0.02876562625169754, + 0.01513415202498436, + 0.018408752977848053, + 0.008826157078146935, + 0.008143770508468151, + 0.01687207631766796, + 0.019728846848011017, + 0.012578596360981464, + 0.02221602573990822, + 0.023867439478635788, + 0.022227991372346878, + 0.014783540740609169, + 0.011384371668100357, + 0.012763009406626225, + 0.01696302928030491, + 0.014823295176029205, + 0.022143913432955742, + 0.011326943524181843, + 0.010802291333675385, + 0.008779333904385567, + 0.010657374747097492, + 0.00726799201220274, + 0.01615731604397297, + 0.01878911815583706, + 0.014766087755560875, + 0.016369091346859932, + 0.009537099860608578, + 0.013394668698310852, + 0.006151522509753704, + 0.004928776063024998, + 0.005451398901641369, + 0.006294074933975935, + 0.010218988172709942, + 0.006714504677802324, + 0.003867988707497716, + 0.00454643415287137, + 0.0023941213730722666, + 0.0038092744071036577, + 0.002169949933886528, + 0.003584641497582197, + 0.0039001437835395336, + 0.010495546273887157, + 0.005440101493149996, + 0.013199079781770706, + 0.006255308166146278, + 0.008864331059157848, + 0.007817643694579601, + 0.006688373629003763, + 0.008045910857617855, + 0.004989233799278736, + 0.005157866980880499, + 0.007093985565006733, + 0.0056875357404351234, + 0.006006401032209396, + 0.006046077702194452, + 0.005484756547957659, + 0.008722432889044285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0015220972709357738 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295044422149658 + ], + "opposite_logprob_tokens": [ + -2.237950325012207 + ], + "logprob_reward": -0.11295044422149658, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295044422149658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13874822854995728, + 0.09774497896432877, + 0.11782091856002808, + 0.2038414180278778, + 0.11556529998779297, + 0.11153942346572876, + 0.04962734878063202, + 0.041181836277246475, + 0.05427850782871246, + 0.06965205818414688 + ], + "credit_map": [ + 0.011964636854827404, + 0.020016305148601532, + 0.009674503467977047, + 0.009398517198860645, + 0.012990565039217472, + 0.03272378444671631, + 0.009461180306971073, + 0.03251872956752777, + 0.007287896238267422, + 0.013675599358975887, + 0.008631506003439426, + 0.018934577703475952, + 0.010775436647236347, + 0.016543060541152954, + 0.007303743157535791, + 0.014593160711228848, + 0.007780804298818111, + 0.009516727179288864, + 0.011051231995224953, + 0.009095442481338978, + 0.01415176596492529, + 0.02697259932756424, + 0.011514904908835888, + 0.0277374517172575, + 0.0295218825340271, + 0.037330321967601776, + 0.015063962899148464, + 0.014507091604173183, + 0.026015864685177803, + 0.028325779363512993, + 0.01776975393295288, + 0.03530675545334816, + 0.015356048941612244, + 0.02000730112195015, + 0.012144207954406738, + 0.01028785202652216, + 0.012149873189628124, + 0.014529773965477943, + 0.01270213071256876, + 0.018388114869594574, + 0.01138074230402708, + 0.016831131651997566, + 0.011360320262610912, + 0.012087496928870678, + 0.00891850609332323, + 0.01464557833969593, + 0.020973622798919678, + 0.015342023223638535, + 0.006650483701378107, + 0.0064269378781318665, + 0.006778139155358076, + 0.004933357238769531, + 0.00546115729957819, + 0.007172978948801756, + 0.005317536648362875, + 0.006886757910251617, + 0.008015168830752373, + 0.004880425985902548, + 0.006212815176695585, + 0.00354834645986557, + 0.004214176908135414, + 0.003237893572077155, + 0.005286519881337881, + 0.005786488298326731, + 0.006433582864701748, + 0.005728452932089567, + 0.010227255523204803, + 0.006042041350156069, + 0.007660217583179474, + 0.005359189119189978, + 0.006963545456528664, + 0.005864221602678299, + 0.005595758091658354, + 0.008954230695962906, + 0.009166961535811424, + 0.00851274374872446, + 0.008638397790491581, + 0.010532212443649769, + 0.008330997079610825, + 0.009920750744640827 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.11295044422149658 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134486675262451 + ], + "opposite_logprob_tokens": [ + -1.3134486675262451 + ], + "logprob_reward": -0.3134486675262451, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134486675262451, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17646928131580353, + 0.1016230508685112, + 0.10835669189691544, + 0.04773552343249321, + 0.19072043895721436, + 0.058201346546411514, + 0.0981338769197464, + 0.029074227437376976, + 0.0345979668200016, + 0.15508760511875153 + ], + "credit_map": [ + 0.015672076493501663, + 0.022739531472325325, + 0.01936205103993416, + 0.014864225871860981, + 0.02228868193924427, + 0.026823174208402634, + 0.026207545772194862, + 0.02851199172437191, + 0.008247560821473598, + 0.015288271009922028, + 0.009061305783689022, + 0.018173763528466225, + 0.014794263057410717, + 0.015781870111823082, + 0.006490482948720455, + 0.0137855289503932, + 0.008983940817415714, + 0.010831452906131744, + 0.016620878130197525, + 0.013676793314516544, + 0.011937611736357212, + 0.017821721732616425, + 0.010534395463764668, + 0.017949890345335007, + 0.004701240919530392, + 0.007092780899256468, + 0.0035904236137866974, + 0.0033099583815783262, + 0.005863165017217398, + 0.008581344969570637, + 0.004826809279620647, + 0.009769801050424576, + 0.043791286647319794, + 0.024669120088219643, + 0.016062607988715172, + 0.016622064635157585, + 0.019524281844496727, + 0.020687974989414215, + 0.017050452530384064, + 0.03231264650821686, + 0.011737036518752575, + 0.007738817948848009, + 0.006848281715065241, + 0.005276696290820837, + 0.005498548038303852, + 0.007167045492678881, + 0.005593621637672186, + 0.008341296575963497, + 0.019851980730891228, + 0.011041000485420227, + 0.0165796410292387, + 0.007313106674700975, + 0.008848875761032104, + 0.009635205380618572, + 0.010897894389927387, + 0.013966170139610767, + 0.005242388695478439, + 0.003068487858399749, + 0.0037307399325072765, + 0.0025206152349710464, + 0.003264268161728978, + 0.0033244448713958263, + 0.003971868194639683, + 0.003951415419578552, + 0.003493078052997589, + 0.004766921047121286, + 0.005656693130731583, + 0.0031404350884258747, + 0.005493581295013428, + 0.0038989733438938856, + 0.004420225974172354, + 0.003728054463863373, + 0.022463567554950714, + 0.018506905063986778, + 0.023520369082689285, + 0.022516053169965744, + 0.015543642453849316, + 0.014953911304473877, + 0.01619355008006096, + 0.021389616653323174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.3134486675262451 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, captured in a fluid, orbiting motion, focuses on an ancient copper lock cylinder suspended in a pure black void. Cast from bronze, its surface is mottled with age-worn patina and minute scratches. A mysterious blue glow permeates its inner workings, imbuing the cold metal with a breath of life. As the camera slowly pans, the cylinder begins to disintegrate with a grace that defies the laws of physics. Countless delicate gears, springs, pins, and levers float and spread silently outward, like stars imbued with life, forming a dazzling mechanical nebula in the darkness. Each tiny component rotates independently, reflecting the faint glow of its core. Then, as the camera continues to track, the scattered components, as if drawn by an invisible force, begin to converge steadily and deliberately towards the center along a precise trajectory. Gears mesh seamlessly, pins slide precisely into their holes, and finally, perfectly reassemble into a complete lock cylinder. The glow within becomes introverted and profound. The whole process, set against a deep background, creates a mysterious atmosphere of focus, tranquility and full of fantasy mechanical aesthetics.", + "seed": 16564, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_0/index_2462_seed_16564_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0736379996407777, + "vlm_logprob": -0.0736379996407777, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.0736379996407777, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06561590731143951 + ], + "opposite_logprob_tokens": [ + -2.8156158924102783 + ], + "logprob_reward": -0.06561590731143951, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06561590731143951, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10519421845674515, + 0.165246844291687, + 0.05608046054840088, + 0.3294197618961334, + 0.029367579147219658, + 0.021528545767068863, + 0.031207561492919922, + 0.16692547500133514, + 0.016061145812273026, + 0.07896844297647476 + ], + "credit_map": [ + 0.012617718428373337, + 0.006638025399297476, + 0.022076647728681564, + 0.006451970431953669, + 0.004274864215403795, + 0.01290555763989687, + 0.010932529345154762, + 0.029296908527612686, + 0.018350135535001755, + 0.026148905977606773, + 0.01133786141872406, + 0.015560817904770374, + 0.027532683685421944, + 0.023606402799487114, + 0.019063714891672134, + 0.02364632487297058, + 0.004411780275404453, + 0.004112157505005598, + 0.004966911859810352, + 0.0034231487661600113, + 0.005143712740391493, + 0.014844950288534164, + 0.0035117247607558966, + 0.015666067600250244, + 0.010907596908509731, + 0.008531256578862667, + 0.0209046620875597, + 0.010633567348122597, + 0.007715195883065462, + 0.1423507183790207, + 0.0065119448117911816, + 0.12186479568481445, + 0.0029181509744375944, + 0.0029669112991541624, + 0.005786329973489046, + 0.004720632452517748, + 0.0026982491835951805, + 0.004333357326686382, + 0.0026673441752791405, + 0.003276602830737829, + 0.0018152116099372506, + 0.004314900375902653, + 0.00331883504986763, + 0.001800517551600933, + 0.0016580282244831324, + 0.003268508706241846, + 0.0016595894703641534, + 0.0036929536145180464, + 0.004199954215437174, + 0.0044713388197124004, + 0.004810539074242115, + 0.003354820888489485, + 0.0026112261693924665, + 0.0045260414481163025, + 0.0019693023059517145, + 0.005264338105916977, + 0.015972932800650597, + 0.016966715455055237, + 0.01823669858276844, + 0.01513929944485426, + 0.033302657306194305, + 0.02470838837325573, + 0.023167124018073082, + 0.01943165808916092, + 0.0022364668548107147, + 0.0010427974630147219, + 0.002029523951932788, + 0.0016287622274830937, + 0.0021215006709098816, + 0.002121253637596965, + 0.0031136495526880026, + 0.0017671912210062146, + 0.016366302967071533, + 0.0093366215005517, + 0.00902873557060957, + 0.004571403376758099, + 0.011588801629841328, + 0.011923681013286114, + 0.006799961905926466, + 0.009352934546768665 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.06561590731143951 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637348137795925 + ], + "opposite_logprob_tokens": [ + -3.6513733863830566 + ], + "logprob_reward": -0.02637348137795925, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637348137795925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10253266245126724, + 0.2244030088186264, + 0.05578712746500969, + 0.22424103319644928, + 0.0603790283203125, + 0.030454149469733238, + 0.04385538399219513, + 0.2030395120382309, + 0.01615339145064354, + 0.03915473073720932 + ], + "credit_map": [ + 0.013957837596535683, + 0.008222105912864208, + 0.012638107873499393, + 0.009025105275213718, + 0.006846284028142691, + 0.01802780292928219, + 0.018883133307099342, + 0.014932284131646156, + 0.025685990229249, + 0.030922412872314453, + 0.01997821405529976, + 0.018817994743585587, + 0.030906163156032562, + 0.03971987962722778, + 0.02362680993974209, + 0.03474557027220726, + 0.00660276971757412, + 0.0075931428000330925, + 0.0057905991561710835, + 0.005717793013900518, + 0.006486073601990938, + 0.008676927536725998, + 0.005586592014878988, + 0.009333230555057526, + 0.007361563388258219, + 0.007671314291656017, + 0.012588771991431713, + 0.009904971346259117, + 0.008252115920186043, + 0.0917009562253952, + 0.009577854536473751, + 0.07718348503112793, + 0.0071934424340724945, + 0.005640439689159393, + 0.01078199502080679, + 0.008443568833172321, + 0.004988592118024826, + 0.008495958521962166, + 0.006436658091843128, + 0.00839837547391653, + 0.002830656711012125, + 0.007053175941109657, + 0.004880536813288927, + 0.0027392178308218718, + 0.0019564577378332615, + 0.004589477088302374, + 0.002085696440190077, + 0.004318928811699152, + 0.003504885593429208, + 0.006141585297882557, + 0.007175964768975973, + 0.004045861307531595, + 0.004580693785101175, + 0.006169716827571392, + 0.005930160637944937, + 0.006306517869234085, + 0.019521698355674744, + 0.021566886454820633, + 0.021368494257330894, + 0.017960229888558388, + 0.0431344173848629, + 0.029798541218042374, + 0.028446007519960403, + 0.02124324068427086, + 0.004158850759267807, + 0.0011973610380664468, + 0.0021393809001892805, + 0.0018034176900982857, + 0.0015657851472496986, + 0.0023651032242923975, + 0.001577557297423482, + 0.0013459344627335668, + 0.0062795160338282585, + 0.0038494043983519077, + 0.004993177484720945, + 0.0031364981550723314, + 0.004720150958746672, + 0.005627837032079697, + 0.004672171548008919, + 0.005875976290553808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 7 + ], + "selected_reward": -0.02637348137795925 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008207766804844141 + ], + "opposite_logprob_tokens": [ + -7.125820636749268 + ], + "logprob_reward": -0.0008207766804844141, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008207766804844141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05032319203019142, + 0.218904510140419, + 0.042068831622600555, + 0.24202312529087067, + 0.029983775690197945, + 0.027457797899842262, + 0.09563237428665161, + 0.1248217523097992, + 0.019500834867358208, + 0.14928381145000458 + ], + "credit_map": [ + 0.008753054775297642, + 0.00504555506631732, + 0.006712461821734905, + 0.004592886194586754, + 0.0029664982575923204, + 0.009354835376143456, + 0.004590573254972696, + 0.00830733124166727, + 0.02885434404015541, + 0.02544706128537655, + 0.0235360786318779, + 0.015171565115451813, + 0.025737011805176735, + 0.045613501220941544, + 0.01927485130727291, + 0.03527011349797249, + 0.004231715574860573, + 0.004103632178157568, + 0.003139094216749072, + 0.0029357702005654573, + 0.00506143132224679, + 0.009198663756251335, + 0.003203013679012656, + 0.010195511393249035, + 0.00756332091987133, + 0.0038451740983873606, + 0.012340941466391087, + 0.006429745815694332, + 0.0045795803889632225, + 0.1071760430932045, + 0.004940373357385397, + 0.09514795243740082, + 0.003010219195857644, + 0.0034265753347426653, + 0.006021624896675348, + 0.002967407461255789, + 0.003344385651871562, + 0.005097526125609875, + 0.0025139464996755123, + 0.0036020921543240547, + 0.002532519865781069, + 0.0038582617416977882, + 0.004915457684546709, + 0.0035585900768637657, + 0.002297036349773407, + 0.003817621385678649, + 0.0021991722751408815, + 0.0042791408486664295, + 0.012384693138301373, + 0.01408978458493948, + 0.012928815558552742, + 0.00938104186207056, + 0.008009598590433598, + 0.014220193959772587, + 0.007220997009426355, + 0.017397264018654823, + 0.011931875720620155, + 0.013171117752790451, + 0.013049927540123463, + 0.010831693187355995, + 0.02588745951652527, + 0.01855654828250408, + 0.01767122559249401, + 0.01372191496193409, + 0.004513754975050688, + 0.0012756618671119213, + 0.002518868539482355, + 0.0017510299803689122, + 0.0020501124672591686, + 0.0029556197114288807, + 0.002432681620121002, + 0.0020031083840876818, + 0.032307226210832596, + 0.018554089590907097, + 0.01614649035036564, + 0.008190442807972431, + 0.021387359127402306, + 0.02275480329990387, + 0.012427382171154022, + 0.01751602254807949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.0008207766804844141 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023284580558538437 + ], + "opposite_logprob_tokens": [ + -3.773284673690796 + ], + "logprob_reward": -0.023284580558538437, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023284580558538437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08462463319301605, + 0.2071867436170578, + 0.054863303899765015, + 0.1429172307252884, + 0.04125978425145149, + 0.04218221455812454, + 0.1911519318819046, + 0.07201528549194336, + 0.037551384419202805, + 0.12624743580818176 + ], + "credit_map": [ + 0.011651868931949139, + 0.009737729094922543, + 0.01438180636614561, + 0.005802929867058992, + 0.006046090740710497, + 0.01093688141554594, + 0.009179025888442993, + 0.016888301819562912, + 0.024670200422406197, + 0.021749600768089294, + 0.022483257576823235, + 0.013363748788833618, + 0.02135554328560829, + 0.04726947844028473, + 0.020688854157924652, + 0.03560606762766838, + 0.005618746858090162, + 0.0055532739497721195, + 0.004130195360630751, + 0.004551890771836042, + 0.005762156099081039, + 0.010593299753963947, + 0.007006252184510231, + 0.011647489853203297, + 0.011394078843295574, + 0.010328652337193489, + 0.01941252313554287, + 0.021795092150568962, + 0.007988707162439823, + 0.0313570573925972, + 0.007183445151895285, + 0.0334576740860939, + 0.004950986243784428, + 0.0038903746753931046, + 0.007910960353910923, + 0.0039167446084320545, + 0.00403839023783803, + 0.006218690425157547, + 0.0034433023538440466, + 0.006890335585922003, + 0.005090195219963789, + 0.006604035384953022, + 0.0077441358007490635, + 0.004511674400418997, + 0.0029394212178885937, + 0.005997460335493088, + 0.0027755333576351404, + 0.006519755814224482, + 0.02360810711979866, + 0.02444448694586754, + 0.021583054214715958, + 0.022648002952337265, + 0.01622900366783142, + 0.027667110785841942, + 0.022140473127365112, + 0.03283169865608215, + 0.009094617329537868, + 0.006375852040946484, + 0.0078073712065815926, + 0.0057263243943452835, + 0.012679960578680038, + 0.011968717910349369, + 0.008266210556030273, + 0.010096237063407898, + 0.00677384901791811, + 0.0031250889878720045, + 0.0037001976743340492, + 0.005284492392092943, + 0.005740968510508537, + 0.004796232096850872, + 0.0042058550752699375, + 0.003924702759832144, + 0.014835869893431664, + 0.01331212930381298, + 0.017911329865455627, + 0.00978111382573843, + 0.01959902234375477, + 0.018853748217225075, + 0.013219988904893398, + 0.018734237179160118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.023284580558538437 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520952522754669 + ], + "opposite_logprob_tokens": [ + -1.5020952224731445 + ], + "logprob_reward": -0.2520952522754669, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520952522754669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07327474653720856, + 0.10375051200389862, + 0.09331716597080231, + 0.14675773680210114, + 0.04099973663687706, + 0.04916173964738846, + 0.16709846258163452, + 0.06552288681268692, + 0.06365473568439484, + 0.19646230340003967 + ], + "credit_map": [ + 0.012767452746629715, + 0.006776791997253895, + 0.010186544619500637, + 0.007635198533535004, + 0.004995762836188078, + 0.010132379829883575, + 0.009866621345281601, + 0.010913997888565063, + 0.012545737437903881, + 0.013326913118362427, + 0.008793727494776249, + 0.010186211206018925, + 0.01479482650756836, + 0.01809725910425186, + 0.010595797561109066, + 0.015410037711262703, + 0.008193458430469036, + 0.006725873798131943, + 0.007171765435487032, + 0.006531837396323681, + 0.00706533994525671, + 0.024651067331433296, + 0.009942673146724701, + 0.02303515002131462, + 0.010226123966276646, + 0.01787872239947319, + 0.020181788131594658, + 0.030548935756087303, + 0.01321813277900219, + 0.017525365576148033, + 0.013407189399003983, + 0.02377147413790226, + 0.005266592372208834, + 0.005120276007801294, + 0.006507331971079111, + 0.004767342936247587, + 0.004062647465616465, + 0.005905705038458109, + 0.0032803677022457123, + 0.00608947454020381, + 0.005031600594520569, + 0.0058358414098620415, + 0.011024652048945427, + 0.004693755879998207, + 0.00342193222604692, + 0.007587389554828405, + 0.003274827031418681, + 0.008291740901768208, + 0.020835500210523605, + 0.02024454064667225, + 0.022499287500977516, + 0.019110746681690216, + 0.012668639421463013, + 0.026638561859726906, + 0.014983918517827988, + 0.03011726401746273, + 0.008214855566620827, + 0.006733946967869997, + 0.0068867686204612255, + 0.005685778800398111, + 0.011729910038411617, + 0.010040435940027237, + 0.007873758673667908, + 0.008357432670891285, + 0.015294517390429974, + 0.00380687671713531, + 0.007311089430004358, + 0.006839504931122065, + 0.00760825676843524, + 0.00864011887460947, + 0.007946889847517014, + 0.006207481026649475, + 0.043256472796201706, + 0.023836877197027206, + 0.021406127139925957, + 0.010411390103399754, + 0.028360635042190552, + 0.029494909569621086, + 0.016703497618436813, + 0.022992393001914024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.2520952522754669 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, captured in a fluid, orbiting motion, focuses on an ancient copper lock cylinder suspended in a pure black void. Cast from bronze, its surface is mottled with age-worn patina and minute scratches. A mysterious blue glow permeates its inner workings, imbuing the cold metal with a breath of life. As the camera slowly pans, the cylinder begins to disintegrate with a grace that defies the laws of physics. Countless delicate gears, springs, pins, and levers float and spread silently outward, like stars imbued with life, forming a dazzling mechanical nebula in the darkness. Each tiny component rotates independently, reflecting the faint glow of its core. Then, as the camera continues to track, the scattered components, as if drawn by an invisible force, begin to converge steadily and deliberately towards the center along a precise trajectory. Gears mesh seamlessly, pins slide precisely into their holes, and finally, perfectly reassemble into a complete lock cylinder. The glow within becomes introverted and profound. The whole process, set against a deep background, creates a mysterious atmosphere of focus, tranquility and full of fantasy mechanical aesthetics.", + "seed": 16565, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_0/index_2462_seed_16565_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11470505781471729, + "vlm_logprob": -0.11470505781471729, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.11470505781471729, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2303302139043808 + ], + "opposite_logprob_tokens": [ + -1.605330228805542 + ], + "logprob_reward": -0.2303302139043808, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2303302139043808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2249341458082199, + 0.1486700177192688, + 0.052900686860084534, + 0.0876321941614151, + 0.04506422579288483, + 0.10702523589134216, + 0.06961171329021454, + 0.025727413594722748, + 0.05771501734852791, + 0.18071934580802917 + ], + "credit_map": [ + 0.026704953983426094, + 0.042612865567207336, + 0.021742943674325943, + 0.024163221940398216, + 0.022061701864004135, + 0.03274118900299072, + 0.02863651141524315, + 0.026270750910043716, + 0.02919001318514347, + 0.013110527768731117, + 0.01858440227806568, + 0.009292862378060818, + 0.01132506225258112, + 0.025467460975050926, + 0.015168572776019573, + 0.026531126350164413, + 0.004863406531512737, + 0.005185083951801062, + 0.009669497609138489, + 0.004113622475415468, + 0.0048035322688519955, + 0.010368154384195805, + 0.003940131515264511, + 0.009957256726920605, + 0.007497834973037243, + 0.007868129760026932, + 0.009405084885656834, + 0.007407324854284525, + 0.009488244540989399, + 0.012468584813177586, + 0.01400515716522932, + 0.019491838291287422, + 0.00416191341355443, + 0.006681936793029308, + 0.006212673615664244, + 0.005776165518909693, + 0.005560901947319508, + 0.00561239430680871, + 0.006230562459677458, + 0.00482767540961504, + 0.004063189961016178, + 0.007662721909582615, + 0.0038926703855395317, + 0.004992110654711723, + 0.00500120734795928, + 0.03793508559465408, + 0.005449463613331318, + 0.038028787821531296, + 0.008747796528041363, + 0.008334813639521599, + 0.0123368501663208, + 0.0062175034545362, + 0.008959004655480385, + 0.009108652360737324, + 0.00839068740606308, + 0.007516403216868639, + 0.00461095105856657, + 0.0030795629136264324, + 0.002848029835149646, + 0.002025774447247386, + 0.0027260344941169024, + 0.004257815890014172, + 0.0019499545451253653, + 0.004229290410876274, + 0.009459824301302433, + 0.0032911240123212337, + 0.00848563015460968, + 0.005414693616330624, + 0.003196900011971593, + 0.011106055229902267, + 0.005542647559195757, + 0.011218145489692688, + 0.014060216024518013, + 0.00864277221262455, + 0.022562319412827492, + 0.019504835829138756, + 0.008953566662967205, + 0.05032450333237648, + 0.01287919469177723, + 0.04379193112254143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.2303302139043808 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500563606619835 + ], + "opposite_logprob_tokens": [ + -2.9300055503845215 + ], + "logprob_reward": -0.05500563606619835, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500563606619835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25728994607925415, + 0.08685780316591263, + 0.05220916494727135, + 0.04891888052225113, + 0.04033726081252098, + 0.11239643394947052, + 0.17360149323940277, + 0.03201979398727417, + 0.026513773947954178, + 0.16985538601875305 + ], + "credit_map": [ + 0.020763032138347626, + 0.04567236825823784, + 0.025154070928692818, + 0.019074557349085808, + 0.03518021106719971, + 0.03718528896570206, + 0.033416371792554855, + 0.04084405303001404, + 0.015376524068415165, + 0.011967062950134277, + 0.009062185883522034, + 0.010620318353176117, + 0.008960279636085033, + 0.011365631595253944, + 0.009708143770694733, + 0.009797659702599049, + 0.005952497012913227, + 0.005851665511727333, + 0.011016413569450378, + 0.00864582508802414, + 0.004242955707013607, + 0.006368027068674564, + 0.004769507329910994, + 0.005362276919186115, + 0.0038930345326662064, + 0.0061134579591453075, + 0.0045707328245043755, + 0.0042646899819374084, + 0.008247259072959423, + 0.0058477334678173065, + 0.009569288231432438, + 0.006412681192159653, + 0.0036908360198140144, + 0.0051764617674052715, + 0.005118860863149166, + 0.004104166757315397, + 0.005234118085354567, + 0.005998339504003525, + 0.00469979178160429, + 0.006314687896519899, + 0.004552528262138367, + 0.00619502505287528, + 0.004345100838690996, + 0.0058166091330349445, + 0.004171090666204691, + 0.04125552251935005, + 0.005223271436989307, + 0.04083729535341263, + 0.02603800594806671, + 0.0252422746270895, + 0.017537258565425873, + 0.01586180180311203, + 0.022830914705991745, + 0.022947082296013832, + 0.021339958533644676, + 0.021804211661219597, + 0.003906799480319023, + 0.004587946925312281, + 0.003294982248917222, + 0.004930073395371437, + 0.004711826331913471, + 0.003785923356190324, + 0.0026897876523435116, + 0.00411245645955205, + 0.004318705759942532, + 0.002500420669093728, + 0.005229639355093241, + 0.0026202464941889048, + 0.002380084479227662, + 0.003772084368392825, + 0.0026027089916169643, + 0.0030898849945515394, + 0.012389109469950199, + 0.005527610890567303, + 0.04820248484611511, + 0.009884477593004704, + 0.004482124466449022, + 0.06170804053544998, + 0.007390543818473816, + 0.02027099020779133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.05500563606619835 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022066310048103333 + ], + "opposite_logprob_tokens": [ + -6.127206802368164 + ], + "logprob_reward": -0.0022066310048103333, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022066310048103333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0994335487484932, + 0.13478079438209534, + 0.09003911912441254, + 0.07992759346961975, + 0.022857511416077614, + 0.26466014981269836, + 0.15954351425170898, + 0.022817742079496384, + 0.038718175143003464, + 0.08722180873155594 + ], + "credit_map": [ + 0.010654437355697155, + 0.022315552458167076, + 0.008627721108496189, + 0.011440138332545757, + 0.011491751298308372, + 0.01196705736219883, + 0.013770625926554203, + 0.009166264906525612, + 0.029799634590744972, + 0.01301196962594986, + 0.018436387181282043, + 0.010897132568061352, + 0.009466622024774551, + 0.018451889976859093, + 0.01271695178002119, + 0.022000202909111977, + 0.00685308501124382, + 0.006042061373591423, + 0.027373891323804855, + 0.005904858931899071, + 0.003995220642536879, + 0.02474130131304264, + 0.005565067287534475, + 0.00956363882869482, + 0.007254971191287041, + 0.005902869626879692, + 0.011878772638738155, + 0.00582923786714673, + 0.007109304424375296, + 0.011878075078129768, + 0.013989381492137909, + 0.016084972769021988, + 0.002275691367685795, + 0.003069633385166526, + 0.0036653883289545774, + 0.002056080847978592, + 0.0025121027138084173, + 0.0031896696891635656, + 0.002854278078302741, + 0.003234667470678687, + 0.00801544263958931, + 0.007963395677506924, + 0.009017650969326496, + 0.008037189021706581, + 0.007709777448326349, + 0.10869568586349487, + 0.009475672617554665, + 0.1057453453540802, + 0.02394231967628002, + 0.024145208299160004, + 0.015287351794540882, + 0.0137092936784029, + 0.021464595571160316, + 0.020630473271012306, + 0.020809777081012726, + 0.019554510712623596, + 0.0022113104350864887, + 0.0023014782927930355, + 0.002723144367337227, + 0.002089032204821706, + 0.0017632150556892157, + 0.00458862679079175, + 0.002555995713919401, + 0.004584940616041422, + 0.005699596367776394, + 0.00288791093043983, + 0.0073759411461651325, + 0.006323415320366621, + 0.0029102363623678684, + 0.005114035680890083, + 0.004204615950584412, + 0.004202418494969606, + 0.007874307222664356, + 0.003814764553681016, + 0.02072477526962757, + 0.00851156935095787, + 0.0035404374357312918, + 0.025319794192910194, + 0.00569605827331543, + 0.011740104295313358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.0022066310048103333 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637083038687706 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.004637083038687706, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637083038687706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08058955520391464, + 0.13902714848518372, + 0.04699896648526192, + 0.08257302641868591, + 0.034518562257289886, + 0.2566489577293396, + 0.11218183487653732, + 0.023829635232686996, + 0.08780580013990402, + 0.1358264535665512 + ], + "credit_map": [ + 0.009907545521855354, + 0.01519466470927, + 0.008406243287026882, + 0.005384763237088919, + 0.010275325737893581, + 0.01150905154645443, + 0.01057406049221754, + 0.009337907657027245, + 0.028206095099449158, + 0.018245844170451164, + 0.016340170055627823, + 0.013784305192530155, + 0.013563579879701138, + 0.01706261746585369, + 0.0157895740121603, + 0.016034970059990883, + 0.008812541142106056, + 0.0066391900181770325, + 0.009133420884609222, + 0.003909142222255468, + 0.0030715824104845524, + 0.006786849815398455, + 0.0039778780192136765, + 0.00466836616396904, + 0.009683957323431969, + 0.009229757823050022, + 0.010121624916791916, + 0.0061262608505785465, + 0.01137827429920435, + 0.011098223738372326, + 0.008321123197674751, + 0.016613811254501343, + 0.0030521841254085302, + 0.006038145627826452, + 0.003216506214812398, + 0.003778151934966445, + 0.005294281058013439, + 0.004628462716937065, + 0.0036372567992657423, + 0.0048735798336565495, + 0.007034520152956247, + 0.013466152362525463, + 0.006586757022887468, + 0.008814924396574497, + 0.008158041164278984, + 0.10231572389602661, + 0.007617327384650707, + 0.10265550762414932, + 0.014607887715101242, + 0.019786575809121132, + 0.010297607630491257, + 0.008595236577093601, + 0.018166329711675644, + 0.013583173044025898, + 0.013460331596434116, + 0.013684692792594433, + 0.0028308038599789143, + 0.002526012482121587, + 0.0028699233662337065, + 0.0027494519017636776, + 0.0022755006793886423, + 0.003968673292547464, + 0.0025500718038529158, + 0.004059201572090387, + 0.015546334907412529, + 0.005133908241987228, + 0.013627659529447556, + 0.01089877262711525, + 0.0068666026927530766, + 0.014417411759495735, + 0.00942602101713419, + 0.011889094486832619, + 0.011404532939195633, + 0.005653783679008484, + 0.030929628759622574, + 0.01235225610435009, + 0.005238719284534454, + 0.04096971079707146, + 0.010126973502337933, + 0.019150855019688606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 9 + ], + "selected_reward": -0.004637083038687706 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813457250595093 + ], + "opposite_logprob_tokens": [ + -1.4063457250595093 + ], + "logprob_reward": -0.2813457250595093, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813457250595093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16075626015663147, + 0.19731082022190094, + 0.1448425054550171, + 0.0749959647655487, + 0.039706140756607056, + 0.133988156914711, + 0.04248526692390442, + 0.0423218309879303, + 0.058027513325214386, + 0.10556559264659882 + ], + "credit_map": [ + 0.02559417486190796, + 0.013247765600681305, + 0.021388370543718338, + 0.014402521774172783, + 0.012623020447790623, + 0.027824439108371735, + 0.023369066417217255, + 0.02230687625706196, + 0.03551178425550461, + 0.019415387883782387, + 0.02869095467031002, + 0.018867624923586845, + 0.01632644794881344, + 0.02952355146408081, + 0.021009409800171852, + 0.027965648099780083, + 0.015214573591947556, + 0.014888627454638481, + 0.022884303703904152, + 0.015341280959546566, + 0.012825490906834602, + 0.01805824041366577, + 0.015942849218845367, + 0.02968713641166687, + 0.007576297968626022, + 0.007853908464312553, + 0.011947949416935444, + 0.00676472345367074, + 0.008934017270803452, + 0.010256013832986355, + 0.011455625295639038, + 0.010207422077655792, + 0.0028162302915006876, + 0.007496108300983906, + 0.006561434827744961, + 0.0037338922265917063, + 0.004613462369889021, + 0.00462326081469655, + 0.004184933844953775, + 0.0056768134236335754, + 0.005440841428935528, + 0.006101881619542837, + 0.0070188953541219234, + 0.00445103133097291, + 0.0058469087816774845, + 0.048751458525657654, + 0.006026384420692921, + 0.05035074055194855, + 0.00599314458668232, + 0.005154626443982124, + 0.005818261299282312, + 0.002810586942359805, + 0.00492863729596138, + 0.006339600775390863, + 0.005775112193077803, + 0.005665290635079145, + 0.005793606396764517, + 0.004061708692461252, + 0.004293334204703569, + 0.0044136615470051765, + 0.005374368280172348, + 0.007172572426497936, + 0.004643249791115522, + 0.006569326389580965, + 0.008255712687969208, + 0.004451198037713766, + 0.01002577506005764, + 0.008367905393242836, + 0.0037739581894129515, + 0.008042586036026478, + 0.005907519720494747, + 0.00920285377651453, + 0.007078671362251043, + 0.0024106139317154884, + 0.03035879135131836, + 0.0035461322404444218, + 0.002142466139048338, + 0.04011198878288269, + 0.0036033946089446545, + 0.016313528642058372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.2813457250595093 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, captured in a fluid, orbiting motion, focuses on an ancient copper lock cylinder suspended in a pure black void. Cast from bronze, its surface is mottled with age-worn patina and minute scratches. A mysterious blue glow permeates its inner workings, imbuing the cold metal with a breath of life. As the camera slowly pans, the cylinder begins to disintegrate with a grace that defies the laws of physics. Countless delicate gears, springs, pins, and levers float and spread silently outward, like stars imbued with life, forming a dazzling mechanical nebula in the darkness. Each tiny component rotates independently, reflecting the faint glow of its core. Then, as the camera continues to track, the scattered components, as if drawn by an invisible force, begin to converge steadily and deliberately towards the center along a precise trajectory. Gears mesh seamlessly, pins slide precisely into their holes, and finally, perfectly reassemble into a complete lock cylinder. The glow within becomes introverted and profound. The whole process, set against a deep background, creates a mysterious atmosphere of focus, tranquility and full of fantasy mechanical aesthetics.", + "seed": 16566, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_0/index_2462_seed_16566_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2500633244402707, + "vlm_logprob": -0.2500633244402707, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.2500633244402707, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0550473928451538 + ], + "opposite_logprob_tokens": [ + -0.4300473928451538 + ], + "logprob_reward": -1.0550473928451538, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0550473928451538, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07535997033119202, + 0.059614844620227814, + 0.512625515460968, + 0.1538570076227188, + 0.02197725884616375, + 0.03645065426826477, + 0.01428285613656044, + 0.022679796442389488, + 0.023273708298802376, + 0.07987837493419647 + ], + "credit_map": [ + 0.006710527930408716, + 0.013307676650583744, + 0.008448494598269463, + 0.007467637769877911, + 0.010620451532304287, + 0.009568375535309315, + 0.009757990948855877, + 0.009478812105953693, + 0.0036413688212633133, + 0.005894871894270182, + 0.010008897632360458, + 0.004473939538002014, + 0.004942058119922876, + 0.012327237986028194, + 0.006937829311937094, + 0.011388644576072693, + 0.04330763965845108, + 0.040688179433345795, + 0.11801017820835114, + 0.06020740419626236, + 0.03431571647524834, + 0.05421556904911995, + 0.0787859708070755, + 0.08309484273195267, + 0.006489178165793419, + 0.007259049452841282, + 0.017710743471980095, + 0.014337859116494656, + 0.013443771749734879, + 0.03568581864237785, + 0.03305714949965477, + 0.025873439386487007, + 0.0015200768830254674, + 0.0028011444956064224, + 0.002778042107820511, + 0.0025961913634091616, + 0.002481423784047365, + 0.003395475447177887, + 0.0032110337633639574, + 0.003193870186805725, + 0.0030030556954443455, + 0.0038960317615419626, + 0.003976230043917894, + 0.0030074245296418667, + 0.003114894963800907, + 0.00652647390961647, + 0.0028148514684289694, + 0.010111688636243343, + 0.0013898739125579596, + 0.0017419374780729413, + 0.001050034537911415, + 0.0012792168417945504, + 0.001683042966760695, + 0.0022547414992004633, + 0.001921771327033639, + 0.002962237922474742, + 0.0026593448128551245, + 0.0021575200371444225, + 0.002499188994988799, + 0.002302763517946005, + 0.0016864999197423458, + 0.0034688543528318405, + 0.0029559778049588203, + 0.004949647467583418, + 0.0023349991533905268, + 0.004194285254925489, + 0.0018794722855091095, + 0.0030125847551971674, + 0.0021750035230070353, + 0.0024620110634714365, + 0.0029250045772641897, + 0.004290347918868065, + 0.007398385554552078, + 0.025433005765080452, + 0.007015387527644634, + 0.002281522611156106, + 0.005878150928765535, + 0.007154237478971481, + 0.0020893653854727745, + 0.022628316655755043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 9 + ], + "selected_reward": -1.0550473928451538 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0233369879424572 + ], + "opposite_logprob_tokens": [ + -3.773336887359619 + ], + "logprob_reward": -0.0233369879424572, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0233369879424572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2643049955368042, + 0.06906240433454514, + 0.12522414326667786, + 0.17117613554000854, + 0.09090417623519897, + 0.04175526648759842, + 0.0424644909799099, + 0.029784299433231354, + 0.04184715822339058, + 0.12347689270973206 + ], + "credit_map": [ + 0.049422599375247955, + 0.03243778645992279, + 0.024962367489933968, + 0.019320320338010788, + 0.03459307178854942, + 0.040940042585134506, + 0.018717341125011444, + 0.04391147196292877, + 0.00451216846704483, + 0.008132848888635635, + 0.013742604292929173, + 0.005351305939257145, + 0.005544876214116812, + 0.009308569133281708, + 0.0077200219966471195, + 0.014750010333955288, + 0.010863137431442738, + 0.01018330454826355, + 0.02798517420887947, + 0.0149388974532485, + 0.008506495505571365, + 0.012814732268452644, + 0.02092374674975872, + 0.019008656963706017, + 0.008012623526155949, + 0.0087121045216918, + 0.018025634810328484, + 0.016122248023748398, + 0.016339393332600594, + 0.0401785634458065, + 0.035015206784009933, + 0.02877037040889263, + 0.01257014088332653, + 0.012768193148076534, + 0.012344889342784882, + 0.0071249413304030895, + 0.012658358551561832, + 0.012641502544283867, + 0.007667279336601496, + 0.013128869235515594, + 0.0027619798202067614, + 0.006302795838564634, + 0.004034174606204033, + 0.004403237719088793, + 0.004203554708510637, + 0.006997986230999231, + 0.0034704345744103193, + 0.00958110298961401, + 0.004064387641847134, + 0.007208181079477072, + 0.004842371214181185, + 0.00650066277012229, + 0.00417314050719142, + 0.004916567821055651, + 0.004719456657767296, + 0.006039722822606564, + 0.003924206364899874, + 0.0042441366240382195, + 0.003244153456762433, + 0.003605964593589306, + 0.0034947250969707966, + 0.004170078784227371, + 0.0025694500654935837, + 0.004531583748757839, + 0.003916084300726652, + 0.009769055992364883, + 0.0037858174182474613, + 0.004592167213559151, + 0.004550967365503311, + 0.005039875395596027, + 0.00430144090205431, + 0.005891749635338783, + 0.011385022662580013, + 0.03899846598505974, + 0.01101760845631361, + 0.003195337485522032, + 0.009603054262697697, + 0.011051100678741932, + 0.0032414011657238007, + 0.03498491272330284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0233369879424572 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006755134556442499 + ], + "opposite_logprob_tokens": [ + -5.006755352020264 + ], + "logprob_reward": -0.006755134556442499, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006755134556442499, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1573486179113388, + 0.0928557738661766, + 0.20036402344703674, + 0.2167147547006607, + 0.08933091908693314, + 0.049613628536462784, + 0.0401659794151783, + 0.050850801169872284, + 0.047865357249975204, + 0.054890114814043045 + ], + "credit_map": [ + 0.01978227123618126, + 0.025083743035793304, + 0.013074438087642193, + 0.015506863594055176, + 0.024143608286976814, + 0.020634301006793976, + 0.015875203534960747, + 0.02324819751083851, + 0.007738566026091576, + 0.009167708456516266, + 0.021748069673776627, + 0.006480247247964144, + 0.006775994319468737, + 0.017894454300403595, + 0.008412610739469528, + 0.014638126827776432, + 0.017378203570842743, + 0.015623625367879868, + 0.04598896577954292, + 0.02434188313782215, + 0.01301069837063551, + 0.020514778792858124, + 0.030825788155198097, + 0.03268007934093475, + 0.01465736236423254, + 0.01580594852566719, + 0.026870228350162506, + 0.01709824427962303, + 0.019957823678851128, + 0.05102955922484398, + 0.0333540253341198, + 0.0379415825009346, + 0.010365490801632404, + 0.011964035220444202, + 0.01201632246375084, + 0.008753456175327301, + 0.01093241199851036, + 0.013331038877367973, + 0.010101092047989368, + 0.011867069639265537, + 0.005057808477431536, + 0.0077026779763400555, + 0.005027958657592535, + 0.005058446899056435, + 0.0045691016130149364, + 0.007394915446639061, + 0.005186828318983316, + 0.00961589440703392, + 0.004014025907963514, + 0.00754691893234849, + 0.004167116712778807, + 0.005077846813946962, + 0.003919965587556362, + 0.005079914350062609, + 0.004458459559828043, + 0.005901732482016087, + 0.008318345062434673, + 0.007619188167154789, + 0.006213818211108446, + 0.0057866438291966915, + 0.00572837283834815, + 0.0063483864068984985, + 0.004378003999590874, + 0.006458046846091747, + 0.00438712676987052, + 0.008376635611057281, + 0.00446389289572835, + 0.005316684488207102, + 0.0047760927118361, + 0.006539378315210342, + 0.0055837370455265045, + 0.008421809412539005, + 0.005288347136229277, + 0.016682639718055725, + 0.005021660588681698, + 0.0022382766474038363, + 0.004302683286368847, + 0.004717240110039711, + 0.0023185531608760357, + 0.01432071439921856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.006755134556442499 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810521960258484 + ], + "opposite_logprob_tokens": [ + -3.2881052494049072 + ], + "logprob_reward": -0.03810521960258484, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810521960258484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12647032737731934, + 0.052946221083402634, + 0.1143687441945076, + 0.1649240404367447, + 0.07113620638847351, + 0.09669667482376099, + 0.06850755959749222, + 0.0748550146818161, + 0.12633252143859863, + 0.10376280546188354 + ], + "credit_map": [ + 0.024370798841118813, + 0.013286760076880455, + 0.013485060073435307, + 0.009335189126431942, + 0.014650636352598667, + 0.01858874037861824, + 0.009442523121833801, + 0.02331061288714409, + 0.005342450924217701, + 0.008121015504002571, + 0.008213446475565434, + 0.005659053567796946, + 0.006694742478430271, + 0.005837318021804094, + 0.00443728081882, + 0.008640911430120468, + 0.009422365576028824, + 0.007503173314034939, + 0.028397534042596817, + 0.013324991799890995, + 0.00663755601271987, + 0.014051098376512527, + 0.01585979014635086, + 0.01917222887277603, + 0.013305753469467163, + 0.015042779967188835, + 0.02031836472451687, + 0.012347566895186901, + 0.016171030700206757, + 0.033997200429439545, + 0.025665899738669395, + 0.0280754454433918, + 0.007112371269613504, + 0.008625248447060585, + 0.010319785214960575, + 0.008561614900827408, + 0.006210635881870985, + 0.011085284873843193, + 0.007775953039526939, + 0.011445313692092896, + 0.010499956086277962, + 0.014271536841988564, + 0.008577389642596245, + 0.010094735771417618, + 0.018033785745501518, + 0.014568758197128773, + 0.009446305222809315, + 0.011204201728105545, + 0.007492220960557461, + 0.012249913066625595, + 0.006269010715186596, + 0.007051857654005289, + 0.009737600572407246, + 0.009274707175791264, + 0.006893983576446772, + 0.009538265876471996, + 0.013244221918284893, + 0.011278459802269936, + 0.009938863106071949, + 0.007180131506174803, + 0.009081689640879631, + 0.010206887498497963, + 0.006278960965573788, + 0.007645799778401852, + 0.012931468896567822, + 0.025506850332021713, + 0.01059512235224247, + 0.01654563657939434, + 0.013789364136755466, + 0.014905166812241077, + 0.013735438697040081, + 0.018323473632335663, + 0.009790215641260147, + 0.03179614618420601, + 0.00843156035989523, + 0.004705082159489393, + 0.009065311402082443, + 0.008902476169168949, + 0.0033357280772179365, + 0.027736280113458633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.03810521960258484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an ancient copper lock cylinder suspended in a pure black void?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cylinder surface show age-worn patina and minute scratches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the cylinder disintegrate into floating gears, springs, pins, and levers that form a mechanical nebula?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the scattered components converge and reassemble into a complete lock cylinder along a precise trajectory?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow pan and track motion throughout the sequence?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707188725471497 + ], + "opposite_logprob_tokens": [ + -2.1270718574523926 + ], + "logprob_reward": -0.12707188725471497, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707188725471497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08188466727733612, + 0.06109455227851868, + 0.40793877840042114, + 0.1766376495361328, + 0.06813155859708786, + 0.03275282680988312, + 0.041907548904418945, + 0.03910312429070473, + 0.03411390259861946, + 0.05643533542752266 + ], + "credit_map": [ + 0.010833202861249447, + 0.011134618893265724, + 0.008135327138006687, + 0.008530833758413792, + 0.009505921974778175, + 0.012739652767777443, + 0.008836355060338974, + 0.012168755754828453, + 0.004358681384474039, + 0.007770051248371601, + 0.012353005819022655, + 0.004011548589915037, + 0.007133111823350191, + 0.011509588919579983, + 0.006495945155620575, + 0.007462621666491032, + 0.03362133726477623, + 0.03217808157205582, + 0.09155785292387009, + 0.048300620168447495, + 0.02723371423780918, + 0.04144250229001045, + 0.07234041392803192, + 0.0612642839550972, + 0.007335666101425886, + 0.00909334048628807, + 0.020969271659851074, + 0.01596221700310707, + 0.018122559413313866, + 0.0406123623251915, + 0.03457804396748543, + 0.029964182525873184, + 0.005756340455263853, + 0.012216378003358841, + 0.00797277968376875, + 0.006945971865206957, + 0.006812567822635174, + 0.010230776853859425, + 0.008830178529024124, + 0.009366568177938461, + 0.0027088504284620285, + 0.0053939297795295715, + 0.004423712380230427, + 0.003694569692015648, + 0.004419846460223198, + 0.004976807162165642, + 0.003427077317610383, + 0.0037080354522913694, + 0.004558708984404802, + 0.005070759914815426, + 0.004804816097021103, + 0.00541619723662734, + 0.0036172298714518547, + 0.006233525462448597, + 0.004440565127879381, + 0.007765745744109154, + 0.006174049340188503, + 0.0044476049952209, + 0.005566059611737728, + 0.003914556931704283, + 0.00375903258100152, + 0.00550714461132884, + 0.004207507707178593, + 0.005527164321392775, + 0.003776089521124959, + 0.0060602943412959576, + 0.0034357619006186724, + 0.003536272095516324, + 0.004610447678714991, + 0.0042225974611938, + 0.003088932717218995, + 0.005383510608226061, + 0.005294483620673418, + 0.017157400026917458, + 0.004898225888609886, + 0.0023372319992631674, + 0.004716366529464722, + 0.005092259030789137, + 0.0021326416172087193, + 0.014806726016104221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.12707188725471497 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_1/index_tensor([2361, 2361])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_1/index_tensor([2361, 2361])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c340f37df6dc41d38768e2742b103df06bb2de8b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_1/index_tensor([2361, 2361])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tracking shot, shot from a slightly upward angle, follows a motorcycle in profile, the scene imbued with a rich, vivid painterly style, its brushstrokes clearly visible. The vintage cruiser, painted a deep crimson, gleams in the golden evening sun, its chrome engine and exhaust reflecting impasto-like highlights. The rider, clad in a wrinkled black leather jacket and a matte black open-face helmet, leans slightly forward, his hands steady on the handlebars, his posture focused and resolute. The motorcycle travels at a steady, powerful pace along a winding country road, its surface composed of rough brushstrokes of gray and brown. On one side, golden wheat fields sway in the wind, on the other, rolling green hills in the distance, while the sky is a magnificent canvas of bold, swirling oranges, purples, and deep blues. The low-hanging sun casts long, twisting shadows across the ground, and the wheels' rotation seems to stir the paint on the canvas. The entire picture is full of strong textures and saturated colors, creating a dynamic atmosphere that combines freedom, strength and rich artistic flavor.", + "seed": 15753, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_1/index_2361_seed_15753_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02468289928510785, + "vlm_logprob": -0.02468289928510785, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.02468289928510785, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06421308219432831 + ], + "opposite_logprob_tokens": [ + -2.8142130374908447 + ], + "logprob_reward": -0.06421308219432831, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06421308219432831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09809326380491257, + 0.06910935044288635, + 0.06573580205440521, + 0.4464363753795624, + 0.05892867222428322, + 0.03029576689004898, + 0.026742098852992058, + 0.025840777903795242, + 0.052951402962207794, + 0.12586653232574463 + ], + "credit_map": [ + 0.011700453236699104, + 0.008759327232837677, + 0.023775074630975723, + 0.012611374258995056, + 0.007276465184986591, + 0.009670997969806194, + 0.014291804283857346, + 0.010007753036916256, + 0.01362286414951086, + 0.00645856000483036, + 0.018786566331982613, + 0.007004480343312025, + 0.0046220202930271626, + 0.005965798161923885, + 0.00682161096483469, + 0.005827444605529308, + 0.008442997932434082, + 0.007722599897533655, + 0.008281631395220757, + 0.009862150065600872, + 0.006968038622289896, + 0.008736152201890945, + 0.008305422961711884, + 0.007416806649416685, + 0.12390977144241333, + 0.0683583989739418, + 0.11758840084075928, + 0.03133009001612663, + 0.020245935767889023, + 0.027020759880542755, + 0.032688457518815994, + 0.025294505059719086, + 0.008637812919914722, + 0.00952434167265892, + 0.008840929716825485, + 0.007935240864753723, + 0.00530257448554039, + 0.004841186571866274, + 0.008145654574036598, + 0.005700923502445221, + 0.006371998693794012, + 0.003056929213926196, + 0.0065978108905255795, + 0.0036273933947086334, + 0.0020520752295851707, + 0.0027891939971596003, + 0.003185476642102003, + 0.002614886499941349, + 0.004531831480562687, + 0.0024909668136388063, + 0.005669451784342527, + 0.003105844371020794, + 0.002097097458317876, + 0.00286151934415102, + 0.00297563336789608, + 0.0030097512062639, + 0.003880947595462203, + 0.0021942444145679474, + 0.004074338357895613, + 0.0031917025335133076, + 0.00461449334397912, + 0.00196836213581264, + 0.002939981408417225, + 0.0029767060186713934, + 0.005929258186370134, + 0.00445806048810482, + 0.006868143565952778, + 0.011686268262565136, + 0.0051560187712311745, + 0.007286603096872568, + 0.00644311960786581, + 0.0051239230670034885, + 0.024050407111644745, + 0.022744664922356606, + 0.009821794927120209, + 0.01646764762699604, + 0.008894214406609535, + 0.029692061245441437, + 0.006268093828111887, + 0.007927634753286839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.06421308219432831 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02059515379369259 + ], + "opposite_logprob_tokens": [ + -3.895595073699951 + ], + "logprob_reward": -0.02059515379369259, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02059515379369259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0418725349009037, + 0.05188317224383354, + 0.032792236655950546, + 0.26058149337768555, + 0.04910357668995857, + 0.057624734938144684, + 0.05100717395544052, + 0.050325457006692886, + 0.06295552849769592, + 0.34185412526130676 + ], + "credit_map": [ + 0.005853455513715744, + 0.005561820697039366, + 0.008325929753482342, + 0.004568081349134445, + 0.003030864056199789, + 0.00422382028773427, + 0.005640449933707714, + 0.004668107721954584, + 0.0066130440682172775, + 0.003743406618013978, + 0.014551208354532719, + 0.0058184098452329636, + 0.004360548686236143, + 0.004449589177966118, + 0.00801159255206585, + 0.004335371777415276, + 0.003684091614559293, + 0.004422626458108425, + 0.007318485528230667, + 0.0036513316445052624, + 0.0029948577284812927, + 0.00336982449516654, + 0.005221545230597258, + 0.0021294739563018084, + 0.07297024130821228, + 0.0330541618168354, + 0.08433188498020172, + 0.01485497411340475, + 0.008904758840799332, + 0.01167629286646843, + 0.02411198988556862, + 0.010677160695195198, + 0.006979513913393021, + 0.0038745540659874678, + 0.010646515525877476, + 0.005931416992098093, + 0.005674764979630709, + 0.005370860919356346, + 0.0059527005068957806, + 0.0046732439659535885, + 0.012587262317538261, + 0.007024597376585007, + 0.01481917779892683, + 0.005229453090578318, + 0.003308937419205904, + 0.004841443616896868, + 0.005442599765956402, + 0.004371259827166796, + 0.008472179993987083, + 0.00599666079506278, + 0.009939705021679401, + 0.005552695132791996, + 0.004661528393626213, + 0.00536236260086298, + 0.00610312120988965, + 0.004918916150927544, + 0.011216111481189728, + 0.003961688838899136, + 0.014517892152071, + 0.00395392207428813, + 0.004162263125181198, + 0.0037979702465236187, + 0.005380784627050161, + 0.00333481770940125, + 0.015072229318320751, + 0.005887382198125124, + 0.021070320159196854, + 0.005443261936306953, + 0.002593077253550291, + 0.004493081010878086, + 0.005549486260861158, + 0.0028466787189245224, + 0.028486832976341248, + 0.04311559349298477, + 0.0878666415810585, + 0.03241538256406784, + 0.017926497384905815, + 0.09212472289800644, + 0.02423306182026863, + 0.015685392543673515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.02059515379369259 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002488494385033846 + ], + "opposite_logprob_tokens": [ + -6.002488613128662 + ], + "logprob_reward": -0.002488494385033846, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002488494385033846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1210215613245964, + 0.08721023052930832, + 0.05421710014343262, + 0.0519634410738945, + 0.031774502247571945, + 0.12978677451610565, + 0.07021266967058182, + 0.06159006804227829, + 0.0862196683883667, + 0.30600404739379883 + ], + "credit_map": [ + 0.01805940456688404, + 0.007580310106277466, + 0.03476528823375702, + 0.012322334572672844, + 0.007250483147799969, + 0.017191730439662933, + 0.012100061401724815, + 0.011751944199204445, + 0.011383538134396076, + 0.010919550433754921, + 0.016800105571746826, + 0.010936856269836426, + 0.008200311101973057, + 0.01071506179869175, + 0.012541675940155983, + 0.005713123362511396, + 0.004933295771479607, + 0.004265754483640194, + 0.019045425578951836, + 0.005072848405689001, + 0.003766341134905815, + 0.00382418860681355, + 0.00845644623041153, + 0.004852792713791132, + 0.009928222745656967, + 0.010405154898762703, + 0.004300407599657774, + 0.006447916850447655, + 0.004270885605365038, + 0.00620055990293622, + 0.006041607819497585, + 0.004368680063635111, + 0.0063540143892169, + 0.004305047914385796, + 0.00644797133281827, + 0.003928444813936949, + 0.002682383870705962, + 0.0022915026638656855, + 0.0030163719784468412, + 0.002748762723058462, + 0.033260978758335114, + 0.013594645075500011, + 0.04023423790931702, + 0.0078017269261181355, + 0.005008170381188393, + 0.010959404520690441, + 0.011352384462952614, + 0.007575216703116894, + 0.013412277214229107, + 0.009424841031432152, + 0.014124569483101368, + 0.007484184578061104, + 0.0053472924046218395, + 0.007378413341939449, + 0.00643286295235157, + 0.006608220748603344, + 0.013369616121053696, + 0.0047595989890396595, + 0.01841861940920353, + 0.004915265832096338, + 0.005706261843442917, + 0.005553280934691429, + 0.005703428294509649, + 0.0031639914959669113, + 0.012956872582435608, + 0.01383716519922018, + 0.020634932443499565, + 0.010432939976453781, + 0.0067682648077607155, + 0.011609218083322048, + 0.005943978205323219, + 0.004036290105432272, + 0.051761988550424576, + 0.061917033046483994, + 0.0373939648270607, + 0.04999037832021713, + 0.02466687373816967, + 0.03737549111247063, + 0.023445885628461838, + 0.019452394917607307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.002488494385033846 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026341434568166733 + ], + "opposite_logprob_tokens": [ + -3.651341438293457 + ], + "logprob_reward": -0.026341434568166733, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026341434568166733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10192003101110458, + 0.066643625497818, + 0.06207073852419853, + 0.04043368622660637, + 0.03573368117213249, + 0.4284937381744385, + 0.09794636815786362, + 0.046763572841882706, + 0.053307704627513885, + 0.06668683886528015 + ], + "credit_map": [ + 0.012411019764840603, + 0.01261717639863491, + 0.01921728253364563, + 0.011576259508728981, + 0.00828549824655056, + 0.020272113382816315, + 0.009439817629754543, + 0.008100862614810467, + 0.009336973540484905, + 0.008537200279533863, + 0.011018646880984306, + 0.00940481573343277, + 0.007287314627319574, + 0.007798708509653807, + 0.008498679846525192, + 0.004761287942528725, + 0.006211998872458935, + 0.006183641497045755, + 0.016700534150004387, + 0.007477348204702139, + 0.0058415899984538555, + 0.004470497369766235, + 0.010711578652262688, + 0.004473549779504538, + 0.011714333668351173, + 0.007339888717979193, + 0.007682092487812042, + 0.0034017893485724926, + 0.002380161313340068, + 0.002460879273712635, + 0.002998573239892721, + 0.002455970970913768, + 0.004572897218167782, + 0.004516254644840956, + 0.004886905197054148, + 0.0038970988243818283, + 0.004216792527586222, + 0.005922501441091299, + 0.003882790682837367, + 0.003838441800326109, + 0.11252735555171967, + 0.04665251076221466, + 0.11998746544122696, + 0.03351622447371483, + 0.02163419872522354, + 0.028451913967728615, + 0.036445051431655884, + 0.029279008507728577, + 0.021103687584400177, + 0.010172855108976364, + 0.0248569305986166, + 0.008147348649799824, + 0.006413449067622423, + 0.009183176793158054, + 0.009734910912811756, + 0.008334006182849407, + 0.009998759254813194, + 0.0036262855865061283, + 0.014784838072955608, + 0.0038809669204056263, + 0.0037790241185575724, + 0.003069300437346101, + 0.0047795334830880165, + 0.0028448649682104588, + 0.01333074551075697, + 0.004260459449142218, + 0.017887528985738754, + 0.004100635182112455, + 0.002244043629616499, + 0.004411717411130667, + 0.004777323454618454, + 0.002295244950801134, + 0.014613076113164425, + 0.006252493243664503, + 0.0067898789420723915, + 0.008326319977641106, + 0.004413717892020941, + 0.016517072916030884, + 0.004639096558094025, + 0.005135184619575739 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.026341434568166733 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00977633148431778 + ], + "opposite_logprob_tokens": [ + -4.6347761154174805 + ], + "logprob_reward": -0.00977633148431778, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00977633148431778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1185779869556427, + 0.13201195001602173, + 0.057576123625040054, + 0.12959466874599457, + 0.09061738103628159, + 0.06404591351747513, + 0.02783529832959175, + 0.17037253081798553, + 0.08447963744401932, + 0.12488853931427002 + ], + "credit_map": [ + 0.018340760841965675, + 0.01748090423643589, + 0.023640749976038933, + 0.015453541651368141, + 0.008844063617289066, + 0.013684382662177086, + 0.010928381234407425, + 0.01020519994199276, + 0.012938891537487507, + 0.013350604102015495, + 0.02754996530711651, + 0.01498626172542572, + 0.011676827445626259, + 0.013584116473793983, + 0.023294439539313316, + 0.014630839228630066, + 0.004801398143172264, + 0.005995510146021843, + 0.015917254611849785, + 0.0060028210282325745, + 0.005517386831343174, + 0.005014027934521437, + 0.008833074010908604, + 0.005494654178619385, + 0.027328375726938248, + 0.0136701799929142, + 0.037073273211717606, + 0.01207747682929039, + 0.007546959910541773, + 0.00596280675381422, + 0.016981136053800583, + 0.008954452350735664, + 0.012391998432576656, + 0.00951897632330656, + 0.018589308485388756, + 0.011990121565759182, + 0.006496912334114313, + 0.006970385555177927, + 0.014948221854865551, + 0.009711455553770065, + 0.015765942633152008, + 0.005876308307051659, + 0.017320869490504265, + 0.005523955449461937, + 0.0032738952431827784, + 0.00480162026360631, + 0.006928000599145889, + 0.0045553212985396385, + 0.003106969641521573, + 0.0026058864314109087, + 0.003958138171583414, + 0.00285474699921906, + 0.003319211071357131, + 0.0025962567888200283, + 0.004317047074437141, + 0.005077043082565069, + 0.03626534342765808, + 0.01536674052476883, + 0.0499720461666584, + 0.014564894139766693, + 0.013508416712284088, + 0.010774356313049793, + 0.01677636429667473, + 0.013144368305802345, + 0.01612473838031292, + 0.004948596004396677, + 0.029100093990564346, + 0.009527933783829212, + 0.005381044466048479, + 0.005110468715429306, + 0.008942028507590294, + 0.00534473080188036, + 0.0212123803794384, + 0.02067086473107338, + 0.03079364262521267, + 0.016164932399988174, + 0.0071749561466276646, + 0.011760036461055279, + 0.011253729462623596, + 0.0058579896576702595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.00977633148431778 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking shot, shot from a slightly upward angle, follows a motorcycle in profile, the scene imbued with a rich, vivid painterly style, its brushstrokes clearly visible. The vintage cruiser, painted a deep crimson, gleams in the golden evening sun, its chrome engine and exhaust reflecting impasto-like highlights. The rider, clad in a wrinkled black leather jacket and a matte black open-face helmet, leans slightly forward, his hands steady on the handlebars, his posture focused and resolute. The motorcycle travels at a steady, powerful pace along a winding country road, its surface composed of rough brushstrokes of gray and brown. On one side, golden wheat fields sway in the wind, on the other, rolling green hills in the distance, while the sky is a magnificent canvas of bold, swirling oranges, purples, and deep blues. The low-hanging sun casts long, twisting shadows across the ground, and the wheels' rotation seems to stir the paint on the canvas. The entire picture is full of strong textures and saturated colors, creating a dynamic atmosphere that combines freedom, strength and rich artistic flavor.", + "seed": 15754, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_1/index_2361_seed_15754_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.012252043746411801, + "vlm_logprob": -0.012252043746411801, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.012252043746411801, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028822878375649452 + ], + "opposite_logprob_tokens": [ + -3.653822898864746 + ], + "logprob_reward": -0.028822878375649452, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.028822878375649452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06804478168487549, + 0.02510528638958931, + 0.006840401794761419, + 0.020900573581457138, + 0.7626111507415771, + 0.01582960970699787, + 0.006046555005013943, + 0.05037324130535126, + 0.007480923552066088, + 0.03676743805408478 + ], + "credit_map": [ + 0.015789970755577087, + 0.00850715022534132, + 0.02185603603720665, + 0.004592208657413721, + 0.003503648564219475, + 0.0032144999131560326, + 0.006255462300032377, + 0.004325807094573975, + 0.0034914903808385134, + 0.0015067282365635037, + 0.003375635016709566, + 0.002325729699805379, + 0.002582474844530225, + 0.0029942479450255632, + 0.006577253341674805, + 0.002251726109534502, + 0.0008037794614210725, + 0.0008086126181297004, + 0.0010646764421835542, + 0.000849968462716788, + 0.0007624011486768723, + 0.0006271369638852775, + 0.0009003946324810386, + 0.0010234316578134894, + 0.0037361241411417723, + 0.0032891829032450914, + 0.002071191556751728, + 0.002130554523319006, + 0.0021164529025554657, + 0.002290009520947933, + 0.0031751564238220453, + 0.002091901609674096, + 0.20696064829826355, + 0.060053590685129166, + 0.23683412373065948, + 0.05099651962518692, + 0.034200094640254974, + 0.053532861173152924, + 0.06240218132734299, + 0.057631149888038635, + 0.002237862441688776, + 0.001482380903325975, + 0.0032751932740211487, + 0.0013924463419243693, + 0.0012198081240057945, + 0.002041012281551957, + 0.002571342047303915, + 0.0016095652244985104, + 0.0007406299118883908, + 0.0007720174617134035, + 0.0006717870710417628, + 0.0006247757701203227, + 0.0006713505717925727, + 0.0006242586532607675, + 0.0012711264425888658, + 0.0006706091808155179, + 0.004331484436988831, + 0.004486633464694023, + 0.007440993096679449, + 0.007059626281261444, + 0.004709663335233927, + 0.005851760506629944, + 0.01013969536870718, + 0.006353386212140322, + 0.0015564532950520515, + 0.0015696394257247448, + 0.0008275965810753405, + 0.0006507677608169615, + 0.00056967738782987, + 0.0006673509487882257, + 0.0009073325782082975, + 0.0007321052253246307, + 0.009585771709680557, + 0.003221483202651143, + 0.012044300325214863, + 0.0023595893289893866, + 0.0014763770159333944, + 0.001648484030738473, + 0.003176777856424451, + 0.0032546543516218662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.028822878375649452 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550657615065575 + ], + "opposite_logprob_tokens": [ + -4.3875508308410645 + ], + "logprob_reward": -0.012550657615065575, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550657615065575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42449429631233215, + 0.05939492583274841, + 0.017321618273854256, + 0.053759459406137466, + 0.27404963970184326, + 0.013803272508084774, + 0.039567962288856506, + 0.04137139022350311, + 0.018444249406456947, + 0.05779312551021576 + ], + "credit_map": [ + 0.11697141826152802, + 0.03401271253824234, + 0.1616152971982956, + 0.025219250470399857, + 0.013668148778378963, + 0.013606183230876923, + 0.03654636815190315, + 0.022854942828416824, + 0.01405742485076189, + 0.007963503710925579, + 0.014502284117043018, + 0.00436404999345541, + 0.004778440110385418, + 0.004065813962370157, + 0.00520765595138073, + 0.004455755464732647, + 0.0028842405881732702, + 0.0030312545131891966, + 0.0020775343291461468, + 0.0016433929558843374, + 0.0023620338179171085, + 0.0017440993105992675, + 0.0019143680110573769, + 0.001664695213548839, + 0.007381115574389696, + 0.007380838040262461, + 0.004688077140599489, + 0.006683196872472763, + 0.004883522633463144, + 0.005363815929740667, + 0.008332375437021255, + 0.009046517312526703, + 0.0750746876001358, + 0.022165026515722275, + 0.08398620039224625, + 0.017009733244776726, + 0.014832021668553352, + 0.01775074191391468, + 0.02163601852953434, + 0.021595194935798645, + 0.0022844444029033184, + 0.0016126391710713506, + 0.002181251998990774, + 0.0011795442551374435, + 0.0018407809548079967, + 0.0014346896205097437, + 0.001498863217420876, + 0.0017710586544126272, + 0.0046730050817132, + 0.0046666949056088924, + 0.008798531256616116, + 0.003684125142171979, + 0.003916739486157894, + 0.005975706502795219, + 0.0031434502452611923, + 0.004709708504378796, + 0.004116034600883722, + 0.004526584409177303, + 0.006649333517998457, + 0.005373711232095957, + 0.0034031509421765804, + 0.006087578367441893, + 0.006941798608750105, + 0.004273201338946819, + 0.0026253266260027885, + 0.003794283838942647, + 0.002741819014772773, + 0.0016653676284477115, + 0.001461915671825409, + 0.00267915241420269, + 0.0015539119485765696, + 0.0019224720308557153, + 0.019472306594252586, + 0.019968412816524506, + 0.003758709877729416, + 0.0024979892186820507, + 0.0029987699817866087, + 0.0020666501950472593, + 0.0034605085384100676, + 0.003569779684767127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.012550657615065575 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00248873233795166 + ], + "opposite_logprob_tokens": [ + -6.002488613128662 + ], + "logprob_reward": -0.00248873233795166, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00248873233795166, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44624483585357666, + 0.06082189455628395, + 0.02936863899230957, + 0.04349532723426819, + 0.07509805262088776, + 0.02839203178882599, + 0.07460511475801468, + 0.05789615586400032, + 0.01589129865169525, + 0.16818663477897644 + ], + "credit_map": [ + 0.11835551261901855, + 0.027316613122820854, + 0.16984088718891144, + 0.025053882971405983, + 0.012861860916018486, + 0.022807303816080093, + 0.040482297539711, + 0.0295264832675457, + 0.015600736252963543, + 0.005565752740949392, + 0.014252816326916218, + 0.004318373277783394, + 0.0029572993516921997, + 0.003334087086841464, + 0.010911731980741024, + 0.003881096839904785, + 0.004654312040656805, + 0.004562786314636469, + 0.005393170285969973, + 0.002013582969084382, + 0.002755134366452694, + 0.004427558742463589, + 0.002503523137420416, + 0.0030585702043026686, + 0.009448669850826263, + 0.004154777154326439, + 0.007906030863523483, + 0.002681349404156208, + 0.0035223790910094976, + 0.007113187108188868, + 0.004019066691398621, + 0.004649864975363016, + 0.01780204102396965, + 0.006343332584947348, + 0.02191738598048687, + 0.0058081308379769325, + 0.0045656561851501465, + 0.005154786631464958, + 0.007690440863370895, + 0.005816278047859669, + 0.0042193117551505566, + 0.004273670259863138, + 0.005629261489957571, + 0.0025939173065125942, + 0.0021236035972833633, + 0.0030117498245090246, + 0.003282456425949931, + 0.0032580618280917406, + 0.009598223492503166, + 0.006658064667135477, + 0.015799066051840782, + 0.006256883032619953, + 0.007062298245728016, + 0.014674915000796318, + 0.00577267212793231, + 0.008782994002103806, + 0.008750059641897678, + 0.010443701408803463, + 0.007869433611631393, + 0.005913556553423405, + 0.005538888741284609, + 0.00854691956192255, + 0.004242859780788422, + 0.006590735167264938, + 0.002283520996570587, + 0.0030892156064510345, + 0.002328429603949189, + 0.0015178621979430318, + 0.00143003452103585, + 0.0018402052810415626, + 0.0017805610550567508, + 0.001621469622477889, + 0.06460151821374893, + 0.06451071053743362, + 0.008791652508080006, + 0.005603334400802851, + 0.0075498600490391254, + 0.003872433677315712, + 0.006307503208518028, + 0.00694961566478014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.00248873233795166 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623625919222832 + ], + "opposite_logprob_tokens": [ + -4.882623672485352 + ], + "logprob_reward": -0.007623625919222832, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623625919222832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7196389436721802, + 0.07947918772697449, + 0.010977866128087044, + 0.009343549609184265, + 0.05351105332374573, + 0.006730339024215937, + 0.021487141028046608, + 0.01808888278901577, + 0.0498468279838562, + 0.03089616447687149 + ], + "credit_map": [ + 0.2039041519165039, + 0.031167898327112198, + 0.29235154390335083, + 0.0387590229511261, + 0.016778381541371346, + 0.028324035927653313, + 0.0677984431385994, + 0.04055549576878548, + 0.021379759535193443, + 0.005312672350555658, + 0.019050493836402893, + 0.003981112036854029, + 0.005296795628964901, + 0.010969157330691814, + 0.006191066466271877, + 0.007298125419765711, + 0.0019510929705575109, + 0.0012773140333592892, + 0.001986747607588768, + 0.0006370464689098299, + 0.0009495739941485226, + 0.002133444417268038, + 0.000850304146297276, + 0.001192342839203775, + 0.00178559182677418, + 0.0014433993492275476, + 0.0018143105553463101, + 0.0007675251690670848, + 0.0007497440674342215, + 0.001123496680520475, + 0.0008679156308062375, + 0.0007915662135928869, + 0.013143431395292282, + 0.004431705456227064, + 0.015486610122025013, + 0.004160785116255283, + 0.0027122346218675375, + 0.0043522692285478115, + 0.005072496831417084, + 0.004151520784944296, + 0.0008820551447570324, + 0.0009619005722925067, + 0.0015245743561536074, + 0.0006103765335865319, + 0.0005150553770363331, + 0.0008601677254773676, + 0.0005591865046881139, + 0.0008170224027708173, + 0.0031695240177214146, + 0.0035797017626464367, + 0.004605210851877928, + 0.0020124707370996475, + 0.0016577334608882666, + 0.0024515134282410145, + 0.0018040427239611745, + 0.0022069436963647604, + 0.00280980602838099, + 0.0026841929648071527, + 0.0027638250030577183, + 0.0017425530822947621, + 0.001971424790099263, + 0.0026692324317991734, + 0.0015991916880011559, + 0.001848657033406198, + 0.00981393363326788, + 0.010392005555331707, + 0.006047867704182863, + 0.003511907299980521, + 0.00319495121948421, + 0.007527264300733805, + 0.004043682478368282, + 0.00531521812081337, + 0.010127240791916847, + 0.011827380396425724, + 0.003052511252462864, + 0.0010067956754937768, + 0.0015712676104158163, + 0.0007209919858723879, + 0.0013292118674144149, + 0.0012607643147930503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.007623625919222832 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774324484169483 + ], + "opposite_logprob_tokens": [ + -4.634774208068848 + ], + "logprob_reward": -0.009774324484169483, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774324484169483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36180317401885986, + 0.1173734962940216, + 0.05934615805745125, + 0.04263952746987343, + 0.05933481082320213, + 0.04438391700387001, + 0.043876808136701584, + 0.07262971252202988, + 0.02798168733716011, + 0.17063072323799133 + ], + "credit_map": [ + 0.06758669018745422, + 0.08640391379594803, + 0.06396320462226868, + 0.026675576344132423, + 0.025964729487895966, + 0.029812633991241455, + 0.03044731169939041, + 0.030949074774980545, + 0.02189251221716404, + 0.024747855961322784, + 0.020350381731987, + 0.008702763356268406, + 0.00975609291344881, + 0.011057119816541672, + 0.011233293451368809, + 0.009633475914597511, + 0.008742470294237137, + 0.006010417826473713, + 0.008270041085779667, + 0.00550035061314702, + 0.007135780993849039, + 0.010018418543040752, + 0.006324803456664085, + 0.007343873381614685, + 0.007259702775627375, + 0.00270038447342813, + 0.004922025371342897, + 0.004764885641634464, + 0.004178308416157961, + 0.0056363963522017, + 0.0059463391080498695, + 0.007231482770293951, + 0.008676315657794476, + 0.006241483148187399, + 0.01039392314851284, + 0.00568728893995285, + 0.007040857337415218, + 0.005699790082871914, + 0.00968010164797306, + 0.005915046203881502, + 0.0061515746638178825, + 0.008176968432962894, + 0.005706429947167635, + 0.004806543234735727, + 0.006101442966610193, + 0.003617730690166354, + 0.00531949894502759, + 0.004503723233938217, + 0.00503036193549633, + 0.004807207267731428, + 0.004640028811991215, + 0.005659796763211489, + 0.006962598767131567, + 0.003408011281862855, + 0.007963836193084717, + 0.005404964555054903, + 0.00961246620863676, + 0.010256597772240639, + 0.009630070999264717, + 0.00836616475135088, + 0.006067080423235893, + 0.010234009474515915, + 0.010717073455452919, + 0.007746248506009579, + 0.005921682808548212, + 0.003995849285274744, + 0.004208598751574755, + 0.0031924075447022915, + 0.0025139653589576483, + 0.0017495297361165285, + 0.0035170840565115213, + 0.00288256979547441, + 0.05749186500906944, + 0.039512913674116135, + 0.03460659459233284, + 0.008614769205451012, + 0.005183978006243706, + 0.006183268036693335, + 0.010336083360016346, + 0.008701237849891186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.009774324484169483 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking shot, shot from a slightly upward angle, follows a motorcycle in profile, the scene imbued with a rich, vivid painterly style, its brushstrokes clearly visible. The vintage cruiser, painted a deep crimson, gleams in the golden evening sun, its chrome engine and exhaust reflecting impasto-like highlights. The rider, clad in a wrinkled black leather jacket and a matte black open-face helmet, leans slightly forward, his hands steady on the handlebars, his posture focused and resolute. The motorcycle travels at a steady, powerful pace along a winding country road, its surface composed of rough brushstrokes of gray and brown. On one side, golden wheat fields sway in the wind, on the other, rolling green hills in the distance, while the sky is a magnificent canvas of bold, swirling oranges, purples, and deep blues. The low-hanging sun casts long, twisting shadows across the ground, and the wheels' rotation seems to stir the paint on the canvas. The entire picture is full of strong textures and saturated colors, creating a dynamic atmosphere that combines freedom, strength and rich artistic flavor.", + "seed": 15755, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_1/index_2361_seed_15755_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03690402451902628, + "vlm_logprob": -0.03690402451902628, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.03690402451902628, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11565513908863068 + ], + "opposite_logprob_tokens": [ + -2.2406551837921143 + ], + "logprob_reward": -0.11565513908863068, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11565513908863068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04987773671746254, + 0.016770359128713608, + 0.012498165480792522, + 0.054629500955343246, + 0.5077628493309021, + 0.029856964945793152, + 0.04165727645158768, + 0.20223447680473328, + 0.046898581087589264, + 0.037814121693372726 + ], + "credit_map": [ + 0.008209671825170517, + 0.005071509629487991, + 0.005492149852216244, + 0.006871891673654318, + 0.007808378431946039, + 0.004550945945084095, + 0.0066599370911717415, + 0.0052132499404251575, + 0.003197260433807969, + 0.0032643936574459076, + 0.0016198899829760194, + 0.00200615543872118, + 0.0020472847390919924, + 0.001504470594227314, + 0.0016248669708147645, + 0.0015060356818139553, + 0.0016877128509804606, + 0.001664794865064323, + 0.0010922442888841033, + 0.002388432389125228, + 0.0018075043335556984, + 0.0008290651021525264, + 0.0018517646240070462, + 0.0011766478419303894, + 0.013544026762247086, + 0.006028446834534407, + 0.006330640055239201, + 0.006176781840622425, + 0.004900828003883362, + 0.004853455349802971, + 0.008426155894994736, + 0.004369169007986784, + 0.10782013088464737, + 0.08356087654829025, + 0.05033862218260765, + 0.04685554280877113, + 0.06978791952133179, + 0.04809972271323204, + 0.06745654344558716, + 0.0338435098528862, + 0.003531030844897032, + 0.00307018612511456, + 0.004652907606214285, + 0.006285081151872873, + 0.003355354070663452, + 0.0024230785202234983, + 0.004021818283945322, + 0.002517511835321784, + 0.008504735305905342, + 0.0043552350252866745, + 0.0048765335232019424, + 0.006092004012316465, + 0.004059444181621075, + 0.0037164983805269003, + 0.0054506040178239346, + 0.004602219443768263, + 0.0342634916305542, + 0.031027019023895264, + 0.02381577156484127, + 0.01820261776447296, + 0.027189645916223526, + 0.027054455131292343, + 0.01963437907397747, + 0.021047085523605347, + 0.003486176487058401, + 0.004015413112938404, + 0.00302779208868742, + 0.004077476914972067, + 0.0036924441810697317, + 0.0037533200811594725, + 0.01728261075913906, + 0.0075633469969034195, + 0.0053263199515640736, + 0.008459676057100296, + 0.0028511600103229284, + 0.0035439664497971535, + 0.00409518601372838, + 0.0044914246536791325, + 0.005090167745947838, + 0.003956221509724855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 3 + ], + "selected_reward": -0.11565513908863068 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023305077105760574 + ], + "opposite_logprob_tokens": [ + -3.7733051776885986 + ], + "logprob_reward": -0.023305077105760574, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023305077105760574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09539423137903214, + 0.08536418527364731, + 0.06985239684581757, + 0.11393505334854126, + 0.21610844135284424, + 0.09928284585475922, + 0.048986323177814484, + 0.08536992222070694, + 0.10074103623628616, + 0.08496558666229248 + ], + "credit_map": [ + 0.012651722878217697, + 0.007899785414338112, + 0.010109852068126202, + 0.009538652375340462, + 0.021143432706594467, + 0.007763538509607315, + 0.01773880235850811, + 0.008548440411686897, + 0.011384140700101852, + 0.008694373071193695, + 0.008151661604642868, + 0.01148736197501421, + 0.018635204061865807, + 0.007298472803086042, + 0.01031761895865202, + 0.009395348839461803, + 0.006480504758656025, + 0.005778375547379255, + 0.0047808862291276455, + 0.007335606962442398, + 0.022358573973178864, + 0.005010541062802076, + 0.010987130925059319, + 0.007120779249817133, + 0.01404762826859951, + 0.009193191304802895, + 0.009697478264570236, + 0.014229122549295425, + 0.028956525027751923, + 0.007466079667210579, + 0.01768375374376774, + 0.012661279179155827, + 0.052449580281972885, + 0.03808600828051567, + 0.02478814497590065, + 0.014856455847620964, + 0.02010171115398407, + 0.023612042888998985, + 0.025491587817668915, + 0.016722915694117546, + 0.025076400488615036, + 0.009666495025157928, + 0.009417655877768993, + 0.007787259295582771, + 0.014373444020748138, + 0.008170517161488533, + 0.015408732928335667, + 0.009382338263094425, + 0.009430757723748684, + 0.004891711752861738, + 0.004566967487335205, + 0.0043803309090435505, + 0.006816119886934757, + 0.006002919282764196, + 0.008164891973137856, + 0.004732620902359486, + 0.012347747571766376, + 0.008411462418735027, + 0.0064020827412605286, + 0.007325861603021622, + 0.015185711905360222, + 0.010859078727662563, + 0.014266664162278175, + 0.010571308434009552, + 0.00914008915424347, + 0.011696932837367058, + 0.008996357209980488, + 0.00781301036477089, + 0.010583781637251377, + 0.013038893230259418, + 0.027113717049360275, + 0.01235826313495636, + 0.010040812194347382, + 0.011344325728714466, + 0.006919474806636572, + 0.005672297906130552, + 0.006118985824286938, + 0.006519561167806387, + 0.018239974975585938, + 0.020110147073864937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.023305077105760574 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010419422760605812 + ], + "opposite_logprob_tokens": [ + -6.876041889190674 + ], + "logprob_reward": -0.0010419422760605812, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010419422760605812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12714742124080658, + 0.03725742921233177, + 0.04912744462490082, + 0.13289856910705566, + 0.3467717170715332, + 0.04803943634033203, + 0.08382011204957962, + 0.03949970379471779, + 0.05121283605694771, + 0.08422534912824631 + ], + "credit_map": [ + 0.02035006508231163, + 0.013211837038397789, + 0.016403423622250557, + 0.019581014290452003, + 0.016942720860242844, + 0.01614728756248951, + 0.012402044609189034, + 0.012109021656215191, + 0.005583667661994696, + 0.0031568342819809914, + 0.00464550731703639, + 0.005670040380209684, + 0.0061767431907355785, + 0.003562561934813857, + 0.004105127416551113, + 0.0043569463305175304, + 0.0063384524546563625, + 0.009068960323929787, + 0.008065708912909031, + 0.006533510982990265, + 0.00550474226474762, + 0.003829322522506118, + 0.005030033644288778, + 0.004756715148687363, + 0.028022324666380882, + 0.01767832040786743, + 0.014803458005189896, + 0.011845888569951057, + 0.010353295132517815, + 0.016576534137129784, + 0.018393659964203835, + 0.015225086361169815, + 0.078091099858284, + 0.053106606006622314, + 0.037640489637851715, + 0.024508094415068626, + 0.04789075627923012, + 0.034647125750780106, + 0.045657604932785034, + 0.025229914113879204, + 0.01010073535144329, + 0.004229643382132053, + 0.003797036362811923, + 0.007876279763877392, + 0.007797191385179758, + 0.0038463331293314695, + 0.005105128046125174, + 0.005287091247737408, + 0.018887514248490334, + 0.007781023625284433, + 0.00975178461521864, + 0.006681547965854406, + 0.0068830219097435474, + 0.008149854838848114, + 0.015426918864250183, + 0.010258440859615803, + 0.003653064602985978, + 0.0035214030649513006, + 0.004129263572394848, + 0.004537336993962526, + 0.007600743323564529, + 0.005361004266887903, + 0.005549840163439512, + 0.005147046409547329, + 0.006009741686284542, + 0.008769826963543892, + 0.004999232944101095, + 0.004323968663811684, + 0.00395998964086175, + 0.0061488766223192215, + 0.008313169702887535, + 0.008688032627105713, + 0.013161477632820606, + 0.018312394618988037, + 0.006914679892361164, + 0.006576562765985727, + 0.006612811237573624, + 0.00813255924731493, + 0.009586820378899574, + 0.014928046613931656 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.0010419422760605812 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634189836680889 + ], + "opposite_logprob_tokens": [ + -3.6513419151306152 + ], + "logprob_reward": -0.02634189836680889, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634189836680889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048397451639175415, + 0.019971752539277077, + 0.03897586092352867, + 0.022889550775289536, + 0.38641467690467834, + 0.016325607895851135, + 0.023606732487678528, + 0.07837309688329697, + 0.11342934519052505, + 0.2516159415245056 + ], + "credit_map": [ + 0.004973686300218105, + 0.005742300767451525, + 0.005747654940932989, + 0.007971180602908134, + 0.008847011253237724, + 0.003978051710873842, + 0.006505848374217749, + 0.0046317121013998985, + 0.00301030813716352, + 0.003823606064543128, + 0.0024733152240514755, + 0.0025397874414920807, + 0.0027254780288785696, + 0.0015423796139657497, + 0.002118028700351715, + 0.0017388475826010108, + 0.006424490828067064, + 0.006404147949069738, + 0.005002462770789862, + 0.006131893955171108, + 0.003911539912223816, + 0.002631210023537278, + 0.005348668433725834, + 0.0031214402988553047, + 0.005766232963651419, + 0.0024964159820228815, + 0.003765589091926813, + 0.0026136934757232666, + 0.0019456482259556651, + 0.0018238738412037492, + 0.0027047605253756046, + 0.0017733355052769184, + 0.08717871457338333, + 0.06872805953025818, + 0.04043354094028473, + 0.02507241815328598, + 0.05023074895143509, + 0.03752340376377106, + 0.0494505949318409, + 0.027797173708677292, + 0.002104551298543811, + 0.0021547535434365273, + 0.0017368922708556056, + 0.001654524472542107, + 0.003479492152109742, + 0.0013359944568946958, + 0.001869434374384582, + 0.0019899639301002026, + 0.0019439831376075745, + 0.0019040302140638232, + 0.0018481137230992317, + 0.0023135116789489985, + 0.006709249224513769, + 0.002735268557444215, + 0.0028420300222933292, + 0.0033105441834777594, + 0.012409165501594543, + 0.007772400509566069, + 0.007569292560219765, + 0.006119153928011656, + 0.01586836762726307, + 0.01120680384337902, + 0.008494389243423939, + 0.008933515287935734, + 0.010408422909677029, + 0.016294827684760094, + 0.0097114946693182, + 0.01129602175205946, + 0.012680089101195335, + 0.013473243452608585, + 0.027553413063287735, + 0.012011817656457424, + 0.036853060126304626, + 0.04184513911604881, + 0.02223321795463562, + 0.014889257960021496, + 0.019029736518859863, + 0.020545726642012596, + 0.04398634284734726, + 0.052233465015888214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.02634189836680889 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176065757870674 + ], + "opposite_logprob_tokens": [ + -4.018176078796387 + ], + "logprob_reward": -0.018176065757870674, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018176065757870674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08129460364580154, + 0.019763866439461708, + 0.01415456086397171, + 0.0432891771197319, + 0.4828915596008301, + 0.14338776469230652, + 0.05231954902410507, + 0.052911922335624695, + 0.07400976121425629, + 0.03597722947597504 + ], + "credit_map": [ + 0.007174106314778328, + 0.00694437138736248, + 0.009461387991905212, + 0.015254970639944077, + 0.014279302209615707, + 0.007536010816693306, + 0.009710392914712429, + 0.010934056714177132, + 0.0028072029817849398, + 0.0019535517785698175, + 0.003343455959111452, + 0.002550694392994046, + 0.002614774974063039, + 0.0015405130106955767, + 0.002622955711558461, + 0.002330717397853732, + 0.0018384645227342844, + 0.001662523951381445, + 0.002161423908546567, + 0.001955246552824974, + 0.0020571614149957895, + 0.0011306770611554384, + 0.0020263250917196274, + 0.0013227378949522972, + 0.00946177076548338, + 0.013222231529653072, + 0.002798810601234436, + 0.003215974662452936, + 0.0043442402966320515, + 0.003463316708803177, + 0.0038831178098917007, + 0.0028997163753956556, + 0.10568394511938095, + 0.08156949281692505, + 0.05150613933801651, + 0.03293271362781525, + 0.0695284828543663, + 0.04617219418287277, + 0.060554035007953644, + 0.034944575279951096, + 0.030787350609898567, + 0.014903627336025238, + 0.014068887569010258, + 0.015888366848230362, + 0.014403455890715122, + 0.013154162093997002, + 0.02655307948589325, + 0.013628832064568996, + 0.006766248028725386, + 0.005639163311570883, + 0.006370767019689083, + 0.00832456722855568, + 0.007029156666249037, + 0.006419527344405651, + 0.005777574144303799, + 0.005992545746266842, + 0.0080023892223835, + 0.009363148361444473, + 0.006040635518729687, + 0.004916210658848286, + 0.007683728821575642, + 0.007049884181469679, + 0.004399859346449375, + 0.005456067156046629, + 0.005723708309233189, + 0.00671763252466917, + 0.004263796843588352, + 0.006625453941524029, + 0.005747351795434952, + 0.004673785995692015, + 0.026962075382471085, + 0.013295959681272507, + 0.0028421503957360983, + 0.004980101715773344, + 0.003028259612619877, + 0.0031418190337717533, + 0.00438955333083868, + 0.0030184586066752672, + 0.008643782697618008, + 0.005933104548603296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.018176065757870674 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tracking shot, shot from a slightly upward angle, follows a motorcycle in profile, the scene imbued with a rich, vivid painterly style, its brushstrokes clearly visible. The vintage cruiser, painted a deep crimson, gleams in the golden evening sun, its chrome engine and exhaust reflecting impasto-like highlights. The rider, clad in a wrinkled black leather jacket and a matte black open-face helmet, leans slightly forward, his hands steady on the handlebars, his posture focused and resolute. The motorcycle travels at a steady, powerful pace along a winding country road, its surface composed of rough brushstrokes of gray and brown. On one side, golden wheat fields sway in the wind, on the other, rolling green hills in the distance, while the sky is a magnificent canvas of bold, swirling oranges, purples, and deep blues. The low-hanging sun casts long, twisting shadows across the ground, and the wheels' rotation seems to stir the paint on the canvas. The entire picture is full of strong textures and saturated colors, creating a dynamic atmosphere that combines freedom, strength and rich artistic flavor.", + "seed": 15756, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_1/index_2361_seed_15756_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02955097397789359, + "vlm_logprob": -0.02955097397789359, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.02955097397789359, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08113549649715424 + ], + "opposite_logprob_tokens": [ + -2.5811355113983154 + ], + "logprob_reward": -0.08113549649715424, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08113549649715424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.050092414021492004, + 0.05332551151514053, + 0.25280290842056274, + 0.06867311894893646, + 0.15467505156993866, + 0.094668447971344, + 0.19738198816776276, + 0.05793004855513573, + 0.021701117977499962, + 0.048749376088380814 + ], + "credit_map": [ + 0.005755005870014429, + 0.005525322165340185, + 0.004144734237343073, + 0.006902117282152176, + 0.005923128221184015, + 0.006434939801692963, + 0.008837235160171986, + 0.006569927092641592, + 0.013783067464828491, + 0.003331855870783329, + 0.0034814225509762764, + 0.004485563840717077, + 0.005291726905852556, + 0.0031468623783439398, + 0.004477062728255987, + 0.015327947214245796, + 0.06884663552045822, + 0.01775263249874115, + 0.01749683916568756, + 0.013867619447410107, + 0.022449661046266556, + 0.07212569564580917, + 0.015138168819248676, + 0.025125637650489807, + 0.012698208913207054, + 0.007847288623452187, + 0.005179086700081825, + 0.007176476996392012, + 0.004940955899655819, + 0.014963868074119091, + 0.009522723034024239, + 0.0063445111736655235, + 0.011618904769420624, + 0.012323050759732723, + 0.01626833714544773, + 0.018680281937122345, + 0.012839086353778839, + 0.018711132928729057, + 0.04320769011974335, + 0.021026575937867165, + 0.010332909412682056, + 0.008938103914260864, + 0.010655652731657028, + 0.018786339089274406, + 0.007280547637492418, + 0.015575578436255455, + 0.013264596462249756, + 0.009834724478423595, + 0.02198331244289875, + 0.018887916579842567, + 0.04221563786268234, + 0.015115015208721161, + 0.01628652587532997, + 0.02667548693716526, + 0.027381468564271927, + 0.02883661910891533, + 0.006207500584423542, + 0.006323415320366621, + 0.007674320600926876, + 0.006131272297352552, + 0.0063255284912884235, + 0.00570380175486207, + 0.008355902507901192, + 0.0112083088606596, + 0.002631670795381069, + 0.002887418493628502, + 0.004240058362483978, + 0.002657544333487749, + 0.0028013093397021294, + 0.0016097795451059937, + 0.0030827356968075037, + 0.0017906019929796457, + 0.005766111891716719, + 0.005007550120353699, + 0.0049309637397527695, + 0.006268338765949011, + 0.008954666554927826, + 0.007038382347673178, + 0.004921597428619862, + 0.0058617671020329 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 4 + ], + "selected_reward": -0.08113549649715424 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018198776990175247 + ], + "opposite_logprob_tokens": [ + -4.0181989669799805 + ], + "logprob_reward": -0.018198776990175247, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018198776990175247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07527326792478561, + 0.08838409185409546, + 0.24391032755374908, + 0.025683004409074783, + 0.2953473627567291, + 0.03640686720609665, + 0.12573997676372528, + 0.029307061806321144, + 0.02548057772219181, + 0.05446745082736015 + ], + "credit_map": [ + 0.010279328562319279, + 0.004797943867743015, + 0.00826977752149105, + 0.006494201719760895, + 0.008364562876522541, + 0.012520486488938332, + 0.00880483165383339, + 0.01574213244020939, + 0.017243923619389534, + 0.008475187234580517, + 0.005954364780336618, + 0.010413545183837414, + 0.008725462481379509, + 0.009073385037481785, + 0.010614326223731041, + 0.017883900552988052, + 0.05794670432806015, + 0.023632986471056938, + 0.013222374022006989, + 0.015083406120538712, + 0.020920949056744576, + 0.07292769104242325, + 0.016360661014914513, + 0.023815566673874855, + 0.003077065572142601, + 0.0041990941390395164, + 0.003743815701454878, + 0.0026421714574098587, + 0.0024678094778209925, + 0.002908459398895502, + 0.003440742613747716, + 0.003203846048563719, + 0.04374690353870392, + 0.019935958087444305, + 0.034338172525167465, + 0.01877426914870739, + 0.030085047706961632, + 0.0718436911702156, + 0.032103031873703, + 0.04452028498053551, + 0.003695912193506956, + 0.003927134443074465, + 0.00380159099586308, + 0.0034874884877353907, + 0.004672112874686718, + 0.008081173524260521, + 0.004880537744611502, + 0.0038609171751886606, + 0.01306089200079441, + 0.012113718315958977, + 0.02170402742922306, + 0.011487461626529694, + 0.015440699644386768, + 0.016016704961657524, + 0.015654759481549263, + 0.02026171050965786, + 0.0031634506303817034, + 0.003567130072042346, + 0.0032411443535238504, + 0.0034221597015857697, + 0.00397940119728446, + 0.0032025729306042194, + 0.0041792429983615875, + 0.004551959224045277, + 0.002716943621635437, + 0.0033904770389199257, + 0.003815016243606806, + 0.003228421788662672, + 0.003570111934095621, + 0.003309203078970313, + 0.0026492448523640633, + 0.0028011598624289036, + 0.008011212572455406, + 0.005918573588132858, + 0.006063762120902538, + 0.006746275816112757, + 0.008471989072859287, + 0.006468979641795158, + 0.004151970148086548, + 0.008634687401354313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.018198776990175247 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015123370103538036 + ], + "opposite_logprob_tokens": [ + -6.50151252746582 + ], + "logprob_reward": -0.0015123370103538036, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015123370103538036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11657710373401642, + 0.15746690332889557, + 0.1527096927165985, + 0.03193305805325508, + 0.22171875834465027, + 0.037594497203826904, + 0.07818643003702164, + 0.06527569144964218, + 0.031915269792079926, + 0.1066226214170456 + ], + "credit_map": [ + 0.01574224792420864, + 0.005729555152356625, + 0.010328113101422787, + 0.008255882188677788, + 0.01702767238020897, + 0.021453430876135826, + 0.010010645724833012, + 0.028029559180140495, + 0.04728942736983299, + 0.006640701089054346, + 0.01066161785274744, + 0.005155185237526894, + 0.012822399847209454, + 0.010880505666136742, + 0.010310117155313492, + 0.053706955164670944, + 0.044137142598629, + 0.007878179661929607, + 0.011198658496141434, + 0.006732531823217869, + 0.011288410052657127, + 0.04842003062367439, + 0.007157564163208008, + 0.01589718461036682, + 0.0058282543905079365, + 0.00260946131311357, + 0.004679787904024124, + 0.003308299696072936, + 0.0024091354571282864, + 0.005078702699393034, + 0.0033069418277591467, + 0.004712472669780254, + 0.03928646072745323, + 0.01453925296664238, + 0.020632939413189888, + 0.013963328674435616, + 0.02067345380783081, + 0.0687052458524704, + 0.014727666974067688, + 0.02919040620326996, + 0.002769731916487217, + 0.004057725425809622, + 0.00435588788241148, + 0.00646804366260767, + 0.003625426674261689, + 0.004335269331932068, + 0.005955856293439865, + 0.006026553921401501, + 0.006912083365023136, + 0.008415515534579754, + 0.010088798590004444, + 0.007447769399732351, + 0.008922341279685497, + 0.013007480651140213, + 0.01153628434985876, + 0.011856156401336193, + 0.0069628083147108555, + 0.007603371981531382, + 0.009066540747880936, + 0.0077687702141702175, + 0.007625048514455557, + 0.005739681888371706, + 0.009476731531322002, + 0.011032736860215664, + 0.0038886242546141148, + 0.004190023988485336, + 0.004239722155034542, + 0.005065394565463066, + 0.004557081498205662, + 0.002271901583299041, + 0.0038510560989379883, + 0.0038514649495482445, + 0.011971982195973396, + 0.010170468129217625, + 0.0147333312779665, + 0.012245865538716316, + 0.01478757057338953, + 0.012205938808619976, + 0.009104011580348015, + 0.021403441205620766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.0015123370103538036 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575299859046936 + ], + "opposite_logprob_tokens": [ + -3.8955752849578857 + ], + "logprob_reward": -0.020575299859046936, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575299859046936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05124147981405258, + 0.38151195645332336, + 0.1095178872346878, + 0.07464447617530823, + 0.1377517282962799, + 0.015848567709326744, + 0.08598082512617111, + 0.02852223813533783, + 0.06370512396097183, + 0.05127565190196037 + ], + "credit_map": [ + 0.007437373511493206, + 0.003007314633578062, + 0.003637796500697732, + 0.0040836636908352375, + 0.00617199158295989, + 0.00567694241181016, + 0.007874608971178532, + 0.01335178967565298, + 0.11482690274715424, + 0.015636499971151352, + 0.025213655084371567, + 0.012282632291316986, + 0.034906670451164246, + 0.026858855038881302, + 0.02536173351109028, + 0.12642499804496765, + 0.02977830357849598, + 0.004644499626010656, + 0.008257851004600525, + 0.004634220153093338, + 0.008137673139572144, + 0.03750025853514671, + 0.003693017642945051, + 0.012872064486145973, + 0.009385545738041401, + 0.0063370405696332455, + 0.013300823047757149, + 0.007427219767123461, + 0.004932981915771961, + 0.01239179540425539, + 0.006415803451091051, + 0.014453260228037834, + 0.021050559356808662, + 0.009052691049873829, + 0.016310101374983788, + 0.00846092775464058, + 0.01317815575748682, + 0.03555132821202278, + 0.013094714842736721, + 0.0210532546043396, + 0.0020139366388320923, + 0.0016125766560435295, + 0.0024107114877551794, + 0.002192582469433546, + 0.0016932743601500988, + 0.002772926352918148, + 0.0017694210400804877, + 0.001383140217512846, + 0.00836093444377184, + 0.008597040548920631, + 0.012059811502695084, + 0.008327065967023373, + 0.010728476569056511, + 0.011182921007275581, + 0.010574682615697384, + 0.016149889677762985, + 0.002538001397624612, + 0.003382343566045165, + 0.0031473319977521896, + 0.0034517506137490273, + 0.003585705766454339, + 0.002711513079702854, + 0.004263905808329582, + 0.005441684741526842, + 0.00795911904424429, + 0.009490525349974632, + 0.011637374758720398, + 0.009120863862335682, + 0.008313771337270737, + 0.004993153735995293, + 0.00575917586684227, + 0.006431139539927244, + 0.007593558169901371, + 0.007464043330401182, + 0.009162076748907566, + 0.004685817752033472, + 0.007892164401710033, + 0.004664659965783358, + 0.004096912685781717, + 0.005716416984796524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.020575299859046936 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a vintage cruiser motorcycle painted deep crimson?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the rider wearing a wrinkled black leather jacket and a matte black open-face helmet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene rendered in a rich, vivid painterly style with clearly visible brushstrokes and impasto-like highlights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the setting feature a winding country road with golden wheat fields on one side and rolling green hills in the distance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a slightly upward angle tracking the motorcycle in profile?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026332959532737732 + ], + "opposite_logprob_tokens": [ + -3.6513328552246094 + ], + "logprob_reward": -0.026332959532737732, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026332959532737732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09291993081569672, + 0.043045271188020706, + 0.6127968430519104, + 0.029476908966898918, + 0.06636510789394379, + 0.017892371863126755, + 0.06023279204964638, + 0.018147587776184082, + 0.015640782192349434, + 0.043482456356287 + ], + "credit_map": [ + 0.015467210672795773, + 0.005913592409342527, + 0.0076094684191048145, + 0.007239270489662886, + 0.011489425785839558, + 0.015558351762592793, + 0.010247865691781044, + 0.019394736737012863, + 0.009578824043273926, + 0.0038024818059056997, + 0.0030807522125542164, + 0.003227849956601858, + 0.005859813187271357, + 0.0033445642329752445, + 0.0034393626265227795, + 0.010711622424423695, + 0.1872493475675583, + 0.025850659236311913, + 0.03650248795747757, + 0.023343458771705627, + 0.048500869423151016, + 0.20805874466896057, + 0.024484751746058464, + 0.05880647152662277, + 0.004278488922864199, + 0.00293197319842875, + 0.0036738624330610037, + 0.0029484096448868513, + 0.0020864515099674463, + 0.005405377596616745, + 0.0029356940649449825, + 0.005216649267822504, + 0.010855796746909618, + 0.005643701646476984, + 0.007650758605450392, + 0.005156068596988916, + 0.0072591365315020084, + 0.018143920227885246, + 0.004856642335653305, + 0.0067990729585289955, + 0.0016330670332536101, + 0.002316672122105956, + 0.0015551490942016244, + 0.00232712528668344, + 0.002945126499980688, + 0.002847404219210148, + 0.002296908525750041, + 0.0019709179177880287, + 0.005402420647442341, + 0.007721361704170704, + 0.008947948925197124, + 0.006715922150760889, + 0.006842573639005423, + 0.009008497931063175, + 0.0076556941494345665, + 0.007938366383314133, + 0.0018041882431134582, + 0.0024147082585841417, + 0.0027254933957010508, + 0.002565680304542184, + 0.0031166556291282177, + 0.0016616006614640355, + 0.002413923153653741, + 0.0014453354524448514, + 0.0015459258574992418, + 0.001720250234939158, + 0.002391914837062359, + 0.002202336909249425, + 0.0019734418019652367, + 0.0018901553703472018, + 0.002212497405707836, + 0.0017042587278410792, + 0.006036209873855114, + 0.003776853671297431, + 0.005124069284647703, + 0.004528232384473085, + 0.009854655712842941, + 0.005956652574241161, + 0.0039632744155824184, + 0.0042425040155649185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.026332959532737732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_10/index_tensor([1036, 1036])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_10/index_tensor([1036, 1036])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9a583dfcaef2daf071ba44f730abb97b792bfa09 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_10/index_tensor([1036, 1036])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a stunning macro close-up shot, the camera tilts down at an almost imperceptible speed, focusing on an exquisite antique ring as it begins to slide down a sheet of deep sapphire silk, stretched across its gently folds. Crafted from gold with a warm, aged patina, the ring's band is engraved with intricate, delicate hand-engraved filigree patterns, revealing the marks of meticulous craftsmanship etched over time. At the ring's apex, a large, lustrous cream pearl rests securely in an elegant prong setting, its surface reflecting a soft, diffuse halo of light. With the gentle pull of gravity, the ring begins its graceful, unstoppable descent, skimming almost frictionlessly across the silk's smooth surface, sliding steadily and deliberately down one of its gentle slopes. A single, warm light source, emanating from the side, casts long, soft shadows across the scene, converging into a bright highlight on the pearl's curved surface and outlining the delicate contours of the gold filigree. The whole picture is immersed in a quiet, slightly melancholic and elegant atmosphere, full of the texture of time and silent stories.", + "seed": 17033, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_10/index_1036_seed_17033_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.054233203362673524, + "vlm_logprob": -0.054233203362673524, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.054233203362673524, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0064387014135718346 + ], + "opposite_logprob_tokens": [ + -5.631438732147217 + ], + "logprob_reward": -0.0064387014135718346, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0064387014135718346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3601456582546234, + 0.06703776866197586, + 0.06876785308122635, + 0.027884235605597496, + 0.08828210085630417, + 0.080401711165905, + 0.03293420374393463, + 0.030916595831513405, + 0.053464796394109726, + 0.1901649683713913 + ], + "credit_map": [ + 0.025817932561039925, + 0.0340704619884491, + 0.05180446058511734, + 0.06055770441889763, + 0.03960468992590904, + 0.05265680328011513, + 0.05670369789004326, + 0.03892989456653595, + 0.005199367180466652, + 0.005516136530786753, + 0.009191107004880905, + 0.016394352540373802, + 0.008396274410188198, + 0.008778424933552742, + 0.008644334971904755, + 0.004917770624160767, + 0.005823875777423382, + 0.0031110169366002083, + 0.016691243276000023, + 0.006692614406347275, + 0.006294337101280689, + 0.006084642838686705, + 0.007984306663274765, + 0.016085823997855186, + 0.0029543705750256777, + 0.002905570901930332, + 0.004932606592774391, + 0.0029918537475168705, + 0.003277734387665987, + 0.0035492207389324903, + 0.0029051187448203564, + 0.00436776177957654, + 0.008193878456950188, + 0.006955495569854975, + 0.023390671238303185, + 0.008377688005566597, + 0.009167157113552094, + 0.009735609404742718, + 0.00911603681743145, + 0.013345560990273952, + 0.005586258135735989, + 0.0028846291825175285, + 0.027604877948760986, + 0.005454512313008308, + 0.003098153043538332, + 0.0038271343801170588, + 0.005501868668943644, + 0.02644428424537182, + 0.004273910075426102, + 0.002849854063242674, + 0.004060059320181608, + 0.00392854493111372, + 0.005074833054095507, + 0.004634752869606018, + 0.0041394284926354885, + 0.003972823265939951, + 0.004306686110794544, + 0.0027055309619754553, + 0.0040755486115813255, + 0.0047104209661483765, + 0.005577555857598782, + 0.0025257172528654337, + 0.0031012066174298525, + 0.003913926891982555, + 0.009184492751955986, + 0.003766697132959962, + 0.0073628914542496204, + 0.005719716660678387, + 0.006001465953886509, + 0.007674707565456629, + 0.005115249659866095, + 0.00863957591354847, + 0.020356036722660065, + 0.017322272062301636, + 0.02378244325518608, + 0.0260036401450634, + 0.02680962160229683, + 0.023087341338396072, + 0.02683688886463642, + 0.025966718792915344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.0064387014135718346 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004110222216695547 + ], + "opposite_logprob_tokens": [ + -5.504110336303711 + ], + "logprob_reward": -0.004110222216695547, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004110222216695547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1361338347196579, + 0.2042393982410431, + 0.09961845725774765, + 0.0709320679306984, + 0.06058025360107422, + 0.15265649557113647, + 0.041833631694316864, + 0.03323571756482124, + 0.03106963261961937, + 0.16970039904117584 + ], + "credit_map": [ + 0.010213389061391354, + 0.014213643968105316, + 0.01887517422437668, + 0.022890273481607437, + 0.012070290744304657, + 0.019348513334989548, + 0.02258406952023506, + 0.015938496217131615, + 0.022194411605596542, + 0.019850146025419235, + 0.026101911440491676, + 0.03736148774623871, + 0.022449098527431488, + 0.030789470300078392, + 0.02972695603966713, + 0.015765950083732605, + 0.009563429281115532, + 0.005451384466141462, + 0.02683934010565281, + 0.010784878395497799, + 0.007810947950929403, + 0.007412551902234554, + 0.010285134427249432, + 0.02147080935537815, + 0.008133194409310818, + 0.006603316869586706, + 0.010525226593017578, + 0.0073204198852181435, + 0.009522711858153343, + 0.010233189910650253, + 0.006340932101011276, + 0.012253081426024437, + 0.006930888630449772, + 0.004768593702465296, + 0.011211695149540901, + 0.006664568558335304, + 0.007687411271035671, + 0.00830694567412138, + 0.007331373170018196, + 0.007678783964365721, + 0.011766734533011913, + 0.0055374507792294025, + 0.04869278147816658, + 0.009722542017698288, + 0.006691865157335997, + 0.01326743233948946, + 0.010151282884180546, + 0.04682642221450806, + 0.006833957973867655, + 0.004168614279478788, + 0.006304183974862099, + 0.004551993682980537, + 0.004910472314804792, + 0.004485448356717825, + 0.004987288732081652, + 0.005591672845184803, + 0.0038944873958826065, + 0.0023406981490552425, + 0.00385898444801569, + 0.004832118283957243, + 0.0046003530733287334, + 0.004914537537842989, + 0.004043907392770052, + 0.0047506350092589855, + 0.006027041468769312, + 0.0021204324439167976, + 0.003313555847853422, + 0.003891734639182687, + 0.0034381437581032515, + 0.003780202241614461, + 0.004615503828972578, + 0.0038830204866826534, + 0.018054720014333725, + 0.016044216230511665, + 0.01862443797290325, + 0.02660403773188591, + 0.023680588230490685, + 0.020378930494189262, + 0.022849617525935173, + 0.023463871330022812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 5 + ], + "selected_reward": -0.004110222216695547 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026384394615888596 + ], + "opposite_logprob_tokens": [ + -3.6513843536376953 + ], + "logprob_reward": -0.026384394615888596, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026384394615888596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1597503274679184, + 0.07334411144256592, + 0.21430985629558563, + 0.03861088305711746, + 0.11611152440309525, + 0.13233327865600586, + 0.05258005112409592, + 0.05854782834649086, + 0.04707116261124611, + 0.10734102129936218 + ], + "credit_map": [ + 0.01196659542620182, + 0.013071347959339619, + 0.026717061176896095, + 0.03112551011145115, + 0.013113150373101234, + 0.01976575143635273, + 0.026883799582719803, + 0.017107097432017326, + 0.005505757872015238, + 0.007484888657927513, + 0.009360061958432198, + 0.01602107658982277, + 0.011507922783493996, + 0.008604618720710278, + 0.008753739297389984, + 0.006106043234467506, + 0.018187131732702255, + 0.009495098143815994, + 0.06868300586938858, + 0.015245557762682438, + 0.01454828679561615, + 0.01917603611946106, + 0.013369496911764145, + 0.05560522526502609, + 0.004195447079837322, + 0.003988933749496937, + 0.00615389458835125, + 0.005207262467592955, + 0.003768965369090438, + 0.0047748880460858345, + 0.004725969862192869, + 0.005795517936348915, + 0.013083323836326599, + 0.010323862545192242, + 0.01899787038564682, + 0.01346495933830738, + 0.012735666707158089, + 0.013143308460712433, + 0.013134563341736794, + 0.021227966994047165, + 0.012191577814519405, + 0.0041369302198290825, + 0.04016012325882912, + 0.0095734978094697, + 0.004886127542704344, + 0.007906466722488403, + 0.008242093026638031, + 0.045236457139253616, + 0.008072391152381897, + 0.004182053729891777, + 0.007969137281179428, + 0.006245502270758152, + 0.005483985412865877, + 0.006658014375716448, + 0.0058047655038535595, + 0.008164196275174618, + 0.007597693707793951, + 0.004590652417391539, + 0.00655027199536562, + 0.006780481431633234, + 0.009527290239930153, + 0.010425935499370098, + 0.005350097548216581, + 0.007725405041128397, + 0.006868717726320028, + 0.004354800563305616, + 0.004974480252712965, + 0.0064451307989656925, + 0.007184337358921766, + 0.004936470650136471, + 0.006673169322311878, + 0.005634052213281393, + 0.009621623903512955, + 0.00900418683886528, + 0.013239902444183826, + 0.016969121992588043, + 0.014944487251341343, + 0.009663375094532967, + 0.020616542547941208, + 0.013281780295073986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.026384394615888596 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864111352711916 + ], + "opposite_logprob_tokens": [ + -4.758641242980957 + ], + "logprob_reward": -0.00864111352711916, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864111352711916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29664698243141174, + 0.08256740868091583, + 0.09059727936983109, + 0.12072862684726715, + 0.07283942401409149, + 0.11299499124288559, + 0.01994490437209606, + 0.06527817249298096, + 0.036792024970054626, + 0.10161019116640091 + ], + "credit_map": [ + 0.02173755317926407, + 0.028774118050932884, + 0.04035086929798126, + 0.05034000426530838, + 0.03269438073039055, + 0.040796101093292236, + 0.0463399775326252, + 0.035613980144262314, + 0.009221525862812996, + 0.0062723662704229355, + 0.01450397539883852, + 0.013133121654391289, + 0.009784791618585587, + 0.009855574928224087, + 0.012237537652254105, + 0.007558519020676613, + 0.0114248376339674, + 0.006664969492703676, + 0.014303464442491531, + 0.01738850586116314, + 0.00791021529585123, + 0.00800198595970869, + 0.013583989813923836, + 0.011319311335682869, + 0.011208080686628819, + 0.006780676543712616, + 0.030878353863954544, + 0.015517215244472027, + 0.008238173089921474, + 0.013745507225394249, + 0.014469414949417114, + 0.019891204312443733, + 0.007955817505717278, + 0.0050935591571033, + 0.015571694821119308, + 0.009040589444339275, + 0.005535185802727938, + 0.007880146615207195, + 0.006513003259897232, + 0.015249424614012241, + 0.007857599295675755, + 0.003741468070074916, + 0.038326650857925415, + 0.008100029081106186, + 0.0038547737058252096, + 0.004506660625338554, + 0.007346570026129484, + 0.039261236786842346, + 0.0021562662441283464, + 0.0017329982947558165, + 0.003588965628296137, + 0.0024840529076755047, + 0.002105769934132695, + 0.0025170466396957636, + 0.0023538803216069937, + 0.003005924401804805, + 0.010375483892858028, + 0.005879578646272421, + 0.006700281985104084, + 0.007832158356904984, + 0.009301119484007359, + 0.0095210000872612, + 0.005901943426579237, + 0.00976660754531622, + 0.00753808906301856, + 0.00240720366127789, + 0.004391040652990341, + 0.003804153995588422, + 0.0037329767365008593, + 0.004084768705070019, + 0.006130935624241829, + 0.004702859092503786, + 0.013944927603006363, + 0.009499196894466877, + 0.010383914224803448, + 0.011973307467997074, + 0.01476222276687622, + 0.011269317008554935, + 0.013721956871449947, + 0.01605534739792347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.00864111352711916 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559158504009247 + ], + "opposite_logprob_tokens": [ + -1.6005915403366089 + ], + "logprob_reward": -0.22559158504009247, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559158504009247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07126474380493164, + 0.08364564925432205, + 0.16808658838272095, + 0.13083890080451965, + 0.08941429853439331, + 0.05351778119802475, + 0.02875874564051628, + 0.127383291721344, + 0.17046993970870972, + 0.07661997526884079 + ], + "credit_map": [ + 0.007901252247393131, + 0.008574465289711952, + 0.007098549976944923, + 0.010277314111590385, + 0.008436247706413269, + 0.008870461024343967, + 0.011480247601866722, + 0.008626209571957588, + 0.007911822758615017, + 0.00768086314201355, + 0.01245177537202835, + 0.015177934430539608, + 0.0103961993008852, + 0.01117549929767847, + 0.011267541907727718, + 0.007584024220705032, + 0.015761282294988632, + 0.007615538313984871, + 0.05054868012666702, + 0.013741767965257168, + 0.010002467781305313, + 0.0158559437841177, + 0.011872024275362492, + 0.042688898742198944, + 0.013141022995114326, + 0.012358512729406357, + 0.023155128583312035, + 0.0171176977455616, + 0.018018145114183426, + 0.01716017909348011, + 0.013358024880290031, + 0.016530200839042664, + 0.008199639618396759, + 0.0069257598370313644, + 0.018900210037827492, + 0.009176448918879032, + 0.009256424382328987, + 0.011369505897164345, + 0.007888155058026314, + 0.017698165029287338, + 0.004568194504827261, + 0.005088365636765957, + 0.011440237052738667, + 0.007426097057759762, + 0.0050259800627827644, + 0.005117448046803474, + 0.004353441298007965, + 0.010498027317225933, + 0.003973113372921944, + 0.00338439317420125, + 0.004483633674681187, + 0.0029589172918349504, + 0.0039023023564368486, + 0.002718183444812894, + 0.0034647465217858553, + 0.0038734597619622946, + 0.019546829164028168, + 0.008024505339562893, + 0.014157655648887157, + 0.015384426340460777, + 0.015917522832751274, + 0.019185595214366913, + 0.013142190873622894, + 0.022024579346179962, + 0.026798268780112267, + 0.011092432774603367, + 0.022609226405620575, + 0.01909220777451992, + 0.020283229649066925, + 0.02396373823285103, + 0.024893613532185555, + 0.021737249568104744, + 0.011593414470553398, + 0.009017132222652435, + 0.0070135281421244144, + 0.009339777752757072, + 0.01197376661002636, + 0.007936052046716213, + 0.008811124600470066, + 0.01093518640846014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 3 + ], + "selected_reward": -0.22559158504009247 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning macro close-up shot, the camera tilts down at an almost imperceptible speed, focusing on an exquisite antique ring as it begins to slide down a sheet of deep sapphire silk, stretched across its gently folds. Crafted from gold with a warm, aged patina, the ring's band is engraved with intricate, delicate hand-engraved filigree patterns, revealing the marks of meticulous craftsmanship etched over time. At the ring's apex, a large, lustrous cream pearl rests securely in an elegant prong setting, its surface reflecting a soft, diffuse halo of light. With the gentle pull of gravity, the ring begins its graceful, unstoppable descent, skimming almost frictionlessly across the silk's smooth surface, sliding steadily and deliberately down one of its gentle slopes. A single, warm light source, emanating from the side, casts long, soft shadows across the scene, converging into a bright highlight on the pearl's curved surface and outlining the delicate contours of the gold filigree. The whole picture is immersed in a quiet, slightly melancholic and elegant atmosphere, full of the texture of time and silent stories.", + "seed": 17034, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_10/index_1036_seed_17034_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.033472828869707885, + "vlm_logprob": -0.033472828869707885, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.033472828869707885, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005613515619188547 + ], + "opposite_logprob_tokens": [ + -5.505613327026367 + ], + "logprob_reward": -0.005613515619188547, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005613515619188547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02289280667901039, + 0.062497761100530624, + 0.03425423428416252, + 0.15593992173671722, + 0.019266530871391296, + 0.08242736756801605, + 0.25493094325065613, + 0.254549115896225, + 0.09656470268964767, + 0.016676610335707664 + ], + "credit_map": [ + 0.0020523061975836754, + 0.00332733360119164, + 0.0035347845405340195, + 0.0031624664552509785, + 0.002668965607881546, + 0.0028433585539460182, + 0.0022534984163939953, + 0.003050093771889806, + 0.0036707192193716764, + 0.0024762845132499933, + 0.012792054563760757, + 0.002454638946801424, + 0.003365351352840662, + 0.003061100607737899, + 0.022092115134000778, + 0.01258549839258194, + 0.0041421507485210896, + 0.00731487525627017, + 0.00399111257866025, + 0.002589339157566428, + 0.004481381271034479, + 0.003786802524700761, + 0.0052872467786073685, + 0.0026613252703100443, + 0.025824930518865585, + 0.009273280389606953, + 0.01118852011859417, + 0.021526020020246506, + 0.00596286915242672, + 0.008823409676551819, + 0.0493527427315712, + 0.023988138884305954, + 0.003595762886106968, + 0.001829040702432394, + 0.0018990575335919857, + 0.0013230032054707408, + 0.0013287231558933854, + 0.0033622346818447113, + 0.0023854607716202736, + 0.0035432479344308376, + 0.010838404297828674, + 0.007542911451309919, + 0.023009704425930977, + 0.006712932139635086, + 0.008052037097513676, + 0.00983071606606245, + 0.0065308623015880585, + 0.009909795597195625, + 0.013714061118662357, + 0.009482314810156822, + 0.08330156654119492, + 0.01725403405725956, + 0.009686163626611233, + 0.0196541715413332, + 0.01495844405144453, + 0.08688018471002579, + 0.04006783291697502, + 0.02059454843401909, + 0.04446561634540558, + 0.023869911208748817, + 0.02440379559993744, + 0.05179763212800026, + 0.02463466115295887, + 0.024715140461921692, + 0.010232529602944851, + 0.01507232803851366, + 0.008492588065564632, + 0.009234950877726078, + 0.013467500917613506, + 0.010499154217541218, + 0.0197242833673954, + 0.0098413722589612, + 0.0019183217082172632, + 0.0017388500273227692, + 0.00220129475928843, + 0.0019816653802990913, + 0.0026874649338424206, + 0.0016364740440621972, + 0.002792521147057414, + 0.001720018102787435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.005613515619188547 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001522692502476275 + ], + "opposite_logprob_tokens": [ + -6.501522541046143 + ], + "logprob_reward": -0.001522692502476275, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.001522692502476275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056078147143125534, + 0.20308606326580048, + 0.051622938364744186, + 0.056469906121492386, + 0.038866523653268814, + 0.10026060789823532, + 0.1252623349428177, + 0.2844778299331665, + 0.053103603422641754, + 0.030772099271416664 + ], + "credit_map": [ + 0.004991953261196613, + 0.0065747471526265144, + 0.01105866115540266, + 0.007352171000093222, + 0.006595958024263382, + 0.007480046711862087, + 0.005348660983145237, + 0.006675945594906807, + 0.012405499815940857, + 0.006642790976911783, + 0.04698595032095909, + 0.00698471162468195, + 0.006246425677090883, + 0.007169776130467653, + 0.07472030818462372, + 0.041930582374334335, + 0.0068833795376122, + 0.0053534312173724174, + 0.0076289246790111065, + 0.006390019319951534, + 0.005059502087533474, + 0.006087114103138447, + 0.008132127113640308, + 0.006088435184210539, + 0.01165307592600584, + 0.0028306073509156704, + 0.004002653528004885, + 0.00567789888009429, + 0.0023149102926254272, + 0.004244484938681126, + 0.01422426849603653, + 0.011522003449499607, + 0.009049876593053341, + 0.003440234577283263, + 0.0032924108672887087, + 0.0024774903431534767, + 0.003278734628111124, + 0.009721515700221062, + 0.0027238966431468725, + 0.004882362205535173, + 0.014123110100626945, + 0.009061946533620358, + 0.025736931711435318, + 0.009125377982854843, + 0.009857797995209694, + 0.011894886381924152, + 0.007875786162912846, + 0.012584769167006016, + 0.008393260650336742, + 0.005377184599637985, + 0.03604539856314659, + 0.008625130169093609, + 0.0049850838258862495, + 0.014324355870485306, + 0.007127794437110424, + 0.04038412123918533, + 0.04282519966363907, + 0.024648530408740044, + 0.052837833762168884, + 0.020911291241645813, + 0.023730263113975525, + 0.06571486592292786, + 0.024237362667918205, + 0.029572458937764168, + 0.005617083515971899, + 0.006599816028028727, + 0.008966082707047462, + 0.004604902118444443, + 0.007377576548606157, + 0.005787573289126158, + 0.005552348680794239, + 0.008598214946687222, + 0.003174126846715808, + 0.003281288780272007, + 0.004640277940779924, + 0.0036860036198049784, + 0.0033645883668214083, + 0.0050697047263383865, + 0.003614541608840227, + 0.00394156388938427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.001522692502476275 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009796281345188618 + ], + "opposite_logprob_tokens": [ + -4.634796142578125 + ], + "logprob_reward": -0.009796281345188618, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009796281345188618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03866618871688843, + 0.09201016277074814, + 0.04672430083155632, + 0.13008011877536774, + 0.08690184354782104, + 0.05525800958275795, + 0.18201012909412384, + 0.24329397082328796, + 0.10155696421861649, + 0.023498235270380974 + ], + "credit_map": [ + 0.00393707724288106, + 0.004989767912775278, + 0.007290896959602833, + 0.0052508823573589325, + 0.003524302737787366, + 0.0048574768006801605, + 0.004308284260332584, + 0.004507501609623432, + 0.00993292685598135, + 0.00386568671092391, + 0.03606955707073212, + 0.0037784737069159746, + 0.004118295852094889, + 0.005586028099060059, + 0.020437968894839287, + 0.008221223019063473, + 0.0066304514184594154, + 0.010304594412446022, + 0.004509259946644306, + 0.002628727350383997, + 0.006883664522320032, + 0.004191833548247814, + 0.008092978969216347, + 0.00348279089666903, + 0.02127881534397602, + 0.008132434450089931, + 0.007892164401710033, + 0.01569582335650921, + 0.0053651658818125725, + 0.007056724280118942, + 0.043163686990737915, + 0.02149529941380024, + 0.030537918210029602, + 0.003932863939553499, + 0.0052903140895068645, + 0.0035471459850668907, + 0.003914604429155588, + 0.032950542867183685, + 0.003537720302119851, + 0.00319073349237442, + 0.007089640013873577, + 0.004809233359992504, + 0.01566692255437374, + 0.00457994919270277, + 0.005397256929427385, + 0.006645947694778442, + 0.005124065559357405, + 0.00594499334692955, + 0.010022252798080444, + 0.0073803807608783245, + 0.05684543401002884, + 0.012703178450465202, + 0.008131753653287888, + 0.014895289205014706, + 0.012531823478639126, + 0.05950000882148743, + 0.03747168555855751, + 0.017985794693231583, + 0.04219008982181549, + 0.021647650748491287, + 0.02371397241950035, + 0.05149083957076073, + 0.0229271799325943, + 0.02586675062775612, + 0.012114754877984524, + 0.01665668748319149, + 0.008984283544123173, + 0.009808795526623726, + 0.013819116167724133, + 0.009545394219458103, + 0.01974816620349884, + 0.01087977085262537, + 0.002544738119468093, + 0.002418545773252845, + 0.0029779677279293537, + 0.0027678357437253, + 0.0034187347628176212, + 0.0026257846038788557, + 0.0034159687347710133, + 0.003328659338876605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.009796281345188618 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007626346778124571 + ], + "opposite_logprob_tokens": [ + -4.882626533508301 + ], + "logprob_reward": -0.007626346778124571, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007626346778124571, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038186218589544296, + 0.1151796504855156, + 0.029418939724564552, + 0.02229321375489235, + 0.09031399339437485, + 0.057798128575086594, + 0.5189698338508606, + 0.06272388994693756, + 0.04868265241384506, + 0.016433512791991234 + ], + "credit_map": [ + 0.0034514765720814466, + 0.00509723462164402, + 0.0065086157992482185, + 0.00517289387062192, + 0.004238621797412634, + 0.005303516052663326, + 0.004249996040016413, + 0.0041638626717031, + 0.012493908405303955, + 0.004608671646565199, + 0.04583269730210304, + 0.00533381849527359, + 0.0038322596810758114, + 0.005066371988505125, + 0.02866649068892002, + 0.00934543740004301, + 0.003641217714175582, + 0.006428986322134733, + 0.0033525642938911915, + 0.0022023573983460665, + 0.004804439842700958, + 0.0028320043347775936, + 0.004017761442810297, + 0.0021396088413894176, + 0.003367530182003975, + 0.002031005686149001, + 0.002295532263815403, + 0.0032928951550275087, + 0.0014703641645610332, + 0.0017300727777183056, + 0.005841025151312351, + 0.0022647902369499207, + 0.024228667840361595, + 0.0062989043071866035, + 0.008997109718620777, + 0.00629410520195961, + 0.006622061133384705, + 0.02711961418390274, + 0.0052324989810585976, + 0.005521029699593782, + 0.007214690558612347, + 0.005966393277049065, + 0.006663036532700062, + 0.00828363373875618, + 0.006383501458913088, + 0.00815138965845108, + 0.008207021281123161, + 0.006928459741175175, + 0.028390655294060707, + 0.018157685175538063, + 0.17774930596351624, + 0.03015020489692688, + 0.016997655853629112, + 0.038816336542367935, + 0.026593711227178574, + 0.18211424350738525, + 0.009697322733700275, + 0.005473157856613398, + 0.011368042789399624, + 0.005309279076755047, + 0.005197383463382721, + 0.013092372566461563, + 0.00554575864225626, + 0.007040574215352535, + 0.005209430120885372, + 0.007126485463231802, + 0.004198108334094286, + 0.0051659573800861835, + 0.007372869644314051, + 0.004363623913377523, + 0.01127779670059681, + 0.003968378528952599, + 0.0018639519112184644, + 0.0016682458808645606, + 0.0029005412943661213, + 0.0024502607993781567, + 0.0018880682764574885, + 0.0017977113602682948, + 0.0022281615529209375, + 0.0016365719493478537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.007626346778124571 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428053081035614 + ], + "opposite_logprob_tokens": [ + -2.017805337905884 + ], + "logprob_reward": -0.1428053081035614, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428053081035614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.012220540083944798, + 0.051497794687747955, + 0.010620511136949062, + 0.10363177210092545, + 0.1436842530965805, + 0.0992693081498146, + 0.5428512096405029, + 0.008986705914139748, + 0.012816261500120163, + 0.014421671628952026 + ], + "credit_map": [ + 0.0012173590948805213, + 0.001728871138766408, + 0.0019258320098742843, + 0.001994177233427763, + 0.001589256920851767, + 0.0013101400109007955, + 0.0009712179889902472, + 0.0014836846385151148, + 0.004832861945033073, + 0.0016923068324103951, + 0.01448903139680624, + 0.0019439522875472903, + 0.00174552365206182, + 0.0023326557129621506, + 0.008362392894923687, + 0.01609906181693077, + 0.0015765553107485175, + 0.0016119437059387565, + 0.0010762391611933708, + 0.0008532641222700477, + 0.0012425435706973076, + 0.0011563628213480115, + 0.002153444802388549, + 0.0009501577005721629, + 0.016054322943091393, + 0.005958770867437124, + 0.008203594945371151, + 0.014602135866880417, + 0.004106426145881414, + 0.00628718500956893, + 0.03321940079331398, + 0.015199922025203705, + 0.04878173768520355, + 0.005894298665225506, + 0.009019719436764717, + 0.0057286666706204414, + 0.00653385603800416, + 0.05366778373718262, + 0.007377331145107746, + 0.006680850870907307, + 0.013450992293655872, + 0.008369582705199718, + 0.025829294696450233, + 0.008005987852811813, + 0.009425856173038483, + 0.013222158886492252, + 0.008317647501826286, + 0.012647784315049648, + 0.03244698792695999, + 0.020230161026120186, + 0.17729103565216064, + 0.03401719778776169, + 0.019533393904566765, + 0.048988115042448044, + 0.029441803693771362, + 0.18090246617794037, + 0.001025123638100922, + 0.0008499917457811534, + 0.0013691454660147429, + 0.0009898917051032186, + 0.0009583043283782899, + 0.001539397519081831, + 0.0009242493542842567, + 0.0013306003529578447, + 0.0017373532755300403, + 0.002318567829206586, + 0.0013095609610900283, + 0.001129882293753326, + 0.0016763577004894614, + 0.0010534084867686033, + 0.002029163995757699, + 0.0015619650948792696, + 0.0016733293887227774, + 0.0016476279124617577, + 0.0031711843330413103, + 0.0017583873122930527, + 0.001422034576535225, + 0.001647423836402595, + 0.0017337616300210357, + 0.0013679205439984798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 3 + ], + "selected_reward": -0.1428053081035614 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning macro close-up shot, the camera tilts down at an almost imperceptible speed, focusing on an exquisite antique ring as it begins to slide down a sheet of deep sapphire silk, stretched across its gently folds. Crafted from gold with a warm, aged patina, the ring's band is engraved with intricate, delicate hand-engraved filigree patterns, revealing the marks of meticulous craftsmanship etched over time. At the ring's apex, a large, lustrous cream pearl rests securely in an elegant prong setting, its surface reflecting a soft, diffuse halo of light. With the gentle pull of gravity, the ring begins its graceful, unstoppable descent, skimming almost frictionlessly across the silk's smooth surface, sliding steadily and deliberately down one of its gentle slopes. A single, warm light source, emanating from the side, casts long, soft shadows across the scene, converging into a bright highlight on the pearl's curved surface and outlining the delicate contours of the gold filigree. The whole picture is immersed in a quiet, slightly melancholic and elegant atmosphere, full of the texture of time and silent stories.", + "seed": 17035, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_10/index_1036_seed_17035_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04456602497957647, + "vlm_logprob": -0.04456602497957647, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.04456602497957647, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005315575283020735 + ], + "opposite_logprob_tokens": [ + -6.005315780639648 + ], + "logprob_reward": -0.005315575283020735, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.005315575283020735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28924912214279175, + 0.052480004727840424, + 0.03882511705160141, + 0.037960588932037354, + 0.1080414205789566, + 0.02714758738875389, + 0.0467245988547802, + 0.025661999359726906, + 0.3214222490787506, + 0.05248728767037392 + ], + "credit_map": [ + 0.026152731850743294, + 0.05144887790083885, + 0.03434400632977486, + 0.04816211760044098, + 0.031809329986572266, + 0.032994162291288376, + 0.03444749861955643, + 0.029890412464737892, + 0.003972148522734642, + 0.006476573646068573, + 0.010113467462360859, + 0.00803206954151392, + 0.008714630268514156, + 0.004849645309150219, + 0.005984706338495016, + 0.004336763639003038, + 0.006705558858811855, + 0.003931121435016394, + 0.004201493691653013, + 0.004706330131739378, + 0.004204959608614445, + 0.005394116975367069, + 0.004682612139731646, + 0.004998926073312759, + 0.004687439184635878, + 0.003976883832365274, + 0.006345449946820736, + 0.005031402222812176, + 0.004019836895167828, + 0.004480467643588781, + 0.003908219281584024, + 0.005510891322046518, + 0.013062967918813229, + 0.004837770480662584, + 0.008553232066333294, + 0.006506304256618023, + 0.009732667356729507, + 0.03383888676762581, + 0.005149859003722668, + 0.02635973133146763, + 0.0031887798104435205, + 0.002245852956548333, + 0.006272594444453716, + 0.0028778451960533857, + 0.0021922835148870945, + 0.0024233581498265266, + 0.00252626184374094, + 0.005420610774308443, + 0.006515271030366421, + 0.004488798324018717, + 0.010429260320961475, + 0.004675935488194227, + 0.0047002993524074554, + 0.003812147304415703, + 0.004352440591901541, + 0.0077504441142082214, + 0.0021424831356853247, + 0.0023492879699915648, + 0.00452613877132535, + 0.002832879777997732, + 0.003445342183113098, + 0.0031311295460909605, + 0.0030193983111530542, + 0.004215338733047247, + 0.022518200799822807, + 0.008027316071093082, + 0.1274595409631729, + 0.01263359747827053, + 0.015572882257401943, + 0.023723365738987923, + 0.012381796725094318, + 0.09910555183887482, + 0.005150550976395607, + 0.004302217625081539, + 0.009066008031368256, + 0.006190459709614515, + 0.003995279315859079, + 0.006454253103584051, + 0.006819724105298519, + 0.010508798994123936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.005315575283020735 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002502645133063197 + ], + "opposite_logprob_tokens": [ + -6.00250244140625 + ], + "logprob_reward": -0.002502645133063197, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002502645133063197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2463224083185196, + 0.044569630175828934, + 0.04046104848384857, + 0.038664091378450394, + 0.06864584237337112, + 0.08568950742483139, + 0.10272824019193649, + 0.06778422743082047, + 0.2528754770755768, + 0.05225953459739685 + ], + "credit_map": [ + 0.020467180758714676, + 0.05349994823336601, + 0.02186826430261135, + 0.03581300377845764, + 0.03469303622841835, + 0.030286584049463272, + 0.025945516303181648, + 0.023748870939016342, + 0.0036173141561448574, + 0.004870356526225805, + 0.005613627377897501, + 0.004848610144108534, + 0.007120458874851465, + 0.004357258789241314, + 0.00841907225549221, + 0.005722926463931799, + 0.0082041434943676, + 0.00430397130548954, + 0.005660071037709713, + 0.00473747355863452, + 0.004030533134937286, + 0.0037149505224078894, + 0.0035848391707986593, + 0.006225063931196928, + 0.004500842187553644, + 0.0023996341042220592, + 0.00957776140421629, + 0.004122217185795307, + 0.003134612226858735, + 0.00442667817696929, + 0.003550152061507106, + 0.006952190771698952, + 0.0073625328950583935, + 0.004040271509438753, + 0.006553941406309605, + 0.005445889662951231, + 0.004631055984646082, + 0.015989063307642937, + 0.004144991748034954, + 0.02047809027135372, + 0.006562147289514542, + 0.005480367224663496, + 0.029222112149000168, + 0.006130126770585775, + 0.005548545625060797, + 0.006051032338291407, + 0.004167336039245129, + 0.02252783067524433, + 0.007300262805074453, + 0.00557341892272234, + 0.030369535088539124, + 0.008118020370602608, + 0.008799628354609013, + 0.00915494654327631, + 0.007400390692055225, + 0.026012029498815536, + 0.0061714015901088715, + 0.007805166766047478, + 0.009714019484817982, + 0.010499696247279644, + 0.009151090867817402, + 0.007788613438606262, + 0.008925335481762886, + 0.007728897966444492, + 0.021546710282564163, + 0.01212848350405693, + 0.08163092285394669, + 0.0159661415964365, + 0.01820535399019718, + 0.02479744143784046, + 0.018419895321130753, + 0.060180507600307465, + 0.005019727163016796, + 0.004366045817732811, + 0.010128786787390709, + 0.004229349549859762, + 0.004392941482365131, + 0.008596234954893589, + 0.005991209764033556, + 0.009535238146781921 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.002502645133063197 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009805607609450817 + ], + "opposite_logprob_tokens": [ + -4.634805679321289 + ], + "logprob_reward": -0.009805607609450817, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009805607609450817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15313702821731567, + 0.04101231321692467, + 0.05636119842529297, + 0.12399620562791824, + 0.04822041466832161, + 0.04081934317946434, + 0.10652242600917816, + 0.06398918479681015, + 0.2917742133140564, + 0.07416769862174988 + ], + "credit_map": [ + 0.012037991546094418, + 0.02082596905529499, + 0.021572619676589966, + 0.02800823003053665, + 0.015097499825060368, + 0.013939169235527515, + 0.018389903008937836, + 0.023265646770596504, + 0.0049339826218783855, + 0.005174465477466583, + 0.005764043889939785, + 0.0058680567890405655, + 0.004444369580596685, + 0.00476114172488451, + 0.005140585359185934, + 0.004925665911287069, + 0.007599735166877508, + 0.00506677757948637, + 0.005594353191554546, + 0.00687485933303833, + 0.005443857982754707, + 0.010230260901153088, + 0.006219667848199606, + 0.009331687353551388, + 0.021068867295980453, + 0.009904172271490097, + 0.017497308552265167, + 0.014464596286416054, + 0.010242274962365627, + 0.015584582462906837, + 0.010053492151200771, + 0.025180907920002937, + 0.004781220108270645, + 0.003401744645088911, + 0.011767702177166939, + 0.004889574833214283, + 0.0036657853052020073, + 0.007364163640886545, + 0.003437596606090665, + 0.008912624791264534, + 0.003628494217991829, + 0.005974949803203344, + 0.00724972365424037, + 0.004917733371257782, + 0.004782443400472403, + 0.0056348987855017185, + 0.0037190886214375496, + 0.0049120099283754826, + 0.007182571571320295, + 0.006829388905316591, + 0.03279399499297142, + 0.007006773259490728, + 0.009887240827083588, + 0.008916136808693409, + 0.0069887349382042885, + 0.026917582377791405, + 0.005144158843904734, + 0.005340249743312597, + 0.009314008057117462, + 0.008679652586579323, + 0.009123699739575386, + 0.009632780216634274, + 0.008208264596760273, + 0.008546368218958378, + 0.023506509140133858, + 0.013298539444804192, + 0.09964647889137268, + 0.01713089644908905, + 0.019282810389995575, + 0.026986712589859962, + 0.018684759736061096, + 0.07323750853538513, + 0.0056467060931026936, + 0.007033506408333778, + 0.01292840950191021, + 0.007146322634071112, + 0.007596283219754696, + 0.010140195488929749, + 0.008353164419531822, + 0.015323116444051266 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.009805607609450817 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036170552484691143 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.0036170552484691143, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036170552484691143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1842864453792572, + 0.150798961520195, + 0.03089303895831108, + 0.03775634989142418, + 0.11724518984556198, + 0.04185861349105835, + 0.04421716183423996, + 0.06909321248531342, + 0.2303672581911087, + 0.0934838131070137 + ], + "credit_map": [ + 0.016681363806128502, + 0.021808689460158348, + 0.02993147447705269, + 0.032420989125967026, + 0.018453139811754227, + 0.01772008091211319, + 0.02297106571495533, + 0.024299653246998787, + 0.019757714122533798, + 0.028591737151145935, + 0.017814120277762413, + 0.02866193652153015, + 0.012843552976846695, + 0.01224520057439804, + 0.01598018780350685, + 0.014904512092471123, + 0.003758050501346588, + 0.002511534607037902, + 0.0035921495873481035, + 0.004090765491127968, + 0.003525426611304283, + 0.0045876274816691875, + 0.003371038008481264, + 0.00545644573867321, + 0.008650501258671284, + 0.0025976011529564857, + 0.005946496967226267, + 0.0031174826435744762, + 0.0033941264264285564, + 0.0033585121855139732, + 0.0030327585991472006, + 0.007658868562430143, + 0.009845429100096226, + 0.004120219964534044, + 0.01075916551053524, + 0.00653431098908186, + 0.008676809258759022, + 0.03451353311538696, + 0.005604300182312727, + 0.037191424518823624, + 0.003240969032049179, + 0.004573550075292587, + 0.008137187920510769, + 0.005139612127095461, + 0.004247387871146202, + 0.006514879409223795, + 0.0036563409958034754, + 0.006348687689751387, + 0.0033218965400010347, + 0.004328833427280188, + 0.00905465055257082, + 0.00438085338100791, + 0.007921067997813225, + 0.003713238751515746, + 0.0043283249251544476, + 0.007168293930590153, + 0.006325643043965101, + 0.004287424962967634, + 0.010760700330138206, + 0.009438708424568176, + 0.00832283589988947, + 0.012186906300485134, + 0.008222384378314018, + 0.009548608213663101, + 0.018289776518940926, + 0.010649275034666061, + 0.075892873108387, + 0.013792733661830425, + 0.016580991446971893, + 0.02450311928987503, + 0.016144830733537674, + 0.05451365187764168, + 0.011267982423305511, + 0.007351589854806662, + 0.021019823849201202, + 0.007380248978734016, + 0.007970902137458324, + 0.014486344531178474, + 0.007939621806144714, + 0.01606729067862034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0036170552484691143 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20158924162387848 + ], + "opposite_logprob_tokens": [ + -1.7015892267227173 + ], + "logprob_reward": -0.20158924162387848, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20158924162387848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.141204372048378, + 0.08268728852272034, + 0.04993661865592003, + 0.05615237355232239, + 0.040563903748989105, + 0.03492619842290878, + 0.13550017774105072, + 0.16366437077522278, + 0.17523856461048126, + 0.12012617290019989 + ], + "credit_map": [ + 0.009564145468175411, + 0.01946994662284851, + 0.018277974799275398, + 0.02850906178355217, + 0.020884687080979347, + 0.013820527121424675, + 0.0158124677836895, + 0.014865553006529808, + 0.008595962077379227, + 0.012280984781682491, + 0.008864344097673893, + 0.008511362597346306, + 0.012328656390309334, + 0.01031892653554678, + 0.01165169570595026, + 0.010135351680219173, + 0.00650169188156724, + 0.005101238377392292, + 0.008985745720565319, + 0.006298622582107782, + 0.003935899585485458, + 0.0062659974209964275, + 0.004262800794094801, + 0.008584625087678432, + 0.00972413457930088, + 0.0035166891757398844, + 0.015991803258657455, + 0.005233211908489466, + 0.0035415873862802982, + 0.004648820497095585, + 0.005132074002176523, + 0.008364051580429077, + 0.0048673115670681, + 0.003621339099481702, + 0.009097239933907986, + 0.005227586254477501, + 0.0029287459328770638, + 0.00485036289319396, + 0.0030832146294414997, + 0.006888100411742926, + 0.0033444829750806093, + 0.004058003891259432, + 0.006214065011590719, + 0.0054283756762743, + 0.004144008737057447, + 0.0038075086195021868, + 0.003242975100874901, + 0.0046867793425917625, + 0.013635317794978619, + 0.011376092210412025, + 0.02978249080479145, + 0.014865068718791008, + 0.014367788098752499, + 0.013696657493710518, + 0.01357622817158699, + 0.024200530722737312, + 0.01926102675497532, + 0.010983607731759548, + 0.030294543132185936, + 0.013161585666239262, + 0.023352567106485367, + 0.028612565249204636, + 0.015273821540176868, + 0.02272464707493782, + 0.013687285594642162, + 0.006096495781093836, + 0.06565666198730469, + 0.009301016107201576, + 0.009985605254769325, + 0.014249113388359547, + 0.009063556790351868, + 0.047198835760354996, + 0.010905400849878788, + 0.008532198145985603, + 0.01863884925842285, + 0.011853300966322422, + 0.009333689697086811, + 0.022288748994469643, + 0.012229938060045242, + 0.026344047859311104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.20158924162387848 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a stunning macro close-up shot, the camera tilts down at an almost imperceptible speed, focusing on an exquisite antique ring as it begins to slide down a sheet of deep sapphire silk, stretched across its gently folds. Crafted from gold with a warm, aged patina, the ring's band is engraved with intricate, delicate hand-engraved filigree patterns, revealing the marks of meticulous craftsmanship etched over time. At the ring's apex, a large, lustrous cream pearl rests securely in an elegant prong setting, its surface reflecting a soft, diffuse halo of light. With the gentle pull of gravity, the ring begins its graceful, unstoppable descent, skimming almost frictionlessly across the silk's smooth surface, sliding steadily and deliberately down one of its gentle slopes. A single, warm light source, emanating from the side, casts long, soft shadows across the scene, converging into a bright highlight on the pearl's curved surface and outlining the delicate contours of the gold filigree. The whole picture is immersed in a quiet, slightly melancholic and elegant atmosphere, full of the texture of time and silent stories.", + "seed": 17036, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_10/index_1036_seed_17036_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03817796451039612, + "vlm_logprob": -0.03817796451039612, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.03817796451039612, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005687957163900137 + ], + "opposite_logprob_tokens": [ + -6.005688190460205 + ], + "logprob_reward": -0.005687957163900137, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.005687957163900137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15289996564388275, + 0.1409599632024765, + 0.060163550078868866, + 0.09281434118747711, + 0.09436050802469254, + 0.050872065126895905, + 0.21104787290096283, + 0.05294930562376976, + 0.06324290484189987, + 0.08068951219320297 + ], + "credit_map": [ + 0.02234577015042305, + 0.019242720678448677, + 0.016904814168810844, + 0.012722494080662727, + 0.023859448730945587, + 0.02606988698244095, + 0.012134005315601826, + 0.01962081529200077, + 0.025273730978369713, + 0.015770915895700455, + 0.018168829381465912, + 0.011110209859907627, + 0.025940019637346268, + 0.014319958165287971, + 0.010850219987332821, + 0.01952606812119484, + 0.011287057772278786, + 0.008106478489935398, + 0.005096723325550556, + 0.007954715751111507, + 0.009650320746004581, + 0.005886595696210861, + 0.003470393130555749, + 0.008711263537406921, + 0.006879551336169243, + 0.014165621250867844, + 0.007577728014439344, + 0.011706368997693062, + 0.019058113917708397, + 0.008895736187696457, + 0.011390055529773235, + 0.013141165487468243, + 0.008767043240368366, + 0.009964187629520893, + 0.010461065918207169, + 0.011638236232101917, + 0.02135804109275341, + 0.007673922926187515, + 0.010834407061338425, + 0.01366360392421484, + 0.007364272605627775, + 0.005521118175238371, + 0.005936859175562859, + 0.0072335717268288136, + 0.006134240422397852, + 0.004336775280535221, + 0.006521826144307852, + 0.007823402993381023, + 0.025041375309228897, + 0.02653397060930729, + 0.018989527598023415, + 0.037714872509241104, + 0.024103563278913498, + 0.02522757649421692, + 0.021888887509703636, + 0.031548112630844116, + 0.00482577458024025, + 0.005083415657281876, + 0.006213181186467409, + 0.006973279174417257, + 0.005742795765399933, + 0.006174616049975157, + 0.0074401237070560455, + 0.01049611996859312, + 0.005053784232586622, + 0.005401390139013529, + 0.0042646038345992565, + 0.009613582864403725, + 0.009068994782865047, + 0.004721521865576506, + 0.014674180187284946, + 0.010444846004247665, + 0.007586831692606211, + 0.0071376715786755085, + 0.009799830615520477, + 0.01622491516172886, + 0.011203259229660034, + 0.005774575751274824, + 0.008770584128797054, + 0.014191846363246441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.005687957163900137 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025000290479511023 + ], + "opposite_logprob_tokens": [ + -6.002500057220459 + ], + "logprob_reward": -0.0025000290479511023, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025000290479511023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10861095041036606, + 0.08079706132411957, + 0.04498874023556709, + 0.05930806323885918, + 0.05084129422903061, + 0.09565035998821259, + 0.17696352303028107, + 0.16299034655094147, + 0.10222376137971878, + 0.11762596666812897 + ], + "credit_map": [ + 0.017255587503314018, + 0.014444337226450443, + 0.007946786470711231, + 0.017258666455745697, + 0.01660873554646969, + 0.012424230575561523, + 0.008605213835835457, + 0.014067392796278, + 0.011039064265787601, + 0.008528931997716427, + 0.011105424724519253, + 0.008058859966695309, + 0.01606062240898609, + 0.008643059991300106, + 0.006372039206326008, + 0.010989063419401646, + 0.009640347212553024, + 0.005259404890239239, + 0.004944412037730217, + 0.003931222017854452, + 0.006225908175110817, + 0.0055361054837703705, + 0.003541166428476572, + 0.00591017073020339, + 0.006332674995064735, + 0.009220144711434841, + 0.0039044071454554796, + 0.007149917539209127, + 0.012665963731706142, + 0.004493766464293003, + 0.00667948042973876, + 0.008861702866852283, + 0.00985707901418209, + 0.006170045118778944, + 0.006377164740115404, + 0.005411782301962376, + 0.006197674199938774, + 0.00527493329718709, + 0.004703929647803307, + 0.006848685909062624, + 0.014812692068517208, + 0.010768054984509945, + 0.010026931762695312, + 0.012515382841229439, + 0.011676578782498837, + 0.007505347020924091, + 0.01251422893255949, + 0.01583113707602024, + 0.019041333347558975, + 0.020508799701929092, + 0.01656777784228325, + 0.02984706684947014, + 0.020054906606674194, + 0.024281127378344536, + 0.02023058757185936, + 0.026431914418935776, + 0.021492570638656616, + 0.012512188404798508, + 0.017482172697782516, + 0.017248500138521194, + 0.01657862402498722, + 0.024420449510216713, + 0.022682858631014824, + 0.03057296946644783, + 0.009808406233787537, + 0.008060311898589134, + 0.009153364226222038, + 0.012857360765337944, + 0.011614994145929813, + 0.0083618788048625, + 0.02665700390934944, + 0.01571044512093067, + 0.011958491057157516, + 0.010753082111477852, + 0.014978129416704178, + 0.022404983639717102, + 0.018242405727505684, + 0.009547977708280087, + 0.012176130898296833, + 0.01756477728486061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.0025000290479511023 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012560428120195866 + ], + "opposite_logprob_tokens": [ + -4.3875603675842285 + ], + "logprob_reward": -0.012560428120195866, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012560428120195866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08791430294513702, + 0.19713324308395386, + 0.10034188628196716, + 0.10515817999839783, + 0.07878922671079636, + 0.05304954573512077, + 0.06725430488586426, + 0.12580998241901398, + 0.08404061943292618, + 0.10050876438617706 + ], + "credit_map": [ + 0.010673238895833492, + 0.009554753080010414, + 0.007843178696930408, + 0.008264117874205112, + 0.017774760723114014, + 0.01429940015077591, + 0.007416984997689724, + 0.012087859213352203, + 0.040386274456977844, + 0.022610628977417946, + 0.01887555979192257, + 0.01896815188229084, + 0.036098506301641464, + 0.01876344531774521, + 0.014859011396765709, + 0.026571646332740784, + 0.013099193572998047, + 0.016078347340226173, + 0.009943993762135506, + 0.01066549587994814, + 0.01875835470855236, + 0.010753784328699112, + 0.0072853947058320045, + 0.013757310807704926, + 0.008409019559621811, + 0.016072317957878113, + 0.007734707556664944, + 0.013742905110120773, + 0.020892338827252388, + 0.00975368544459343, + 0.012759174220263958, + 0.015794022008776665, + 0.010271595790982246, + 0.0077278814278542995, + 0.011890873312950134, + 0.007551339454948902, + 0.009373591281473637, + 0.013655818067491055, + 0.007227771449834108, + 0.011090350337326527, + 0.007354883011430502, + 0.0060951909981667995, + 0.005274645052850246, + 0.007979203015565872, + 0.007438731379806995, + 0.005081718787550926, + 0.006237811874598265, + 0.007587358355522156, + 0.00888456217944622, + 0.0068761021830141544, + 0.008252300322055817, + 0.011783129535615444, + 0.007249524816870689, + 0.006893352139741182, + 0.006878191605210304, + 0.010437141172587872, + 0.01804894208908081, + 0.008779209107160568, + 0.009038279764354229, + 0.0108192078769207, + 0.00977491308003664, + 0.009748117998242378, + 0.014326934702694416, + 0.04527436941862106, + 0.0069542741402983665, + 0.005873506423085928, + 0.006545977666974068, + 0.011274313554167747, + 0.009046689607203007, + 0.005449165124446154, + 0.024584747850894928, + 0.014311936683952808, + 0.01128322072327137, + 0.009413330815732479, + 0.010119155049324036, + 0.020000800490379333, + 0.014602055773139, + 0.007254905067384243, + 0.011866210959851742, + 0.01596907526254654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 3 + ], + "selected_reward": -0.012560428120195866 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783769026398659 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783769026398659, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783769026398659, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2512342035770416, + 0.12993638217449188, + 0.05077004432678223, + 0.03636600449681282, + 0.10805562138557434, + 0.05364763364195824, + 0.0644790530204773, + 0.1223372295498848, + 0.0740099549293518, + 0.10916391015052795 + ], + "credit_map": [ + 0.037408944219350815, + 0.03262046352028847, + 0.027194995433092117, + 0.031365372240543365, + 0.030435340479016304, + 0.03850388526916504, + 0.017979780212044716, + 0.0357254333794117, + 0.028201874345541, + 0.01365931797772646, + 0.0133983064442873, + 0.012403740547597408, + 0.02219419740140438, + 0.013103283010423183, + 0.009865391068160534, + 0.017110276967287064, + 0.007106635253876448, + 0.007589528802782297, + 0.005848952569067478, + 0.006808712612837553, + 0.009359349496662617, + 0.005236943252384663, + 0.003400985151529312, + 0.005418932065367699, + 0.003376067616045475, + 0.004662986379116774, + 0.0035149019677191973, + 0.0049513475969433784, + 0.005679076071828604, + 0.004536333493888378, + 0.004071523901075125, + 0.005573769565671682, + 0.02459346503019333, + 0.010629500262439251, + 0.015218956395983696, + 0.010592259466648102, + 0.009488136507570744, + 0.009352165274322033, + 0.010039307177066803, + 0.018141835927963257, + 0.006144464015960693, + 0.007718435022979975, + 0.00598907982930541, + 0.007154884748160839, + 0.007511470932513475, + 0.005258800927549601, + 0.005796711426228285, + 0.008073789067566395, + 0.006140836980193853, + 0.008518313057720661, + 0.005455044563859701, + 0.012781677767634392, + 0.009708237834274769, + 0.00576513959094882, + 0.0059100608341395855, + 0.010199748910963535, + 0.015539036132395267, + 0.008939165621995926, + 0.008945398963987827, + 0.012156515382230282, + 0.010261845774948597, + 0.009124742820858955, + 0.015074390918016434, + 0.04229613393545151, + 0.006207800470292568, + 0.006889214739203453, + 0.005310302600264549, + 0.013041269034147263, + 0.009807570837438107, + 0.005187669303268194, + 0.01419964898377657, + 0.013366473838686943, + 0.012614944018423557, + 0.009181322529911995, + 0.009494252502918243, + 0.021846598014235497, + 0.015968654304742813, + 0.009351776912808418, + 0.012350006960332394, + 0.018356354907155037 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.009783769026398659 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up focusing on an antique gold ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ring sliding down a sheet of deep sapphire silk with gentle folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the ring feature a large cream pearl in a prong setting with intricate hand-engraved filigree patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a single warm light source from the side casting long soft shadows and a bright highlight on the pearl?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera tilt down at an almost imperceptible speed while the ring slides down the silk?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035763919353485 + ], + "opposite_logprob_tokens": [ + -1.9103575944900513 + ], + "logprob_reward": -0.16035763919353485, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035763919353485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15636122226715088, + 0.056821923702955246, + 0.044190794229507446, + 0.1314207911491394, + 0.09423357993364334, + 0.06141771003603935, + 0.1758739948272705, + 0.07077449560165405, + 0.09891600906848907, + 0.10998952388763428 + ], + "credit_map": [ + 0.025659073144197464, + 0.017374485731124878, + 0.01812594197690487, + 0.01648494228720665, + 0.020833810791373253, + 0.021316125988960266, + 0.01211500633507967, + 0.024451829493045807, + 0.006321430206298828, + 0.006148204673081636, + 0.005773395299911499, + 0.0062048048712313175, + 0.01314101368188858, + 0.006782702635973692, + 0.0039684586226940155, + 0.00848191138356924, + 0.005172032862901688, + 0.004434859845787287, + 0.006205802783370018, + 0.006568173412233591, + 0.007099221460521221, + 0.005662947427481413, + 0.0031566971447318792, + 0.005891057662665844, + 0.015057326294481754, + 0.017987437546253204, + 0.008668357506394386, + 0.01762671209871769, + 0.027209747582674026, + 0.010679054073989391, + 0.01552511565387249, + 0.018667031079530716, + 0.010719009675085545, + 0.00938711129128933, + 0.0088889691978693, + 0.01348771620541811, + 0.01750444620847702, + 0.011551677249372005, + 0.009059017524123192, + 0.013635626994073391, + 0.006351293995976448, + 0.011038917116820812, + 0.0063819740898907185, + 0.007339856121689081, + 0.010652071796357632, + 0.006244752090424299, + 0.0050187804736196995, + 0.008390058763325214, + 0.025690363720059395, + 0.028139125555753708, + 0.015631580725312233, + 0.025642504915595055, + 0.01870127022266388, + 0.020385228097438812, + 0.016574159264564514, + 0.025109754875302315, + 0.00930763315409422, + 0.005566461943089962, + 0.00930110178887844, + 0.010531713254749775, + 0.009135843254625797, + 0.0056367251090705395, + 0.007509083952754736, + 0.01378593035042286, + 0.009327298030257225, + 0.007258822210133076, + 0.010488300584256649, + 0.012078528292477131, + 0.010906873270869255, + 0.018041390925645828, + 0.01479626540094614, + 0.016018521040678024, + 0.010778779163956642, + 0.007964767515659332, + 0.011585120111703873, + 0.015707779675722122, + 0.016791364178061485, + 0.011606375686824322, + 0.013488519005477428, + 0.022066812962293625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.16035763919353485 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_11/index_tensor([310, 310])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_11/index_tensor([310, 310])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a1bd647f0c116e93881ad0813b5712e10a6e1325 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_11/index_tensor([310, 310])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static shot, taken from a medium close-up and slightly from an upward angle, captures a man in extreme exhaustion. His messy, short, dark hair, slightly damp with sweat, clings to his pale forehead. His sunken eyes are etched with dark circles, and every fine line on his face bears the weight of mental strain. He stands before a rough concrete wall. A single, harsh, cold-blue light source, emanating from the side and below, sharply cuts his outline and casts a vast, dramatically elongating shadow behind him. As he slowly, almost in submission, closes his heavy eyes, his facial muscles instantly relax as all support is released. The once clear shadow behind him wriggles and deforms, stretching infinitely upwards, seemingly consuming the space. The entire image is shrouded in a cold, blue hue. The light creates sharp highlights on the man's high nose and cheekbones, contrasting sharply with the deep shadows on his cheeks and neck, creating an atmosphere of profound loneliness and mental exhaustion, imbued with a cinematic sense of oppression.", + "seed": 18788, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_11/index_310_seed_18788_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08428900875151157, + "vlm_logprob": -0.08428900875151157, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.08428900875151157, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022376900538802147 + ], + "opposite_logprob_tokens": [ + -4.397377014160156 + ], + "logprob_reward": -0.022376900538802147, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.022376900538802147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17770503461360931, + 0.10907108336687088, + 0.04800019785761833, + 0.03436823934316635, + 0.03803236410021782, + 0.0300563033670187, + 0.0386916883289814, + 0.18186068534851074, + 0.2836359739303589, + 0.058578457683324814 + ], + "credit_map": [ + 0.022304408252239227, + 0.017851343378424644, + 0.021520676091313362, + 0.0162375345826149, + 0.016937024891376495, + 0.02464553713798523, + 0.021766209974884987, + 0.03644230216741562, + 0.017298560589551926, + 0.013624940998852253, + 0.015194329433143139, + 0.01179527584463358, + 0.011960522271692753, + 0.010528375394642353, + 0.01477076020091772, + 0.013898330740630627, + 0.005032201763242483, + 0.005711728241294622, + 0.008115417324006557, + 0.006580155808478594, + 0.004916222766041756, + 0.0059378561563789845, + 0.006297747604548931, + 0.005408868659287691, + 0.0036178925074636936, + 0.004066881723701954, + 0.0037075597792863846, + 0.003398740431293845, + 0.0036306031979620457, + 0.005159273743629456, + 0.002917358186095953, + 0.007869929075241089, + 0.006013788282871246, + 0.004672383889555931, + 0.005337430629879236, + 0.005275330040603876, + 0.004729009233415127, + 0.0038540870882570744, + 0.0034196407068520784, + 0.004730693995952606, + 0.00404007826000452, + 0.004866155330091715, + 0.0033349534496665, + 0.0038080753292888403, + 0.0029695588164031506, + 0.003306078491732478, + 0.0029834879096597433, + 0.004747916478663683, + 0.004244134295731783, + 0.0044902111403644085, + 0.006772309076040983, + 0.005103878676891327, + 0.005558316130191088, + 0.005001996643841267, + 0.002899727551266551, + 0.0046211127191782, + 0.02004542015492916, + 0.03442395105957985, + 0.029984449967741966, + 0.02335178479552269, + 0.023282315582036972, + 0.01984715834259987, + 0.01503437664359808, + 0.015891235321760178, + 0.030689764767885208, + 0.02722351811826229, + 0.02606532908976078, + 0.02744327485561371, + 0.03991291671991348, + 0.04043371602892876, + 0.02600615844130516, + 0.0658613070845604, + 0.011750943027436733, + 0.006582343485206366, + 0.007424134761095047, + 0.007413654122501612, + 0.00627537677064538, + 0.005803151987493038, + 0.005786148831248283, + 0.00754270376637578 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.022376900538802147 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029828868806362152 + ], + "opposite_logprob_tokens": [ + -3.5298287868499756 + ], + "logprob_reward": -0.029828868806362152, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029828868806362152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2886824905872345, + 0.05190247669816017, + 0.023781297728419304, + 0.05495252460241318, + 0.0172524806112051, + 0.034253865480422974, + 0.06264860183000565, + 0.369615375995636, + 0.07017799466848373, + 0.026732970029115677 + ], + "credit_map": [ + 0.03595610707998276, + 0.031246475875377655, + 0.03434767201542854, + 0.02695104107260704, + 0.028292465955018997, + 0.042318396270275116, + 0.03114517219364643, + 0.05842513218522072, + 0.008293217979371548, + 0.005747360642999411, + 0.0077192592434585094, + 0.005663132760673761, + 0.005080596078187227, + 0.005374305881559849, + 0.005891006905585527, + 0.0081336023285985, + 0.0029525351710617542, + 0.003134885337203741, + 0.0035351074766367674, + 0.0028615030460059643, + 0.0021104749757796526, + 0.0028176642954349518, + 0.002662772312760353, + 0.0037063551135361195, + 0.005755185149610043, + 0.007303295191377401, + 0.008245597593486309, + 0.007494248915463686, + 0.005194474011659622, + 0.0073873731307685375, + 0.004881209693849087, + 0.008691138587892056, + 0.002905396744608879, + 0.0021242410875856876, + 0.002258328953757882, + 0.0023085344582796097, + 0.0016376065323129296, + 0.001827880972996354, + 0.001820604084059596, + 0.002369887428358197, + 0.0032633927185088396, + 0.005672022234648466, + 0.0041941325180232525, + 0.005320833995938301, + 0.0032407690305262804, + 0.0037119348999112844, + 0.003254030365496874, + 0.00559674808755517, + 0.008191962726414204, + 0.007384601514786482, + 0.013472398743033409, + 0.007659814320504665, + 0.005914219189435244, + 0.0070379567332565784, + 0.0057199872098863125, + 0.007267665583640337, + 0.04355445131659508, + 0.06357154995203018, + 0.06546679139137268, + 0.05072304606437683, + 0.03802401199936867, + 0.04005800187587738, + 0.03106151707470417, + 0.03715598210692406, + 0.006683212239295244, + 0.010595137253403664, + 0.006653128657490015, + 0.008467159233987331, + 0.00967478659003973, + 0.008660625666379929, + 0.0064886282198131084, + 0.012955320067703724, + 0.0034791177604347467, + 0.003948615398257971, + 0.002711312146857381, + 0.0028458284214138985, + 0.004515401553362608, + 0.0023810439743101597, + 0.0039349570870399475, + 0.002916694385930896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.029828868806362152 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809856250882149 + ], + "opposite_logprob_tokens": [ + -3.2880985736846924 + ], + "logprob_reward": -0.03809856250882149, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809856250882149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42065438628196716, + 0.07421513646841049, + 0.05859914422035217, + 0.043295685201883316, + 0.027306990697979927, + 0.07685612887144089, + 0.03284164145588875, + 0.10779872536659241, + 0.130796417593956, + 0.02763582579791546 + ], + "credit_map": [ + 0.059121273458004, + 0.028759567067027092, + 0.03387577459216118, + 0.032553691416978836, + 0.031602077186107635, + 0.07463665306568146, + 0.034035824239254, + 0.12606950104236603, + 0.011335046961903572, + 0.00787811167538166, + 0.010649521835148335, + 0.008724856190383434, + 0.005334039684385061, + 0.009125445038080215, + 0.009555734694004059, + 0.011612379923462868, + 0.0068539888598024845, + 0.0068070245906710625, + 0.0076342616230249405, + 0.00605385098606348, + 0.005770555697381496, + 0.008738134056329727, + 0.005074675194919109, + 0.011666649952530861, + 0.005638544913381338, + 0.005536830984055996, + 0.00535969203338027, + 0.004273758269846439, + 0.0035356960725039244, + 0.0057161590084433556, + 0.0036925957538187504, + 0.009542407467961311, + 0.005024257116019726, + 0.0036557859275490046, + 0.003539866069331765, + 0.0033384424168616533, + 0.0030301178339868784, + 0.0028548717964440584, + 0.0023418550845235586, + 0.00352179491892457, + 0.009506774134933949, + 0.010467742569744587, + 0.008150394074618816, + 0.009096338413655758, + 0.007638401817530394, + 0.010468916967511177, + 0.007194035686552525, + 0.014333529397845268, + 0.0047752745449543, + 0.003309940453618765, + 0.0035941158421337605, + 0.0026052333414554596, + 0.002669607987627387, + 0.005422245245426893, + 0.002834073267877102, + 0.0076311491429805756, + 0.013294778764247894, + 0.018197419121861458, + 0.015551663003861904, + 0.013028275221586227, + 0.013060090132057667, + 0.012321165762841702, + 0.007398595102131367, + 0.014946751296520233, + 0.013857555575668812, + 0.016668671742081642, + 0.011352143250405788, + 0.01509028673171997, + 0.01817546784877777, + 0.017007213085889816, + 0.011637825518846512, + 0.02700725756585598, + 0.00491179246455431, + 0.002878002356737852, + 0.004071274306625128, + 0.003151901764795184, + 0.00332418130710721, + 0.002774960594251752, + 0.002812531776726246, + 0.003711181692779064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.03809856250882149 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899792492389679 + ], + "opposite_logprob_tokens": [ + -2.578997850418091 + ], + "logprob_reward": -0.07899792492389679, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899792492389679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21022282540798187, + 0.07302668690681458, + 0.0837993249297142, + 0.08768148720264435, + 0.03418894112110138, + 0.26054394245147705, + 0.04360552504658699, + 0.054714400321245193, + 0.13714556396007538, + 0.015071279369294643 + ], + "credit_map": [ + 0.029782453551888466, + 0.01771215721964836, + 0.01893564499914646, + 0.01752626895904541, + 0.015050973743200302, + 0.03716258332133293, + 0.01468635257333517, + 0.0593663826584816, + 0.013291065581142902, + 0.0070730471052229404, + 0.01171350758522749, + 0.00856672041118145, + 0.0069917780347168446, + 0.007867047563195229, + 0.007858088240027428, + 0.009665432386100292, + 0.009663833305239677, + 0.008491883985698223, + 0.012808473780751228, + 0.008498001843690872, + 0.0047357697039842606, + 0.013198211789131165, + 0.005708072800189257, + 0.02069508284330368, + 0.010898253880441189, + 0.008582521229982376, + 0.009647143073379993, + 0.01032257080078125, + 0.006059596315026283, + 0.013062132522463799, + 0.007630170788615942, + 0.021479101851582527, + 0.00605251919478178, + 0.004236322361975908, + 0.004887106362730265, + 0.0040900204330682755, + 0.004013373050838709, + 0.003614156972616911, + 0.003968832548707724, + 0.003326611826196313, + 0.023085806518793106, + 0.04716019704937935, + 0.03803135082125664, + 0.045711107552051544, + 0.030891159549355507, + 0.02317681722342968, + 0.02210366725921631, + 0.030383827164769173, + 0.00634814752265811, + 0.0050355540588498116, + 0.00561939412727952, + 0.0038871595170348883, + 0.004715751390904188, + 0.006051512900739908, + 0.004266258329153061, + 0.007681748829782009, + 0.0066987574100494385, + 0.008288320153951645, + 0.007622038945555687, + 0.008437429554760456, + 0.004681712482124567, + 0.006198874209076166, + 0.003985248506069183, + 0.008802019990980625, + 0.015742667019367218, + 0.017344241961836815, + 0.011444559320807457, + 0.016354437917470932, + 0.01608777977526188, + 0.019139042124152184, + 0.011617348529398441, + 0.029415488243103027, + 0.0019956580363214016, + 0.0019368203356862068, + 0.0019351162482053041, + 0.001683748560026288, + 0.001479104976169765, + 0.002008314710110426, + 0.0015257431659847498, + 0.0025067736860364676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.07899792492389679 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521427869796753 + ], + "opposite_logprob_tokens": [ + -1.5021427869796753 + ], + "logprob_reward": -0.2521427869796753, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521427869796753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25117257237434387, + 0.13124173879623413, + 0.07653743773698807, + 0.07630525529384613, + 0.0632530152797699, + 0.06578393280506134, + 0.04244383051991463, + 0.12964509427547455, + 0.1227920800447464, + 0.04082507640123367 + ], + "credit_map": [ + 0.03134988248348236, + 0.02127731405198574, + 0.024227360263466835, + 0.02307237684726715, + 0.0213822890073061, + 0.041759736835956573, + 0.017999207600951195, + 0.07010439038276672, + 0.021991988644003868, + 0.01557333767414093, + 0.02002457156777382, + 0.016727037727832794, + 0.012280462309718132, + 0.015327821485698223, + 0.014710838906466961, + 0.014605669304728508, + 0.008282188326120377, + 0.008497407659888268, + 0.010981042869389057, + 0.006922028958797455, + 0.007472914643585682, + 0.010885794647037983, + 0.0070334747433662415, + 0.016462581232190132, + 0.0103373434394598, + 0.008184555917978287, + 0.010591822676360607, + 0.00678458996117115, + 0.006539136171340942, + 0.009340205229818821, + 0.008814229629933834, + 0.015713360160589218, + 0.01053659152239561, + 0.00923334900289774, + 0.009838545694947243, + 0.008597969077527523, + 0.007554410025477409, + 0.005496652331203222, + 0.006135560106486082, + 0.00585993193089962, + 0.005750087555497885, + 0.014231660403311253, + 0.007721432484686375, + 0.011089194566011429, + 0.00716017372906208, + 0.006003838963806629, + 0.006692016031593084, + 0.007135526277124882, + 0.004857776686549187, + 0.005114935804158449, + 0.005870037712156773, + 0.00453692302107811, + 0.0033731895964592695, + 0.0070619103498756886, + 0.003155820071697235, + 0.008473233319818974, + 0.012302929535508156, + 0.016869990155100822, + 0.014404639601707458, + 0.015119445510208607, + 0.014257071539759636, + 0.017720455303788185, + 0.013561739586293697, + 0.025408808141946793, + 0.014585270546376705, + 0.015901867300271988, + 0.013063939288258553, + 0.015099179930984974, + 0.01504378579556942, + 0.015445885248482227, + 0.012036443687975407, + 0.021615691483020782, + 0.004979510325938463, + 0.005583826918154955, + 0.0057688429951667786, + 0.004692367743700743, + 0.006311590783298016, + 0.0037690133322030306, + 0.005121462047100067, + 0.004598459228873253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.2521427869796753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, taken from a medium close-up and slightly from an upward angle, captures a man in extreme exhaustion. His messy, short, dark hair, slightly damp with sweat, clings to his pale forehead. His sunken eyes are etched with dark circles, and every fine line on his face bears the weight of mental strain. He stands before a rough concrete wall. A single, harsh, cold-blue light source, emanating from the side and below, sharply cuts his outline and casts a vast, dramatically elongating shadow behind him. As he slowly, almost in submission, closes his heavy eyes, his facial muscles instantly relax as all support is released. The once clear shadow behind him wriggles and deforms, stretching infinitely upwards, seemingly consuming the space. The entire image is shrouded in a cold, blue hue. The light creates sharp highlights on the man's high nose and cheekbones, contrasting sharply with the deep shadows on his cheeks and neck, creating an atmosphere of profound loneliness and mental exhaustion, imbued with a cinematic sense of oppression.", + "seed": 18789, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_11/index_310_seed_18789_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03446500021964312, + "vlm_logprob": -0.03446500021964312, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.03446500021964312, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0265978891402483 + ], + "opposite_logprob_tokens": [ + -5.15159797668457 + ], + "logprob_reward": -0.0265978891402483, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0265978891402483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08083291351795197, + 0.11798232048749924, + 0.05277404934167862, + 0.09462404996156693, + 0.266936331987381, + 0.07107847183942795, + 0.051931023597717285, + 0.0476195402443409, + 0.1380634903907776, + 0.07815790921449661 + ], + "credit_map": [ + 0.01528365071862936, + 0.009628764353692532, + 0.013934547081589699, + 0.006076637655496597, + 0.00823257677257061, + 0.012343306094408035, + 0.007340248674154282, + 0.00799318216741085, + 0.027848757803440094, + 0.00939017441123724, + 0.013260402716696262, + 0.00673547200858593, + 0.007727200631052256, + 0.03444516286253929, + 0.008232942782342434, + 0.010342208668589592, + 0.0066823698580265045, + 0.007337468210607767, + 0.008433816954493523, + 0.00879093911498785, + 0.006145845633000135, + 0.006000868044793606, + 0.0047396207228302956, + 0.0046431212686002254, + 0.013223077170550823, + 0.01569962315261364, + 0.014928885735571384, + 0.012636574916541576, + 0.009502701461315155, + 0.011122439056634903, + 0.009403537958860397, + 0.008107214234769344, + 0.0693279504776001, + 0.023996887728571892, + 0.018930984660983086, + 0.021387426182627678, + 0.017021125182509422, + 0.08549891412258148, + 0.012259949930012226, + 0.01851307787001133, + 0.009353186003863811, + 0.007554603274911642, + 0.00792978797107935, + 0.010377664119005203, + 0.007451195735484362, + 0.008835501037538052, + 0.010662245564162731, + 0.008914288133382797, + 0.007366114296019077, + 0.004615722689777613, + 0.008004696108400822, + 0.005432546604424715, + 0.006333874072879553, + 0.008471343666315079, + 0.0042102038860321045, + 0.007496522273868322, + 0.007360753137618303, + 0.0052241855300962925, + 0.0077133383601903915, + 0.007742315996438265, + 0.003933376166969538, + 0.005463432986289263, + 0.005589729640632868, + 0.004592408426105976, + 0.013897750526666641, + 0.009625517763197422, + 0.03165490925312042, + 0.009636923670768738, + 0.011882965452969074, + 0.022798212245106697, + 0.010045022703707218, + 0.028522195294499397, + 0.0069074323400855064, + 0.01097303256392479, + 0.008981778286397457, + 0.013903290033340454, + 0.009121705777943134, + 0.009638151153922081, + 0.00650052260607481, + 0.012131997384130955 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 1 + ], + "selected_reward": -0.0265978891402483 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255913358181715 + ], + "opposite_logprob_tokens": [ + -4.387558937072754 + ], + "logprob_reward": -0.01255913358181715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255913358181715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10084767639636993, + 0.07143272459506989, + 0.06830315291881561, + 0.02874354086816311, + 0.06069091707468033, + 0.2406836599111557, + 0.10744903236627579, + 0.07277244329452515, + 0.143161803483963, + 0.10591507703065872 + ], + "credit_map": [ + 0.019111670553684235, + 0.01239961851388216, + 0.012430012226104736, + 0.009532098658382893, + 0.010639743879437447, + 0.017065025866031647, + 0.009783297777175903, + 0.009886214509606361, + 0.008848657831549644, + 0.009754790924489498, + 0.00966095644980669, + 0.009577074088156223, + 0.007638250477612019, + 0.010485232807695866, + 0.007909787818789482, + 0.0075579737313091755, + 0.007275443058460951, + 0.010557631962001324, + 0.00969292689114809, + 0.010129495523869991, + 0.007285078521817923, + 0.009504219517111778, + 0.007264448329806328, + 0.006593913771212101, + 0.00376028916798532, + 0.003811619011685252, + 0.00434983940795064, + 0.002823792863637209, + 0.00417459849268198, + 0.003161435015499592, + 0.004144360311329365, + 0.0025176063645631075, + 0.007502770982682705, + 0.010072752833366394, + 0.0054706730879843235, + 0.01049740519374609, + 0.005239331163465977, + 0.009281325154006481, + 0.004756845999509096, + 0.007869812659919262, + 0.037682536989450455, + 0.01021178811788559, + 0.02975231595337391, + 0.01873423345386982, + 0.01483593974262476, + 0.08956712484359741, + 0.013046505860984325, + 0.02685319446027279, + 0.011966104619204998, + 0.014005624689161777, + 0.010883488692343235, + 0.013829465955495834, + 0.016510244458913803, + 0.014465837739408016, + 0.01119694858789444, + 0.014591312035918236, + 0.007946285419166088, + 0.006695383694022894, + 0.012284375727176666, + 0.013772941194474697, + 0.007777410093694925, + 0.009906241670250893, + 0.007735440507531166, + 0.006654368247836828, + 0.014282925985753536, + 0.011221688240766525, + 0.03342081606388092, + 0.012920920737087727, + 0.01673876866698265, + 0.020453443750739098, + 0.013137667439877987, + 0.020985577255487442, + 0.010049032047390938, + 0.014224730432033539, + 0.016071969643235207, + 0.02087586745619774, + 0.01003390271216631, + 0.012892216444015503, + 0.009001899510622025, + 0.012765458784997463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 6 + ], + "selected_reward": -0.01255913358181715 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607493683695793 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.01607493683695793, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607493683695793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06832022964954376, + 0.07408501207828522, + 0.09549394994974136, + 0.09973116219043732, + 0.21030999720096588, + 0.1305173635482788, + 0.08479151129722595, + 0.04146089404821396, + 0.12431838363409042, + 0.07097149640321732 + ], + "credit_map": [ + 0.008987300097942352, + 0.01028726901859045, + 0.011812739074230194, + 0.006149887107312679, + 0.008741826750338078, + 0.006199593655765057, + 0.008982391096651554, + 0.007159218657761812, + 0.011923123151063919, + 0.00811123475432396, + 0.011187226511538029, + 0.006539086811244488, + 0.0069167474284768105, + 0.015488668344914913, + 0.006821710616350174, + 0.007097211200743914, + 0.011603595688939095, + 0.012169011868536472, + 0.02115151286125183, + 0.0115636195987463, + 0.0074310628697276115, + 0.012433416210114956, + 0.009516573511064053, + 0.00962516013532877, + 0.011911032721400261, + 0.014011732302606106, + 0.020103054121136665, + 0.010472551919519901, + 0.010415459983050823, + 0.011294648982584476, + 0.012266219593584538, + 0.009256464429199696, + 0.048856690526008606, + 0.02579335868358612, + 0.019162621349096298, + 0.018743842840194702, + 0.015245192684233189, + 0.0552121140062809, + 0.012661647982895374, + 0.014634527266025543, + 0.027075689285993576, + 0.006913806311786175, + 0.012153112329542637, + 0.007148630451411009, + 0.0064213513396680355, + 0.049608636647462845, + 0.006426720879971981, + 0.01476940605789423, + 0.00974463764578104, + 0.0071063051000237465, + 0.01634528860449791, + 0.011523004621267319, + 0.010427327826619148, + 0.011929293163120747, + 0.007524424232542515, + 0.010191230103373528, + 0.007972447201609612, + 0.0029633392114192247, + 0.007392311468720436, + 0.0049850898794829845, + 0.003888026112690568, + 0.0057089366018772125, + 0.003675573505461216, + 0.004875166807323694, + 0.015979386866092682, + 0.009723847731947899, + 0.025977641344070435, + 0.010932298377156258, + 0.012375295162200928, + 0.01854829303920269, + 0.009593273513019085, + 0.021188344806432724, + 0.0065583051182329655, + 0.008490919135510921, + 0.01020744163542986, + 0.011190921068191528, + 0.00954811554402113, + 0.008730852976441383, + 0.006330482661724091, + 0.00991445779800415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.01607493683695793 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811153024435043 + ], + "opposite_logprob_tokens": [ + -3.288111448287964 + ], + "logprob_reward": -0.03811153024435043, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811153024435043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07002658396959305, + 0.05214626342058182, + 0.07394403219223022, + 0.09352909028530121, + 0.08682417124509811, + 0.1156863421201706, + 0.10914303362369537, + 0.15102477371692657, + 0.1616211235523224, + 0.08605460077524185 + ], + "credit_map": [ + 0.009940729476511478, + 0.009828032925724983, + 0.01074298471212387, + 0.00779452221468091, + 0.007938826456665993, + 0.007050099316984415, + 0.007599906995892525, + 0.009131481871008873, + 0.007691511884331703, + 0.006060593295842409, + 0.005079158581793308, + 0.006008966825902462, + 0.0052186427637934685, + 0.009431266225874424, + 0.0058703175745904446, + 0.006785808131098747, + 0.008818761445581913, + 0.010534862987697124, + 0.013157387264072895, + 0.008007204160094261, + 0.008392889983952045, + 0.00813442561775446, + 0.007768547628074884, + 0.009129956364631653, + 0.016562003642320633, + 0.00901635643094778, + 0.01675310730934143, + 0.009376843459904194, + 0.009970209561288357, + 0.01622111164033413, + 0.007663119118660688, + 0.007966347970068455, + 0.015531033277511597, + 0.011234026402235031, + 0.008213013410568237, + 0.010731223970651627, + 0.00681106885895133, + 0.018450984731316566, + 0.006710582412779331, + 0.009142246097326279, + 0.018029268831014633, + 0.005704209208488464, + 0.014285949058830738, + 0.009315253235399723, + 0.007428730838000774, + 0.0419747494161129, + 0.006474149879068136, + 0.012474020943045616, + 0.01124847773462534, + 0.017698021605610847, + 0.012968389317393303, + 0.01424434408545494, + 0.014905862510204315, + 0.011176405474543571, + 0.012135382741689682, + 0.014766144566237926, + 0.0164450965821743, + 0.012369843199849129, + 0.030240604653954506, + 0.026536863297224045, + 0.009732616133987904, + 0.026783205568790436, + 0.014663323760032654, + 0.014253220520913601, + 0.015031248331069946, + 0.01615573838353157, + 0.03140905126929283, + 0.02146613597869873, + 0.01606069877743721, + 0.02075710892677307, + 0.017723752185702324, + 0.023017384111881256, + 0.007735242135822773, + 0.010407846421003342, + 0.008397744968533516, + 0.014376730658113956, + 0.015327705070376396, + 0.00986133236438036, + 0.010095176286995411, + 0.009852821938693523 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.03811153024435043 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898151129484177 + ], + "opposite_logprob_tokens": [ + -2.578981399536133 + ], + "logprob_reward": -0.07898151129484177, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898151129484177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07210676372051239, + 0.055440496653318405, + 0.07705958932638168, + 0.03771202266216278, + 0.05173295736312866, + 0.14018206298351288, + 0.07045943289995193, + 0.11571138352155685, + 0.3224424123764038, + 0.05715286359190941 + ], + "credit_map": [ + 0.013533313758671284, + 0.009639604948461056, + 0.007362599950283766, + 0.006921245716512203, + 0.0076211197301745415, + 0.012201344594359398, + 0.007615362759679556, + 0.0072121708653867245, + 0.010132969357073307, + 0.006823292467743158, + 0.004425263497978449, + 0.00562912505120039, + 0.006018536631017923, + 0.011952446773648262, + 0.005857671145349741, + 0.0046011898666620255, + 0.007192431949079037, + 0.01343204453587532, + 0.0074703274294734, + 0.012461579404771328, + 0.009409726597368717, + 0.010652218014001846, + 0.00905723124742508, + 0.007384035736322403, + 0.006871814839541912, + 0.003949229139834642, + 0.006784067954868078, + 0.0041285790503025055, + 0.0032558084931224585, + 0.0067504397593438625, + 0.0028488547541201115, + 0.00312322867102921, + 0.011342120356857777, + 0.0042043947614729404, + 0.005102080758661032, + 0.004527340643107891, + 0.003601962933316827, + 0.014637481421232224, + 0.0032336395233869553, + 0.005083934403955936, + 0.029306013137102127, + 0.005200148094445467, + 0.013008266687393188, + 0.007350856903940439, + 0.009084988385438919, + 0.051605235785245895, + 0.008898149244487286, + 0.01572839356958866, + 0.0073099564760923386, + 0.009562513791024685, + 0.01013607531785965, + 0.008786968886852264, + 0.007580361794680357, + 0.007846084423363209, + 0.007027156185358763, + 0.01221031416207552, + 0.01248792465776205, + 0.010863369330763817, + 0.023435376584529877, + 0.01885107532143593, + 0.007778332103043795, + 0.021025758236646652, + 0.009678775444626808, + 0.011590773239731789, + 0.033288780599832535, + 0.025183623656630516, + 0.07263646274805069, + 0.02723781019449234, + 0.03785217925906181, + 0.04657606780529022, + 0.03163023665547371, + 0.048037271946668625, + 0.006661512888967991, + 0.005108135286718607, + 0.011934882029891014, + 0.007022918201982975, + 0.006394616328179836, + 0.008084218949079514, + 0.004936356097459793, + 0.007010222412645817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.07898151129484177 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, taken from a medium close-up and slightly from an upward angle, captures a man in extreme exhaustion. His messy, short, dark hair, slightly damp with sweat, clings to his pale forehead. His sunken eyes are etched with dark circles, and every fine line on his face bears the weight of mental strain. He stands before a rough concrete wall. A single, harsh, cold-blue light source, emanating from the side and below, sharply cuts his outline and casts a vast, dramatically elongating shadow behind him. As he slowly, almost in submission, closes his heavy eyes, his facial muscles instantly relax as all support is released. The once clear shadow behind him wriggles and deforms, stretching infinitely upwards, seemingly consuming the space. The entire image is shrouded in a cold, blue hue. The light creates sharp highlights on the man's high nose and cheekbones, contrasting sharply with the deep shadows on his cheeks and neck, creating an atmosphere of profound loneliness and mental exhaustion, imbued with a cinematic sense of oppression.", + "seed": 18790, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_11/index_310_seed_18790_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04041740335524082, + "vlm_logprob": -0.04041740335524082, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.04041740335524082, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032204609364271164 + ], + "opposite_logprob_tokens": [ + -4.157204627990723 + ], + "logprob_reward": -0.032204609364271164, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.032204609364271164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07471641898155212, + 0.042347390204668045, + 0.06478478014469147, + 0.018850017338991165, + 0.025237826630473137, + 0.10090862214565277, + 0.0507110059261322, + 0.04860097914934158, + 0.0687924325466156, + 0.5050504803657532 + ], + "credit_map": [ + 0.006207215134054422, + 0.002983472077175975, + 0.012948105111718178, + 0.00371599686332047, + 0.011241573840379715, + 0.005536139011383057, + 0.011023818515241146, + 0.021060099825263023, + 0.004725336562842131, + 0.0018049528589472175, + 0.011914415284991264, + 0.0025983192026615143, + 0.004745775368064642, + 0.002157001057639718, + 0.0049682618118822575, + 0.009433326311409473, + 0.007405134383589029, + 0.0018108097137883306, + 0.02412291057407856, + 0.002991998568177223, + 0.002456594491377473, + 0.003502104664221406, + 0.0027836011722683907, + 0.019711630418896675, + 0.0019877273589372635, + 0.0015750867314636707, + 0.0037993374280631542, + 0.0014094710350036621, + 0.002761341165751219, + 0.00167681067250669, + 0.0025082107167690992, + 0.00313203246332705, + 0.0031204463448375463, + 0.0024838398676365614, + 0.0034976790193468332, + 0.0017574891680851579, + 0.003399468492716551, + 0.002870977856218815, + 0.0031060855835676193, + 0.0050018406473100185, + 0.009734006598591805, + 0.009407663717865944, + 0.010403968393802643, + 0.007543051615357399, + 0.008662433363497257, + 0.014458276331424713, + 0.014195863157510757, + 0.026503365486860275, + 0.004526985343545675, + 0.0036476021632552147, + 0.006463765632361174, + 0.0027828405145555735, + 0.004650838207453489, + 0.008379796519875526, + 0.006626402959227562, + 0.01363277342170477, + 0.006058411207050085, + 0.002419480588287115, + 0.006584236864000559, + 0.002496682107448578, + 0.0025655806530267, + 0.008389848284423351, + 0.0044616893865168095, + 0.01562504842877388, + 0.01338297687470913, + 0.003874262562021613, + 0.01760639064013958, + 0.004108770284801722, + 0.004440692253410816, + 0.008473127149045467, + 0.004485566169023514, + 0.012420653365552425, + 0.15585452318191528, + 0.013319593854248524, + 0.2129831612110138, + 0.0246969535946846, + 0.017315221950411797, + 0.03467682749032974, + 0.009645448997616768, + 0.03655879572033882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.032204609364271164 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016094176098704338 + ], + "opposite_logprob_tokens": [ + -4.141094207763672 + ], + "logprob_reward": -0.016094176098704338, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016094176098704338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12746697664260864, + 0.06623746454715729, + 0.27665022015571594, + 0.03867094963788986, + 0.06964094936847687, + 0.08682464808225632, + 0.06360703706741333, + 0.05041446164250374, + 0.16099879145622253, + 0.05948852002620697 + ], + "credit_map": [ + 0.012863511219620705, + 0.010791345499455929, + 0.014880738221108913, + 0.007911073043942451, + 0.027894502505660057, + 0.01226375624537468, + 0.024274274706840515, + 0.016587767750024796, + 0.00806187093257904, + 0.006194787565618753, + 0.007654616609215736, + 0.005101221613585949, + 0.009771076031029224, + 0.006604878231883049, + 0.010806338861584663, + 0.012042676098644733, + 0.028723347932100296, + 0.010689113289117813, + 0.09335487335920334, + 0.014055737294256687, + 0.013334850780665874, + 0.014602446928620338, + 0.02181662805378437, + 0.08007322251796722, + 0.0049219997599720955, + 0.004439563024789095, + 0.007103877607733011, + 0.0035316424909979105, + 0.004664323758333921, + 0.0033987718634307384, + 0.005904635414481163, + 0.004706141073256731, + 0.007012802176177502, + 0.00748166861012578, + 0.010943597182631493, + 0.005208455491811037, + 0.006194487679749727, + 0.00850007589906454, + 0.008285683579742908, + 0.016014177352190018, + 0.00939016230404377, + 0.008185476064682007, + 0.009242626838386059, + 0.005237140692770481, + 0.007250976283103228, + 0.012364467605948448, + 0.011660745367407799, + 0.02349306270480156, + 0.006764102727174759, + 0.005205916706472635, + 0.0109248710796237, + 0.0048760236240923405, + 0.006494180765002966, + 0.007063623983412981, + 0.010673397220671177, + 0.011604918166995049, + 0.009928196668624878, + 0.0046534230932593346, + 0.00494564650580287, + 0.003166364738717675, + 0.0034230665769428015, + 0.00869273766875267, + 0.004393107257783413, + 0.011211915872991085, + 0.027178803458809853, + 0.008330436423420906, + 0.05208444595336914, + 0.00972655974328518, + 0.015000722371041775, + 0.017711620777845383, + 0.01822705566883087, + 0.012739153578877449, + 0.009665670804679394, + 0.0040705460123717785, + 0.010529679246246815, + 0.004163546487689018, + 0.0068749417550861835, + 0.004787486512213945, + 0.009907272644340992, + 0.009489374235272408 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.016094176098704338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865659400820732 + ], + "opposite_logprob_tokens": [ + -3.048656702041626 + ], + "logprob_reward": -0.04865659400820732, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865659400820732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07940910756587982, + 0.14549550414085388, + 0.23914240300655365, + 0.022877054288983345, + 0.043217796832323074, + 0.12800155580043793, + 0.043755531311035156, + 0.06571388989686966, + 0.10320937633514404, + 0.1291777640581131 + ], + "credit_map": [ + 0.009009785018861294, + 0.006780713330954313, + 0.012227150611579418, + 0.006779203191399574, + 0.013776187784969807, + 0.007403995841741562, + 0.010288222692906857, + 0.013143854215741158, + 0.015754330903291702, + 0.010424570180475712, + 0.04440929740667343, + 0.012533406727015972, + 0.008747825399041176, + 0.006697040516883135, + 0.011857597157359123, + 0.03507143259048462, + 0.03015078231692314, + 0.021311860531568527, + 0.060972198843955994, + 0.019662339240312576, + 0.020861461758613586, + 0.017305949702858925, + 0.027951734140515327, + 0.040926091372966766, + 0.0030222823843359947, + 0.0024434830993413925, + 0.004739054013043642, + 0.0018664150265976787, + 0.00340366386808455, + 0.001512653543613851, + 0.001906018704175949, + 0.003983485046774149, + 0.004189899656921625, + 0.004359485115855932, + 0.007369727361947298, + 0.0028707515448331833, + 0.0057448735460639, + 0.004678621422499418, + 0.006134068127721548, + 0.007870371453464031, + 0.014794952236115932, + 0.012551010586321354, + 0.016412317752838135, + 0.008992131799459457, + 0.013117040507495403, + 0.01690084859728813, + 0.018701763823628426, + 0.026531491428613663, + 0.00519753759726882, + 0.004290449898689985, + 0.005541296675801277, + 0.0035252217203378677, + 0.0036059485282748938, + 0.005780435167253017, + 0.006959442049264908, + 0.008855199441313744, + 0.010816244408488274, + 0.004548695404082537, + 0.008266709744930267, + 0.0030208469834178686, + 0.006234390195459127, + 0.010838288813829422, + 0.005276946816593409, + 0.01671176590025425, + 0.018588237464427948, + 0.005298294592648745, + 0.027458131313323975, + 0.006003382615745068, + 0.007321582175791264, + 0.012314149178564548, + 0.007905705831944942, + 0.01831989362835884, + 0.040430858731269836, + 0.004399141762405634, + 0.04940102994441986, + 0.006285890471190214, + 0.004770045168697834, + 0.00829822663217783, + 0.0036224776413291693, + 0.011970099993050098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.04865659400820732 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308178275823593 + ], + "opposite_logprob_tokens": [ + -3.168081760406494 + ], + "logprob_reward": -0.04308178275823593, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308178275823593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043203067034482956, + 0.021787337958812714, + 0.09612863510847092, + 0.03805229067802429, + 0.12178005278110504, + 0.051514845341444016, + 0.03782733157277107, + 0.087058424949646, + 0.04824669659137726, + 0.45440131425857544 + ], + "credit_map": [ + 0.004273540806025267, + 0.003107482800260186, + 0.0067282323725521564, + 0.0025382046587765217, + 0.008044983260333538, + 0.00362192839384079, + 0.007886470295488834, + 0.007002221420407295, + 0.0027072352822870016, + 0.0024460800923407078, + 0.0033610721584409475, + 0.002022993052378297, + 0.003063580486923456, + 0.0019839489832520485, + 0.0038709568325430155, + 0.0023314706049859524, + 0.010074802674353123, + 0.0034117395989596844, + 0.033822935074567795, + 0.004553533159196377, + 0.004679965320974588, + 0.005618471186608076, + 0.004663724917918444, + 0.029303455725312233, + 0.004248605575412512, + 0.0018730657175183296, + 0.009143108502030373, + 0.002141504315659404, + 0.0031417361460626125, + 0.003856455907225609, + 0.004909642972052097, + 0.008738167583942413, + 0.017314573749899864, + 0.00660901702940464, + 0.02082766965031624, + 0.005659669637680054, + 0.014016193337738514, + 0.014468922279775143, + 0.011752947233617306, + 0.031131049618124962, + 0.004896860104054213, + 0.005672025959938765, + 0.006776658818125725, + 0.005711264908313751, + 0.005151828750967979, + 0.006372873205691576, + 0.00697236368432641, + 0.009960965253412724, + 0.005601733457297087, + 0.0023365712258964777, + 0.007782080210745335, + 0.002192703541368246, + 0.002910915995016694, + 0.004963140469044447, + 0.005026007071137428, + 0.007014174945652485, + 0.00979089830070734, + 0.004810533951967955, + 0.011226448230445385, + 0.004000616259872913, + 0.0037947376258671284, + 0.015793105587363243, + 0.006525163073092699, + 0.031116921454668045, + 0.009614181704819202, + 0.0031220088712871075, + 0.008317437954246998, + 0.00307669909670949, + 0.003809853456914425, + 0.005200076848268509, + 0.004294841084629297, + 0.010811597108840942, + 0.14023834466934204, + 0.011685733683407307, + 0.19132767617702484, + 0.02207382395863533, + 0.015360606834292412, + 0.032018031924963, + 0.009131819941103458, + 0.032565295696258545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 2 + ], + "selected_reward": -0.04308178275823593 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062049854546785355 + ], + "opposite_logprob_tokens": [ + -2.8120498657226562 + ], + "logprob_reward": -0.062049854546785355, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062049854546785355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19179174304008484, + 0.03121204860508442, + 0.1409309357404709, + 0.06937901675701141, + 0.05830825865268707, + 0.18795524537563324, + 0.1314123272895813, + 0.022641010582447052, + 0.05483679845929146, + 0.11153263598680496 + ], + "credit_map": [ + 0.017645498737692833, + 0.008413305506110191, + 0.039064060896635056, + 0.0076576280407607555, + 0.021812666207551956, + 0.0128502007573843, + 0.02283589541912079, + 0.061512477695941925, + 0.003913357388228178, + 0.0020239935256540775, + 0.008009846322238445, + 0.0019199000671505928, + 0.0037798420526087284, + 0.00200799573212862, + 0.003939727786928415, + 0.005617382936179638, + 0.014494284056127071, + 0.0036950535140931606, + 0.05186590179800987, + 0.005827055312693119, + 0.00501048332080245, + 0.00800199806690216, + 0.005267689470201731, + 0.04676845669746399, + 0.0068736569955945015, + 0.0032720929011702538, + 0.019761724397540092, + 0.0031523643992841244, + 0.004729066044092178, + 0.005286612547934055, + 0.0066012185998260975, + 0.019702279940247536, + 0.005317727569490671, + 0.007094617001712322, + 0.008726994507014751, + 0.004726163577288389, + 0.008731906302273273, + 0.006937358994036913, + 0.005740126594901085, + 0.01103336364030838, + 0.015797536820173264, + 0.018130550161004066, + 0.027728553861379623, + 0.01159963570535183, + 0.016165927052497864, + 0.025660939514636993, + 0.02256474643945694, + 0.050307340919971466, + 0.013748607598245144, + 0.008796704933047295, + 0.025120478123426437, + 0.007771536707878113, + 0.01216336339712143, + 0.01762024126946926, + 0.01730979047715664, + 0.02888159081339836, + 0.0029255200643092394, + 0.0021315263584256172, + 0.002492828294634819, + 0.0014308409299701452, + 0.00225004437379539, + 0.0031229995656758547, + 0.0023441831581294537, + 0.0059430645778775215, + 0.011269286274909973, + 0.0035372001584619284, + 0.008296414278447628, + 0.0038006161339581013, + 0.004150536842644215, + 0.005704853683710098, + 0.005461788736283779, + 0.012616102583706379, + 0.033522557467222214, + 0.00374541524797678, + 0.043825872242450714, + 0.0058943708427250385, + 0.004194618668407202, + 0.007738967426121235, + 0.003566149855032563, + 0.009044676087796688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.062049854546785355 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static shot, taken from a medium close-up and slightly from an upward angle, captures a man in extreme exhaustion. His messy, short, dark hair, slightly damp with sweat, clings to his pale forehead. His sunken eyes are etched with dark circles, and every fine line on his face bears the weight of mental strain. He stands before a rough concrete wall. A single, harsh, cold-blue light source, emanating from the side and below, sharply cuts his outline and casts a vast, dramatically elongating shadow behind him. As he slowly, almost in submission, closes his heavy eyes, his facial muscles instantly relax as all support is released. The once clear shadow behind him wriggles and deforms, stretching infinitely upwards, seemingly consuming the space. The entire image is shrouded in a cold, blue hue. The light creates sharp highlights on the man's high nose and cheekbones, contrasting sharply with the deep shadows on his cheeks and neck, creating an atmosphere of profound loneliness and mental exhaustion, imbued with a cinematic sense of oppression.", + "seed": 18791, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_11/index_310_seed_18791_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07295147180557252, + "vlm_logprob": -0.07295147180557252, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.07295147180557252, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.032165247946977615 + ], + "opposite_logprob_tokens": [ + -4.157165050506592 + ], + "logprob_reward": -0.032165247946977615, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.032165247946977615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17094267904758453, + 0.11675937473773956, + 0.07242539525032043, + 0.03995252028107643, + 0.09871479868888855, + 0.2114437073469162, + 0.09159731864929199, + 0.024652333930134773, + 0.10986597836017609, + 0.06364592164754868 + ], + "credit_map": [ + 0.027525221928954124, + 0.022623082622885704, + 0.02304898016154766, + 0.02097112499177456, + 0.015675801783800125, + 0.024711618199944496, + 0.0142845893278718, + 0.022102253511548042, + 0.015776174142956734, + 0.01421453058719635, + 0.016782857477664948, + 0.024751614779233932, + 0.012181883677840233, + 0.006892791483551264, + 0.013656471855938435, + 0.01250305213034153, + 0.01023748330771923, + 0.008549980819225311, + 0.010545649565756321, + 0.013062680140137672, + 0.007791885640472174, + 0.007108849473297596, + 0.008820083923637867, + 0.006308786105364561, + 0.00495029566809535, + 0.004431010223925114, + 0.004487491678446531, + 0.007522658444941044, + 0.005339958239346743, + 0.003926955163478851, + 0.006027983967214823, + 0.003266167128458619, + 0.009898878633975983, + 0.01116789598017931, + 0.011444007977843285, + 0.017020052298903465, + 0.014241572469472885, + 0.012644557282328606, + 0.01336664892733097, + 0.008931177668273449, + 0.02245398610830307, + 0.023243265226483345, + 0.025707244873046875, + 0.0308547280728817, + 0.028217479586601257, + 0.03191637247800827, + 0.025022761896252632, + 0.024027863517403603, + 0.009294213727116585, + 0.0076811593025922775, + 0.009565106593072414, + 0.013120496645569801, + 0.013302169740200043, + 0.01680891588330269, + 0.01215661782771349, + 0.009668637998402119, + 0.0042558335699141026, + 0.002252285834401846, + 0.003291298868134618, + 0.0027399465907365084, + 0.0026270390953868628, + 0.0033999113366007805, + 0.0032396491151303053, + 0.002846368122845888, + 0.012686328031122684, + 0.018419906497001648, + 0.012533760629594326, + 0.02156892418861389, + 0.01560505386441946, + 0.007722037844359875, + 0.012802219949662685, + 0.008527744561433792, + 0.006144343409687281, + 0.008450053632259369, + 0.007447248790413141, + 0.008652337826788425, + 0.010273798368871212, + 0.006422612350434065, + 0.009584669023752213, + 0.006670854985713959 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 1 + ], + "selected_reward": -0.032165247946977615 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813987225294113 + ], + "opposite_logprob_tokens": [ + -3.288139820098877 + ], + "logprob_reward": -0.03813987225294113, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813987225294113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.196045383810997, + 0.13186661899089813, + 0.048996612429618835, + 0.05170677229762077, + 0.15949952602386475, + 0.16222773492336273, + 0.034864526242017746, + 0.03039609268307686, + 0.13738679885864258, + 0.0470099151134491 + ], + "credit_map": [ + 0.028410781174898148, + 0.023789627477526665, + 0.02314363606274128, + 0.029400603845715523, + 0.020678440108895302, + 0.032446105033159256, + 0.01791750267148018, + 0.02025868371129036, + 0.01508273184299469, + 0.01663937233388424, + 0.014420383609831333, + 0.022066254168748856, + 0.018700525164604187, + 0.010823228396475315, + 0.017078835517168045, + 0.017055289819836617, + 0.005411400459706783, + 0.008135342970490456, + 0.005129500757902861, + 0.008426687680184841, + 0.0062039257027208805, + 0.00518184807151556, + 0.0060569243505597115, + 0.004450981505215168, + 0.006409968249499798, + 0.005245818756520748, + 0.005695858970284462, + 0.01079502236098051, + 0.006683620624244213, + 0.006460563279688358, + 0.006935006473213434, + 0.003480913583189249, + 0.010492450557649136, + 0.014591433107852936, + 0.015711156651377678, + 0.021138304844498634, + 0.018244454637169838, + 0.03265124186873436, + 0.023636752739548683, + 0.023033738136291504, + 0.015089141204953194, + 0.02003728412091732, + 0.01865067519247532, + 0.026738783344626427, + 0.02126273699104786, + 0.02202375791966915, + 0.0204771738499403, + 0.017948180437088013, + 0.004707980901002884, + 0.003313750959932804, + 0.004084086045622826, + 0.006124103907495737, + 0.004429563879966736, + 0.003992925398051739, + 0.004981327801942825, + 0.0032307878136634827, + 0.003488924354314804, + 0.0024081633891910315, + 0.0036144007463008165, + 0.00426413444802165, + 0.0035539167001843452, + 0.004986971616744995, + 0.00415375642478466, + 0.0039258250035345554, + 0.02313021384179592, + 0.016514258459210396, + 0.015950875356793404, + 0.014044410549104214, + 0.019715771079063416, + 0.017778072506189346, + 0.013624090701341629, + 0.016629114747047424, + 0.00493755005300045, + 0.006119871977716684, + 0.005223348271101713, + 0.006617200560867786, + 0.007215343415737152, + 0.004124782048165798, + 0.0074407984502613544, + 0.005331019870936871 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.03813987225294113 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370547667145729 + ], + "opposite_logprob_tokens": [ + -3.408705472946167 + ], + "logprob_reward": -0.03370547667145729, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370547667145729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15059344470500946, + 0.08188308775424957, + 0.06040104478597641, + 0.051308855414390564, + 0.12371589988470078, + 0.19244644045829773, + 0.08652523905038834, + 0.06210705637931824, + 0.10427405685186386, + 0.08674488216638565 + ], + "credit_map": [ + 0.023475833237171173, + 0.019308697432279587, + 0.018237916752696037, + 0.024556834250688553, + 0.0167230311781168, + 0.015218346379697323, + 0.015538560226559639, + 0.017534220591187477, + 0.013875514268875122, + 0.00910566933453083, + 0.014229550957679749, + 0.01502747368067503, + 0.007933971472084522, + 0.004361604806035757, + 0.010153280571103096, + 0.007196018006652594, + 0.008578713983297348, + 0.007807161193341017, + 0.008866493590176105, + 0.008361565880477428, + 0.007081661373376846, + 0.004929245915263891, + 0.006875364109873772, + 0.007900835946202278, + 0.007954421453177929, + 0.0038295614067465067, + 0.008606288582086563, + 0.006562255322933197, + 0.007052531000226736, + 0.005758955609053373, + 0.006523873656988144, + 0.005020968150347471, + 0.010756304487586021, + 0.010778562165796757, + 0.0111233526840806, + 0.018020177260041237, + 0.014554754830896854, + 0.024474620819091797, + 0.018575351685285568, + 0.015432772226631641, + 0.04412403330206871, + 0.023745890706777573, + 0.03331064060330391, + 0.026909328997135162, + 0.018513847142457962, + 0.018066519871354103, + 0.01579792983829975, + 0.01197823230177164, + 0.00871636439114809, + 0.009060985408723354, + 0.010009201243519783, + 0.013378674164414406, + 0.01281367614865303, + 0.013029662892222404, + 0.011848590336740017, + 0.007668077014386654, + 0.006427778396755457, + 0.005422892514616251, + 0.00810793787240982, + 0.008357136510312557, + 0.007139919325709343, + 0.0103813037276268, + 0.00778233865275979, + 0.008487747050821781, + 0.02133754827082157, + 0.007281193509697914, + 0.014909404329955578, + 0.010157234966754913, + 0.014471775852143764, + 0.0138244042173028, + 0.010386491194367409, + 0.01190599612891674, + 0.018487948924303055, + 0.010670003481209278, + 0.014403821900486946, + 0.01024610921740532, + 0.00981472060084343, + 0.005245185922831297, + 0.00763644278049469, + 0.010240644216537476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.03370547667145729 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10034331679344177 + ], + "opposite_logprob_tokens": [ + -2.3503432273864746 + ], + "logprob_reward": -0.10034331679344177, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10034331679344177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15885357558727264, + 0.0687737911939621, + 0.11060471087694168, + 0.07903321832418442, + 0.14214110374450684, + 0.08747189491987228, + 0.058302272111177444, + 0.05064670741558075, + 0.20478133857250214, + 0.039391353726387024 + ], + "credit_map": [ + 0.028956333175301552, + 0.015446660108864307, + 0.031118519604206085, + 0.01741166226565838, + 0.01417036633938551, + 0.021551180630922318, + 0.01336121466010809, + 0.01683763600885868, + 0.009773272089660168, + 0.006986006163060665, + 0.010235692374408245, + 0.00833713449537754, + 0.009238224476575851, + 0.005645398981869221, + 0.010266871191561222, + 0.008291186764836311, + 0.015558804385364056, + 0.013711984269320965, + 0.017542799934744835, + 0.019224101677536964, + 0.012953083962202072, + 0.008703114464879036, + 0.014799417927861214, + 0.00811140425503254, + 0.00906062126159668, + 0.008200634270906448, + 0.009877432137727737, + 0.014983667060732841, + 0.011338424868881702, + 0.009079938754439354, + 0.010710818693041801, + 0.005781681276857853, + 0.012083392590284348, + 0.012946492061018944, + 0.015038182027637959, + 0.01707577332854271, + 0.01451351772993803, + 0.030756451189517975, + 0.019878527149558067, + 0.019848762080073357, + 0.008617118000984192, + 0.007119493093341589, + 0.009217862971127033, + 0.016373516991734505, + 0.012951318174600601, + 0.012104415334761143, + 0.012910186313092709, + 0.008177981711924076, + 0.00739546213299036, + 0.005227189045399427, + 0.007041678763926029, + 0.010135565884411335, + 0.007867148146033287, + 0.006066716741770506, + 0.007746192626655102, + 0.0068223183043301105, + 0.0047265696339309216, + 0.0035523034166544676, + 0.0050497050397098064, + 0.006091617047786713, + 0.007849195972084999, + 0.008885566145181656, + 0.007101318333297968, + 0.007390433456748724, + 0.023874597623944283, + 0.026250993832945824, + 0.020454714074730873, + 0.04432178661227226, + 0.026199035346508026, + 0.02304922603070736, + 0.019581563770771027, + 0.021049417555332184, + 0.005417855456471443, + 0.0034164180979132652, + 0.005236208438873291, + 0.004048312082886696, + 0.006152153015136719, + 0.0035522079560905695, + 0.005252805072814226, + 0.006315397098660469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.10034331679344177 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium close-up taken from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's hair messy, short, dark, and slightly damp with sweat clinging to his forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single harsh cold-blue light source coming from the side and below?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the shadow behind the man wriggle and deform while stretching upwards?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the man close his heavy eyes and relax his facial muscles as the shadow deforms?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16040344536304474 + ], + "opposite_logprob_tokens": [ + -1.9104034900665283 + ], + "logprob_reward": -0.16040344536304474, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16040344536304474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1483699083328247, + 0.052609626203775406, + 0.06766947358846664, + 0.05923433229327202, + 0.13031470775604248, + 0.08302310854196548, + 0.09094209223985672, + 0.03330797329545021, + 0.29288235306739807, + 0.041646361351013184 + ], + "credit_map": [ + 0.020194552838802338, + 0.018228353932499886, + 0.01934511587023735, + 0.02438628487288952, + 0.018268143758177757, + 0.010211740620434284, + 0.02093750797212124, + 0.016798213124275208, + 0.006417613010853529, + 0.007102583069354296, + 0.007312419358640909, + 0.007303635124117136, + 0.0062465667724609375, + 0.004001011606305838, + 0.007996139116585255, + 0.0062296586111187935, + 0.00949487742036581, + 0.007340596057474613, + 0.010045030154287815, + 0.012088967487215996, + 0.0076236785389482975, + 0.0056832036934792995, + 0.009221906773746014, + 0.006171212997287512, + 0.007479513995349407, + 0.005744975060224533, + 0.006199803203344345, + 0.012137837707996368, + 0.007532901596277952, + 0.008065827190876007, + 0.008165227249264717, + 0.003908243961632252, + 0.01512114517390728, + 0.015718430280685425, + 0.016984689980745316, + 0.016161108389496803, + 0.021903324872255325, + 0.01495401095598936, + 0.020268168300390244, + 0.009203830733895302, + 0.008195177651941776, + 0.009554841555655003, + 0.010148923844099045, + 0.01436145044863224, + 0.009816336445510387, + 0.011653410270810127, + 0.009429823607206345, + 0.009863143786787987, + 0.007747530471533537, + 0.006960724946111441, + 0.011122477240860462, + 0.013639861717820168, + 0.012666109018027782, + 0.017571430653333664, + 0.011351368390023708, + 0.009882588870823383, + 0.00466001033782959, + 0.0026538928505033255, + 0.004322056658565998, + 0.0035244778264313936, + 0.004608426708728075, + 0.004402471240609884, + 0.00445313286036253, + 0.004683504346758127, + 0.022990094497799873, + 0.03140997886657715, + 0.02600480057299137, + 0.04302404075860977, + 0.05827195569872856, + 0.041652075946331024, + 0.04027751833200455, + 0.02925189398229122, + 0.005384796764701605, + 0.004649503156542778, + 0.0047835600562393665, + 0.005198150407522917, + 0.005597952287644148, + 0.003877925220876932, + 0.005730703007429838, + 0.006423771847039461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.16040344536304474 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_12/index_tensor([3944, 3944])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_12/index_tensor([3944, 3944])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7ba12e149cc396a4e3baef28d14c7541617548f1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_12/index_tensor([3944, 3944])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic drone shot follows closely, following an Asian man in a crisp, dark charcoal-gray suit as he glides smoothly through the vast air between two skyscrapers with a grace that defies the laws of physics. Slightly above his horizon, the camera captures his elongated, steady form, as if he were stepping on invisible air, his movements reminiscent of a martial arts master. Beneath his impeccably tailored suit jacket lies a crisp white shirt and a narrow, dark silk tie that billows backwards in the wind, its hems undulating but never a touch disheveled. His face is calm and focused, his gaze fixed firmly on the platform of the next building ahead. The fading light of the evening sun casts a sharp golden outline on his dark suit, casting a shifting, warm glow across the surrounding glass walls. Below is a bustling city canyon. The lights of countless vehicles have merged into a flowing river of light. The sky is dyed a gradient of warm orange-red and deep purple. Thin clouds and mist are lingering between the buildings. The whole picture combines the elegance of martial arts, the grandeur of surrealism and the coldness of the modern city, creating a breathtaking visual spectacle.", + "seed": 30076, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_12/index_3944_seed_30076_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19408349376171827, + "vlm_logprob": -0.19408349376171827, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.19408349376171827, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22878120839595795 + ], + "opposite_logprob_tokens": [ + -1.6037812232971191 + ], + "logprob_reward": -0.22878120839595795, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22878120839595795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13245925307273865, + 0.04150179401040077, + 0.04688182473182678, + 0.10425911098718643, + 0.1149686872959137, + 0.04573974385857582, + 0.10523984581232071, + 0.13368278741836548, + 0.18940995633602142, + 0.08585700392723083 + ], + "credit_map": [ + 0.019665265455842018, + 0.01586201600730419, + 0.01734805852174759, + 0.01263347640633583, + 0.016769446432590485, + 0.02007921412587166, + 0.01834019459784031, + 0.011761578731238842, + 0.0069288103841245174, + 0.0058961957693099976, + 0.004492942243814468, + 0.004168402403593063, + 0.005334125831723213, + 0.006387878675013781, + 0.004702787380665541, + 0.003590649226680398, + 0.009866460226476192, + 0.008425362408161163, + 0.005802466068416834, + 0.005116644781082869, + 0.003378473222255707, + 0.0067842029966413975, + 0.0032620998099446297, + 0.004246115684509277, + 0.011601979844272137, + 0.010384087450802326, + 0.0158710777759552, + 0.014049042947590351, + 0.009436195716261864, + 0.0202085692435503, + 0.012419831939041615, + 0.010288326069712639, + 0.025875002145767212, + 0.011844036169350147, + 0.027211779728531837, + 0.007419496309012175, + 0.009916850365698338, + 0.01399816945195198, + 0.0065977079793810844, + 0.012105644680559635, + 0.003899316070601344, + 0.004940604791045189, + 0.007363425567746162, + 0.003905684920027852, + 0.005580442026257515, + 0.00802542082965374, + 0.005975507199764252, + 0.006049343850463629, + 0.015057237818837166, + 0.015815822407603264, + 0.01616084948182106, + 0.010723876766860485, + 0.013425251469016075, + 0.009435184299945831, + 0.00953265093266964, + 0.015088966116309166, + 0.019385825842618942, + 0.023603923618793488, + 0.013928284868597984, + 0.0158279687166214, + 0.02157675102353096, + 0.014467882923781872, + 0.014141235500574112, + 0.010750924237072468, + 0.026117855682969093, + 0.009337710216641426, + 0.017111359164118767, + 0.009584100916981697, + 0.040216442197561264, + 0.051454707980155945, + 0.016239706426858902, + 0.019348077476024628, + 0.01652442291378975, + 0.0058184219524264336, + 0.015802226960659027, + 0.007624211721122265, + 0.009965743869543076, + 0.008105196990072727, + 0.009318910539150238, + 0.012697862461209297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.22878120839595795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934544444084167 + ], + "opposite_logprob_tokens": [ + -0.6934544444084167 + ], + "logprob_reward": -0.6934544444084167, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934544444084167, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17921622097492218, + 0.04989268258213997, + 0.022649621590971947, + 0.08114363998174667, + 0.0632990151643753, + 0.11317639797925949, + 0.04644273221492767, + 0.042216479778289795, + 0.3355987071990967, + 0.06636446714401245 + ], + "credit_map": [ + 0.017149273306131363, + 0.008749271742999554, + 0.028462380170822144, + 0.008500725962221622, + 0.00819364283233881, + 0.07533689588308334, + 0.006556735839694738, + 0.02626729942858219, + 0.006392146460711956, + 0.004479687195271254, + 0.005743775516748428, + 0.005745472386479378, + 0.00450272299349308, + 0.009130613878369331, + 0.006296362262219191, + 0.007601904217153788, + 0.004263488110154867, + 0.001978624379262328, + 0.003764505498111248, + 0.002592697273939848, + 0.001952210790477693, + 0.003877473995089531, + 0.002338492078706622, + 0.0018821293488144875, + 0.009140025824308395, + 0.006690761540085077, + 0.01805732026696205, + 0.007530523464083672, + 0.005091117229312658, + 0.020090259611606598, + 0.006803891621530056, + 0.007739743683487177, + 0.013271893374621868, + 0.006121421232819557, + 0.016148678958415985, + 0.004543581046164036, + 0.005079262889921665, + 0.006730775814503431, + 0.004255341365933418, + 0.007148061413317919, + 0.016578130424022675, + 0.006476508919149637, + 0.014749659225344658, + 0.00568750174716115, + 0.012934153899550438, + 0.03409595787525177, + 0.010283403098583221, + 0.012371079996228218, + 0.006088484078645706, + 0.007222538813948631, + 0.005182171240448952, + 0.007765715941786766, + 0.005436804611235857, + 0.005985795054584742, + 0.004960447549819946, + 0.003800771664828062, + 0.006028022617101669, + 0.006194675341248512, + 0.004971856251358986, + 0.005162052344530821, + 0.004295563790947199, + 0.0049912650138139725, + 0.0047999280504882336, + 0.0057731145061552525, + 0.040626224130392075, + 0.015215499326586723, + 0.03843129426240921, + 0.010555718094110489, + 0.06204118952155113, + 0.07999297231435776, + 0.037858761847019196, + 0.05087703466415405, + 0.010604574345052242, + 0.003315156325697899, + 0.006165438797324896, + 0.0030558446887880564, + 0.016144588589668274, + 0.015491027384996414, + 0.005821969825774431, + 0.005765870213508606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.6934544444084167 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026357805356383324 + ], + "opposite_logprob_tokens": [ + -3.651357889175415 + ], + "logprob_reward": -0.026357805356383324, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026357805356383324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24115954339504242, + 0.044406574219465256, + 0.060947485268116, + 0.09138166904449463, + 0.05859173834323883, + 0.07456283271312714, + 0.06371571868658066, + 0.08191432058811188, + 0.22112029790878296, + 0.062199775129556656 + ], + "credit_map": [ + 0.02532106824219227, + 0.011658040806651115, + 0.04028387740254402, + 0.012499300763010979, + 0.012012913823127747, + 0.09736563265323639, + 0.008698329329490662, + 0.03332039341330528, + 0.007363309618085623, + 0.004824868869036436, + 0.005365174263715744, + 0.004865061957389116, + 0.004842639435082674, + 0.006108600180596113, + 0.005250775255262852, + 0.005786149762570858, + 0.012902642600238323, + 0.006707506719976664, + 0.010229856707155704, + 0.006641610059887171, + 0.004473074804991484, + 0.009764370508491993, + 0.004221339710056782, + 0.006007089279592037, + 0.010633363388478756, + 0.007395695894956589, + 0.022448472678661346, + 0.008928283117711544, + 0.005469959229230881, + 0.02049250900745392, + 0.00761338509619236, + 0.008400005288422108, + 0.009888501837849617, + 0.006544473115354776, + 0.01166848000138998, + 0.005034403409808874, + 0.005483527667820454, + 0.008928047493100166, + 0.004827835597097874, + 0.006216477137058973, + 0.007546309381723404, + 0.01132527831941843, + 0.008399272337555885, + 0.010351400822401047, + 0.011415982618927956, + 0.01188816037029028, + 0.007407009601593018, + 0.006229425314813852, + 0.009361924603581429, + 0.00995410606265068, + 0.008159555494785309, + 0.010098490864038467, + 0.00716136721894145, + 0.00855337269604206, + 0.005567545536905527, + 0.004859358072280884, + 0.02046036720275879, + 0.008558047004044056, + 0.007769614923745394, + 0.007732908241450787, + 0.008203706704080105, + 0.015180948190391064, + 0.006829271093010902, + 0.007179465144872665, + 0.02969628944993019, + 0.007849572226405144, + 0.018760791048407555, + 0.008280964568257332, + 0.04912073165178299, + 0.06330178678035736, + 0.019011329859495163, + 0.02509884722530842, + 0.006321210414171219, + 0.003737723222002387, + 0.0072997258976101875, + 0.0029862630181014538, + 0.01669653132557869, + 0.01711263321340084, + 0.004059209488332272, + 0.003986484371125698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.026357805356383324 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628357969224453 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628357969224453, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628357969224453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3641025125980377, + 0.06363068521022797, + 0.07073690742254257, + 0.04964762181043625, + 0.08649926632642746, + 0.0766448974609375, + 0.048662446439266205, + 0.03240826725959778, + 0.16424717009067535, + 0.04342019185423851 + ], + "credit_map": [ + 0.04446336254477501, + 0.019326481968164444, + 0.057119276374578476, + 0.0168176107108593, + 0.017029043287038803, + 0.1513955146074295, + 0.010357527993619442, + 0.047593697905540466, + 0.012673066928982735, + 0.009327100589871407, + 0.006951390765607357, + 0.006539465393871069, + 0.004593492951244116, + 0.01103338785469532, + 0.006301931571215391, + 0.006210849620401859, + 0.015575585886836052, + 0.008171929977834225, + 0.010562890209257603, + 0.008113265968859196, + 0.004884349647909403, + 0.010635649785399437, + 0.00569379236549139, + 0.007099441718310118, + 0.005631726700812578, + 0.003995859995484352, + 0.01041833683848381, + 0.005190393887460232, + 0.0030832644551992416, + 0.012225253507494926, + 0.004331383854150772, + 0.004771405830979347, + 0.009726090356707573, + 0.006719617173075676, + 0.017071271315217018, + 0.005249016918241978, + 0.0077107148244977, + 0.026655249297618866, + 0.0066849240101873875, + 0.0066823894158005714, + 0.01332323532551527, + 0.007296685129404068, + 0.00891178473830223, + 0.0061526247300207615, + 0.006217649206519127, + 0.021376991644501686, + 0.006164322141557932, + 0.007201599422842264, + 0.006992153823375702, + 0.004933411255478859, + 0.007337982300668955, + 0.0056349546648561954, + 0.00454078521579504, + 0.007255902048200369, + 0.005135863088071346, + 0.006831395905464888, + 0.00421402882784605, + 0.0020658927969634533, + 0.006508748047053814, + 0.0026650261133909225, + 0.002604016102850437, + 0.005178644321858883, + 0.0039804899133741856, + 0.005191422067582607, + 0.020515576004981995, + 0.008499739691615105, + 0.025625737383961678, + 0.007334919646382332, + 0.025265024974942207, + 0.031415749341249466, + 0.019305400550365448, + 0.026285018771886826, + 0.009531780146062374, + 0.0031575981993228197, + 0.005881975870579481, + 0.0031416991259902716, + 0.005585226695984602, + 0.004380325321108103, + 0.005133320111781359, + 0.0066082654520869255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.007628357969224453 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195652678608894 + ], + "opposite_logprob_tokens": [ + -4.264195442199707 + ], + "logprob_reward": -0.014195652678608894, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195652678608894, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18529921770095825, + 0.04446739703416824, + 0.04600859805941582, + 0.06971120834350586, + 0.07778611779212952, + 0.07818379253149033, + 0.060015011578798294, + 0.06983181834220886, + 0.2926100790500641, + 0.07608677446842194 + ], + "credit_map": [ + 0.035182416439056396, + 0.015366447158157825, + 0.023689011111855507, + 0.012496447190642357, + 0.017586631700396538, + 0.05364257097244263, + 0.01776895672082901, + 0.009566712193191051, + 0.00758568337187171, + 0.004381289705634117, + 0.005255591124296188, + 0.005065219942480326, + 0.004446563310921192, + 0.006332273129373789, + 0.005245450884103775, + 0.006155320443212986, + 0.006110711954534054, + 0.00431853299960494, + 0.0077819726429879665, + 0.005564270541071892, + 0.004315131343901157, + 0.009928460232913494, + 0.004859136417508125, + 0.0031303742434829473, + 0.0090177645906806, + 0.007647226098924875, + 0.010416124947369099, + 0.01200344692915678, + 0.005648597609251738, + 0.009519496001303196, + 0.009794331155717373, + 0.005664209835231304, + 0.010567067191004753, + 0.008294502273201942, + 0.019066883251070976, + 0.008457401767373085, + 0.005669554695487022, + 0.013003287836909294, + 0.005594433285295963, + 0.007132982835173607, + 0.01216891873627901, + 0.0070561901666224, + 0.01048485841602087, + 0.005407485645264387, + 0.006004609167575836, + 0.018358053639531136, + 0.009062198922038078, + 0.009641473181545734, + 0.008345973677933216, + 0.010647698305547237, + 0.010818547569215298, + 0.005600393749773502, + 0.00824011117219925, + 0.005977036897093058, + 0.005524730775505304, + 0.004860512912273407, + 0.01442932989448309, + 0.0074354237876832485, + 0.008838367648422718, + 0.0069024041295051575, + 0.005619523581117392, + 0.01543037686496973, + 0.0053167990408837795, + 0.005859590135514736, + 0.04096301272511482, + 0.010933377780020237, + 0.023755040019750595, + 0.0090797683224082, + 0.06889517605304718, + 0.09026755392551422, + 0.020025596022605896, + 0.028690533712506294, + 0.011966531164944172, + 0.006157041061669588, + 0.01099864300340414, + 0.004892236553132534, + 0.01108898688107729, + 0.009324443526566029, + 0.009120041504502296, + 0.012538839131593704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.014195652678608894 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic drone shot follows closely, following an Asian man in a crisp, dark charcoal-gray suit as he glides smoothly through the vast air between two skyscrapers with a grace that defies the laws of physics. Slightly above his horizon, the camera captures his elongated, steady form, as if he were stepping on invisible air, his movements reminiscent of a martial arts master. Beneath his impeccably tailored suit jacket lies a crisp white shirt and a narrow, dark silk tie that billows backwards in the wind, its hems undulating but never a touch disheveled. His face is calm and focused, his gaze fixed firmly on the platform of the next building ahead. The fading light of the evening sun casts a sharp golden outline on his dark suit, casting a shifting, warm glow across the surrounding glass walls. Below is a bustling city canyon. The lights of countless vehicles have merged into a flowing river of light. The sky is dyed a gradient of warm orange-red and deep purple. Thin clouds and mist are lingering between the buildings. The whole picture combines the elegance of martial arts, the grandeur of surrealism and the coldness of the modern city, creating a breathtaking visual spectacle.", + "seed": 30077, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_12/index_3944_seed_30077_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.035030425526201725, + "vlm_logprob": -0.035030425526201725, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.035030425526201725, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07357057183980942 + ], + "opposite_logprob_tokens": [ + -2.698570489883423 + ], + "logprob_reward": -0.07357057183980942, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07357057183980942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23232409358024597, + 0.049302347004413605, + 0.05752583220601082, + 0.09017045795917511, + 0.08309967815876007, + 0.09818451851606369, + 0.0935852974653244, + 0.06375396996736526, + 0.08471548557281494, + 0.14733830094337463 + ], + "credit_map": [ + 0.0250311940908432, + 0.01970161311328411, + 0.05044715851545334, + 0.030258530750870705, + 0.01527341827750206, + 0.017229510471224785, + 0.031117282807826996, + 0.04326540604233742, + 0.004600056447088718, + 0.007777300663292408, + 0.007844366133213043, + 0.009648220613598824, + 0.004425369668751955, + 0.0033289887942373753, + 0.006950695998966694, + 0.004727353807538748, + 0.010261459276080132, + 0.005343044642359018, + 0.008628224022686481, + 0.009210740216076374, + 0.006436225026845932, + 0.006787327118217945, + 0.005800517741590738, + 0.0050583030097186565, + 0.008586334995925426, + 0.006794223096221685, + 0.020686935633420944, + 0.01266444567590952, + 0.005829810630530119, + 0.01343595888465643, + 0.009372580796480179, + 0.01280017290264368, + 0.007264554500579834, + 0.005468348041176796, + 0.013133500702679157, + 0.011957557871937752, + 0.0062813046388328075, + 0.006182602606713772, + 0.020121660083532333, + 0.012690162286162376, + 0.00695143174380064, + 0.006600810214877129, + 0.01637766882777214, + 0.01213981956243515, + 0.015216201543807983, + 0.008962997235357761, + 0.021130219101905823, + 0.010805382393300533, + 0.006684323772788048, + 0.008458808064460754, + 0.018466392531991005, + 0.011928475461900234, + 0.00821029394865036, + 0.01598474383354187, + 0.011821984313428402, + 0.012030293233692646, + 0.010240260511636734, + 0.007961281575262547, + 0.009850642643868923, + 0.006417320109903812, + 0.0076767392456531525, + 0.009752742946147919, + 0.006108436267822981, + 0.0057465448044240475, + 0.012363455258309841, + 0.01012750156223774, + 0.007948864251375198, + 0.009076925925910473, + 0.010278250090777874, + 0.008036474697291851, + 0.01536584086716175, + 0.011518184095621109, + 0.010432973504066467, + 0.007096289657056332, + 0.01934441551566124, + 0.017613714560866356, + 0.00889370683580637, + 0.007093651220202446, + 0.05366705358028412, + 0.023196518421173096 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.07357057183980942 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813666105270386 + ], + "opposite_logprob_tokens": [ + -3.2881367206573486 + ], + "logprob_reward": -0.03813666105270386, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813666105270386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07391685992479324, + 0.04595515877008438, + 0.03818666934967041, + 0.08594607561826706, + 0.13921846449375153, + 0.08940207213163376, + 0.10814619809389114, + 0.13026434183120728, + 0.22089633345603943, + 0.06806782633066177 + ], + "credit_map": [ + 0.009240684099495411, + 0.005488166119903326, + 0.009948444552719593, + 0.014966162852942944, + 0.006646319758147001, + 0.004676852375268936, + 0.01227602269500494, + 0.01067421119660139, + 0.0061630974523723125, + 0.00556159159168601, + 0.006950203329324722, + 0.00882758293300867, + 0.003928452264517546, + 0.0032351824920624495, + 0.006534461863338947, + 0.004754586145281792, + 0.0036504913587123156, + 0.0029102966655045748, + 0.004998604767024517, + 0.004759809002280235, + 0.005349739920347929, + 0.004812128841876984, + 0.0053897094912827015, + 0.006315890233963728, + 0.01054416410624981, + 0.007277590222656727, + 0.016060873866081238, + 0.015378241427242756, + 0.006546419579535723, + 0.008147059939801693, + 0.011327396146953106, + 0.010664327070116997, + 0.012238884344696999, + 0.01077849417924881, + 0.02088485099375248, + 0.021930310875177383, + 0.01679147779941559, + 0.014250407926738262, + 0.02392725646495819, + 0.018416786566376686, + 0.007120230235159397, + 0.008331398479640484, + 0.021402975544333458, + 0.011577129364013672, + 0.006311768665909767, + 0.011365961283445358, + 0.01740036904811859, + 0.005892233923077583, + 0.009511265903711319, + 0.011335139162838459, + 0.014889380894601345, + 0.018237318843603134, + 0.010167970322072506, + 0.010567022487521172, + 0.02037259377539158, + 0.013065505772829056, + 0.021145137026906013, + 0.014882946386933327, + 0.014453519135713577, + 0.013577842153608799, + 0.02023801952600479, + 0.01697872206568718, + 0.014062724076211452, + 0.014925436116755009, + 0.03335709497332573, + 0.019174687564373016, + 0.025077704340219498, + 0.022627824917435646, + 0.023574184626340866, + 0.016777144744992256, + 0.049065712839365005, + 0.031241990625858307, + 0.0067391968332231045, + 0.005996052175760269, + 0.01079957839101553, + 0.008848327212035656, + 0.003948161844164133, + 0.006332414224743843, + 0.016708852723240852, + 0.008695241995155811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.03813666105270386 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018189528957009315 + ], + "opposite_logprob_tokens": [ + -4.018189430236816 + ], + "logprob_reward": -0.018189528957009315, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018189528957009315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14319899678230286, + 0.04903578385710716, + 0.08858592808246613, + 0.11832977831363678, + 0.13191810250282288, + 0.10910631716251373, + 0.05529322847723961, + 0.14466693997383118, + 0.0863075703382492, + 0.07355741411447525 + ], + "credit_map": [ + 0.01776033639907837, + 0.010209480300545692, + 0.028681296855211258, + 0.026734545826911926, + 0.014832560904324055, + 0.010683576576411724, + 0.015810806304216385, + 0.01848638616502285, + 0.005944989621639252, + 0.010234025306999683, + 0.007488680072128773, + 0.007982254959642887, + 0.0051956092938780785, + 0.0027181762270629406, + 0.0058072712272405624, + 0.0036647780798375607, + 0.009786010719835758, + 0.007812543772161007, + 0.013125771656632423, + 0.01995571330189705, + 0.009778960607945919, + 0.008852213621139526, + 0.011954771354794502, + 0.007319941185414791, + 0.009993184357881546, + 0.007684285286813974, + 0.014384553767740726, + 0.0235697403550148, + 0.008998853154480457, + 0.01234640646725893, + 0.021373923867940903, + 0.019978830590844154, + 0.009205794893205166, + 0.010851647704839706, + 0.01620899885892868, + 0.021105291321873665, + 0.013924263417720795, + 0.01754046604037285, + 0.025825241580605507, + 0.017256390303373337, + 0.011131472885608673, + 0.011249479837715626, + 0.0274003017693758, + 0.012475750409066677, + 0.007049313746392727, + 0.012547088786959648, + 0.01916644163429737, + 0.008086469024419785, + 0.005648717749863863, + 0.005417957436293364, + 0.009162147529423237, + 0.006547568831592798, + 0.005261401645839214, + 0.009353565983474255, + 0.006101037375628948, + 0.007800830993801355, + 0.023798353970050812, + 0.015616022050380707, + 0.01679631881415844, + 0.016149725764989853, + 0.021317949518561363, + 0.01680980622768402, + 0.015839211642742157, + 0.018339555710554123, + 0.011175005696713924, + 0.014750699512660503, + 0.00869572814553976, + 0.011005708947777748, + 0.00808682106435299, + 0.01186957024037838, + 0.013627300038933754, + 0.00709674134850502, + 0.013478773646056652, + 0.007615814916789532, + 0.010909037664532661, + 0.009467601776123047, + 0.005011560395359993, + 0.011572843417525291, + 0.006384463980793953, + 0.009117317385971546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.018189528957009315 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021968772634863853 + ], + "opposite_logprob_tokens": [ + -6.127196788787842 + ], + "logprob_reward": -0.0021968772634863853, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021968772634863853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08855970948934555, + 0.08110011368989944, + 0.04666462540626526, + 0.24537010490894318, + 0.08153057843446732, + 0.050213154405355453, + 0.040353961288928986, + 0.06429000198841095, + 0.06925085932016373, + 0.23266679048538208 + ], + "credit_map": [ + 0.015940768644213676, + 0.008853008970618248, + 0.020846541970968246, + 0.009624707512557507, + 0.008808793500065804, + 0.006387816742062569, + 0.009350895881652832, + 0.008747185580432415, + 0.010465244762599468, + 0.013397164642810822, + 0.009662330150604248, + 0.012780264019966125, + 0.010229486972093582, + 0.009661589749157429, + 0.008620164357125759, + 0.0062838769517838955, + 0.005595832131803036, + 0.003968291450291872, + 0.006213882006704807, + 0.007471924182027578, + 0.007533184252679348, + 0.005412862170487642, + 0.004654790740460157, + 0.0058138612657785416, + 0.026970835402607918, + 0.018414171412587166, + 0.06866323202848434, + 0.03944722190499306, + 0.010709310881793499, + 0.03166128322482109, + 0.02267463132739067, + 0.02682945877313614, + 0.008725747466087341, + 0.01151219755411148, + 0.009628417901694775, + 0.010939712636172771, + 0.008426844142377377, + 0.010745598934590816, + 0.012089263647794724, + 0.00946281012147665, + 0.0039856587536633015, + 0.0047721778973937035, + 0.013354240916669369, + 0.00647379457950592, + 0.003227459965273738, + 0.005106319207698107, + 0.009785720147192478, + 0.003507788060232997, + 0.004020615480840206, + 0.004375233314931393, + 0.008715485222637653, + 0.004449971951544285, + 0.0028577083721756935, + 0.007322484627366066, + 0.003482918720692396, + 0.005129550117999315, + 0.009741929359734058, + 0.01124629657715559, + 0.0076704914681613445, + 0.006397459656000137, + 0.006961192935705185, + 0.007522917352616787, + 0.006067141890525818, + 0.008682580664753914, + 0.008821775205433369, + 0.010751912370324135, + 0.006674090400338173, + 0.011036988347768784, + 0.00554400822147727, + 0.009399275295436382, + 0.010479552671313286, + 0.006543260999023914, + 0.06640142947435379, + 0.016536055132746696, + 0.02746042050421238, + 0.019907141104340553, + 0.010863295756280422, + 0.0324859544634819, + 0.027366498485207558, + 0.031646016985177994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.0021968772634863853 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305848851799965 + ], + "opposite_logprob_tokens": [ + -3.168058395385742 + ], + "logprob_reward": -0.04305848851799965, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305848851799965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0666196271777153, + 0.04954381659626961, + 0.04150385037064552, + 0.10101858526468277, + 0.13128003478050232, + 0.07782347500324249, + 0.10860767960548401, + 0.1292438507080078, + 0.15951132774353027, + 0.13484777510166168 + ], + "credit_map": [ + 0.008774752728641033, + 0.007489562500268221, + 0.010322528891265392, + 0.012312102131545544, + 0.005121739115566015, + 0.005680879577994347, + 0.010470773093402386, + 0.006447285413742065, + 0.004872384015470743, + 0.006545146461576223, + 0.007267195265740156, + 0.008600539527833462, + 0.004877013154327869, + 0.0033948062919080257, + 0.008245681412518024, + 0.005741048138588667, + 0.007230738643556833, + 0.003996389918029308, + 0.006505285389721394, + 0.005909134168177843, + 0.00499157328158617, + 0.00442881416529417, + 0.004235365893691778, + 0.004206546116620302, + 0.009366698563098907, + 0.0073961177840828896, + 0.017069801688194275, + 0.015095698647201061, + 0.006375395692884922, + 0.015886710956692696, + 0.013048725202679634, + 0.016779426485300064, + 0.01075836829841137, + 0.011737321503460407, + 0.016686543822288513, + 0.019985657185316086, + 0.014612337574362755, + 0.01649555191397667, + 0.02388438582420349, + 0.017119867727160454, + 0.0077452706173062325, + 0.008747913874685764, + 0.011654696427285671, + 0.013182492926716805, + 0.008546174503862858, + 0.010437422432005405, + 0.011681001633405685, + 0.005828497931361198, + 0.007098125293850899, + 0.01039628405123949, + 0.014163964428007603, + 0.018212106078863144, + 0.011256212368607521, + 0.012851307168602943, + 0.023011159151792526, + 0.011618517339229584, + 0.022372422739863396, + 0.017763378098607063, + 0.01633954979479313, + 0.009711254388093948, + 0.012615171261131763, + 0.019235875457525253, + 0.015972917899489403, + 0.015233286656439304, + 0.02416878007352352, + 0.023708408698439598, + 0.013211438432335854, + 0.01964123547077179, + 0.01784350909292698, + 0.020844589918851852, + 0.020330989733338356, + 0.01976238191127777, + 0.022952280938625336, + 0.006781038828194141, + 0.020007196813821793, + 0.015813078731298447, + 0.0073887137696146965, + 0.018091943114995956, + 0.026869818568229675, + 0.016943691298365593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.04305848851799965 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic drone shot follows closely, following an Asian man in a crisp, dark charcoal-gray suit as he glides smoothly through the vast air between two skyscrapers with a grace that defies the laws of physics. Slightly above his horizon, the camera captures his elongated, steady form, as if he were stepping on invisible air, his movements reminiscent of a martial arts master. Beneath his impeccably tailored suit jacket lies a crisp white shirt and a narrow, dark silk tie that billows backwards in the wind, its hems undulating but never a touch disheveled. His face is calm and focused, his gaze fixed firmly on the platform of the next building ahead. The fading light of the evening sun casts a sharp golden outline on his dark suit, casting a shifting, warm glow across the surrounding glass walls. Below is a bustling city canyon. The lights of countless vehicles have merged into a flowing river of light. The sky is dyed a gradient of warm orange-red and deep purple. Thin clouds and mist are lingering between the buildings. The whole picture combines the elegance of martial arts, the grandeur of surrealism and the coldness of the modern city, creating a breathtaking visual spectacle.", + "seed": 30078, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_12/index_3944_seed_30078_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05451295431703329, + "vlm_logprob": -0.05451295431703329, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.05451295431703329, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12940610945224762 + ], + "opposite_logprob_tokens": [ + -2.129406213760376 + ], + "logprob_reward": -0.12940610945224762, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12940610945224762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39451533555984497, + 0.11408180743455887, + 0.1894530951976776, + 0.025648722425103188, + 0.0238886047154665, + 0.03267616406083107, + 0.014887687750160694, + 0.049105413258075714, + 0.08181633055210114, + 0.07392686605453491 + ], + "credit_map": [ + 0.025660192593932152, + 0.14352720975875854, + 0.020238233730196953, + 0.019611908122897148, + 0.017247185111045837, + 0.13688984513282776, + 0.01371346227824688, + 0.017627330496907234, + 0.016328074038028717, + 0.026411985978484154, + 0.020451245829463005, + 0.01568986475467682, + 0.006275124382227659, + 0.00979224406182766, + 0.006121539045125246, + 0.013011723756790161, + 0.04146447405219078, + 0.042297977954149246, + 0.01873316429555416, + 0.011650573462247849, + 0.009982113726437092, + 0.03601609170436859, + 0.013728685677051544, + 0.015580005943775177, + 0.0040154410526156425, + 0.004998424556106329, + 0.0035269514191895723, + 0.0018760283710435033, + 0.003419410903006792, + 0.0016514305025339127, + 0.003924807533621788, + 0.0022362275049090385, + 0.002209289465099573, + 0.0019994035828858614, + 0.003655014792457223, + 0.0025779891293495893, + 0.0040845307521522045, + 0.002455072943121195, + 0.003916444722563028, + 0.00299085839651525, + 0.0020043428521603346, + 0.002148132538422942, + 0.003560007782652974, + 0.003443940309807658, + 0.007030264474451542, + 0.004514435771852732, + 0.00504498602822423, + 0.004930053371936083, + 0.0014307510573416948, + 0.0016253317007794976, + 0.0017679790034890175, + 0.001394503633491695, + 0.0025055299047380686, + 0.0019773722160607576, + 0.0020844077225774527, + 0.0021018146071583033, + 0.006431390531361103, + 0.003639612812548876, + 0.006697553209960461, + 0.002491293242201209, + 0.003096409374848008, + 0.007258193101733923, + 0.00422936724498868, + 0.015261590480804443, + 0.006384377367794514, + 0.010809056460857391, + 0.014765025116503239, + 0.014485912397503853, + 0.00833768118172884, + 0.009568901732563972, + 0.007253161631524563, + 0.010212215594947338, + 0.006598439067602158, + 0.011925414204597473, + 0.0069071706384420395, + 0.007967526093125343, + 0.009978308342397213, + 0.012705465778708458, + 0.011805353686213493, + 0.006039188243448734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.12940610945224762 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907795697450638 + ], + "opposite_logprob_tokens": [ + -2.464077949523926 + ], + "logprob_reward": -0.08907795697450638, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907795697450638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18558812141418457, + 0.19820648431777954, + 0.10451100021600723, + 0.022740760818123817, + 0.035809390246868134, + 0.019677726551890373, + 0.04476785287261009, + 0.05287487804889679, + 0.2130037248134613, + 0.12282012403011322 + ], + "credit_map": [ + 0.012474622577428818, + 0.06497388333082199, + 0.01113842148333788, + 0.010148459114134312, + 0.009859672747552395, + 0.06139469891786575, + 0.007436653133481741, + 0.008161713369190693, + 0.03217747062444687, + 0.06230872869491577, + 0.031637296080589294, + 0.01666918583214283, + 0.006756471004337072, + 0.025084450840950012, + 0.004564141388982534, + 0.019008727744221687, + 0.021886134520173073, + 0.022404465824365616, + 0.01021735928952694, + 0.006517603062093258, + 0.006587005220353603, + 0.01941189542412758, + 0.008851458318531513, + 0.008635071106255054, + 0.0024641586933285, + 0.0027421533595770597, + 0.0031292526982724667, + 0.00203161989338696, + 0.0037310675252228975, + 0.0026705237105488777, + 0.0033928658813238144, + 0.002579116029664874, + 0.004807927180081606, + 0.004019497893750668, + 0.007197787519544363, + 0.003351386869326234, + 0.002991759218275547, + 0.004085017368197441, + 0.003507767105475068, + 0.00584824476391077, + 0.0023065570276230574, + 0.0024955293629318476, + 0.0024160805623978376, + 0.0021522454917430878, + 0.001908562728203833, + 0.00287227937951684, + 0.0025500536430627108, + 0.0029764159116894007, + 0.004230158403515816, + 0.0035943659022450447, + 0.0065970709547400475, + 0.002830138197168708, + 0.0035838503390550613, + 0.005179956555366516, + 0.004577228799462318, + 0.014175078831613064, + 0.00459520798176527, + 0.0041655623354017735, + 0.0053066350519657135, + 0.00461244024336338, + 0.008481927216053009, + 0.006169866304844618, + 0.009513363242149353, + 0.010029871948063374, + 0.0232857596129179, + 0.02039553038775921, + 0.03274225443601608, + 0.02158619649708271, + 0.024071549996733665, + 0.037480589002370834, + 0.023102084174752235, + 0.030339734628796577, + 0.01798180863261223, + 0.01987096108496189, + 0.025418389588594437, + 0.011413752101361752, + 0.010520555078983307, + 0.011821513064205647, + 0.013421669602394104, + 0.012371474876999855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.08907795697450638 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033699944615364075 + ], + "opposite_logprob_tokens": [ + -3.4086999893188477 + ], + "logprob_reward": -0.033699944615364075, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033699944615364075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27891993522644043, + 0.08656248450279236, + 0.11325400322675705, + 0.035400938242673874, + 0.024158526211977005, + 0.021909553557634354, + 0.09974343329668045, + 0.20189140737056732, + 0.10563312470912933, + 0.0325266569852829 + ], + "credit_map": [ + 0.018934784457087517, + 0.1078580766916275, + 0.01138534490019083, + 0.00842328555881977, + 0.01040123775601387, + 0.10235784202814102, + 0.008413203060626984, + 0.011146164499223232, + 0.012492828071117401, + 0.025529753416776657, + 0.011928725056350231, + 0.008922228589653969, + 0.00377025012858212, + 0.012172919698059559, + 0.003912729676812887, + 0.007833040319383144, + 0.023425254970788956, + 0.022944392636418343, + 0.011660108342766762, + 0.006752655375748873, + 0.006347726099193096, + 0.021568989381194115, + 0.010252363979816437, + 0.010302508249878883, + 0.006436501629650593, + 0.006248071789741516, + 0.002917500678449869, + 0.0026489647570997477, + 0.004031634423881769, + 0.0032627924811095, + 0.006415816023945808, + 0.0034396578557789326, + 0.003417643718421459, + 0.002461905125528574, + 0.003034314839169383, + 0.0022798050194978714, + 0.003245240543037653, + 0.0031247935257852077, + 0.004204526077955961, + 0.0023902980610728264, + 0.0018898365087807178, + 0.0022117875050753355, + 0.0030119356233626604, + 0.003719822969287634, + 0.0022247706074267626, + 0.0032611372880637646, + 0.0027651817072182894, + 0.0028250827454030514, + 0.008885820396244526, + 0.006652167532593012, + 0.01549601275473833, + 0.007155288476496935, + 0.009839541278779507, + 0.014293385669589043, + 0.008435126394033432, + 0.028986087068915367, + 0.016673700883984566, + 0.011266169138252735, + 0.022529197856783867, + 0.015755362808704376, + 0.020588833838701248, + 0.04747665673494339, + 0.021229123696684837, + 0.046372365206480026, + 0.012281091883778572, + 0.00987769104540348, + 0.013192341662943363, + 0.014722013846039772, + 0.011114832945168018, + 0.019377851858735085, + 0.012343090958893299, + 0.012724212370812893, + 0.004273887258023024, + 0.004223411902785301, + 0.0033128911163657904, + 0.0037291061598807573, + 0.004516736138612032, + 0.003745113266631961, + 0.005114805418998003, + 0.003610704792663455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.033699944615364075 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002195212058722973 + ], + "opposite_logprob_tokens": [ + -6.127195358276367 + ], + "logprob_reward": -0.002195212058722973, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002195212058722973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11746463924646378, + 0.15677303075790405, + 0.2439294308423996, + 0.020221179351210594, + 0.07360901683568954, + 0.03885749727487564, + 0.08685021102428436, + 0.1657426804304123, + 0.04309132322669029, + 0.05346103385090828 + ], + "credit_map": [ + 0.008004128001630306, + 0.02469779923558235, + 0.013485322706401348, + 0.013083415105938911, + 0.010280290618538857, + 0.030799562111496925, + 0.008390283212065697, + 0.008723829872906208, + 0.02659202180802822, + 0.04839375242590904, + 0.02850799635052681, + 0.009345753118395805, + 0.006023375317454338, + 0.018656153231859207, + 0.004131706897169352, + 0.015122251585125923, + 0.05076824501156807, + 0.06728924810886383, + 0.025360891595482826, + 0.015523560345172882, + 0.011187892407178879, + 0.042976781725883484, + 0.010833430103957653, + 0.019989363849163055, + 0.003173317527398467, + 0.003586966311559081, + 0.002177128568291664, + 0.0017421070951968431, + 0.0023934110067784786, + 0.002431144705042243, + 0.002712808083742857, + 0.002004296751692891, + 0.008545568212866783, + 0.007751288823783398, + 0.01818941719830036, + 0.009277479723095894, + 0.005980361253023148, + 0.007586752530187368, + 0.008886560797691345, + 0.007391588296741247, + 0.0037457006983458996, + 0.0032609584741294384, + 0.009229378774762154, + 0.004696996882557869, + 0.0026031150482594967, + 0.005682963412255049, + 0.0033401562832295895, + 0.006298228166997433, + 0.008678280748426914, + 0.007786646019667387, + 0.011402427218854427, + 0.009056505747139454, + 0.009475850500166416, + 0.013404467143118382, + 0.009390735067427158, + 0.017655296251177788, + 0.015341299585998058, + 0.010100790299475193, + 0.024044645950198174, + 0.013207383453845978, + 0.014525092206895351, + 0.04089231416583061, + 0.013830790296196938, + 0.03380037099123001, + 0.0034750725608319044, + 0.006196227390319109, + 0.007125514093786478, + 0.005379870999604464, + 0.004953683353960514, + 0.005276239477097988, + 0.005018586292862892, + 0.0056661320850253105, + 0.007645001634955406, + 0.009227481670677662, + 0.013384475372731686, + 0.005419041495770216, + 0.0029738289304077625, + 0.006108960602432489, + 0.0032837209291756153, + 0.005418520886451006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.002195212058722973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818554848432541 + ], + "opposite_logprob_tokens": [ + -4.018185615539551 + ], + "logprob_reward": -0.01818554848432541, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818554848432541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37341517210006714, + 0.14742513000965118, + 0.08285365253686905, + 0.014873399399220943, + 0.025363821536302567, + 0.012739138677716255, + 0.03447326645255089, + 0.05367138609290123, + 0.2049340456724167, + 0.05025096237659454 + ], + "credit_map": [ + 0.028439771384000778, + 0.14562322199344635, + 0.013266106136143208, + 0.013465658761560917, + 0.015164714306592941, + 0.1333692967891693, + 0.01120149064809084, + 0.012884900905191898, + 0.026323316618800163, + 0.050452832132577896, + 0.02031390182673931, + 0.008443351835012436, + 0.005038972478359938, + 0.021649284288287163, + 0.0035413748119026423, + 0.011662089265882969, + 0.01643562875688076, + 0.021138550713658333, + 0.007108472287654877, + 0.0036224001087248325, + 0.005713969003409147, + 0.016916424036026, + 0.005783114116638899, + 0.006135089788585901, + 0.001485271961428225, + 0.0015405506128445268, + 0.0018633910221979022, + 0.0011699385941028595, + 0.003086430486291647, + 0.001417064806446433, + 0.002884277841076255, + 0.001426474074833095, + 0.002016549464315176, + 0.0021297484636306763, + 0.0019362003076821566, + 0.0021325000561773777, + 0.0046053798869252205, + 0.0032941391691565514, + 0.004530034959316254, + 0.004719268996268511, + 0.0010508298873901367, + 0.0009937662398442626, + 0.002749449573457241, + 0.0015323967672884464, + 0.0015578590100631118, + 0.0018161743646487594, + 0.0012242242228239775, + 0.0018144381465390325, + 0.0026984596624970436, + 0.002321874490007758, + 0.005487933289259672, + 0.003135186154395342, + 0.004256127402186394, + 0.0047894436866045, + 0.0038493478205055, + 0.007934890687465668, + 0.003476299811154604, + 0.0039895265363156796, + 0.007960455492138863, + 0.004671513102948666, + 0.006616842932999134, + 0.014538908377289772, + 0.005191316362470388, + 0.007226521149277687, + 0.022398943081498146, + 0.02330010198056698, + 0.02898021787405014, + 0.019413437694311142, + 0.0247383713722229, + 0.03674893453717232, + 0.02144588902592659, + 0.02790815755724907, + 0.004650793503969908, + 0.006980241276323795, + 0.005220229271799326, + 0.006874249316751957, + 0.008519555442035198, + 0.0050432113930583, + 0.008568338118493557, + 0.004394345451146364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.01818554848432541 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic drone shot follows closely, following an Asian man in a crisp, dark charcoal-gray suit as he glides smoothly through the vast air between two skyscrapers with a grace that defies the laws of physics. Slightly above his horizon, the camera captures his elongated, steady form, as if he were stepping on invisible air, his movements reminiscent of a martial arts master. Beneath his impeccably tailored suit jacket lies a crisp white shirt and a narrow, dark silk tie that billows backwards in the wind, its hems undulating but never a touch disheveled. His face is calm and focused, his gaze fixed firmly on the platform of the next building ahead. The fading light of the evening sun casts a sharp golden outline on his dark suit, casting a shifting, warm glow across the surrounding glass walls. Below is a bustling city canyon. The lights of countless vehicles have merged into a flowing river of light. The sky is dyed a gradient of warm orange-red and deep purple. Thin clouds and mist are lingering between the buildings. The whole picture combines the elegance of martial arts, the grandeur of surrealism and the coldness of the modern city, creating a breathtaking visual spectacle.", + "seed": 30079, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_12/index_3944_seed_30079_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04496216818224639, + "vlm_logprob": -0.04496216818224639, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.04496216818224639, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09255245327949524 + ], + "opposite_logprob_tokens": [ + -2.467552423477173 + ], + "logprob_reward": -0.09255245327949524, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09255245327949524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38110822439193726, + 0.06808793544769287, + 0.08834639936685562, + 0.11389046162366867, + 0.07191254198551178, + 0.05250020697712898, + 0.06816685199737549, + 0.05311084911227226, + 0.04743111878633499, + 0.055445458739995956 + ], + "credit_map": [ + 0.13943356275558472, + 0.1393715888261795, + 0.01531151868402958, + 0.013540251180529594, + 0.019414668902754784, + 0.022432053461670876, + 0.011689617298543453, + 0.019914954900741577, + 0.00948107149451971, + 0.010044815018773079, + 0.005945783108472824, + 0.00860773678869009, + 0.009829448536038399, + 0.007990848273038864, + 0.009738458320498466, + 0.0064497715793550014, + 0.008986500091850758, + 0.00755435461178422, + 0.01023312658071518, + 0.010145370848476887, + 0.01107989251613617, + 0.008999145589768887, + 0.01497469563037157, + 0.01637331210076809, + 0.014690165407955647, + 0.010321051813662052, + 0.017155248671770096, + 0.01882511004805565, + 0.011292989365756512, + 0.012521086260676384, + 0.009971081279218197, + 0.01911371946334839, + 0.007806615438312292, + 0.0085190050303936, + 0.007342619821429253, + 0.008581237867474556, + 0.006956617813557386, + 0.00928027369081974, + 0.012250388041138649, + 0.011175778694450855, + 0.006061641499400139, + 0.009848413057625294, + 0.00573006272315979, + 0.007325495593249798, + 0.005374555476009846, + 0.007265808060765266, + 0.005032286513596773, + 0.005861942190676928, + 0.009463985450565815, + 0.016381939873099327, + 0.008053178898990154, + 0.01095620822161436, + 0.0036834629718214273, + 0.011563640087842941, + 0.004347267094999552, + 0.0037171656731516123, + 0.0066290125250816345, + 0.004964015446603298, + 0.008923763409256935, + 0.006962866988033056, + 0.007993663661181927, + 0.007719571236521006, + 0.005255585536360741, + 0.004662367515265942, + 0.00738936522975564, + 0.007850181311368942, + 0.005711320787668228, + 0.005609462969005108, + 0.004922147374600172, + 0.007472928613424301, + 0.0038892142474651337, + 0.004586493596434593, + 0.007358376402407885, + 0.009063554927706718, + 0.007660732604563236, + 0.005952209699898958, + 0.008609275333583355, + 0.006366044748574495, + 0.004858687985688448, + 0.005576575640588999 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.09255245327949524 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890750139951706 + ], + "opposite_logprob_tokens": [ + -2.4640750885009766 + ], + "logprob_reward": -0.0890750139951706, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890750139951706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060332104563713074, + 0.032050516456365585, + 0.08154745399951935, + 0.3388293981552124, + 0.07185475528240204, + 0.10841400921344757, + 0.14958782494068146, + 0.08876118063926697, + 0.01982330158352852, + 0.04879944398999214 + ], + "credit_map": [ + 0.006737705785781145, + 0.007125762291252613, + 0.004553704988211393, + 0.0053183166310191154, + 0.010584603995084763, + 0.009470436722040176, + 0.008140724152326584, + 0.008400848135352135, + 0.005748232360929251, + 0.005928867030888796, + 0.0025456405710428953, + 0.0064051803201437, + 0.0032903114333748817, + 0.002010826487094164, + 0.0026392266154289246, + 0.0034822323359549046, + 0.008225635625422001, + 0.0065051354467868805, + 0.009546572342514992, + 0.010145657695829868, + 0.010496986098587513, + 0.008027040399610996, + 0.012844488024711609, + 0.015755947679281235, + 0.030604397878050804, + 0.02441084384918213, + 0.0565672293305397, + 0.08219044655561447, + 0.022691091522574425, + 0.03759635239839554, + 0.027977166697382927, + 0.05679186061024666, + 0.011533369310200214, + 0.01281990110874176, + 0.009857832454144955, + 0.010298256762325764, + 0.004054057411849499, + 0.009194054640829563, + 0.005944672506302595, + 0.0081526143476367, + 0.013805018737912178, + 0.024141360074281693, + 0.013174287043511868, + 0.015612741932272911, + 0.007631095591932535, + 0.01636580377817154, + 0.006789850071072578, + 0.010893848724663258, + 0.025038547813892365, + 0.03260108828544617, + 0.017947202548384666, + 0.023221297189593315, + 0.0102410102263093, + 0.025171633809804916, + 0.006805758457630873, + 0.008561300113797188, + 0.011627829633653164, + 0.014384930022060871, + 0.017063336446881294, + 0.011538548395037651, + 0.006734773516654968, + 0.013126177713274956, + 0.006730433087795973, + 0.00755514670163393, + 0.0031627677381038666, + 0.0033966819755733013, + 0.002558849286288023, + 0.0019128145650029182, + 0.0019555564504116774, + 0.003253704169765115, + 0.001880088122561574, + 0.0017028398578986526, + 0.006972650997340679, + 0.007046941667795181, + 0.008403158746659756, + 0.006456017028540373, + 0.005140442401170731, + 0.005380287300795317, + 0.0037626547273248434, + 0.005637289024889469 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.0890750139951706 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818976178765297 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.01818976178765297, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818976178765297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1280180811882019, + 0.039604417979717255, + 0.08698549121618271, + 0.37592586874961853, + 0.04705454036593437, + 0.08123808354139328, + 0.0527862049639225, + 0.06672348827123642, + 0.050611864775419235, + 0.0710519403219223 + ], + "credit_map": [ + 0.03219924122095108, + 0.03385760262608528, + 0.010645922273397446, + 0.019094198942184448, + 0.007820554077625275, + 0.007800040300935507, + 0.006986080203205347, + 0.009614434093236923, + 0.005596539471298456, + 0.005945275537669659, + 0.004063326865434647, + 0.0047154841013252735, + 0.0044920435175299644, + 0.0049411458894610405, + 0.005401586648076773, + 0.004449015483260155, + 0.008695747703313828, + 0.006786313373595476, + 0.015303957276046276, + 0.009835943579673767, + 0.014258196577429771, + 0.008343837223947048, + 0.014393867924809456, + 0.009367627091705799, + 0.04224993288516998, + 0.030659537762403488, + 0.05919402837753296, + 0.07151594758033752, + 0.030853239819407463, + 0.04489341378211975, + 0.037630222737789154, + 0.05892954021692276, + 0.006902199704200029, + 0.006667276844382286, + 0.007387290243059397, + 0.004557726439088583, + 0.004569855518639088, + 0.007112571969628334, + 0.004695378243923187, + 0.005162241403013468, + 0.010577346198260784, + 0.018128959462046623, + 0.011237693019211292, + 0.00980391725897789, + 0.005973665043711662, + 0.012340409681200981, + 0.005555690266191959, + 0.0076204021461308, + 0.008192712441086769, + 0.00840253196656704, + 0.0063585140742361546, + 0.008614819496870041, + 0.00547774089500308, + 0.006639410275965929, + 0.00468138512223959, + 0.004419086501002312, + 0.011400410905480385, + 0.013146334327757359, + 0.010548613965511322, + 0.007660822477191687, + 0.004421103745698929, + 0.009453680366277695, + 0.004099143669009209, + 0.005993381142616272, + 0.008906325325369835, + 0.007798966020345688, + 0.005860555451363325, + 0.005989340133965015, + 0.005559083539992571, + 0.008048229850828648, + 0.004775022156536579, + 0.003674344625324011, + 0.0072740293107926846, + 0.015249112620949745, + 0.01150185614824295, + 0.008713589049875736, + 0.008008935488760471, + 0.007876453921198845, + 0.005816836841404438, + 0.006611129734665155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.01818976178765297 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001712642959319055 + ], + "opposite_logprob_tokens": [ + -6.376712799072266 + ], + "logprob_reward": -0.001712642959319055, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001712642959319055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.188106968998909, + 0.06957320868968964, + 0.09901905804872513, + 0.10109831392765045, + 0.16298268735408783, + 0.06910599023103714, + 0.14212504029273987, + 0.0832258090376854, + 0.04391322657465935, + 0.04084968939423561 + ], + "credit_map": [ + 0.06304328143596649, + 0.06186337023973465, + 0.01073975395411253, + 0.016983386129140854, + 0.009745918214321136, + 0.008531018160283566, + 0.005618573632091284, + 0.011581671424210072, + 0.008161736652255058, + 0.007602033670991659, + 0.011212881654500961, + 0.010921669192612171, + 0.009394261986017227, + 0.005956160370260477, + 0.008429931476712227, + 0.007894533686339855, + 0.015885869041085243, + 0.014540242962539196, + 0.022291874513030052, + 0.010097199119627476, + 0.006491560954600573, + 0.00954191479831934, + 0.006909333635121584, + 0.01326106395572424, + 0.015725618228316307, + 0.010235714726150036, + 0.014214254915714264, + 0.01112376619130373, + 0.011498779058456421, + 0.011170294135808945, + 0.012563338503241539, + 0.014566545374691486, + 0.028894612565636635, + 0.029809731990098953, + 0.021661942824721336, + 0.02443578653037548, + 0.009679487906396389, + 0.024612993001937866, + 0.00895751267671585, + 0.014930618926882744, + 0.008936564438045025, + 0.017780005931854248, + 0.007229000795632601, + 0.008706626482307911, + 0.006347157526761293, + 0.010256767272949219, + 0.004619274754077196, + 0.005230589769780636, + 0.020203161984682083, + 0.02542346715927124, + 0.02036454528570175, + 0.024684278294444084, + 0.011026752181351185, + 0.020877420902252197, + 0.009640263393521309, + 0.009905148297548294, + 0.009669691324234009, + 0.011223775334656239, + 0.015506151132285595, + 0.008095020428299904, + 0.010621974244713783, + 0.011357715353369713, + 0.006639378610998392, + 0.010112103074789047, + 0.006652404088526964, + 0.009176602587103844, + 0.005062985699623823, + 0.0043372055515646935, + 0.004199329297989607, + 0.0077484529465436935, + 0.002927898894995451, + 0.0038083477411419153, + 0.006969149224460125, + 0.005333771929144859, + 0.008066077716648579, + 0.004929875954985619, + 0.005036441143602133, + 0.003918183036148548, + 0.0037769072223454714, + 0.0028192840982228518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.001712642959319055 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an Asian man wearing a dark charcoal-gray suit with a white shirt and a dark silk tie?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man gliding smoothly through the air between two skyscrapers without visible support?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera positioned slightly above the man's horizon, capturing his elongated and steady form?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a city canyon with a sky showing a gradient of warm orange-red and deep purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a sharp golden outline on the man's suit caused by the fading evening sun?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280968889594078 + ], + "opposite_logprob_tokens": [ + -3.7732808589935303 + ], + "logprob_reward": -0.023280968889594078, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280968889594078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12099423259496689, + 0.1039728969335556, + 0.0551731213927269, + 0.13730670511722565, + 0.07238306850194931, + 0.12197894603013992, + 0.2034396082162857, + 0.08541375398635864, + 0.035765498876571655, + 0.06357219815254211 + ], + "credit_map": [ + 0.026213405653834343, + 0.026492996141314507, + 0.011649275198578835, + 0.017901409417390823, + 0.009594644419848919, + 0.006673657335340977, + 0.006693275645375252, + 0.01577557809650898, + 0.016699213534593582, + 0.01762530207633972, + 0.013666568323969841, + 0.015693292021751404, + 0.01201694831252098, + 0.009029693901538849, + 0.00937047228217125, + 0.009871409274637699, + 0.005100752227008343, + 0.0052055474370718, + 0.010474102571606636, + 0.008423921652138233, + 0.006784228608012199, + 0.005088974721729755, + 0.005518096033483744, + 0.0085774976760149, + 0.009500118903815746, + 0.010619946755468845, + 0.021301213651895523, + 0.036569103598594666, + 0.009082968346774578, + 0.014562947675585747, + 0.010716656222939491, + 0.024953749030828476, + 0.010227815248072147, + 0.012929034419357777, + 0.010543745011091232, + 0.006936328019946814, + 0.010409709066152573, + 0.008828436955809593, + 0.007035928312689066, + 0.005472078453749418, + 0.01156244520097971, + 0.02123388461768627, + 0.017997747287154198, + 0.014480906538665295, + 0.018339671194553375, + 0.015824900940060616, + 0.01175402570515871, + 0.010785361751914024, + 0.019435301423072815, + 0.02814233861863613, + 0.023997854441404343, + 0.02283610589802265, + 0.03653515502810478, + 0.027829574421048164, + 0.030209125950932503, + 0.01445415336638689, + 0.007815615274012089, + 0.009317411109805107, + 0.011717764660716057, + 0.00835807342082262, + 0.018046032637357712, + 0.011282152496278286, + 0.008482327684760094, + 0.010394379496574402, + 0.003508215770125389, + 0.004520015325397253, + 0.006408393383026123, + 0.0037476078141480684, + 0.005673480685800314, + 0.0052165924571454525, + 0.003405191469937563, + 0.0032860003411769867, + 0.008402564562857151, + 0.009573258459568024, + 0.009916662238538265, + 0.008351442404091358, + 0.009343109093606472, + 0.005945625249296427, + 0.006918451748788357, + 0.0051210843957960606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.023280968889594078 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_13/index_tensor([6245, 6245])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_13/index_tensor([6245, 6245])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b615047d89a2623cdb1227b1801071326a640e46 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_13/index_tensor([6245, 6245])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a close-up shot in stark black and white, the camera remains still, focused on a massive watermelon about to be split open. A sharp, high-contrast overhead light casts a dazzling metallic white on the broad blade of a heavy carving knife, while the dark outline of the handle almost blends into the pure black background. The alternating light and dark stripes on the watermelon's smooth skin transform into a vibrant gray and nearly black banding pattern against the stark monochrome. With a swift, decisive arc, the blade cuts down without hesitation, sinking deeply into the watermelon with a clean, sharp stroke, instantly bisecting it. In the moment of the split, countless drops of crystalline juice are violently hurled in all directions by the immense force of the impact, each drop captured against the pitch-black background like a tiny diamond bursting into flame. The once vibrant red, juicy flesh takes on a grainy, dark gray. The entire image is imbued with extreme dynamic tension and sculptural beauty, transforming a simple action into a powerful visual poem.", + "seed": 7333, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_13/index_6245_seed_7333_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.20795759898610414, + "vlm_logprob": -0.20795759898610414, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.20795759898610414, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8028125166893005 + ], + "opposite_logprob_tokens": [ + -0.6778125166893005 + ], + "logprob_reward": -0.8028125166893005, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.8028125166893005, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.049814507365226746, + 0.07916579395532608, + 0.281637966632843, + 0.21118472516536713, + 0.02556697279214859, + 0.1339316964149475, + 0.0383995957672596, + 0.05986994504928589, + 0.08929590880870819, + 0.03113287128508091 + ], + "credit_map": [ + 0.008454492315649986, + 0.006267551798373461, + 0.00646737776696682, + 0.010371520183980465, + 0.004697584081441164, + 0.004497324116528034, + 0.005901657044887543, + 0.0031569944694638252, + 0.012062828987836838, + 0.0046946145594120026, + 0.0065173315815627575, + 0.011844107881188393, + 0.006195402704179287, + 0.005711916834115982, + 0.025456102564930916, + 0.006683490704745054, + 0.09494110941886902, + 0.015714306384325027, + 0.012003169395029545, + 0.009668989107012749, + 0.013973522000014782, + 0.10288700461387634, + 0.019486641511321068, + 0.012963230721652508, + 0.029727717861533165, + 0.02105122245848179, + 0.01616082340478897, + 0.022324485704302788, + 0.03144485503435135, + 0.03442469239234924, + 0.03929266706109047, + 0.01675826869904995, + 0.003263626480475068, + 0.003209203714504838, + 0.004015210084617138, + 0.0022452655248343945, + 0.0029432903975248337, + 0.0037031625397503376, + 0.002121636178344488, + 0.004065577406436205, + 0.04629764333367348, + 0.00841915886849165, + 0.004321699496358633, + 0.0027926000766456127, + 0.00776311382651329, + 0.0550033338367939, + 0.002539928536862135, + 0.006794213782995939, + 0.007492201868444681, + 0.003615954425185919, + 0.003445013426244259, + 0.0028658215887844563, + 0.009610084816813469, + 0.003944484516978264, + 0.0034380254801362753, + 0.003988009411841631, + 0.007765783462673426, + 0.010042857378721237, + 0.005448529031127691, + 0.005035781767219305, + 0.009001943282783031, + 0.009638155810534954, + 0.006410128436982632, + 0.006526764947921038, + 0.01602991856634617, + 0.011960084550082684, + 0.0038215790409594774, + 0.004176996648311615, + 0.03214556351304054, + 0.011601613834500313, + 0.003622745629400015, + 0.005937401670962572, + 0.0024923160672187805, + 0.0050848727114498615, + 0.0036827526055276394, + 0.002027131849899888, + 0.0033462464343756437, + 0.004713745787739754, + 0.0030124594923108816, + 0.006773344706743956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 5 + ], + "selected_reward": -0.8028125166893005 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560566663742065 + ], + "opposite_logprob_tokens": [ + -1.6006057262420654 + ], + "logprob_reward": -0.22560566663742065, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560566663742065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05962540581822395, + 0.08417798578739166, + 0.061609525233507156, + 0.1094323992729187, + 0.05431678146123886, + 0.32391849160194397, + 0.04713165760040283, + 0.06515703350305557, + 0.11612670868635178, + 0.0785040408372879 + ], + "credit_map": [ + 0.009571325965225697, + 0.008439208380877972, + 0.005173955112695694, + 0.009011207148432732, + 0.005914671812206507, + 0.006403586361557245, + 0.007667074911296368, + 0.007444377988576889, + 0.007870357483625412, + 0.004255739971995354, + 0.02014322206377983, + 0.013866564258933067, + 0.0055613890290260315, + 0.00788580346852541, + 0.018166163936257362, + 0.006428748369216919, + 0.012467294000089169, + 0.004995169583708048, + 0.003790087066590786, + 0.007006118539720774, + 0.0057581751607358456, + 0.014148971997201443, + 0.00773128867149353, + 0.005712416954338551, + 0.016865653917193413, + 0.006902644876390696, + 0.006487577222287655, + 0.008288041688501835, + 0.01096328441053629, + 0.016830863431096077, + 0.03351035714149475, + 0.009583975188434124, + 0.008779407478868961, + 0.006655609235167503, + 0.006451200228184462, + 0.003757152706384659, + 0.006882634945213795, + 0.009446435607969761, + 0.00535876490175724, + 0.00698557635769248, + 0.12178509682416916, + 0.013907047919929028, + 0.008571214973926544, + 0.00617031566798687, + 0.023666968569159508, + 0.12791688740253448, + 0.006798508111387491, + 0.015102454461157322, + 0.008382735773921013, + 0.0036714153829962015, + 0.004924749489873648, + 0.003640329698100686, + 0.00868845172226429, + 0.008303129114210606, + 0.004287625662982464, + 0.005233220290392637, + 0.008149622939527035, + 0.012162496335804462, + 0.006338637787848711, + 0.00715105514973402, + 0.009221208281815052, + 0.008497356437146664, + 0.0037145595997571945, + 0.009922095574438572, + 0.01605999656021595, + 0.016984395682811737, + 0.0060837664641439915, + 0.006780151277780533, + 0.03565611690282822, + 0.015076936222612858, + 0.00591911980882287, + 0.013566230423748493, + 0.00789568666368723, + 0.013225547969341278, + 0.007405634503811598, + 0.004530760459601879, + 0.010551384650170803, + 0.015499305911362171, + 0.010302847251296043, + 0.00909287016838789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.22560566663742065 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024986020289361477 + ], + "opposite_logprob_tokens": [ + -6.002498626708984 + ], + "logprob_reward": -0.0024986020289361477, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024986020289361477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09057142585515976, + 0.1005365252494812, + 0.10260193794965744, + 0.4189806282520294, + 0.039160799235105515, + 0.1027521938085556, + 0.015253730118274689, + 0.03929172828793526, + 0.05640312656760216, + 0.034447889775037766 + ], + "credit_map": [ + 0.011872520670294762, + 0.013203060254454613, + 0.008884335868060589, + 0.02080625109374523, + 0.008100480772554874, + 0.00856137927621603, + 0.012321075424551964, + 0.006822321563959122, + 0.012766660191118717, + 0.010828687809407711, + 0.01288029458373785, + 0.01919442042708397, + 0.011801381595432758, + 0.012979282066226006, + 0.01154166366904974, + 0.008544141426682472, + 0.031520918011665344, + 0.007273880299180746, + 0.00538151478394866, + 0.005006751511245966, + 0.006147474516183138, + 0.034484345465898514, + 0.007815356366336346, + 0.004971700720489025, + 0.05918252095580101, + 0.04246668517589569, + 0.0403580442070961, + 0.05446281656622887, + 0.05742032080888748, + 0.07911691814661026, + 0.050331857055425644, + 0.03564147278666496, + 0.0065963682718575, + 0.004911643918603659, + 0.0030864470172673464, + 0.0025245756842195988, + 0.007269265595823526, + 0.008802690543234348, + 0.002423770958557725, + 0.0035460367798805237, + 0.034927550703287125, + 0.007326203864067793, + 0.0023970690090209246, + 0.002262506168335676, + 0.0062102703377604485, + 0.04268969967961311, + 0.0019647174049168825, + 0.004974179435521364, + 0.002602317137643695, + 0.0012901542941108346, + 0.0012937730643898249, + 0.0009811008349061012, + 0.0037462711334228516, + 0.002431289292871952, + 0.001224918058142066, + 0.0016839070012792945, + 0.004585438407957554, + 0.005092978477478027, + 0.003725714748725295, + 0.0027571129612624645, + 0.005944776348769665, + 0.005471451207995415, + 0.0034643802791833878, + 0.008249876089394093, + 0.009250306524336338, + 0.007402458228170872, + 0.0028238659724593163, + 0.003283552825450897, + 0.01994713954627514, + 0.006547919940203428, + 0.003195687662810087, + 0.0039521935395896435, + 0.0038909048307687044, + 0.005231830757111311, + 0.005967649165540934, + 0.0023158977273851633, + 0.003936907276511192, + 0.00457686698064208, + 0.0034110103733837605, + 0.005116820335388184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.0024986020289361477 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036193120758980513 + ], + "opposite_logprob_tokens": [ + -5.628619194030762 + ], + "logprob_reward": -0.0036193120758980513, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036193120758980513, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15893220901489258, + 0.04025815799832344, + 0.10981082916259766, + 0.04097578302025795, + 0.1385650932788849, + 0.10925731062889099, + 0.05020904168486595, + 0.18555480241775513, + 0.07597991824150085, + 0.09045683592557907 + ], + "credit_map": [ + 0.02305077761411667, + 0.01685364916920662, + 0.020255746319890022, + 0.03563406690955162, + 0.01632292941212654, + 0.012259039096534252, + 0.018939033150672913, + 0.01561698131263256, + 0.005145269446074963, + 0.0048898691311478615, + 0.005308798514306545, + 0.006205626297742128, + 0.004634863696992397, + 0.005969178397208452, + 0.003942734096199274, + 0.00416182167828083, + 0.03414260223507881, + 0.0065727136097848415, + 0.004966218024492264, + 0.004083506762981415, + 0.006434181705117226, + 0.03852091729640961, + 0.010223793797194958, + 0.004866898059844971, + 0.0066096968948841095, + 0.003447022056207061, + 0.0037075441796332598, + 0.0031566419638693333, + 0.003946445882320404, + 0.012881678529083729, + 0.004428785759955645, + 0.0027979703154414892, + 0.02222864143550396, + 0.019771479070186615, + 0.016992058604955673, + 0.012097199447453022, + 0.017293665558099747, + 0.024153321981430054, + 0.011837219819426537, + 0.014191524125635624, + 0.03543448448181152, + 0.005927359219640493, + 0.006672961171716452, + 0.004571473225951195, + 0.008226550184190273, + 0.038123466074466705, + 0.004421531222760677, + 0.005879497621208429, + 0.008875217288732529, + 0.0045651886612176895, + 0.0066956463269889355, + 0.004040989559143782, + 0.006678642705082893, + 0.007174050901085138, + 0.006555481813848019, + 0.005623828619718552, + 0.02256685309112072, + 0.04153329133987427, + 0.009424867108464241, + 0.021770257502794266, + 0.02430466189980507, + 0.022891340777277946, + 0.011671441607177258, + 0.031392090022563934, + 0.007887978106737137, + 0.009612650610506535, + 0.00460090721026063, + 0.0046880473382771015, + 0.023751650005578995, + 0.013645479455590248, + 0.004822034854441881, + 0.006971179042011499, + 0.01017744094133377, + 0.012771371752023697, + 0.015500068664550781, + 0.007264379411935806, + 0.01066549401730299, + 0.013379808515310287, + 0.008004275150597095, + 0.012693996541202068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.0036193120758980513 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251897498965263 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.005251897498965263, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251897498965263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09675230085849762, + 0.06804739683866501, + 0.19857138395309448, + 0.05193779990077019, + 0.09983891993761063, + 0.1858488768339157, + 0.03210277855396271, + 0.12892048060894012, + 0.05645424872636795, + 0.08152581006288528 + ], + "credit_map": [ + 0.009136727079749107, + 0.01193381566554308, + 0.012400646694004536, + 0.02522497996687889, + 0.00804352480918169, + 0.009110802784562111, + 0.012713728472590446, + 0.008188062347471714, + 0.004848436452448368, + 0.004993726499378681, + 0.013402543030679226, + 0.008155583404004574, + 0.006099604535847902, + 0.004496641922742128, + 0.021343765780329704, + 0.004707097075879574, + 0.06751570105552673, + 0.011840512044727802, + 0.008910020813345909, + 0.006894455291330814, + 0.010197934694588184, + 0.07188977301120758, + 0.012322749942541122, + 0.009000244550406933, + 0.007793806027621031, + 0.004203383810818195, + 0.0037845962215214968, + 0.005970392841845751, + 0.005514676682651043, + 0.01459440030157566, + 0.006260222289711237, + 0.003816321026533842, + 0.016928907483816147, + 0.011254302226006985, + 0.0075111109763383865, + 0.0050619738176465034, + 0.019485175609588623, + 0.02057829312980175, + 0.006182522512972355, + 0.012836630456149578, + 0.03866355121135712, + 0.029758768156170845, + 0.00615338608622551, + 0.0063316128216683865, + 0.019345642998814583, + 0.06907689571380615, + 0.006160605698823929, + 0.01035841554403305, + 0.00557026406750083, + 0.002199329901486635, + 0.003235810436308384, + 0.002168501727283001, + 0.0062661608681082726, + 0.004855600651353598, + 0.0025837398134171963, + 0.005223368760198355, + 0.015286893583834171, + 0.01801259070634842, + 0.005517676472663879, + 0.013049047440290451, + 0.02189156413078308, + 0.02003420703113079, + 0.007780741434544325, + 0.027347758412361145, + 0.006632229778915644, + 0.012645323760807514, + 0.004007156938314438, + 0.005245037376880646, + 0.007620942313224077, + 0.008291592821478844, + 0.004067485220730305, + 0.007944479584693909, + 0.009969168342649937, + 0.013293711468577385, + 0.01193646527826786, + 0.00506925443187356, + 0.009445925243198872, + 0.013455048203468323, + 0.006044964771717787, + 0.012311271391808987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 7 + ], + "selected_reward": -0.005251897498965263 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a close-up shot in stark black and white, the camera remains still, focused on a massive watermelon about to be split open. A sharp, high-contrast overhead light casts a dazzling metallic white on the broad blade of a heavy carving knife, while the dark outline of the handle almost blends into the pure black background. The alternating light and dark stripes on the watermelon's smooth skin transform into a vibrant gray and nearly black banding pattern against the stark monochrome. With a swift, decisive arc, the blade cuts down without hesitation, sinking deeply into the watermelon with a clean, sharp stroke, instantly bisecting it. In the moment of the split, countless drops of crystalline juice are violently hurled in all directions by the immense force of the impact, each drop captured against the pitch-black background like a tiny diamond bursting into flame. The once vibrant red, juicy flesh takes on a grainy, dark gray. The entire image is imbued with extreme dynamic tension and sculptural beauty, transforming a simple action into a powerful visual poem.", + "seed": 7334, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_13/index_6245_seed_7334_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07153650987893342, + "vlm_logprob": -0.07153650987893342, + "vlm_yes_no_margin": 4.25, + "vlm_token_credit": -0.07153650987893342, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3070891201496124 + ], + "opposite_logprob_tokens": [ + -1.55708909034729 + ], + "logprob_reward": -0.3070891201496124, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.3070891201496124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06530658900737762, + 0.0874788835644722, + 0.07331279665231705, + 0.043314166367053986, + 0.12811608612537384, + 0.1463455706834793, + 0.052660513669252396, + 0.10917999595403671, + 0.1954839527606964, + 0.09880142658948898 + ], + "credit_map": [ + 0.008536388166248798, + 0.006171198561787605, + 0.014035625383257866, + 0.007662140764296055, + 0.004080642480403185, + 0.005749580450356007, + 0.005598980933427811, + 0.013472030870616436, + 0.0149745624512434, + 0.008171230554580688, + 0.02310345135629177, + 0.008106370456516743, + 0.005860878620296717, + 0.009203379042446613, + 0.007745202165096998, + 0.010313799604773521, + 0.01120935007929802, + 0.008014237508177757, + 0.018359528854489326, + 0.007430787198245525, + 0.004078315105289221, + 0.005613427143543959, + 0.006017804611474276, + 0.01258935034275055, + 0.006603327114135027, + 0.003939791116863489, + 0.00591645622625947, + 0.00665963813662529, + 0.005060300696641207, + 0.0042577628046274185, + 0.0054729776456952095, + 0.005403916817158461, + 0.02184569463133812, + 0.012836663983762264, + 0.026444436982274055, + 0.01318077277392149, + 0.012236704118549824, + 0.013595320284366608, + 0.01122372504323721, + 0.01675277017056942, + 0.014682195149362087, + 0.015547228045761585, + 0.019904140383005142, + 0.021515261381864548, + 0.012511907145380974, + 0.017419682815670967, + 0.025422876700758934, + 0.019342267885804176, + 0.006141037680208683, + 0.005738799460232258, + 0.0067613814026117325, + 0.006917674094438553, + 0.003424703609198332, + 0.007991726510226727, + 0.007820188999176025, + 0.007865002378821373, + 0.013322202488780022, + 0.0137791084125638, + 0.019464120268821716, + 0.015235081315040588, + 0.011686677113175392, + 0.01140410266816616, + 0.011729667894542217, + 0.012559032998979092, + 0.037836916744709015, + 0.03269572928547859, + 0.01539267972111702, + 0.03388424590229988, + 0.037586845457553864, + 0.009307896718382835, + 0.012420013546943665, + 0.01635962724685669, + 0.011804345063865185, + 0.011414604261517525, + 0.007509003393352032, + 0.02001560851931572, + 0.01759110949933529, + 0.01001112163066864, + 0.01118552591651678, + 0.009270102716982365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.3070891201496124 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03815501928329468 + ], + "opposite_logprob_tokens": [ + -3.2881550788879395 + ], + "logprob_reward": -0.03815501928329468, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03815501928329468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19323354959487915, + 0.0781870111823082, + 0.10138170421123505, + 0.0461575984954834, + 0.15025044977664948, + 0.07283022254705429, + 0.09994541108608246, + 0.054840460419654846, + 0.10390887409448624, + 0.09926468878984451 + ], + "credit_map": [ + 0.025686396285891533, + 0.02494567260146141, + 0.0315667986869812, + 0.020557958632707596, + 0.011474515311419964, + 0.02399229258298874, + 0.022125108167529106, + 0.032884806394577026, + 0.012502865865826607, + 0.009026171639561653, + 0.012061994522809982, + 0.00975063443183899, + 0.008000296540558338, + 0.0074754757806658745, + 0.008646467700600624, + 0.010723101906478405, + 0.017781496047973633, + 0.011983837932348251, + 0.019902698695659637, + 0.011801610700786114, + 0.007062714081257582, + 0.009827508591115475, + 0.008748616091907024, + 0.014273233711719513, + 0.009591374546289444, + 0.0043863714672625065, + 0.00567634217441082, + 0.0048583257012069225, + 0.005029973573982716, + 0.004402743652462959, + 0.005059396382421255, + 0.007153074257075787, + 0.03433657065033913, + 0.012346861883997917, + 0.02810715325176716, + 0.017236948013305664, + 0.015950439497828484, + 0.013940908946096897, + 0.009974145330488682, + 0.018357424065470695, + 0.007925091311335564, + 0.009486878290772438, + 0.0070209624245762825, + 0.009626176208257675, + 0.00868170615285635, + 0.010772679932415485, + 0.008806640282273293, + 0.010510098189115524, + 0.014759484678506851, + 0.011835873126983643, + 0.011661054566502571, + 0.012836034409701824, + 0.00617827707901597, + 0.021567029878497124, + 0.010009405203163624, + 0.011098254472017288, + 0.0074666691944003105, + 0.0068702055141329765, + 0.011614852584898472, + 0.008199325762689114, + 0.006760112009942532, + 0.004383548628538847, + 0.0039251986891031265, + 0.005620546638965607, + 0.011370418593287468, + 0.016919197514653206, + 0.008407674729824066, + 0.021670615300536156, + 0.02142072096467018, + 0.004078860394656658, + 0.009238812141120434, + 0.010802581906318665, + 0.008965626358985901, + 0.012644617818295956, + 0.010945182293653488, + 0.0077211735770106316, + 0.010508929379284382, + 0.016752168536186218, + 0.010262127965688705, + 0.021464858204126358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.03815501928329468 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00250252615660429 + ], + "opposite_logprob_tokens": [ + -6.00250244140625 + ], + "logprob_reward": -0.00250252615660429, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.00250252615660429, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05878762900829315, + 0.09033002704381943, + 0.08977603167295456, + 0.08953455090522766, + 0.05959063768386841, + 0.10050574690103531, + 0.04995273798704147, + 0.10587466508150101, + 0.3136804401874542, + 0.0419674888253212 + ], + "credit_map": [ + 0.009789912961423397, + 0.00760055985301733, + 0.008784919045865536, + 0.007509379647672176, + 0.005280925892293453, + 0.004380934406071901, + 0.007191584911197424, + 0.008249413222074509, + 0.01589442417025566, + 0.008424361236393452, + 0.023884156718850136, + 0.008672462776303291, + 0.006284561939537525, + 0.0081416554749012, + 0.008194679394364357, + 0.010833724401891232, + 0.015149835497140884, + 0.009178943932056427, + 0.021258579567074776, + 0.009581468999385834, + 0.0054775080643594265, + 0.006363241001963615, + 0.00745230121538043, + 0.015314148738980293, + 0.010641243308782578, + 0.009816695936024189, + 0.009184529073536396, + 0.017436817288398743, + 0.009491982869803905, + 0.008292102254927158, + 0.01168659795075655, + 0.012984584085643291, + 0.012860141694545746, + 0.006052105221897364, + 0.00991628784686327, + 0.008401872590184212, + 0.004069134127348661, + 0.00553501769900322, + 0.005259580910205841, + 0.007496499456465244, + 0.011712929233908653, + 0.010098297148942947, + 0.01429370604455471, + 0.01801268942654133, + 0.008805293589830399, + 0.01175497006624937, + 0.012915494851768017, + 0.012912366539239883, + 0.00672384025529027, + 0.0048915920779109, + 0.0051782457157969475, + 0.0038719316944479942, + 0.002538428409025073, + 0.014131501317024231, + 0.005466115195304155, + 0.0071510858833789825, + 0.01347783301025629, + 0.019025005400180817, + 0.010332969017326832, + 0.009396219626069069, + 0.01252961065620184, + 0.009930713102221489, + 0.011312318034470081, + 0.019869999960064888, + 0.08003285527229309, + 0.05743349343538284, + 0.01934664510190487, + 0.04044925048947334, + 0.06725197285413742, + 0.013072549365460873, + 0.01675177551805973, + 0.019341904670000076, + 0.005024042446166277, + 0.005469141062349081, + 0.0028224075213074684, + 0.005292744841426611, + 0.008594151586294174, + 0.005581439472734928, + 0.004337338265031576, + 0.004846225026994944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.00250252615660429 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031987475231289864 + ], + "opposite_logprob_tokens": [ + -5.753198623657227 + ], + "logprob_reward": -0.0031987475231289864, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031987475231289864, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0702391117811203, + 0.06716855615377426, + 0.05677705258131027, + 0.03997548669576645, + 0.10451363027095795, + 0.20400495827198029, + 0.06523975729942322, + 0.06988947093486786, + 0.2339092642068863, + 0.0882827639579773 + ], + "credit_map": [ + 0.011317954398691654, + 0.010732939466834068, + 0.009198644198477268, + 0.010440574027597904, + 0.004361608996987343, + 0.00733811454847455, + 0.0075592235662043095, + 0.009290045127272606, + 0.013638665899634361, + 0.005612798035144806, + 0.02111007273197174, + 0.004677494987845421, + 0.0035943605471402407, + 0.006273768376559019, + 0.004239377100020647, + 0.008022013120353222, + 0.009466770105063915, + 0.006328060757368803, + 0.013311588205397129, + 0.00673655467107892, + 0.003595707006752491, + 0.0036053138319402933, + 0.004744863137602806, + 0.00898818951100111, + 0.004987273830920458, + 0.004386872984468937, + 0.003475972916930914, + 0.007414962165057659, + 0.004726742394268513, + 0.0034749815240502357, + 0.005543186794966459, + 0.005965490825474262, + 0.026080802083015442, + 0.008673863485455513, + 0.023742392659187317, + 0.00794169120490551, + 0.006921276915818453, + 0.008365928195416927, + 0.006968893576413393, + 0.015818780288100243, + 0.01742781698703766, + 0.033837780356407166, + 0.017148001119494438, + 0.0265829935669899, + 0.023937029764056206, + 0.023900754749774933, + 0.033392466604709625, + 0.027778107672929764, + 0.008253445848822594, + 0.010876375250518322, + 0.010660496540367603, + 0.008247362449765205, + 0.0043893130496144295, + 0.008956224657595158, + 0.005408586468547583, + 0.00844794511795044, + 0.010979373008012772, + 0.009001871570944786, + 0.008918830193579197, + 0.009921449236571789, + 0.006989010609686375, + 0.0058858972042799, + 0.007745708804577589, + 0.010447321459650993, + 0.06236109510064125, + 0.04405202344059944, + 0.010974768549203873, + 0.027804633602499962, + 0.05268557369709015, + 0.009929648600518703, + 0.012669840827584267, + 0.01343168132007122, + 0.011626563034951687, + 0.009951087646186352, + 0.007715960498899221, + 0.01051959116011858, + 0.01700974628329277, + 0.011783616617321968, + 0.009006980806589127, + 0.010669207200407982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.0031987475231289864 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737136282026768 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.006737136282026768, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737136282026768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041249096393585205, + 0.05875137448310852, + 0.05498106777667999, + 0.03802670165896416, + 0.05022139474749565, + 0.0838218480348587, + 0.04607711359858513, + 0.18523521721363068, + 0.31036096811294556, + 0.13127519190311432 + ], + "credit_map": [ + 0.006722048856317997, + 0.005553746595978737, + 0.007062773685902357, + 0.005461090710014105, + 0.0027422106359153986, + 0.004470545798540115, + 0.0034954636357724667, + 0.005741214379668236, + 0.012151854112744331, + 0.004536765161901712, + 0.018618253991007805, + 0.004582531284540892, + 0.0028635081835091114, + 0.005195367615669966, + 0.003672323888167739, + 0.0071307700127363205, + 0.00871089193969965, + 0.005783536471426487, + 0.012024017050862312, + 0.0066232215613126755, + 0.0033797291107475758, + 0.004290285985916853, + 0.005351576954126358, + 0.00881781056523323, + 0.004527028650045395, + 0.0038308785296976566, + 0.0033875606022775173, + 0.007913276553153992, + 0.004214955028146505, + 0.0030710843857377768, + 0.005501011852174997, + 0.005580907687544823, + 0.00819134060293436, + 0.006125297863036394, + 0.005787169095128775, + 0.006861528381705284, + 0.005246370565146208, + 0.006748387590050697, + 0.004622327629476786, + 0.006638974882662296, + 0.00926834624260664, + 0.009598664939403534, + 0.012516248971223831, + 0.014318286441266537, + 0.007118659093976021, + 0.009597714059054852, + 0.010737690143287182, + 0.010666241869330406, + 0.004943401552736759, + 0.007807118818163872, + 0.007226654794067144, + 0.0055209859274327755, + 0.003468241775408387, + 0.006956762168556452, + 0.004243492614477873, + 0.00591045618057251, + 0.025008348748087883, + 0.036041975021362305, + 0.01956936903297901, + 0.03469400480389595, + 0.018960848450660706, + 0.011047356761991978, + 0.015541866421699524, + 0.024371437728405, + 0.08097024261951447, + 0.058732785284519196, + 0.017179694026708603, + 0.04089611396193504, + 0.0681896060705185, + 0.011491064913570881, + 0.016375254839658737, + 0.016526207327842712, + 0.01716451346874237, + 0.014259726740419865, + 0.01590571179986, + 0.012596562504768372, + 0.019826805219054222, + 0.01621103100478649, + 0.013510895892977715, + 0.02179994247853756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.006737136282026768 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a close-up shot in stark black and white, the camera remains still, focused on a massive watermelon about to be split open. A sharp, high-contrast overhead light casts a dazzling metallic white on the broad blade of a heavy carving knife, while the dark outline of the handle almost blends into the pure black background. The alternating light and dark stripes on the watermelon's smooth skin transform into a vibrant gray and nearly black banding pattern against the stark monochrome. With a swift, decisive arc, the blade cuts down without hesitation, sinking deeply into the watermelon with a clean, sharp stroke, instantly bisecting it. In the moment of the split, countless drops of crystalline juice are violently hurled in all directions by the immense force of the impact, each drop captured against the pitch-black background like a tiny diamond bursting into flame. The once vibrant red, juicy flesh takes on a grainy, dark gray. The entire image is imbued with extreme dynamic tension and sculptural beauty, transforming a simple action into a powerful visual poem.", + "seed": 7335, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_13/index_6245_seed_7335_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2504496508743614, + "vlm_logprob": -0.2504496508743614, + "vlm_yes_no_margin": 3.200000011920929, + "vlm_token_credit": -0.2504496508743614, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16314974427223206 + ], + "opposite_logprob_tokens": [ + -2.163149833679199 + ], + "logprob_reward": -0.16314974427223206, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.16314974427223206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19984467327594757, + 0.08536940068006516, + 0.03435058146715164, + 0.052587248384952545, + 0.14728915691375732, + 0.1294349581003189, + 0.07702554762363434, + 0.06044905260205269, + 0.0903463289141655, + 0.12330304831266403 + ], + "credit_map": [ + 0.04139697179198265, + 0.007126964163035154, + 0.012127134017646313, + 0.012575635686516762, + 0.01046021282672882, + 0.07419992983341217, + 0.021891580894589424, + 0.020066240802407265, + 0.01302076131105423, + 0.007132226601243019, + 0.008733413182199001, + 0.00915051531046629, + 0.00744719710201025, + 0.008248433470726013, + 0.019059481099247932, + 0.012577368877828121, + 0.00631303247064352, + 0.0026344629004597664, + 0.0041709658689796925, + 0.002988374326378107, + 0.0025596830528229475, + 0.0037624880205839872, + 0.005830381065607071, + 0.0060911900363862514, + 0.008646918460726738, + 0.004033040255308151, + 0.004558844957500696, + 0.007181303109973669, + 0.0035708360373973846, + 0.006159520708024502, + 0.011930253356695175, + 0.006506531033664942, + 0.02359369397163391, + 0.012799681164324284, + 0.012964023277163506, + 0.0173055250197649, + 0.014419394545257092, + 0.022468602284789085, + 0.027215968817472458, + 0.016522277146577835, + 0.020635491237044334, + 0.011327194981276989, + 0.011735708452761173, + 0.012223273515701294, + 0.00918861199170351, + 0.01846904680132866, + 0.01698301173746586, + 0.02887262962758541, + 0.011522465385496616, + 0.006622715387493372, + 0.006895695813000202, + 0.00967028271406889, + 0.007807008922100067, + 0.011349123902618885, + 0.014044337905943394, + 0.009113914333283901, + 0.006435717921704054, + 0.00557137094438076, + 0.005382266361266375, + 0.007097228430211544, + 0.006072745658457279, + 0.012125114910304546, + 0.008170132525265217, + 0.009594476781785488, + 0.01183068286627531, + 0.010005046613514423, + 0.009959354996681213, + 0.01527627743780613, + 0.009074660018086433, + 0.009968397207558155, + 0.011320022866129875, + 0.012911887839436531, + 0.00810420885682106, + 0.012463928200304508, + 0.011566318571567535, + 0.01751568540930748, + 0.01548189576715231, + 0.018111469224095345, + 0.025197017937898636, + 0.014862527139484882 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.16314974427223206 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539560317993164 + ], + "opposite_logprob_tokens": [ + -0.4289560914039612 + ], + "logprob_reward": -1.0539560317993164, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.0539560317993164, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2306094765663147, + 0.19832377135753632, + 0.06053195148706436, + 0.027952125295996666, + 0.2065380960702896, + 0.03279556706547737, + 0.05270441249012947, + 0.04556930810213089, + 0.062386348843574524, + 0.08258894830942154 + ], + "credit_map": [ + 0.0403362438082695, + 0.009692452847957611, + 0.014462591148912907, + 0.021373672410845757, + 0.01252340991050005, + 0.06837230920791626, + 0.03825070708990097, + 0.025598088279366493, + 0.03195885568857193, + 0.009915112517774105, + 0.020909124985337257, + 0.025344757363200188, + 0.010849046520888805, + 0.015276946127414703, + 0.055327434092760086, + 0.028742477297782898, + 0.009435047395527363, + 0.0037300842814147472, + 0.005607278551906347, + 0.009316680021584034, + 0.003963529132306576, + 0.00547811621800065, + 0.016150251030921936, + 0.0068509639240801334, + 0.006054234690964222, + 0.00257475720718503, + 0.003185092471539974, + 0.0025944169610738754, + 0.0020142446737736464, + 0.0022925687953829765, + 0.00556983333081007, + 0.003666974836960435, + 0.04798252880573273, + 0.015725312754511833, + 0.01237721461802721, + 0.009005943313241005, + 0.02016252465546131, + 0.05631573498249054, + 0.018336355686187744, + 0.026632467284798622, + 0.005485353525727987, + 0.002935377648100257, + 0.0033460434060543776, + 0.004953504540026188, + 0.002424656879156828, + 0.004770477768033743, + 0.004442874807864428, + 0.004437274765223265, + 0.008730840869247913, + 0.0049775512889027596, + 0.006056601647287607, + 0.00804342981427908, + 0.00575917586684227, + 0.0065771727822721004, + 0.007354143541306257, + 0.005205495748668909, + 0.008205192163586617, + 0.004906278569251299, + 0.004375353921204805, + 0.006054799072444439, + 0.003855242859572172, + 0.006146291736513376, + 0.008049247786402702, + 0.003976898733526468, + 0.009364468976855278, + 0.005285177379846573, + 0.0070593166165053844, + 0.009827092289924622, + 0.0072863404639065266, + 0.008420972153544426, + 0.008974527008831501, + 0.006168446969240904, + 0.00680814404040575, + 0.010195225477218628, + 0.010760295204818249, + 0.012269399128854275, + 0.012259301729500294, + 0.009579573757946491, + 0.010642741806805134, + 0.010074259713292122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -1.0539560317993164 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028325694147497416 + ], + "opposite_logprob_tokens": [ + -5.877832412719727 + ], + "logprob_reward": -0.0028325694147497416, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028325694147497416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12151335179805756, + 0.08893387019634247, + 0.1647350937128067, + 0.08624555170536041, + 0.14520876109600067, + 0.02956080064177513, + 0.09542914479970932, + 0.04044438526034355, + 0.05070291832089424, + 0.17722609639167786 + ], + "credit_map": [ + 0.018823521211743355, + 0.008502309210598469, + 0.008876778185367584, + 0.014674338512122631, + 0.008665898814797401, + 0.026359742507338524, + 0.0215451717376709, + 0.01406558696180582, + 0.015551222488284111, + 0.005342421587556601, + 0.009894846007227898, + 0.009425561875104904, + 0.006125084590166807, + 0.006413809489458799, + 0.023091381415724754, + 0.013089552521705627, + 0.025929972529411316, + 0.00913365837186575, + 0.01366132590919733, + 0.027067333459854126, + 0.009991091676056385, + 0.014645143412053585, + 0.04575355723500252, + 0.01855301298201084, + 0.013312458992004395, + 0.0065917749889194965, + 0.009402661584317684, + 0.010853390209376812, + 0.005826852284371853, + 0.009152348153293133, + 0.02049523964524269, + 0.010610830970108509, + 0.030996227636933327, + 0.01060960628092289, + 0.011954709887504578, + 0.009697044268250465, + 0.015348266810178757, + 0.035744644701480865, + 0.018701588734984398, + 0.012156665325164795, + 0.0041104997508227825, + 0.0029627534095197916, + 0.0028589521534740925, + 0.003820476122200489, + 0.0026230188086628914, + 0.003856176510453224, + 0.004971674643456936, + 0.004357249941676855, + 0.014363909140229225, + 0.007329766638576984, + 0.009408628568053246, + 0.011464842595160007, + 0.010046567767858505, + 0.015170680359005928, + 0.013387143611907959, + 0.01425760518759489, + 0.005849272944033146, + 0.004521256312727928, + 0.004709426313638687, + 0.005860430654138327, + 0.004287058021873236, + 0.006101259030401707, + 0.005297048017382622, + 0.0038186318706721067, + 0.006255437154322863, + 0.0044643753208220005, + 0.005674340762197971, + 0.007307809311896563, + 0.0059729344211518764, + 0.005262965336441994, + 0.0077729905024170876, + 0.00799206830561161, + 0.01487918570637703, + 0.019244764000177383, + 0.024696219712495804, + 0.02190415933728218, + 0.03717804700136185, + 0.02206200361251831, + 0.021450357511639595, + 0.015811366960406303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.0028325694147497416 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956398788839579 + ], + "opposite_logprob_tokens": [ + -5.130956172943115 + ], + "logprob_reward": -0.005956398788839579, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956398788839579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12023669481277466, + 0.06136123836040497, + 0.03230932354927063, + 0.035948291420936584, + 0.17751623690128326, + 0.15111571550369263, + 0.05719700828194618, + 0.03426828980445862, + 0.18326538801193237, + 0.14678184688091278 + ], + "credit_map": [ + 0.02142402157187462, + 0.005234378855675459, + 0.01008063368499279, + 0.00861729308962822, + 0.007389849983155727, + 0.027864260599017143, + 0.018823526799678802, + 0.02080271765589714, + 0.009673494845628738, + 0.005041323136538267, + 0.006179135292768478, + 0.0064368960447609425, + 0.0054609160870313644, + 0.0062363711185753345, + 0.013079079799354076, + 0.009254024364054203, + 0.005068381782621145, + 0.0024621058255434036, + 0.003111175959929824, + 0.005249985959380865, + 0.001891268533654511, + 0.002279214560985565, + 0.008583242073655128, + 0.003663947805762291, + 0.004626675043255091, + 0.004317117389291525, + 0.0031802942976355553, + 0.00440272968262434, + 0.0037629182916134596, + 0.004818718880414963, + 0.00671067601069808, + 0.0041291615925729275, + 0.04170835763216019, + 0.012514844536781311, + 0.01249033585190773, + 0.008062557317316532, + 0.017620768398046494, + 0.051137711852788925, + 0.020175227895379066, + 0.013806439936161041, + 0.0174948051571846, + 0.015076580457389355, + 0.010340617038309574, + 0.017045902088284492, + 0.0173188503831625, + 0.02734202705323696, + 0.02502840757369995, + 0.02146853134036064, + 0.011128858663141727, + 0.004251812119036913, + 0.005408644210547209, + 0.008828301914036274, + 0.0046903654001653194, + 0.007042073644697666, + 0.009465523064136505, + 0.006381429731845856, + 0.005312574096024036, + 0.004505847115069628, + 0.003475356847047806, + 0.004431223962455988, + 0.002755085239186883, + 0.0049468884244561195, + 0.005059355869889259, + 0.0037819587159901857, + 0.02478126622736454, + 0.018383776769042015, + 0.021225808188319206, + 0.03540283814072609, + 0.018468821421265602, + 0.02062448300421238, + 0.0248393751680851, + 0.01953902095556259, + 0.007055993657559156, + 0.017546596005558968, + 0.007704500108957291, + 0.015533976256847382, + 0.043302860110998154, + 0.025939878076314926, + 0.012872994877398014, + 0.01682504266500473 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 5 + ], + "selected_reward": -0.005956398788839579 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026353510096669197 + ], + "opposite_logprob_tokens": [ + -3.651353597640991 + ], + "logprob_reward": -0.026353510096669197, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026353510096669197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13801875710487366, + 0.03105013072490692, + 0.02152336575090885, + 0.034381695091724396, + 0.10673884302377701, + 0.20202091336250305, + 0.15613892674446106, + 0.045965466648340225, + 0.06587312370538712, + 0.1982887089252472 + ], + "credit_map": [ + 0.025777027010917664, + 0.007589593529701233, + 0.011035888455808163, + 0.006395032629370689, + 0.008457034826278687, + 0.04207305610179901, + 0.018475202843546867, + 0.01821592077612877, + 0.005264091305434704, + 0.0026199771091341972, + 0.0037064196076244116, + 0.0036334737669676542, + 0.002361451741307974, + 0.0023836151231080294, + 0.005672193132340908, + 0.005408910568803549, + 0.004141924902796745, + 0.0014009077567607164, + 0.002035643672570586, + 0.002140078926458955, + 0.0016834643902257085, + 0.0016639907844364643, + 0.005383638199418783, + 0.0030737172346562147, + 0.006091263610869646, + 0.002501771552488208, + 0.002886386588215828, + 0.004849719814956188, + 0.002233124803751707, + 0.003848780645057559, + 0.008268133737146854, + 0.003702513175085187, + 0.023234976455569267, + 0.00814652070403099, + 0.008931012824177742, + 0.009585989639163017, + 0.009267318993806839, + 0.026626603677868843, + 0.01287350058555603, + 0.008072914555668831, + 0.0373971201479435, + 0.015960393473505974, + 0.01502136792987585, + 0.022886337712407112, + 0.013277936726808548, + 0.020393218845129013, + 0.029026128351688385, + 0.0480584055185318, + 0.023296233266592026, + 0.01632521115243435, + 0.019579505547881126, + 0.026457907631993294, + 0.015165554359555244, + 0.0185295008122921, + 0.023787496611475945, + 0.012997519224882126, + 0.0067633939906954765, + 0.0038338014855980873, + 0.00461957510560751, + 0.006526818033307791, + 0.005105830263346434, + 0.006505094934254885, + 0.00680161826312542, + 0.005809336435049772, + 0.007716374006122351, + 0.007720216643065214, + 0.007586304564028978, + 0.012900689616799355, + 0.005932813510298729, + 0.007270185276865959, + 0.009498517960309982, + 0.0072480239905416965, + 0.016004497185349464, + 0.024218587204813957, + 0.019306033849716187, + 0.023104839026927948, + 0.03962975740432739, + 0.03053523413836956, + 0.026754047721624374, + 0.018735714256763458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.026353510096669197 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a close-up shot in stark black and white, the camera remains still, focused on a massive watermelon about to be split open. A sharp, high-contrast overhead light casts a dazzling metallic white on the broad blade of a heavy carving knife, while the dark outline of the handle almost blends into the pure black background. The alternating light and dark stripes on the watermelon's smooth skin transform into a vibrant gray and nearly black banding pattern against the stark monochrome. With a swift, decisive arc, the blade cuts down without hesitation, sinking deeply into the watermelon with a clean, sharp stroke, instantly bisecting it. In the moment of the split, countless drops of crystalline juice are violently hurled in all directions by the immense force of the impact, each drop captured against the pitch-black background like a tiny diamond bursting into flame. The once vibrant red, juicy flesh takes on a grainy, dark gray. The entire image is imbued with extreme dynamic tension and sculptural beauty, transforming a simple action into a powerful visual poem.", + "seed": 7336, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_13/index_6245_seed_7336_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10222536567598581, + "vlm_logprob": -0.10222536567598581, + "vlm_yes_no_margin": 3.950000011920929, + "vlm_token_credit": -0.10222536567598581, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.46983665227890015 + ], + "opposite_logprob_tokens": [ + -1.094836711883545 + ], + "logprob_reward": -0.46983665227890015, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.46983665227890015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07453010231256485, + 0.07699202001094818, + 0.07181745767593384, + 0.06357009708881378, + 0.04436875134706497, + 0.13875050842761993, + 0.14751195907592773, + 0.1269174963235855, + 0.13093747198581696, + 0.12460407614707947 + ], + "credit_map": [ + 0.004669676534831524, + 0.006715844850987196, + 0.006549314130097628, + 0.011272488161921501, + 0.010162128135561943, + 0.006654831580817699, + 0.016411226242780685, + 0.012094603851437569, + 0.0069979336112737656, + 0.007246802560985088, + 0.007220730185508728, + 0.010511058382689953, + 0.008474874310195446, + 0.006588563323020935, + 0.015960974618792534, + 0.013991079293191433, + 0.006243075709789991, + 0.00709183095023036, + 0.008275429718196392, + 0.009028811939060688, + 0.006411511451005936, + 0.006292396225035191, + 0.015718400478363037, + 0.012756003998219967, + 0.005039669573307037, + 0.008138244971632957, + 0.005857257638126612, + 0.00924623291939497, + 0.00574177410453558, + 0.008082921616733074, + 0.011141981929540634, + 0.010322018526494503, + 0.0031395573168992996, + 0.003488008165732026, + 0.006206235382705927, + 0.007971588522195816, + 0.003026597434654832, + 0.003552832407876849, + 0.008548954501748085, + 0.008434976451098919, + 0.011139553040266037, + 0.012099046260118484, + 0.013307123444974422, + 0.015211718156933784, + 0.01272611040621996, + 0.025880279019474983, + 0.022218655794858932, + 0.026168040931224823, + 0.012081793509423733, + 0.015959173440933228, + 0.019601671025156975, + 0.024636058136820793, + 0.01764870248734951, + 0.016706295311450958, + 0.027623141184449196, + 0.013255133293569088, + 0.012168332003057003, + 0.01682629995048046, + 0.014280454255640507, + 0.015989409759640694, + 0.024103228002786636, + 0.014374020509421825, + 0.00935450941324234, + 0.01982124336063862, + 0.021206850185990334, + 0.01919073797762394, + 0.026031792163848877, + 0.012392671778798103, + 0.012466306798160076, + 0.011186429299414158, + 0.015539588406682014, + 0.012923112139105797, + 0.017354629933834076, + 0.024156557396054268, + 0.010846257209777832, + 0.014415646903216839, + 0.009646857157349586, + 0.01416821125894785, + 0.012776274234056473, + 0.02123965322971344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.46983665227890015 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023334426805377007 + ], + "opposite_logprob_tokens": [ + -3.773334503173828 + ], + "logprob_reward": -0.023334426805377007, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023334426805377007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13158857822418213, + 0.024745063856244087, + 0.029895758256316185, + 0.03134799003601074, + 0.30443620681762695, + 0.031045161187648773, + 0.041133590042591095, + 0.05335630476474762, + 0.19270192086696625, + 0.15974944829940796 + ], + "credit_map": [ + 0.012670296244323254, + 0.010478368028998375, + 0.012231653556227684, + 0.01401585154235363, + 0.01399585697799921, + 0.012890249490737915, + 0.03164024278521538, + 0.02366606704890728, + 0.0021517544519156218, + 0.003152331570163369, + 0.0026782709173858166, + 0.0042579988949000835, + 0.0029398447368294, + 0.0021086791530251503, + 0.004123160615563393, + 0.0033330246806144714, + 0.003092753002420068, + 0.00207086862064898, + 0.002775185275822878, + 0.004109661560505629, + 0.0024476638063788414, + 0.004857327789068222, + 0.005609809886664152, + 0.004932488780468702, + 0.0032826035749167204, + 0.003103390336036682, + 0.0019660366233438253, + 0.003989819902926683, + 0.003184282686561346, + 0.006931505165994167, + 0.004618479870259762, + 0.004271873738616705, + 0.018738074228167534, + 0.008000073954463005, + 0.06160254403948784, + 0.019873255863785744, + 0.011336877010762691, + 0.011759313754737377, + 0.06564577668905258, + 0.10748032480478287, + 0.0024286825209856033, + 0.0030614971183240414, + 0.00263249222189188, + 0.0043090395629405975, + 0.0025919233448803425, + 0.005349528975784779, + 0.005942036863416433, + 0.004729960113763809, + 0.004214349668473005, + 0.0034501738846302032, + 0.003381140064448118, + 0.004919275175780058, + 0.005841964390128851, + 0.006416448391973972, + 0.007276827469468117, + 0.005633410066366196, + 0.005116933025419712, + 0.006903138477355242, + 0.005633202847093344, + 0.0062208049930632114, + 0.012491063214838505, + 0.005545527674257755, + 0.007065753918141127, + 0.004379881080240011, + 0.0332188680768013, + 0.015042203478515148, + 0.06556309014558792, + 0.015102626755833626, + 0.012473045848309994, + 0.012612961232662201, + 0.01946774311363697, + 0.01922137290239334, + 0.01672407239675522, + 0.03156444430351257, + 0.010304853320121765, + 0.01796761341392994, + 0.01190151646733284, + 0.031513575464487076, + 0.015981676056981087, + 0.023791706189513206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.023334426805377007 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005263755563646555 + ], + "opposite_logprob_tokens": [ + -5.255263805389404 + ], + "logprob_reward": -0.005263755563646555, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005263755563646555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10310503840446472, + 0.06333687156438828, + 0.05563771724700928, + 0.2207222431898117, + 0.04030687361955643, + 0.031921789050102234, + 0.2883754372596741, + 0.05135849863290787, + 0.06605026125907898, + 0.07918527722358704 + ], + "credit_map": [ + 0.008763146586716175, + 0.010827652178704739, + 0.00940720085054636, + 0.01581474579870701, + 0.012086689472198486, + 0.006785588804632425, + 0.02294892817735672, + 0.01647108420729637, + 0.006519611459225416, + 0.007055793423205614, + 0.005444003269076347, + 0.01007881760597229, + 0.006882596295326948, + 0.006349983625113964, + 0.01122238114476204, + 0.009783685207366943, + 0.006326866336166859, + 0.005943843629211187, + 0.005561209283769131, + 0.008418578654527664, + 0.00410240376368165, + 0.006144020240753889, + 0.010766660794615746, + 0.008374136872589588, + 0.015078351832926273, + 0.026033898815512657, + 0.021059410646557808, + 0.027936415746808052, + 0.018358439207077026, + 0.019278263673186302, + 0.04749513044953346, + 0.04548234865069389, + 0.003500966588035226, + 0.003618633607402444, + 0.006049464922398329, + 0.005960787646472454, + 0.002858218736946583, + 0.004212791100144386, + 0.00703501608222723, + 0.0070709907449781895, + 0.0026222257874906063, + 0.00521619850769639, + 0.002598054939880967, + 0.004255553707480431, + 0.0034832446835935116, + 0.005286884494125843, + 0.004754120483994484, + 0.0037055062130093575, + 0.027650900185108185, + 0.03422542288899422, + 0.02454187162220478, + 0.049553509801626205, + 0.04571548104286194, + 0.033933788537979126, + 0.04176066815853119, + 0.030993813648819923, + 0.008705473504960537, + 0.0057134609669446945, + 0.005266127642244101, + 0.0061109634116292, + 0.008416586555540562, + 0.0049955337308347225, + 0.006767562124878168, + 0.00538278603926301, + 0.01027219370007515, + 0.007041103672236204, + 0.014189017005264759, + 0.00603904016315937, + 0.01132740918546915, + 0.005405569914728403, + 0.0054346853867173195, + 0.006341241300106049, + 0.012482996098697186, + 0.019083745777606964, + 0.005197327584028244, + 0.006207697093486786, + 0.006721248850226402, + 0.011285698972642422, + 0.005951912607997656, + 0.012254653498530388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.005263755563646555 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595355499535799 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.00595355499535799, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595355499535799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062289848923683167, + 0.025631802156567574, + 0.03875840827822685, + 0.06487217545509338, + 0.1769210398197174, + 0.07045963406562805, + 0.05485276132822037, + 0.09562566876411438, + 0.34599533677101135, + 0.06459330022335052 + ], + "credit_map": [ + 0.004844957496970892, + 0.006927526090294123, + 0.004861895926296711, + 0.010173497721552849, + 0.008009185083210468, + 0.004975865129381418, + 0.012894146144390106, + 0.009602782316505909, + 0.002442616503685713, + 0.0029331890400499105, + 0.0022235200740396976, + 0.0041978987865149975, + 0.002683813450857997, + 0.0020613085944205523, + 0.0050978828221559525, + 0.003991575445979834, + 0.00291653536260128, + 0.003323768964037299, + 0.00326390634290874, + 0.004678591154515743, + 0.0030942759476602077, + 0.007830760441720486, + 0.007326755207031965, + 0.006323818117380142, + 0.004912347067147493, + 0.007407356984913349, + 0.004834351129829884, + 0.008154597133398056, + 0.006858744192868471, + 0.01234058104455471, + 0.01157564390450716, + 0.00878855399787426, + 0.009444404393434525, + 0.00906684622168541, + 0.02352823317050934, + 0.022642357274889946, + 0.007979419082403183, + 0.009391925297677517, + 0.04109625518321991, + 0.053771596401929855, + 0.0058217234909534454, + 0.0060316892340779305, + 0.005769113078713417, + 0.006490138825029135, + 0.006234121974557638, + 0.014626475051045418, + 0.010867755860090256, + 0.01461862027645111, + 0.005847148597240448, + 0.003987965639680624, + 0.006117961835116148, + 0.006978983525186777, + 0.007756768725812435, + 0.00736257154494524, + 0.008785552345216274, + 0.008015813305974007, + 0.010487543419003487, + 0.012203891761600971, + 0.011162959970533848, + 0.010712171904742718, + 0.019450165331363678, + 0.010572783648967743, + 0.010213898494839668, + 0.010822261683642864, + 0.05643508583307266, + 0.030590754002332687, + 0.10677912086248398, + 0.02571314014494419, + 0.02069336734712124, + 0.02553323097527027, + 0.037001051008701324, + 0.043249599635601044, + 0.009881886653602123, + 0.010248871520161629, + 0.006141218356788158, + 0.010702649131417274, + 0.006238151341676712, + 0.006307463627308607, + 0.005753592122346163, + 0.009319476783275604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 7 + ], + "selected_reward": -0.00595355499535799 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot in a close-up view with a still camera?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video presented in stark black and white with a high-contrast monochrome palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a massive watermelon visible with a smooth skin featuring alternating light and dark stripes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a heavy carving knife with a sharp blade and dark handle cut down in a swift arc to bisect the watermelon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless drops of juice violently hurled in all directions against a pitch-black background at the moment of the split?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067384387366473675 + ], + "opposite_logprob_tokens": [ + -5.006738662719727 + ], + "logprob_reward": -0.0067384387366473675, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067384387366473675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06759774684906006, + 0.022109854966402054, + 0.02364286780357361, + 0.026632055640220642, + 0.042834945023059845, + 0.06599032878875732, + 0.03714144229888916, + 0.05854860693216324, + 0.41541385650634766, + 0.24008828401565552 + ], + "credit_map": [ + 0.006423240061849356, + 0.005611492786556482, + 0.007107954937964678, + 0.009689807891845703, + 0.008466897532343864, + 0.004607265815138817, + 0.015179122798144817, + 0.010511967353522778, + 0.0022886067163199186, + 0.0020661635790020227, + 0.001648339326493442, + 0.0035301269963383675, + 0.001944124698638916, + 0.001893091481178999, + 0.0049877227284014225, + 0.003751679789274931, + 0.0019014824647456408, + 0.0017582582077011466, + 0.0020877658389508724, + 0.002969620283693075, + 0.0018334242049604654, + 0.002916892757639289, + 0.005560838617384434, + 0.004614587407559156, + 0.0023532863706350327, + 0.0027325258124619722, + 0.0023517131339758635, + 0.003095893654972315, + 0.002664408180862665, + 0.003445022739470005, + 0.005275967996567488, + 0.004713236819952726, + 0.002645788947120309, + 0.0015914771938696504, + 0.008384442888200283, + 0.0024737652856856585, + 0.0017460286617279053, + 0.0028080842457711697, + 0.009100374765694141, + 0.014084981754422188, + 0.004667453467845917, + 0.005875422153621912, + 0.005317394156008959, + 0.007429356221109629, + 0.007220136001706123, + 0.01200505718588829, + 0.010553397238254547, + 0.012922114692628384, + 0.003645667340606451, + 0.004455108195543289, + 0.005315921735018492, + 0.006017687730491161, + 0.004151674918830395, + 0.004635171499103308, + 0.005300805903971195, + 0.003619404509663582, + 0.005312023684382439, + 0.0067159212194383144, + 0.006246304139494896, + 0.008012672886252403, + 0.010073761455714703, + 0.006886359304189682, + 0.0062108878046274185, + 0.009090673178434372, + 0.07370612770318985, + 0.04186958819627762, + 0.10471244901418686, + 0.0366390198469162, + 0.029123641550540924, + 0.03288290277123451, + 0.04651713743805885, + 0.049962982535362244, + 0.031229553744196892, + 0.04664973169565201, + 0.021576637402176857, + 0.0310643520206213, + 0.02022281289100647, + 0.028038522228598595, + 0.027543650940060616, + 0.03376302495598793 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0067384387366473675 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_14/index_tensor([515, 515])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_14/index_tensor([515, 515])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f82d3f2521df8d03afaa79a5de484a7c4e32697f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_14/index_tensor([515, 515])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extremely low-angle tracking shot, positioned just inches from the ground, glides smoothly alongside a severely crumpled lottery ticket as it crosses a weathered concrete sidewalk. The once tightly held ticket is now partially unfolded, its surface riddled with sharp creases and soft wrinkles, the design's plain, faded primary colors barely visible beneath a layer of grime. Harsh, direct sunlight from late afternoon sweeps across the scene, transforming the cracks and gravel texture of the sidewalk into dramatic topography and casting long, sharp shadows from the invisible urban fabric that stretch across the frame like dark fingers. A faint, intermittent breeze, the sole source of movement, causes the ticket to tumble erratically, skittering sideways, occasionally rising delicately into the air, stirring up tiny whirlwinds of fine dust and small, dry leaf fragments that swirl around it. The entire scene is bathed in a warm, golden-orange glow, yet the slow, lonely dance of the discarded ticket across the desolate landscape evokes a potent atmosphere of quiet melancholy and forgotten hope.", + "seed": 9319, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_14/index_515_seed_9319_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13825349230319262, + "vlm_logprob": -0.13825349230319262, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.13825349230319262, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12655754387378693 + ], + "opposite_logprob_tokens": [ + -2.2515575885772705 + ], + "logprob_reward": -0.12655754387378693, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12655754387378693, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09694234281778336, + 0.06773445010185242, + 0.03652065247297287, + 0.08334348350763321, + 0.04429806023836136, + 0.1932898610830307, + 0.06222435086965561, + 0.2832350730895996, + 0.050452981144189835, + 0.08195874094963074 + ], + "credit_map": [ + 0.007474302314221859, + 0.009770181030035019, + 0.009730980731546879, + 0.0183517187833786, + 0.00880630873143673, + 0.010359426960349083, + 0.013426298275589943, + 0.019023122265934944, + 0.0033354281913489103, + 0.009044468402862549, + 0.004131898749619722, + 0.007341293152421713, + 0.017176242545247078, + 0.007338316645473242, + 0.01270159799605608, + 0.006665208376944065, + 0.0028003619518131018, + 0.005200263112783432, + 0.004353536292910576, + 0.007297954522073269, + 0.005246707703918219, + 0.004349116235971451, + 0.0036936611868441105, + 0.0035790535621345043, + 0.006951679941266775, + 0.012336939573287964, + 0.007137208245694637, + 0.015047358348965645, + 0.009783215820789337, + 0.016911551356315613, + 0.007285456173121929, + 0.007890075445175171, + 0.005581801291555166, + 0.00804147683084011, + 0.004528829827904701, + 0.004503391683101654, + 0.0041181729175150394, + 0.008071953430771828, + 0.003430936485528946, + 0.006021498702466488, + 0.048245273530483246, + 0.02919141948223114, + 0.012199054472148418, + 0.009949631057679653, + 0.01073369663208723, + 0.016612110659480095, + 0.02172657661139965, + 0.04463211074471474, + 0.013432027772068977, + 0.005922539159655571, + 0.003615857334807515, + 0.00392371928319335, + 0.002126484178006649, + 0.009322523139417171, + 0.0031495923176407814, + 0.02073160745203495, + 0.024399064481258392, + 0.048099372535943985, + 0.02752397209405899, + 0.0365464948117733, + 0.034895896911621094, + 0.027492418885231018, + 0.03403182700276375, + 0.05024602264165878, + 0.006519857794046402, + 0.005112082231789827, + 0.007454024162143469, + 0.0036347631830722094, + 0.005803951062262058, + 0.009408208541572094, + 0.00413300609216094, + 0.008387086912989616, + 0.012493062764406204, + 0.011153568513691425, + 0.005378821399062872, + 0.0066462852992117405, + 0.006710730958729982, + 0.027662400156259537, + 0.006023810710757971, + 0.005890059284865856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.12655754387378693 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233526825904846 + ], + "opposite_logprob_tokens": [ + -0.8983526825904846 + ], + "logprob_reward": -0.5233526825904846, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233526825904846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08176239579916, + 0.018364571034908295, + 0.03405786678195, + 0.06374738365411758, + 0.05547412484884262, + 0.3098442852497101, + 0.024200811982154846, + 0.19197052717208862, + 0.04097281023859978, + 0.17960524559020996 + ], + "credit_map": [ + 0.006826359312981367, + 0.007403329014778137, + 0.006256395485252142, + 0.013279414735734463, + 0.007397213950753212, + 0.006018654443323612, + 0.016126640141010284, + 0.01845439523458481, + 0.0013454229338094592, + 0.002363189123570919, + 0.0014577468391507864, + 0.002240526955574751, + 0.004639682825654745, + 0.00251524499617517, + 0.0024630180560052395, + 0.0013397386064752936, + 0.004019429907202721, + 0.0063453735783696175, + 0.004704983904957771, + 0.004343684762716293, + 0.004820230882614851, + 0.004146804101765156, + 0.002564032096415758, + 0.003113323822617531, + 0.005617773625999689, + 0.016158869490027428, + 0.007010696455836296, + 0.006770334672182798, + 0.0064023579470813274, + 0.013744931668043137, + 0.004396368283778429, + 0.0036460498813539743, + 0.0036741613876074553, + 0.006767482962459326, + 0.006115179043263197, + 0.006283405236899853, + 0.015273078344762325, + 0.006848829332739115, + 0.004749576095491648, + 0.0057624103501439095, + 0.06854484230279922, + 0.041906144469976425, + 0.03249518200755119, + 0.012395447120070457, + 0.018445881083607674, + 0.02864612080156803, + 0.03607361018657684, + 0.07133703678846359, + 0.004330661613494158, + 0.00281068729236722, + 0.0032110055908560753, + 0.002029514405876398, + 0.0017607719637453556, + 0.0032115187495946884, + 0.0015588079113513231, + 0.005287845153361559, + 0.01799917034804821, + 0.032333727926015854, + 0.01973819173872471, + 0.023627491667866707, + 0.020777877420186996, + 0.01899801567196846, + 0.024068813771009445, + 0.034427233040332794, + 0.0052180541679263115, + 0.003086542012169957, + 0.006109689362347126, + 0.004647971596568823, + 0.004768719896674156, + 0.00555969774723053, + 0.0038445577956736088, + 0.007737576030194759, + 0.03726962208747864, + 0.018207700923085213, + 0.011929736472666264, + 0.018417714163661003, + 0.013955110684037209, + 0.04442375898361206, + 0.021325258538126945, + 0.0140763521194458 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.5233526825904846 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607704907655716 + ], + "opposite_logprob_tokens": [ + -4.141077041625977 + ], + "logprob_reward": -0.01607704907655716, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607704907655716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23524773120880127, + 0.03968518599867821, + 0.053786445409059525, + 0.11313337087631226, + 0.047212403267621994, + 0.06313483417034149, + 0.031871456652879715, + 0.29562756419181824, + 0.10721239447593689, + 0.013088627718389034 + ], + "credit_map": [ + 0.012060975655913353, + 0.018852274864912033, + 0.021801024675369263, + 0.03686954081058502, + 0.01752433553338051, + 0.017579587176442146, + 0.03850680962204933, + 0.07205317169427872, + 0.005167142488062382, + 0.005579277407377958, + 0.003903469070792198, + 0.004725003149360418, + 0.00821653287857771, + 0.0034262717235833406, + 0.005407370626926422, + 0.003260119818150997, + 0.00334484432823956, + 0.007705298252403736, + 0.006343503016978502, + 0.009537691250443459, + 0.007589769084006548, + 0.005893738940358162, + 0.005984142888337374, + 0.0073874592781066895, + 0.010546250268816948, + 0.031041795387864113, + 0.009985718876123428, + 0.010005911812186241, + 0.010446901433169842, + 0.025883980095386505, + 0.00802748091518879, + 0.007195329759269953, + 0.003628758480772376, + 0.0060513014905154705, + 0.005838742945343256, + 0.006639184895902872, + 0.009896557778120041, + 0.0063104890286922455, + 0.0033538558054715395, + 0.005493514705449343, + 0.00625170674175024, + 0.009933381341397762, + 0.010528569109737873, + 0.00906831119209528, + 0.007713743951171637, + 0.006523463875055313, + 0.0046873316168785095, + 0.008428323082625866, + 0.0041272807866334915, + 0.005386639386415482, + 0.004831749014556408, + 0.0035298378206789494, + 0.0029887580312788486, + 0.004090157803148031, + 0.003096964443102479, + 0.003820069134235382, + 0.041904766112565994, + 0.04047897830605507, + 0.02082441747188568, + 0.025450006127357483, + 0.04263942316174507, + 0.06984733790159225, + 0.02006719261407852, + 0.03441541641950607, + 0.0128835653886199, + 0.010313443839550018, + 0.01743183843791485, + 0.007707573939114809, + 0.013892075978219509, + 0.020456798374652863, + 0.007951416075229645, + 0.016575688496232033, + 0.001780437189154327, + 0.0021692148875445127, + 0.0016984790563583374, + 0.0020105200819671154, + 0.001044910284690559, + 0.0015552437398582697, + 0.0017633694224059582, + 0.0010664532892405987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.01607704907655716 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086649261415005 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.011086649261415005, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086649261415005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07193689048290253, + 0.02507743053138256, + 0.07919631153345108, + 0.04412757605314255, + 0.030831368640065193, + 0.07598372548818588, + 0.08601793646812439, + 0.520367443561554, + 0.033142272382974625, + 0.033319056034088135 + ], + "credit_map": [ + 0.00887238048017025, + 0.009220262989401817, + 0.008380524814128876, + 0.009933983907103539, + 0.007843763567507267, + 0.006932790391147137, + 0.012973586097359657, + 0.007779598701745272, + 0.002602027729153633, + 0.002651938470080495, + 0.0025078249163925648, + 0.003108208766207099, + 0.004420748446136713, + 0.0035660674329847097, + 0.0038614615332335234, + 0.0023591534700244665, + 0.006946144159883261, + 0.012123788706958294, + 0.011032499372959137, + 0.012873134575784206, + 0.012098872102797031, + 0.009119881317019463, + 0.006474080961197615, + 0.008527912199497223, + 0.0037540507037192583, + 0.008097548969089985, + 0.0036286686081439257, + 0.008430265821516514, + 0.005337154492735863, + 0.007701757363975048, + 0.0031107887625694275, + 0.004067337140440941, + 0.004280284978449345, + 0.0051156566478312016, + 0.0035250133369117975, + 0.002782257040962577, + 0.0034090217668563128, + 0.004382147919386625, + 0.0022899992763996124, + 0.005046986974775791, + 0.017448537051677704, + 0.012031467631459236, + 0.00880963634699583, + 0.004412597976624966, + 0.004509922582656145, + 0.007122694049030542, + 0.006508639082312584, + 0.015140224248170853, + 0.016175707802176476, + 0.009896713308990002, + 0.006476215552538633, + 0.005472627468407154, + 0.004021046683192253, + 0.013280817307531834, + 0.004502505995333195, + 0.026192305609583855, + 0.05879781022667885, + 0.07375162094831467, + 0.052409008145332336, + 0.057736776769161224, + 0.04632863402366638, + 0.08564453572034836, + 0.05927080661058426, + 0.08642825484275818, + 0.0031643062829971313, + 0.0031259143725037575, + 0.0041468264535069466, + 0.004179670475423336, + 0.0045464797876775265, + 0.005841290112584829, + 0.0030009583570063114, + 0.005136825144290924, + 0.005109789315611124, + 0.004461747128516436, + 0.0033793749753385782, + 0.005227676127105951, + 0.0026555380318313837, + 0.004201419185847044, + 0.004948387388139963, + 0.003335120389238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 2 + ], + "selected_reward": -0.011086649261415005 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193536713719368 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193536713719368, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193536713719368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17103923857212067, + 0.09142801910638809, + 0.057746756821870804, + 0.03245893493294716, + 0.16392533481121063, + 0.1454307734966278, + 0.026571232825517654, + 0.21064768731594086, + 0.0398184172809124, + 0.06093361973762512 + ], + "credit_map": [ + 0.011570527218282223, + 0.022484226152300835, + 0.020401816815137863, + 0.03413110971450806, + 0.014260867610573769, + 0.009327467530965805, + 0.03284187987446785, + 0.02602134831249714, + 0.01430805679410696, + 0.013362092897295952, + 0.010697804391384125, + 0.011782548390328884, + 0.017111817374825478, + 0.007470861077308655, + 0.010200658813118935, + 0.006494176108390093, + 0.004802923649549484, + 0.010824300348758698, + 0.006342378444969654, + 0.009030562825500965, + 0.008596090599894524, + 0.006475597154349089, + 0.0048503633588552475, + 0.006824538577347994, + 0.003155240323394537, + 0.006965591572225094, + 0.0028913638088852167, + 0.004103632178157568, + 0.003006627317517996, + 0.0066468371078372, + 0.002457743277773261, + 0.0032319005113095045, + 0.01642148196697235, + 0.024714741855859756, + 0.014201589860022068, + 0.020666733384132385, + 0.015557177364826202, + 0.027656253427267075, + 0.01435843389481306, + 0.030348917469382286, + 0.03243764489889145, + 0.024052800610661507, + 0.016671588644385338, + 0.0072477832436561584, + 0.008288932964205742, + 0.010754658840596676, + 0.015564529225230217, + 0.030412837862968445, + 0.003947167191654444, + 0.003575950860977173, + 0.003523028688505292, + 0.0019359400030225515, + 0.0018475622637197375, + 0.003796870354562998, + 0.0019468682585284114, + 0.005997845437377691, + 0.024068990722298622, + 0.028567206114530563, + 0.020604081451892853, + 0.022020746022462845, + 0.018884271383285522, + 0.037986937910318375, + 0.022427627816796303, + 0.03608783707022667, + 0.00349519494920969, + 0.00519226910546422, + 0.004840932320803404, + 0.005686130374670029, + 0.005109051242470741, + 0.004354014992713928, + 0.005081434268504381, + 0.006059391889721155, + 0.010782718658447266, + 0.006912657059729099, + 0.002609595190733671, + 0.005330270156264305, + 0.005242032464593649, + 0.020282777026295662, + 0.005199866369366646, + 0.004573699552565813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.014193536713719368 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely low-angle tracking shot, positioned just inches from the ground, glides smoothly alongside a severely crumpled lottery ticket as it crosses a weathered concrete sidewalk. The once tightly held ticket is now partially unfolded, its surface riddled with sharp creases and soft wrinkles, the design's plain, faded primary colors barely visible beneath a layer of grime. Harsh, direct sunlight from late afternoon sweeps across the scene, transforming the cracks and gravel texture of the sidewalk into dramatic topography and casting long, sharp shadows from the invisible urban fabric that stretch across the frame like dark fingers. A faint, intermittent breeze, the sole source of movement, causes the ticket to tumble erratically, skittering sideways, occasionally rising delicately into the air, stirring up tiny whirlwinds of fine dust and small, dry leaf fragments that swirl around it. The entire scene is bathed in a warm, golden-orange glow, yet the slow, lonely dance of the discarded ticket across the desolate landscape evokes a potent atmosphere of quiet melancholy and forgotten hope.", + "seed": 9320, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_14/index_515_seed_9320_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1677677409723401, + "vlm_logprob": -0.1677677409723401, + "vlm_yes_no_margin": 3.2, + "vlm_token_credit": -0.1677677409723401, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17771849036216736 + ], + "opposite_logprob_tokens": [ + -1.9277185201644897 + ], + "logprob_reward": -0.17771849036216736, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17771849036216736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4395504295825958, + 0.05123184621334076, + 0.08367464691400528, + 0.07381857931613922, + 0.03543081134557724, + 0.07085087150335312, + 0.07810136675834656, + 0.023709803819656372, + 0.04062722995877266, + 0.10300441831350327 + ], + "credit_map": [ + 0.06994059681892395, + 0.07246291637420654, + 0.06138426810503006, + 0.036153171211481094, + 0.06555182486772537, + 0.0299252700060606, + 0.06107107177376747, + 0.04306130111217499, + 0.006600530352443457, + 0.007856563664972782, + 0.006427249405533075, + 0.005668942350894213, + 0.004635367076843977, + 0.005164468195289373, + 0.008229103870689869, + 0.0066496203653514385, + 0.011422710493206978, + 0.011526701971888542, + 0.010942727327346802, + 0.00883136410266161, + 0.010887634009122849, + 0.010723590850830078, + 0.010194385424256325, + 0.009145529940724373, + 0.009224056266248226, + 0.00723956897854805, + 0.00981801375746727, + 0.009455754421651363, + 0.007492509670555592, + 0.009406298398971558, + 0.011440356262028217, + 0.009742017835378647, + 0.004986175801604986, + 0.002643237356096506, + 0.00412254873663187, + 0.0034424166660755873, + 0.004176530987024307, + 0.0045317173935472965, + 0.002697011921554804, + 0.008831175044178963, + 0.01400794368237257, + 0.004590126220136881, + 0.006487073842436075, + 0.007341926917433739, + 0.008481418713927269, + 0.015384197235107422, + 0.005059713497757912, + 0.009498469531536102, + 0.013582777231931686, + 0.005582176148891449, + 0.005323090590536594, + 0.005292812362313271, + 0.006693185772746801, + 0.014844812452793121, + 0.0054063028655946255, + 0.021376214921474457, + 0.003912255167961121, + 0.001899626455269754, + 0.003615048946812749, + 0.00354597344994545, + 0.002289228141307831, + 0.002286122879013419, + 0.003384791314601898, + 0.002776756649836898, + 0.0054915957152843475, + 0.0033752431627362967, + 0.0049845632165670395, + 0.0037308998871594667, + 0.005417321342974901, + 0.005587700288742781, + 0.005963513627648354, + 0.006076394114643335, + 0.011836744844913483, + 0.012609654106199741, + 0.014975130558013916, + 0.013565498404204845, + 0.010441259481012821, + 0.009742479771375656, + 0.015335473231971264, + 0.014498167671263218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.17771849036216736 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328349113464355 + ], + "opposite_logprob_tokens": [ + -0.7578349113464355 + ], + "logprob_reward": -0.6328349113464355, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328349113464355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21130579710006714, + 0.08433563262224197, + 0.05361774563789368, + 0.030517246574163437, + 0.03403184562921524, + 0.11673714220523834, + 0.15837109088897705, + 0.066481813788414, + 0.0761023759841919, + 0.16849938035011292 + ], + "credit_map": [ + 0.033780042082071304, + 0.027472976595163345, + 0.02406410686671734, + 0.020202701911330223, + 0.030521972104907036, + 0.01771175116300583, + 0.034166064113378525, + 0.023386189714074135, + 0.012191957794129848, + 0.008229968138039112, + 0.014707458205521107, + 0.00902075320482254, + 0.011288049630820751, + 0.008026614785194397, + 0.012408020906150341, + 0.008462811820209026, + 0.007910073734819889, + 0.006497083697468042, + 0.008617745712399483, + 0.006031740922480822, + 0.005083413794636726, + 0.003855930408462882, + 0.007657471112906933, + 0.007964287884533405, + 0.004376525059342384, + 0.003757745958864689, + 0.004546077456325293, + 0.0029926972929388285, + 0.0030104336328804493, + 0.00399495754390955, + 0.003979542292654514, + 0.0038592673372477293, + 0.0032263253815472126, + 0.005694415885955095, + 0.005210082978010178, + 0.0036780324298888445, + 0.00487281009554863, + 0.003619995666667819, + 0.00309390714392066, + 0.0046362741850316525, + 0.024853942915797234, + 0.007909536361694336, + 0.009699173271656036, + 0.009790626354515553, + 0.013466251082718372, + 0.028832560405135155, + 0.007778496015816927, + 0.014406556263566017, + 0.01955440081655979, + 0.007594794500619173, + 0.007864834740757942, + 0.005970033351331949, + 0.006951208226382732, + 0.04875699430704117, + 0.0058447932824492455, + 0.05583402141928673, + 0.010363621637225151, + 0.005360347684472799, + 0.006564008072018623, + 0.005643305368721485, + 0.010386375710368156, + 0.007085544057190418, + 0.010634531266987324, + 0.01044408418238163, + 0.007036508992314339, + 0.007537614554166794, + 0.008012812584638596, + 0.007971387356519699, + 0.010708355344831944, + 0.011491210199892521, + 0.012993027456104755, + 0.010351455770432949, + 0.025141723453998566, + 0.01966448687016964, + 0.01941823400557041, + 0.01955064944922924, + 0.020060507580637932, + 0.015139203518629074, + 0.029219310730695724, + 0.020305251702666283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.6328349113464355 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954977124929428 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.005954977124929428, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954977124929428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06867491453886032, + 0.10599059611558914, + 0.05289756506681442, + 0.04600133001804352, + 0.17151013016700745, + 0.08865143358707428, + 0.05068108066916466, + 0.042582474648952484, + 0.07669568061828613, + 0.2963148057460785 + ], + "credit_map": [ + 0.010133270174264908, + 0.009553729556500912, + 0.007100965362042189, + 0.006689222063869238, + 0.009525265544652939, + 0.00569328386336565, + 0.009068136103451252, + 0.010911044664680958, + 0.012802519835531712, + 0.009992209263145924, + 0.016747601330280304, + 0.013950484804809093, + 0.010524521581828594, + 0.010060703381896019, + 0.019106466323137283, + 0.012806090526282787, + 0.007474189158529043, + 0.007836430333554745, + 0.008045997470617294, + 0.005537603981792927, + 0.0053798058070242405, + 0.005297502037137747, + 0.007160912733525038, + 0.0061651268042624, + 0.006631731521338224, + 0.004772843327373266, + 0.006185496691614389, + 0.005145321600139141, + 0.00503922626376152, + 0.006560791749507189, + 0.005476229824125767, + 0.0061896867118775845, + 0.020529335364699364, + 0.010398881509900093, + 0.015972303226590157, + 0.011054103262722492, + 0.01359250396490097, + 0.030370598658919334, + 0.011332262307405472, + 0.05826013535261154, + 0.014257194474339485, + 0.007107640150934458, + 0.014483976177871227, + 0.009008361026644707, + 0.01169170904904604, + 0.013554811477661133, + 0.005619450937956572, + 0.012928294949233532, + 0.005184503737837076, + 0.0024270194116979837, + 0.002886719536036253, + 0.0028896781150251627, + 0.0036773562896996737, + 0.015290182083845139, + 0.002139107324182987, + 0.016186516731977463, + 0.005657705944031477, + 0.003552067093551159, + 0.006673881784081459, + 0.006333319935947657, + 0.004774985369294882, + 0.0048149870708584785, + 0.006222902797162533, + 0.004552624188363552, + 0.007748595904558897, + 0.00954871904104948, + 0.010049319826066494, + 0.01115355920046568, + 0.009321928955614567, + 0.009250250644981861, + 0.010667800903320312, + 0.008955515921115875, + 0.05110760033130646, + 0.038965582847595215, + 0.03697820007801056, + 0.04609294235706329, + 0.029164357110857964, + 0.017895378172397614, + 0.04231465235352516, + 0.03379608318209648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 1 + ], + "selected_reward": -0.005954977124929428 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791087359189987 + ], + "opposite_logprob_tokens": [ + -4.634790897369385 + ], + "logprob_reward": -0.009791087359189987, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791087359189987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1748749166727066, + 0.11930707097053528, + 0.08143433183431625, + 0.081748828291893, + 0.13188409805297852, + 0.07424995303153992, + 0.09692904353141785, + 0.07270693778991699, + 0.052764758467674255, + 0.11410005390644073 + ], + "credit_map": [ + 0.018690641969442368, + 0.019904492422938347, + 0.02608557976782322, + 0.019092822447419167, + 0.019291624426841736, + 0.011249806731939316, + 0.03918469697237015, + 0.021375248208642006, + 0.013606124557554722, + 0.014102031476795673, + 0.015940846875309944, + 0.014228140003979206, + 0.012423188425600529, + 0.011166241951286793, + 0.021604184061288834, + 0.0162363164126873, + 0.011234188452363014, + 0.01034197025001049, + 0.011946323327720165, + 0.009932619519531727, + 0.010786090977489948, + 0.007319716736674309, + 0.010938356630504131, + 0.008935065008699894, + 0.010940386913716793, + 0.007985010743141174, + 0.01079829502850771, + 0.010387595742940903, + 0.008520342409610748, + 0.009547730907797813, + 0.012992600910365582, + 0.010576868429780006, + 0.012980292551219463, + 0.0070059411227703094, + 0.006793583743274212, + 0.008158840239048004, + 0.008172559551894665, + 0.02836923860013485, + 0.010837986133992672, + 0.04956565424799919, + 0.015954356640577316, + 0.0057840533554553986, + 0.006047307979315519, + 0.005539710633456707, + 0.008187818340957165, + 0.01950175128877163, + 0.005124739836901426, + 0.008110214956104755, + 0.007250404916703701, + 0.005472953896969557, + 0.006658360827714205, + 0.008732511661946774, + 0.01262999139726162, + 0.026987003162503242, + 0.007109424099326134, + 0.022088393568992615, + 0.013607127591967583, + 0.004300939850509167, + 0.00616997666656971, + 0.006832494400441647, + 0.009054605849087238, + 0.00996975228190422, + 0.012794885784387589, + 0.009977155365049839, + 0.00840934831649065, + 0.004576485138386488, + 0.0069260974414646626, + 0.00432608975097537, + 0.0060818493366241455, + 0.006465385667979717, + 0.005394923035055399, + 0.010584582574665546, + 0.01484187226742506, + 0.015279004350304604, + 0.016575895249843597, + 0.01580675132572651, + 0.0158159751445055, + 0.008612442761659622, + 0.015413575805723667, + 0.011754547245800495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.009791087359189987 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539238668978214 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539238668978214, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539238668978214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2585757076740265, + 0.10475324839353561, + 0.13377267122268677, + 0.06518466025590897, + 0.06613891571760178, + 0.11708851158618927, + 0.06194629147648811, + 0.07232341170310974, + 0.05232832953333855, + 0.0678882822394371 + ], + "credit_map": [ + 0.04322325065732002, + 0.03330501914024353, + 0.03665853664278984, + 0.024257851764559746, + 0.023039814084768295, + 0.020411869511008263, + 0.04305064305663109, + 0.03462871164083481, + 0.02097744680941105, + 0.01756737008690834, + 0.018157940357923508, + 0.009547380730509758, + 0.00868863333016634, + 0.00769134471192956, + 0.010407974012196064, + 0.011715160682797432, + 0.0207790806889534, + 0.015985017642378807, + 0.01781190000474453, + 0.014006043784320354, + 0.01673763245344162, + 0.011154635809361935, + 0.021138550713658333, + 0.016159819439053535, + 0.006801690440624952, + 0.007461745757609606, + 0.011691169813275337, + 0.008554639294743538, + 0.0069164312444627285, + 0.005737947300076485, + 0.009442288428544998, + 0.008578748442232609, + 0.013803314417600632, + 0.009748348034918308, + 0.011391299776732922, + 0.007582942023873329, + 0.0059427376836538315, + 0.007391711231321096, + 0.004558257758617401, + 0.005720307119190693, + 0.024415673688054085, + 0.008929958567023277, + 0.009454994462430477, + 0.010879130102694035, + 0.011060263961553574, + 0.029097769409418106, + 0.007466919254511595, + 0.015783801674842834, + 0.005423733498901129, + 0.002667434746399522, + 0.004103066399693489, + 0.002867825562134385, + 0.005339094903320074, + 0.01858045533299446, + 0.0032226690091192722, + 0.01974201388657093, + 0.012824565172195435, + 0.004684085492044687, + 0.006100581958889961, + 0.00669868104159832, + 0.009024033322930336, + 0.010106772184371948, + 0.012297045439481735, + 0.01058764848858118, + 0.0056070550344884396, + 0.0058242278173565865, + 0.005980610381811857, + 0.007112508174031973, + 0.0066995685920119286, + 0.006015701685100794, + 0.008114397525787354, + 0.006974261254072189, + 0.010411577299237251, + 0.006511378567665815, + 0.008444083854556084, + 0.008085987530648708, + 0.007935979403555393, + 0.006309563294053078, + 0.007216454949229956, + 0.012973257340490818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.012539238668978214 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely low-angle tracking shot, positioned just inches from the ground, glides smoothly alongside a severely crumpled lottery ticket as it crosses a weathered concrete sidewalk. The once tightly held ticket is now partially unfolded, its surface riddled with sharp creases and soft wrinkles, the design's plain, faded primary colors barely visible beneath a layer of grime. Harsh, direct sunlight from late afternoon sweeps across the scene, transforming the cracks and gravel texture of the sidewalk into dramatic topography and casting long, sharp shadows from the invisible urban fabric that stretch across the frame like dark fingers. A faint, intermittent breeze, the sole source of movement, causes the ticket to tumble erratically, skittering sideways, occasionally rising delicately into the air, stirring up tiny whirlwinds of fine dust and small, dry leaf fragments that swirl around it. The entire scene is bathed in a warm, golden-orange glow, yet the slow, lonely dance of the discarded ticket across the desolate landscape evokes a potent atmosphere of quiet melancholy and forgotten hope.", + "seed": 9321, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_14/index_515_seed_9321_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.16789066027849914, + "vlm_logprob": -0.16789066027849914, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.16789066027849914, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11239490658044815 + ], + "opposite_logprob_tokens": [ + -2.3623948097229004 + ], + "logprob_reward": -0.11239490658044815, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.11239490658044815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0679672509431839, + 0.05438675358891487, + 0.05644706264138222, + 0.2570700943470001, + 0.18985377252101898, + 0.023336797952651978, + 0.09788643568754196, + 0.0549466498196125, + 0.11379223316907883, + 0.08431290835142136 + ], + "credit_map": [ + 0.009251751936972141, + 0.008520405739545822, + 0.007346093654632568, + 0.008670474402606487, + 0.007706604897975922, + 0.006322299595922232, + 0.008031647652387619, + 0.01211797259747982, + 0.012125412933528423, + 0.008455978706479073, + 0.005877693649381399, + 0.004925508052110672, + 0.005161067470908165, + 0.005496833939105272, + 0.005211593583226204, + 0.007132663857191801, + 0.007364198099821806, + 0.0112658916041255, + 0.0038271895609796047, + 0.006458480376750231, + 0.01024380698800087, + 0.00601896783336997, + 0.004657307639718056, + 0.006611220072954893, + 0.059666018933057785, + 0.020710257813334465, + 0.011553751304745674, + 0.006921559106558561, + 0.018315335735678673, + 0.10940883308649063, + 0.01433568075299263, + 0.016158677637577057, + 0.036806363612413406, + 0.029322903603315353, + 0.0395490787923336, + 0.014917034655809402, + 0.019202956929802895, + 0.020985282957553864, + 0.011875451542437077, + 0.017194708809256554, + 0.002572680125012994, + 0.006305963732302189, + 0.002906775800511241, + 0.0023569955956190825, + 0.003966334275901318, + 0.0022051124833524227, + 0.0014505472499877214, + 0.0015723874093964696, + 0.011618045158684254, + 0.005824192427098751, + 0.006090137641876936, + 0.017051376402378082, + 0.03607350215315819, + 0.011744833551347256, + 0.005207289010286331, + 0.0042770602740347385, + 0.01550135388970375, + 0.0033564213663339615, + 0.006075597368180752, + 0.01624794863164425, + 0.0029008574783802032, + 0.0028374576941132545, + 0.004180280026048422, + 0.003846733830869198, + 0.013542603701353073, + 0.016231128945946693, + 0.018431175500154495, + 0.01458973903208971, + 0.014291560277342796, + 0.01352839358150959, + 0.010052932426333427, + 0.013124705292284489, + 0.004222644027322531, + 0.008467193692922592, + 0.010823046788573265, + 0.0032849907875061035, + 0.034736063331365585, + 0.015822144225239754, + 0.0036349063739180565, + 0.0033219198230654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.11239490658044815 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933355331420898 + ], + "opposite_logprob_tokens": [ + -0.6933355331420898 + ], + "logprob_reward": -0.6933355331420898, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933355331420898, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.03733007609844208, + 0.030626187101006508, + 0.06218487024307251, + 0.042839668691158295, + 0.08933214843273163, + 0.10945046693086624, + 0.4203116297721863, + 0.11832442134618759, + 0.062178898602724075, + 0.027421634644269943 + ], + "credit_map": [ + 0.004902668297290802, + 0.003985860385000706, + 0.005775023251771927, + 0.00424339110031724, + 0.0048066070303320885, + 0.003588446881622076, + 0.004481247626245022, + 0.005546828731894493, + 0.004548895638436079, + 0.004218990448862314, + 0.0037587361875921488, + 0.0037869762163609266, + 0.003009234322234988, + 0.0029198862612247467, + 0.003952085506170988, + 0.004431379493325949, + 0.006452202796936035, + 0.013509446755051613, + 0.004638562444597483, + 0.007378558628261089, + 0.01182741392403841, + 0.005741261877119541, + 0.00544744823127985, + 0.007189971394836903, + 0.007078514434397221, + 0.003956265281885862, + 0.003416213905438781, + 0.0019206886645406485, + 0.0038490761071443558, + 0.015727253630757332, + 0.002890508621931076, + 0.004001141991466284, + 0.018443606793880463, + 0.011601854115724564, + 0.019706230610609055, + 0.007186324335634708, + 0.008184175007045269, + 0.008924899622797966, + 0.00532965874299407, + 0.00995539128780365, + 0.009425817057490349, + 0.006557823624461889, + 0.006414053961634636, + 0.00884721614420414, + 0.03858162835240364, + 0.007776265498250723, + 0.009659778326749802, + 0.02218787558376789, + 0.11319146305322647, + 0.021438803523778915, + 0.02808702364563942, + 0.106070376932621, + 0.07008916884660721, + 0.03992884233593941, + 0.026562076061964035, + 0.014943838119506836, + 0.033990807831287384, + 0.006869412027299404, + 0.010300889611244202, + 0.030271437019109726, + 0.014731251634657383, + 0.005993273109197617, + 0.008535576984286308, + 0.007631761021912098, + 0.0072404081001877785, + 0.009248022921383381, + 0.007488924078643322, + 0.004873168189078569, + 0.016476904973387718, + 0.008508953265845776, + 0.003895336529240012, + 0.004447173327207565, + 0.002117492025718093, + 0.0022787051275372505, + 0.005954456981271505, + 0.0012910728110000491, + 0.007719994056969881, + 0.004108823835849762, + 0.0016711998032405972, + 0.0022798869758844376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.6933355331420898 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543358840048313 + ], + "opposite_logprob_tokens": [ + -4.387543201446533 + ], + "logprob_reward": -0.012543358840048313, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543358840048313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18677860498428345, + 0.04729870334267616, + 0.029013164341449738, + 0.05659065395593643, + 0.15871307253837585, + 0.09410335123538971, + 0.1317116618156433, + 0.13017971813678741, + 0.13283173739910126, + 0.032779306173324585 + ], + "credit_map": [ + 0.029857946559786797, + 0.02107342705130577, + 0.01922188326716423, + 0.03361721336841583, + 0.018901091068983078, + 0.016134532168507576, + 0.021515311673283577, + 0.026457197964191437, + 0.007687884848564863, + 0.008668717928230762, + 0.004511801991611719, + 0.005210812669247389, + 0.0044125718995928764, + 0.006958567537367344, + 0.004877403378486633, + 0.004970946814864874, + 0.0031045146752148867, + 0.005150176119059324, + 0.0024953146930783987, + 0.0032086120918393135, + 0.0038375803269445896, + 0.003981405403465033, + 0.0025470005348324776, + 0.004688560031354427, + 0.013132740743458271, + 0.003982884343713522, + 0.0027344003319740295, + 0.0017044333508238196, + 0.0041436259634792805, + 0.02267265133559704, + 0.0035247609484940767, + 0.004695162642747164, + 0.033046573400497437, + 0.018456468358635902, + 0.03409744054079056, + 0.013351643458008766, + 0.014550238847732544, + 0.017924021929502487, + 0.01010191161185503, + 0.017184771597385406, + 0.008039097301661968, + 0.011177151463925838, + 0.0077677019871771336, + 0.009621767327189445, + 0.01591820828616619, + 0.00872380193322897, + 0.006999300327152014, + 0.025856314226984978, + 0.011801932007074356, + 0.00922917015850544, + 0.011528069153428078, + 0.020183805376291275, + 0.04748745635151863, + 0.01779523305594921, + 0.0072616636753082275, + 0.006424326449632645, + 0.0382799468934536, + 0.007404105272144079, + 0.013578275218605995, + 0.04126947745680809, + 0.005479570012539625, + 0.005038049072027206, + 0.009782139211893082, + 0.00934816524386406, + 0.015916980803012848, + 0.019732341170310974, + 0.02101842686533928, + 0.015529109165072441, + 0.019153427332639694, + 0.015604274347424507, + 0.009397079236805439, + 0.0164800938218832, + 0.002656971337273717, + 0.002904701279476285, + 0.005914733745157719, + 0.0014815088361501694, + 0.010451039299368858, + 0.005258184392005205, + 0.0015809694305062294, + 0.002531198551878333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.012543358840048313 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544653378427029 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544653378427029, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544653378427029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07284101098775864, + 0.06573258340358734, + 0.11117091774940491, + 0.06816989183425903, + 0.06819851696491241, + 0.0369376614689827, + 0.359460711479187, + 0.07174646109342575, + 0.10325202345848083, + 0.042490243911743164 + ], + "credit_map": [ + 0.007545387372374535, + 0.007993465289473534, + 0.006411390379071236, + 0.010524371638894081, + 0.016389939934015274, + 0.00794020015746355, + 0.00865665078163147, + 0.007379601243883371, + 0.01153000257909298, + 0.009735173545777798, + 0.0072635686956346035, + 0.007149967830628157, + 0.00752649363130331, + 0.006348081864416599, + 0.007551272865384817, + 0.008628016337752342, + 0.012548482976853848, + 0.020451514050364494, + 0.009875125251710415, + 0.013225998729467392, + 0.022505570203065872, + 0.010813437402248383, + 0.00948288757354021, + 0.012267897836863995, + 0.01367307547479868, + 0.008080669678747654, + 0.0034601339139044285, + 0.002914740703999996, + 0.005066629033535719, + 0.025568068027496338, + 0.004278331063687801, + 0.005128244869410992, + 0.010664151981472969, + 0.013869628310203552, + 0.010724804364144802, + 0.005462603643536568, + 0.010129583068192005, + 0.005783368833363056, + 0.004950748756527901, + 0.006613622419536114, + 0.005105458665639162, + 0.007817746140062809, + 0.0043522147461771965, + 0.004986939486116171, + 0.003519895253702998, + 0.004266624804586172, + 0.00335882930085063, + 0.003529947018250823, + 0.06556598097085953, + 0.022669339552521706, + 0.029911432415246964, + 0.13386228680610657, + 0.048513855785131454, + 0.01372018363326788, + 0.027791200205683708, + 0.01742641255259514, + 0.01876373030245304, + 0.004568864591419697, + 0.0082570631057024, + 0.017839666455984116, + 0.008112345822155476, + 0.0032745087519288063, + 0.004950487986207008, + 0.005979793146252632, + 0.012488508597016335, + 0.015451455488801003, + 0.008363243192434311, + 0.008125434629619122, + 0.031142128631472588, + 0.014122292399406433, + 0.006973631680011749, + 0.006585327442735434, + 0.0029056270141154528, + 0.0037884721532464027, + 0.0058854189701378345, + 0.0017360046040266752, + 0.01544419676065445, + 0.007828662171959877, + 0.0018922275630757213, + 0.0030096310656517744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.012544653378427029 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634849451482296 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008634849451482296, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634849451482296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06543297320604324, + 0.04183434322476387, + 0.03186574578285217, + 0.08989106118679047, + 0.08701186627149582, + 0.05734611302614212, + 0.3236919343471527, + 0.07255536317825317, + 0.04095843806862831, + 0.18941223621368408 + ], + "credit_map": [ + 0.008330199867486954, + 0.008328196592628956, + 0.009707644581794739, + 0.007470167241990566, + 0.008092614822089672, + 0.006126743275672197, + 0.0069721536710858345, + 0.010405243374407291, + 0.0063469065353274345, + 0.005651693791151047, + 0.005459835287183523, + 0.004575539380311966, + 0.0042596617713570595, + 0.00488299410790205, + 0.005220948718488216, + 0.005436762701719999, + 0.00378861790522933, + 0.005833623930811882, + 0.002567155985161662, + 0.004079482052475214, + 0.005794689524918795, + 0.003180735744535923, + 0.0028938420582562685, + 0.003727592760697007, + 0.02246067486703396, + 0.006209089420735836, + 0.0036042784340679646, + 0.0023695281706750393, + 0.0059531270526349545, + 0.03853253275156021, + 0.004971803165972233, + 0.00579002546146512, + 0.020039396360516548, + 0.0076760295778512955, + 0.020882559940218925, + 0.007834752090275288, + 0.008199629373848438, + 0.009435679763555527, + 0.0043877060525119305, + 0.008556102402508259, + 0.009224677458405495, + 0.004588787443935871, + 0.0030061528086662292, + 0.0041510965675115585, + 0.023037023842334747, + 0.005636619403958321, + 0.003501984989270568, + 0.004199767019599676, + 0.08698920905590057, + 0.016919584944844246, + 0.02322820946574211, + 0.09323141723871231, + 0.041092999279499054, + 0.027830179780721664, + 0.02183612994849682, + 0.012564181350171566, + 0.0195308905094862, + 0.004994331859052181, + 0.0063795000314712524, + 0.018284594640135765, + 0.009428155608475208, + 0.003480354556813836, + 0.0050930543802678585, + 0.005364483222365379, + 0.004742856603115797, + 0.0044918907806277275, + 0.0035759401507675648, + 0.00516049237921834, + 0.00903255119919777, + 0.006329166702926159, + 0.004212726838886738, + 0.0034128117840737104, + 0.01042989082634449, + 0.01846924051642418, + 0.025334736332297325, + 0.007791371550410986, + 0.07465637475252151, + 0.03535905480384827, + 0.007749335840344429, + 0.009622226469218731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.008634849451482296 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extremely low-angle tracking shot, positioned just inches from the ground, glides smoothly alongside a severely crumpled lottery ticket as it crosses a weathered concrete sidewalk. The once tightly held ticket is now partially unfolded, its surface riddled with sharp creases and soft wrinkles, the design's plain, faded primary colors barely visible beneath a layer of grime. Harsh, direct sunlight from late afternoon sweeps across the scene, transforming the cracks and gravel texture of the sidewalk into dramatic topography and casting long, sharp shadows from the invisible urban fabric that stretch across the frame like dark fingers. A faint, intermittent breeze, the sole source of movement, causes the ticket to tumble erratically, skittering sideways, occasionally rising delicately into the air, stirring up tiny whirlwinds of fine dust and small, dry leaf fragments that swirl around it. The entire scene is bathed in a warm, golden-orange glow, yet the slow, lonely dance of the discarded ticket across the desolate landscape evokes a potent atmosphere of quiet melancholy and forgotten hope.", + "seed": 9322, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_14/index_515_seed_9322_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.12499214857816696, + "vlm_logprob": -0.12499214857816696, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.12499214857816696, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17552217841148376 + ], + "opposite_logprob_tokens": [ + -1.9255222082138062 + ], + "logprob_reward": -0.17552217841148376, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17552217841148376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25669389963150024, + 0.2957373261451721, + 0.03957221657037735, + 0.1302718222141266, + 0.05423847213387489, + 0.05648414045572281, + 0.038448914885520935, + 0.026508299633860588, + 0.019017189741134644, + 0.0830276757478714 + ], + "credit_map": [ + 0.026034189388155937, + 0.028263675048947334, + 0.04148278385400772, + 0.04172586277127266, + 0.0313820019364357, + 0.0331561453640461, + 0.03214716538786888, + 0.022502074018120766, + 0.017564786598086357, + 0.11308301985263824, + 0.08013281226158142, + 0.025294028222560883, + 0.019108960404992104, + 0.016585681587457657, + 0.014578052796423435, + 0.009389972314238548, + 0.004068498499691486, + 0.004317996092140675, + 0.009416831657290459, + 0.0027200868353247643, + 0.004950335249304771, + 0.007355499546974897, + 0.0032208363991230726, + 0.003522136714309454, + 0.01041486021131277, + 0.038081295788288116, + 0.007212947588413954, + 0.0049878074787557125, + 0.0512874573469162, + 0.010542474687099457, + 0.004857172723859549, + 0.002887803129851818, + 0.004753907211124897, + 0.0057775890454649925, + 0.013287699781358242, + 0.005133467726409435, + 0.010007156990468502, + 0.00661506736651063, + 0.004427612759172916, + 0.0042359670624136925, + 0.007431747857481241, + 0.004783584736287594, + 0.009381079114973545, + 0.007001234218478203, + 0.007155354134738445, + 0.008896511979401112, + 0.006212783511728048, + 0.005621846299618483, + 0.0034512944985181093, + 0.005621870048344135, + 0.004733676556497812, + 0.004616694059222937, + 0.004141520243138075, + 0.005949717015028, + 0.004425553139299154, + 0.00550859235227108, + 0.003601928474381566, + 0.002545645460486412, + 0.0036606364883482456, + 0.0035525434650480747, + 0.003363344818353653, + 0.002712324494495988, + 0.003625748911872506, + 0.0034461277537047863, + 0.0021103157196193933, + 0.0021056816913187504, + 0.0032603940926492214, + 0.0026264344342052937, + 0.0017469661543145776, + 0.0023905299603939056, + 0.002546230098232627, + 0.002230635844171047, + 0.009071149863302708, + 0.00786842592060566, + 0.01251261867582798, + 0.01656811684370041, + 0.007098915055394173, + 0.008652016520500183, + 0.01147853210568428, + 0.009777899831533432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.17552217841148376 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38711249828338623 + ], + "opposite_logprob_tokens": [ + -1.1371124982833862 + ], + "logprob_reward": -0.38711249828338623, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38711249828338623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1114543154835701, + 0.19873416423797607, + 0.037745192646980286, + 0.19631342589855194, + 0.08535011857748032, + 0.11612937599420547, + 0.06096579506993294, + 0.048943161964416504, + 0.09915231168270111, + 0.04521210491657257 + ], + "credit_map": [ + 0.007632798980921507, + 0.011312506161630154, + 0.010412581264972687, + 0.010622266680002213, + 0.0183683093637228, + 0.013534029014408588, + 0.026639342308044434, + 0.01293247938156128, + 0.01209288276731968, + 0.07283218950033188, + 0.05490054190158844, + 0.016920998692512512, + 0.01514154952019453, + 0.013096993789076805, + 0.007656649220734835, + 0.006092374678701162, + 0.003718338906764984, + 0.004667631350457668, + 0.005414996296167374, + 0.0028292108327150345, + 0.005519903730601072, + 0.008959321305155754, + 0.003450088668614626, + 0.0031857003923505545, + 0.014068415388464928, + 0.05620494484901428, + 0.016366012394428253, + 0.0110649224370718, + 0.06683507561683655, + 0.012302402406930923, + 0.012823279947042465, + 0.006648378446698189, + 0.008074509911239147, + 0.008224237710237503, + 0.013349984772503376, + 0.009419163689017296, + 0.013810970820486546, + 0.013307960703969002, + 0.010129556059837341, + 0.009033743292093277, + 0.011667445302009583, + 0.016757341101765633, + 0.018323864787817, + 0.01533111184835434, + 0.015388410538434982, + 0.014473460614681244, + 0.01095595769584179, + 0.013231795281171799, + 0.004018654115498066, + 0.006143244449049234, + 0.006513095460832119, + 0.005339256953448057, + 0.004617366939783096, + 0.013795014470815659, + 0.0049736290238797665, + 0.015565535984933376, + 0.005001704208552837, + 0.005086414981633425, + 0.00603394815698266, + 0.009503287263214588, + 0.00521471630781889, + 0.0052327984012663364, + 0.006954406388103962, + 0.005915890913456678, + 0.00761456647887826, + 0.008129513822495937, + 0.020409904420375824, + 0.009867401793599129, + 0.007244979031383991, + 0.009619268588721752, + 0.025997357442975044, + 0.010269330814480782, + 0.005063771270215511, + 0.004651772323995829, + 0.00628954591229558, + 0.007339083589613438, + 0.003401416353881359, + 0.006577510852366686, + 0.00627095764502883, + 0.005618048831820488 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.38711249828338623 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075171530246735 + ], + "opposite_logprob_tokens": [ + -4.141075134277344 + ], + "logprob_reward": -0.016075171530246735, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075171530246735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3506039083003998, + 0.16138043999671936, + 0.040921419858932495, + 0.07668835669755936, + 0.10806424915790558, + 0.10476968437433243, + 0.029952451586723328, + 0.06549299508333206, + 0.0238838791847229, + 0.038242533802986145 + ], + "credit_map": [ + 0.03149488568305969, + 0.0420990064740181, + 0.053163617849349976, + 0.05723867937922478, + 0.04421525076031685, + 0.03862832114100456, + 0.058544471859931946, + 0.02521970309317112, + 0.009623502381145954, + 0.05524531379342079, + 0.040690965950489044, + 0.014166479930281639, + 0.01119297742843628, + 0.01116547454148531, + 0.011977971531450748, + 0.007317753974348307, + 0.0032681170850992203, + 0.00380394677631557, + 0.0053597185760736465, + 0.005193295422941446, + 0.005400494672358036, + 0.006438825745135546, + 0.00571999279782176, + 0.005737029481679201, + 0.013677229173481464, + 0.011903645470738411, + 0.010905077680945396, + 0.00587727315723896, + 0.010431678034365177, + 0.01343624573200941, + 0.006686043459922075, + 0.003771166317164898, + 0.009310794062912464, + 0.01177854835987091, + 0.026090314611792564, + 0.010779459029436111, + 0.019969308748841286, + 0.013060405850410461, + 0.007469932083040476, + 0.009605484083294868, + 0.010336191393435001, + 0.00896173994988203, + 0.025202829390764236, + 0.009161929599940777, + 0.015521114692091942, + 0.01520185824483633, + 0.009865731932222843, + 0.010518294759094715, + 0.003805687418207526, + 0.0032253852114081383, + 0.004956800024956465, + 0.003069324418902397, + 0.0030377116054296494, + 0.0051948814652860165, + 0.002671850612387061, + 0.003990810830146074, + 0.00673667062073946, + 0.005203417036682367, + 0.01315050944685936, + 0.010691803880035877, + 0.00570992985740304, + 0.008353828452527523, + 0.006844541057944298, + 0.008802293799817562, + 0.0019303102744743228, + 0.003268368309363723, + 0.0033795202616602182, + 0.003760253544896841, + 0.0020698020234704018, + 0.002295165089890361, + 0.004326989874243736, + 0.002853469690307975, + 0.005170437972992659, + 0.0037088242825120687, + 0.004951842594891787, + 0.006434457376599312, + 0.004073415417224169, + 0.004350949078798294, + 0.004932847805321217, + 0.004619760438799858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.016075171530246735 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553717941045761 + ], + "opposite_logprob_tokens": [ + -4.387553691864014 + ], + "logprob_reward": -0.012553717941045761, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553717941045761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07543668895959854, + 0.15865686535835266, + 0.05375128611922264, + 0.44886788725852966, + 0.03887302055954933, + 0.054068539291620255, + 0.0467861145734787, + 0.06894727051258087, + 0.02234121412038803, + 0.032271139323711395 + ], + "credit_map": [ + 0.007582519203424454, + 0.008884552866220474, + 0.009406042285263538, + 0.01124931126832962, + 0.00826972909271717, + 0.013675828464329243, + 0.009671006351709366, + 0.0066977012902498245, + 0.009924540296196938, + 0.05982249602675438, + 0.044302452355623245, + 0.013162734918296337, + 0.01234462857246399, + 0.008287789300084114, + 0.005926642566919327, + 0.004885573871433735, + 0.007861942052841187, + 0.006862188223749399, + 0.015839433297514915, + 0.004403889644891024, + 0.005283969920128584, + 0.00603677099570632, + 0.003523111343383789, + 0.003939982037991285, + 0.03253921866416931, + 0.15062934160232544, + 0.016143176704645157, + 0.012849051505327225, + 0.18501053750514984, + 0.026567500084638596, + 0.015057221055030823, + 0.010071823373436928, + 0.003461419139057398, + 0.004455973394215107, + 0.005668982397764921, + 0.004103363025933504, + 0.0046057733707129955, + 0.006448409054428339, + 0.005880942568182945, + 0.0042481571435928345, + 0.007877299562096596, + 0.005652374122291803, + 0.009290290996432304, + 0.005809751804918051, + 0.007490538991987705, + 0.006958246696740389, + 0.00468285009264946, + 0.006307187024503946, + 0.003710850840434432, + 0.003985336981713772, + 0.005030061583966017, + 0.0034973931033164263, + 0.00357191008515656, + 0.013044910505414009, + 0.0043712337501347065, + 0.009574417024850845, + 0.005806609056890011, + 0.005308555904775858, + 0.011830287985503674, + 0.00990060530602932, + 0.0065171560272574425, + 0.00885824952274561, + 0.011628823354840279, + 0.009096982888877392, + 0.0018138362793251872, + 0.0018719498766586185, + 0.004187366459518671, + 0.0023611101787537336, + 0.001482947263866663, + 0.002104652114212513, + 0.006108482833951712, + 0.0024108681827783585, + 0.004081329796463251, + 0.0031508542597293854, + 0.004140905104577541, + 0.004460452124476433, + 0.0031152009032666683, + 0.004037604667246342, + 0.004844394978135824, + 0.004440398421138525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.012553717941045761 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at an extremely low angle, just inches from the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera performing a smooth tracking shot alongside a severely crumpled lottery ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a weathered concrete sidewalk with visible cracks and gravel texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the lottery ticket partially unfolded with sharp creases, soft wrinkles, and a surface covered in grime?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature harsh, direct late afternoon sunlight casting long, sharp shadows and a warm, golden-orange glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369717672467232 + ], + "opposite_logprob_tokens": [ + -3.4086971282958984 + ], + "logprob_reward": -0.03369717672467232, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369717672467232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22617816925048828, + 0.14919596910476685, + 0.10983462631702423, + 0.06336048245429993, + 0.14519089460372925, + 0.06930845975875854, + 0.027004340663552284, + 0.06103760376572609, + 0.06406795978546143, + 0.08482149988412857 + ], + "credit_map": [ + 0.023679999634623528, + 0.020434750244021416, + 0.03652246668934822, + 0.03412008658051491, + 0.031870316714048386, + 0.03275468572974205, + 0.029600029811263084, + 0.017195841297507286, + 0.008998250588774681, + 0.05558677762746811, + 0.04075312241911888, + 0.014169531874358654, + 0.010530323721468449, + 0.006939845625311136, + 0.007799133658409119, + 0.0044189877808094025, + 0.011497413739562035, + 0.013724224641919136, + 0.012618359178304672, + 0.0100230872631073, + 0.017299901694059372, + 0.014895661734044552, + 0.012410754337906837, + 0.0173652246594429, + 0.0071014659479260445, + 0.012867107056081295, + 0.012494834139943123, + 0.004691122565418482, + 0.01271793432533741, + 0.005288471933454275, + 0.005614951718598604, + 0.0025845938362181187, + 0.012956690974533558, + 0.014584425836801529, + 0.03136910870671272, + 0.014039585366845131, + 0.019403917714953423, + 0.017678827047348022, + 0.014776065945625305, + 0.020382273942232132, + 0.007875761948525906, + 0.00700976001098752, + 0.010062696412205696, + 0.009098519571125507, + 0.009646211750805378, + 0.010280326940119267, + 0.0069968909956514835, + 0.008338293991982937, + 0.0018272515153512359, + 0.001553763053379953, + 0.0024510223884135485, + 0.0017216195119544864, + 0.0016805175691843033, + 0.007754040416330099, + 0.001451617106795311, + 0.008564510382711887, + 0.005991669371724129, + 0.005114191677421331, + 0.008773292414844036, + 0.008367392234504223, + 0.00646716495975852, + 0.008855510503053665, + 0.009719864465296268, + 0.00774851581081748, + 0.004712534602731466, + 0.005902048200368881, + 0.011792031116783619, + 0.006997983902692795, + 0.004780580755323172, + 0.005025009158998728, + 0.017992885783314705, + 0.006864883471280336, + 0.011259155347943306, + 0.010849135927855968, + 0.010782774537801743, + 0.009915675967931747, + 0.008513736538589, + 0.012175079435110092, + 0.009412121959030628, + 0.01191382110118866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.03369717672467232 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_15/index_tensor([9742, 9742])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_15/index_tensor([9742, 9742])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6ff405518e40078937c902f1e1d25d001dd986dd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_15/index_tensor([9742, 9742])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle, slow-tracking shot smoothly glides across a surreal, flooded English garden, where the water, perfectly still and crystal clear, acts as a mirror against a soft, overcast sky. Beneath the surface, a submerged maze of manicured hedges and stone paths looms. The camera gradually descends, following a solitary figure seated in a sleek, dark, high-gloss mahogany canoe. The figure, whose features remain blurred, wears a simple cream-colored linen tunic, their posture poised and contemplative. They row with a steady, deliberate rhythm, gracefully dipping the slender wooden oars into the tranquil water. Each stroke sends gossamer ripples spreading outward, gently nudging the heavy, blossoming heads of the submerged crimson and pink rose bushes. Droplets of water cling like tiny gemstones to the velvety petals of the flowers shattered by the surface. The scene is bathed in a soft, diffuse light that casts no harsh shadows, creating an atmosphere of profound tranquility and melancholic beauty, the vibrant red of the roses contrasting beautifully with the deep green of the underwater foliage and the dark wood of the boat.", + "seed": 18006, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_15/index_9742_seed_18006_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04909320045262575, + "vlm_logprob": -0.04909320045262575, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.04909320045262575, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03442949801683426 + ], + "opposite_logprob_tokens": [ + -3.5344295501708984 + ], + "logprob_reward": -0.03442949801683426, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03442949801683426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2838054299354553, + 0.12673141062259674, + 0.048121023923158646, + 0.04520658403635025, + 0.052596453577280045, + 0.03875497356057167, + 0.056281477212905884, + 0.11962941288948059, + 0.10905010998249054, + 0.11982318758964539 + ], + "credit_map": [ + 0.038001902401447296, + 0.029865730553865433, + 0.047431740909814835, + 0.025983409956097603, + 0.031196853145956993, + 0.030969804152846336, + 0.042821332812309265, + 0.03753466159105301, + 0.010747716762125492, + 0.010408003814518452, + 0.009414155036211014, + 0.009990178979933262, + 0.03705967962741852, + 0.012740583159029484, + 0.028584660962224007, + 0.007786439266055822, + 0.005211316514760256, + 0.005122450180351734, + 0.005997838918119669, + 0.0062621477991342545, + 0.006655014120042324, + 0.00651409151032567, + 0.0069625480100512505, + 0.005395617336034775, + 0.004197827540338039, + 0.006435397081077099, + 0.004953611642122269, + 0.0033714021556079388, + 0.006657033693045378, + 0.008359456434845924, + 0.005995221436023712, + 0.005236634984612465, + 0.005654654931277037, + 0.006956953555345535, + 0.004830371588468552, + 0.004027184098958969, + 0.006042989436537027, + 0.010670436546206474, + 0.006662868428975344, + 0.007750995457172394, + 0.00660208985209465, + 0.004349537659436464, + 0.0037466869689524174, + 0.0032580916304141283, + 0.005351031199097633, + 0.00557402940467, + 0.004521866329014301, + 0.00535163888707757, + 0.006431151181459427, + 0.0049322424456477165, + 0.004685867577791214, + 0.006420646794140339, + 0.008001779206097126, + 0.012385028414428234, + 0.004220794420689344, + 0.009203968569636345, + 0.011655603535473347, + 0.02289857715368271, + 0.007191954180598259, + 0.011842070147395134, + 0.022346720099449158, + 0.017234938219189644, + 0.014478985220193863, + 0.011980561539530754, + 0.012921856716275215, + 0.014551389031112194, + 0.016750479117035866, + 0.0157995093613863, + 0.013021157123148441, + 0.012624779716134071, + 0.010385547764599323, + 0.012995388358831406, + 0.013506856746971607, + 0.020903661847114563, + 0.015844568610191345, + 0.02276061661541462, + 0.013569105416536331, + 0.008495025336742401, + 0.01022021472454071, + 0.014523142948746681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.03442949801683426 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020608583465218544 + ], + "opposite_logprob_tokens": [ + -3.89560866355896 + ], + "logprob_reward": -0.020608583465218544, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020608583465218544, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14455696940422058, + 0.11731834709644318, + 0.1049884632229805, + 0.059440068900585175, + 0.013151098974049091, + 0.044797658920288086, + 0.07606028020381927, + 0.07415726035833359, + 0.17640724778175354, + 0.18912261724472046 + ], + "credit_map": [ + 0.018053637817502022, + 0.015593506395816803, + 0.02625148743391037, + 0.014928923919796944, + 0.013602318242192268, + 0.015306180343031883, + 0.020769260823726654, + 0.02005165070295334, + 0.013035205192863941, + 0.017529716715216637, + 0.010732237249612808, + 0.016986949369311333, + 0.02029081992805004, + 0.010550144128501415, + 0.015773795545101166, + 0.012419474311172962, + 0.012773292139172554, + 0.010303160175681114, + 0.00990037526935339, + 0.009632740169763565, + 0.013289546594023705, + 0.018960682675242424, + 0.011585719883441925, + 0.018542945384979248, + 0.004668483044952154, + 0.0053990562446415424, + 0.003039652481675148, + 0.0032467516139149666, + 0.007878344506025314, + 0.017586009576916695, + 0.0059481472708284855, + 0.011673619970679283, + 0.0015228529227897525, + 0.0016517098993062973, + 0.0014443148393183947, + 0.0019125367980450392, + 0.0014115982921794057, + 0.001960235182195902, + 0.0017267377115786076, + 0.0015211139107123017, + 0.008829283528029919, + 0.00494051119312644, + 0.006165634840726852, + 0.003808982903137803, + 0.004510731901973486, + 0.006570146884769201, + 0.004377155564725399, + 0.005595211870968342, + 0.007893524132668972, + 0.007832072675228119, + 0.006710444577038288, + 0.007948353886604309, + 0.010216661728918552, + 0.01636805199086666, + 0.007009470369666815, + 0.012081698514521122, + 0.007545525673776865, + 0.011763718910515308, + 0.004603262525051832, + 0.007392833475023508, + 0.012519573792815208, + 0.016405848786234856, + 0.007743620313704014, + 0.006182877346873283, + 0.020559361204504967, + 0.022209294140338898, + 0.03000841848552227, + 0.024472158402204514, + 0.02006629668176174, + 0.01809786446392536, + 0.01661325804889202, + 0.024380594491958618, + 0.023716844618320465, + 0.037847090512514114, + 0.02546614781022072, + 0.03329847380518913, + 0.021793222054839134, + 0.011594974435865879, + 0.014893079176545143, + 0.02051277458667755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.020608583465218544 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011107279919087887 + ], + "opposite_logprob_tokens": [ + -4.511107444763184 + ], + "logprob_reward": -0.011107279919087887, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011107279919087887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08052773773670197, + 0.04232014715671539, + 0.0954912081360817, + 0.03442957252264023, + 0.04246128723025322, + 0.035579148679971695, + 0.2108139544725418, + 0.20148496329784393, + 0.0751902386546135, + 0.18170173466205597 + ], + "credit_map": [ + 0.009395170956850052, + 0.01096317172050476, + 0.011766398325562477, + 0.008214091882109642, + 0.009518281556665897, + 0.01046949252486229, + 0.010119096376001835, + 0.010082035325467587, + 0.0037858071736991405, + 0.004814316984266043, + 0.003197690239176154, + 0.004772183019667864, + 0.010850793682038784, + 0.00504447054117918, + 0.00392682570964098, + 0.005928057711571455, + 0.01239323616027832, + 0.0115975895896554, + 0.008285409770905972, + 0.009776762686669827, + 0.012000162154436111, + 0.015526990406215191, + 0.011173349805176258, + 0.01473770011216402, + 0.0028580257203429937, + 0.0035443375818431377, + 0.004394975956529379, + 0.004385497886687517, + 0.0053670951165258884, + 0.0061364201828837395, + 0.004286416806280613, + 0.0034568007104098797, + 0.004063124302774668, + 0.004417911171913147, + 0.004721327684819698, + 0.003548920853063464, + 0.0052323974668979645, + 0.008004466071724892, + 0.006820063106715679, + 0.005653079599142075, + 0.0048973592929542065, + 0.0053992182947695255, + 0.003547952277585864, + 0.004464092198759317, + 0.004876133054494858, + 0.004211327992379665, + 0.00386384897865355, + 0.004319217521697283, + 0.04280495271086693, + 0.02234750986099243, + 0.022771719843149185, + 0.019771220162510872, + 0.02474035881459713, + 0.027289463207125664, + 0.015202355571091175, + 0.035886362195014954, + 0.019764943048357964, + 0.04020652920007706, + 0.01169629767537117, + 0.01853768713772297, + 0.038440462201833725, + 0.031211216002702713, + 0.02314905822277069, + 0.01847878098487854, + 0.008682015351951122, + 0.010552267543971539, + 0.01251230575144291, + 0.009721284732222557, + 0.008242466486990452, + 0.007566355634480715, + 0.006556981708854437, + 0.011356563307344913, + 0.021379409357905388, + 0.04425375908613205, + 0.02182542346417904, + 0.025731870904564857, + 0.02300090156495571, + 0.015690680593252182, + 0.013193085789680481, + 0.016626602038741112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.011107279919087887 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003280058503151 + ], + "opposite_logprob_tokens": [ + -2.350327968597412 + ], + "logprob_reward": -0.1003280058503151, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003280058503151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0900368019938469, + 0.07395780086517334, + 0.06151221692562103, + 0.05731940269470215, + 0.05257029831409454, + 0.04743610695004463, + 0.09997989982366562, + 0.32967090606689453, + 0.13522273302078247, + 0.0522937998175621 + ], + "credit_map": [ + 0.0099550262093544, + 0.00990346446633339, + 0.01604325883090496, + 0.010696490295231342, + 0.010468624532222748, + 0.009316357783973217, + 0.012070205993950367, + 0.011583377607166767, + 0.006696383468806744, + 0.006969472859054804, + 0.005590712185949087, + 0.010171199217438698, + 0.009607811458408833, + 0.005509238690137863, + 0.019838839769363403, + 0.00957415159791708, + 0.00545093696564436, + 0.006337577942758799, + 0.005307350307703018, + 0.006927157286554575, + 0.011074270121753216, + 0.010901698842644691, + 0.00833172257989645, + 0.007181508932262659, + 0.008077352307736874, + 0.009872264228761196, + 0.007020261138677597, + 0.0041341218166053295, + 0.006785796955227852, + 0.008223076350986958, + 0.007915288209915161, + 0.005291247274726629, + 0.005610353779047728, + 0.005337937269359827, + 0.005518246442079544, + 0.004276349674910307, + 0.006381699349731207, + 0.010490380227565765, + 0.008452880196273327, + 0.006502452772110701, + 0.007262811996042728, + 0.004868337418884039, + 0.005606492515653372, + 0.0076478454284369946, + 0.004791506100445986, + 0.00652112765237689, + 0.004793415777385235, + 0.005944572854787111, + 0.010391422547399998, + 0.009300497360527515, + 0.010342434048652649, + 0.012446013279259205, + 0.014770803973078728, + 0.018380623310804367, + 0.008613893762230873, + 0.015734214335680008, + 0.029265090823173523, + 0.06965793669223785, + 0.021391138434410095, + 0.028634846210479736, + 0.06717907637357712, + 0.04322702810168266, + 0.03588239103555679, + 0.034433409571647644, + 0.01718594692647457, + 0.023295802995562553, + 0.0198532547801733, + 0.014649580232799053, + 0.016280509531497955, + 0.013897370547056198, + 0.012505416758358479, + 0.01755484752357006, + 0.006792270112782717, + 0.010740699246525764, + 0.006502156145870686, + 0.010004434734582901, + 0.006264272145926952, + 0.0036546410992741585, + 0.004083066247403622, + 0.0042522624135017395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.1003280058503151 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899263501167297 + ], + "opposite_logprob_tokens": [ + -2.5789926052093506 + ], + "logprob_reward": -0.07899263501167297, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899263501167297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13399773836135864, + 0.06782498955726624, + 0.03870310261845589, + 0.052974194288253784, + 0.06849676370620728, + 0.10313820093870163, + 0.13130231201648712, + 0.14443665742874146, + 0.17369936406612396, + 0.08542664349079132 + ], + "credit_map": [ + 0.016395829617977142, + 0.014238441362977028, + 0.01780891790986061, + 0.015352493152022362, + 0.018132196739315987, + 0.013520486652851105, + 0.019984262064099312, + 0.0185650996863842, + 0.0056690191850066185, + 0.007533067371696234, + 0.004830819554626942, + 0.008844315074384212, + 0.0106315016746521, + 0.008083188906311989, + 0.011790139600634575, + 0.010442938655614853, + 0.00602692412212491, + 0.003152805380523205, + 0.00356120802462101, + 0.002939337631687522, + 0.0075223143212497234, + 0.005894417874515057, + 0.003870354499667883, + 0.005735739599913359, + 0.004948105663061142, + 0.008181335404515266, + 0.004231528844684362, + 0.004654090851545334, + 0.009580772370100021, + 0.008527912199497223, + 0.006967170629650354, + 0.005883278325200081, + 0.006928711663931608, + 0.007365136872977018, + 0.0074224709533154964, + 0.007713499944657087, + 0.009140275418758392, + 0.012203683145344257, + 0.008583270013332367, + 0.009139714762568474, + 0.0136155029758811, + 0.010118929669260979, + 0.008057958446443081, + 0.010128788650035858, + 0.01729782670736313, + 0.014788306318223476, + 0.016310017555952072, + 0.01282086968421936, + 0.016909195110201836, + 0.014779712073504925, + 0.008812621235847473, + 0.013319995254278183, + 0.022990381345152855, + 0.02035871520638466, + 0.011479925364255905, + 0.022651782259345055, + 0.015185708180069923, + 0.027404308319091797, + 0.009265678934752941, + 0.0119945602491498, + 0.027783842757344246, + 0.023503068834543228, + 0.015634968876838684, + 0.013664519414305687, + 0.022788409143686295, + 0.035923559218645096, + 0.016880139708518982, + 0.021155856549739838, + 0.020032167434692383, + 0.019595252349972725, + 0.018260890617966652, + 0.019063087180256844, + 0.011131939478218555, + 0.014665454626083374, + 0.010258306749165058, + 0.013418419286608696, + 0.009670428931713104, + 0.00852635782212019, + 0.007567541673779488, + 0.010188188403844833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.07899263501167297 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, slow-tracking shot smoothly glides across a surreal, flooded English garden, where the water, perfectly still and crystal clear, acts as a mirror against a soft, overcast sky. Beneath the surface, a submerged maze of manicured hedges and stone paths looms. The camera gradually descends, following a solitary figure seated in a sleek, dark, high-gloss mahogany canoe. The figure, whose features remain blurred, wears a simple cream-colored linen tunic, their posture poised and contemplative. They row with a steady, deliberate rhythm, gracefully dipping the slender wooden oars into the tranquil water. Each stroke sends gossamer ripples spreading outward, gently nudging the heavy, blossoming heads of the submerged crimson and pink rose bushes. Droplets of water cling like tiny gemstones to the velvety petals of the flowers shattered by the surface. The scene is bathed in a soft, diffuse light that casts no harsh shadows, creating an atmosphere of profound tranquility and melancholic beauty, the vibrant red of the roses contrasting beautifully with the deep green of the underwater foliage and the dark wood of the boat.", + "seed": 18007, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_15/index_9742_seed_18007_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03225352428853512, + "vlm_logprob": -0.03225352428853512, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.03225352428853512, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05140351504087448 + ], + "opposite_logprob_tokens": [ + -3.051403522491455 + ], + "logprob_reward": -0.05140351504087448, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05140351504087448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10532309114933014, + 0.18493938446044922, + 0.09079064428806305, + 0.03617232292890549, + 0.05921207368373871, + 0.10108737647533417, + 0.08764529228210449, + 0.18144822120666504, + 0.06619532406330109, + 0.08718625456094742 + ], + "credit_map": [ + 0.013321669772267342, + 0.008526579476892948, + 0.019576745107769966, + 0.010526922531425953, + 0.014120402745902538, + 0.01889505609869957, + 0.008961754851043224, + 0.011393962427973747, + 0.049955639988183975, + 0.013959569856524467, + 0.015899518504738808, + 0.013078205287456512, + 0.01580796018242836, + 0.0420425646007061, + 0.017599696293473244, + 0.016596226021647453, + 0.015304961241781712, + 0.015760494396090508, + 0.012236058712005615, + 0.007984827272593975, + 0.010081889107823372, + 0.010588457807898521, + 0.007898573763668537, + 0.010935381054878235, + 0.004852257668972015, + 0.003945596050471067, + 0.00588770629838109, + 0.004269902128726244, + 0.0040209596045315266, + 0.005342039279639721, + 0.0029485521372407675, + 0.004905308596789837, + 0.010381774976849556, + 0.007806053850799799, + 0.007216406986117363, + 0.005004964768886566, + 0.006647207774221897, + 0.008315767161548138, + 0.005366299767047167, + 0.008473597466945648, + 0.017732292413711548, + 0.021177606657147408, + 0.008272497914731503, + 0.009135288186371326, + 0.00938720814883709, + 0.013244381174445152, + 0.008815051056444645, + 0.013323056511580944, + 0.007881570607423782, + 0.030848441645503044, + 0.02715824730694294, + 0.0036602893378585577, + 0.003228672780096531, + 0.0025577445048838854, + 0.004831143654882908, + 0.007479182444512844, + 0.0467199869453907, + 0.014014397747814655, + 0.0731702446937561, + 0.00916384719312191, + 0.008622021414339542, + 0.007101043127477169, + 0.008251732215285301, + 0.014404958114027977, + 0.014928232878446579, + 0.0058052390813827515, + 0.0077238865196704865, + 0.004200725816190243, + 0.0044716643169522285, + 0.004379832651466131, + 0.00470733130350709, + 0.019978417083621025, + 0.008732261136174202, + 0.011186218820512295, + 0.010738374665379524, + 0.013315379619598389, + 0.009850580245256424, + 0.008020657114684582, + 0.013444853946566582, + 0.011897936463356018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.05140351504087448 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026371274143457413 + ], + "opposite_logprob_tokens": [ + -3.6513712406158447 + ], + "logprob_reward": -0.026371274143457413, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026371274143457413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1234939694404602, + 0.050202708691358566, + 0.06309909373521805, + 0.08446406573057175, + 0.017974086105823517, + 0.06320147216320038, + 0.035490747541189194, + 0.38358256220817566, + 0.11932457983493805, + 0.059166718274354935 + ], + "credit_map": [ + 0.014473328366875648, + 0.010518941096961498, + 0.016895495355129242, + 0.012848368845880032, + 0.01997273787856102, + 0.017911816015839577, + 0.01798742264509201, + 0.012885860167443752, + 0.006416807416826487, + 0.00650023203343153, + 0.006830895785242319, + 0.005215589888393879, + 0.007077137473970652, + 0.005638510454446077, + 0.005753315519541502, + 0.0067702182568609715, + 0.008740440011024475, + 0.006726512685418129, + 0.006898227147758007, + 0.005745498929172754, + 0.007912734523415565, + 0.010575609281659126, + 0.009133348241448402, + 0.0073667289689183235, + 0.006944866385310888, + 0.007269113324582577, + 0.00953010842204094, + 0.008089683949947357, + 0.013653187081217766, + 0.016209008172154427, + 0.012209794484078884, + 0.010558304376900196, + 0.00309500633738935, + 0.0028993680607527494, + 0.0019124719547107816, + 0.0013698943657800555, + 0.001719121471978724, + 0.0024341498501598835, + 0.0019589411094784737, + 0.0025851326063275337, + 0.009746847674250603, + 0.00820440985262394, + 0.006155938841402531, + 0.005615546368062496, + 0.006776140537112951, + 0.010057276114821434, + 0.008252490311861038, + 0.008392823860049248, + 0.004460729192942381, + 0.008614752441644669, + 0.008334769867360592, + 0.002002718625590205, + 0.002216146094724536, + 0.0035603095311671495, + 0.0036158154252916574, + 0.002685511251911521, + 0.09949100017547607, + 0.030719269067049026, + 0.15115365386009216, + 0.024018265306949615, + 0.01651042327284813, + 0.014823691919445992, + 0.019145095720887184, + 0.027721168473362923, + 0.026139643043279648, + 0.009902208112180233, + 0.008202929981052876, + 0.006421387195587158, + 0.007616438437253237, + 0.008677163161337376, + 0.01420403178781271, + 0.0381607860326767, + 0.0073221223428845406, + 0.008686709217727184, + 0.008541054092347622, + 0.009243047796189785, + 0.005417320411652327, + 0.004880872089415789, + 0.008821429684758186, + 0.006254165433347225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.026371274143457413 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609640382230282 + ], + "opposite_logprob_tokens": [ + -4.141096591949463 + ], + "logprob_reward": -0.01609640382230282, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609640382230282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14344534277915955, + 0.09709704667329788, + 0.12545371055603027, + 0.04986345395445824, + 0.02598286233842373, + 0.07269487529993057, + 0.054193489253520966, + 0.20927011966705322, + 0.0737125352025032, + 0.14828655123710632 + ], + "credit_map": [ + 0.01976151019334793, + 0.01137515902519226, + 0.014338439330458641, + 0.014539222232997417, + 0.02105090022087097, + 0.028248244896531105, + 0.01921231858432293, + 0.014919552020728588, + 0.014235633425414562, + 0.006662360392510891, + 0.01277051493525505, + 0.016517214477062225, + 0.013701463118195534, + 0.011871370486915112, + 0.011255156248807907, + 0.010083328932523727, + 0.014749804511666298, + 0.012129259295761585, + 0.015542419627308846, + 0.012772132642567158, + 0.027997201308608055, + 0.013571737334132195, + 0.014513062313199043, + 0.014178094454109669, + 0.003926299978047609, + 0.005141322035342455, + 0.006113081239163876, + 0.005034439731389284, + 0.00977990124374628, + 0.008662156760692596, + 0.00589007930830121, + 0.005316169001162052, + 0.0055210343562066555, + 0.0041634137742221355, + 0.0036523870658129454, + 0.0018896845867857337, + 0.0021342369727790356, + 0.0027932520024478436, + 0.0026013904716819525, + 0.0032274648547172546, + 0.012491816654801369, + 0.013528215698897839, + 0.0068093277513980865, + 0.005246688611805439, + 0.00830821506679058, + 0.0098394975066185, + 0.007215958554297686, + 0.009255161508917809, + 0.003985138610005379, + 0.018612314015626907, + 0.015689514577388763, + 0.0025026395451277494, + 0.003078611334785819, + 0.002950792433694005, + 0.003567129373550415, + 0.003807348432019353, + 0.054656852036714554, + 0.01636338420212269, + 0.08356896787881851, + 0.012041199952363968, + 0.008447309024631977, + 0.007795802783221006, + 0.010465439409017563, + 0.015931162983179092, + 0.012258229777216911, + 0.009051511064171791, + 0.007674605585634708, + 0.0047832876443862915, + 0.004612290766090155, + 0.006188022438436747, + 0.009247081354260445, + 0.01989750564098358, + 0.0238138809800148, + 0.023080680519342422, + 0.01628163270652294, + 0.02016507461667061, + 0.014681640081107616, + 0.013122747652232647, + 0.02000533789396286, + 0.01713557168841362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.01609640382230282 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370121121406555 + ], + "opposite_logprob_tokens": [ + -3.408701181411743 + ], + "logprob_reward": -0.03370121121406555, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370121121406555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0747959166765213, + 0.05862167850136757, + 0.09889387339353561, + 0.05462504178285599, + 0.11220746487379074, + 0.060421016067266464, + 0.21262788772583008, + 0.107333704829216, + 0.09796033054590225, + 0.12251310795545578 + ], + "credit_map": [ + 0.0111221419647336, + 0.006272932514548302, + 0.014346200972795486, + 0.0073113529942929745, + 0.007403889670968056, + 0.010112117975950241, + 0.009889071807265282, + 0.008338210172951221, + 0.009753379039466381, + 0.004365946631878614, + 0.005619732663035393, + 0.0075945365242660046, + 0.006987540051341057, + 0.012418953701853752, + 0.0065069920383393764, + 0.005374597385525703, + 0.018409475684165955, + 0.019083524122834206, + 0.012913558632135391, + 0.007771780714392662, + 0.008206959813833237, + 0.012998812831938267, + 0.007291567046195269, + 0.012218194082379341, + 0.005291558802127838, + 0.0066236951388418674, + 0.00836181454360485, + 0.005828524939715862, + 0.01010117493569851, + 0.006200382020324469, + 0.005132229067385197, + 0.007085664197802544, + 0.0198073573410511, + 0.0156377162784338, + 0.011945116333663464, + 0.010282326489686966, + 0.013834581710398197, + 0.015675116330385208, + 0.012431089766323566, + 0.012594162486493587, + 0.01069982536137104, + 0.009305468760430813, + 0.007109754718840122, + 0.005632556043565273, + 0.006022116169333458, + 0.0077109988778829575, + 0.006776455324143171, + 0.007163836620748043, + 0.01628037914633751, + 0.07704196870326996, + 0.06860887259244919, + 0.008634518831968307, + 0.010650156997144222, + 0.006572794169187546, + 0.010262098163366318, + 0.014577099122107029, + 0.030885828658938408, + 0.005864256992936134, + 0.044295694679021835, + 0.004955371841788292, + 0.004308826290071011, + 0.004281854256987572, + 0.005161709152162075, + 0.007580171339213848, + 0.01292804628610611, + 0.0150937270373106, + 0.012790258042514324, + 0.01613304391503334, + 0.006245079915970564, + 0.007388683967292309, + 0.009695782326161861, + 0.01768570765852928, + 0.018405450507998466, + 0.019367355853319168, + 0.013359970413148403, + 0.01801242120563984, + 0.011754642240703106, + 0.01094721257686615, + 0.01695306971669197, + 0.013712991960346699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.03370121121406555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369521722197533 + ], + "opposite_logprob_tokens": [ + -3.4086952209472656 + ], + "logprob_reward": -0.03369521722197533, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369521722197533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048044852912425995, + 0.06028910353779793, + 0.03127294406294823, + 0.021352799609303474, + 0.01669497974216938, + 0.09632539749145508, + 0.11761272698640823, + 0.12694573402404785, + 0.14177799224853516, + 0.339683473110199 + ], + "credit_map": [ + 0.008196882903575897, + 0.004642628598958254, + 0.005305691156536341, + 0.004013415891677141, + 0.005277006886899471, + 0.008963269181549549, + 0.004757754039019346, + 0.006888200528919697, + 0.013432689942419529, + 0.004040860570967197, + 0.008045916445553303, + 0.005300778895616531, + 0.005804428830742836, + 0.011361637152731419, + 0.0064735389314591885, + 0.005829249508678913, + 0.005058908835053444, + 0.003924441523849964, + 0.003943428862839937, + 0.0029133390635252, + 0.003636088455095887, + 0.005016875918954611, + 0.002996552037075162, + 0.0037833047099411488, + 0.002361094579100609, + 0.0028204498812556267, + 0.003510050242766738, + 0.00266688852570951, + 0.0027287385892122984, + 0.002196029294282198, + 0.0024035314563661814, + 0.002666014712303877, + 0.0026164541486650705, + 0.002035082085058093, + 0.0020503064151853323, + 0.0018565321806818247, + 0.0021899379789829254, + 0.0025631438475102186, + 0.0015894845128059387, + 0.0017940382240340114, + 0.018945297226309776, + 0.01796548441052437, + 0.009694519452750683, + 0.0061689261347055435, + 0.006386460270732641, + 0.012175600044429302, + 0.01155292708426714, + 0.013436167500913143, + 0.009924931451678276, + 0.04372015222907066, + 0.0387449748814106, + 0.0045955851674079895, + 0.0042662532068789005, + 0.002991541288793087, + 0.0053756278939545155, + 0.007993659004569054, + 0.029037881642580032, + 0.00900937058031559, + 0.04103701189160347, + 0.01502696331590414, + 0.008967938832938671, + 0.005756288301199675, + 0.00868895836174488, + 0.009421315044164658, + 0.016549916937947273, + 0.020517775788903236, + 0.03097677230834961, + 0.019889594987034798, + 0.008881889283657074, + 0.010250020772218704, + 0.018265556544065475, + 0.01644645258784294, + 0.061674509197473526, + 0.059129051864147186, + 0.04225882515311241, + 0.04354919120669365, + 0.027392109856009483, + 0.02542145363986492, + 0.047127414494752884, + 0.03313089534640312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.03369521722197533 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, slow-tracking shot smoothly glides across a surreal, flooded English garden, where the water, perfectly still and crystal clear, acts as a mirror against a soft, overcast sky. Beneath the surface, a submerged maze of manicured hedges and stone paths looms. The camera gradually descends, following a solitary figure seated in a sleek, dark, high-gloss mahogany canoe. The figure, whose features remain blurred, wears a simple cream-colored linen tunic, their posture poised and contemplative. They row with a steady, deliberate rhythm, gracefully dipping the slender wooden oars into the tranquil water. Each stroke sends gossamer ripples spreading outward, gently nudging the heavy, blossoming heads of the submerged crimson and pink rose bushes. Droplets of water cling like tiny gemstones to the velvety petals of the flowers shattered by the surface. The scene is bathed in a soft, diffuse light that casts no harsh shadows, creating an atmosphere of profound tranquility and melancholic beauty, the vibrant red of the roses contrasting beautifully with the deep green of the underwater foliage and the dark wood of the boat.", + "seed": 18008, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_15/index_9742_seed_18008_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03622807525098324, + "vlm_logprob": -0.03622807525098324, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.03622807525098324, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04436904937028885 + ], + "opposite_logprob_tokens": [ + -3.1693689823150635 + ], + "logprob_reward": -0.04436904937028885, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04436904937028885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16456936299800873, + 0.1706773340702057, + 0.200150266289711, + 0.03465474024415016, + 0.03128788620233536, + 0.08086470514535904, + 0.01983669400215149, + 0.21651852130889893, + 0.03648567944765091, + 0.04495485499501228 + ], + "credit_map": [ + 0.026035001501441002, + 0.017549991607666016, + 0.01938081718981266, + 0.016557423397898674, + 0.021348146721720695, + 0.020661043003201485, + 0.02420930005609989, + 0.01882763020694256, + 0.021135244518518448, + 0.02026277780532837, + 0.017073962837457657, + 0.01425281073898077, + 0.031323034316301346, + 0.02237642928957939, + 0.024080561473965645, + 0.020172521471977234, + 0.05222994461655617, + 0.020068202167749405, + 0.005625603254884481, + 0.008782705292105675, + 0.011678067035973072, + 0.07926847785711288, + 0.01460647489875555, + 0.007890787906944752, + 0.0035514063201844692, + 0.004446482751518488, + 0.0031659319065511227, + 0.0032133287750184536, + 0.00759906368330121, + 0.004040377214550972, + 0.003323216224089265, + 0.005314934067428112, + 0.005156330764293671, + 0.002317926613613963, + 0.003741987282410264, + 0.003044811775907874, + 0.003440352389588952, + 0.004086635075509548, + 0.0037124718073755503, + 0.005787367932498455, + 0.007384075783193111, + 0.0060269529931247234, + 0.00946851447224617, + 0.010840637609362602, + 0.011763646267354488, + 0.009310062974691391, + 0.017451908439397812, + 0.008618902415037155, + 0.0017436505295336246, + 0.0018937054555863142, + 0.002614029450342059, + 0.0031261553522199392, + 0.0022464615758508444, + 0.0020472656469792128, + 0.0036141234450042248, + 0.0025513023138046265, + 0.014986623078584671, + 0.011755839921534061, + 0.01290073897689581, + 0.08126220852136612, + 0.009186587296426296, + 0.011343455873429775, + 0.017342306673526764, + 0.05774076282978058, + 0.005876501556485891, + 0.0043390714563429356, + 0.003569243708625436, + 0.0037581261713057756, + 0.005552950780838728, + 0.004601927474141121, + 0.005153831094503403, + 0.0036340265069156885, + 0.007526646833866835, + 0.005065376404672861, + 0.003919954877346754, + 0.003907104022800922, + 0.006666117813438177, + 0.004945287946611643, + 0.006900016218423843, + 0.006024348083883524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.04436904937028885 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598305389285088 + ], + "opposite_logprob_tokens": [ + -3.8955984115600586 + ], + "logprob_reward": -0.020598305389285088, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598305389285088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07275374978780746, + 0.08771100640296936, + 0.24073877930641174, + 0.059316832572221756, + 0.06414848566055298, + 0.03448841720819473, + 0.024037865921854973, + 0.13413318991661072, + 0.22871585190296173, + 0.05395585298538208 + ], + "credit_map": [ + 0.010783644393086433, + 0.00721572944894433, + 0.007289186120033264, + 0.008417638018727303, + 0.010388286784291267, + 0.009083692915737629, + 0.01102596614509821, + 0.00854959711432457, + 0.011104363948106766, + 0.008637825958430767, + 0.010276319459080696, + 0.011333230882883072, + 0.00948372669517994, + 0.011594148352742195, + 0.013488458469510078, + 0.011792927049100399, + 0.021473009139299393, + 0.055993009358644485, + 0.017910992726683617, + 0.028683101758360863, + 0.0219801627099514, + 0.0580286830663681, + 0.02361227013170719, + 0.013057532720267773, + 0.0057317703031003475, + 0.0067818379029631615, + 0.007514881901443005, + 0.007136608939617872, + 0.012944594025611877, + 0.006057536229491234, + 0.006044583860784769, + 0.007105018012225628, + 0.01112127210944891, + 0.004968640394508839, + 0.006907325237989426, + 0.005782357417047024, + 0.01924317516386509, + 0.006364405620843172, + 0.005302070640027523, + 0.004459234420210123, + 0.004434517119079828, + 0.0027543387841433287, + 0.0034542668145149946, + 0.004540803376585245, + 0.004449713043868542, + 0.003906240686774254, + 0.006404634565114975, + 0.004543900489807129, + 0.002535642357543111, + 0.0023713987320661545, + 0.0027050746139138937, + 0.0032528773881495, + 0.004390530288219452, + 0.002802298404276371, + 0.0030288200359791517, + 0.0029512217734009027, + 0.00898099783807993, + 0.00900206994265318, + 0.009000959806144238, + 0.04612008109688759, + 0.007066636811941862, + 0.007342533208429813, + 0.012601243332028389, + 0.03401864692568779, + 0.04231686145067215, + 0.023866713047027588, + 0.02388058975338936, + 0.023679319769144058, + 0.029922310262918472, + 0.02776212804019451, + 0.034945033490657806, + 0.022342873737215996, + 0.007816790603101254, + 0.005256251432001591, + 0.0042751203291118145, + 0.006100302096456289, + 0.009344913065433502, + 0.006375663448125124, + 0.007018072064965963, + 0.00776874041184783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.020598305389285088 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012564077973365784 + ], + "opposite_logprob_tokens": [ + -4.387564182281494 + ], + "logprob_reward": -0.012564077973365784, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012564077973365784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.049200765788555145, + 0.03666296601295471, + 0.4029103219509125, + 0.021669425070285797, + 0.04029015451669693, + 0.054013464599847794, + 0.032440945506095886, + 0.17127984762191772, + 0.1666412055492401, + 0.02489095740020275 + ], + "credit_map": [ + 0.006522543728351593, + 0.005953473504632711, + 0.0052738552913069725, + 0.006736230570822954, + 0.006157136522233486, + 0.005883532576262951, + 0.0058731622993946075, + 0.0068008266389369965, + 0.004183424171060324, + 0.003519432619214058, + 0.0028123483061790466, + 0.003704102709889412, + 0.004805425181984901, + 0.005741151515394449, + 0.0066571966744959354, + 0.005239882972091436, + 0.0782819539308548, + 0.06804295629262924, + 0.018765978515148163, + 0.026326535269618034, + 0.026115285232663155, + 0.14304903149604797, + 0.027503877878189087, + 0.014824675396084785, + 0.0019641623366624117, + 0.00266034877859056, + 0.0025916346348822117, + 0.0027755340561270714, + 0.004414600320160389, + 0.0023003399837762117, + 0.0026882842648774385, + 0.0022745197638869286, + 0.004000535234808922, + 0.0022048642858862877, + 0.003221957478672266, + 0.003173590637743473, + 0.014941813424229622, + 0.0024836882948875427, + 0.0036453239154070616, + 0.0066183763556182384, + 0.007964219897985458, + 0.003735597711056471, + 0.0049989596009254456, + 0.007579331286251545, + 0.007065943442285061, + 0.005985510069876909, + 0.008179724216461182, + 0.008504172787070274, + 0.0025867812801152468, + 0.003023226046934724, + 0.004368743859231472, + 0.003953520208597183, + 0.005681079812347889, + 0.0036613873671740294, + 0.005418967455625534, + 0.003747238777577877, + 0.012184999883174896, + 0.009938526898622513, + 0.009214920923113823, + 0.06273595243692398, + 0.00791865959763527, + 0.0087751355022192, + 0.014344324357807636, + 0.046167321503162384, + 0.020034415647387505, + 0.020193053409457207, + 0.01531208772212267, + 0.01658867858350277, + 0.028889043256640434, + 0.02182498574256897, + 0.02815520390868187, + 0.015643713995814323, + 0.004181351512670517, + 0.0026221859734505415, + 0.002077684039250016, + 0.0023657118435949087, + 0.003907329402863979, + 0.0025771160144358873, + 0.00327499327249825, + 0.003884583245962858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.012564077973365784 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495745688676834 + ], + "opposite_logprob_tokens": [ + -2.929957389831543 + ], + "logprob_reward": -0.05495745688676834, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495745688676834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16663293540477753, + 0.0814453586935997, + 0.20546244084835052, + 0.05556023120880127, + 0.09484800696372986, + 0.2025926262140274, + 0.01801074482500553, + 0.13271847367286682, + 0.024146215990185738, + 0.018583031371235847 + ], + "credit_map": [ + 0.04025896638631821, + 0.01738334447145462, + 0.012633060105144978, + 0.01844061352312565, + 0.020488392561674118, + 0.01898849382996559, + 0.02239658124744892, + 0.016043467447161674, + 0.010009757243096828, + 0.012380863539874554, + 0.006338616833090782, + 0.006765031721442938, + 0.021699970588088036, + 0.006609761156141758, + 0.009789193980395794, + 0.007852156646549702, + 0.050934530794620514, + 0.022833311930298805, + 0.006780920550227165, + 0.010921726934611797, + 0.013053417205810547, + 0.07863974571228027, + 0.013481240719556808, + 0.00881753396242857, + 0.005207831505686045, + 0.006762096658349037, + 0.0054958234541118145, + 0.008518637157976627, + 0.008761879988014698, + 0.005830659996718168, + 0.006288229953497648, + 0.008695068769156933, + 0.012161658145487309, + 0.0041601755656301975, + 0.006541109643876553, + 0.006842870730906725, + 0.030607953667640686, + 0.007181047461926937, + 0.005439962726086378, + 0.02191321738064289, + 0.013156140223145485, + 0.016173841431736946, + 0.024420155212283134, + 0.03203858807682991, + 0.028031000867486, + 0.022347310557961464, + 0.04295844957232475, + 0.023467103019356728, + 0.0019038637401536107, + 0.0015633750008419156, + 0.002387849846854806, + 0.002874043071642518, + 0.0026898514479398727, + 0.0014511204790323973, + 0.003080532420426607, + 0.0020601064898073673, + 0.00896493811160326, + 0.007573178969323635, + 0.00730883190408349, + 0.05063295736908913, + 0.005967401899397373, + 0.006698952056467533, + 0.009613028727471828, + 0.03595917671918869, + 0.005368413403630257, + 0.00311447586864233, + 0.0027332103345543146, + 0.002073355484753847, + 0.003605266334488988, + 0.002193224849179387, + 0.002493488835170865, + 0.0025647797156125307, + 0.0029110959731042385, + 0.0023627481423318386, + 0.002234320854768157, + 0.0017104005673900247, + 0.0027304140385240316, + 0.0019948510453104973, + 0.0025221831165254116, + 0.002117015654221177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.05495745688676834 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865148663520813 + ], + "opposite_logprob_tokens": [ + -3.0486514568328857 + ], + "logprob_reward": -0.04865148663520813, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865148663520813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05549907311797142, + 0.0640132874250412, + 0.4385136663913727, + 0.03480379283428192, + 0.12827667593955994, + 0.022418135777115822, + 0.01589326560497284, + 0.03624826297163963, + 0.07386711984872818, + 0.13046668469905853 + ], + "credit_map": [ + 0.011146416887640953, + 0.0056327641941607, + 0.0047771260142326355, + 0.005592510104179382, + 0.007505747955292463, + 0.006397695746272802, + 0.007675028871744871, + 0.006771789863705635, + 0.008274312131106853, + 0.005956357344985008, + 0.005676796194165945, + 0.005916382186114788, + 0.012390526942908764, + 0.008383161388337612, + 0.008193578571081161, + 0.009222179651260376, + 0.14970563352108002, + 0.022504501044750214, + 0.010787662118673325, + 0.012736440636217594, + 0.027658427134156227, + 0.16534370183944702, + 0.031514909118413925, + 0.01826244592666626, + 0.004404135048389435, + 0.0035642345901578665, + 0.004563512746244669, + 0.003785034641623497, + 0.006755390670150518, + 0.0036766931880265474, + 0.003051087260246277, + 0.005003707483410835, + 0.012356472201645374, + 0.005858557298779488, + 0.009480933658778667, + 0.007332222070544958, + 0.05535814166069031, + 0.00910767912864685, + 0.0067479549907147884, + 0.02203473076224327, + 0.0022548881825059652, + 0.0019387216307222843, + 0.0026777538005262613, + 0.00310693122446537, + 0.0029514022171497345, + 0.00183676031883806, + 0.005139308515936136, + 0.0025123702362179756, + 0.0011542121646925807, + 0.0017318662721663713, + 0.0018806066364049911, + 0.0021585968788713217, + 0.0021491101942956448, + 0.0015049073845148087, + 0.003483040491119027, + 0.0018309272127225995, + 0.002293715951964259, + 0.0031387771014124155, + 0.0027463685255497694, + 0.010771187953650951, + 0.002960184356197715, + 0.0028406521305441856, + 0.0034434970002621412, + 0.008053882047533989, + 0.01315808854997158, + 0.007630379870533943, + 0.007254018448293209, + 0.006764474790543318, + 0.012319518253207207, + 0.00926966406404972, + 0.010012098588049412, + 0.007458882872015238, + 0.012422625906765461, + 0.017944255843758583, + 0.012500355020165443, + 0.012634603306651115, + 0.01636560820043087, + 0.012131568975746632, + 0.02308688871562481, + 0.023380795493721962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.04865148663520813 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, slow-tracking shot smoothly glides across a surreal, flooded English garden, where the water, perfectly still and crystal clear, acts as a mirror against a soft, overcast sky. Beneath the surface, a submerged maze of manicured hedges and stone paths looms. The camera gradually descends, following a solitary figure seated in a sleek, dark, high-gloss mahogany canoe. The figure, whose features remain blurred, wears a simple cream-colored linen tunic, their posture poised and contemplative. They row with a steady, deliberate rhythm, gracefully dipping the slender wooden oars into the tranquil water. Each stroke sends gossamer ripples spreading outward, gently nudging the heavy, blossoming heads of the submerged crimson and pink rose bushes. Droplets of water cling like tiny gemstones to the velvety petals of the flowers shattered by the surface. The scene is bathed in a soft, diffuse light that casts no harsh shadows, creating an atmosphere of profound tranquility and melancholic beauty, the vibrant red of the roses contrasting beautifully with the deep green of the underwater foliage and the dark wood of the boat.", + "seed": 18009, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_15/index_9742_seed_18009_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0340042095631361, + "vlm_logprob": -0.0340042095631361, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.0340042095631361, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025218134745955467 + ], + "opposite_logprob_tokens": [ + -3.7752182483673096 + ], + "logprob_reward": -0.025218134745955467, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025218134745955467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19537995755672455, + 0.03974051773548126, + 0.08835705369710922, + 0.04162954166531563, + 0.0756569355726242, + 0.11496733129024506, + 0.03174912557005882, + 0.048837557435035706, + 0.1329079568386078, + 0.23077403008937836 + ], + "credit_map": [ + 0.029882576316595078, + 0.021423963829874992, + 0.018820052966475487, + 0.011633719317615032, + 0.021411223337054253, + 0.014227050356566906, + 0.03769411891698837, + 0.040287237614393234, + 0.006062271073460579, + 0.005380621179938316, + 0.006237864028662443, + 0.004214687272906303, + 0.00504736602306366, + 0.0034744024742394686, + 0.004432902671396732, + 0.004890404175966978, + 0.01959075964987278, + 0.01806192845106125, + 0.010756746865808964, + 0.007457197178155184, + 0.006877423729747534, + 0.0076967207714915276, + 0.0074280910193920135, + 0.010488183237612247, + 0.007444582413882017, + 0.004028453025966883, + 0.004425678867846727, + 0.0046361652202904224, + 0.0035444938112050295, + 0.004787276033312082, + 0.0060893455520272255, + 0.006673547904938459, + 0.00767536973580718, + 0.00598341366276145, + 0.014991888776421547, + 0.014707707799971104, + 0.004388636909425259, + 0.004491876810789108, + 0.016825836151838303, + 0.006592206656932831, + 0.011509071104228497, + 0.009313196875154972, + 0.023319192230701447, + 0.015427238307893276, + 0.007983207702636719, + 0.006289443001151085, + 0.0301143079996109, + 0.011011676862835884, + 0.003638545051217079, + 0.0026032563764601946, + 0.004362904001027346, + 0.0054026818834245205, + 0.0027016066014766693, + 0.0031986383255571127, + 0.006898056715726852, + 0.0029434377793222666, + 0.003400377696380019, + 0.0030472208745777607, + 0.009397740475833416, + 0.007808363065123558, + 0.004717918578535318, + 0.005552834365516901, + 0.008750761859118938, + 0.006162339821457863, + 0.007590420078486204, + 0.005071297753602266, + 0.01879573054611683, + 0.01098534744232893, + 0.007887281477451324, + 0.011505414731800556, + 0.03138023987412453, + 0.03969221189618111, + 0.0327102467417717, + 0.02420174330472946, + 0.04373222589492798, + 0.049650877714157104, + 0.017639247700572014, + 0.020651232451200485, + 0.02186259813606739, + 0.02032586932182312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.025218134745955467 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062070805579423904 + ], + "opposite_logprob_tokens": [ + -2.812070846557617 + ], + "logprob_reward": -0.062070805579423904, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062070805579423904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12678250670433044, + 0.1440882831811905, + 0.03805842623114586, + 0.08525760471820831, + 0.04358305037021637, + 0.07633736729621887, + 0.04627165198326111, + 0.11211525648832321, + 0.19600500166416168, + 0.13150082528591156 + ], + "credit_map": [ + 0.015740901231765747, + 0.011150335893034935, + 0.016613010317087173, + 0.013356496579945087, + 0.021105503663420677, + 0.0123574985191226, + 0.023138543590903282, + 0.013320215046405792, + 0.017195863649249077, + 0.014927943237125874, + 0.015978818759322166, + 0.019307667389512062, + 0.016625650227069855, + 0.022988956421613693, + 0.017967695370316505, + 0.019095690920948982, + 0.004915330559015274, + 0.005379368085414171, + 0.004088898189365864, + 0.006558596156537533, + 0.004053943790495396, + 0.0040351166389882565, + 0.003943486139178276, + 0.005083684343844652, + 0.015803515911102295, + 0.009192555211484432, + 0.010239108465611935, + 0.012579991482198238, + 0.007870910689234734, + 0.007760521490126848, + 0.012470070272684097, + 0.009340936318039894, + 0.0043095615692436695, + 0.003134680213406682, + 0.007478915620595217, + 0.006248421035706997, + 0.003564347280189395, + 0.004161125980317593, + 0.008289829827845097, + 0.006396167445927858, + 0.0069183786399662495, + 0.0037558446638286114, + 0.021171335130929947, + 0.012123068794608116, + 0.003356065135449171, + 0.003859699936583638, + 0.012266022153198719, + 0.012886951677501202, + 0.006861823610961437, + 0.004125738982111216, + 0.0064764441922307014, + 0.006303131114691496, + 0.004838135093450546, + 0.005085529759526253, + 0.008371802978217602, + 0.0042090448550879955, + 0.009579123929142952, + 0.010639897547662258, + 0.011178331449627876, + 0.016319215297698975, + 0.016468552872538567, + 0.018080802634358406, + 0.016901718452572823, + 0.012947604060173035, + 0.012463035061955452, + 0.006967825815081596, + 0.017993459478020668, + 0.011995725333690643, + 0.013697906397283077, + 0.0167898777872324, + 0.057289548218250275, + 0.058807626366615295, + 0.019479546695947647, + 0.015428131446242332, + 0.02149646356701851, + 0.0254023689776659, + 0.012435605749487877, + 0.014721927233040333, + 0.012064889073371887, + 0.010471883229911327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 9 + ], + "selected_reward": -0.062070805579423904 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005963034927845001 + ], + "opposite_logprob_tokens": [ + -5.13096284866333 + ], + "logprob_reward": -0.005963034927845001, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005963034927845001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22846554219722748, + 0.11253499239683151, + 0.10071362555027008, + 0.04262066259980202, + 0.13677112758159637, + 0.12001495063304901, + 0.022970402613282204, + 0.0876428633928299, + 0.035163845866918564, + 0.11310195922851562 + ], + "credit_map": [ + 0.030871137976646423, + 0.02678397297859192, + 0.02668953873217106, + 0.018966371193528175, + 0.033192720264196396, + 0.021021541208028793, + 0.030507957562804222, + 0.04043230414390564, + 0.022358261048793793, + 0.011113950051367283, + 0.017544809728860855, + 0.013449854217469692, + 0.011548933573067188, + 0.009561649523675442, + 0.016320928931236267, + 0.010636613704264164, + 0.022730279713869095, + 0.0165764968842268, + 0.012493169866502285, + 0.010117324069142342, + 0.008746474049985409, + 0.009658031165599823, + 0.009569300338625908, + 0.01082255132496357, + 0.004940045066177845, + 0.004340613726526499, + 0.004486914724111557, + 0.004499461967498064, + 0.00545915961265564, + 0.007883964106440544, + 0.004341961350291967, + 0.006668538320809603, + 0.011731854639947414, + 0.011666413396596909, + 0.013375603593885899, + 0.019723745062947273, + 0.0069091543555259705, + 0.020229466259479523, + 0.03597709909081459, + 0.01715780235826969, + 0.012420408427715302, + 0.005253722425550222, + 0.033509232103824615, + 0.017294391989707947, + 0.005390978883951902, + 0.006586136296391487, + 0.022074634209275246, + 0.017485443502664566, + 0.0023113160859793425, + 0.0019164917757734656, + 0.003224140964448452, + 0.0034827329218387604, + 0.003338584676384926, + 0.0028128186240792274, + 0.003967322409152985, + 0.0019169955048710108, + 0.00794211309403181, + 0.008100948296487331, + 0.01383898314088583, + 0.011592864990234375, + 0.012150641530752182, + 0.010534077882766724, + 0.014336386695504189, + 0.009146847762167454, + 0.00293494900688529, + 0.0025779621209949255, + 0.006188312545418739, + 0.005076720379292965, + 0.005640900228172541, + 0.004692158196121454, + 0.004762917757034302, + 0.003289927961304784, + 0.01333248894661665, + 0.01404382660984993, + 0.020131465047597885, + 0.025144295766949654, + 0.007941413670778275, + 0.011025759391486645, + 0.012274018488824368, + 0.00920869316905737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.005963034927845001 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043067168444395065 + ], + "opposite_logprob_tokens": [ + -3.168067216873169 + ], + "logprob_reward": -0.043067168444395065, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043067168444395065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06857410818338394, + 0.02891194447875023, + 0.059061478823423386, + 0.0346856489777565, + 0.12643973529338837, + 0.2060893028974533, + 0.0638471394777298, + 0.04238982871174812, + 0.20604020357131958, + 0.16396059095859528 + ], + "credit_map": [ + 0.01013532280921936, + 0.01139453612267971, + 0.006365463603287935, + 0.00556793762370944, + 0.008035264909267426, + 0.0038496016059070826, + 0.009424194693565369, + 0.013801787979900837, + 0.005149104166775942, + 0.00426557706668973, + 0.003990912344306707, + 0.003659596433863044, + 0.0025022884365171194, + 0.0034968906547874212, + 0.0029243575409054756, + 0.002923219231888652, + 0.012922178022563457, + 0.010824398137629032, + 0.007066618651151657, + 0.005814132746309042, + 0.0048414054326713085, + 0.004976407624781132, + 0.006570195779204369, + 0.006046148482710123, + 0.004937958437949419, + 0.002822810085490346, + 0.004032570868730545, + 0.0038185997400432825, + 0.0032138400711119175, + 0.006174600683152676, + 0.0037334044463932514, + 0.005951865576207638, + 0.014609770849347115, + 0.00932067260146141, + 0.015785371884703636, + 0.02616780623793602, + 0.004997061099857092, + 0.007578249089419842, + 0.029885010793805122, + 0.01809580810368061, + 0.021557962521910667, + 0.013649978674948215, + 0.05555092170834541, + 0.028473585844039917, + 0.01562250591814518, + 0.015665186569094658, + 0.03159242868423462, + 0.02397674322128296, + 0.0038009537383913994, + 0.005606743972748518, + 0.009431730024516582, + 0.009990205988287926, + 0.007041127886623144, + 0.006799509283155203, + 0.01407504454255104, + 0.00710183335468173, + 0.003966818563640118, + 0.0034973870497196913, + 0.004630655515938997, + 0.004505644086748362, + 0.005925729870796204, + 0.0054125008173286915, + 0.008012707345187664, + 0.006438388489186764, + 0.01970694214105606, + 0.00581315066665411, + 0.01636255532503128, + 0.016171185299754143, + 0.012391624972224236, + 0.014788058586418629, + 0.06866025179624557, + 0.05214645713567734, + 0.02742718905210495, + 0.01608593948185444, + 0.02737060748040676, + 0.036172907799482346, + 0.012183264829218388, + 0.015298431739211082, + 0.015817996114492416, + 0.013604262843728065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.043067168444395065 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a surreal flooded English garden with submerged manicured hedges and stone paths?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a solitary figure seated in a sleek, dark, high-gloss mahogany canoe visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a simple cream-colored linen tunic with blurred features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera start with a high-angle slow-tracking shot and then gradually descend?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are submerged crimson and pink rose bushes visible beneath the water surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033701904118061066 + ], + "opposite_logprob_tokens": [ + -3.4087018966674805 + ], + "logprob_reward": -0.033701904118061066, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033701904118061066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06500104814767838, + 0.03280508518218994, + 0.08052410185337067, + 0.05835767462849617, + 0.05260583385825157, + 0.055412743240594864, + 0.02284884825348854, + 0.037204571068286896, + 0.1995418220758438, + 0.39569827914237976 + ], + "credit_map": [ + 0.010730714537203312, + 0.007119001820683479, + 0.012429162859916687, + 0.0059693376533687115, + 0.006215826142579317, + 0.004812788218259811, + 0.00878486130386591, + 0.00893936026841402, + 0.004833484999835491, + 0.004948951303958893, + 0.003992051817476749, + 0.0031487613450735807, + 0.004189138766378164, + 0.00427526980638504, + 0.0031929786782711744, + 0.004224451258778572, + 0.019325708970427513, + 0.014166047796607018, + 0.009952681139111519, + 0.006101489998400211, + 0.006260058376938105, + 0.006810379214584827, + 0.007533359341323376, + 0.010374373756349087, + 0.010170990601181984, + 0.005837409291416407, + 0.006029653828591108, + 0.007496584206819534, + 0.005364229436963797, + 0.009069547057151794, + 0.00693199597299099, + 0.007457262370735407, + 0.00615369388833642, + 0.003473079763352871, + 0.006512693595141172, + 0.008773052133619785, + 0.002417772077023983, + 0.003048158250749111, + 0.014319834299385548, + 0.007907549850642681, + 0.005109632387757301, + 0.006057410500943661, + 0.005499131046235561, + 0.011598250828683376, + 0.005387089680880308, + 0.004811914172023535, + 0.009818402118980885, + 0.007130913436412811, + 0.002086852677166462, + 0.0021901526488363743, + 0.0037957397289574146, + 0.004854629747569561, + 0.0023174602538347244, + 0.0013992536114528775, + 0.00467262789607048, + 0.0015321314567700028, + 0.003066206816583872, + 0.0034556291066110134, + 0.007314743474125862, + 0.005344475153833628, + 0.0046088011004030704, + 0.0035008410923182964, + 0.005994271021336317, + 0.003919604234397411, + 0.020787212997674942, + 0.006354421842843294, + 0.022883139550685883, + 0.016293799504637718, + 0.013029067777097225, + 0.013201123103499413, + 0.0666186586022377, + 0.04037441313266754, + 0.0569024384021759, + 0.036079563200473785, + 0.0791589692234993, + 0.09552731364965439, + 0.02075619250535965, + 0.024056769907474518, + 0.05327485874295235, + 0.02994217537343502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.033701904118061066 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_2/index_tensor([6307, 6307])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_2/index_tensor([6307, 6307])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..add6c0037dd9a530895bee598ad52677925b4293 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_2/index_tensor([6307, 6307])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, beginning from a slightly overhead angle, the camera slowly zooms in, creating a suffocating, oppressive feeling. In the frame, a young delivery driver hunches over, his bright yellow waterproof uniform glaring against the darkness of the late-night alley, its zipper stained with grease. Cold sweat drenches his messy, short black hair, leaving strands clinging to his skin. The only light in the scene comes from the smartphone he clutches tightly in his hand. The screen's cold blue glow carves deep, tense shadows across his weary face. His gaze is fixed on the order details popping up in the center of the screen, the address \"44 Wangsheng Road\" clearly displayed in bold, bold letters. His index finger, the tip of his fingernail trembling slightly with hesitation, the grime of the day still lingering between his nails, hovers mere millimeters from the screen. A visible, nervous swallow causes his Adam's apple to roll violently. At that moment, the number representing the subsidy amount at the bottom of the screen, initially displaying \"12.8 yuan,\" flickered without warning. The digits blurred and stretched, then, amid a faint electronic hum, instantly reconstructed into an absurdly large number: \"128 yuan.\" His pupils suddenly constricted, his lips parted slightly, his face a mixture of extreme horror, greed, and a deep-seated fear. The entire scene was frozen in an eerie yet alluring atmosphere of suspense.", + "seed": 18594, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_2/index_6307_seed_18594_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.040558209526352584, + "vlm_logprob": -0.040558209526352584, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.040558209526352584, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024159319698810577 + ], + "opposite_logprob_tokens": [ + -4.0241594314575195 + ], + "logprob_reward": -0.024159319698810577, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.024159319698810577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07579514384269714, + 0.05125368386507034, + 0.2140299528837204, + 0.05179707333445549, + 0.11124344170093536, + 0.05270998552441597, + 0.11334706842899323, + 0.13881906867027283, + 0.12416162341833115, + 0.0668429359793663 + ], + "credit_map": [ + 0.008162437006831169, + 0.006579661276191473, + 0.013510561548173428, + 0.008811342529952526, + 0.009250333532691002, + 0.006550392601639032, + 0.012974708341062069, + 0.00995570793747902, + 0.005878221709281206, + 0.00461185397580266, + 0.012928079813718796, + 0.003010418964549899, + 0.00282068969681859, + 0.006481530144810677, + 0.004814710468053818, + 0.010708179324865341, + 0.020598405972123146, + 0.009729476645588875, + 0.07984330505132675, + 0.007484366651624441, + 0.0265332143753767, + 0.010240578092634678, + 0.007349808234721422, + 0.05225080996751785, + 0.005357409827411175, + 0.005755609832704067, + 0.005462342873215675, + 0.006245030555874109, + 0.009630911983549595, + 0.008130616508424282, + 0.004437597934156656, + 0.006777556147426367, + 0.011312647722661495, + 0.014992743730545044, + 0.016759159043431282, + 0.016552308574318886, + 0.012832676991820335, + 0.008707838132977486, + 0.015936540439724922, + 0.014149530790746212, + 0.005351015366613865, + 0.007261778227984905, + 0.007144306320697069, + 0.007828970439732075, + 0.006202511489391327, + 0.006444682367146015, + 0.0065082344226539135, + 0.0059684873558580875, + 0.01103475596755743, + 0.017468580976128578, + 0.011251530610024929, + 0.02016221545636654, + 0.012160668149590492, + 0.017330337315797806, + 0.011606615968048573, + 0.0123323705047369, + 0.017316995188593864, + 0.027462828904390335, + 0.010556711815297604, + 0.021816259250044823, + 0.01832161657512188, + 0.024720335379242897, + 0.010074049234390259, + 0.008550283499062061, + 0.019111737608909607, + 0.009757224470376968, + 0.010099893435835838, + 0.012595172971487045, + 0.02453504502773285, + 0.016353584825992584, + 0.016374962404370308, + 0.015334004536271095, + 0.007610928267240524, + 0.00844463799148798, + 0.00855216570198536, + 0.011125011369585991, + 0.003978627733886242, + 0.011959625408053398, + 0.004726956598460674, + 0.010444979183375835 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.024159319698810577 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009809385053813457 + ], + "opposite_logprob_tokens": [ + -4.634809494018555 + ], + "logprob_reward": -0.009809385053813457, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009809385053813457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07353100180625916, + 0.04936159402132034, + 0.21488681435585022, + 0.13259156048297882, + 0.034957364201545715, + 0.10601396858692169, + 0.12919317185878754, + 0.15855509042739868, + 0.07382611185312271, + 0.027083341032266617 + ], + "credit_map": [ + 0.004224134609103203, + 0.0054856897331774235, + 0.014537986367940903, + 0.0059363506734371185, + 0.008603428490459919, + 0.004862223286181688, + 0.01503536757081747, + 0.014845818281173706, + 0.00608034199103713, + 0.006130651570856571, + 0.00883340835571289, + 0.004391571041196585, + 0.0037505165673792362, + 0.008700191043317318, + 0.0044895680621266365, + 0.006985343527048826, + 0.018222706392407417, + 0.016337255015969276, + 0.059818606823682785, + 0.012143027037382126, + 0.056184254586696625, + 0.021319443359971046, + 0.014175870455801487, + 0.01668565161526203, + 0.015766561031341553, + 0.012890156358480453, + 0.014500392600893974, + 0.021125657483935356, + 0.014112371020019054, + 0.01789151318371296, + 0.009538396261632442, + 0.026766516268253326, + 0.004768913146108389, + 0.004326358903199434, + 0.004549642093479633, + 0.004945958033204079, + 0.003918547183275223, + 0.0025649936869740486, + 0.004287275951355696, + 0.005595674738287926, + 0.010442744940519333, + 0.019061019644141197, + 0.015237836167216301, + 0.013378625735640526, + 0.012071063742041588, + 0.011705704964697361, + 0.011599625460803509, + 0.012517346069216728, + 0.01040936354547739, + 0.022164391353726387, + 0.010882641188800335, + 0.020312387496232986, + 0.01194645557552576, + 0.018365254625678062, + 0.018075495958328247, + 0.017037177458405495, + 0.02050170488655567, + 0.023565033450722694, + 0.016849499195814133, + 0.018276499584317207, + 0.013023282401263714, + 0.027703838422894478, + 0.024631861597299576, + 0.014003364369273186, + 0.00775756174698472, + 0.008051319047808647, + 0.00862320140004158, + 0.008033568039536476, + 0.010639091022312641, + 0.009281424805521965, + 0.012070161290466785, + 0.00936979055404663, + 0.0020343316718935966, + 0.0030462504364550114, + 0.0018659258494153619, + 0.003744624089449644, + 0.0024911488872021437, + 0.007721481379121542, + 0.0015872649382799864, + 0.004592313431203365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 3 + ], + "selected_reward": -0.009809385053813457 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000927137560211122 + ], + "opposite_logprob_tokens": [ + -7.000926971435547 + ], + "logprob_reward": -0.000927137560211122, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000927137560211122, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09894400835037231, + 0.0795450210571289, + 0.12238574028015137, + 0.05926061421632767, + 0.14576323330402374, + 0.12966805696487427, + 0.0860179215669632, + 0.1125248372554779, + 0.11439775675535202, + 0.05149281397461891 + ], + "credit_map": [ + 0.007345546502619982, + 0.008474190719425678, + 0.016748182475566864, + 0.011018610559403896, + 0.013297933153808117, + 0.009583361446857452, + 0.01766250655055046, + 0.014813666231930256, + 0.006353490054607391, + 0.011255085468292236, + 0.013166170567274094, + 0.009109921753406525, + 0.00969867967069149, + 0.0075012026354670525, + 0.007926968857645988, + 0.014533495530486107, + 0.019989579916000366, + 0.015234329737722874, + 0.02196146547794342, + 0.009441192261874676, + 0.01881532557308674, + 0.013844754546880722, + 0.00929170846939087, + 0.013807384297251701, + 0.007190180942416191, + 0.004415480885654688, + 0.005503122229129076, + 0.00498467730358243, + 0.013199410401284695, + 0.008742873556911945, + 0.004511497914791107, + 0.0107133649289608, + 0.01397947408258915, + 0.022186502814292908, + 0.017879610881209373, + 0.01669144257903099, + 0.02259235642850399, + 0.01284858025610447, + 0.01893981359899044, + 0.020645441487431526, + 0.012550425715744495, + 0.01994664967060089, + 0.01737196370959282, + 0.017232082784175873, + 0.015916595235466957, + 0.0123596815392375, + 0.015408214181661606, + 0.018882425501942635, + 0.008754323236644268, + 0.011655027978122234, + 0.007022157311439514, + 0.015705201774835587, + 0.009871912188827991, + 0.013928341679275036, + 0.009269940666854382, + 0.009811002761125565, + 0.01407091785222292, + 0.0238956268876791, + 0.008122862316668034, + 0.016457386314868927, + 0.011876438744366169, + 0.01750379242002964, + 0.011871117167174816, + 0.008726680651307106, + 0.011296671815216541, + 0.01077333651483059, + 0.012475784868001938, + 0.016967108473181725, + 0.018465422093868256, + 0.016355756670236588, + 0.012080715037882328, + 0.01598295383155346, + 0.006580683868378401, + 0.006829801481217146, + 0.008297703228890896, + 0.006991666276007891, + 0.0038853005971759558, + 0.00857661385089159, + 0.003764878958463669, + 0.006566158961504698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 2 + ], + "selected_reward": -0.000927137560211122 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294384300708771 + ], + "opposite_logprob_tokens": [ + -2.2379438877105713 + ], + "logprob_reward": -0.11294384300708771, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294384300708771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3025699257850647, + 0.053122177720069885, + 0.20832867920398712, + 0.09099724888801575, + 0.05957338586449623, + 0.07954036444425583, + 0.07314573973417282, + 0.054922252893447876, + 0.04961194470524788, + 0.028188209980726242 + ], + "credit_map": [ + 0.01991897076368332, + 0.020354637876152992, + 0.056938737630844116, + 0.028858831152319908, + 0.03462371975183487, + 0.024660253897309303, + 0.06949102878570557, + 0.04772375896573067, + 0.006034948863089085, + 0.006721304263919592, + 0.01063612848520279, + 0.0042464411817491055, + 0.005225108936429024, + 0.006788826547563076, + 0.005257782526314259, + 0.00821163970977068, + 0.01401064358651638, + 0.007946429774165154, + 0.07525019347667694, + 0.010037199594080448, + 0.029557151719927788, + 0.010513884015381336, + 0.012330529280006886, + 0.04868265613913536, + 0.011233268305659294, + 0.0089826425537467, + 0.010353646241128445, + 0.013916563242673874, + 0.009003235958516598, + 0.01235945988446474, + 0.008599766530096531, + 0.016548678278923035, + 0.0075749121606349945, + 0.006476858630776405, + 0.01009173970669508, + 0.009089021943509579, + 0.007345001213252544, + 0.004345051944255829, + 0.009188770316541195, + 0.005462028086185455, + 0.008117319084703922, + 0.011964772827923298, + 0.013986499980092049, + 0.012571030296385288, + 0.006795007269829512, + 0.006999318953603506, + 0.009214513003826141, + 0.009891913272440434, + 0.006301520857959986, + 0.011756397783756256, + 0.00696515291929245, + 0.011893185786902905, + 0.009211675263941288, + 0.014007631689310074, + 0.0059360116720199585, + 0.007074161898344755, + 0.005969524383544922, + 0.010136745870113373, + 0.0028935286682099104, + 0.00635512126609683, + 0.008945615030825138, + 0.008895286358892918, + 0.006499254144728184, + 0.005227180663496256, + 0.005295616574585438, + 0.004603464622050524, + 0.00289922789670527, + 0.004954216070473194, + 0.015107480809092522, + 0.007999633438885212, + 0.003333465429022908, + 0.00541883846744895, + 0.004181083291769028, + 0.003986657131463289, + 0.001887368387542665, + 0.0031357635743916035, + 0.002026859438046813, + 0.004988675471395254, + 0.0029636197723448277, + 0.005018184427171946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.11294384300708771 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495136231184006 + ], + "opposite_logprob_tokens": [ + -2.9299514293670654 + ], + "logprob_reward": -0.05495136231184006, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495136231184006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11875846982002258, + 0.04514651000499725, + 0.10912644863128662, + 0.14407366514205933, + 0.09750545024871826, + 0.0630241110920906, + 0.17401674389839172, + 0.08530710637569427, + 0.107636459171772, + 0.05540505424141884 + ], + "credit_map": [ + 0.011211882345378399, + 0.018221601843833923, + 0.0211767740547657, + 0.012947670184075832, + 0.010922607034444809, + 0.01256025955080986, + 0.011959637515246868, + 0.019758038222789764, + 0.005710757803171873, + 0.005951649509370327, + 0.008574317209422588, + 0.0034134648740291595, + 0.0029925976414233446, + 0.00749744288623333, + 0.0035236289259046316, + 0.0074826511554419994, + 0.01023548748344183, + 0.014532255008816719, + 0.010895122773945332, + 0.01421236153692007, + 0.018587123602628708, + 0.014719611965119839, + 0.006061347667127848, + 0.019883135333657265, + 0.01724749058485031, + 0.018067309632897377, + 0.01488515269011259, + 0.022114235907793045, + 0.019726432859897614, + 0.01921234093606472, + 0.010392190888524055, + 0.022428497672080994, + 0.010115440934896469, + 0.013079317286610603, + 0.014767293818295002, + 0.011728253215551376, + 0.01140148751437664, + 0.009283851832151413, + 0.013951178640127182, + 0.013178618624806404, + 0.005758925341069698, + 0.01012793555855751, + 0.010131044313311577, + 0.008694122545421124, + 0.006404319778084755, + 0.0060216207057237625, + 0.007444881368428469, + 0.0084412582218647, + 0.018936971202492714, + 0.025105517357587814, + 0.015704961493611336, + 0.03165370598435402, + 0.02526242658495903, + 0.02926977165043354, + 0.013570956885814667, + 0.014512441121041775, + 0.011329294182360172, + 0.016608748584985733, + 0.0065151117742061615, + 0.011603238992393017, + 0.00969710387289524, + 0.013593020848929882, + 0.007995432242751122, + 0.00796515867114067, + 0.01569698005914688, + 0.009385882876813412, + 0.007310471497476101, + 0.008078488521277905, + 0.03195150941610336, + 0.01579255238175392, + 0.008221850730478764, + 0.01119872648268938, + 0.005766937043517828, + 0.007122131995856762, + 0.003737580729648471, + 0.007853330112993717, + 0.005488223396241665, + 0.012700053863227367, + 0.003789595328271389, + 0.008947202004492283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.05495136231184006 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, beginning from a slightly overhead angle, the camera slowly zooms in, creating a suffocating, oppressive feeling. In the frame, a young delivery driver hunches over, his bright yellow waterproof uniform glaring against the darkness of the late-night alley, its zipper stained with grease. Cold sweat drenches his messy, short black hair, leaving strands clinging to his skin. The only light in the scene comes from the smartphone he clutches tightly in his hand. The screen's cold blue glow carves deep, tense shadows across his weary face. His gaze is fixed on the order details popping up in the center of the screen, the address \"44 Wangsheng Road\" clearly displayed in bold, bold letters. His index finger, the tip of his fingernail trembling slightly with hesitation, the grime of the day still lingering between his nails, hovers mere millimeters from the screen. A visible, nervous swallow causes his Adam's apple to roll violently. At that moment, the number representing the subsidy amount at the bottom of the screen, initially displaying \"12.8 yuan,\" flickered without warning. The digits blurred and stretched, then, amid a faint electronic hum, instantly reconstructed into an absurdly large number: \"128 yuan.\" His pupils suddenly constricted, his lips parted slightly, his face a mixture of extreme horror, greed, and a deep-seated fear. The entire scene was frozen in an eerie yet alluring atmosphere of suspense.", + "seed": 18595, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_2/index_6307_seed_18595_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03529247913975268, + "vlm_logprob": -0.03529247913975268, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.03529247913975268, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0326358862221241 + ], + "opposite_logprob_tokens": [ + -3.5326359272003174 + ], + "logprob_reward": -0.0326358862221241, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0326358862221241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21238458156585693, + 0.04678153246641159, + 0.047289907932281494, + 0.05668603628873825, + 0.08957048505544662, + 0.050136175006628036, + 0.1771436631679535, + 0.09899523109197617, + 0.12313597649335861, + 0.09787631779909134 + ], + "credit_map": [ + 0.02498159557580948, + 0.022967197000980377, + 0.03275802358984947, + 0.018695522099733353, + 0.026941422373056412, + 0.03216708451509476, + 0.023668378591537476, + 0.0302053801715374, + 0.006110436748713255, + 0.004706263076514006, + 0.008401686325669289, + 0.003918438218533993, + 0.00496635539457202, + 0.007461462635546923, + 0.00481306342408061, + 0.006403835956007242, + 0.00605706125497818, + 0.006675138603895903, + 0.005835545249283314, + 0.005420259200036526, + 0.005874374881386757, + 0.004915646277368069, + 0.0072652739472687244, + 0.005246612709015608, + 0.006102243438363075, + 0.008717614226043224, + 0.006360701750963926, + 0.008495520800352097, + 0.007755667436867952, + 0.0070370836183428764, + 0.007374854758381844, + 0.004842354916036129, + 0.0062693641521036625, + 0.012589829042553902, + 0.011007927358150482, + 0.011864698491990566, + 0.013027184642851353, + 0.016578705981373787, + 0.01146694365888834, + 0.006765847560018301, + 0.004541187547147274, + 0.0054227011278271675, + 0.005026594735682011, + 0.005924610886722803, + 0.008260256610810757, + 0.011152259074151516, + 0.004955959971994162, + 0.0048526134341955185, + 0.016229169443249702, + 0.03489216789603233, + 0.024150311946868896, + 0.016298646107316017, + 0.033010490238666534, + 0.01163901761174202, + 0.024216558784246445, + 0.016707314178347588, + 0.00609710905700922, + 0.020924542099237442, + 0.00534421531483531, + 0.014527860097587109, + 0.018197624012827873, + 0.005666655953973532, + 0.023666031658649445, + 0.004571206867694855, + 0.00694930600002408, + 0.03366370126605034, + 0.007597420364618301, + 0.01793610490858555, + 0.020217595621943474, + 0.017006151378154755, + 0.01250369194895029, + 0.007262023165822029, + 0.011784474365413189, + 0.01808679662644863, + 0.011812158860266209, + 0.00755568640306592, + 0.013719402253627777, + 0.008646252565085888, + 0.012728842906653881, + 0.01354271825402975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.0326358862221241 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869486019015312 + ], + "opposite_logprob_tokens": [ + -3.0486948490142822 + ], + "logprob_reward": -0.04869486019015312, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869486019015312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08174221962690353, + 0.02546280063688755, + 0.05853145942091942, + 0.05721616372466087, + 0.0343109592795372, + 0.06812875717878342, + 0.1860838383436203, + 0.12436456233263016, + 0.1827152669429779, + 0.1814439445734024 + ], + "credit_map": [ + 0.014616808854043484, + 0.006455647759139538, + 0.01570970192551613, + 0.004902136046439409, + 0.008224604651331902, + 0.013357785530388355, + 0.0064461808651685715, + 0.012029360048472881, + 0.0033918162807822227, + 0.0020828768610954285, + 0.005542887840420008, + 0.0018444419838488102, + 0.0034657418727874756, + 0.00437595508992672, + 0.0015698082279413939, + 0.0031892736442387104, + 0.009819039143621922, + 0.0058696395717561245, + 0.008852514438331127, + 0.008049007505178452, + 0.005477048456668854, + 0.005246972665190697, + 0.007682254072278738, + 0.0075349826365709305, + 0.005921980831772089, + 0.00900049414485693, + 0.00560712767764926, + 0.0119329784065485, + 0.0064222910441458225, + 0.004883569665253162, + 0.008792337030172348, + 0.004655385855585337, + 0.0025914888828992844, + 0.004827369004487991, + 0.0041271476075053215, + 0.003695770865306258, + 0.005361436866223812, + 0.007084000390022993, + 0.0041209012269973755, + 0.0025028435047715902, + 0.008246304467320442, + 0.008337998762726784, + 0.007976825349032879, + 0.008381303399801254, + 0.01143554039299488, + 0.01039669569581747, + 0.0061824731528759, + 0.007171616889536381, + 0.013284260407090187, + 0.03759210929274559, + 0.020468350499868393, + 0.018331320956349373, + 0.029972361400723457, + 0.026932496577501297, + 0.023723136633634567, + 0.01577978953719139, + 0.015980517491698265, + 0.018393514677882195, + 0.013956555165350437, + 0.017422255128622055, + 0.019820157438516617, + 0.00790877640247345, + 0.016847258433699608, + 0.014035521075129509, + 0.014203354716300964, + 0.043117132037878036, + 0.010253815911710262, + 0.016426770016551018, + 0.028588898479938507, + 0.04207705706357956, + 0.016993509605526924, + 0.011054744943976402, + 0.021856894716620445, + 0.03705712407827377, + 0.023529520258307457, + 0.01620749570429325, + 0.02476978302001953, + 0.015700053423643112, + 0.01949303038418293, + 0.02283005230128765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.04869486019015312 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019459851318970323 + ], + "opposite_logprob_tokens": [ + -6.251945972442627 + ], + "logprob_reward": -0.0019459851318970323, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019459851318970323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2439044713973999, + 0.06929980218410492, + 0.04803328961133957, + 0.0373043417930603, + 0.06925305724143982, + 0.05543021857738495, + 0.18347929418087006, + 0.1813986450433731, + 0.054852765053510666, + 0.057044029235839844 + ], + "credit_map": [ + 0.028548432514071465, + 0.04419412836432457, + 0.029840137809515, + 0.02166871726512909, + 0.024236809462308884, + 0.028611820191144943, + 0.032528024166822433, + 0.03427641838788986, + 0.0070472233928740025, + 0.007644381374120712, + 0.01224363874644041, + 0.008802056312561035, + 0.008776391856372356, + 0.006999293342232704, + 0.009590365923941135, + 0.008196453563869, + 0.005753558594733477, + 0.006529050413519144, + 0.00804617628455162, + 0.006164427846670151, + 0.004717092029750347, + 0.00528381485491991, + 0.006224777549505234, + 0.005314395762979984, + 0.00328349438495934, + 0.0053879981860518456, + 0.0040909224189817905, + 0.004767225589603186, + 0.0036320388317108154, + 0.005812330171465874, + 0.007501477375626564, + 0.002828859491273761, + 0.0035386942327022552, + 0.010491638444364071, + 0.004379895515739918, + 0.008840887807309628, + 0.010333431884646416, + 0.01535892952233553, + 0.010598079301416874, + 0.005711505189538002, + 0.003964511211961508, + 0.007912286557257175, + 0.0057872687466442585, + 0.007911643013358116, + 0.009073039516806602, + 0.008547157980501652, + 0.0073578269220888615, + 0.004876487422734499, + 0.011459301225841045, + 0.03580208495259285, + 0.01631985604763031, + 0.01311691664159298, + 0.027526482939720154, + 0.045439694076776505, + 0.017952237278223038, + 0.015862727537751198, + 0.026435263454914093, + 0.024213267490267754, + 0.023705067113041878, + 0.021268755197525024, + 0.025351176038384438, + 0.013449875637888908, + 0.022338010370731354, + 0.024637244641780853, + 0.0038888920098543167, + 0.01144622452557087, + 0.004686570260673761, + 0.00563702080398798, + 0.007786351256072521, + 0.011611416935920715, + 0.005638871807605028, + 0.004157421179115772, + 0.005929107312113047, + 0.008058731444180012, + 0.006789274979382753, + 0.003929818514734507, + 0.007331840693950653, + 0.011208700016140938, + 0.007839720696210861, + 0.00595683790743351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0019459851318970323 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899153977632523 + ], + "opposite_logprob_tokens": [ + -2.578991651535034 + ], + "logprob_reward": -0.07899153977632523, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899153977632523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.266608864068985, + 0.044697534292936325, + 0.10715702176094055, + 0.0725766122341156, + 0.05304957926273346, + 0.06918828189373016, + 0.09795284271240234, + 0.1529427021741867, + 0.04519917070865631, + 0.09062737226486206 + ], + "credit_map": [ + 0.037235524505376816, + 0.022559236735105515, + 0.04544447362422943, + 0.022861236706376076, + 0.036096103489398956, + 0.0372173935174942, + 0.02703685499727726, + 0.03815803676843643, + 0.006935156416147947, + 0.005013574380427599, + 0.00597655912861228, + 0.003251433139666915, + 0.006230010651051998, + 0.00751254428178072, + 0.0042457589879632, + 0.005532498005777597, + 0.01770973950624466, + 0.011044890619814396, + 0.014703388325870037, + 0.01804857887327671, + 0.013644074089825153, + 0.00903836078941822, + 0.010695576667785645, + 0.012272407300770283, + 0.005614559631794691, + 0.012215632945299149, + 0.008237755857408047, + 0.012086105532944202, + 0.010339462198317051, + 0.007530279457569122, + 0.01177236344665289, + 0.004780451767146587, + 0.004415575880557299, + 0.007559081539511681, + 0.006426132284104824, + 0.006381531246006489, + 0.008076069876551628, + 0.009728724136948586, + 0.006206793710589409, + 0.004255668260157108, + 0.006920722778886557, + 0.007578723598271608, + 0.007365162018686533, + 0.00843749474734068, + 0.011592091992497444, + 0.013598131947219372, + 0.006815718486905098, + 0.006880233529955149, + 0.006856684107333422, + 0.020376961678266525, + 0.006225110962986946, + 0.013219386339187622, + 0.018820207566022873, + 0.008680303581058979, + 0.018475746735930443, + 0.005298434756696224, + 0.02206028811633587, + 0.022248400375247, + 0.02122163586318493, + 0.018013421446084976, + 0.02104193903505802, + 0.011096932925283909, + 0.016233598813414574, + 0.02102648839354515, + 0.005066378973424435, + 0.010143784806132317, + 0.0032718644943088293, + 0.00511457584798336, + 0.006146136671304703, + 0.007826386950910091, + 0.004496391862630844, + 0.003133650403469801, + 0.010485240258276463, + 0.01775471307337284, + 0.010183111764490604, + 0.008693765848875046, + 0.014100907370448112, + 0.008265488781034946, + 0.011087242513895035, + 0.01005689799785614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.07899153977632523 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419412437826395 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.01419412437826395, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419412437826395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32416433095932007, + 0.03024095669388771, + 0.04593901336193085, + 0.0811043381690979, + 0.04622650891542435, + 0.08071360737085342, + 0.17887134850025177, + 0.06142895668745041, + 0.08017893135547638, + 0.07113204896450043 + ], + "credit_map": [ + 0.04657750576734543, + 0.01825495809316635, + 0.0790034681558609, + 0.016772693023085594, + 0.03837474435567856, + 0.05036788806319237, + 0.026106514036655426, + 0.0487065352499485, + 0.0037033739499747753, + 0.0033934800885617733, + 0.005139919929206371, + 0.003331872634589672, + 0.00475638173520565, + 0.0029689737129956484, + 0.00348537415266037, + 0.0034615793265402317, + 0.00734378956258297, + 0.006338262930512428, + 0.005317628383636475, + 0.004700545687228441, + 0.008257506415247917, + 0.005165230482816696, + 0.0038715607952326536, + 0.004944485612213612, + 0.009334739297628403, + 0.012891469523310661, + 0.009564723819494247, + 0.013140277937054634, + 0.00934636127203703, + 0.010065495036542416, + 0.011023270897567272, + 0.005738005507737398, + 0.005995584186166525, + 0.005913731642067432, + 0.00506672915071249, + 0.005488288123160601, + 0.0056922342628240585, + 0.007621697150170803, + 0.006311127450317144, + 0.0041371178813278675, + 0.006845494266599417, + 0.010747356340289116, + 0.008410069160163403, + 0.012566219083964825, + 0.012492877431213856, + 0.010874330066144466, + 0.011093607172369957, + 0.007683658041059971, + 0.016947679221630096, + 0.0264983382076025, + 0.013198204338550568, + 0.018247736617922783, + 0.028883056715130806, + 0.036617837846279144, + 0.020325850695371628, + 0.01815263368189335, + 0.009198347106575966, + 0.011133587919175625, + 0.0071340082213282585, + 0.006536387838423252, + 0.00826801173388958, + 0.004336814396083355, + 0.008153035305440426, + 0.006668761372566223, + 0.006105062086135149, + 0.022288408130407333, + 0.0053812917321920395, + 0.01176824513822794, + 0.013579859398305416, + 0.007147477474063635, + 0.008025937713682652, + 0.005882654804736376, + 0.006611553952097893, + 0.01431423518806696, + 0.006644207052886486, + 0.007526474539190531, + 0.010523131117224693, + 0.008938014507293701, + 0.011361481621861458, + 0.005212951917201281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.01419412437826395 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, beginning from a slightly overhead angle, the camera slowly zooms in, creating a suffocating, oppressive feeling. In the frame, a young delivery driver hunches over, his bright yellow waterproof uniform glaring against the darkness of the late-night alley, its zipper stained with grease. Cold sweat drenches his messy, short black hair, leaving strands clinging to his skin. The only light in the scene comes from the smartphone he clutches tightly in his hand. The screen's cold blue glow carves deep, tense shadows across his weary face. His gaze is fixed on the order details popping up in the center of the screen, the address \"44 Wangsheng Road\" clearly displayed in bold, bold letters. His index finger, the tip of his fingernail trembling slightly with hesitation, the grime of the day still lingering between his nails, hovers mere millimeters from the screen. A visible, nervous swallow causes his Adam's apple to roll violently. At that moment, the number representing the subsidy amount at the bottom of the screen, initially displaying \"12.8 yuan,\" flickered without warning. The digits blurred and stretched, then, amid a faint electronic hum, instantly reconstructed into an absurdly large number: \"128 yuan.\" His pupils suddenly constricted, his lips parted slightly, his face a mixture of extreme horror, greed, and a deep-seated fear. The entire scene was frozen in an eerie yet alluring atmosphere of suspense.", + "seed": 18596, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_2/index_6307_seed_18596_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02451242906972766, + "vlm_logprob": -0.02451242906972766, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.02451242906972766, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029259085655212402 + ], + "opposite_logprob_tokens": [ + -3.779259204864502 + ], + "logprob_reward": -0.029259085655212402, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.029259085655212402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03572612628340721, + 0.057493869215250015, + 0.28626787662506104, + 0.04716223105788231, + 0.11048656702041626, + 0.2730647027492523, + 0.06773926317691803, + 0.04476498067378998, + 0.04198998212814331, + 0.03530437871813774 + ], + "credit_map": [ + 0.0038693032693117857, + 0.003204246051609516, + 0.004863179754465818, + 0.0032733960542827845, + 0.006053215358406305, + 0.0034515622537583113, + 0.0028578799683600664, + 0.008153343573212624, + 0.005682097282260656, + 0.0057902950793504715, + 0.009551565162837505, + 0.004780881106853485, + 0.009332237765192986, + 0.006231075152754784, + 0.00670973164960742, + 0.009415984153747559, + 0.012510635890066624, + 0.01718015968799591, + 0.0230172798037529, + 0.010100162588059902, + 0.06514982879161835, + 0.11563991010189056, + 0.01776895858347416, + 0.024900928139686584, + 0.0028016783762723207, + 0.0038128287997096777, + 0.0043574171140789986, + 0.004121947567909956, + 0.01433013565838337, + 0.0033178552985191345, + 0.010234211571514606, + 0.004186158534139395, + 0.00910339318215847, + 0.013760295696556568, + 0.01156881544739008, + 0.013668342493474483, + 0.023376340046525, + 0.01339095737785101, + 0.015536211431026459, + 0.010082216002047062, + 0.02742401883006096, + 0.04849342629313469, + 0.026169292628765106, + 0.02911650761961937, + 0.05638928338885307, + 0.021394100040197372, + 0.043946027755737305, + 0.02013206109404564, + 0.0056481314823031425, + 0.010481211356818676, + 0.00627329433336854, + 0.008703080005943775, + 0.014816045761108398, + 0.0054682036861777306, + 0.009897442534565926, + 0.006451848428696394, + 0.004888157360255718, + 0.004391741007566452, + 0.0033652102574706078, + 0.00635554688051343, + 0.008500013500452042, + 0.0054248543456196785, + 0.006273726467043161, + 0.005565730854868889, + 0.007784214802086353, + 0.0059107327833771706, + 0.0038156474474817514, + 0.00300814351066947, + 0.005835443269461393, + 0.003468831768259406, + 0.006892488803714514, + 0.005274481140077114, + 0.008459849283099174, + 0.00485504325479269, + 0.0024065726902335882, + 0.00319864135235548, + 0.0039579859003424644, + 0.002431720495223999, + 0.005310532171279192, + 0.004684030544012785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.029259085655212402 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00980454497039318 + ], + "opposite_logprob_tokens": [ + -4.634804725646973 + ], + "logprob_reward": -0.00980454497039318, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00980454497039318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08779513090848923, + 0.056818921118974686, + 0.5216423273086548, + 0.028765344992280006, + 0.032853927463293076, + 0.07567966729402542, + 0.04157833382487297, + 0.08053845912218094, + 0.042328841984272, + 0.031998973339796066 + ], + "credit_map": [ + 0.008869652636349201, + 0.009918584488332272, + 0.010921841487288475, + 0.008734812960028648, + 0.01638103649020195, + 0.006955283228307962, + 0.011917470023036003, + 0.014096450060606003, + 0.005281293299049139, + 0.00487356586381793, + 0.010003922507166862, + 0.0036278932821005583, + 0.006999138742685318, + 0.006090900860726833, + 0.005026871804147959, + 0.014915340580046177, + 0.022935256361961365, + 0.03186924383044243, + 0.04326576367020607, + 0.019274229183793068, + 0.11505249887704849, + 0.21156685054302216, + 0.03176248073577881, + 0.04591604694724083, + 0.0018125168280676007, + 0.0025112887378782034, + 0.004224343691021204, + 0.003944197669625282, + 0.005381079390645027, + 0.0027207229286432266, + 0.004185715690255165, + 0.0039854818023741245, + 0.002296225167810917, + 0.0035536661744117737, + 0.005445485003292561, + 0.0038723545148968697, + 0.0054504345171153545, + 0.0039035405497998, + 0.004616312682628632, + 0.003715913277119398, + 0.0059862094931304455, + 0.010810894891619682, + 0.007025276776403189, + 0.00796017237007618, + 0.017850082367658615, + 0.007030480541288853, + 0.012873764149844646, + 0.006142794620245695, + 0.00432773120701313, + 0.005103467032313347, + 0.003475877922028303, + 0.004921541083604097, + 0.008738177828490734, + 0.004320620559155941, + 0.0057006580755114555, + 0.0049902633763849735, + 0.007101454772055149, + 0.01025722362101078, + 0.005876830313354731, + 0.01165139488875866, + 0.013367698527872562, + 0.008290225639939308, + 0.01596630923449993, + 0.008027327246963978, + 0.006186481099575758, + 0.005013579502701759, + 0.006445750128477812, + 0.003821135498583317, + 0.004131263121962547, + 0.004892513155937195, + 0.0046200440265238285, + 0.007218078710138798, + 0.004703069571405649, + 0.003572333138436079, + 0.004706212785094976, + 0.004496362991631031, + 0.0036046956665813923, + 0.0034023087937384844, + 0.004986822139471769, + 0.0025271715130656958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.00980454497039318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052597238682210445 + ], + "opposite_logprob_tokens": [ + -5.2552595138549805 + ], + "logprob_reward": -0.0052597238682210445, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052597238682210445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06828685104846954, + 0.05225788429379463, + 0.6819599866867065, + 0.014831805601716042, + 0.026987353339791298, + 0.03967639431357384, + 0.025958865880966187, + 0.036597080528736115, + 0.035275865346193314, + 0.018167922273278236 + ], + "credit_map": [ + 0.006580384448170662, + 0.006849844008684158, + 0.008334669284522533, + 0.007216958794742823, + 0.011351731605827808, + 0.0063612209632992744, + 0.010298388078808784, + 0.011293653398752213, + 0.004216019995510578, + 0.004458008334040642, + 0.008273125626146793, + 0.003857414470985532, + 0.007224030327051878, + 0.0055073522962629795, + 0.005454315338283777, + 0.01326762419193983, + 0.032063230872154236, + 0.04507298395037651, + 0.05625681206583977, + 0.02901615761220455, + 0.1416385918855667, + 0.268604040145874, + 0.04842763394117355, + 0.06088053062558174, + 0.0011787660187110305, + 0.0015325313434004784, + 0.002677727025002241, + 0.0014249967643991113, + 0.002825675532221794, + 0.001121952896937728, + 0.0024891782086342573, + 0.0015809769975021482, + 0.0035448556300252676, + 0.0029044640250504017, + 0.002772182459011674, + 0.0033040703274309635, + 0.004708894062787294, + 0.00241879397071898, + 0.004374805837869644, + 0.002959286328405142, + 0.004718395881354809, + 0.005498421378433704, + 0.003999293316155672, + 0.004549258854240179, + 0.007872391492128372, + 0.0036864150315523148, + 0.006050786469131708, + 0.003301430493593216, + 0.0026710249949246645, + 0.00427953340113163, + 0.0022391846869140863, + 0.0036101965233683586, + 0.0049582431092858315, + 0.0017673232359811664, + 0.003594714915379882, + 0.0028386444319039583, + 0.00413040304556489, + 0.006544426549226046, + 0.002779804402962327, + 0.004651399329304695, + 0.006059015169739723, + 0.003642151365056634, + 0.005603718105703592, + 0.0031861618626862764, + 0.00557877030223608, + 0.003633195534348488, + 0.004168392159044743, + 0.002843925729393959, + 0.004386535380035639, + 0.004638268146663904, + 0.004819116555154324, + 0.005207662936300039, + 0.0038505345582962036, + 0.002180866664275527, + 0.0011385052930563688, + 0.0022904102224856615, + 0.001675017992965877, + 0.0013149271253496408, + 0.0039029086474329233, + 0.0018147514201700687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0052597238682210445 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495700612664223 + ], + "opposite_logprob_tokens": [ + -2.9299569129943848 + ], + "logprob_reward": -0.05495700612664223, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495700612664223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13426841795444489, + 0.06489729136228561, + 0.3614111840724945, + 0.06661227345466614, + 0.028639771044254303, + 0.1931380331516266, + 0.03647124394774437, + 0.06423023343086243, + 0.0285048745572567, + 0.02182670682668686 + ], + "credit_map": [ + 0.012097502127289772, + 0.012677290476858616, + 0.02202225849032402, + 0.013105577789247036, + 0.019633198156952858, + 0.009805799461901188, + 0.02160727232694626, + 0.02331952191889286, + 0.007240313105285168, + 0.008559109643101692, + 0.009252271614968777, + 0.004800476599484682, + 0.011439352296292782, + 0.007210067939013243, + 0.007339335512369871, + 0.009056366048753262, + 0.01782572828233242, + 0.026923874393105507, + 0.03818981721997261, + 0.02146899700164795, + 0.05045883730053902, + 0.13951285183429718, + 0.03338366746902466, + 0.03364739939570427, + 0.0042635295540094376, + 0.007267278153449297, + 0.010783584788441658, + 0.0069860126823186874, + 0.013418051414191723, + 0.006068338640034199, + 0.009894690476357937, + 0.007930790074169636, + 0.0025133281014859676, + 0.0033567894715815783, + 0.0033366442658007145, + 0.00342297600582242, + 0.005022935103625059, + 0.0034281492698937654, + 0.0046933903358876705, + 0.0028655589558184147, + 0.016968466341495514, + 0.033906240016222, + 0.01768013834953308, + 0.020840127021074295, + 0.04232601448893547, + 0.015112506225705147, + 0.03240032121539116, + 0.013904216699302197, + 0.003922141157090664, + 0.004949652589857578, + 0.003911306150257587, + 0.005220573395490646, + 0.005997847765684128, + 0.0034178015775978565, + 0.005073873326182365, + 0.003978049848228693, + 0.009618021547794342, + 0.00824244786053896, + 0.004313938319683075, + 0.00785932969301939, + 0.010159972123801708, + 0.006135464645922184, + 0.007476005703210831, + 0.010425057262182236, + 0.004390465561300516, + 0.0038544617127627134, + 0.002811037702485919, + 0.0025040977634489536, + 0.0043631792068481445, + 0.002487622434273362, + 0.004417556803673506, + 0.0036764536052942276, + 0.0029906618874520063, + 0.002732254099100828, + 0.0029800445772707462, + 0.0036821237299591303, + 0.002506009303033352, + 0.0020374569576233625, + 0.00333596533164382, + 0.0015621909406036139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.05495700612664223 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328178472816944 + ], + "opposite_logprob_tokens": [ + -3.7732818126678467 + ], + "logprob_reward": -0.02328178472816944, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328178472816944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12239432334899902, + 0.1502029448747635, + 0.17193792760372162, + 0.06822218000888824, + 0.11358622461557388, + 0.110201396048069, + 0.08225924521684647, + 0.0729796290397644, + 0.0701332613825798, + 0.03808292746543884 + ], + "credit_map": [ + 0.01679905131459236, + 0.015033265575766563, + 0.020623227581381798, + 0.01162048988044262, + 0.017241034656763077, + 0.011624693870544434, + 0.012190945446491241, + 0.017261624336242676, + 0.017579995095729828, + 0.015023943036794662, + 0.02899377979338169, + 0.007977165281772614, + 0.02133849821984768, + 0.016739368438720703, + 0.011914484202861786, + 0.03063570335507393, + 0.013358647003769875, + 0.029036933556199074, + 0.02309015765786171, + 0.011664021760225296, + 0.043643493205308914, + 0.025861401110887527, + 0.01452317088842392, + 0.010760094970464706, + 0.006461306009441614, + 0.00651146424934268, + 0.011145157739520073, + 0.006741945166140795, + 0.015655478462576866, + 0.005147813353687525, + 0.011520982719957829, + 0.005038035102188587, + 0.01013757474720478, + 0.012752238661050797, + 0.014675132930278778, + 0.01414164062589407, + 0.021473390981554985, + 0.01396376732736826, + 0.016109608113765717, + 0.010332872159779072, + 0.01351187564432621, + 0.016744667664170265, + 0.012333190999925137, + 0.01159599144011736, + 0.01976650208234787, + 0.010494327172636986, + 0.01545153558254242, + 0.010303303599357605, + 0.010257845744490623, + 0.010131195187568665, + 0.006843356415629387, + 0.009778984822332859, + 0.018804317340254784, + 0.0072287400253117085, + 0.012430816888809204, + 0.0067839836701750755, + 0.007973366416990757, + 0.010928100906312466, + 0.006229345221072435, + 0.010828613303601742, + 0.012503020465373993, + 0.00818623136729002, + 0.009431183338165283, + 0.006899770814925432, + 0.010673919692635536, + 0.0062913899309933186, + 0.008375400677323341, + 0.004987090360373259, + 0.00788012333214283, + 0.011489679105579853, + 0.009000963531434536, + 0.011434696614742279, + 0.0058255321346223354, + 0.0033802867401391268, + 0.004088589921593666, + 0.0057314541190862656, + 0.0032088765874505043, + 0.00396354403346777, + 0.006051685195416212, + 0.005832962691783905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.02328178472816944 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, beginning from a slightly overhead angle, the camera slowly zooms in, creating a suffocating, oppressive feeling. In the frame, a young delivery driver hunches over, his bright yellow waterproof uniform glaring against the darkness of the late-night alley, its zipper stained with grease. Cold sweat drenches his messy, short black hair, leaving strands clinging to his skin. The only light in the scene comes from the smartphone he clutches tightly in his hand. The screen's cold blue glow carves deep, tense shadows across his weary face. His gaze is fixed on the order details popping up in the center of the screen, the address \"44 Wangsheng Road\" clearly displayed in bold, bold letters. His index finger, the tip of his fingernail trembling slightly with hesitation, the grime of the day still lingering between his nails, hovers mere millimeters from the screen. A visible, nervous swallow causes his Adam's apple to roll violently. At that moment, the number representing the subsidy amount at the bottom of the screen, initially displaying \"12.8 yuan,\" flickered without warning. The digits blurred and stretched, then, amid a faint electronic hum, instantly reconstructed into an absurdly large number: \"128 yuan.\" His pupils suddenly constricted, his lips parted slightly, his face a mixture of extreme horror, greed, and a deep-seated fear. The entire scene was frozen in an eerie yet alluring atmosphere of suspense.", + "seed": 18597, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_2/index_6307_seed_18597_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.024775570863857864, + "vlm_logprob": -0.024775570863857864, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.024775570863857864, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0207232516258955 + ], + "opposite_logprob_tokens": [ + -4.145723342895508 + ], + "logprob_reward": -0.0207232516258955, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0207232516258955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13956859707832336, + 0.08196283131837845, + 0.059478141367435455, + 0.03220490738749504, + 0.10964752733707428, + 0.16203679144382477, + 0.10900789499282837, + 0.048303112387657166, + 0.1072784885764122, + 0.15051162242889404 + ], + "credit_map": [ + 0.015610948204994202, + 0.01671881414949894, + 0.027577733621001244, + 0.011219620704650879, + 0.02029031701385975, + 0.02504086308181286, + 0.00853395089507103, + 0.014576362445950508, + 0.005571356974542141, + 0.008594397455453873, + 0.0057128132320940495, + 0.0035068036522716284, + 0.017866432666778564, + 0.026630768552422523, + 0.003873901441693306, + 0.010206365957856178, + 0.006544914096593857, + 0.01139089372009039, + 0.006966401822865009, + 0.005313579458743334, + 0.0110380994156003, + 0.006747619714587927, + 0.004292973317205906, + 0.0071836658753454685, + 0.0022075362503528595, + 0.004412039648741484, + 0.0038085097912698984, + 0.005850127432495356, + 0.0039231181144714355, + 0.003977353218942881, + 0.003998932894319296, + 0.004027295857667923, + 0.011839841492474079, + 0.015002152882516384, + 0.005825767293572426, + 0.012605124153196812, + 0.017195284366607666, + 0.005306517239660025, + 0.03441736474633217, + 0.007455501705408096, + 0.010836939327418804, + 0.018735550343990326, + 0.011203894391655922, + 0.026852678507566452, + 0.029705718159675598, + 0.012193968519568443, + 0.02644278109073639, + 0.026065293699502945, + 0.01019211020320654, + 0.013190759345889091, + 0.010501407086849213, + 0.013353761285543442, + 0.019911209121346474, + 0.008726152591407299, + 0.020192274823784828, + 0.012940246611833572, + 0.0037498113233596087, + 0.005706162191927433, + 0.004013659432530403, + 0.005935159977525473, + 0.009034346789121628, + 0.004479832015931606, + 0.010195577517151833, + 0.005188574083149433, + 0.010775781236588955, + 0.017700929194688797, + 0.011255296878516674, + 0.013306641019880772, + 0.01676926203072071, + 0.009691826067864895, + 0.019007399678230286, + 0.008771362714469433, + 0.011824900284409523, + 0.01161830686032772, + 0.011403415352106094, + 0.02960261143743992, + 0.020711800083518028, + 0.030798181891441345, + 0.025283878669142723, + 0.009268553927540779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.0207232516258955 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018225232139229774 + ], + "opposite_logprob_tokens": [ + -4.018225193023682 + ], + "logprob_reward": -0.018225232139229774, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018225232139229774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04439608380198479, + 0.0813140869140625, + 0.025598062202334404, + 0.033264387398958206, + 0.4147232174873352, + 0.03691891208291054, + 0.04940970987081528, + 0.05798816680908203, + 0.041256967931985855, + 0.21513032913208008 + ], + "credit_map": [ + 0.004646476358175278, + 0.005540329497307539, + 0.006007448770105839, + 0.005250074900686741, + 0.006599081680178642, + 0.00488229189068079, + 0.0053497920744121075, + 0.006120595149695873, + 0.008862336166203022, + 0.011814531870186329, + 0.011860957369208336, + 0.007391652557998896, + 0.011150017380714417, + 0.014812727458775043, + 0.005902436096221209, + 0.009519437327980995, + 0.0023231005761772394, + 0.0036048428155481815, + 0.0029048407450318336, + 0.0026936728972941637, + 0.006169622763991356, + 0.002580696949735284, + 0.002428105566650629, + 0.002893180353567004, + 0.0030978433787822723, + 0.0041976263746619225, + 0.004363340325653553, + 0.004231817089021206, + 0.0043934485875070095, + 0.0036412538029253483, + 0.0035889979917556047, + 0.0057500614784657955, + 0.0565672367811203, + 0.060720473527908325, + 0.0253764521330595, + 0.035031937062740326, + 0.056604329496622086, + 0.022849805653095245, + 0.1246897354722023, + 0.03288327157497406, + 0.0022316479589790106, + 0.004567498806864023, + 0.0025875470601022243, + 0.005519383121281862, + 0.005959087982773781, + 0.003192722564563155, + 0.006588619202375412, + 0.006272407714277506, + 0.005617561750113964, + 0.004595233593136072, + 0.00434331689029932, + 0.0061728861182928085, + 0.007624957710504532, + 0.004703516606241465, + 0.008099524304270744, + 0.00825271476060152, + 0.0054443213157355785, + 0.005740699358284473, + 0.005407634191215038, + 0.005489050410687923, + 0.010701494291424751, + 0.00683668814599514, + 0.009200229309499264, + 0.009168055839836597, + 0.004932962823659182, + 0.005929022561758757, + 0.0038872479926794767, + 0.00518398592248559, + 0.006700438447296619, + 0.0039392211474478245, + 0.007410693913698196, + 0.0032733972184360027, + 0.023530473932623863, + 0.019288649782538414, + 0.021324288100004196, + 0.054222121834754944, + 0.019082747399806976, + 0.03700277954339981, + 0.02661174163222313, + 0.014067540876567364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.018225232139229774 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003624300705268979 + ], + "opposite_logprob_tokens": [ + -5.628624439239502 + ], + "logprob_reward": -0.003624300705268979, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003624300705268979, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0676049217581749, + 0.4480791687965393, + 0.04764104261994362, + 0.0259371530264616, + 0.08058011531829834, + 0.0888996571302414, + 0.03983185440301895, + 0.042620670050382614, + 0.021689651533961296, + 0.1371157467365265 + ], + "credit_map": [ + 0.006076687481254339, + 0.007692520972341299, + 0.009956834837794304, + 0.006016395054757595, + 0.009268315508961678, + 0.011876984499394894, + 0.00717758946120739, + 0.009539591148495674, + 0.04895155131816864, + 0.06198890134692192, + 0.05518929660320282, + 0.05200564116239548, + 0.05520595982670784, + 0.06239110603928566, + 0.057004280388355255, + 0.05534243583679199, + 0.004093267489224672, + 0.005677664652466774, + 0.004376738797873259, + 0.006274392828345299, + 0.007530085276812315, + 0.0043128966353833675, + 0.007722793612629175, + 0.00765320286154747, + 0.0019468613900244236, + 0.0024117419961839914, + 0.0026061832904815674, + 0.004131103865802288, + 0.004186619538813829, + 0.002706164726987481, + 0.0035444791428744793, + 0.004403999075293541, + 0.013826747424900532, + 0.007401519920676947, + 0.012962451204657555, + 0.008545822463929653, + 0.007974658161401749, + 0.005629816558212042, + 0.00727531872689724, + 0.01696377620100975, + 0.006165745668113232, + 0.010106037370860577, + 0.006239649374037981, + 0.01538233831524849, + 0.01101372204720974, + 0.008357866667211056, + 0.015742793679237366, + 0.015891501680016518, + 0.003289689077064395, + 0.005527789704501629, + 0.0036273610312491655, + 0.004313609097152948, + 0.006878117565065622, + 0.0051605296321213245, + 0.007231304422020912, + 0.0038034520111978054, + 0.003249239642173052, + 0.0047166370786726475, + 0.0032821076456457376, + 0.004943337291479111, + 0.008701353333890438, + 0.005846648011356592, + 0.0066749220713973045, + 0.005206423345953226, + 0.002573268720880151, + 0.002771563595160842, + 0.0018944598268717527, + 0.0017380027566105127, + 0.003603525459766388, + 0.0033813787158578634, + 0.0034769750200212, + 0.002250476973131299, + 0.013681518845260143, + 0.012246191501617432, + 0.009022588841617107, + 0.03817275911569595, + 0.014182521030306816, + 0.024397145956754684, + 0.016746971756219864, + 0.008666048757731915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 5 + ], + "selected_reward": -0.003624300705268979 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496061593294144 + ], + "opposite_logprob_tokens": [ + -2.9299607276916504 + ], + "logprob_reward": -0.05496061593294144, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496061593294144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06819895654916763, + 0.15395331382751465, + 0.05230829492211342, + 0.11116500198841095, + 0.25149300694465637, + 0.07629972696304321, + 0.04044097661972046, + 0.031120922416448593, + 0.033334068953990936, + 0.18168576061725616 + ], + "credit_map": [ + 0.0074331993237137794, + 0.009162410162389278, + 0.010029124096035957, + 0.006339897867292166, + 0.01133815385401249, + 0.009123479016125202, + 0.005545147694647312, + 0.009227543137967587, + 0.016080765053629875, + 0.02219320833683014, + 0.014937407337129116, + 0.018274083733558655, + 0.016728093847632408, + 0.02120736800134182, + 0.021427983418107033, + 0.02310439571738243, + 0.005296645220369101, + 0.006148214917629957, + 0.005680252332240343, + 0.010130472481250763, + 0.00659334659576416, + 0.00472245691344142, + 0.006635147612541914, + 0.007101761177182198, + 0.007903260178864002, + 0.011015627533197403, + 0.011451608501374722, + 0.0202175360172987, + 0.013646573759615421, + 0.008875706233084202, + 0.019298899918794632, + 0.018755782395601273, + 0.03753669932484627, + 0.03094092383980751, + 0.02531547099351883, + 0.02637942135334015, + 0.029471682384610176, + 0.01560270693153143, + 0.05188621208071709, + 0.03435986861586571, + 0.005004669073969126, + 0.009008761495351791, + 0.005756203550845385, + 0.012430740520358086, + 0.01131616160273552, + 0.006963266059756279, + 0.011635374277830124, + 0.01418454572558403, + 0.005331003572791815, + 0.0046892412938177586, + 0.003562111873179674, + 0.004438192117959261, + 0.006834672763943672, + 0.003639114322140813, + 0.005591406021267176, + 0.006355236750096083, + 0.002388071035966277, + 0.004461080767214298, + 0.0025155311450362206, + 0.0035397792235016823, + 0.006206497084349394, + 0.0021514787804335356, + 0.007062290329486132, + 0.002796194050461054, + 0.0035579227842390537, + 0.005513370502740145, + 0.002124143997207284, + 0.0025423462502658367, + 0.006467872299253941, + 0.004454934969544411, + 0.00555597385391593, + 0.0031175038311630487, + 0.015625933185219765, + 0.010488989762961864, + 0.008027494885027409, + 0.03279385715723038, + 0.03491291031241417, + 0.045303210616111755, + 0.023086067289114, + 0.011447298340499401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 1 + ], + "selected_reward": -0.05496061593294144 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a young delivery driver hunched over?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the driver wearing a bright yellow waterproof uniform with a grease-stained zipper?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature a smartphone with a cold blue glow illuminating the driver's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the screen display the address '44 Wangsheng Road' and the subsidy amount changing from '12.8 yuan' to '128 yuan'?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly overhead angle and slowly zoom in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026344453915953636 + ], + "opposite_logprob_tokens": [ + -3.6513445377349854 + ], + "logprob_reward": -0.026344453915953636, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026344453915953636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05637189373373985, + 0.07408580929040909, + 0.02622421272099018, + 0.01767178624868393, + 0.2989484369754791, + 0.31553128361701965, + 0.046408332884311676, + 0.014555346220731735, + 0.03299324959516525, + 0.11720962077379227 + ], + "credit_map": [ + 0.006084458436816931, + 0.007260399404913187, + 0.008809923194348812, + 0.005389045458287001, + 0.01199513953179121, + 0.00709203677251935, + 0.004169438499957323, + 0.00557145569473505, + 0.0068927062675356865, + 0.009190944954752922, + 0.007794850040227175, + 0.006769618019461632, + 0.007899833843111992, + 0.013180761598050594, + 0.009011080488562584, + 0.01334602851420641, + 0.002617263700813055, + 0.005412860307842493, + 0.0028369957581162453, + 0.0025177327916026115, + 0.005718484055250883, + 0.003665195545181632, + 0.0016567836282774806, + 0.0017989010084420443, + 0.0016687583411112428, + 0.0027542675379663706, + 0.002172641223296523, + 0.0024215520825237036, + 0.0023558717221021652, + 0.0019228599267080426, + 0.00218886393122375, + 0.002186974510550499, + 0.049605049192905426, + 0.03623885661363602, + 0.027703994885087013, + 0.025967413559556007, + 0.04429018869996071, + 0.022167181596159935, + 0.04755989834666252, + 0.04541587829589844, + 0.02143724635243416, + 0.0393640473484993, + 0.022524518892169, + 0.05331652984023094, + 0.04671832546591759, + 0.025970758870244026, + 0.049656473100185394, + 0.056543417274951935, + 0.004087948240339756, + 0.007194458041340113, + 0.003127335337921977, + 0.005974491126835346, + 0.007689072750508785, + 0.004698410164564848, + 0.007306243758648634, + 0.006330377422273159, + 0.001228206790983677, + 0.001577517599798739, + 0.0012551977997645736, + 0.001512321294285357, + 0.0027692466974258423, + 0.001178985694423318, + 0.003810224821791053, + 0.0012236477341502905, + 0.003017245326191187, + 0.00562099227681756, + 0.0026250623632222414, + 0.003988249227404594, + 0.005552741698920727, + 0.002249703276902437, + 0.007390839979052544, + 0.0025484177749603987, + 0.010361799038946629, + 0.007670689839869738, + 0.006272133905440569, + 0.016489626839756966, + 0.029973311349749565, + 0.016279658302664757, + 0.022486431524157524, + 0.0076759811490774155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.026344453915953636 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_3/index_tensor([7692, 7692])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_3/index_tensor([7692, 7692])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..54977dba4058bb4e4a040fa3f7566166a7fc565b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_3/index_tensor([7692, 7692])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "smooth tracking shot, starting from a slightly upward angle, gracefully circles a white-haired young woman in a kimono embellished with ice crystal patterns. She rises steadily and deliberately from a surface of water as deep and perfectly calm as obsidian. A cool overhead light, almost moonlight, shines down vertically, precisely outlining every detail of her fair skin and the long, silvery hair that clings wetly to her cheeks and neck. Her eyes are closed, her face serene and ethereal, with glistening water droplets dangling from her eyelashes. She wears a dark blue silk kimono, the soaked fabric clinging to her slender figure. The sleeves and bodice are meticulously embroidered with intricate, shimmering ice crystal patterns in silver and light blue thread, which reflect a subtle glow in the overhead light. As her upper body continues to rise, the water flows silently down her skin and the folds of her kimono, and ray tracing technology is used to its extreme at this moment, capturing countless tiny water droplets being lifted from the water. They form a layer of shimmering light mist around her body. Each independent, airborne water droplet is like a miniature prism, refracting and reflecting the single beam of overhead light, emitting a delicate and changing rainbow-like brilliance. Against the backdrop of endless darkness, the entire picture presents a mysterious atmosphere intertwined with extreme realism and an otherworldly and tranquil aesthetic.", + "seed": 788, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_3/index_7692_seed_788_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.032894156267866494, + "vlm_logprob": -0.032894156267866494, + "vlm_yes_no_margin": 4.174999952316284, + "vlm_token_credit": -0.032894156267866494, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07760372012853622 + ], + "opposite_logprob_tokens": [ + -2.702603816986084 + ], + "logprob_reward": -0.07760372012853622, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07760372012853622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10853421688079834, + 0.03182215988636017, + 0.02192917838692665, + 0.012568996287882328, + 0.21451815962791443, + 0.013976541347801685, + 0.01311179343611002, + 0.5517334342002869, + 0.017461178824305534, + 0.014344289898872375 + ], + "credit_map": [ + 0.011521405540406704, + 0.013981403782963753, + 0.009068495593965054, + 0.012388234958052635, + 0.011775112710893154, + 0.015464500524103642, + 0.019631942734122276, + 0.014703127555549145, + 0.004766900558024645, + 0.004990814719349146, + 0.0023356880992650986, + 0.002911454066634178, + 0.00392258120700717, + 0.005337063688784838, + 0.003974521532654762, + 0.0035831371787935495, + 0.002932449569925666, + 0.0027506977785378695, + 0.0016857465961948037, + 0.0023779976181685925, + 0.003751573618501425, + 0.0038011905271559954, + 0.002017138758674264, + 0.00261238357052207, + 0.001624155673198402, + 0.0011981342686340213, + 0.0010293642990291119, + 0.001349440193735063, + 0.0018361654365435243, + 0.0025986928958445787, + 0.0012507073115557432, + 0.0016823369078338146, + 0.030528616160154343, + 0.012271867133677006, + 0.025967219844460487, + 0.017362941056489944, + 0.01603824459016323, + 0.03219068795442581, + 0.03939945623278618, + 0.0407591313123703, + 0.0019280207343399525, + 0.0019943094812333584, + 0.0011106778401881456, + 0.0018692818703129888, + 0.002050592564046383, + 0.0021033992525190115, + 0.0014636231353506446, + 0.0014566362369805574, + 0.002386614913120866, + 0.0009747095173224807, + 0.0008066937443800271, + 0.0006649300921708345, + 0.0020216137636452913, + 0.004037420731037855, + 0.0009844184387475252, + 0.0012353912461549044, + 0.21458889544010162, + 0.018745217472314835, + 0.02560664340853691, + 0.010178795084357262, + 0.02026144601404667, + 0.23336808383464813, + 0.009345894679427147, + 0.019638482481241226, + 0.0038907069247215986, + 0.0017455477500334382, + 0.0024733657483011484, + 0.0015974233392626047, + 0.002466612495481968, + 0.002252028789371252, + 0.0010635824874043465, + 0.001971910707652569, + 0.0019211621256545186, + 0.0012789744650945067, + 0.001938316971063614, + 0.0007750420481897891, + 0.0034783247392624617, + 0.0019466654630377889, + 0.0011402246309444308, + 0.0018655791645869613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.07760372012853622 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018218208104372025 + ], + "opposite_logprob_tokens": [ + -4.018218040466309 + ], + "logprob_reward": -0.018218208104372025, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018218208104372025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01374017633497715, + 0.016007229685783386, + 0.08555128425359726, + 0.005389714147895575, + 0.05270924046635628, + 0.0032085557468235493, + 0.010856933891773224, + 0.8064211010932922, + 0.003537506330758333, + 0.002578230807557702 + ], + "credit_map": [ + 0.0014985540183261037, + 0.001742115244269371, + 0.0010761162266135216, + 0.0019023482454940677, + 0.0013601613463833928, + 0.0023276018910109997, + 0.0021686735562980175, + 0.0016646067379042506, + 0.0016590304439887404, + 0.002190707018598914, + 0.001005240948870778, + 0.002494516083970666, + 0.002154847141355276, + 0.0020320650655776262, + 0.002848895499482751, + 0.0016219286480918527, + 0.01454548817127943, + 0.0067636496387422085, + 0.00473321508616209, + 0.01030465867370367, + 0.007617922965437174, + 0.013540380634367466, + 0.010375847108662128, + 0.017670124769210815, + 0.0007232108036987484, + 0.0004351672250777483, + 0.0003571649722289294, + 0.0005980449495837092, + 0.0006640212959609926, + 0.0012446242617443204, + 0.0006442207959480584, + 0.0007232602802105248, + 0.007387314923107624, + 0.0031847748905420303, + 0.006209952291101217, + 0.0042118048295378685, + 0.004077200777828693, + 0.007792395073920488, + 0.009677513502538204, + 0.010168289765715599, + 0.0006061434978619218, + 0.0003230793809052557, + 0.000250471115577966, + 0.00031895743450149894, + 0.0003031147934962064, + 0.0007109914440661669, + 0.00035304194898344576, + 0.00034275639336556196, + 0.002354935510084033, + 0.0005752693978138268, + 0.0005762191140092909, + 0.000432606932008639, + 0.0014114264631643891, + 0.004124874714761972, + 0.000493704981636256, + 0.0008878991939127445, + 0.31808188557624817, + 0.024773165583610535, + 0.039049990475177765, + 0.01354423351585865, + 0.02329901047050953, + 0.3478429615497589, + 0.01290060207247734, + 0.02692931704223156, + 0.00046389459748752415, + 0.0004145610728301108, + 0.0004150948370806873, + 0.000339708843966946, + 0.0006865953910164535, + 0.00045757615589536726, + 0.000322599895298481, + 0.00043747559539042413, + 0.0004051702271681279, + 0.00021223154908511788, + 0.0002892879128921777, + 0.00021670656860806048, + 0.0005866416031494737, + 0.00032291735988110304, + 0.00027362548280507326, + 0.00027165032224729657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.018218208104372025 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205713748931885 + ], + "opposite_logprob_tokens": [ + -2.8120570182800293 + ], + "logprob_reward": -0.06205713748931885, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205713748931885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037004124373197556, + 0.03634347394108772, + 0.09428979456424713, + 0.01326743047684431, + 0.03679652512073517, + 0.015296248719096184, + 0.0876508578658104, + 0.6438307762145996, + 0.02737312950193882, + 0.008147578686475754 + ], + "credit_map": [ + 0.004217181820422411, + 0.004132759757339954, + 0.003271395107731223, + 0.0040208520367741585, + 0.0035857688635587692, + 0.007126510143280029, + 0.005482992623001337, + 0.005166670773178339, + 0.0041906447149813175, + 0.004763626027852297, + 0.0022877517621964216, + 0.005832680035382509, + 0.004516688175499439, + 0.00528758205473423, + 0.005870105233043432, + 0.003594399197027087, + 0.015396184287965298, + 0.008037970401346684, + 0.0049025267362594604, + 0.010142366401851177, + 0.009240513667464256, + 0.013757116161286831, + 0.012727602384984493, + 0.02008552849292755, + 0.0021459825802594423, + 0.001436758553609252, + 0.0007595491479150951, + 0.001029284088872373, + 0.0012156222946941853, + 0.0037049136590212584, + 0.0018106066854670644, + 0.0011647152714431286, + 0.004473203327506781, + 0.00396930705755949, + 0.0032637896947562695, + 0.0025521842762827873, + 0.006752990186214447, + 0.00523692462593317, + 0.004977714270353317, + 0.005570416338741779, + 0.002018066355958581, + 0.001803696621209383, + 0.0012816897360607982, + 0.0020506810396909714, + 0.0018074304098263383, + 0.0027659465558826923, + 0.001767451991327107, + 0.0018012874061241746, + 0.0189500842243433, + 0.004852685146033764, + 0.004859066568315029, + 0.0033641282934695482, + 0.010625172406435013, + 0.03334050625562668, + 0.004503997974097729, + 0.007155223749577999, + 0.25331979990005493, + 0.020197998732328415, + 0.03172043338418007, + 0.010526390746235847, + 0.018941989168524742, + 0.27738022804260254, + 0.010437331162393093, + 0.021306701004505157, + 0.006816196255385876, + 0.0021540075540542603, + 0.004381662700325251, + 0.0016980969812721014, + 0.0033517677802592516, + 0.0035788374952971935, + 0.0020359118934720755, + 0.003356651868671179, + 0.001644224626943469, + 0.0006727914442308247, + 0.0010847024386748672, + 0.0005052502965554595, + 0.001482713152654469, + 0.0011029375018551946, + 0.0006130221881903708, + 0.0010419378522783518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.06205713748931885 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101199563592672 + ], + "opposite_logprob_tokens": [ + -5.504101276397705 + ], + "logprob_reward": -0.004101199563592672, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101199563592672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02457238733768463, + 0.0307494904845953, + 0.04929988831281662, + 0.014338713139295578, + 0.050050295889377594, + 0.007461713161319494, + 0.03974597901105881, + 0.7683473825454712, + 0.008331185206770897, + 0.007103028707206249 + ], + "credit_map": [ + 0.002970469184219837, + 0.002350343856960535, + 0.0016298630507662892, + 0.0023386343382298946, + 0.0028683310374617577, + 0.004285755101591349, + 0.0051749469712376595, + 0.0029540425166487694, + 0.003983964677900076, + 0.004774799570441246, + 0.003274187445640564, + 0.0033614037092775106, + 0.003734905505552888, + 0.004028256516903639, + 0.0037926388904452324, + 0.003799332771450281, + 0.008743386715650558, + 0.005095442291349173, + 0.0031128886621445417, + 0.005248941481113434, + 0.005592812784016132, + 0.007429881952702999, + 0.005717514548450708, + 0.00835901778191328, + 0.0017598093254491687, + 0.0017064000712707639, + 0.0008812773739919066, + 0.001858748379163444, + 0.0018005338497459888, + 0.0022985099349170923, + 0.001783067942596972, + 0.0022503656800836325, + 0.006112307775765657, + 0.004103514831513166, + 0.004676414653658867, + 0.003699901746585965, + 0.007834994234144688, + 0.006532449275255203, + 0.008711555041372776, + 0.00837915949523449, + 0.0010278308764100075, + 0.0009314448107033968, + 0.0005923611461184919, + 0.0005938121466897428, + 0.0009381918935105205, + 0.0013651144690811634, + 0.0012109681265428662, + 0.0008019895758479834, + 0.008666528388857841, + 0.0018780879909172654, + 0.002240701112896204, + 0.0013969086576253176, + 0.0049536521546542645, + 0.015362139791250229, + 0.0016224157297983766, + 0.0036255440209060907, + 0.303290456533432, + 0.02358875796198845, + 0.037164751440286636, + 0.012710950337350368, + 0.022074036300182343, + 0.3317425549030304, + 0.012235142290592194, + 0.02554069645702839, + 0.001351367449387908, + 0.0010281866416335106, + 0.001109127071686089, + 0.0006181825301609933, + 0.0015194740844890475, + 0.0010504767997190356, + 0.0007565535488538444, + 0.0008978168480098248, + 0.0012720845406875014, + 0.000525181065313518, + 0.0008818142232485116, + 0.0004956582561135292, + 0.0015634284354746342, + 0.000960740027949214, + 0.0006247544661164284, + 0.0007793678087182343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.004101199563592672 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024905160535126925 + ], + "opposite_logprob_tokens": [ + -6.002490520477295 + ], + "logprob_reward": -0.0024905160535126925, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024905160535126925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053342826664447784, + 0.03175070881843567, + 0.18483716249465942, + 0.012534638866782188, + 0.2089463323354721, + 0.016415946185588837, + 0.11977661401033401, + 0.33457499742507935, + 0.023842304944992065, + 0.01397842075675726 + ], + "credit_map": [ + 0.006405423395335674, + 0.007375237997621298, + 0.005423770751804113, + 0.005828279070556164, + 0.005931418389081955, + 0.0063194213435053825, + 0.009600698947906494, + 0.006458577699959278, + 0.004316996317356825, + 0.004008202813565731, + 0.00239744083955884, + 0.003881711745634675, + 0.0039664460346102715, + 0.005030933767557144, + 0.005308473017066717, + 0.0028405040502548218, + 0.030548961833119392, + 0.015784339979290962, + 0.01073843240737915, + 0.02165159396827221, + 0.017137514427304268, + 0.025937415659427643, + 0.023184021934866905, + 0.039854880422353745, + 0.0019295624224469066, + 0.0017678869189694524, + 0.0007749007199890912, + 0.0009462728048674762, + 0.0019452093401923776, + 0.0029361313208937645, + 0.001029362203553319, + 0.0012053130194544792, + 0.027132729068398476, + 0.015418454073369503, + 0.0209700595587492, + 0.014957822859287262, + 0.02756447158753872, + 0.029927510768175125, + 0.036483559757471085, + 0.03649172931909561, + 0.0025075506418943405, + 0.0018905468750745058, + 0.0013801628956571221, + 0.0023953707423061132, + 0.0018341625109314919, + 0.003143643494695425, + 0.001607318758033216, + 0.0016571892192587256, + 0.025382813066244125, + 0.004908237140625715, + 0.006713998969644308, + 0.003507198765873909, + 0.01664407178759575, + 0.04628210514783859, + 0.00538205448538065, + 0.010956128127872944, + 0.12940022349357605, + 0.011097365990281105, + 0.01722201704978943, + 0.00555943138897419, + 0.011536380276083946, + 0.142046719789505, + 0.006298911292105913, + 0.011413974687457085, + 0.006188858300447464, + 0.0018109101802110672, + 0.004175201989710331, + 0.0012967761140316725, + 0.002935795346274972, + 0.0033944537863135338, + 0.0014521351549774408, + 0.0025881724432110786, + 0.002160408301278949, + 0.0009005606989376247, + 0.0021338972728699446, + 0.0007808036752976477, + 0.0031047456432133913, + 0.0020553660579025745, + 0.0009775160579010844, + 0.0018651235150173306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.0024905160535126925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth tracking shot, starting from a slightly upward angle, gracefully circles a white-haired young woman in a kimono embellished with ice crystal patterns. She rises steadily and deliberately from a surface of water as deep and perfectly calm as obsidian. A cool overhead light, almost moonlight, shines down vertically, precisely outlining every detail of her fair skin and the long, silvery hair that clings wetly to her cheeks and neck. Her eyes are closed, her face serene and ethereal, with glistening water droplets dangling from her eyelashes. She wears a dark blue silk kimono, the soaked fabric clinging to her slender figure. The sleeves and bodice are meticulously embroidered with intricate, shimmering ice crystal patterns in silver and light blue thread, which reflect a subtle glow in the overhead light. As her upper body continues to rise, the water flows silently down her skin and the folds of her kimono, and ray tracing technology is used to its extreme at this moment, capturing countless tiny water droplets being lifted from the water. They form a layer of shimmering light mist around her body. Each independent, airborne water droplet is like a miniature prism, refracting and reflecting the single beam of overhead light, emitting a delicate and changing rainbow-like brilliance. Against the backdrop of endless darkness, the entire picture presents a mysterious atmosphere intertwined with extreme realism and an otherworldly and tranquil aesthetic.", + "seed": 789, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_3/index_7692_seed_789_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.019505420140922068, + "vlm_logprob": -0.019505420140922068, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.019505420140922068, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029252950102090836 + ], + "opposite_logprob_tokens": [ + -3.9042530059814453 + ], + "logprob_reward": -0.029252950102090836, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.029252950102090836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16318780183792114, + 0.09842319786548615, + 0.0680592954158783, + 0.09300859272480011, + 0.06805931031703949, + 0.04185776039958, + 0.13133279979228973, + 0.07720495760440826, + 0.20645181834697723, + 0.052414435893297195 + ], + "credit_map": [ + 0.01948971301317215, + 0.01640894263982773, + 0.020325541496276855, + 0.020645370706915855, + 0.019006310030817986, + 0.01917036436498165, + 0.015554026700556278, + 0.03258754685521126, + 0.016165614128112793, + 0.01766275428235531, + 0.013291502371430397, + 0.009813807904720306, + 0.009638930670917034, + 0.012715290300548077, + 0.00987701304256916, + 0.009258300065994263, + 0.006879535503685474, + 0.007228718604892492, + 0.0052240099757909775, + 0.006819050759077072, + 0.00744576146826148, + 0.013325857929885387, + 0.005330459214746952, + 0.01580590195953846, + 0.007172833196818829, + 0.010485874488949776, + 0.008054030127823353, + 0.012115460820496082, + 0.005980327725410461, + 0.01995156891644001, + 0.007615308277308941, + 0.021633192896842957, + 0.006672675721347332, + 0.006280982866883278, + 0.00790758989751339, + 0.010851250030100346, + 0.00570214306935668, + 0.011686013080179691, + 0.006042391527444124, + 0.01291626412421465, + 0.0029272204264998436, + 0.004856155253946781, + 0.004929693415760994, + 0.00606918428093195, + 0.00602237693965435, + 0.008414502255618572, + 0.0034693293273448944, + 0.00516930129379034, + 0.008219030685722828, + 0.010351368226110935, + 0.009796109050512314, + 0.011724602431058884, + 0.03030179813504219, + 0.02204112894833088, + 0.019988704472780228, + 0.01891007460653782, + 0.007691123988479376, + 0.010040218010544777, + 0.005158740561455488, + 0.009093403816223145, + 0.010883667506277561, + 0.014556589536368847, + 0.008933545090258121, + 0.010847668163478374, + 0.019079195335507393, + 0.0423872135579586, + 0.011092687956988811, + 0.01665390282869339, + 0.022125642746686935, + 0.038674138486385345, + 0.015583602711558342, + 0.040855444967746735, + 0.009422538802027702, + 0.007606974802911282, + 0.00448344461619854, + 0.006773735862225294, + 0.006185435224324465, + 0.006015603430569172, + 0.0056602079421281815, + 0.006266497075557709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.029252950102090836 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060823142528534 + ], + "opposite_logprob_tokens": [ + -3.8956081867218018 + ], + "logprob_reward": -0.02060823142528534, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060823142528534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07850434631109238, + 0.0512525700032711, + 0.11078748106956482, + 0.08795719593763351, + 0.09765494614839554, + 0.04545213282108307, + 0.1114296242594719, + 0.09898684918880463, + 0.27370062470436096, + 0.044274233281612396 + ], + "credit_map": [ + 0.011202993802726269, + 0.0069913375191390514, + 0.010854807682335377, + 0.014254760928452015, + 0.011209971271455288, + 0.006456988863646984, + 0.010258628986775875, + 0.007274858187884092, + 0.009007180109620094, + 0.0041567618027329445, + 0.009170515462756157, + 0.006249624770134687, + 0.004847268108278513, + 0.007318450137972832, + 0.004863473121076822, + 0.005639304872602224, + 0.012883003801107407, + 0.012423790991306305, + 0.013345349580049515, + 0.01272462960332632, + 0.01207039225846529, + 0.018795786425471306, + 0.010261518880724907, + 0.01828301139175892, + 0.007612177170813084, + 0.01015431061387062, + 0.0055372584611177444, + 0.01812247931957245, + 0.008121396414935589, + 0.013919382356107235, + 0.010773385874927044, + 0.01371681410819292, + 0.007202234119176865, + 0.009382287971675396, + 0.010453563183546066, + 0.02429896965622902, + 0.009633827023208141, + 0.01388629898428917, + 0.008245760574936867, + 0.014552013948559761, + 0.0037077004089951515, + 0.006330252159386873, + 0.005187357310205698, + 0.005511499010026455, + 0.005953195504844189, + 0.008941568434238434, + 0.0034962932113558054, + 0.006324268877506256, + 0.0074785007163882256, + 0.008588195778429508, + 0.007977197878062725, + 0.006410763133317232, + 0.01780129410326481, + 0.03275122120976448, + 0.008939850144088268, + 0.02148260921239853, + 0.007623475510627031, + 0.013344665989279747, + 0.006419699173420668, + 0.013346043415367603, + 0.011480210348963737, + 0.018064765259623528, + 0.016602111980319023, + 0.012105891481041908, + 0.021204311400651932, + 0.016615869477391243, + 0.008979197591543198, + 0.013417555019259453, + 0.019937677308917046, + 0.09043684601783752, + 0.011743998154997826, + 0.09136518836021423, + 0.01000219490379095, + 0.005697005428373814, + 0.004430945031344891, + 0.005722308065742254, + 0.004486996680498123, + 0.0049651190638542175, + 0.004008762538433075, + 0.004960907157510519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 2 + ], + "selected_reward": -0.02060823142528534 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810384124517441 + ], + "opposite_logprob_tokens": [ + -3.2881038188934326 + ], + "logprob_reward": -0.03810384124517441, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810384124517441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10939977318048477, + 0.11785145103931427, + 0.22298020124435425, + 0.04924871399998665, + 0.037376195192337036, + 0.07902228087186813, + 0.0628528743982315, + 0.025961991399526596, + 0.17503809928894043, + 0.12026837468147278 + ], + "credit_map": [ + 0.016294347122311592, + 0.00802040845155716, + 0.015271835960447788, + 0.01934289000928402, + 0.011797505430877209, + 0.011279651895165443, + 0.012028065510094166, + 0.015365061350166798, + 0.033772509545087814, + 0.00990053080022335, + 0.01896488480269909, + 0.008612235076725483, + 0.008337545208632946, + 0.01773003302514553, + 0.010051392018795013, + 0.010482323355972767, + 0.020901480689644814, + 0.030710220336914062, + 0.024333935230970383, + 0.027372680604457855, + 0.024935288354754448, + 0.03854554146528244, + 0.019515886902809143, + 0.0366651676595211, + 0.00484346691519022, + 0.005154795479029417, + 0.0026692249812185764, + 0.004783914890140295, + 0.004431623034179211, + 0.009662583470344543, + 0.006002062000334263, + 0.011701040901243687, + 0.002745268866419792, + 0.0051657892763614655, + 0.0025796506088227034, + 0.0063805291429162025, + 0.004307111259549856, + 0.006144993007183075, + 0.004253126215189695, + 0.005799728445708752, + 0.004298628307878971, + 0.008174382150173187, + 0.007331541273742914, + 0.008049055933952332, + 0.008553936146199703, + 0.019047746434807777, + 0.006949989590793848, + 0.016616998240351677, + 0.005055700894445181, + 0.0045236265286803246, + 0.004953529220074415, + 0.005390843376517296, + 0.008541411720216274, + 0.01635044626891613, + 0.005011821165680885, + 0.013025498017668724, + 0.0027146246284246445, + 0.0039457837119698524, + 0.0016325413016602397, + 0.002489353297278285, + 0.004739866126328707, + 0.0036620174068957567, + 0.002939656376838684, + 0.003838149132207036, + 0.011203423142433167, + 0.012732471339404583, + 0.007292490918189287, + 0.006384451873600483, + 0.013152572326362133, + 0.05810145288705826, + 0.007680465467274189, + 0.058490775525569916, + 0.024752089753746986, + 0.012678191997110844, + 0.010114418342709541, + 0.012352571822702885, + 0.015187311917543411, + 0.015159184113144875, + 0.013569191098213196, + 0.01645541377365589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.03810384124517441 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006742346566170454 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.006742346566170454, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006742346566170454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1016593649983406, + 0.04207446798682213, + 0.15654823184013367, + 0.058769553899765015, + 0.05411132797598839, + 0.04724650830030441, + 0.27708229422569275, + 0.0597040094435215, + 0.12684227526187897, + 0.07596195489168167 + ], + "credit_map": [ + 0.013311202637851238, + 0.009866020642220974, + 0.01513840351253748, + 0.018368816003203392, + 0.010498730465769768, + 0.011730631813406944, + 0.009138901717960835, + 0.013606663793325424, + 0.005951836239546537, + 0.006373335141688585, + 0.0054700542241334915, + 0.004320613108575344, + 0.005112724378705025, + 0.0053685009479522705, + 0.004493465181440115, + 0.004983937367796898, + 0.014949426054954529, + 0.014546442776918411, + 0.020847748965024948, + 0.01495045330375433, + 0.017215223982930183, + 0.02826954424381256, + 0.01568247191607952, + 0.0300869382917881, + 0.004215604159981012, + 0.004881414584815502, + 0.003188665257766843, + 0.0069898455403745174, + 0.004706432577222586, + 0.012843855656683445, + 0.006509570404887199, + 0.015434171073138714, + 0.00459953723475337, + 0.006211473606526852, + 0.005105506628751755, + 0.011647014878690243, + 0.005489112343639135, + 0.007465280592441559, + 0.005941647104918957, + 0.007651756517589092, + 0.00329203880392015, + 0.004528116900473833, + 0.005980226211249828, + 0.004671747330576181, + 0.009090237319469452, + 0.007779582869261503, + 0.005707541014999151, + 0.006197023205459118, + 0.01319125946611166, + 0.021843155845999718, + 0.01776055246591568, + 0.01707073487341404, + 0.06863130629062653, + 0.06678397953510284, + 0.019670682027935982, + 0.05213063582777977, + 0.005178391933441162, + 0.006898813880980015, + 0.005004864186048508, + 0.007807201240211725, + 0.00927301961928606, + 0.009095526300370693, + 0.009593913331627846, + 0.006852282676845789, + 0.01542219240218401, + 0.020839298143982887, + 0.008879844099283218, + 0.007739502005279064, + 0.01861036941409111, + 0.024925490841269493, + 0.011113668791949749, + 0.019311919808387756, + 0.014475696720182896, + 0.008653191849589348, + 0.007412756327539682, + 0.009341186843812466, + 0.008356599137187004, + 0.008212688378989697, + 0.009703514166176319, + 0.009806322865188122 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.006742346566170454 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819731365889311 + ], + "opposite_logprob_tokens": [ + -5.877819538116455 + ], + "logprob_reward": -0.002819731365889311, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819731365889311, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17329581081867218, + 0.12592580914497375, + 0.06529560685157776, + 0.06335967779159546, + 0.04421120882034302, + 0.04858662560582161, + 0.08372896909713745, + 0.08582387119531631, + 0.14465266466140747, + 0.16511981189250946 + ], + "credit_map": [ + 0.0159232746809721, + 0.020734716206789017, + 0.020930325612425804, + 0.028961339965462685, + 0.027721522375941277, + 0.019319843500852585, + 0.018809322267770767, + 0.02089546248316765, + 0.019218837842345238, + 0.011511159129440784, + 0.019361376762390137, + 0.017374461516737938, + 0.016026124358177185, + 0.01664927788078785, + 0.013592229224741459, + 0.012192336842417717, + 0.005966609809547663, + 0.007562065962702036, + 0.006525018718093634, + 0.007279158569872379, + 0.007843793369829655, + 0.012797953560948372, + 0.0055655487813055515, + 0.011755455285310745, + 0.004808072000741959, + 0.006351337768137455, + 0.004532069433480501, + 0.009823634289205074, + 0.0052300686948001385, + 0.011873265728354454, + 0.007729646284133196, + 0.013011579401791096, + 0.004434450063854456, + 0.004443466197699308, + 0.0038425689563155174, + 0.0077283382415771484, + 0.004282236099243164, + 0.006929953116923571, + 0.005130202975124121, + 0.007419992703944445, + 0.002660255879163742, + 0.00365350185893476, + 0.00379534182138741, + 0.004727295134216547, + 0.005700035486370325, + 0.01235975418239832, + 0.004437597468495369, + 0.011252837255597115, + 0.007012820336967707, + 0.0072291335090994835, + 0.0059474557638168335, + 0.005437573883682489, + 0.019821589812636375, + 0.01732737384736538, + 0.008479698561131954, + 0.012473327107727528, + 0.006366652902215719, + 0.011085796169936657, + 0.005651259794831276, + 0.009382877498865128, + 0.010709271766245365, + 0.018683912232518196, + 0.013301822356879711, + 0.010642272420227528, + 0.01629488356411457, + 0.016748959198594093, + 0.008696404285728931, + 0.011703708209097385, + 0.015160900540649891, + 0.03686470910906792, + 0.009423552080988884, + 0.02975952997803688, + 0.04937519133090973, + 0.022612739354372025, + 0.011326927691698074, + 0.011671208776533604, + 0.016977904364466667, + 0.02078133076429367, + 0.011417335830628872, + 0.020957166329026222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.002819731365889311 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth tracking shot, starting from a slightly upward angle, gracefully circles a white-haired young woman in a kimono embellished with ice crystal patterns. She rises steadily and deliberately from a surface of water as deep and perfectly calm as obsidian. A cool overhead light, almost moonlight, shines down vertically, precisely outlining every detail of her fair skin and the long, silvery hair that clings wetly to her cheeks and neck. Her eyes are closed, her face serene and ethereal, with glistening water droplets dangling from her eyelashes. She wears a dark blue silk kimono, the soaked fabric clinging to her slender figure. The sleeves and bodice are meticulously embroidered with intricate, shimmering ice crystal patterns in silver and light blue thread, which reflect a subtle glow in the overhead light. As her upper body continues to rise, the water flows silently down her skin and the folds of her kimono, and ray tracing technology is used to its extreme at this moment, capturing countless tiny water droplets being lifted from the water. They form a layer of shimmering light mist around her body. Each independent, airborne water droplet is like a miniature prism, refracting and reflecting the single beam of overhead light, emitting a delicate and changing rainbow-like brilliance. Against the backdrop of endless darkness, the entire picture presents a mysterious atmosphere intertwined with extreme realism and an otherworldly and tranquil aesthetic.", + "seed": 790, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_3/index_7692_seed_790_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017346821445971727, + "vlm_logprob": -0.017346821445971727, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.017346821445971727, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03890414908528328 + ], + "opposite_logprob_tokens": [ + -3.6639041900634766 + ], + "logprob_reward": -0.03890414908528328, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03890414908528328, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14531466364860535, + 0.10770118236541748, + 0.037912506610155106, + 0.164038747549057, + 0.16506963968276978, + 0.0941225066781044, + 0.06696324050426483, + 0.0698896050453186, + 0.0903077945113182, + 0.05868016555905342 + ], + "credit_map": [ + 0.023959049955010414, + 0.015973122790455818, + 0.01844949461519718, + 0.01404298935085535, + 0.012469255365431309, + 0.020756350830197334, + 0.015500493347644806, + 0.024163899943232536, + 0.014088204130530357, + 0.01900317706167698, + 0.014674119651317596, + 0.011035137809813023, + 0.012623485177755356, + 0.012839292176067829, + 0.012579765170812607, + 0.010857990942895412, + 0.0062248981557786465, + 0.005187411326915026, + 0.0027061111759394407, + 0.004387750755995512, + 0.004010419826954603, + 0.00525608379393816, + 0.0033626346848905087, + 0.006777193862944841, + 0.015247241593897343, + 0.02305639162659645, + 0.01396774873137474, + 0.03724292293190956, + 0.012173805385828018, + 0.020563721656799316, + 0.02454277314245701, + 0.0172441303730011, + 0.023801181465387344, + 0.011529631912708282, + 0.014630106277763844, + 0.026750050485134125, + 0.009859000332653522, + 0.02509925328195095, + 0.01570863649249077, + 0.03769176825881004, + 0.01042230986058712, + 0.010913318023085594, + 0.0083644799888134, + 0.010085683315992355, + 0.009533288888633251, + 0.011397644877433777, + 0.01791127398610115, + 0.015494497492909431, + 0.00589553639292717, + 0.007053961977362633, + 0.007579469587653875, + 0.011926877312362194, + 0.005873542744666338, + 0.008151175454258919, + 0.012917469255626202, + 0.007565201260149479, + 0.009229524992406368, + 0.005451733246445656, + 0.018620453774929047, + 0.00297291437163949, + 0.004167434759438038, + 0.007241805084049702, + 0.003089378122240305, + 0.019116360694169998, + 0.010031742043793201, + 0.013646258041262627, + 0.011346797458827496, + 0.013869719579815865, + 0.011211594566702843, + 0.01211385615170002, + 0.007279177196323872, + 0.010808645747601986, + 0.006945845205336809, + 0.006030416581779718, + 0.005810366477817297, + 0.006498732604086399, + 0.006760045420378447, + 0.007068098522722721, + 0.011837797239422798, + 0.007728861179202795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.03890414908528328 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026368720456957817 + ], + "opposite_logprob_tokens": [ + -3.6513686180114746 + ], + "logprob_reward": -0.026368720456957817, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026368720456957817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18658335506916046, + 0.08419082313776016, + 0.09279665350914001, + 0.15691716969013214, + 0.12017448246479034, + 0.09699764847755432, + 0.061973489820957184, + 0.1389440894126892, + 0.032256461679935455, + 0.029165828600525856 + ], + "credit_map": [ + 0.034251175820827484, + 0.015248352661728859, + 0.024260444566607475, + 0.021619407460093498, + 0.012666197493672371, + 0.027386102825403214, + 0.017771147191524506, + 0.033380523324012756, + 0.011129642836749554, + 0.014071501791477203, + 0.011609235778450966, + 0.008846652694046497, + 0.008472666144371033, + 0.011043133214116096, + 0.009066351689398289, + 0.00995163805782795, + 0.017022186890244484, + 0.01136582251638174, + 0.009074036963284016, + 0.010267064906656742, + 0.007217646576464176, + 0.01395412441343069, + 0.007647618185728788, + 0.016248149797320366, + 0.013798482716083527, + 0.017115673050284386, + 0.012646401301026344, + 0.04502783343195915, + 0.01056214515119791, + 0.022137045860290527, + 0.018831240013241768, + 0.01679835468530655, + 0.017281098291277885, + 0.006069817580282688, + 0.014630953781306744, + 0.011567366309463978, + 0.004729097709059715, + 0.023141955956816673, + 0.007745642215013504, + 0.035008542239665985, + 0.01291736401617527, + 0.010266188532114029, + 0.01309512834995985, + 0.009153839200735092, + 0.012340091168880463, + 0.01019606739282608, + 0.01733008772134781, + 0.011698891408741474, + 0.011319024488329887, + 0.006155057810246944, + 0.006532934959977865, + 0.0049530756659805775, + 0.006123214028775692, + 0.012499557808041573, + 0.006104154512286186, + 0.008286470547318459, + 0.010517865419387817, + 0.006551600527018309, + 0.02221936546266079, + 0.005611997563391924, + 0.005869368091225624, + 0.033500757068395615, + 0.006442296784371138, + 0.048230838030576706, + 0.003571677254512906, + 0.005445469636470079, + 0.003758092876523733, + 0.0038912517484277487, + 0.0036654428113251925, + 0.005496959201991558, + 0.0030015327502042055, + 0.0034260316751897335, + 0.0032539074309170246, + 0.002383328741416335, + 0.003366631455719471, + 0.0029348833486437798, + 0.0035565586294978857, + 0.006160771939903498, + 0.0031335968524217606, + 0.004376152064651251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.026368720456957817 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607024483382702 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.01607024483382702, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607024483382702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25124749541282654, + 0.15745379030704498, + 0.038200099021196365, + 0.06609600782394409, + 0.21919895708560944, + 0.07541344314813614, + 0.07211977988481522, + 0.05186847224831581, + 0.039764322340488434, + 0.028637617826461792 + ], + "credit_map": [ + 0.04641500860452652, + 0.01850075088441372, + 0.038353513926267624, + 0.026590198278427124, + 0.016464712098240852, + 0.03659864887595177, + 0.02316410467028618, + 0.04516055807471275, + 0.03356636315584183, + 0.023060888051986694, + 0.018038440495729446, + 0.012664577923715115, + 0.017447900027036667, + 0.023793460801243782, + 0.013423374854028225, + 0.015458788722753525, + 0.0061406721360981464, + 0.004411418456584215, + 0.003192862030118704, + 0.004933727439492941, + 0.00286728423088789, + 0.005193429533392191, + 0.003962545655667782, + 0.007498159073293209, + 0.008892787620425224, + 0.007446983829140663, + 0.005807730834931135, + 0.012815373949706554, + 0.00512559013441205, + 0.009666104800999165, + 0.007971630431711674, + 0.008369806222617626, + 0.03112187795341015, + 0.008779865689575672, + 0.02664312906563282, + 0.020872067660093307, + 0.008130131289362907, + 0.043779876083135605, + 0.014776458963751793, + 0.06509555131196976, + 0.011357858777046204, + 0.009777520783245564, + 0.009230989031493664, + 0.005656799767166376, + 0.010280652903020382, + 0.009557844139635563, + 0.010150729678571224, + 0.009401049464941025, + 0.0070717474445700645, + 0.009232548996806145, + 0.0059984782710671425, + 0.007472529541701078, + 0.007404144387692213, + 0.0150088956579566, + 0.007268533576279879, + 0.012662898749113083, + 0.00620036106556654, + 0.004222100134938955, + 0.006136645562946796, + 0.002852494828402996, + 0.003613903420045972, + 0.013098503462970257, + 0.003206385998055339, + 0.012538080103695393, + 0.004336124751716852, + 0.006018638610839844, + 0.0048844460397958755, + 0.004564348142594099, + 0.0049934228882193565, + 0.006879420951008797, + 0.0036679659970104694, + 0.004419953096657991, + 0.002657902194187045, + 0.002369966357946396, + 0.003142984351143241, + 0.0028742016293108463, + 0.003022853285074234, + 0.005550149828195572, + 0.004290517419576645, + 0.004729042295366526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.01607024483382702 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002199494047090411 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.002199494047090411, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002199494047090411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17652752995491028, + 0.06143388897180557, + 0.03517784923315048, + 0.0304891187697649, + 0.12745513021945953, + 0.15018120408058167, + 0.1678571105003357, + 0.15392814576625824, + 0.04885329306125641, + 0.048096805810928345 + ], + "credit_map": [ + 0.02530919387936592, + 0.015889113768935204, + 0.022555997595191002, + 0.018517374992370605, + 0.017652412876486778, + 0.026015549898147583, + 0.013614777475595474, + 0.03697311878204346, + 0.015382975339889526, + 0.006290819961577654, + 0.006613812875002623, + 0.005288304295390844, + 0.004217897076159716, + 0.010717789642512798, + 0.0047039445489645, + 0.00821834709495306, + 0.005543416365981102, + 0.0036917617544531822, + 0.002515884116292, + 0.0033854658249765635, + 0.0026031294837594032, + 0.005322627257555723, + 0.0030724701937288046, + 0.009043095633387566, + 0.004164817743003368, + 0.0033472164068371058, + 0.0029984896536916494, + 0.0051992908120155334, + 0.0018380009569227695, + 0.0040308949537575245, + 0.0030666179955005646, + 0.005843790713697672, + 0.01811029762029648, + 0.007728454191237688, + 0.01485089398920536, + 0.012180396355688572, + 0.006232835818082094, + 0.023542746901512146, + 0.012317328713834286, + 0.032492175698280334, + 0.03354443982243538, + 0.016736190766096115, + 0.015214781276881695, + 0.013199720531702042, + 0.01243711356073618, + 0.019727949053049088, + 0.01868792437016964, + 0.020633095875382423, + 0.02800767682492733, + 0.014725401066243649, + 0.011792760342359543, + 0.01243120152503252, + 0.013163268566131592, + 0.03962872549891472, + 0.013314073905348778, + 0.03479400649666786, + 0.024493012577295303, + 0.00314162066206336, + 0.059384554624557495, + 0.003952817991375923, + 0.006012863013893366, + 0.03978503867983818, + 0.004312518984079361, + 0.012845722958445549, + 0.006075325887650251, + 0.007316195871680975, + 0.005817859899252653, + 0.005438970867544413, + 0.006005492992699146, + 0.00790857058018446, + 0.004438554402440786, + 0.005852323025465012, + 0.007375351153314114, + 0.0031398122664541006, + 0.006989422254264355, + 0.0039476053789258, + 0.0035934492480009794, + 0.012080573476850986, + 0.0036088330671191216, + 0.007361759431660175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.002199494047090411 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031914988067001104 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031914988067001104, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031914988067001104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2714417576789856, + 0.08914587646722794, + 0.04514872282743454, + 0.06143607571721077, + 0.19737324118614197, + 0.06282318383455276, + 0.06250744313001633, + 0.13404962420463562, + 0.03959352895617485, + 0.03648059070110321 + ], + "credit_map": [ + 0.0351857952773571, + 0.03313835710287094, + 0.03657575696706772, + 0.02841297909617424, + 0.03764118254184723, + 0.035544686019420624, + 0.02791709639132023, + 0.03702591359615326, + 0.012286094017326832, + 0.014886464923620224, + 0.01287104468792677, + 0.00820230133831501, + 0.010643282905220985, + 0.011898884549736977, + 0.00843066070228815, + 0.00992713961750269, + 0.007829863578081131, + 0.00542611675336957, + 0.004192947410047054, + 0.004926895722746849, + 0.003726700786501169, + 0.006602414883673191, + 0.004568226635456085, + 0.00787555892020464, + 0.008240578696131706, + 0.00666720001026988, + 0.005948259960860014, + 0.008798133581876755, + 0.004868084564805031, + 0.009080782532691956, + 0.006757738534361124, + 0.011075296439230442, + 0.030356390401721, + 0.00978973601013422, + 0.02159959450364113, + 0.01545997429639101, + 0.012034743092954159, + 0.038392819464206696, + 0.014241117052733898, + 0.055498868227005005, + 0.007024503778666258, + 0.01003311388194561, + 0.009306063875555992, + 0.005530789494514465, + 0.007180172950029373, + 0.007759875152260065, + 0.007409297861158848, + 0.008579366840422153, + 0.0068346429616212845, + 0.007055165711790323, + 0.006652495823800564, + 0.005050355568528175, + 0.0076029179617762566, + 0.012120125815272331, + 0.00926888082176447, + 0.007922863587737083, + 0.020748160779476166, + 0.003648237092420459, + 0.04957406967878342, + 0.00426620664075017, + 0.006261858157813549, + 0.03507756441831589, + 0.004322349093854427, + 0.010151169262826443, + 0.005229151342064142, + 0.0057119582779705524, + 0.004725888837128878, + 0.004068517126142979, + 0.005732642952352762, + 0.0061167278327047825, + 0.0038154502399265766, + 0.0041931928135454655, + 0.0039101229049265385, + 0.002885847818106413, + 0.005105431191623211, + 0.0036699725314974785, + 0.0036541956942528486, + 0.009641684591770172, + 0.0019023427739739418, + 0.005710992030799389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.0031914988067001104 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "smooth tracking shot, starting from a slightly upward angle, gracefully circles a white-haired young woman in a kimono embellished with ice crystal patterns. She rises steadily and deliberately from a surface of water as deep and perfectly calm as obsidian. A cool overhead light, almost moonlight, shines down vertically, precisely outlining every detail of her fair skin and the long, silvery hair that clings wetly to her cheeks and neck. Her eyes are closed, her face serene and ethereal, with glistening water droplets dangling from her eyelashes. She wears a dark blue silk kimono, the soaked fabric clinging to her slender figure. The sleeves and bodice are meticulously embroidered with intricate, shimmering ice crystal patterns in silver and light blue thread, which reflect a subtle glow in the overhead light. As her upper body continues to rise, the water flows silently down her skin and the folds of her kimono, and ray tracing technology is used to its extreme at this moment, capturing countless tiny water droplets being lifted from the water. They form a layer of shimmering light mist around her body. Each independent, airborne water droplet is like a miniature prism, refracting and reflecting the single beam of overhead light, emitting a delicate and changing rainbow-like brilliance. Against the backdrop of endless darkness, the entire picture presents a mysterious atmosphere intertwined with extreme realism and an otherworldly and tranquil aesthetic.", + "seed": 791, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_3/index_7692_seed_791_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03434230447746813, + "vlm_logprob": -0.03434230447746813, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.03434230447746813, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04785303398966789 + ], + "opposite_logprob_tokens": [ + -3.2978529930114746 + ], + "logprob_reward": -0.04785303398966789, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04785303398966789, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15715567767620087, + 0.06063735857605934, + 0.16462796926498413, + 0.11357921361923218, + 0.11903781443834305, + 0.08862456679344177, + 0.09752088040113449, + 0.09596755355596542, + 0.05599893629550934, + 0.04685003310441971 + ], + "credit_map": [ + 0.01803283579647541, + 0.01285118330270052, + 0.02323441207408905, + 0.020023059099912643, + 0.012667435221374035, + 0.02035416103899479, + 0.018841391429305077, + 0.031151192262768745, + 0.009991982020437717, + 0.007758487947285175, + 0.009252532385289669, + 0.0059170350432395935, + 0.007853885181248188, + 0.006698225159198046, + 0.005518126301467419, + 0.007647082209587097, + 0.016281604766845703, + 0.017359234392642975, + 0.04356495290994644, + 0.01388110313564539, + 0.016296453773975372, + 0.011653922498226166, + 0.01293467078357935, + 0.03265603259205818, + 0.014672771096229553, + 0.010901353321969509, + 0.015994146466255188, + 0.028191685676574707, + 0.011758209206163883, + 0.00956704281270504, + 0.013993802480399609, + 0.00850020069628954, + 0.011342911049723625, + 0.014738338068127632, + 0.01606646366417408, + 0.014754507690668106, + 0.02139076218008995, + 0.013268415816128254, + 0.016011260449886322, + 0.011465159244835377, + 0.01483234390616417, + 0.012584928423166275, + 0.01049429178237915, + 0.008061042055487633, + 0.014315460808575153, + 0.010210352018475533, + 0.010428053326904774, + 0.007698097266256809, + 0.008531523868441582, + 0.016177751123905182, + 0.014063278213143349, + 0.013747641816735268, + 0.007845781743526459, + 0.012285261414945126, + 0.013241792097687721, + 0.011627855710685253, + 0.009461346082389355, + 0.010019810870289803, + 0.011646619066596031, + 0.009926618076860905, + 0.011347933672368526, + 0.013776839710772038, + 0.013764994218945503, + 0.016023386269807816, + 0.008999450132250786, + 0.009457824751734734, + 0.0054970430210232735, + 0.00412718765437603, + 0.006200161762535572, + 0.010190248489379883, + 0.005209171213209629, + 0.006317847408354282, + 0.007170549593865871, + 0.008125943131744862, + 0.005023610778152943, + 0.0054583000019192696, + 0.0038855618331581354, + 0.007874376140534878, + 0.003773848759010434, + 0.005537842866033316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.04785303398966789 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308804258704185 + ], + "opposite_logprob_tokens": [ + -3.773308753967285 + ], + "logprob_reward": -0.023308804258704185, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308804258704185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08575456589460373, + 0.261365681886673, + 0.032631609588861465, + 0.07854703068733215, + 0.06099306046962738, + 0.29704973101615906, + 0.07254742085933685, + 0.058176759630441666, + 0.02009189873933792, + 0.03284219652414322 + ], + "credit_map": [ + 0.0072538419626653194, + 0.00550253689289093, + 0.011211252771317959, + 0.008038239553570747, + 0.008711906149983406, + 0.01404369343072176, + 0.01535274088382721, + 0.01564035564661026, + 0.07738685607910156, + 0.012295487336814404, + 0.013156789354979992, + 0.007798843551427126, + 0.014257160946726799, + 0.01879889890551567, + 0.024691546335816383, + 0.09298012405633926, + 0.0035529229789972305, + 0.002636043122038245, + 0.00876703392714262, + 0.003632492385804653, + 0.0029617983382195234, + 0.0023613397497683764, + 0.0027781305834650993, + 0.005941847804933786, + 0.009772912599146366, + 0.007864617742598057, + 0.010197227820754051, + 0.01992114447057247, + 0.00792816560715437, + 0.006277799606323242, + 0.009145691059529781, + 0.007439473643898964, + 0.00836604181677103, + 0.01133856549859047, + 0.009297686628997326, + 0.0043792081996798515, + 0.009164728224277496, + 0.008464471437036991, + 0.005299626849591732, + 0.004682730883359909, + 0.023320309817790985, + 0.02160753682255745, + 0.03480742126703262, + 0.03722082078456879, + 0.03746572881937027, + 0.03283175081014633, + 0.03914479911327362, + 0.07065136730670929, + 0.006091252434998751, + 0.006106922402977943, + 0.01212708093225956, + 0.008374286815524101, + 0.006794694345444441, + 0.009994220919907093, + 0.015408198349177837, + 0.0076507641933858395, + 0.00571876997128129, + 0.010044385679066181, + 0.007714931853115559, + 0.00567856477573514, + 0.00635944027453661, + 0.0076311565935611725, + 0.005884766113013029, + 0.009144745767116547, + 0.003192460397258401, + 0.002010376425459981, + 0.0015297641512006521, + 0.001943251583725214, + 0.0023229371290653944, + 0.004637009929865599, + 0.0017684715567156672, + 0.0026876286137849092, + 0.005087332800030708, + 0.0038532332982867956, + 0.0023061924148350954, + 0.003110331017524004, + 0.002773463958874345, + 0.007630994543433189, + 0.0037335120141506195, + 0.004347138572484255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.023308804258704185 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890466570854187 + ], + "opposite_logprob_tokens": [ + -2.4640467166900635 + ], + "logprob_reward": -0.0890466570854187, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890466570854187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15393570065498352, + 0.27778974175453186, + 0.038886360824108124, + 0.12373903393745422, + 0.0858636423945427, + 0.09363792836666107, + 0.039475809782743454, + 0.10980305075645447, + 0.037511833012104034, + 0.03935694694519043 + ], + "credit_map": [ + 0.014946206472814083, + 0.012834389694035053, + 0.018813712522387505, + 0.015575510449707508, + 0.026346106082201004, + 0.016699114814400673, + 0.027934439480304718, + 0.02078619971871376, + 0.06953556090593338, + 0.02360568754374981, + 0.021924173459410667, + 0.01566288247704506, + 0.01865537092089653, + 0.031076639890670776, + 0.019699012860655785, + 0.07763040065765381, + 0.00442514568567276, + 0.003349453443661332, + 0.007715786807239056, + 0.003861956764012575, + 0.0037959387991577387, + 0.003792887320742011, + 0.002612221287563443, + 0.009332969784736633, + 0.015398722141981125, + 0.011296970769762993, + 0.015293160453438759, + 0.02576286345720291, + 0.015276182442903519, + 0.012061930261552334, + 0.01462564617395401, + 0.014023547992110252, + 0.008421066217124462, + 0.010988665744662285, + 0.009371347725391388, + 0.017211364582180977, + 0.009790435433387756, + 0.008213059045374393, + 0.013442616909742355, + 0.00842507928609848, + 0.009171916171908379, + 0.006639184430241585, + 0.011185282841324806, + 0.011662150733172894, + 0.012017368339002132, + 0.010442235507071018, + 0.011747653596103191, + 0.02077212929725647, + 0.0029313168488442898, + 0.009666243568062782, + 0.0043770014308393, + 0.004312773235142231, + 0.00327284331433475, + 0.005118703935295343, + 0.006103479769080877, + 0.0036934446543455124, + 0.010324571281671524, + 0.01016998291015625, + 0.009918794967234135, + 0.008238420821726322, + 0.015613962896168232, + 0.017353547737002373, + 0.010559335350990295, + 0.027624424546957016, + 0.006484877783805132, + 0.004576332867145538, + 0.002602177672088146, + 0.0029970991890877485, + 0.0032527337316423655, + 0.010695516131818295, + 0.0022484231740236282, + 0.004654668737202883, + 0.007588046137243509, + 0.0028007011860609055, + 0.0021718526259064674, + 0.0021894345991313457, + 0.003036735812202096, + 0.010735948570072651, + 0.0035210493952035904, + 0.0073131779208779335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.0890466570854187 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789789095520973 + ], + "opposite_logprob_tokens": [ + -4.634789943695068 + ], + "logprob_reward": -0.009789789095520973, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789789095520973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1682676523923874, + 0.12278608232736588, + 0.05637817457318306, + 0.14396069943904877, + 0.10768291354179382, + 0.042677607387304306, + 0.056126341223716736, + 0.20206907391548157, + 0.04950453341007233, + 0.050546955317258835 + ], + "credit_map": [ + 0.020400097593665123, + 0.017271218821406364, + 0.017735160887241364, + 0.018213925883173943, + 0.024205146357417107, + 0.023461682721972466, + 0.025134842842817307, + 0.021845579147338867, + 0.016986748203635216, + 0.017958445474505424, + 0.017249057069420815, + 0.01218690536916256, + 0.012632925994694233, + 0.0182708241045475, + 0.015149402432143688, + 0.012351770885288715, + 0.005947601515799761, + 0.007842258550226688, + 0.011099635623395443, + 0.0046922434121370316, + 0.006069354712963104, + 0.0068015712313354015, + 0.0037488543894141912, + 0.01017665769904852, + 0.02112836390733719, + 0.013493968173861504, + 0.018111713230609894, + 0.028045056387782097, + 0.015522349625825882, + 0.013701608404517174, + 0.01671290025115013, + 0.017244741320610046, + 0.009372983127832413, + 0.014741982333362103, + 0.012860688380897045, + 0.018080245703458786, + 0.014111938886344433, + 0.010752425529062748, + 0.01862121932208538, + 0.009141435846686363, + 0.003914671950042248, + 0.004014546051621437, + 0.005398083943873644, + 0.006679699756205082, + 0.003713544923812151, + 0.00472228741273284, + 0.00694061815738678, + 0.007294155191630125, + 0.004390326328575611, + 0.010631000623106956, + 0.00532521540299058, + 0.007069109007716179, + 0.004807128105312586, + 0.007532571908086538, + 0.00902512576431036, + 0.007345862686634064, + 0.02030276507139206, + 0.01960316300392151, + 0.01932237669825554, + 0.015860483050346375, + 0.033202871680259705, + 0.029043134301900864, + 0.025482086464762688, + 0.03925219178199768, + 0.00884284358471632, + 0.005938702262938023, + 0.003742737229913473, + 0.00333185913041234, + 0.00448747118934989, + 0.01323742140084505, + 0.003909028135240078, + 0.0060144709423184395, + 0.006848693825304508, + 0.007601965218782425, + 0.004726535174995661, + 0.006471650674939156, + 0.004463949706405401, + 0.010966740548610687, + 0.0031629716977477074, + 0.006304447539150715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.009789789095520973 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a smooth tracking shot that circles a white-haired young woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman wearing a dark blue silk kimono with silver and light blue ice crystal embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman rising steadily from a surface of water that is deep, calm, and black like obsidian?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a cool overhead light resembling moonlight that outlines her skin and wet hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are countless tiny water droplets visible flowing down her skin and kimono, refracting light to create a shimmering mist?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017132379580289125 + ], + "opposite_logprob_tokens": [ + -6.376713275909424 + ], + "logprob_reward": -0.0017132379580289125, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017132379580289125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09400417655706406, + 0.44075876474380493, + 0.042455341666936874, + 0.08345326036214828, + 0.051464758813381195, + 0.11571897566318512, + 0.02844974584877491, + 0.04914821684360504, + 0.03166333585977554, + 0.06288344413042068 + ], + "credit_map": [ + 0.01127214077860117, + 0.009607937186956406, + 0.011794066056609154, + 0.014083709567785263, + 0.012368736788630486, + 0.012434128671884537, + 0.012101233005523682, + 0.010342229157686234, + 0.13187706470489502, + 0.0175782423466444, + 0.022404056042432785, + 0.011393209919333458, + 0.02282715030014515, + 0.03273564577102661, + 0.03962862864136696, + 0.162314772605896, + 0.004429228138178587, + 0.0047851973213255405, + 0.010017198510468006, + 0.0038524875417351723, + 0.004015919286757708, + 0.003780054161325097, + 0.0028206591960042715, + 0.008754598908126354, + 0.010874691419303417, + 0.007291433867067099, + 0.009844889864325523, + 0.012830850668251514, + 0.011544173583388329, + 0.008993311785161495, + 0.011659392155706882, + 0.010414517484605312, + 0.004287124611437321, + 0.00806481298059225, + 0.006737698335200548, + 0.008018418215215206, + 0.00749237323179841, + 0.0055987839587032795, + 0.007167057599872351, + 0.004098494071513414, + 0.012344781309366226, + 0.009102641604840755, + 0.013996468856930733, + 0.013461262919008732, + 0.01560206525027752, + 0.014004183001816273, + 0.013678587041795254, + 0.02352898195385933, + 0.002778306370601058, + 0.006029897835105658, + 0.002624806249514222, + 0.003225657157599926, + 0.0024475734680891037, + 0.0040007540956139565, + 0.0034467927180230618, + 0.0038959586527198553, + 0.005515325348824263, + 0.008773974142968655, + 0.0037768124602735043, + 0.0039018909446895123, + 0.007492052856832743, + 0.006993139628320932, + 0.0044648656621575356, + 0.00823015533387661, + 0.004861034452915192, + 0.005717339459806681, + 0.0036630930844694376, + 0.003790757153183222, + 0.0028422714676707983, + 0.003891812404617667, + 0.0034056573640555143, + 0.0034913700073957443, + 0.009520146064460278, + 0.007384927477687597, + 0.004873719532042742, + 0.005484799388796091, + 0.0059376247227191925, + 0.016773374751210213, + 0.0038931886665523052, + 0.009015662595629692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.0017132379580289125 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_4/index_tensor([7532, 7532])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_4/index_tensor([7532, 7532])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4179e8edab1e60c38584126a983db33e994bab80 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_4/index_tensor([7532, 7532])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a high-angle, ultra-wide shot captured in the grandiose style of Andreas Gursky, the camera tracks smoothly to the right, following a turbulent current resembling a subterranean river. This murky, silt-laden torrent, churning with white foam and forming powerful whirlpools in the gloom, relentlessly erodes the ancient maze-like foundations of the Angkor Wat temple complex. Rows of majestic sandstone pillars, covered in moss and water stains, with intricate relief carvings of ancient gods and nymphs (Apsaras), appear in the dim light. The torrent crashes violently against the pillars' bases, stirring up high waves. The water level rises rapidly, steadily and deliberately swallowing up the exquisite carvings one by one, submerging them beneath the murky yellow surface. A few rays of light from invisible cracks above cast shafts of light in the dusty air, illuminating the restless surface of the water and the dull texture of the wet stone. The entire picture presents a cold tone composed of dark brown, moss green and dark gray, creating a grand, tragic and unstoppable sense of historical destiny, perfectly embodying the eternal erosion of natural forces on the remains of human civilization.", + "seed": 2597, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_4/index_7532_seed_2597_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04728506188839674, + "vlm_logprob": -0.04728506188839674, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.04728506188839674, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13235926628112793 + ], + "opposite_logprob_tokens": [ + -2.632359266281128 + ], + "logprob_reward": -0.13235926628112793, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.13235926628112793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19648049771785736, + 0.05956055968999863, + 0.05497506260871887, + 0.10506436973810196, + 0.10545577108860016, + 0.05330117046833038, + 0.054000623524188995, + 0.0682644173502922, + 0.057544007897377014, + 0.24535344541072845 + ], + "credit_map": [ + 0.015943925827741623, + 0.01633494906127453, + 0.02558540739119053, + 0.018230969086289406, + 0.016530191525816917, + 0.014665411785244942, + 0.041027575731277466, + 0.04816209152340889, + 0.00992356613278389, + 0.006331768352538347, + 0.007232330739498138, + 0.005496775731444359, + 0.006112042348831892, + 0.007889652624726295, + 0.009515384212136269, + 0.007059046998620033, + 0.0036358910147100687, + 0.005191328469663858, + 0.006178478244692087, + 0.00901830568909645, + 0.004979652818292379, + 0.007893471047282219, + 0.007878083735704422, + 0.01019985694438219, + 0.008833428844809532, + 0.011767941527068615, + 0.009477431885898113, + 0.009530064649879932, + 0.01836380548775196, + 0.01819075085222721, + 0.02022840641438961, + 0.00867254938930273, + 0.009297531098127365, + 0.011991383507847786, + 0.010251334868371487, + 0.011113380081951618, + 0.017469506710767746, + 0.016213180497288704, + 0.01806657575070858, + 0.011052887886762619, + 0.006786600686609745, + 0.009211232885718346, + 0.006841700058430433, + 0.0063381502404809, + 0.005508042871952057, + 0.005137948784977198, + 0.007308883126825094, + 0.00616862066090107, + 0.00700252503156662, + 0.005469124298542738, + 0.004986596759408712, + 0.008691363036632538, + 0.007828596979379654, + 0.007628574036061764, + 0.005383672192692757, + 0.007010175380855799, + 0.007589265238493681, + 0.011213276535272598, + 0.009485060349106789, + 0.009156808257102966, + 0.0098201222717762, + 0.005814556032419205, + 0.007849583402276039, + 0.007335752248764038, + 0.008059969171881676, + 0.01215931586921215, + 0.005768958944827318, + 0.004701902624219656, + 0.009085729718208313, + 0.005746232811361551, + 0.006102588027715683, + 0.005919313989579678, + 0.03902122750878334, + 0.031377989798784256, + 0.02091570384800434, + 0.025667335838079453, + 0.01781819760799408, + 0.02433326467871666, + 0.040034450590610504, + 0.04618528112769127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.13235926628112793 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04868759214878082 + ], + "opposite_logprob_tokens": [ + -3.048687696456909 + ], + "logprob_reward": -0.04868759214878082, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04868759214878082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47688233852386475, + 0.09173990041017532, + 0.05908984690904617, + 0.06234452873468399, + 0.07002997398376465, + 0.06294172257184982, + 0.025561107322573662, + 0.05686553940176964, + 0.04180380329489708, + 0.05274123325943947 + ], + "credit_map": [ + 0.04222273826599121, + 0.01927371509373188, + 0.028553007170557976, + 0.013075275346636772, + 0.01935136318206787, + 0.016322992742061615, + 0.17084112763404846, + 0.1672421246767044, + 0.011272996664047241, + 0.016375357285141945, + 0.012246849946677685, + 0.01081051304936409, + 0.009017862379550934, + 0.00894581526517868, + 0.012986552901566029, + 0.010083948262035847, + 0.006093167699873447, + 0.0066715991124510765, + 0.005427608732134104, + 0.013061082921922207, + 0.005347622558474541, + 0.007276177406311035, + 0.009441398084163666, + 0.0057711852714419365, + 0.005466261878609657, + 0.010128895752131939, + 0.005587148480117321, + 0.006082191132009029, + 0.008737586438655853, + 0.009966554120182991, + 0.009911676868796349, + 0.006464215461164713, + 0.006578696891665459, + 0.008211317472159863, + 0.007512344978749752, + 0.007439333014190197, + 0.01321116741746664, + 0.009958328679203987, + 0.009503098204731941, + 0.007615691050887108, + 0.004532482009381056, + 0.004901235923171043, + 0.005721230059862137, + 0.005274055525660515, + 0.016266396269202232, + 0.008088020607829094, + 0.01407904177904129, + 0.0040792603977024555, + 0.002923233201727271, + 0.004424327984452248, + 0.0021494782995432615, + 0.00350299384444952, + 0.0038300123997032642, + 0.002488254802301526, + 0.0033292153384536505, + 0.0029135907534509897, + 0.006967449095100164, + 0.007272965274751186, + 0.00909495074301958, + 0.007516886107623577, + 0.00849782396107912, + 0.0035435925237834454, + 0.007131279446184635, + 0.006840586196631193, + 0.005869947373867035, + 0.006595304235816002, + 0.004309799987822771, + 0.0032412034925073385, + 0.005908208899199963, + 0.007851843722164631, + 0.004715241026133299, + 0.0033122538588941097, + 0.011787141673266888, + 0.007018164731562138, + 0.005824561230838299, + 0.005205527879297733, + 0.004488032311201096, + 0.006199135910719633, + 0.005793936084955931, + 0.00642473716288805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04868759214878082 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014208580367267132 + ], + "opposite_logprob_tokens": [ + -4.264208793640137 + ], + "logprob_reward": -0.014208580367267132, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014208580367267132, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45935216546058655, + 0.09487513452768326, + 0.06562826782464981, + 0.10525419563055038, + 0.1230410486459732, + 0.02749713324010372, + 0.026420319452881813, + 0.03827228397130966, + 0.0170463677495718, + 0.042613159865140915 + ], + "credit_map": [ + 0.041472651064395905, + 0.0209813192486763, + 0.030489061027765274, + 0.0187322236597538, + 0.020481940358877182, + 0.01742599532008171, + 0.1615474969148636, + 0.1482214480638504, + 0.010534384287893772, + 0.015510615892708302, + 0.006246987264603376, + 0.00913431216031313, + 0.020257553085684776, + 0.009084241464734077, + 0.0122548071667552, + 0.011852232739329338, + 0.007607514504343271, + 0.007332174107432365, + 0.007656019646674395, + 0.007749606389552355, + 0.0075070820748806, + 0.010501067154109478, + 0.008977295830845833, + 0.008297507651150227, + 0.007174848113209009, + 0.013924708589911461, + 0.009297364391386509, + 0.010937903076410294, + 0.012789785861968994, + 0.017560718581080437, + 0.02140967734158039, + 0.01215919479727745, + 0.011163097806274891, + 0.012868004851043224, + 0.011391658335924149, + 0.012359189800918102, + 0.02268189750611782, + 0.01886214315891266, + 0.020009580999612808, + 0.013705474324524403, + 0.0026414149906486273, + 0.003101048292592168, + 0.002220212947577238, + 0.0027755701448768377, + 0.006793406791985035, + 0.0025776817928999662, + 0.005089617799967527, + 0.002298180479556322, + 0.0032485434785485268, + 0.0037054368294775486, + 0.002211350714787841, + 0.003265783190727234, + 0.004218237474560738, + 0.0038561869878321886, + 0.0025409949012100697, + 0.0033737856429070234, + 0.0031517355237156153, + 0.00515521178022027, + 0.004666426219046116, + 0.004172109998762608, + 0.0058524878695607185, + 0.005489785689860582, + 0.0031864289194345474, + 0.006598098203539848, + 0.0015069415094330907, + 0.002783501287922263, + 0.002409448614344001, + 0.0014023983385413885, + 0.002216736786067486, + 0.0025939615443348885, + 0.002259763889014721, + 0.0018736169440671802, + 0.006866360083222389, + 0.00519271194934845, + 0.005206013098359108, + 0.0033926856704056263, + 0.0041533540934324265, + 0.007654368411749601, + 0.00562251266092062, + 0.004525153432041407 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.014208580367267132 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588496699929237 + ], + "opposite_logprob_tokens": [ + -3.8955883979797363 + ], + "logprob_reward": -0.020588496699929237, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588496699929237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7099217772483826, + 0.04190583899617195, + 0.02831932157278061, + 0.0457616001367569, + 0.05880328267812729, + 0.03692879527807236, + 0.023755187168717384, + 0.020052596926689148, + 0.011103841476142406, + 0.023447683081030846 + ], + "credit_map": [ + 0.06225420534610748, + 0.02822040393948555, + 0.04507331922650337, + 0.017103491351008415, + 0.029081525281071663, + 0.021801069378852844, + 0.2571440637111664, + 0.24924370646476746, + 0.004556017927825451, + 0.005156093742698431, + 0.004605385009199381, + 0.006753417197614908, + 0.006682199891656637, + 0.004145502578467131, + 0.005063367076218128, + 0.004943853709846735, + 0.0028008364606648684, + 0.0026258258149027824, + 0.0032448654528707266, + 0.004504088778048754, + 0.0035220033023506403, + 0.003840524470433593, + 0.0047618369571864605, + 0.0030193382408469915, + 0.003717204788699746, + 0.005006360821425915, + 0.005118322558701038, + 0.005065490957349539, + 0.005789151415228844, + 0.008173516020178795, + 0.007687067613005638, + 0.005204485729336739, + 0.005741396453231573, + 0.0064596510492265224, + 0.004799390211701393, + 0.005219067446887493, + 0.014326431788504124, + 0.00987744890153408, + 0.006746286991983652, + 0.0056336126290261745, + 0.004100841004401445, + 0.0044675301760435104, + 0.004132560919970274, + 0.004153270274400711, + 0.006005138158798218, + 0.003243690589442849, + 0.0068069216795265675, + 0.004018841311335564, + 0.0030748704448342323, + 0.0038674052339047194, + 0.0023836472537368536, + 0.0024359414819628, + 0.003656391752883792, + 0.0026271396782249212, + 0.0021153746638447046, + 0.003594414796680212, + 0.0018360966350883245, + 0.0023875022307038307, + 0.0029960807878524065, + 0.0033979041036218405, + 0.0028378472197800875, + 0.0022047883830964565, + 0.0025732084177434444, + 0.0018191671697422862, + 0.001137120765633881, + 0.0016530813882127404, + 0.001455468824133277, + 0.0010963449021801353, + 0.0013291140785440803, + 0.0013704224256798625, + 0.0013071232242509723, + 0.0017551661003381014, + 0.0035994609352201223, + 0.003011951921507716, + 0.0016885151853784919, + 0.0028931007254868746, + 0.0024459147825837135, + 0.002988476539030671, + 0.0031986923422664404, + 0.0036215721629559994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.020588496699929237 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581373944878578 + ], + "opposite_logprob_tokens": [ + -3.8955814838409424 + ], + "logprob_reward": -0.020581373944878578, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581373944878578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18544290959835052, + 0.04045892506837845, + 0.0270062368363142, + 0.26832643151283264, + 0.10639642179012299, + 0.2463352084159851, + 0.01592113822698593, + 0.026326501742005348, + 0.031090419739484787, + 0.05269573628902435 + ], + "credit_map": [ + 0.015413830988109112, + 0.009266779758036137, + 0.013447973877191544, + 0.0065227411687374115, + 0.010236415080726147, + 0.011504942551255226, + 0.06096934154629707, + 0.05808087810873985, + 0.008611677214503288, + 0.004704949911683798, + 0.004431499633938074, + 0.00365038076415658, + 0.0035367736127227545, + 0.0051583764143288136, + 0.004992937669157982, + 0.005372330546379089, + 0.003269474022090435, + 0.0029892276506870985, + 0.0029164906591176987, + 0.004152656998485327, + 0.0040334537625312805, + 0.00365613610483706, + 0.00344319106079638, + 0.002545607043430209, + 0.02252950333058834, + 0.02798801101744175, + 0.026686396449804306, + 0.030203549191355705, + 0.036714550107717514, + 0.044525984674692154, + 0.0436822772026062, + 0.03599615395069122, + 0.008653262630105019, + 0.013234944082796574, + 0.008454559370875359, + 0.009494801051914692, + 0.02061852253973484, + 0.01764935627579689, + 0.016945218667387962, + 0.011345762759447098, + 0.019126592203974724, + 0.015341848134994507, + 0.021675024181604385, + 0.02100091241300106, + 0.05218671262264252, + 0.0209022406488657, + 0.07676076889038086, + 0.019341127946972847, + 0.00229915464296937, + 0.0023583939764648676, + 0.0014904856216162443, + 0.0019392739050090313, + 0.0021306073758751154, + 0.001672538579441607, + 0.0016838001320138574, + 0.0023468828294426203, + 0.002914383076131344, + 0.006268139462918043, + 0.0036213109269738197, + 0.002794775180518627, + 0.004575259983539581, + 0.001781103084795177, + 0.002129348460584879, + 0.0022421821486204863, + 0.002270073164254427, + 0.0026103132404386997, + 0.003032751614227891, + 0.0033347313292324543, + 0.0036624528001993895, + 0.006649717688560486, + 0.005964788142591715, + 0.0035655959509313107, + 0.0077606067061424255, + 0.0066680582240223885, + 0.00407029315829277, + 0.0041579208336770535, + 0.004185836296528578, + 0.005555200390517712, + 0.00932795088738203, + 0.010969871655106544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.020581373944878578 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-angle, ultra-wide shot captured in the grandiose style of Andreas Gursky, the camera tracks smoothly to the right, following a turbulent current resembling a subterranean river. This murky, silt-laden torrent, churning with white foam and forming powerful whirlpools in the gloom, relentlessly erodes the ancient maze-like foundations of the Angkor Wat temple complex. Rows of majestic sandstone pillars, covered in moss and water stains, with intricate relief carvings of ancient gods and nymphs (Apsaras), appear in the dim light. The torrent crashes violently against the pillars' bases, stirring up high waves. The water level rises rapidly, steadily and deliberately swallowing up the exquisite carvings one by one, submerging them beneath the murky yellow surface. A few rays of light from invisible cracks above cast shafts of light in the dusty air, illuminating the restless surface of the water and the dull texture of the wet stone. The entire picture presents a cold tone composed of dark brown, moss green and dark gray, creating a grand, tragic and unstoppable sense of historical destiny, perfectly embodying the eternal erosion of natural forces on the remains of human civilization.", + "seed": 2598, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_4/index_7532_seed_2598_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.055331938341259955, + "vlm_logprob": -0.055331938341259955, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.055331938341259955, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17952391505241394 + ], + "opposite_logprob_tokens": [ + -2.1795239448547363 + ], + "logprob_reward": -0.17952391505241394, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.17952391505241394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2872028946876526, + 0.109837606549263, + 0.14151433110237122, + 0.0534181110560894, + 0.1454937607049942, + 0.03260437026619911, + 0.055589620023965836, + 0.10264033079147339, + 0.04272725433111191, + 0.028971819207072258 + ], + "credit_map": [ + 0.03481488674879074, + 0.03983280435204506, + 0.032832443714141846, + 0.02658804878592491, + 0.04857643321156502, + 0.03453020751476288, + 0.04750100150704384, + 0.022527024149894714, + 0.01762920245528221, + 0.014312083832919598, + 0.011207628063857555, + 0.00969625823199749, + 0.017942097038030624, + 0.015277650207281113, + 0.013735962100327015, + 0.01003669761121273, + 0.01562623865902424, + 0.023149002343416214, + 0.013034036383032799, + 0.027624204754829407, + 0.02151239663362503, + 0.02232692949473858, + 0.012074199505150318, + 0.006167293526232243, + 0.009154838509857655, + 0.006250013131648302, + 0.006111037917435169, + 0.005609809421002865, + 0.0071092876605689526, + 0.007913987152278423, + 0.005706840194761753, + 0.005562284961342812, + 0.024022962898015976, + 0.017209971323609352, + 0.02810707315802574, + 0.01316503994166851, + 0.02479458600282669, + 0.018056996166706085, + 0.008495409041643143, + 0.011641684919595718, + 0.0035890128929167986, + 0.005572936963289976, + 0.002591744065284729, + 0.0027836235240101814, + 0.005507635418325663, + 0.004989508539438248, + 0.004038423765450716, + 0.003531479276716709, + 0.006761961616575718, + 0.009530854411423206, + 0.0035452842712402344, + 0.0037095278967171907, + 0.011803572997450829, + 0.00739369448274374, + 0.006840990390628576, + 0.006003721617162228, + 0.011941598728299141, + 0.012698840349912643, + 0.013224129565060139, + 0.007886006496846676, + 0.012641494162380695, + 0.015017971396446228, + 0.01410820335149765, + 0.015122071839869022, + 0.005022536963224411, + 0.006222825031727552, + 0.003985322080552578, + 0.0037855824921280146, + 0.007489591836929321, + 0.0068976785987615585, + 0.0031560310162603855, + 0.0061676776967942715, + 0.0032215924002230167, + 0.0032060490921139717, + 0.003428249852731824, + 0.002886188682168722, + 0.004139590077102184, + 0.004603188950568438, + 0.0026014551986008883, + 0.004885499831289053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.17952391505241394 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03374166786670685 + ], + "opposite_logprob_tokens": [ + -3.4087417125701904 + ], + "logprob_reward": -0.03374166786670685, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03374166786670685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10501419007778168, + 0.06694190949201584, + 0.07026541233062744, + 0.06530827283859253, + 0.09188568592071533, + 0.08252095431089401, + 0.14060306549072266, + 0.22034882009029388, + 0.12644484639167786, + 0.03066677413880825 + ], + "credit_map": [ + 0.01413062121719122, + 0.014187312684953213, + 0.00919837225228548, + 0.009749097749590874, + 0.01632465235888958, + 0.017429659143090248, + 0.013539507053792477, + 0.010454976931214333, + 0.012895094230771065, + 0.009034689515829086, + 0.005850049667060375, + 0.006824858020991087, + 0.007527047768235207, + 0.010920512489974499, + 0.007309436798095703, + 0.006580232642591, + 0.009558344259858131, + 0.008893758989870548, + 0.007368667051196098, + 0.00963107030838728, + 0.01278185285627842, + 0.01015081163495779, + 0.006578246131539345, + 0.00530266622081399, + 0.008260044269263744, + 0.008498593233525753, + 0.008867907337844372, + 0.009289677254855633, + 0.012038221582770348, + 0.008584128692746162, + 0.0045885322615504265, + 0.005181177984923124, + 0.014853599481284618, + 0.010091726668179035, + 0.01742497645318508, + 0.008167349733412266, + 0.015356037765741348, + 0.010290966369211674, + 0.00661685923114419, + 0.009084172546863556, + 0.011428521014750004, + 0.011787104420363903, + 0.008308149874210358, + 0.007268172688782215, + 0.011122167110443115, + 0.012326488271355629, + 0.008726028725504875, + 0.011554333381354809, + 0.016067305579781532, + 0.024065067991614342, + 0.009344655089080334, + 0.009901481680572033, + 0.02812953293323517, + 0.018381915986537933, + 0.01831885054707527, + 0.01639426127076149, + 0.02180813066661358, + 0.025447160005569458, + 0.021183103322982788, + 0.01829477772116661, + 0.04368268698453903, + 0.03242941200733185, + 0.03332799673080444, + 0.024175554513931274, + 0.02741703949868679, + 0.016674617305397987, + 0.01270691305398941, + 0.009129560552537441, + 0.017070528119802475, + 0.019308671355247498, + 0.010904679074883461, + 0.013232855126261711, + 0.003342971671372652, + 0.003405549330636859, + 0.003336995141580701, + 0.003369324840605259, + 0.005297574680298567, + 0.005174270365387201, + 0.0034761112183332443, + 0.003263978287577629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.03374166786670685 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210578985512257 + ], + "opposite_logprob_tokens": [ + -4.2642107009887695 + ], + "logprob_reward": -0.014210578985512257, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210578985512257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13732656836509705, + 0.15604597330093384, + 0.24093155562877655, + 0.08335335552692413, + 0.06529417634010315, + 0.026768933981657028, + 0.09472472965717316, + 0.051337163895368576, + 0.11705255508422852, + 0.02716505527496338 + ], + "credit_map": [ + 0.01753593608736992, + 0.022956136614084244, + 0.013312987051904202, + 0.012582679279148579, + 0.023822970688343048, + 0.018268754705786705, + 0.017168287187814713, + 0.011678803712129593, + 0.02306295931339264, + 0.027046475559473038, + 0.011202997528016567, + 0.022918764501810074, + 0.015882033854722977, + 0.031558889895677567, + 0.014827811159193516, + 0.009546029381453991, + 0.028248362243175507, + 0.06078021973371506, + 0.019690509885549545, + 0.03996862471103668, + 0.02297566831111908, + 0.040081482380628586, + 0.016389571130275726, + 0.012797104194760323, + 0.012731436640024185, + 0.012127969413995743, + 0.010100780986249447, + 0.009563776664435863, + 0.014905668795108795, + 0.012700998224318027, + 0.0056886449456214905, + 0.005534075200557709, + 0.009503386914730072, + 0.008392052724957466, + 0.010667452588677406, + 0.004722733981907368, + 0.009129688143730164, + 0.01001111138612032, + 0.005401983857154846, + 0.007465760223567486, + 0.003573962952941656, + 0.005209574941545725, + 0.0023853855673223734, + 0.001999597530812025, + 0.004513525404036045, + 0.003490173490718007, + 0.0028133115265518427, + 0.0027834007050842047, + 0.013750743120908737, + 0.020708704367280006, + 0.009364563040435314, + 0.00793053861707449, + 0.010550671257078648, + 0.013291055336594582, + 0.009511236101388931, + 0.009617202915251255, + 0.00581909017637372, + 0.007462788838893175, + 0.005427204072475433, + 0.004320046864449978, + 0.006511345971375704, + 0.00710689602419734, + 0.008243775926530361, + 0.006446016486734152, + 0.02479054592549801, + 0.013351627625524998, + 0.012300780043005943, + 0.008438226766884327, + 0.01604531705379486, + 0.01922382414340973, + 0.008031877689063549, + 0.01487034186720848, + 0.0038400546181946993, + 0.00479029631242156, + 0.002420400967821479, + 0.0028944306541234255, + 0.004769130144268274, + 0.004004168324172497, + 0.0020586918108165264, + 0.0023878796491771936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.014210578985512257 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079457588493824 + ], + "opposite_logprob_tokens": [ + -4.51107931137085 + ], + "logprob_reward": -0.011079457588493824, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079457588493824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17865967750549316, + 0.18179231882095337, + 0.0801222026348114, + 0.061687931418418884, + 0.09921108931303024, + 0.03483186289668083, + 0.17267277836799622, + 0.09745195508003235, + 0.05938655510544777, + 0.0341835543513298 + ], + "credit_map": [ + 0.023594995960593224, + 0.024935830384492874, + 0.0175593513995409, + 0.015408952720463276, + 0.022395208477973938, + 0.022614838555455208, + 0.030888468027114868, + 0.02126205898821354, + 0.030846886336803436, + 0.021892771124839783, + 0.014873790554702282, + 0.022028418257832527, + 0.02010476030409336, + 0.023213185369968414, + 0.02769395336508751, + 0.02113858051598072, + 0.010448032058775425, + 0.014591935090720654, + 0.007424640003591776, + 0.00988446269184351, + 0.009683373384177685, + 0.010823763906955719, + 0.012606348842382431, + 0.004659653641283512, + 0.008760572411119938, + 0.00792017113417387, + 0.007066670805215836, + 0.007906362414360046, + 0.01043829508125782, + 0.00874931551516056, + 0.004875052720308304, + 0.0059714955277740955, + 0.01039439719170332, + 0.013177747838199139, + 0.012726719491183758, + 0.010923315770924091, + 0.014087448827922344, + 0.01599441096186638, + 0.009287793189287186, + 0.012619268149137497, + 0.004671388305723667, + 0.005175507161766291, + 0.0030182290356606245, + 0.0026970331091433764, + 0.005258241202682257, + 0.005743002984672785, + 0.003902619006112218, + 0.004365840926766396, + 0.025571556761860847, + 0.04604405164718628, + 0.012782767415046692, + 0.014327908866107464, + 0.022373488172888756, + 0.027175473049283028, + 0.013802829198539257, + 0.010594712570309639, + 0.009693779051303864, + 0.014663374051451683, + 0.01159879844635725, + 0.006312074139714241, + 0.017648639157414436, + 0.016556501388549805, + 0.009675509296357632, + 0.011303286999464035, + 0.00830812193453312, + 0.006842535454779863, + 0.005408062133938074, + 0.004729434847831726, + 0.011152072809636593, + 0.009644915349781513, + 0.005807569250464439, + 0.0074938456527888775, + 0.005526527296751738, + 0.004560486413538456, + 0.002890043193474412, + 0.0030022612772881985, + 0.006877986714243889, + 0.005725360009819269, + 0.0029699676670134068, + 0.0026309245731681585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.011079457588493824 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810407221317291 + ], + "opposite_logprob_tokens": [ + -3.2881040573120117 + ], + "logprob_reward": -0.03810407221317291, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810407221317291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1113249659538269, + 0.2406451255083084, + 0.05513099953532219, + 0.032116614282131195, + 0.09496749937534332, + 0.13542792201042175, + 0.059802111238241196, + 0.1993197500705719, + 0.043683748692274094, + 0.02758127637207508 + ], + "credit_map": [ + 0.017400391399860382, + 0.01565757766366005, + 0.01394771970808506, + 0.00944141298532486, + 0.017491286620497704, + 0.010700256563723087, + 0.014866885729134083, + 0.011819436214864254, + 0.03968902677297592, + 0.02506103366613388, + 0.01364943478256464, + 0.01843331754207611, + 0.03832610324025154, + 0.06225639954209328, + 0.01952243037521839, + 0.023707367479801178, + 0.006459078751504421, + 0.008787057362496853, + 0.008924141526222229, + 0.007480389904230833, + 0.005970504600554705, + 0.0075018578208982944, + 0.005095848347991705, + 0.004912120755761862, + 0.005920193158090115, + 0.003741796361282468, + 0.003958918619900942, + 0.0030541333835572004, + 0.006337218917906284, + 0.004577675834298134, + 0.0019833233673125505, + 0.002543356968089938, + 0.013292125426232815, + 0.01644248142838478, + 0.010790995322167873, + 0.007162760943174362, + 0.020055998116731644, + 0.015785256400704384, + 0.006130959838628769, + 0.005306916777044535, + 0.019982026889920235, + 0.02206897921860218, + 0.01199929416179657, + 0.011475346982479095, + 0.021957868710160255, + 0.018338141962885857, + 0.01451865304261446, + 0.015087598003447056, + 0.006594954524189234, + 0.009639169089496136, + 0.004781126510351896, + 0.004868905059993267, + 0.01349577121436596, + 0.008313227444887161, + 0.007093478459864855, + 0.005015483591705561, + 0.022818801924586296, + 0.026845959946513176, + 0.02512483112514019, + 0.01878066547214985, + 0.024827295914292336, + 0.02607869543135166, + 0.029384758323431015, + 0.02545873448252678, + 0.00853995606303215, + 0.005627247039228678, + 0.004302786663174629, + 0.0027141692116856575, + 0.0069930534809827805, + 0.0073182727210223675, + 0.002563866088166833, + 0.005624399520456791, + 0.0029399951454252005, + 0.003932422026991844, + 0.00237344135530293, + 0.0032395529560744762, + 0.005308710969984531, + 0.004055560566484928, + 0.002389725297689438, + 0.0033418680541217327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 5 + ], + "selected_reward": -0.03810407221317291 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-angle, ultra-wide shot captured in the grandiose style of Andreas Gursky, the camera tracks smoothly to the right, following a turbulent current resembling a subterranean river. This murky, silt-laden torrent, churning with white foam and forming powerful whirlpools in the gloom, relentlessly erodes the ancient maze-like foundations of the Angkor Wat temple complex. Rows of majestic sandstone pillars, covered in moss and water stains, with intricate relief carvings of ancient gods and nymphs (Apsaras), appear in the dim light. The torrent crashes violently against the pillars' bases, stirring up high waves. The water level rises rapidly, steadily and deliberately swallowing up the exquisite carvings one by one, submerging them beneath the murky yellow surface. A few rays of light from invisible cracks above cast shafts of light in the dusty air, illuminating the restless surface of the water and the dull texture of the wet stone. The entire picture presents a cold tone composed of dark brown, moss green and dark gray, creating a grand, tragic and unstoppable sense of historical destiny, perfectly embodying the eternal erosion of natural forces on the remains of human civilization.", + "seed": 2599, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_4/index_7532_seed_2599_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0622958755120635, + "vlm_logprob": -0.0622958755120635, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.0622958755120635, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19778387248516083 + ], + "opposite_logprob_tokens": [ + -2.322783946990967 + ], + "logprob_reward": -0.19778387248516083, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.19778387248516083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21685786545276642, + 0.08677874505519867, + 0.036807406693696976, + 0.293753981590271, + 0.032342080026865005, + 0.04774630814790726, + 0.0680253729224205, + 0.10135606676340103, + 0.09033647179603577, + 0.025995653122663498 + ], + "credit_map": [ + 0.021989675238728523, + 0.027103383094072342, + 0.021911734715104103, + 0.028593849390745163, + 0.029103577136993408, + 0.0242199394851923, + 0.040614962577819824, + 0.023320745676755905, + 0.00806357804685831, + 0.01892395131289959, + 0.007299031130969524, + 0.01153432298451662, + 0.013850254938006401, + 0.008321864530444145, + 0.011210381984710693, + 0.007575359661132097, + 0.005612561479210854, + 0.004149730782955885, + 0.004362409468740225, + 0.0036327443085610867, + 0.006201093550771475, + 0.00287441024556756, + 0.006367586087435484, + 0.0036068703047931194, + 0.10470587760210037, + 0.013800229877233505, + 0.025613250210881233, + 0.0067371223121881485, + 0.011209104210138321, + 0.016708238050341606, + 0.012107337825000286, + 0.1028728038072586, + 0.004753651097416878, + 0.003928163554519415, + 0.00406875042244792, + 0.00646256236359477, + 0.002637167926877737, + 0.0025968621484935284, + 0.003590491134673357, + 0.0043044313788414, + 0.0048889839090406895, + 0.004563353024423122, + 0.006053636781871319, + 0.004401569254696369, + 0.010289572179317474, + 0.006214825436472893, + 0.0066450899466872215, + 0.004689277149736881, + 0.012899376451969147, + 0.0036761476658284664, + 0.01379687525331974, + 0.013360865414142609, + 0.005493595730513334, + 0.005129626952111721, + 0.007903166115283966, + 0.005765723530203104, + 0.013024727813899517, + 0.031517576426267624, + 0.01126965694129467, + 0.012055602855980396, + 0.00617994787171483, + 0.00764221278950572, + 0.009290489368140697, + 0.010375852696597576, + 0.006609614938497543, + 0.0053583355620503426, + 0.025222424417734146, + 0.02242502011358738, + 0.005243260879069567, + 0.005834680050611496, + 0.008165255188941956, + 0.011477884836494923, + 0.002422566758468747, + 0.0024055715184658766, + 0.004179252311587334, + 0.0028827846981585026, + 0.0038573064375668764, + 0.003594903741031885, + 0.0037797896657139063, + 0.0028734784573316574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.19778387248516083 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03813815116882324 + ], + "opposite_logprob_tokens": [ + -3.2881381511688232 + ], + "logprob_reward": -0.03813815116882324, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03813815116882324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1624407172203064, + 0.058051906526088715, + 0.04846569523215294, + 0.14228205382823944, + 0.0367858000099659, + 0.05214061588048935, + 0.0756857693195343, + 0.0682508647441864, + 0.21242427825927734, + 0.14347240328788757 + ], + "credit_map": [ + 0.015383337624371052, + 0.01862304098904133, + 0.014918357133865356, + 0.0201907716691494, + 0.023966215550899506, + 0.016198182478547096, + 0.034615691751241684, + 0.01854509674012661, + 0.007836041040718555, + 0.009449657052755356, + 0.009018627926707268, + 0.006207372061908245, + 0.008650788106024265, + 0.005840484518557787, + 0.006098381709307432, + 0.00495054479688406, + 0.004955959971994162, + 0.0057697296142578125, + 0.0054284147918224335, + 0.006512716878205538, + 0.00994577631354332, + 0.0041144234128296375, + 0.005757772363722324, + 0.0059808967635035515, + 0.04736721143126488, + 0.008720655925571918, + 0.012437042780220509, + 0.0043136849999427795, + 0.006278748158365488, + 0.008844923228025436, + 0.006533696316182613, + 0.04778607562184334, + 0.0054686786606907845, + 0.003958947025239468, + 0.005531456787139177, + 0.006123900413513184, + 0.0033931261859834194, + 0.002622141968458891, + 0.006004714872688055, + 0.003682829672470689, + 0.006642958149313927, + 0.008197423070669174, + 0.006640406791120768, + 0.00612708879634738, + 0.007249808870255947, + 0.006742737255990505, + 0.005595849826931953, + 0.00494433892890811, + 0.018387489020824432, + 0.004759109113365412, + 0.010907256044447422, + 0.01396320853382349, + 0.005024460610002279, + 0.005861591547727585, + 0.007739921100437641, + 0.009042727760970592, + 0.007444652728736401, + 0.022073950618505478, + 0.008045922964811325, + 0.005711236968636513, + 0.004380736034363508, + 0.004590324126183987, + 0.009893638081848621, + 0.006110400427132845, + 0.020514464005827904, + 0.025892769917845726, + 0.02813749387860298, + 0.04510016739368439, + 0.02303052507340908, + 0.025059418752789497, + 0.028041787445545197, + 0.016647623851895332, + 0.010806352831423283, + 0.014277566224336624, + 0.023356646299362183, + 0.027287021279335022, + 0.017589982599020004, + 0.01599813625216484, + 0.017812976613640785, + 0.016343723982572556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.03813815116882324 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812725096940994 + ], + "opposite_logprob_tokens": [ + -3.2881271839141846 + ], + "logprob_reward": -0.03812725096940994, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812725096940994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3190421760082245, + 0.03657999262213707, + 0.05309087410569191, + 0.17827290296554565, + 0.04583509638905525, + 0.06436783075332642, + 0.036941949278116226, + 0.024463308975100517, + 0.07128259539604187, + 0.17012333869934082 + ], + "credit_map": [ + 0.028765957802534103, + 0.038764454424381256, + 0.036382947117090225, + 0.041378702968358994, + 0.044139452278614044, + 0.035808999091386795, + 0.059287574142217636, + 0.03451409563422203, + 0.004129332490265369, + 0.004140700679272413, + 0.006562081631273031, + 0.0033786690328270197, + 0.0042646462097764015, + 0.006165491882711649, + 0.004603441804647446, + 0.003335624933242798, + 0.006024347618222237, + 0.006299016997218132, + 0.008932532742619514, + 0.00610489584505558, + 0.00919489935040474, + 0.0031340327113866806, + 0.008892658166587353, + 0.004508491605520248, + 0.06349388509988785, + 0.008382081054151058, + 0.015440770424902439, + 0.004774702247232199, + 0.007250438444316387, + 0.010629904456436634, + 0.007836463861167431, + 0.06046465411782265, + 0.006635229103267193, + 0.005466745235025883, + 0.008015217259526253, + 0.005189166869968176, + 0.006655532401055098, + 0.004463610239326954, + 0.005462573375552893, + 0.003947024699300528, + 0.005802299827337265, + 0.00809347815811634, + 0.009913106448948383, + 0.007785405032336712, + 0.006904160138219595, + 0.005609060171991587, + 0.008344294503331184, + 0.01191602274775505, + 0.008056050166487694, + 0.0026823640801012516, + 0.005096952430903912, + 0.004115880001336336, + 0.0025254376232624054, + 0.004263201728463173, + 0.004424578510224819, + 0.005777482409030199, + 0.003235628129914403, + 0.0032383876387029886, + 0.0024143310729414225, + 0.0026671262457966805, + 0.0032854985911399126, + 0.003374305088073015, + 0.0036441627889871597, + 0.0026038691867142916, + 0.005602005869150162, + 0.00473759276792407, + 0.019231131300330162, + 0.017011653631925583, + 0.004108313005417585, + 0.006036713719367981, + 0.005976689979434013, + 0.008578496985137463, + 0.011775937862694263, + 0.010403371416032314, + 0.031332727521657944, + 0.05267104506492615, + 0.007727901451289654, + 0.011109513230621815, + 0.015707317739725113, + 0.029395531862974167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.03812725096940994 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081933043897152 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081933043897152, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081933043897152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2496005892753601, + 0.10816168785095215, + 0.05578616261482239, + 0.10827629268169403, + 0.04156740754842758, + 0.04889070615172386, + 0.04699208214879036, + 0.057890743017196655, + 0.08883979916572571, + 0.19399452209472656 + ], + "credit_map": [ + 0.020123548805713654, + 0.025390883907675743, + 0.023541748523712158, + 0.03207229822874069, + 0.03549446910619736, + 0.024672511965036392, + 0.056273672729730606, + 0.03203145042061806, + 0.014777743257582188, + 0.010269694030284882, + 0.01973498985171318, + 0.010276504792273045, + 0.011992436833679676, + 0.015443095937371254, + 0.013928701169788837, + 0.011738521046936512, + 0.0051470245234668255, + 0.006479551550000906, + 0.009258448146283627, + 0.007132761646062136, + 0.007437243591994047, + 0.005326677113771439, + 0.007058105431497097, + 0.007946353405714035, + 0.02479676529765129, + 0.008010836318135262, + 0.014082479290664196, + 0.008152390830218792, + 0.007092194631695747, + 0.0066610523499548435, + 0.014229615218937397, + 0.025250963866710663, + 0.006278978660702705, + 0.0043976823799312115, + 0.004819280933588743, + 0.008274759165942669, + 0.005308772902935743, + 0.003725106595084071, + 0.0045984541065990925, + 0.0041643730364739895, + 0.005420489236712456, + 0.004563435912132263, + 0.005116810090839863, + 0.0057281083427369595, + 0.006132225971668959, + 0.00714867003262043, + 0.006397875491529703, + 0.008383091539144516, + 0.008016015402972698, + 0.0029622975271195173, + 0.008353631012141705, + 0.007882451638579369, + 0.0027127524372190237, + 0.004792834632098675, + 0.005046642851084471, + 0.0072254603728652, + 0.0059302556328475475, + 0.014872143976390362, + 0.008899788372218609, + 0.007568421307951212, + 0.0027472744695842266, + 0.003053915686905384, + 0.009774215519428253, + 0.005044727586209774, + 0.008578536100685596, + 0.008321946486830711, + 0.013300052843987942, + 0.02089453861117363, + 0.008212670683860779, + 0.011656684800982475, + 0.00982444453984499, + 0.00805091205984354, + 0.016052205115556717, + 0.0156421959400177, + 0.03159911930561066, + 0.050275471061468124, + 0.014058208093047142, + 0.015879813581705093, + 0.021482868120074272, + 0.02900465577840805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.011081933043897152 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348169893026352 + ], + "opposite_logprob_tokens": [ + -3.651348114013672 + ], + "logprob_reward": -0.026348169893026352, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348169893026352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14643120765686035, + 0.33075231313705444, + 0.06995842605829239, + 0.0938325896859169, + 0.10190793126821518, + 0.05311153084039688, + 0.05862781032919884, + 0.03491014614701271, + 0.04649144038558006, + 0.06397659331560135 + ], + "credit_map": [ + 0.010699059814214706, + 0.013940812088549137, + 0.013748576864600182, + 0.01803150400519371, + 0.022205719724297523, + 0.01610134169459343, + 0.03521129488945007, + 0.016492893919348717, + 0.04265503212809563, + 0.02618790604174137, + 0.06978316605091095, + 0.029286768287420273, + 0.04028644412755966, + 0.04949769005179405, + 0.04174897447228432, + 0.03130630776286125, + 0.008216368965804577, + 0.009775687009096146, + 0.00957710575312376, + 0.0060004680417478085, + 0.010872500017285347, + 0.007565841544419527, + 0.01163563597947359, + 0.006314815022051334, + 0.015903441235423088, + 0.012534425593912601, + 0.0119165675714612, + 0.007167724426835775, + 0.010612424463033676, + 0.010134929791092873, + 0.012671630829572678, + 0.012891443446278572, + 0.01680924743413925, + 0.012684267945587635, + 0.023745408281683922, + 0.011364651843905449, + 0.011312250979244709, + 0.006008788477629423, + 0.011412120424211025, + 0.008571206592023373, + 0.006681591738015413, + 0.006224783603101969, + 0.006416293792426586, + 0.005048714112490416, + 0.007851431146264076, + 0.009508518502116203, + 0.005201979074627161, + 0.0061782230623066425, + 0.010517172515392303, + 0.004507461097091436, + 0.00760611379519105, + 0.011355213820934296, + 0.004422089084982872, + 0.00722125917673111, + 0.0058807507157325745, + 0.007117754779756069, + 0.00458443583920598, + 0.009509584866464138, + 0.0035420041531324387, + 0.0031398178543895483, + 0.003089813981205225, + 0.003717039478942752, + 0.0037863377947360277, + 0.00354111404158175, + 0.0027930731885135174, + 0.002096359385177493, + 0.012772590853273869, + 0.012987516820430756, + 0.0027549301739782095, + 0.0033055166713893414, + 0.0050289807841181755, + 0.004752475302666426, + 0.005940386559814215, + 0.003417915664613247, + 0.012963310815393925, + 0.015322171151638031, + 0.004386612679809332, + 0.005178287625312805, + 0.008272716775536537, + 0.008495194837450981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.026348169893026352 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-angle, ultra-wide shot captured in the grandiose style of Andreas Gursky, the camera tracks smoothly to the right, following a turbulent current resembling a subterranean river. This murky, silt-laden torrent, churning with white foam and forming powerful whirlpools in the gloom, relentlessly erodes the ancient maze-like foundations of the Angkor Wat temple complex. Rows of majestic sandstone pillars, covered in moss and water stains, with intricate relief carvings of ancient gods and nymphs (Apsaras), appear in the dim light. The torrent crashes violently against the pillars' bases, stirring up high waves. The water level rises rapidly, steadily and deliberately swallowing up the exquisite carvings one by one, submerging them beneath the murky yellow surface. A few rays of light from invisible cracks above cast shafts of light in the dusty air, illuminating the restless surface of the water and the dull texture of the wet stone. The entire picture presents a cold tone composed of dark brown, moss green and dark gray, creating a grand, tragic and unstoppable sense of historical destiny, perfectly embodying the eternal erosion of natural forces on the remains of human civilization.", + "seed": 2600, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_4/index_7532_seed_2600_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05137533657252789, + "vlm_logprob": -0.05137533657252789, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.05137533657252789, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08069297671318054 + ], + "opposite_logprob_tokens": [ + -3.330693006515503 + ], + "logprob_reward": -0.08069297671318054, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.08069297671318054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1855422407388687, + 0.15897966921329498, + 0.14944204688072205, + 0.18441475927829742, + 0.05447900667786598, + 0.04072166234254837, + 0.04401453211903572, + 0.044037219136953354, + 0.08826475590467453, + 0.05010411515831947 + ], + "credit_map": [ + 0.021250184625387192, + 0.03630145266652107, + 0.014348833821713924, + 0.020310482010245323, + 0.028449099510908127, + 0.028122762218117714, + 0.018389025703072548, + 0.01837039738893509, + 0.012835820205509663, + 0.02720893733203411, + 0.012571854516863823, + 0.019585436210036278, + 0.025890236720442772, + 0.028726745396852493, + 0.016763756051659584, + 0.015396879985928535, + 0.015777939930558205, + 0.019740121439099312, + 0.01865551993250847, + 0.020900007337331772, + 0.02449832670390606, + 0.013634700328111649, + 0.023534025996923447, + 0.012701408006250858, + 0.022753968834877014, + 0.02859746478497982, + 0.022391296923160553, + 0.0234193317592144, + 0.026653263717889786, + 0.012463996186852455, + 0.028307026252150536, + 0.01982841268181801, + 0.007074115797877312, + 0.006624649278819561, + 0.006173775065690279, + 0.004058405291289091, + 0.004616847261786461, + 0.004545066971331835, + 0.007615342736244202, + 0.013770801946520805, + 0.006271868012845516, + 0.005133018363267183, + 0.0040473248809576035, + 0.004275470972061157, + 0.005111752077937126, + 0.005760171916335821, + 0.006037597544491291, + 0.004084460437297821, + 0.005467663519084454, + 0.008644134737551212, + 0.002963966690003872, + 0.0044623506255447865, + 0.006025537848472595, + 0.005397296044975519, + 0.00421748124063015, + 0.0068361032754182816, + 0.00613313727080822, + 0.006631982512772083, + 0.005202075932174921, + 0.005030859261751175, + 0.00686218636110425, + 0.005097535904496908, + 0.004799014423042536, + 0.004280430264770985, + 0.009828750044107437, + 0.008244015276432037, + 0.008536769077181816, + 0.01183090079575777, + 0.009665992110967636, + 0.006909295916557312, + 0.02097710594534874, + 0.012271924875676632, + 0.008456741459667683, + 0.006179247982800007, + 0.0057564424350857735, + 0.004392435774207115, + 0.006006390321999788, + 0.008543371222913265, + 0.004994181450456381, + 0.005775306839495897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.08069297671318054 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906356245279312 + ], + "opposite_logprob_tokens": [ + -2.4640636444091797 + ], + "logprob_reward": -0.08906356245279312, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906356245279312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26162102818489075, + 0.0925193503499031, + 0.12940026819705963, + 0.07978888601064682, + 0.04305785149335861, + 0.19650278985500336, + 0.05862170830368996, + 0.05967317149043083, + 0.04125714302062988, + 0.037557780742645264 + ], + "credit_map": [ + 0.024360302835702896, + 0.0381523035466671, + 0.017844971269369125, + 0.039537087082862854, + 0.03618042543530464, + 0.03834528475999832, + 0.03315277770161629, + 0.03404786437749863, + 0.008975334465503693, + 0.018273184075951576, + 0.006600850727409124, + 0.01112827192991972, + 0.01455089170485735, + 0.0155774662271142, + 0.010231215506792068, + 0.007182145491242409, + 0.012762603349983692, + 0.03684534132480621, + 0.009561541490256786, + 0.010047980584204197, + 0.032410621643066406, + 0.011726035736501217, + 0.009032524190843105, + 0.0070136175490915775, + 0.009749706834554672, + 0.010976988822221756, + 0.009080062620341778, + 0.007357600145041943, + 0.011612667702138424, + 0.009088922291994095, + 0.010500866919755936, + 0.011422078125178814, + 0.007314539048820734, + 0.005810060538351536, + 0.006481496151536703, + 0.0034546228125691414, + 0.004350655246526003, + 0.0038658694829791784, + 0.005608548875898123, + 0.0061720628291368484, + 0.031032074242830276, + 0.022898945957422256, + 0.020305316895246506, + 0.023086268454790115, + 0.017118539661169052, + 0.02720983885228634, + 0.028875675052404404, + 0.025976121425628662, + 0.007629311643540859, + 0.012697027064859867, + 0.003941077273339033, + 0.006110471207648516, + 0.008230019360780716, + 0.007390662096440792, + 0.005756683647632599, + 0.006866458337754011, + 0.007961181923747063, + 0.0065156626515090466, + 0.005547964479774237, + 0.010622894391417503, + 0.006735646165907383, + 0.004365380387753248, + 0.011558898724615574, + 0.006365541834384203, + 0.006554632913321257, + 0.007095221430063248, + 0.003345632925629616, + 0.005391509737819433, + 0.0044282288290560246, + 0.005675747059285641, + 0.00402172701433301, + 0.0047444445081055164, + 0.005863953847438097, + 0.004579013679176569, + 0.0037981297355145216, + 0.0044371699914336205, + 0.0039822496473789215, + 0.006111993920058012, + 0.004300805274397135, + 0.004484466277062893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.08906356245279312 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498047545552254 + ], + "opposite_logprob_tokens": [ + -2.929980516433716 + ], + "logprob_reward": -0.05498047545552254, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498047545552254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42041242122650146, + 0.12949086725711823, + 0.10124768316745758, + 0.04629134386777878, + 0.03059866651892662, + 0.053022097796201706, + 0.08568931370973587, + 0.03422367572784424, + 0.06754676252603531, + 0.031477201730012894 + ], + "credit_map": [ + 0.05157836899161339, + 0.06093844771385193, + 0.02517339028418064, + 0.06709617376327515, + 0.05801869183778763, + 0.05840842053294182, + 0.0497901551425457, + 0.04940873384475708, + 0.011900058947503567, + 0.022404314950108528, + 0.007308062165975571, + 0.018047286197543144, + 0.01943698152899742, + 0.020649297162890434, + 0.016900524497032166, + 0.012844348326325417, + 0.009039799682796001, + 0.02074512466788292, + 0.009382413700222969, + 0.009943598881363869, + 0.020429736003279686, + 0.009079094976186752, + 0.013737022876739502, + 0.008890895172953606, + 0.005177550483494997, + 0.00687900185585022, + 0.004068142734467983, + 0.003822026075795293, + 0.007805696222931147, + 0.008693122304975986, + 0.0052867853082716465, + 0.004559019114822149, + 0.005082016810774803, + 0.004218833986669779, + 0.0045782420784235, + 0.0019215343054383993, + 0.0034584100358188152, + 0.00262095732614398, + 0.003466608701273799, + 0.005252063274383545, + 0.006292352452874184, + 0.006660841405391693, + 0.00417632982134819, + 0.00613494822755456, + 0.005083540454506874, + 0.005693495739251375, + 0.011551103554666042, + 0.007429486606270075, + 0.012221870943903923, + 0.014311651699244976, + 0.0056178877130150795, + 0.010495767928659916, + 0.008904265239834785, + 0.01253185048699379, + 0.010125999338924885, + 0.01148002315312624, + 0.006185302976518869, + 0.00722266361117363, + 0.0031639444641768932, + 0.0042783599346876144, + 0.005644520744681358, + 0.003123676171526313, + 0.002504212548956275, + 0.0021009952761232853, + 0.011406492441892624, + 0.015635522082448006, + 0.005025553982704878, + 0.00700525613501668, + 0.00786186009645462, + 0.007335726637393236, + 0.007359561510384083, + 0.005916785914450884, + 0.004708023276180029, + 0.006220563780516386, + 0.004212443251162767, + 0.0035005409736186266, + 0.002518272027373314, + 0.003559437580406666, + 0.003423268673941493, + 0.0033346519339829683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.05498047545552254 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607177034020424 + ], + "opposite_logprob_tokens": [ + -4.141071796417236 + ], + "logprob_reward": -0.01607177034020424, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607177034020424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4446692168712616, + 0.07899218797683716, + 0.09107948839664459, + 0.05061837285757065, + 0.041029442101716995, + 0.15212340652942657, + 0.056384120136499405, + 0.031101416796445847, + 0.038797132670879364, + 0.015205231495201588 + ], + "credit_map": [ + 0.04560001939535141, + 0.062025316059589386, + 0.02647228352725506, + 0.07249975204467773, + 0.06290865689516068, + 0.06557044386863708, + 0.05209611728787422, + 0.05749666690826416, + 0.011965450830757618, + 0.01531998161226511, + 0.006194198504090309, + 0.010988181456923485, + 0.009967256337404251, + 0.011572160758078098, + 0.006986026186496019, + 0.005998930893838406, + 0.007304397411644459, + 0.028192324563860893, + 0.006018481682986021, + 0.004947783891111612, + 0.02695748209953308, + 0.006558433175086975, + 0.007315014488995075, + 0.0037855657283216715, + 0.0048810397274792194, + 0.006898500490933657, + 0.00476029422134161, + 0.0030265802051872015, + 0.005822186358273029, + 0.00807228498160839, + 0.006130571942776442, + 0.011026913300156593, + 0.0089868837967515, + 0.006352832540869713, + 0.004901421722024679, + 0.002712137997150421, + 0.0037585587706416845, + 0.0033648298121988773, + 0.006281948182731867, + 0.0046708290465176105, + 0.02384897693991661, + 0.017215924337506294, + 0.018038611859083176, + 0.024540554732084274, + 0.012516545131802559, + 0.019241977483034134, + 0.019333699718117714, + 0.017387110739946365, + 0.007707174401730299, + 0.0106696467846632, + 0.00358795584179461, + 0.005829161033034325, + 0.010987547226250172, + 0.006545005366206169, + 0.005496567115187645, + 0.005561064928770065, + 0.004078022204339504, + 0.004987049847841263, + 0.0031739906407892704, + 0.0029348868411034346, + 0.006748784799128771, + 0.004218202084302902, + 0.0024875507224351168, + 0.00247292872518301, + 0.005521152634173632, + 0.005676088854670525, + 0.003948666155338287, + 0.004364132881164551, + 0.003238241421058774, + 0.00306723196990788, + 0.0075089409947395325, + 0.005472677294164896, + 0.0021293493919074535, + 0.0023961681872606277, + 0.0023422741796821356, + 0.0015587123343721032, + 0.0011573083465918899, + 0.0018014867091551423, + 0.0017685374477878213, + 0.002051395131275058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.01607177034020424 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot captured from a high-angle, ultra-wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera track smoothly to the right following a turbulent subterranean river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are rows of sandstone pillars with intricate relief carvings of ancient gods and nymphs (Apsaras) visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the water murky, silt-laden, and yellow, actively rising to submerge the carvings one by one?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are shafts of light visible casting through dusty air onto the water and wet stone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067897900938988 + ], + "opposite_logprob_tokens": [ + -4.141067981719971 + ], + "logprob_reward": -0.016067897900938988, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067897900938988, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5886176824569702, + 0.20507311820983887, + 0.1287798285484314, + 0.009559274651110172, + 0.017226548865437508, + 0.013932258822023869, + 0.011687682010233402, + 0.0035221336875110865, + 0.014340338297188282, + 0.007261075545102358 + ], + "credit_map": [ + 0.0903576910495758, + 0.0838586837053299, + 0.03366674482822418, + 0.09055694937705994, + 0.0771496593952179, + 0.07608190178871155, + 0.07262077182531357, + 0.06432527303695679, + 0.016806960105895996, + 0.03870655968785286, + 0.010781383141875267, + 0.027371924370527267, + 0.03386329486966133, + 0.03673737123608589, + 0.019937237724661827, + 0.020868375897407532, + 0.010006249882280827, + 0.03920451179146767, + 0.005994574166834354, + 0.007207285612821579, + 0.03651908412575722, + 0.014823006466031075, + 0.008395270444452763, + 0.00662984699010849, + 0.0011128131300210953, + 0.001356058637611568, + 0.0009088977240025997, + 0.0007554377079941332, + 0.0013734472449868917, + 0.001134232385084033, + 0.0013292150106281042, + 0.0015891732182353735, + 0.003061889670789242, + 0.0024065643083304167, + 0.0016896271845325828, + 0.0016221236437559128, + 0.00199153833091259, + 0.0021762929391115904, + 0.002388426335528493, + 0.001890085986815393, + 0.002243540482595563, + 0.0016283588483929634, + 0.0015170946717262268, + 0.0014389761490747333, + 0.0013963456731289625, + 0.0020828733686357737, + 0.0019304299494251609, + 0.0016946394462138414, + 0.001783300074748695, + 0.002197881927713752, + 0.0009293905459344387, + 0.00113777129445225, + 0.001866753213107586, + 0.0014830573927611113, + 0.0010033772559836507, + 0.0012861513532698154, + 0.00048241016338579357, + 0.00033912729122675955, + 0.00038132184999994934, + 0.0006406389293260872, + 0.00046463150647468865, + 0.00030946143670007586, + 0.0004641615378204733, + 0.0004403809434734285, + 0.0025624565314501524, + 0.0028919021133333445, + 0.0009349830797873437, + 0.0007675985107198358, + 0.001755029079504311, + 0.00287792831659317, + 0.0009541889885440469, + 0.0015962516190484166, + 0.0011440798407420516, + 0.0011167575139552355, + 0.0007382089970633388, + 0.0006436718394979835, + 0.0007873533759266138, + 0.0013239162508398294, + 0.0006257623317651451, + 0.0008813254535198212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.016067897900938988 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_5/index_tensor([8201, 8201])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_5/index_tensor([8201, 8201])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b226fb12d6f4315a9953d30a1db03daea97c526c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_5/index_tensor([8201, 8201])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dramatic, eye-level extreme close-up focuses on a freshly sliced chocolate lava cake, the camera held still to capture every minute detail. The cake, with its dark, tender, slightly matte crumb, sits in the center of a pristine, glossy white ceramic dish. A clean cut reveals its core, and a thick, molten river of dark chocolate begins to ooze out luxuriously. The liquid chocolate, with its incredible sheen, captures the warm, soft light from a single, directional light source just off-camera, creating brilliant, mirrored highlights as it flows steadily and deliberately down the sides of the cake. The lava pools at the bottom, forming an ever-expanding, mirror-like puddle on the bare white dish. Its slow, viscous movement creates an atmosphere of decadent, irresistible indulgence against the softly blurred dark background.", + "seed": 19119, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_5/index_8201_seed_19119_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.30844272500835357, + "vlm_logprob": -0.30844272500835357, + "vlm_yes_no_margin": 2.624999988079071, + "vlm_token_credit": -0.30844272500835357, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029246004298329353 + ], + "opposite_logprob_tokens": [ + -3.9042460918426514 + ], + "logprob_reward": -0.029246004298329353, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.029246004298329353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16495417058467865, + 0.10897917300462723, + 0.0870630145072937, + 0.0428951159119606, + 0.15497663617134094, + 0.06433239579200745, + 0.17821623384952545, + 0.06055906042456627, + 0.06716731190681458, + 0.07085686177015305 + ], + "credit_map": [ + 0.020822634920477867, + 0.018796773627400398, + 0.02571193128824234, + 0.01475139707326889, + 0.027965767309069633, + 0.021253330633044243, + 0.015037165023386478, + 0.020615186542272568, + 0.01580631360411644, + 0.01369744073599577, + 0.013591984286904335, + 0.012387461960315704, + 0.012376436963677406, + 0.013338561169803143, + 0.01059831865131855, + 0.017182668671011925, + 0.011853335425257683, + 0.012175345793366432, + 0.010961515828967094, + 0.010196503251791, + 0.010072566568851471, + 0.012468471191823483, + 0.011077656410634518, + 0.008257628418505192, + 0.006140653509646654, + 0.0057662008330225945, + 0.005022360477596521, + 0.006288711912930012, + 0.004261631052941084, + 0.005608584266155958, + 0.004713186528533697, + 0.005093790590763092, + 0.022907542064785957, + 0.020263250917196274, + 0.02298056147992611, + 0.01385733112692833, + 0.021564314141869545, + 0.021583234891295433, + 0.012184610590338707, + 0.019635792821645737, + 0.006455704104155302, + 0.007781240623444319, + 0.011886275373399258, + 0.007923340424895287, + 0.007512740790843964, + 0.009406877681612968, + 0.005371606443077326, + 0.007994616404175758, + 0.020557481795549393, + 0.01933976449072361, + 0.019192520529031754, + 0.04076765105128288, + 0.02144831418991089, + 0.026003582403063774, + 0.01300851907581091, + 0.017898399382829666, + 0.008309749886393547, + 0.007549561560153961, + 0.005228896625339985, + 0.006570848636329174, + 0.012489697895944118, + 0.008192387409508228, + 0.005713269114494324, + 0.0065046572126448154, + 0.010348348878324032, + 0.008714823983609676, + 0.005621629301458597, + 0.007004826795309782, + 0.015375721268355846, + 0.00862502958625555, + 0.0049692136235535145, + 0.006507718004286289, + 0.007839369587600231, + 0.009699661284685135, + 0.006860914174467325, + 0.0085067143663764, + 0.014370168559253216, + 0.009790366515517235, + 0.006495956797152758, + 0.007293711882084608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.029246004298329353 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42901259660720825 + ], + "opposite_logprob_tokens": [ + -1.0540125370025635 + ], + "logprob_reward": -0.42901259660720825, + "yes_no_margin_reward": 0.6249999403953552, + "token_credit_reward": -0.42901259660720825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2893775403499603, + 0.06121375411748886, + 0.031672559678554535, + 0.048602864146232605, + 0.04057380184531212, + 0.061717282980680466, + 0.0488923043012619, + 0.06098978966474533, + 0.2039773017168045, + 0.15298280119895935 + ], + "credit_map": [ + 0.031173041090369225, + 0.027163105085492134, + 0.03876494988799095, + 0.027972111478447914, + 0.02922358177602291, + 0.034609559923410416, + 0.03342216834425926, + 0.06704899668693542, + 0.010084002278745174, + 0.0062461006455123425, + 0.0063422792591154575, + 0.008808373473584652, + 0.006141996942460537, + 0.007642840035259724, + 0.007822682149708271, + 0.008125470019876957, + 0.004385338630527258, + 0.005592680536210537, + 0.003712988458573818, + 0.0043416377156972885, + 0.002811074024066329, + 0.004587547853589058, + 0.003947831224650145, + 0.002293459139764309, + 0.00605360371991992, + 0.0062013063579797745, + 0.006820472422987223, + 0.0050710588693618774, + 0.006609817035496235, + 0.006678216625005007, + 0.006232148036360741, + 0.004936235956847668, + 0.005764862522482872, + 0.0053181881085038185, + 0.005640091374516487, + 0.0037439633160829544, + 0.007400874048471451, + 0.0060362378135323524, + 0.0027498931158334017, + 0.003919687122106552, + 0.00662566302344203, + 0.0077236369252204895, + 0.007495969068259001, + 0.0073836673982441425, + 0.008965843357145786, + 0.009541746228933334, + 0.005309515167027712, + 0.008671240881085396, + 0.005191209726035595, + 0.005108100362122059, + 0.005514366552233696, + 0.007918951101601124, + 0.009497111663222313, + 0.006939985323697329, + 0.0037248150911182165, + 0.004997759126126766, + 0.008925472386181355, + 0.007050988730043173, + 0.0066371182911098, + 0.00527050020173192, + 0.008908059448003769, + 0.008487800136208534, + 0.006495294626802206, + 0.009214547462761402, + 0.03535287082195282, + 0.023213546723127365, + 0.021176772192120552, + 0.018142499029636383, + 0.027312057092785835, + 0.027509987354278564, + 0.02543691359460354, + 0.025832630693912506, + 0.022775720804929733, + 0.018158867955207825, + 0.014915721490979195, + 0.02254413813352585, + 0.016631927341222763, + 0.02139844000339508, + 0.017990849912166595, + 0.01856711506843567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.42901259660720825 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036305959802120924 + ], + "opposite_logprob_tokens": [ + -5.628630638122559 + ], + "logprob_reward": -0.0036305959802120924, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036305959802120924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3220789134502411, + 0.03607317805290222, + 0.038805775344371796, + 0.0412154495716095, + 0.1265728920698166, + 0.09089528024196625, + 0.08208812773227692, + 0.05270373076200485, + 0.16628308594226837, + 0.0432836189866066 + ], + "credit_map": [ + 0.04058106988668442, + 0.031072162091732025, + 0.05122089013457298, + 0.03537118062376976, + 0.03430847451090813, + 0.03613283485174179, + 0.034511953592300415, + 0.058880340307950974, + 0.007134158164262772, + 0.003992119804024696, + 0.004573459271341562, + 0.0036371033638715744, + 0.005106708966195583, + 0.0050887358374893665, + 0.0026491847820580006, + 0.003891701577231288, + 0.005746785085648298, + 0.005547240376472473, + 0.004481830168515444, + 0.0051467278972268105, + 0.00554684130474925, + 0.005283951759338379, + 0.004157385788857937, + 0.002895006677135825, + 0.005383682902902365, + 0.005452093668282032, + 0.004403463564813137, + 0.007107673678547144, + 0.0041389623656868935, + 0.0051429602317512035, + 0.004407099448144436, + 0.0051795123144984245, + 0.016953501850366592, + 0.02034679241478443, + 0.018127379938960075, + 0.011682027950882912, + 0.01791856810450554, + 0.0170233566313982, + 0.009908212348818779, + 0.014613043516874313, + 0.011660945601761341, + 0.012044197879731655, + 0.014572307467460632, + 0.009952345862984657, + 0.011916186660528183, + 0.012746412307024002, + 0.007382662501186132, + 0.01062021218240261, + 0.013185925781726837, + 0.010464628227055073, + 0.008584647439420223, + 0.007691920269280672, + 0.015865812078118324, + 0.010542646050453186, + 0.006090129725635052, + 0.009662413969635963, + 0.007888193242251873, + 0.005466231144964695, + 0.005411122459918261, + 0.004885435104370117, + 0.010159381665289402, + 0.006995347794145346, + 0.005200437270104885, + 0.006697577890008688, + 0.033976104110479355, + 0.01897537335753441, + 0.016671128571033478, + 0.015623214654624462, + 0.024475762620568275, + 0.021167242899537086, + 0.015476264990866184, + 0.019917981699109077, + 0.006900687236338854, + 0.005314068403095007, + 0.0046005891636013985, + 0.00872811395674944, + 0.005249224603176117, + 0.005672571714967489, + 0.003915613517165184, + 0.002902743872255087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.0036305959802120924 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026367442682385445 + ], + "opposite_logprob_tokens": [ + -3.651367425918579 + ], + "logprob_reward": -0.026367442682385445, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026367442682385445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21437998116016388, + 0.04021440073847771, + 0.03709941357374191, + 0.039490923285484314, + 0.09568336606025696, + 0.1504463106393814, + 0.12998558580875397, + 0.07176388055086136, + 0.15524166822433472, + 0.0656944215297699 + ], + "credit_map": [ + 0.03246293216943741, + 0.022720733657479286, + 0.030527908354997635, + 0.024783197790384293, + 0.02598898857831955, + 0.023414427414536476, + 0.020062152296304703, + 0.03441964462399483, + 0.004391571041196585, + 0.0038141722325235605, + 0.005489672534167767, + 0.004326744470745325, + 0.005455228965729475, + 0.005577385425567627, + 0.0044854809530079365, + 0.0066741458140313625, + 0.004808558616787195, + 0.004271079320460558, + 0.004069975111633539, + 0.004595569800585508, + 0.005173393525183201, + 0.005156488157808781, + 0.003959839232265949, + 0.005064512137323618, + 0.005536101292818785, + 0.0035794833675026894, + 0.004737053997814655, + 0.005115948151797056, + 0.003706261282786727, + 0.004817074630409479, + 0.004411952570080757, + 0.007587049622088671, + 0.0169517882168293, + 0.009289756417274475, + 0.013985437341034412, + 0.007385795470327139, + 0.010868178680539131, + 0.013041268102824688, + 0.009705825708806515, + 0.01445531751960516, + 0.01727033220231533, + 0.017381535843014717, + 0.023490475490689278, + 0.017254846170544624, + 0.01919306255877018, + 0.021890755742788315, + 0.011846602894365788, + 0.022118695080280304, + 0.014646482653915882, + 0.01704980805516243, + 0.01488781813532114, + 0.020628949627280235, + 0.025366611778736115, + 0.016999632120132446, + 0.010365286841988564, + 0.01004099752753973, + 0.008743718266487122, + 0.007985095493495464, + 0.007420816924422979, + 0.005533282645046711, + 0.015593135729432106, + 0.009325575083494186, + 0.007368394173681736, + 0.00979386642575264, + 0.02919004112482071, + 0.01659317873418331, + 0.01798359677195549, + 0.01463315263390541, + 0.023311691358685493, + 0.020396528765559196, + 0.015211428515613079, + 0.0179220512509346, + 0.009770839475095272, + 0.007764861453324556, + 0.006946466397494078, + 0.00984497182071209, + 0.008088522590696812, + 0.008602356538176537, + 0.00723262969404459, + 0.007443778682500124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.026367442682385445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0539569854736328 + ], + "opposite_logprob_tokens": [ + -0.4289570152759552 + ], + "logprob_reward": -1.0539569854736328, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0539569854736328, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.22360943257808685, + 0.0827060341835022, + 0.04899206385016441, + 0.04716731607913971, + 0.03584323450922966, + 0.06839662790298462, + 0.144108846783638, + 0.07739941775798798, + 0.12970827519893646, + 0.1420687437057495 + ], + "credit_map": [ + 0.026676390320062637, + 0.022146563977003098, + 0.028701353818178177, + 0.026436829939484596, + 0.02521560713648796, + 0.027393408119678497, + 0.02460102178156376, + 0.04243827238678932, + 0.01256564911454916, + 0.009843670763075352, + 0.008702024817466736, + 0.0105648348107934, + 0.009840627200901508, + 0.011181550100445747, + 0.008402755483984947, + 0.01160492654889822, + 0.0061375475488603115, + 0.007606158498674631, + 0.004875482525676489, + 0.006476001814007759, + 0.006709180772304535, + 0.006567477714270353, + 0.0060576144605875015, + 0.00456260284408927, + 0.00741167925298214, + 0.006935057230293751, + 0.006059839855879545, + 0.004717146512120962, + 0.0058579337783157825, + 0.00705074705183506, + 0.00460214726626873, + 0.00453276839107275, + 0.0055632940493524075, + 0.003887961385771632, + 0.005254547577351332, + 0.0026529061142355204, + 0.004034134559333324, + 0.005312894005328417, + 0.004123347345739603, + 0.005014151334762573, + 0.00942253228276968, + 0.007136613130569458, + 0.011155069805681705, + 0.009215154685080051, + 0.006920969113707542, + 0.009730233810842037, + 0.005751579999923706, + 0.009064479731023312, + 0.01741670072078705, + 0.01538701169192791, + 0.01662076823413372, + 0.03241661563515663, + 0.01810716651380062, + 0.020629577338695526, + 0.01078042108565569, + 0.01275059673935175, + 0.013207736425101757, + 0.009420590475201607, + 0.008255083113908768, + 0.011744886636734009, + 0.010756283067166805, + 0.009115952998399734, + 0.00706949969753623, + 0.007829387672245502, + 0.014804639853537083, + 0.01830352284014225, + 0.01201253104954958, + 0.016893791034817696, + 0.019482582807540894, + 0.018200146034359932, + 0.015202878043055534, + 0.01480818074196577, + 0.024576576426625252, + 0.016858384013175964, + 0.015533964149653912, + 0.024385053664445877, + 0.019080517813563347, + 0.01839578151702881, + 0.012884072959423065, + 0.010354398749768734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -1.0539569854736328 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "dramatic, eye-level extreme close-up focuses on a freshly sliced chocolate lava cake, the camera held still to capture every minute detail. The cake, with its dark, tender, slightly matte crumb, sits in the center of a pristine, glossy white ceramic dish. A clean cut reveals its core, and a thick, molten river of dark chocolate begins to ooze out luxuriously. The liquid chocolate, with its incredible sheen, captures the warm, soft light from a single, directional light source just off-camera, creating brilliant, mirrored highlights as it flows steadily and deliberately down the sides of the cake. The lava pools at the bottom, forming an ever-expanding, mirror-like puddle on the bare white dish. Its slow, viscous movement creates an atmosphere of decadent, irresistible indulgence against the softly blurred dark background.", + "seed": 19120, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_5/index_8201_seed_19120_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1911132974550128, + "vlm_logprob": -0.1911132974550128, + "vlm_yes_no_margin": 3.025, + "vlm_token_credit": -0.1911132974550128, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035077836364507675 + ], + "opposite_logprob_tokens": [ + -3.6600778102874756 + ], + "logprob_reward": -0.035077836364507675, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.035077836364507675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05796287581324577, + 0.13993120193481445, + 0.05636568367481232, + 0.1892423778772354, + 0.11052010208368301, + 0.08955115079879761, + 0.1454140841960907, + 0.08278838545084, + 0.10625579953193665, + 0.021968374028801918 + ], + "credit_map": [ + 0.0052642980590462685, + 0.010371577925980091, + 0.004557779058814049, + 0.00649633351713419, + 0.012992645613849163, + 0.002384102437645197, + 0.011835671029984951, + 0.004060467705130577, + 0.02374235913157463, + 0.011471492238342762, + 0.008828363381326199, + 0.01307796873152256, + 0.02156863547861576, + 0.012265895493328571, + 0.04015776515007019, + 0.008818713016808033, + 0.010009421966969967, + 0.006114646792411804, + 0.007066542748361826, + 0.006067216396331787, + 0.006829514633864164, + 0.00460161454975605, + 0.009130469523370266, + 0.006546256598085165, + 0.016694745048880577, + 0.027948511764407158, + 0.030722741037607193, + 0.028532342985272408, + 0.0283044446259737, + 0.01508705597370863, + 0.02637595124542713, + 0.015576601028442383, + 0.012661649845540524, + 0.011974549852311611, + 0.026955006644129753, + 0.017293598502874374, + 0.01456766203045845, + 0.008171744644641876, + 0.012765453197062016, + 0.006130438297986984, + 0.006416652351617813, + 0.011140814051032066, + 0.011770770885050297, + 0.015196402557194233, + 0.01913086697459221, + 0.006796883884817362, + 0.013290122151374817, + 0.0058086346834897995, + 0.00988106057047844, + 0.01884157769382, + 0.01831713691353798, + 0.029246114194393158, + 0.027218755334615707, + 0.01178561057895422, + 0.020171625539660454, + 0.009952206164598465, + 0.005845667328685522, + 0.01115841418504715, + 0.010265051387250423, + 0.01689237356185913, + 0.01640000008046627, + 0.004879420157521963, + 0.011226422153413296, + 0.006121031474322081, + 0.00944877602159977, + 0.010416657663881779, + 0.015266881324350834, + 0.02203001081943512, + 0.019618963822722435, + 0.006252529565244913, + 0.015248575247824192, + 0.007973400875926018, + 0.002150529995560646, + 0.0026832432486116886, + 0.0032566809095442295, + 0.003931400831788778, + 0.0033714519813656807, + 0.00127982790581882, + 0.0032393496949225664, + 0.002055891091004014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.035077836364507675 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07008305937051773 + ], + "opposite_logprob_tokens": [ + -2.6950831413269043 + ], + "logprob_reward": -0.07008305937051773, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07008305937051773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0797671303153038, + 0.4319811761379242, + 0.06214727461338043, + 0.11336031556129456, + 0.03987238556146622, + 0.043785348534584045, + 0.06755224615335464, + 0.06097889691591263, + 0.08157487213611603, + 0.018980346620082855 + ], + "credit_map": [ + 0.012296441942453384, + 0.008323089219629765, + 0.009251927025616169, + 0.0072324201464653015, + 0.014236826449632645, + 0.0043290541507303715, + 0.009765813127160072, + 0.014331553131341934, + 0.0945686548948288, + 0.02868570014834404, + 0.04137824475765228, + 0.03512130305171013, + 0.04922031611204147, + 0.024240046739578247, + 0.1302371472120285, + 0.028529765084385872, + 0.007908262312412262, + 0.00781610980629921, + 0.008164352737367153, + 0.008436770178377628, + 0.008595199324190617, + 0.004360520280897617, + 0.010211152955889702, + 0.006654903758317232, + 0.01275604497641325, + 0.017204513773322105, + 0.010261047631502151, + 0.020126860588788986, + 0.020162958651781082, + 0.008549383841454983, + 0.01571672223508358, + 0.00858277641236782, + 0.004735908005386591, + 0.0050630876794457436, + 0.005813310854136944, + 0.008774001151323318, + 0.005679022520780563, + 0.0025699883699417114, + 0.005063590593636036, + 0.002173472661525011, + 0.0032537474762648344, + 0.004653591196984053, + 0.005684462375938892, + 0.007824170403182507, + 0.010710556991398335, + 0.002584564033895731, + 0.006707092747092247, + 0.002367161214351654, + 0.006339667830616236, + 0.0067413183860480785, + 0.009615294635295868, + 0.014418456703424454, + 0.010279137641191483, + 0.005574414040893316, + 0.009698661975562572, + 0.004885285627096891, + 0.0046033491380512714, + 0.008707978762686253, + 0.009908368811011314, + 0.009494639933109283, + 0.011339288204908371, + 0.0046913595870137215, + 0.0075988867320120335, + 0.004635022487491369, + 0.007290556561201811, + 0.008694934658706188, + 0.012942004948854446, + 0.017042962834239006, + 0.013932819478213787, + 0.00563086150214076, + 0.01153959333896637, + 0.004501136485487223, + 0.0015317306388169527, + 0.0023567250464111567, + 0.0023849639110267162, + 0.00377128878608346, + 0.0026997202076017857, + 0.001736804493702948, + 0.002927161054685712, + 0.0015719522489234805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.07008305937051773 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005970382131636143 + ], + "opposite_logprob_tokens": [ + -5.130970478057861 + ], + "logprob_reward": -0.005970382131636143, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005970382131636143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05593340843915939, + 0.0836980789899826, + 0.05717562511563301, + 0.1329020857810974, + 0.13512557744979858, + 0.17244075238704681, + 0.10040014982223511, + 0.1140444204211235, + 0.12425624579191208, + 0.024023637175559998 + ], + "credit_map": [ + 0.006091705057770014, + 0.0076060667634010315, + 0.007514494936913252, + 0.005100418347865343, + 0.009911326691508293, + 0.004329553805291653, + 0.008515586145222187, + 0.006864248309284449, + 0.01737257093191147, + 0.006821237504482269, + 0.007774198893457651, + 0.007561509497463703, + 0.010423863306641579, + 0.0060119666159152985, + 0.0215471014380455, + 0.006185627076774836, + 0.004960350692272186, + 0.008820676244795322, + 0.00415494991466403, + 0.008125207386910915, + 0.010135926306247711, + 0.004895740654319525, + 0.00821594800800085, + 0.007866824977099895, + 0.02052944526076317, + 0.017125297337770462, + 0.023097604513168335, + 0.015932006761431694, + 0.019542574882507324, + 0.010723179206252098, + 0.01729751192033291, + 0.008654454723000526, + 0.016530513763427734, + 0.013970484025776386, + 0.03379961475729942, + 0.02340221032500267, + 0.013441618531942368, + 0.009330087341368198, + 0.016007138416171074, + 0.008643915876746178, + 0.014559820294380188, + 0.020826224237680435, + 0.02797369658946991, + 0.02679392509162426, + 0.037182923406362534, + 0.010443471372127533, + 0.024413110688328743, + 0.010247577913105488, + 0.009485344402492046, + 0.011314960196614265, + 0.020038655027747154, + 0.01441034022718668, + 0.015921343117952347, + 0.008064874447882175, + 0.011367467232048512, + 0.00979716144502163, + 0.008218078874051571, + 0.014496867544949055, + 0.01765311323106289, + 0.02005806565284729, + 0.02447216957807541, + 0.007101650349795818, + 0.014035183936357498, + 0.00800928846001625, + 0.011352548375725746, + 0.012540850788354874, + 0.02149835042655468, + 0.024280143901705742, + 0.01886548474431038, + 0.008531732484698296, + 0.019390353932976723, + 0.007796767167747021, + 0.003738851984962821, + 0.0025195337366312742, + 0.00295656849630177, + 0.005287602543830872, + 0.0024290960282087326, + 0.0015066330088302493, + 0.0034500567708164454, + 0.002135294023901224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.005970382131636143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018211418762803078 + ], + "opposite_logprob_tokens": [ + -4.018211364746094 + ], + "logprob_reward": -0.018211418762803078, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018211418762803078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06338996440172195, + 0.22058174014091492, + 0.10834331810474396, + 0.12514132261276245, + 0.05459069088101387, + 0.09083006531000137, + 0.14211440086364746, + 0.08051122725009918, + 0.053163670003414154, + 0.06133360043168068 + ], + "credit_map": [ + 0.007032189983874559, + 0.009279526770114899, + 0.00454573193565011, + 0.004994896240532398, + 0.014900095760822296, + 0.006163059733808041, + 0.01051350124180317, + 0.005960968788713217, + 0.03334832936525345, + 0.020455241203308105, + 0.023961147293448448, + 0.020780101418495178, + 0.029082786291837692, + 0.014657807536423206, + 0.06208733841776848, + 0.01620897464454174, + 0.012527164071798325, + 0.01832340657711029, + 0.008478913456201553, + 0.01649632677435875, + 0.015740124508738518, + 0.007027890533208847, + 0.017222169786691666, + 0.01252731867134571, + 0.011705121025443077, + 0.017664963379502296, + 0.021366002038121223, + 0.01834435947239399, + 0.016850799322128296, + 0.010964907705783844, + 0.017942069098353386, + 0.010303091257810593, + 0.006802288815379143, + 0.006064070854336023, + 0.012249978259205818, + 0.008980797603726387, + 0.006651869043707848, + 0.004247670527547598, + 0.006203192286193371, + 0.00339082651771605, + 0.008658421225845814, + 0.01041441410779953, + 0.018382776528596878, + 0.01237927284091711, + 0.012638192623853683, + 0.0074165137484669685, + 0.01313034351915121, + 0.0078101251274347305, + 0.013620003126561642, + 0.014591999351978302, + 0.0261753611266613, + 0.01989944837987423, + 0.02211150713264942, + 0.011659574694931507, + 0.018503418192267418, + 0.015553091652691364, + 0.0069308956153690815, + 0.008759547956287861, + 0.01273155864328146, + 0.014052403159439564, + 0.01247282512485981, + 0.008203673176467419, + 0.010600660927593708, + 0.006759666837751865, + 0.005458923056721687, + 0.005742760840803385, + 0.009843869134783745, + 0.00911980401724577, + 0.008995696902275085, + 0.003935645800083876, + 0.007212699390947819, + 0.0028542708605527878, + 0.005707209929823875, + 0.0067715696059167385, + 0.008952968753874302, + 0.010659215971827507, + 0.008523664437234402, + 0.005017551593482494, + 0.008963937871158123, + 0.00673748180270195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 3 + ], + "selected_reward": -0.018211418762803078 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8262237906455994 + ], + "opposite_logprob_tokens": [ + -0.5762237906455994 + ], + "logprob_reward": -0.8262237906455994, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8262237906455994, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09516266733407974, + 0.14463870227336884, + 0.16582998633384705, + 0.10883710533380508, + 0.03257666155695915, + 0.08481159806251526, + 0.14296266436576843, + 0.08256547898054123, + 0.11623769998550415, + 0.0263773575425148 + ], + "credit_map": [ + 0.011829198338091373, + 0.009696141816675663, + 0.0070692552253603935, + 0.0075406888499855995, + 0.022679900750517845, + 0.006521106697618961, + 0.015313811600208282, + 0.01451258547604084, + 0.020316962152719498, + 0.014420599676668644, + 0.014905431307852268, + 0.018572116270661354, + 0.027477160096168518, + 0.010625742375850677, + 0.029066067188978195, + 0.009254639968276024, + 0.01925336755812168, + 0.02129070647060871, + 0.014455202966928482, + 0.024478832259774208, + 0.02109774388372898, + 0.009755291976034641, + 0.03913804143667221, + 0.016360828652977943, + 0.009683283977210522, + 0.016036689281463623, + 0.01745361089706421, + 0.014778665266931057, + 0.022135339677333832, + 0.007761194370687008, + 0.014658581465482712, + 0.006329749245196581, + 0.0041146534495055676, + 0.004043865948915482, + 0.004793923813849688, + 0.005181576125323772, + 0.004919738043099642, + 0.0026365588419139385, + 0.005062680225819349, + 0.0018236730247735977, + 0.009563819505274296, + 0.010362992994487286, + 0.0127548947930336, + 0.01001073606312275, + 0.013518593274056911, + 0.0074202814139425755, + 0.012681816704571247, + 0.008498483337461948, + 0.012824421748518944, + 0.014504734426736832, + 0.02059904672205448, + 0.023363331332802773, + 0.02465992234647274, + 0.012631186284124851, + 0.02064647525548935, + 0.0137335741892457, + 0.005984040908515453, + 0.012055184692144394, + 0.010736631229519844, + 0.01523685548454523, + 0.014782893471419811, + 0.005363571457564831, + 0.011137482710182667, + 0.007268827874213457, + 0.009517583064734936, + 0.012863139621913433, + 0.01697668805718422, + 0.024272017180919647, + 0.021411165595054626, + 0.007152422331273556, + 0.015611733309924603, + 0.008432965725660324, + 0.002799056004732847, + 0.003245157189667225, + 0.003035768400877714, + 0.004895930178463459, + 0.0050432113930583, + 0.0019825242925435305, + 0.0035637945402413607, + 0.001811919268220663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.8262237906455994 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "dramatic, eye-level extreme close-up focuses on a freshly sliced chocolate lava cake, the camera held still to capture every minute detail. The cake, with its dark, tender, slightly matte crumb, sits in the center of a pristine, glossy white ceramic dish. A clean cut reveals its core, and a thick, molten river of dark chocolate begins to ooze out luxuriously. The liquid chocolate, with its incredible sheen, captures the warm, soft light from a single, directional light source just off-camera, creating brilliant, mirrored highlights as it flows steadily and deliberately down the sides of the cake. The lava pools at the bottom, forming an ever-expanding, mirror-like puddle on the bare white dish. Its slow, viscous movement creates an atmosphere of decadent, irresistible indulgence against the softly blurred dark background.", + "seed": 19121, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_5/index_8201_seed_19121_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1424465744756162, + "vlm_logprob": -0.1424465744756162, + "vlm_yes_no_margin": 3.0249999940395353, + "vlm_token_credit": -0.1424465744756162, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03847786411643028 + ], + "opposite_logprob_tokens": [ + -3.538477897644043 + ], + "logprob_reward": -0.03847786411643028, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03847786411643028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3459837734699249, + 0.31034281849861145, + 0.054158758372068405, + 0.06583236157894135, + 0.07807131856679916, + 0.03883747383952141, + 0.024273168295621872, + 0.0220574252307415, + 0.0413065105676651, + 0.01913638785481453 + ], + "credit_map": [ + 0.0901648998260498, + 0.02841142937541008, + 0.023528579622507095, + 0.028536953032016754, + 0.02705800160765648, + 0.023817600682377815, + 0.016038784757256508, + 0.10842752456665039, + 0.10943741351366043, + 0.01803439110517502, + 0.01313627790659666, + 0.016760753467679024, + 0.015612439252436161, + 0.01693810150027275, + 0.011767564341425896, + 0.1086558848619461, + 0.007783735170960426, + 0.004909212701022625, + 0.011209207586944103, + 0.006044827401638031, + 0.004961800295859575, + 0.007181288208812475, + 0.0050188214518129826, + 0.007049869745969772, + 0.007160988636314869, + 0.007304326631128788, + 0.011560440994799137, + 0.010127585381269455, + 0.00640898896381259, + 0.007745085284113884, + 0.009201399981975555, + 0.006323541514575481, + 0.011159297078847885, + 0.008540323935449123, + 0.012882928363978863, + 0.010784931480884552, + 0.00904977135360241, + 0.00777785899117589, + 0.010467653162777424, + 0.007408551871776581, + 0.005535576958209276, + 0.006284283008426428, + 0.0029007678385823965, + 0.00533134862780571, + 0.0068715885281562805, + 0.003914614208042622, + 0.0049778674729168415, + 0.003021428594365716, + 0.0034781554713845253, + 0.0032928013242781162, + 0.0032499847002327442, + 0.002651655813679099, + 0.002828592201694846, + 0.00305473362095654, + 0.003151322016492486, + 0.002565922448411584, + 0.0028703028801828623, + 0.003193933516740799, + 0.0025450061075389385, + 0.003137304913252592, + 0.002988452324643731, + 0.0018622076604515314, + 0.003078550100326538, + 0.00238166656345129, + 0.0045773242600262165, + 0.005005154758691788, + 0.006445007864385843, + 0.0044669643975794315, + 0.005271406844258308, + 0.004587156698107719, + 0.004657957702875137, + 0.006295539438724518, + 0.002165447222068906, + 0.0030608021188527346, + 0.0019851839169859886, + 0.0023860405199229717, + 0.002431137254461646, + 0.002208869205787778, + 0.0025297305546700954, + 0.0023691775277256966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.03847786411643028 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799364537000656 + ], + "opposite_logprob_tokens": [ + -1.804936408996582 + ], + "logprob_reward": -0.1799364537000656, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799364537000656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45004087686538696, + 0.23177610337734222, + 0.047642942517995834, + 0.034948740154504776, + 0.07053717225790024, + 0.04749299958348274, + 0.04809034615755081, + 0.018844660371541977, + 0.03192413970828056, + 0.018702026456594467 + ], + "credit_map": [ + 0.11084845662117004, + 0.027086962014436722, + 0.15329040586948395, + 0.03625874221324921, + 0.034559424966573715, + 0.020939156413078308, + 0.017913775518536568, + 0.04914393275976181, + 0.05002898722887039, + 0.017339691519737244, + 0.020391905680298805, + 0.01837458275258541, + 0.018988801166415215, + 0.02928551472723484, + 0.014470133930444717, + 0.0628964900970459, + 0.007551700342446566, + 0.0036249414552003145, + 0.008077939040958881, + 0.0038664215244352818, + 0.004632817581295967, + 0.009048950858414173, + 0.003215071512386203, + 0.0076251039281487465, + 0.004952451679855585, + 0.0030274512246251106, + 0.008201969787478447, + 0.004538699518889189, + 0.003430284559726715, + 0.0036602914333343506, + 0.0034913120325654745, + 0.0036462778225541115, + 0.010341963730752468, + 0.007333343382924795, + 0.008479895070195198, + 0.012657210230827332, + 0.006411972921341658, + 0.007434315513819456, + 0.008451739326119423, + 0.009426726028323174, + 0.008149612694978714, + 0.004833353217691183, + 0.007189374417066574, + 0.007415253669023514, + 0.00414566183462739, + 0.004193099681288004, + 0.004362526815384626, + 0.007204117719084024, + 0.007380976807326078, + 0.004961362108588219, + 0.009634501300752163, + 0.005384077318012714, + 0.004611678421497345, + 0.0051450300961732864, + 0.004427645355463028, + 0.006545072887092829, + 0.0024227206595242023, + 0.0022320239804685116, + 0.0030819172970950603, + 0.0022892325650900602, + 0.0020303353667259216, + 0.002005375688895583, + 0.0018226506654173136, + 0.0029604046139866114, + 0.003855540882796049, + 0.003458056366071105, + 0.005589789245277643, + 0.0031209730077534914, + 0.003667497308924794, + 0.003912350162863731, + 0.0033881894778460264, + 0.0049317446537315845, + 0.00233279075473547, + 0.0022524790838360786, + 0.003163775894790888, + 0.002076509175822139, + 0.0023491529282182455, + 0.0022655357606709003, + 0.001978771761059761, + 0.0022830097004771233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.1799364537000656 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005272767972201109 + ], + "opposite_logprob_tokens": [ + -5.25527286529541 + ], + "logprob_reward": -0.005272767972201109, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005272767972201109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10658983141183853, + 0.5834476351737976, + 0.06024226173758507, + 0.04374202713370323, + 0.07621970027685165, + 0.05931085720658302, + 0.017474716529250145, + 0.016492599621415138, + 0.02299991436302662, + 0.013480382971465588 + ], + "credit_map": [ + 0.012644626200199127, + 0.011095112189650536, + 0.024053022265434265, + 0.010689057409763336, + 0.011703403666615486, + 0.008111652918159962, + 0.0067358738742768764, + 0.021557090803980827, + 0.20018534362316132, + 0.03222208842635155, + 0.02468700520694256, + 0.023902924731373787, + 0.028829719871282578, + 0.032836198806762695, + 0.02104494720697403, + 0.2197394222021103, + 0.007942664436995983, + 0.006588981486856937, + 0.01019279658794403, + 0.006895382888615131, + 0.0063328552059829235, + 0.008537858724594116, + 0.006497703026980162, + 0.007254021242260933, + 0.005664416588842869, + 0.004812201019376516, + 0.008220926858484745, + 0.0048634945414960384, + 0.004139316733926535, + 0.0066725159995257854, + 0.004486572928726673, + 0.004882583860307932, + 0.012818954885005951, + 0.007870110683143139, + 0.00892720464617014, + 0.007819979451596737, + 0.01015490386635065, + 0.00910468865185976, + 0.010813139379024506, + 0.008710718713700771, + 0.010011736303567886, + 0.006775415502488613, + 0.00711588142439723, + 0.007979137822985649, + 0.007961252704262733, + 0.006705442443490028, + 0.0071145957335829735, + 0.005647401791065931, + 0.002633474301546812, + 0.0018602233612909913, + 0.0033740350045263767, + 0.0017268214141950011, + 0.0018052525119856, + 0.0018793062772601843, + 0.0017255254788324237, + 0.002470080042257905, + 0.0020768840331584215, + 0.0022820117883384228, + 0.001498269964940846, + 0.0018284325487911701, + 0.002352738520130515, + 0.002366139553487301, + 0.0020903744734823704, + 0.001997749786823988, + 0.0031056487932801247, + 0.0032066244166344404, + 0.0029221531003713608, + 0.002684474689885974, + 0.002727100858464837, + 0.002377219730988145, + 0.002563383663073182, + 0.0034133109729737043, + 0.0015108335064724088, + 0.0016509999986737967, + 0.002519728848710656, + 0.00151471805293113, + 0.001544603961519897, + 0.001535533694550395, + 0.0016162469983100891, + 0.0015877189580351114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.005272767972201109 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014211636036634445 + ], + "opposite_logprob_tokens": [ + -4.264211654663086 + ], + "logprob_reward": -0.014211636036634445, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014211636036634445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32149794697761536, + 0.24114568531513214, + 0.01783093251287937, + 0.03672312945127487, + 0.04647623002529144, + 0.04225903004407883, + 0.03278602659702301, + 0.09756524115800858, + 0.12870432436466217, + 0.03501138091087341 + ], + "credit_map": [ + 0.06264094263315201, + 0.023138059303164482, + 0.1021263524889946, + 0.02998224087059498, + 0.024345455691218376, + 0.01619141548871994, + 0.015645181760191917, + 0.04742831736803055, + 0.08909232914447784, + 0.012580012902617455, + 0.009552989155054092, + 0.008719992823898792, + 0.011604919098317623, + 0.01317149493843317, + 0.006924498826265335, + 0.08949945122003555, + 0.002501613460481167, + 0.0023379020858556032, + 0.0023528276942670345, + 0.0020243474282324314, + 0.0019443732453510165, + 0.0027338331565260887, + 0.0017203337047249079, + 0.002215700689703226, + 0.004692289046943188, + 0.004229768645018339, + 0.007183540146797895, + 0.0051071117632091045, + 0.0033241156488656998, + 0.004086136352270842, + 0.00383804296143353, + 0.004262126050889492, + 0.007553372532129288, + 0.004432896617799997, + 0.007867095991969109, + 0.005876616574823856, + 0.0048363530077040195, + 0.006121091544628143, + 0.004454196430742741, + 0.00533460546284914, + 0.005266224499791861, + 0.005420898552983999, + 0.006148333661258221, + 0.0053899940103292465, + 0.005185242276638746, + 0.0042237224988639355, + 0.0064114294946193695, + 0.004213183186948299, + 0.004808109253644943, + 0.00413029408082366, + 0.005341480486094952, + 0.003921094350516796, + 0.0030895613599568605, + 0.004170953296124935, + 0.0034886873327195644, + 0.003835845971480012, + 0.01646931655704975, + 0.01137422677129507, + 0.01682484708726406, + 0.008797867223620415, + 0.010221382603049278, + 0.011688299477100372, + 0.007217936683446169, + 0.014971357770264149, + 0.020650386810302734, + 0.014178089797496796, + 0.022526821121573448, + 0.012815963476896286, + 0.0129767544567585, + 0.014930871315300465, + 0.011198911815881729, + 0.019426528364419937, + 0.005962912924587727, + 0.0037284553982317448, + 0.006400729529559612, + 0.0032529637683182955, + 0.0036595205310732126, + 0.0036307305563241243, + 0.0031572587322443724, + 0.005218809936195612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.014211636036634445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47433415055274963 + ], + "opposite_logprob_tokens": [ + -0.9743341207504272 + ], + "logprob_reward": -0.47433415055274963, + "yes_no_margin_reward": 0.4999999701976776, + "token_credit_reward": -0.47433415055274963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43518760800361633, + 0.3587193787097931, + 0.026330694556236267, + 0.014102887362241745, + 0.02680058963596821, + 0.015383503399789333, + 0.013906150124967098, + 0.06501425802707672, + 0.033720050007104874, + 0.01083491649478674 + ], + "credit_map": [ + 0.11985551565885544, + 0.023644201457500458, + 0.15267804265022278, + 0.035919927060604095, + 0.03177601471543312, + 0.017912806943058968, + 0.014162132516503334, + 0.03923898935317993, + 0.12738420069217682, + 0.019953865557909012, + 0.013470254838466644, + 0.014159677550196648, + 0.018391983583569527, + 0.0189192071557045, + 0.013145876117050648, + 0.13329432904720306, + 0.003203606465831399, + 0.002737695351243019, + 0.004023910034447908, + 0.004371833987534046, + 0.002728931372985244, + 0.002985424594953656, + 0.0025822839234024286, + 0.0036970097571611404, + 0.0016077168984338641, + 0.00189955101814121, + 0.0017719335155561566, + 0.00205237022601068, + 0.0015157453017309308, + 0.0018967364449054003, + 0.0016284460434690118, + 0.0017303884960711002, + 0.0045262654311954975, + 0.002907274290919304, + 0.006357772741466761, + 0.0032285801135003567, + 0.0026281087193638086, + 0.0019555550534278154, + 0.0022159789223223925, + 0.002981058554723859, + 0.0023688534274697304, + 0.0017688500229269266, + 0.0029322353657335043, + 0.0018974222475662827, + 0.0014981620479375124, + 0.0019470916595309973, + 0.001588081126101315, + 0.0013828087830916047, + 0.002267710864543915, + 0.002165964338928461, + 0.0017309149261564016, + 0.001106816460378468, + 0.0019048226531594992, + 0.0014016577042639256, + 0.00173036172054708, + 0.001597901340574026, + 0.00997065007686615, + 0.007548762951046228, + 0.00966040138155222, + 0.007268028799444437, + 0.006753899157047272, + 0.005939054768532515, + 0.00735587440431118, + 0.010517595335841179, + 0.005770607851445675, + 0.0043759653344750404, + 0.0051216846331954, + 0.003286033868789673, + 0.003244770457968116, + 0.003966666758060455, + 0.0034080820623785257, + 0.004546239506453276, + 0.0014479957753792405, + 0.0015002809232100844, + 0.0014122274005785584, + 0.0011390310246497393, + 0.0014209875371307135, + 0.0012560448376461864, + 0.0013345261104404926, + 0.0013238233514130116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.47433415055274963 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dramatic, eye-level extreme close-up focuses on a freshly sliced chocolate lava cake, the camera held still to capture every minute detail. The cake, with its dark, tender, slightly matte crumb, sits in the center of a pristine, glossy white ceramic dish. A clean cut reveals its core, and a thick, molten river of dark chocolate begins to ooze out luxuriously. The liquid chocolate, with its incredible sheen, captures the warm, soft light from a single, directional light source just off-camera, creating brilliant, mirrored highlights as it flows steadily and deliberately down the sides of the cake. The lava pools at the bottom, forming an ever-expanding, mirror-like puddle on the bare white dish. Its slow, viscous movement creates an atmosphere of decadent, irresistible indulgence against the softly blurred dark background.", + "seed": 19122, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_5/index_8201_seed_19122_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.27878691235091535, + "vlm_logprob": -0.27878691235091535, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.27878691235091535, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030911173671483994 + ], + "opposite_logprob_tokens": [ + -3.7809112071990967 + ], + "logprob_reward": -0.030911173671483994, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.030911173671483994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15231306850910187, + 0.044063981622457504, + 0.18309178948402405, + 0.07260514050722122, + 0.06537353992462158, + 0.051106132566928864, + 0.290252685546875, + 0.07398317009210587, + 0.028137974441051483, + 0.03907248005270958 + ], + "credit_map": [ + 0.03512188419699669, + 0.027480967342853546, + 0.017939956858754158, + 0.01396992802619934, + 0.021629750728607178, + 0.015199743211269379, + 0.009823375381529331, + 0.01114746741950512, + 0.01006778422743082, + 0.008069458417594433, + 0.005114514380693436, + 0.003723838599398732, + 0.006045714020729065, + 0.00287039577960968, + 0.003769309027120471, + 0.004402972292155027, + 0.03865816444158554, + 0.03425392508506775, + 0.01362245250493288, + 0.018743472173810005, + 0.025554684922099113, + 0.01305578462779522, + 0.022813783958554268, + 0.016389526426792145, + 0.019157927483320236, + 0.010209646075963974, + 0.007236391305923462, + 0.007862296886742115, + 0.00894977431744337, + 0.005035140551626682, + 0.005699755623936653, + 0.008454222232103348, + 0.011871017515659332, + 0.009562884457409382, + 0.007424019742757082, + 0.011427379213273525, + 0.009801742620766163, + 0.003982877358794212, + 0.007159063592553139, + 0.0041445596143603325, + 0.0067459107376635075, + 0.004845496732741594, + 0.007454379461705685, + 0.006100810598582029, + 0.007972928695380688, + 0.004783382639288902, + 0.007508176378905773, + 0.005695051047950983, + 0.038080938160419464, + 0.019483862444758415, + 0.11435839533805847, + 0.02130761556327343, + 0.046990618109703064, + 0.00830068625509739, + 0.03238014876842499, + 0.009350460022687912, + 0.00955080334097147, + 0.009387131780385971, + 0.014794987626373768, + 0.007793913595378399, + 0.013502932153642178, + 0.004329419694840908, + 0.007814011536538601, + 0.006809981539845467, + 0.003167412942275405, + 0.004338773433119059, + 0.004219538532197475, + 0.0030703695956617594, + 0.006429779808968306, + 0.0022545261308550835, + 0.0021880792919546366, + 0.0024694951716810465, + 0.003681698814034462, + 0.006321536377072334, + 0.005891646258533001, + 0.005329639185220003, + 0.006397981662303209, + 0.0028216801583766937, + 0.005268308334052563, + 0.0033599918242543936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.030911173671483994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42898309230804443 + ], + "opposite_logprob_tokens": [ + -1.0539830923080444 + ], + "logprob_reward": -0.42898309230804443, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42898309230804443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0617661252617836, + 0.04631868377327919, + 0.12253295630216599, + 0.2203805148601532, + 0.03882361575961113, + 0.10513689368963242, + 0.1369255930185318, + 0.15775066614151, + 0.07778729498386383, + 0.03257761523127556 + ], + "credit_map": [ + 0.015408066101372242, + 0.01202345173805952, + 0.0072098965756595135, + 0.005461790133267641, + 0.009259803220629692, + 0.0035490933805704117, + 0.00444989325478673, + 0.004404128063470125, + 0.00984176341444254, + 0.008473783731460571, + 0.0050475639291107655, + 0.004792403429746628, + 0.0077370163053274155, + 0.002862845780327916, + 0.004947375506162643, + 0.002615934470668435, + 0.019168106839060783, + 0.019150611013174057, + 0.010311445221304893, + 0.016381612047553062, + 0.017177758738398552, + 0.006697980221360922, + 0.020503085106611252, + 0.013142364099621773, + 0.03235787898302078, + 0.030342314392328262, + 0.03139127045869827, + 0.028570329770445824, + 0.030506104230880737, + 0.018712185323238373, + 0.023345105350017548, + 0.02515532076358795, + 0.006713844370096922, + 0.004899356979876757, + 0.0047436985187232494, + 0.006066733971238136, + 0.005827856250107288, + 0.0030049565248191357, + 0.004778529983013868, + 0.0027886363677680492, + 0.01523909717798233, + 0.010495582595467567, + 0.011329654604196548, + 0.009531675837934017, + 0.012417727150022984, + 0.01822848618030548, + 0.00812440738081932, + 0.019770262762904167, + 0.014393121935427189, + 0.011803613975644112, + 0.045441579073667526, + 0.011324603110551834, + 0.027349350973963737, + 0.005996681284159422, + 0.012236541137099266, + 0.008380110375583172, + 0.018682966008782387, + 0.01713629812002182, + 0.04190510883927345, + 0.013303590938448906, + 0.027521464973688126, + 0.011755202896893024, + 0.015643395483493805, + 0.011802637949585915, + 0.008128671906888485, + 0.007973470725119114, + 0.0199115052819252, + 0.008214502595365047, + 0.011343101970851421, + 0.005986266769468784, + 0.008889791555702686, + 0.007339982781559229, + 0.0046572876162827015, + 0.004410726483911276, + 0.0047710551880300045, + 0.0034824202302843332, + 0.005508702248334885, + 0.0020854345057159662, + 0.0046847909688949585, + 0.0029771996196359396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 6 + ], + "selected_reward": -0.42898309230804443 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019511011196300387 + ], + "opposite_logprob_tokens": [ + -6.251951217651367 + ], + "logprob_reward": -0.0019511011196300387, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019511011196300387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10942468792200089, + 0.07765720784664154, + 0.07760036736726761, + 0.039092808961868286, + 0.06184348836541176, + 0.1132412925362587, + 0.29519179463386536, + 0.11065535992383957, + 0.036657460033893585, + 0.07863558083772659 + ], + "credit_map": [ + 0.021875819191336632, + 0.0207851380109787, + 0.017705876380205154, + 0.010017113760113716, + 0.015799233689904213, + 0.0067503671161830425, + 0.008113951422274113, + 0.008377181366086006, + 0.016422279179096222, + 0.012727547436952591, + 0.006988209672272205, + 0.007472367491573095, + 0.011638864874839783, + 0.010604510083794594, + 0.006700543686747551, + 0.005102884955704212, + 0.012238143011927605, + 0.011113903485238552, + 0.006162265781313181, + 0.008561767637729645, + 0.010456986725330353, + 0.009837879799306393, + 0.011502240784466267, + 0.007727180141955614, + 0.011086584068834782, + 0.005231720395386219, + 0.0034989879932254553, + 0.003334613749757409, + 0.00557017233222723, + 0.0033284565433859825, + 0.003885807003825903, + 0.0031564647797495127, + 0.008595267310738564, + 0.011735280975699425, + 0.007640980649739504, + 0.010039906948804855, + 0.009803525172173977, + 0.0039234161376953125, + 0.0055377185344696045, + 0.004567391239106655, + 0.013695187866687775, + 0.010618027299642563, + 0.016638049855828285, + 0.011181646026670933, + 0.022732490673661232, + 0.014419685117900372, + 0.01045619323849678, + 0.013500017113983631, + 0.04186413437128067, + 0.02177051454782486, + 0.12030871957540512, + 0.017519323155283928, + 0.04598202928900719, + 0.007537923287600279, + 0.03267187625169754, + 0.007537264842540026, + 0.013222798705101013, + 0.011203986592590809, + 0.028914159163832664, + 0.009325109422206879, + 0.017825203016400337, + 0.00898811686784029, + 0.010968679562211037, + 0.010207302868366241, + 0.005805960390716791, + 0.00790337286889553, + 0.003824664046987891, + 0.0037132410798221827, + 0.00644749915227294, + 0.002511988626793027, + 0.0027774637565016747, + 0.003673271741718054, + 0.0100551201030612, + 0.009326408617198467, + 0.01099217776209116, + 0.007423190400004387, + 0.013346360996365547, + 0.004972725175321102, + 0.016428427770733833, + 0.006091168615967035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.0019511011196300387 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372691571712494 + ], + "opposite_logprob_tokens": [ + -3.408726930618286 + ], + "logprob_reward": -0.03372691571712494, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372691571712494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1849466860294342, + 0.07266631722450256, + 0.05940211936831474, + 0.0495760440826416, + 0.07098947465419769, + 0.12449970096349716, + 0.24399714171886444, + 0.1027250811457634, + 0.06108657643198967, + 0.03011086769402027 + ], + "credit_map": [ + 0.050167642533779144, + 0.03206473961472511, + 0.017199262976646423, + 0.017667008563876152, + 0.03083096630871296, + 0.011410136707127094, + 0.014547426253557205, + 0.01105948444455862, + 0.015107821673154831, + 0.011217697523534298, + 0.00973750464618206, + 0.006721961311995983, + 0.010753319598734379, + 0.005701872520148754, + 0.008743543177843094, + 0.004682597238570452, + 0.010204809717833996, + 0.011438660323619843, + 0.005508858244866133, + 0.0074852644465863705, + 0.008498763665556908, + 0.004771714098751545, + 0.005544943269342184, + 0.005949107930064201, + 0.010037017986178398, + 0.007889775559306145, + 0.0066597131080925465, + 0.004757065791636705, + 0.00626898230984807, + 0.0030466450843960047, + 0.004019801039248705, + 0.0068970429711043835, + 0.012416202574968338, + 0.00961652398109436, + 0.00966986920684576, + 0.011913732625544071, + 0.011096864007413387, + 0.005631907843053341, + 0.0066497656516730785, + 0.0039946166798472404, + 0.01367863267660141, + 0.012079804204404354, + 0.0200851671397686, + 0.016608502715826035, + 0.023632710799574852, + 0.01293756254017353, + 0.013110094703733921, + 0.012367219664156437, + 0.02313864976167679, + 0.02024558000266552, + 0.07343456894159317, + 0.02631402015686035, + 0.0492224395275116, + 0.014468151144683361, + 0.023229427635669708, + 0.013944297097623348, + 0.015307603403925896, + 0.013387374579906464, + 0.0147520387545228, + 0.010191524401307106, + 0.021102525293827057, + 0.009349602274596691, + 0.008049269206821918, + 0.010585138574242592, + 0.007708114571869373, + 0.009900355711579323, + 0.013479593209922314, + 0.004566126968711615, + 0.010957074351608753, + 0.0035824738442897797, + 0.006022811401635408, + 0.0048700240440666676, + 0.00508534861728549, + 0.0033945769537240267, + 0.0034078839235007763, + 0.003577850293368101, + 0.0038312498945742846, + 0.002785350428894162, + 0.0037415942642837763, + 0.004287014715373516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.03372691571712494 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level extreme close-up of a freshly sliced chocolate lava cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the cake sitting in a pristine, glossy white ceramic dish in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a thick, molten river of dark chocolate oozing out from the cut core of the cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene lit by a single directional light source creating brilliant mirrored highlights on the liquid chocolate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983622789382935 + ], + "opposite_logprob_tokens": [ + -0.5233622789382935 + ], + "logprob_reward": -0.8983622789382935, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983622789382935, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09167669713497162, + 0.014549474231898785, + 0.045304957777261734, + 0.036701418459415436, + 0.1120796799659729, + 0.07430386543273926, + 0.4766174256801605, + 0.058295849710702896, + 0.042062096297740936, + 0.048408493399620056 + ], + "credit_map": [ + 0.02292805165052414, + 0.0157693549990654, + 0.010371061973273754, + 0.008655351586639881, + 0.014219284988939762, + 0.0066019692458212376, + 0.007212123367935419, + 0.005919499322772026, + 0.0026421351358294487, + 0.002064249711111188, + 0.0019130792934447527, + 0.0012663399102166295, + 0.0019569117575883865, + 0.0015273867174983025, + 0.001063895644620061, + 0.0021154771093279123, + 0.009345592930912971, + 0.00825682282447815, + 0.0037826618645340204, + 0.005108106881380081, + 0.006601294502615929, + 0.0032521181274205446, + 0.005796265322715044, + 0.003162095556035638, + 0.006785979960113764, + 0.005079165101051331, + 0.004425171297043562, + 0.004037976730614901, + 0.0052599310874938965, + 0.00263894721865654, + 0.0037205854896456003, + 0.004753661807626486, + 0.023434096947312355, + 0.02262123115360737, + 0.012915496714413166, + 0.020777421072125435, + 0.01574460044503212, + 0.004564870614558458, + 0.0083064716309309, + 0.0037154958117753267, + 0.01061453390866518, + 0.008828048594295979, + 0.010689967311918736, + 0.010039891116321087, + 0.016756992787122726, + 0.0039026252925395966, + 0.006874348036944866, + 0.00659746490418911, + 0.06703650206327438, + 0.03316191956400871, + 0.19698317348957062, + 0.02976318448781967, + 0.06676225364208221, + 0.011972271837294102, + 0.05996581166982651, + 0.010972305200994015, + 0.0066827950067818165, + 0.006071374285966158, + 0.01273059006780386, + 0.0061078197322785854, + 0.010327832773327827, + 0.0049667819403111935, + 0.0057046134024858475, + 0.0057040429674088955, + 0.004589809104800224, + 0.005186057183891535, + 0.010284108109772205, + 0.003349174512550235, + 0.007861370220780373, + 0.0029574029613286257, + 0.004219254944473505, + 0.003614920424297452, + 0.00498522724956274, + 0.006858135107904673, + 0.006369487382471561, + 0.006144139915704727, + 0.009245449677109718, + 0.004419354256242514, + 0.006431330926716328, + 0.003955369349569082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.8983622789382935 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_6/index_tensor([3905, 3905])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_6/index_tensor([3905, 3905])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0bd71267a6e12320f88743d67de6744ede124fc3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_6/index_tensor([3905, 3905])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle drone glides slowly and steadily forward, tracing the languid path of a slow-flowing river as it weaves through a dense, old-growth forest. The camera maintains a steady, dreamlike pace above the treetops, capturing the entire canopy, forming a near-continuous ceiling above the water. The interconnected branches and lush foliage are vibrant shades of golden yellow and rich ochre. Below, the river's dark, calm surface acts like a dynamic mirror; it captures the fiery foliage, but the gentle, almost imperceptible current distorts the reflection into soft, twisted, and painterly images of the trees. A thick, ethereal mist slowly billows over the water, its thin tendrils curling and clinging to the surface, creating a profound sense of tranquility and mystery. The entire scene is bathed in a soft, diffuse light, as if filtering through the mist itself, casting a warm golden glow that enhances the mysterious and serene atmosphere.", + "seed": 15424, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_6/index_3905_seed_15424_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.420532594807446, + "vlm_logprob": -0.420532594807446, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.420532594807446, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.713754415512085 + ], + "opposite_logprob_tokens": [ + -0.21375446021556854 + ], + "logprob_reward": -1.713754415512085, + "yes_no_margin_reward": -1.5, + "token_credit_reward": -1.713754415512085, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1871776431798935, + 0.11630106717348099, + 0.12466534972190857, + 0.1407897174358368, + 0.049609456211328506, + 0.06120573729276657, + 0.1552240550518036, + 0.05399654805660248, + 0.042153242975473404, + 0.0688771978020668 + ], + "credit_map": [ + 0.017143014818429947, + 0.01948838122189045, + 0.03502228483557701, + 0.018136803060770035, + 0.030237670987844467, + 0.016887106001377106, + 0.020863043144345284, + 0.029399320483207703, + 0.007464362774044275, + 0.007740330882370472, + 0.012581131421029568, + 0.009946544654667377, + 0.007500019855797291, + 0.014487917535007, + 0.041951004415750504, + 0.014629739336669445, + 0.007108170073479414, + 0.005803182255476713, + 0.012824267148971558, + 0.010934487916529179, + 0.010327732190489769, + 0.020261850208044052, + 0.04313324764370918, + 0.014272397384047508, + 0.010890512727200985, + 0.011845403350889683, + 0.021402746438980103, + 0.012408627197146416, + 0.012062450870871544, + 0.013874712400138378, + 0.030497968196868896, + 0.027807289734482765, + 0.004573683254420757, + 0.006199233699589968, + 0.006229598540812731, + 0.005941423121839762, + 0.006227496545761824, + 0.004874793346971273, + 0.008396608754992485, + 0.007166616152971983, + 0.006180033087730408, + 0.002988559892401099, + 0.005649399012327194, + 0.003584841499105096, + 0.005553442984819412, + 0.005035083740949631, + 0.017272910103201866, + 0.014941458590328693, + 0.01944003440439701, + 0.01562261488288641, + 0.01925501599907875, + 0.01384205836802721, + 0.02361759915947914, + 0.01718621887266636, + 0.023451900109648705, + 0.02280859462916851, + 0.005154098384082317, + 0.0058536953292787075, + 0.009773499332368374, + 0.003896275768056512, + 0.007758922874927521, + 0.005886525381356478, + 0.009095018729567528, + 0.006578509230166674, + 0.003439905121922493, + 0.0032982276752591133, + 0.004240314941853285, + 0.0035702544264495373, + 0.003499653423205018, + 0.00391372200101614, + 0.01110017579048872, + 0.009090987034142017, + 0.004547793418169022, + 0.011057727970182896, + 0.005621327552944422, + 0.005919144954532385, + 0.011664177291095257, + 0.012560485862195492, + 0.010026322677731514, + 0.00748021062463522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -1.713754415512085 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048706669360399246 + ], + "opposite_logprob_tokens": [ + -3.0487067699432373 + ], + "logprob_reward": -0.048706669360399246, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048706669360399246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21717911958694458, + 0.15119050443172455, + 0.060352157801389694, + 0.11494587361812592, + 0.05876254662871361, + 0.06003078073263168, + 0.09130444377660751, + 0.09196513146162033, + 0.07858376950025558, + 0.07568565756082535 + ], + "credit_map": [ + 0.014064846560359001, + 0.020532723516225815, + 0.02891561947762966, + 0.015030946582555771, + 0.024216391146183014, + 0.017406878992915154, + 0.03983468562364578, + 0.057177022099494934, + 0.010291448794305325, + 0.015051975846290588, + 0.021800119429826736, + 0.013747982680797577, + 0.014697094447910786, + 0.024030214175581932, + 0.03886237367987633, + 0.01270930003374815, + 0.004431485198438168, + 0.0044390978291630745, + 0.00996062159538269, + 0.006369482260197401, + 0.004322397988289595, + 0.01037458423525095, + 0.01451870147138834, + 0.005935789551585913, + 0.009102897718548775, + 0.00991385243833065, + 0.017111683264374733, + 0.010949304327368736, + 0.011552756652235985, + 0.01298852451145649, + 0.01967257261276245, + 0.023654280230402946, + 0.004709878005087376, + 0.0059590269811451435, + 0.009097767062485218, + 0.006043898873031139, + 0.0059007746167480946, + 0.007427564822137356, + 0.009870591573417187, + 0.009753040969371796, + 0.009898361749947071, + 0.007838829420506954, + 0.006634275894612074, + 0.004893313627690077, + 0.00846591405570507, + 0.005460833199322224, + 0.009856833145022392, + 0.006982419639825821, + 0.008610294200479984, + 0.00978250801563263, + 0.014002992771565914, + 0.007510325405746698, + 0.011157567612826824, + 0.008959708735346794, + 0.015239397995173931, + 0.016041645780205727, + 0.011100195348262787, + 0.010620640590786934, + 0.0184492040425539, + 0.007891126908361912, + 0.00812977459281683, + 0.008911561220884323, + 0.013631933368742466, + 0.013230694457888603, + 0.005540927872061729, + 0.007957377471029758, + 0.00751561438664794, + 0.008251287043094635, + 0.008481444790959358, + 0.01032333355396986, + 0.01919380947947502, + 0.011319972574710846, + 0.008978787809610367, + 0.00974953081458807, + 0.00750251580029726, + 0.00648981798440218, + 0.012787839397788048, + 0.008875502273440361, + 0.012425944209098816, + 0.008875717408955097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.048706669360399246 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203527010977268 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.014203527010977268, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203527010977268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13005082309246063, + 0.07682061940431595, + 0.07430592179298401, + 0.12814590334892273, + 0.04556223750114441, + 0.25944146513938904, + 0.0647381842136383, + 0.05426238849759102, + 0.08953459560871124, + 0.07713783532381058 + ], + "credit_map": [ + 0.014423692598938942, + 0.019190127030014992, + 0.015811361372470856, + 0.014785747975111008, + 0.01987919583916664, + 0.011533539742231369, + 0.012615730054676533, + 0.021811434999108315, + 0.005852276924997568, + 0.0069686416536569595, + 0.00836374331265688, + 0.006011310499161482, + 0.009397909045219421, + 0.008252975530922413, + 0.01603778637945652, + 0.015935981646180153, + 0.0051226443611085415, + 0.0047738216817379, + 0.009830217808485031, + 0.005745796952396631, + 0.006420517805963755, + 0.013097251765429974, + 0.022027837112545967, + 0.007287836167961359, + 0.011934968642890453, + 0.012603522278368473, + 0.021082140505313873, + 0.011825988069176674, + 0.015666605904698372, + 0.015453114174306393, + 0.015680765733122826, + 0.02389879710972309, + 0.0065484605729579926, + 0.005775244440883398, + 0.004656653851270676, + 0.003563319332897663, + 0.004207737278193235, + 0.005123142618685961, + 0.009337326511740685, + 0.006350353360176086, + 0.02778971567749977, + 0.011426592245697975, + 0.018762512132525444, + 0.01140549499541521, + 0.015438172034919262, + 0.015047961845993996, + 0.09166539460420609, + 0.06790561228990555, + 0.00686826603487134, + 0.00674515450373292, + 0.009516154415905476, + 0.005437609739601612, + 0.008502095006406307, + 0.006598420906811953, + 0.010085925459861755, + 0.010984565131366253, + 0.0052475291304290295, + 0.006001237779855728, + 0.006931130774319172, + 0.0048264614306390285, + 0.010295930318534374, + 0.006121255457401276, + 0.007862759754061699, + 0.006976080592721701, + 0.004041312262415886, + 0.0036995646078139544, + 0.005156373605132103, + 0.004223398864269257, + 0.0061011845245957375, + 0.006621906068176031, + 0.031121009960770607, + 0.02856985107064247, + 0.00638387119397521, + 0.0055830031633377075, + 0.00537867983803153, + 0.00427962327376008, + 0.010790199041366577, + 0.014203116297721863, + 0.017614971846342087, + 0.012904366478323936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.014203527010977268 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539591640233994 + ], + "opposite_logprob_tokens": [ + -4.387539386749268 + ], + "logprob_reward": -0.012539591640233994, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539591640233994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18783272802829742, + 0.06922170519828796, + 0.15807373821735382, + 0.0933360606431961, + 0.14793279767036438, + 0.0618465282022953, + 0.08619450032711029, + 0.038408152759075165, + 0.05036628246307373, + 0.10678746551275253 + ], + "credit_map": [ + 0.010949096642434597, + 0.0195314958691597, + 0.01870541274547577, + 0.0172590184956789, + 0.014964363537728786, + 0.01802314817905426, + 0.031494077295064926, + 0.05690611153841019, + 0.005132550373673439, + 0.0073633985593914986, + 0.014373572543263435, + 0.0075476826168596745, + 0.007524892222136259, + 0.01074968371540308, + 0.009096046909689903, + 0.0074338726699352264, + 0.008240075781941414, + 0.006206259597092867, + 0.008596659637987614, + 0.014598422683775425, + 0.007875684648752213, + 0.030125899240374565, + 0.06322376430034637, + 0.01920696720480919, + 0.007922775112092495, + 0.00834308099001646, + 0.01629697158932686, + 0.00973806157708168, + 0.009187673218548298, + 0.009317081421613693, + 0.016087636351585388, + 0.01644277386367321, + 0.013264721259474754, + 0.013934820890426636, + 0.022405611351132393, + 0.014375980943441391, + 0.01637575402855873, + 0.017611287534236908, + 0.027547374367713928, + 0.0224172193557024, + 0.007479190360754728, + 0.005148411728441715, + 0.004927383735775948, + 0.003850469598546624, + 0.008975828066468239, + 0.0063387020491063595, + 0.015834709629416466, + 0.009291832335293293, + 0.012498144991695881, + 0.007928787730634212, + 0.011860871687531471, + 0.007911640219390392, + 0.011782120913267136, + 0.008543038740754128, + 0.012632042169570923, + 0.013037855736911297, + 0.0026303473860025406, + 0.0038133186753839254, + 0.00630750460550189, + 0.0035889751743525267, + 0.005257079377770424, + 0.003922237548977137, + 0.006924936547875404, + 0.0059637511149048805, + 0.0022789454087615013, + 0.00261684344150126, + 0.003495124401524663, + 0.003543825587257743, + 0.003569835564121604, + 0.003373610321432352, + 0.017012204974889755, + 0.014475890435278416, + 0.0073116496205329895, + 0.008507446385920048, + 0.007764291018247604, + 0.00689212279394269, + 0.016317186877131462, + 0.018615156412124634, + 0.024390365928411484, + 0.016989244148135185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.012539591640233994 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31345877051353455 + ], + "opposite_logprob_tokens": [ + -1.313458800315857 + ], + "logprob_reward": -0.31345877051353455, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31345877051353455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13425715267658234, + 0.04556354880332947, + 0.32040008902549744, + 0.06335634738206863, + 0.1405831128358841, + 0.08595985174179077, + 0.03301345184445381, + 0.021694650873541832, + 0.020545601844787598, + 0.1346261352300644 + ], + "credit_map": [ + 0.009027935564517975, + 0.016900809481739998, + 0.015505030751228333, + 0.010878372006118298, + 0.011916382238268852, + 0.010040621273219585, + 0.024136513471603394, + 0.03585148602724075, + 0.003381777321919799, + 0.004411046858876944, + 0.004018935840576887, + 0.004105201456695795, + 0.004523865878582001, + 0.009680762887001038, + 0.010938632301986217, + 0.004503326956182718, + 0.017184481024742126, + 0.012756947427988052, + 0.01290757954120636, + 0.030365053564310074, + 0.0153766805306077, + 0.060788460075855255, + 0.1309313327074051, + 0.04008955880999565, + 0.004445559345185757, + 0.005161616951227188, + 0.00840589590370655, + 0.0049903555773198605, + 0.005989424884319305, + 0.00702484417706728, + 0.013699306175112724, + 0.01363934576511383, + 0.015255254693329334, + 0.016210461035370827, + 0.020468007773160934, + 0.015787845477461815, + 0.012503379955887794, + 0.012662914581596851, + 0.02454918436706066, + 0.023146063089370728, + 0.00454404391348362, + 0.005528935696929693, + 0.004241274204105139, + 0.004431618377566338, + 0.007624882273375988, + 0.005088341422379017, + 0.028416119515895844, + 0.026084640994668007, + 0.0030743733514100313, + 0.0036265410017222166, + 0.0038484232500195503, + 0.0028560867067426443, + 0.0031273874919861555, + 0.004485440440475941, + 0.005654218140989542, + 0.006340981926769018, + 0.0026678224094212055, + 0.0031485753133893013, + 0.003145156428217888, + 0.002476033288985491, + 0.003095848485827446, + 0.001641349052079022, + 0.002573738107457757, + 0.0029461283702403307, + 0.0014744846848770976, + 0.0014832386514171958, + 0.002687924774363637, + 0.0015299267834052444, + 0.0019364800537005067, + 0.001846985542215407, + 0.0052939485758543015, + 0.004292611498385668, + 0.006634514778852463, + 0.009508434683084488, + 0.005761911626905203, + 0.007136842701584101, + 0.017512917518615723, + 0.022915339097380638, + 0.03473934531211853, + 0.030416835099458694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 9 + ], + "selected_reward": -0.31345877051353455 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle drone glides slowly and steadily forward, tracing the languid path of a slow-flowing river as it weaves through a dense, old-growth forest. The camera maintains a steady, dreamlike pace above the treetops, capturing the entire canopy, forming a near-continuous ceiling above the water. The interconnected branches and lush foliage are vibrant shades of golden yellow and rich ochre. Below, the river's dark, calm surface acts like a dynamic mirror; it captures the fiery foliage, but the gentle, almost imperceptible current distorts the reflection into soft, twisted, and painterly images of the trees. A thick, ethereal mist slowly billows over the water, its thin tendrils curling and clinging to the surface, creating a profound sense of tranquility and mystery. The entire scene is bathed in a soft, diffuse light, as if filtering through the mist itself, casting a warm golden glow that enhances the mysterious and serene atmosphere.", + "seed": 15425, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_6/index_3905_seed_15425_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.515028545819223, + "vlm_logprob": -0.515028545819223, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.515028545819223, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.361076831817627 + ], + "opposite_logprob_tokens": [ + -0.11107682436704636 + ], + "logprob_reward": -2.361076831817627, + "yes_no_margin_reward": -2.25, + "token_credit_reward": -2.361076831817627, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19703152775764465, + 0.4014403223991394, + 0.032485924661159515, + 0.05337094888091087, + 0.05346377193927765, + 0.022445276379585266, + 0.0744943916797638, + 0.06462803483009338, + 0.0567772313952446, + 0.04386265575885773 + ], + "credit_map": [ + 0.026675699278712273, + 0.015520280227065086, + 0.03388366475701332, + 0.01624302752315998, + 0.02715856023132801, + 0.034193675965070724, + 0.0161422211676836, + 0.02721436880528927, + 0.02281040884554386, + 0.013151805847883224, + 0.14407473802566528, + 0.01355145126581192, + 0.024148350581526756, + 0.032445576041936874, + 0.018571753054857254, + 0.13268616795539856, + 0.0027485638856887817, + 0.0035311386454850435, + 0.004408450331538916, + 0.003585048019886017, + 0.0036509153433144093, + 0.004130685701966286, + 0.004371928051114082, + 0.006059185601770878, + 0.005932307336479425, + 0.005035156849771738, + 0.008741115219891071, + 0.0044564600102603436, + 0.006888724863529205, + 0.009208505041897297, + 0.005753536242991686, + 0.007355134934186935, + 0.0038038187194615602, + 0.00407191039994359, + 0.006102876737713814, + 0.0036209949757903814, + 0.012021651491522789, + 0.006303598638623953, + 0.005944500211626291, + 0.011594410054385662, + 0.0030396252404898405, + 0.0025356558617204428, + 0.00294694397598505, + 0.0022854136768728495, + 0.0031750083435326815, + 0.003205162938684225, + 0.002572757424786687, + 0.0026847044937312603, + 0.0054872953332960606, + 0.00656835176050663, + 0.015196707099676132, + 0.0060890368185937405, + 0.011715378612279892, + 0.011012783274054527, + 0.007068675942718983, + 0.011356151662766933, + 0.006509082391858101, + 0.00541438814252615, + 0.011582108214497566, + 0.0073961797170341015, + 0.008282355964183807, + 0.010065861977636814, + 0.007198978681117296, + 0.008179075084626675, + 0.006217154674232006, + 0.0051513779908418655, + 0.008984661661088467, + 0.0036938146222382784, + 0.010390589013695717, + 0.009205392561852932, + 0.0066045597195625305, + 0.006529665552079678, + 0.004915131721645594, + 0.004211873281747103, + 0.00639786571264267, + 0.0038156097289174795, + 0.007119407877326012, + 0.006772349588572979, + 0.005491358693689108, + 0.005139052867889404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -2.361076831817627 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033789265900850296 + ], + "opposite_logprob_tokens": [ + -3.4087891578674316 + ], + "logprob_reward": -0.033789265900850296, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033789265900850296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21587640047073364, + 0.4158739149570465, + 0.020096611231565475, + 0.07939062267541885, + 0.08954459428787231, + 0.02786717377603054, + 0.05886170640587807, + 0.03515816107392311, + 0.02749861590564251, + 0.029832229018211365 + ], + "credit_map": [ + 0.02787405252456665, + 0.023485615849494934, + 0.03927966579794884, + 0.017264019697904587, + 0.03475108742713928, + 0.030426397919654846, + 0.015443108044564724, + 0.027352429926395416, + 0.023730622604489326, + 0.012165888212621212, + 0.15246962010860443, + 0.013060953468084335, + 0.022710783407092094, + 0.032553303986787796, + 0.018534120172262192, + 0.14064857363700867, + 0.001447051065042615, + 0.0025347715709358454, + 0.0027573800180107355, + 0.0015117032453417778, + 0.0034055360592901707, + 0.0026874965988099575, + 0.0028062311466783285, + 0.0029464398976415396, + 0.009258600883185863, + 0.006298823747783899, + 0.012524951249361038, + 0.006422955077141523, + 0.013966252095997334, + 0.01446780189871788, + 0.0077614812180399895, + 0.008689754642546177, + 0.005790481809526682, + 0.007880592718720436, + 0.007898683659732342, + 0.0030012703500688076, + 0.02648070640861988, + 0.009332946501672268, + 0.009493219666182995, + 0.019666682928800583, + 0.003378173103556037, + 0.002995173679664731, + 0.0034332640934735537, + 0.004067922942340374, + 0.003908208571374416, + 0.003747540293261409, + 0.003314646426588297, + 0.0030222453642636538, + 0.0041041201911866665, + 0.004166611470282078, + 0.012416742742061615, + 0.0043939221650362015, + 0.006698657292872667, + 0.007370544131845236, + 0.006947982590645552, + 0.012763123027980328, + 0.0031121468637138605, + 0.003305987222120166, + 0.005402812734246254, + 0.004556972999125719, + 0.0046797809191048145, + 0.005212739575654268, + 0.0043595354072749615, + 0.004528186284005642, + 0.002560970839112997, + 0.0022141088265925646, + 0.00489457044750452, + 0.002632608637213707, + 0.004010309930890799, + 0.004217519890516996, + 0.0034262151457369328, + 0.0035423110239207745, + 0.0032130705658346415, + 0.0029610865749418736, + 0.00368568766862154, + 0.002604805864393711, + 0.004610419739037752, + 0.005804256536066532, + 0.0035536494106054306, + 0.0033992507960647345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.033789265900850296 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029828058555722237 + ], + "opposite_logprob_tokens": [ + -3.5298280715942383 + ], + "logprob_reward": -0.029828058555722237, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029828058555722237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3372192680835724, + 0.14688168466091156, + 0.09628669172525406, + 0.06466137617826462, + 0.03497381508350372, + 0.07404524832963943, + 0.08507440239191055, + 0.04360165446996689, + 0.07579807937145233, + 0.04145779460668564 + ], + "credit_map": [ + 0.02739369124174118, + 0.043352413922548294, + 0.037580106407403946, + 0.016289392486214638, + 0.08010611683130264, + 0.06724436581134796, + 0.03641626983880997, + 0.02883690409362316, + 0.009734313935041428, + 0.007818038575351238, + 0.03812168911099434, + 0.004495281260460615, + 0.022311830893158913, + 0.020855627954006195, + 0.008912104181945324, + 0.034632790833711624, + 0.008773431181907654, + 0.010109099559485912, + 0.014456496573984623, + 0.004754321184009314, + 0.015981169417500496, + 0.015877561643719673, + 0.009237482212483883, + 0.017097124829888344, + 0.005866244435310364, + 0.009154285304248333, + 0.006253221537917852, + 0.0048108953051269054, + 0.014558232389390469, + 0.010759428143501282, + 0.0067122397013008595, + 0.006546829827129841, + 0.003300959710031748, + 0.0030750096775591373, + 0.004905020352452993, + 0.0024816179648041725, + 0.005704719107598066, + 0.0059554665349423885, + 0.00343940407037735, + 0.006111616268754005, + 0.006865819916129112, + 0.011210943572223186, + 0.007774787954986095, + 0.005185422953218222, + 0.014815672300755978, + 0.007941127754747868, + 0.00930043589323759, + 0.010951036587357521, + 0.005743688903748989, + 0.010050452314317226, + 0.014899902045726776, + 0.005225446540862322, + 0.022336775436997414, + 0.011243765242397785, + 0.007704058662056923, + 0.00787031278014183, + 0.004300032742321491, + 0.0043154554441571236, + 0.006678932812064886, + 0.004091349896043539, + 0.007414762396365404, + 0.005708782002329826, + 0.005315796937793493, + 0.00577654130756855, + 0.007326953113079071, + 0.007411051541566849, + 0.012228299863636494, + 0.0044282772578299046, + 0.017180584371089935, + 0.009713966399431229, + 0.00843834038823843, + 0.009070605970919132, + 0.004582257010042667, + 0.003271718742325902, + 0.006684159394353628, + 0.002872990909963846, + 0.00844191387295723, + 0.005490528419613838, + 0.0049716634675860405, + 0.005142563953995705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.029828058555722237 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007633444853127003 + ], + "opposite_logprob_tokens": [ + -4.882633209228516 + ], + "logprob_reward": -0.007633444853127003, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007633444853127003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19812004268169403, + 0.4198666214942932, + 0.04986252263188362, + 0.0456608347594738, + 0.049583129584789276, + 0.026809152215719223, + 0.07043977081775665, + 0.0607767216861248, + 0.055482055991888046, + 0.023399164900183678 + ], + "credit_map": [ + 0.01907297410070896, + 0.01920471526682377, + 0.02722114510834217, + 0.012746092863380909, + 0.037434715777635574, + 0.03883889690041542, + 0.019233286380767822, + 0.02436821535229683, + 0.02712652087211609, + 0.03026663139462471, + 0.11450568586587906, + 0.01260401215404272, + 0.05280286446213722, + 0.05635372921824455, + 0.019982190802693367, + 0.10622497648000717, + 0.004510913975536823, + 0.0061939978040754795, + 0.006037478800863028, + 0.003261665115132928, + 0.00831782165914774, + 0.009185371920466423, + 0.00517522357404232, + 0.0071800472214818, + 0.004918730817735195, + 0.004652561619877815, + 0.0054343645460903645, + 0.004022257402539253, + 0.007977391593158245, + 0.007651593070477247, + 0.005053831730037928, + 0.005950100254267454, + 0.003289529588073492, + 0.004808689933270216, + 0.004643875639885664, + 0.0032779662869870663, + 0.013168854638934135, + 0.005449321586638689, + 0.0052571683190763, + 0.009687723591923714, + 0.003305362071841955, + 0.0025297696702182293, + 0.004306687507778406, + 0.0020436744671314955, + 0.004030529409646988, + 0.0035353568382561207, + 0.003206030698493123, + 0.003851741086691618, + 0.004616902209818363, + 0.005855670664459467, + 0.008585638366639614, + 0.006404119078069925, + 0.008421897888183594, + 0.012180390767753124, + 0.01069712545722723, + 0.013678031973540783, + 0.006273140665143728, + 0.00512360455468297, + 0.01035830844193697, + 0.006321718916296959, + 0.008360720239579678, + 0.008721034042537212, + 0.007627828978002071, + 0.007990366779267788, + 0.006426443345844746, + 0.005389578174799681, + 0.006956454832106829, + 0.004268844611942768, + 0.011113052256405354, + 0.008249708451330662, + 0.00676390016451478, + 0.006314069963991642, + 0.0024818540550768375, + 0.0022814329713582993, + 0.003076237626373768, + 0.0016888411482796073, + 0.00469739455729723, + 0.0037559643387794495, + 0.002670897403731942, + 0.0027465412858873606 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.007633444853127003 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14281512796878815 + ], + "opposite_logprob_tokens": [ + -2.017815113067627 + ], + "logprob_reward": -0.14281512796878815, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14281512796878815, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21009287238121033, + 0.4271332025527954, + 0.021222878247499466, + 0.027255358174443245, + 0.010862328112125397, + 0.016272956505417824, + 0.16760481894016266, + 0.02615462616086006, + 0.07850054651498795, + 0.014900369569659233 + ], + "credit_map": [ + 0.022014256566762924, + 0.027977226302027702, + 0.028885720297694206, + 0.01648808643221855, + 0.030677782371640205, + 0.034085262566804886, + 0.016553275287151337, + 0.03341127187013626, + 0.030294401571154594, + 0.041544508188962936, + 0.09815902262926102, + 0.015271197073161602, + 0.0649208277463913, + 0.06325341016054153, + 0.02171812392771244, + 0.09197171777486801, + 0.0018774777418002486, + 0.0024790766183286905, + 0.0026518101803958416, + 0.0016791082452982664, + 0.004229888319969177, + 0.003429073141887784, + 0.0020667552016675472, + 0.0028096900787204504, + 0.0029273375403136015, + 0.0023791377898305655, + 0.004184344317764044, + 0.00236510974355042, + 0.005410055164247751, + 0.004282982554286718, + 0.002712898887693882, + 0.0029934949707239866, + 0.0008508975151926279, + 0.0010900940978899598, + 0.001171209616586566, + 0.0011703280033543706, + 0.0016420463798567653, + 0.0026397344190627337, + 0.0011065512662753463, + 0.0011914661154150963, + 0.0019395757699385285, + 0.0018882727017626166, + 0.003076371271163225, + 0.0017758014146238565, + 0.002021193504333496, + 0.0020627318881452084, + 0.0016940312925726175, + 0.0018149781972169876, + 0.009795614518225193, + 0.011993247084319592, + 0.03161054849624634, + 0.010391335934400558, + 0.027450045570731163, + 0.01872127689421177, + 0.019097771495580673, + 0.03854496777057648, + 0.0031087666284292936, + 0.0029412119183689356, + 0.0038248798809945583, + 0.0026955888606607914, + 0.003817155957221985, + 0.0033503156155347824, + 0.0031097913160920143, + 0.0033069185446947813, + 0.008514132350683212, + 0.007736925967037678, + 0.0122588025406003, + 0.005441402550786734, + 0.015388180501759052, + 0.011269252747297287, + 0.009050192311406136, + 0.00884165521711111, + 0.0014452800387516618, + 0.00167220423463732, + 0.0026045546401292086, + 0.0011194454273208976, + 0.00246434286236763, + 0.0019265145529061556, + 0.0018673959420993924, + 0.0018006322206929326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.14281512796878815 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle drone glides slowly and steadily forward, tracing the languid path of a slow-flowing river as it weaves through a dense, old-growth forest. The camera maintains a steady, dreamlike pace above the treetops, capturing the entire canopy, forming a near-continuous ceiling above the water. The interconnected branches and lush foliage are vibrant shades of golden yellow and rich ochre. Below, the river's dark, calm surface acts like a dynamic mirror; it captures the fiery foliage, but the gentle, almost imperceptible current distorts the reflection into soft, twisted, and painterly images of the trees. A thick, ethereal mist slowly billows over the water, its thin tendrils curling and clinging to the surface, creating a profound sense of tranquility and mystery. The entire scene is bathed in a soft, diffuse light, as if filtering through the mist itself, casting a warm golden glow that enhances the mysterious and serene atmosphere.", + "seed": 15426, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_6/index_3905_seed_15426_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.36889676470309496, + "vlm_logprob": -0.36889676470309496, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.36889676470309496, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.5142199993133545 + ], + "opposite_logprob_tokens": [ + -0.2642199695110321 + ], + "logprob_reward": -1.5142199993133545, + "yes_no_margin_reward": -1.25, + "token_credit_reward": -1.5142199993133545, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.7598493695259094, + 0.03188113123178482, + 0.03872545436024666, + 0.02295532450079918, + 0.041489072144031525, + 0.02540730871260166, + 0.019571898505091667, + 0.015829604119062424, + 0.016790347173810005, + 0.027500495314598083 + ], + "credit_map": [ + 0.11348192393779755, + 0.0844566598534584, + 0.06885433197021484, + 0.07643922418355942, + 0.12609519064426422, + 0.0839780867099762, + 0.1335304230451584, + 0.0730135515332222, + 0.004326005466282368, + 0.002912201453000307, + 0.004868037533015013, + 0.0031057295855134726, + 0.004606522154062986, + 0.004766132682561874, + 0.0038701980374753475, + 0.0034263036213815212, + 0.0035093713086098433, + 0.004345445893704891, + 0.002916027093306184, + 0.002451893175020814, + 0.007299595046788454, + 0.006386027671396732, + 0.005051286891102791, + 0.00676580797880888, + 0.0033672149293124676, + 0.0029923198744654655, + 0.0022003906778991222, + 0.002137540141120553, + 0.004384199157357216, + 0.0031646129209548235, + 0.002392509486526251, + 0.0023165387101471424, + 0.005887753330171108, + 0.004092419054359198, + 0.005734541453421116, + 0.003655343083664775, + 0.004378917161375284, + 0.007004792336374521, + 0.0050390008836984634, + 0.005696307402104139, + 0.0019891285337507725, + 0.0024316946510225534, + 0.0018380586989223957, + 0.0014512959169223905, + 0.001680668443441391, + 0.007353844586759806, + 0.0019365723710507154, + 0.0067260474897921085, + 0.001695820945315063, + 0.0027738322969526052, + 0.0017502863192930818, + 0.003171265823766589, + 0.002560637891292572, + 0.0028155441395938396, + 0.0025745839811861515, + 0.002229927573353052, + 0.0017188614001497626, + 0.0015505942283198237, + 0.0010512132430449128, + 0.0015146227087825537, + 0.0020795604214072227, + 0.003101842012256384, + 0.001755357370711863, + 0.0030575513374060392, + 0.002164715901017189, + 0.0021809141617268324, + 0.0019033023854717612, + 0.001687122625298798, + 0.001905452343635261, + 0.0024297423660755157, + 0.0021532645914703608, + 0.002365831984207034, + 0.0030314933974295855, + 0.0042367419227957726, + 0.002632079878821969, + 0.0024126155767589808, + 0.0032143075950443745, + 0.004440564662218094, + 0.00398118793964386, + 0.0035515036433935165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -1.5142199993133545 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018220899626612663 + ], + "opposite_logprob_tokens": [ + -4.018220901489258 + ], + "logprob_reward": -0.018220899626612663, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018220899626612663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7462766766548157, + 0.02447550930082798, + 0.038858868181705475, + 0.012709420174360275, + 0.018826594576239586, + 0.03715991973876953, + 0.021123552694916725, + 0.031528789550065994, + 0.04300948604941368, + 0.026031265035271645 + ], + "credit_map": [ + 0.11268924921751022, + 0.083414725959301, + 0.06695939600467682, + 0.0728917196393013, + 0.12493827193975449, + 0.08226905763149261, + 0.1319526880979538, + 0.07116147130727768, + 0.002646049717441201, + 0.0029943780973553658, + 0.002914348617196083, + 0.002623870735988021, + 0.003268448868766427, + 0.003661796450614929, + 0.004022613167762756, + 0.002344002714380622, + 0.003337735077366233, + 0.003912434913218021, + 0.001996910898014903, + 0.0017167062032967806, + 0.006885848473757505, + 0.008998999372124672, + 0.0033489209599792957, + 0.008661307394504547, + 0.0015887211775407195, + 0.0015259168576449156, + 0.0014451098395511508, + 0.0011821662774309516, + 0.002174359280616045, + 0.0015494388062506914, + 0.0019256832310929894, + 0.0013180231908336282, + 0.0027002226561307907, + 0.0021610271651297808, + 0.002811311511322856, + 0.0018630464328452945, + 0.0023349574767053127, + 0.0023553278297185898, + 0.002309519099071622, + 0.0022911792621016502, + 0.002287347801029682, + 0.0018740813247859478, + 0.0015246879775077105, + 0.0012877972330898046, + 0.0027509566862136126, + 0.013608035631477833, + 0.001863196725025773, + 0.011963815428316593, + 0.002008974552154541, + 0.0021150752436369658, + 0.0023568517062813044, + 0.002905741101130843, + 0.0025636162608861923, + 0.0029380936175584793, + 0.003436689730733633, + 0.002798507921397686, + 0.0029114994686096907, + 0.0036560099106281996, + 0.003012206871062517, + 0.0055288407020270824, + 0.004050861578434706, + 0.005248144268989563, + 0.0034126529935747385, + 0.003708571894094348, + 0.006682590581476688, + 0.005535513628274202, + 0.003610400715842843, + 0.007660817354917526, + 0.0039300438947975636, + 0.00464291637763381, + 0.00566685339435935, + 0.0052803452126681805, + 0.004234693944454193, + 0.004088574554771185, + 0.0036145851481705904, + 0.0029514585621654987, + 0.0019137479830533266, + 0.0028358132112771273, + 0.003156977938488126, + 0.0032354106660932302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.018220899626612663 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0125464191660285 + ], + "opposite_logprob_tokens": [ + -4.387546539306641 + ], + "logprob_reward": -0.0125464191660285, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.0125464191660285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.690885066986084, + 0.048781272023916245, + 0.06600242853164673, + 0.015358633361756802, + 0.030413681641221046, + 0.025284551084041595, + 0.024884169921278954, + 0.023121826350688934, + 0.03730623796582222, + 0.03796211630105972 + ], + "credit_map": [ + 0.10673341900110245, + 0.07726111263036728, + 0.06148843094706535, + 0.061524197459220886, + 0.11471173912286758, + 0.07603766024112701, + 0.12470421195030212, + 0.06842432171106339, + 0.006448816973716021, + 0.004685765597969294, + 0.0033258823677897453, + 0.0030552505049854517, + 0.007283088285475969, + 0.010490906424820423, + 0.005352930631488562, + 0.008138634264469147, + 0.0062176864594221115, + 0.00659105833619833, + 0.003541938029229641, + 0.003924185875803232, + 0.013662618584930897, + 0.008416177704930305, + 0.011469248682260513, + 0.012179522775113583, + 0.002131296321749687, + 0.0019325679168105125, + 0.0013301344588398933, + 0.0014066232834011316, + 0.0027820770628750324, + 0.0020950399339199066, + 0.0021321314852684736, + 0.0015487645287066698, + 0.0037643355317413807, + 0.0049001434817910194, + 0.0028175117913633585, + 0.002752449130639434, + 0.005485930014401674, + 0.004466675221920013, + 0.003698056098073721, + 0.0025285810697823763, + 0.0017554659862071276, + 0.0035294261761009693, + 0.0017395556205883622, + 0.001090328092686832, + 0.0037763332948088646, + 0.006069278344511986, + 0.002122309524565935, + 0.005201856140047312, + 0.0018346459837630391, + 0.0021656870376318693, + 0.0015303191030398011, + 0.0019512976286932826, + 0.0025486338417977095, + 0.0069255284033715725, + 0.002144064288586378, + 0.005783994682133198, + 0.002920206869021058, + 0.0022294740192592144, + 0.002510890131816268, + 0.002769030397757888, + 0.0035269137006253004, + 0.0027878147084265947, + 0.002444799989461899, + 0.003932698164135218, + 0.005259186495095491, + 0.003298217197880149, + 0.0034775149542838335, + 0.00590906385332346, + 0.004168463405221701, + 0.004302138928323984, + 0.00561606464907527, + 0.005275593139231205, + 0.004939052276313305, + 0.004682619124650955, + 0.002951259957626462, + 0.0035148770548403263, + 0.004944948013871908, + 0.008234302513301373, + 0.004373624455183744, + 0.004321433138102293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0125464191660285 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818496361374855 + ], + "opposite_logprob_tokens": [ + -4.018185138702393 + ], + "logprob_reward": -0.01818496361374855, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818496361374855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39553213119506836, + 0.18548820912837982, + 0.15813462436199188, + 0.03516874089837074, + 0.04030696675181389, + 0.03354262933135033, + 0.0316983200609684, + 0.04762658104300499, + 0.022593021392822266, + 0.0499088279902935 + ], + "credit_map": [ + 0.05955227464437485, + 0.0428573414683342, + 0.03400442376732826, + 0.04112855717539787, + 0.06870681047439575, + 0.044754523783922195, + 0.06897684931755066, + 0.03555132448673248, + 0.012989914044737816, + 0.013344143517315388, + 0.007451422978192568, + 0.012515373528003693, + 0.019953813403844833, + 0.060832757502794266, + 0.018217116594314575, + 0.04018363356590271, + 0.015592260286211967, + 0.018793746829032898, + 0.008562230505049229, + 0.00962752290070057, + 0.03742806240916252, + 0.03057466447353363, + 0.013439071364700794, + 0.024117035791277885, + 0.004784128163009882, + 0.004187759477645159, + 0.0053596505895257, + 0.004397081211209297, + 0.004781918600201607, + 0.004133233334869146, + 0.004299965221434832, + 0.003224998712539673, + 0.005395814776420593, + 0.0052675483748316765, + 0.005425115115940571, + 0.004787089768797159, + 0.005899760872125626, + 0.0053331623785197735, + 0.0043996465392410755, + 0.0037988258991390467, + 0.002281657885760069, + 0.004197917878627777, + 0.004716099239885807, + 0.004978102166205645, + 0.0034889215603470802, + 0.0059060025960206985, + 0.0037690834142267704, + 0.004204839468002319, + 0.0023774232249706984, + 0.003952283412218094, + 0.0018099788576364517, + 0.002696535550057888, + 0.0033977075945585966, + 0.007549860514700413, + 0.0026903743855655193, + 0.007224154658615589, + 0.0042922599241137505, + 0.004202771931886673, + 0.005832142196595669, + 0.006967253051698208, + 0.005773398093879223, + 0.0068842628970742226, + 0.005622647237032652, + 0.008051843382418156, + 0.0030362929683178663, + 0.0037485354114323854, + 0.002342002000659704, + 0.003224712098017335, + 0.0024071538355201483, + 0.0029978586826473475, + 0.002575267804786563, + 0.002261195797473192, + 0.007216234691441059, + 0.005716667510569096, + 0.005974982399493456, + 0.006030970718711615, + 0.005375936161726713, + 0.006829436402767897, + 0.006991920527070761, + 0.005772674921900034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01818496361374855 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813115417957306 + ], + "opposite_logprob_tokens": [ + -1.4063115119934082 + ], + "logprob_reward": -0.2813115417957306, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813115417957306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7589915990829468, + 0.03743142634630203, + 0.04205523058772087, + 0.00780488783493638, + 0.02188143879175186, + 0.011171267367899418, + 0.031985487788915634, + 0.019140655174851418, + 0.03613061085343361, + 0.033407505601644516 + ], + "credit_map": [ + 0.1150648221373558, + 0.08467680215835571, + 0.06591209769248962, + 0.07200966030359268, + 0.12640145421028137, + 0.08395775407552719, + 0.1375718116760254, + 0.07339707762002945, + 0.0043007987551391125, + 0.003378152148798108, + 0.003811245784163475, + 0.0033404557034373283, + 0.0034637695644050837, + 0.008653748780488968, + 0.0044947341084480286, + 0.0059885126538574696, + 0.004105365369468927, + 0.004132428206503391, + 0.0037170329596847296, + 0.003145623719319701, + 0.006733840797096491, + 0.006066980771720409, + 0.006424264516681433, + 0.007729686331003904, + 0.0009297299548052251, + 0.0010980658698827028, + 0.0006826328462921083, + 0.0007726228795945644, + 0.0011888990411534905, + 0.001089712604880333, + 0.0013980477815493941, + 0.0006451748195104301, + 0.0033338803332298994, + 0.0022145332768559456, + 0.0031782889273017645, + 0.002609775634482503, + 0.002846511546522379, + 0.0028531798161566257, + 0.00272050267085433, + 0.002124762861058116, + 0.0010054158046841621, + 0.0014222111785784364, + 0.0008478782256133854, + 0.001195781398564577, + 0.0015461768489331007, + 0.0020293721463531256, + 0.0010028501274064183, + 0.002121579833328724, + 0.0033873531501740217, + 0.004733314737677574, + 0.0031704208813607693, + 0.004667052533477545, + 0.0031425869092345238, + 0.004136431030929089, + 0.004627171438187361, + 0.004121152684092522, + 0.0019134017638862133, + 0.0016083666123449802, + 0.001459947437979281, + 0.0020981074776500463, + 0.0030622591730207205, + 0.002956787357106805, + 0.002322508255019784, + 0.003719272557646036, + 0.005315894261002541, + 0.005397693254053593, + 0.0038262037560343742, + 0.004912701901048422, + 0.004038085695356131, + 0.00495984498411417, + 0.0038108096923679113, + 0.0038693693932145834, + 0.005091732367873192, + 0.0049813673831522465, + 0.003923638723790646, + 0.00389212928712368, + 0.003337568836286664, + 0.0031704821158200502, + 0.004767803940922022, + 0.004242775961756706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.2813115417957306 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle drone glides slowly and steadily forward, tracing the languid path of a slow-flowing river as it weaves through a dense, old-growth forest. The camera maintains a steady, dreamlike pace above the treetops, capturing the entire canopy, forming a near-continuous ceiling above the water. The interconnected branches and lush foliage are vibrant shades of golden yellow and rich ochre. Below, the river's dark, calm surface acts like a dynamic mirror; it captures the fiery foliage, but the gentle, almost imperceptible current distorts the reflection into soft, twisted, and painterly images of the trees. A thick, ethereal mist slowly billows over the water, its thin tendrils curling and clinging to the surface, creating a profound sense of tranquility and mystery. The entire scene is bathed in a soft, diffuse light, as if filtering through the mist itself, casting a warm golden glow that enhances the mysterious and serene atmosphere.", + "seed": 15427, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_6/index_3905_seed_15427_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.4995186261832714, + "vlm_logprob": -0.4995186261832714, + "vlm_yes_no_margin": 1.925, + "vlm_token_credit": -0.4995186261832714, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.1339187622070312 + ], + "opposite_logprob_tokens": [ + -0.1339186578989029 + ], + "logprob_reward": -2.1339187622070312, + "yes_no_margin_reward": -2.0, + "token_credit_reward": -2.1339187622070312, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2571266293525696, + 0.15680666267871857, + 0.08519384264945984, + 0.06797106564044952, + 0.13575580716133118, + 0.036904476583004, + 0.020710015669465065, + 0.04704979807138443, + 0.033053524792194366, + 0.15942814946174622 + ], + "credit_map": [ + 0.029831407591700554, + 0.029033470898866653, + 0.0303502157330513, + 0.019642388448119164, + 0.03805389255285263, + 0.027899382635951042, + 0.03583899512887001, + 0.04647686332464218, + 0.01549317967146635, + 0.011503995396196842, + 0.022532470524311066, + 0.011577642522752285, + 0.025337880477309227, + 0.021853191778063774, + 0.030760696157813072, + 0.017747605219483376, + 0.009552628733217716, + 0.007402942515909672, + 0.009428762830793858, + 0.006464137230068445, + 0.010537443682551384, + 0.019233450293540955, + 0.009320522658526897, + 0.013253956101834774, + 0.006526658311486244, + 0.007058247458189726, + 0.009203040972352028, + 0.006122102495282888, + 0.009641841053962708, + 0.008785434067249298, + 0.008935722522437572, + 0.011698015034198761, + 0.011097614653408527, + 0.014534627087414265, + 0.010528516955673695, + 0.009222379885613918, + 0.024777252227067947, + 0.022402845323085785, + 0.01747749373316765, + 0.02571507915854454, + 0.00335996993817389, + 0.0031253760680556297, + 0.0021050830837339163, + 0.002175925299525261, + 0.0043191369622945786, + 0.009377080015838146, + 0.004249505698680878, + 0.008192398585379124, + 0.0018287578132003546, + 0.0021010844502598047, + 0.0018679210916161537, + 0.0016956540057435632, + 0.002401145873591304, + 0.003629434620961547, + 0.0031903539784252644, + 0.003995663486421108, + 0.004296023864299059, + 0.005911652930080891, + 0.0046021477319300175, + 0.003189491108059883, + 0.006490980740636587, + 0.008836701512336731, + 0.006220592185854912, + 0.007502205669879913, + 0.002458660863339901, + 0.0033140245359390974, + 0.002601266372948885, + 0.0036113872192800045, + 0.004755288828164339, + 0.006430889014154673, + 0.0043594143353402615, + 0.0055225929245352745, + 0.007826991379261017, + 0.0068697212263941765, + 0.010656604543328285, + 0.011842796579003334, + 0.011017064563930035, + 0.05400910973548889, + 0.010299572721123695, + 0.04690629988908768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -2.1339187622070312 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05502459034323692 + ], + "opposite_logprob_tokens": [ + -2.9300246238708496 + ], + "logprob_reward": -0.05502459034323692, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05502459034323692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20531317591667175, + 0.14484143257141113, + 0.045787423849105835, + 0.05792190879583359, + 0.0716874897480011, + 0.0633717030286789, + 0.059721652418375015, + 0.06758025288581848, + 0.0794353038072586, + 0.2043396383523941 + ], + "credit_map": [ + 0.022080060094594955, + 0.025214921683073044, + 0.022757265716791153, + 0.01732301525771618, + 0.02920590713620186, + 0.0232810340821743, + 0.030198117718100548, + 0.035252854228019714, + 0.015425552614033222, + 0.016102582216262817, + 0.018885347992181778, + 0.011182444170117378, + 0.03134139999747276, + 0.013738302513957024, + 0.023964546620845795, + 0.014201273210346699, + 0.005408960860222578, + 0.004152967594563961, + 0.004520220682024956, + 0.003165643895044923, + 0.005408450961112976, + 0.009425190277397633, + 0.006397181656211615, + 0.007308808621019125, + 0.005909949075430632, + 0.006940715480595827, + 0.0075210644863545895, + 0.004744474310427904, + 0.008773655630648136, + 0.007135753519833088, + 0.007496932987123728, + 0.009399366565048695, + 0.007440832443535328, + 0.009032030589878559, + 0.005251879338175058, + 0.005387418903410435, + 0.013020826503634453, + 0.010198641568422318, + 0.009236238896846771, + 0.01211961917579174, + 0.00506294472143054, + 0.005334281828254461, + 0.005470719188451767, + 0.004876346327364445, + 0.006998961325734854, + 0.014796324074268341, + 0.008355197496712208, + 0.012476926669478416, + 0.006001039408147335, + 0.006386935245245695, + 0.008321619592607021, + 0.006300110369920731, + 0.008156915195286274, + 0.010818731039762497, + 0.006089045666158199, + 0.007647254969924688, + 0.005818458274006844, + 0.006639436818659306, + 0.0037832066882401705, + 0.0040014805272221565, + 0.011915973387658596, + 0.015685614198446274, + 0.006686892360448837, + 0.013049195520579815, + 0.00712984474375844, + 0.009880773723125458, + 0.007908911444246769, + 0.007758268155157566, + 0.012486142106354237, + 0.013043766841292381, + 0.00918498169630766, + 0.012042619287967682, + 0.011613053269684315, + 0.013286679983139038, + 0.016378141939640045, + 0.016930103302001953, + 0.019159885123372078, + 0.058255475014448166, + 0.017614565789699554, + 0.05110174044966698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.05502459034323692 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033749788999557 + ], + "opposite_logprob_tokens": [ + -2.350337505340576 + ], + "logprob_reward": -0.10033749788999557, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033749788999557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2211901992559433, + 0.21221573650836945, + 0.11144720017910004, + 0.07649312913417816, + 0.0946740061044693, + 0.07800791412591934, + 0.07408571988344193, + 0.07917415350675583, + 0.021233050152659416, + 0.03147892281413078 + ], + "credit_map": [ + 0.019589055329561234, + 0.030007410794496536, + 0.011507612653076649, + 0.01360558532178402, + 0.04166235774755478, + 0.032174158841371536, + 0.030126551166176796, + 0.04251744598150253, + 0.016550671309232712, + 0.015525350347161293, + 0.016930507495999336, + 0.011810156516730785, + 0.028086038306355476, + 0.056723084300756454, + 0.01931639015674591, + 0.04727351292967796, + 0.012096354737877846, + 0.011400546878576279, + 0.006790103390812874, + 0.00727567495778203, + 0.026014424860477448, + 0.021377837285399437, + 0.013561424799263477, + 0.012930814176797867, + 0.006827270612120628, + 0.006274120882153511, + 0.007933403365314007, + 0.007847943343222141, + 0.01286966260522604, + 0.014638721942901611, + 0.009637460112571716, + 0.010464543476700783, + 0.006259065121412277, + 0.008111288771033287, + 0.01065747532993555, + 0.009447330608963966, + 0.012959667481482029, + 0.018264608457684517, + 0.011519179679453373, + 0.017455382272601128, + 0.005837191827595234, + 0.005501697771251202, + 0.004739102441817522, + 0.004552076570689678, + 0.007911976426839828, + 0.02088475041091442, + 0.007206997834146023, + 0.021374117583036423, + 0.007015845738351345, + 0.007891306653618813, + 0.009017876349389553, + 0.0051762196235358715, + 0.01287833135575056, + 0.009894434362649918, + 0.009316646493971348, + 0.012895051389932632, + 0.0069089122116565704, + 0.006577090825885534, + 0.005583507940173149, + 0.005329047329723835, + 0.014067686162889004, + 0.014371514320373535, + 0.009512221440672874, + 0.01682417094707489, + 0.0015862233703956008, + 0.0029592057690024376, + 0.0018657295731827617, + 0.0022225426509976387, + 0.003702846122905612, + 0.0029466066043823957, + 0.003337407484650612, + 0.0026124869473278522, + 0.0026973376516252756, + 0.003214339492842555, + 0.0028035594150424004, + 0.003926650155335665, + 0.005154109094291925, + 0.0049656894989311695, + 0.004515252076089382, + 0.004201983567327261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.10033749788999557 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739504635334015 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739504635334015, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739504635334015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2393706738948822, + 0.17937779426574707, + 0.10024435818195343, + 0.03854319453239441, + 0.10461734980344772, + 0.05410078540444374, + 0.07070489972829819, + 0.057991836220026016, + 0.06041208654642105, + 0.0946369618177414 + ], + "credit_map": [ + 0.02721850387752056, + 0.026322176679968834, + 0.019592221826314926, + 0.01809108816087246, + 0.031364668160676956, + 0.03226103633642197, + 0.03803475573658943, + 0.04648624733090401, + 0.010659979656338692, + 0.013462521135807037, + 0.012681462801992893, + 0.008707850240170956, + 0.020200345665216446, + 0.0513794869184494, + 0.01666862890124321, + 0.04561752825975418, + 0.008015790954232216, + 0.005457415245473385, + 0.005327309016138315, + 0.004714952316135168, + 0.009708374738693237, + 0.03338949754834175, + 0.008085221983492374, + 0.02554580569267273, + 0.0035964304115623236, + 0.003988126292824745, + 0.0042574661783874035, + 0.002836340107023716, + 0.00588019285351038, + 0.005940595641732216, + 0.005496062804013491, + 0.006547986529767513, + 0.006805110722780228, + 0.010256659239530563, + 0.007432491984218359, + 0.00639850739389658, + 0.016054976731538773, + 0.022439902648329735, + 0.01287835743278265, + 0.022351356223225594, + 0.004655718337744474, + 0.004115595016628504, + 0.0028256294317543507, + 0.0023758832830935717, + 0.007232624106109142, + 0.01468366477638483, + 0.005629641003906727, + 0.012582040391862392, + 0.005838403943926096, + 0.007423536852002144, + 0.006678449921309948, + 0.005323687568306923, + 0.01279023289680481, + 0.013018961064517498, + 0.009643941186368465, + 0.009987691417336464, + 0.005827576387673616, + 0.005800328217446804, + 0.004248844925314188, + 0.0037364340387284756, + 0.012434707023203373, + 0.009150811471045017, + 0.00569873908534646, + 0.011094402521848679, + 0.00576022220775485, + 0.006291158962994814, + 0.008065863512456417, + 0.009107613936066628, + 0.008445058017969131, + 0.007053818088024855, + 0.007817864418029785, + 0.007870495319366455, + 0.004830331541597843, + 0.006741971243172884, + 0.0089739253744483, + 0.006235460750758648, + 0.010409520007669926, + 0.026499098166823387, + 0.0065644667483866215, + 0.024382203817367554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.006739504635334015 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a high angle, looking down from above the treetops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera move slowly and steadily forward, tracing the path of a river?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the trees and foliage depicted in vibrant shades of golden yellow and rich ochre?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the river surface dark and calm, acting as a mirror that reflects the foliage in soft, twisted, and painterly images?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a thick, ethereal mist visible billowing over the water and clinging to the surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157277584075928 + ], + "opposite_logprob_tokens": [ + -1.7015727758407593 + ], + "logprob_reward": -0.20157277584075928, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157277584075928, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23809270560741425, + 0.22035402059555054, + 0.04918958619236946, + 0.03405275568366051, + 0.038438767194747925, + 0.05572708696126938, + 0.09301723539829254, + 0.1359616070985794, + 0.053591322153806686, + 0.08157483488321304 + ], + "credit_map": [ + 0.029554542154073715, + 0.035619404166936874, + 0.01263982243835926, + 0.013282299973070621, + 0.0348694771528244, + 0.03607858717441559, + 0.029071463271975517, + 0.04697713628411293, + 0.01340297982096672, + 0.006964862812310457, + 0.012036334723234177, + 0.010119927115738392, + 0.01659034751355648, + 0.07755649834871292, + 0.01705452799797058, + 0.06662856787443161, + 0.0033630188554525375, + 0.0024271828588098288, + 0.0037193854805082083, + 0.0031560109928250313, + 0.005767864175140858, + 0.014674435369670391, + 0.004377864766865969, + 0.011703833937644958, + 0.00285072298720479, + 0.0033001615665853024, + 0.0028860277961939573, + 0.0023221182636916637, + 0.006592108402401209, + 0.005711335223168135, + 0.004870670847594738, + 0.005519610829651356, + 0.004883964546024799, + 0.004716932773590088, + 0.003381389193236828, + 0.0027995468117296696, + 0.006257350090891123, + 0.006296172272413969, + 0.00455741910263896, + 0.0055459956638514996, + 0.004495910368859768, + 0.0060767726972699165, + 0.0029340472538024187, + 0.00298036215826869, + 0.009845811873674393, + 0.012220577336847782, + 0.0070877415128052235, + 0.01008586771786213, + 0.007390438113361597, + 0.00988700333982706, + 0.009131280705332756, + 0.0069026462733745575, + 0.01591091975569725, + 0.018179504200816154, + 0.012851505540311337, + 0.01276394072920084, + 0.010584631003439426, + 0.012951771728694439, + 0.008551114238798618, + 0.007895620539784431, + 0.025227217003703117, + 0.02844730019569397, + 0.01773850806057453, + 0.024565458297729492, + 0.003539077937602997, + 0.005236722994595766, + 0.0037627583369612694, + 0.006015755236148834, + 0.006857067812234163, + 0.011854469776153564, + 0.00703660910949111, + 0.009288864210247993, + 0.004637797363102436, + 0.004947332199662924, + 0.00639074994251132, + 0.007186955772340298, + 0.0065056090243160725, + 0.02390407584607601, + 0.007131097372621298, + 0.02087121643126011 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.20157277584075928 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_7/index_tensor([6734, 6734])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_7/index_tensor([6734, 6734])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..49191ef5705c2db25c5794cb8043fc984ee67f9d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_7/index_tensor([6734, 6734])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight head-tracking shot maintains constant focus on the nose of a paper airplane meticulously folded from a dreamy train ticket. The ticket, a warm cream with a pearlescent sheen, bears faint golden text in an elegant, artistic font, the creases of the paper crisp and soft. Slowly and gracefully, the paper airplane ascends into a gentle, fairytale world, its nose firmly pointed skyward, its body swaying gently in an invisible breeze, its gesture both innocent and endearing. The camera follows it through vast, fluffy, cotton-candy-like clouds tinted in pastel pinks, lavenders, and baby blues, their wingtips stirring up wisps of shimmering fluff as they pass. The entire scene is bathed in a dreamy, soft golden light, which penetrates the clouds and casts iridescent shadows on the paper airplane's surface, creating an atmosphere of tranquility, warmth, and whimsy. The overall image evokes the rich, subdued, and delicate style of a fairytale illustration.", + "seed": 5080, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_7/index_6734_seed_5080_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04961116770282388, + "vlm_logprob": -0.04961116770282388, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.04961116770282388, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18423020839691162 + ], + "opposite_logprob_tokens": [ + -1.8092302083969116 + ], + "logprob_reward": -0.18423020839691162, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18423020839691162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11352778971195221, + 0.06443657726049423, + 0.059884317219257355, + 0.25409695506095886, + 0.05198904871940613, + 0.09948976337909698, + 0.03217754885554314, + 0.09560343623161316, + 0.046965114772319794, + 0.18182942271232605 + ], + "credit_map": [ + 0.017132282257080078, + 0.012504849582910538, + 0.01956919953227043, + 0.011969163082540035, + 0.013132978230714798, + 0.017175640910863876, + 0.011676381342113018, + 0.01036730408668518, + 0.0053263939917087555, + 0.006435831543058157, + 0.010610355995595455, + 0.011146052740514278, + 0.00846372451633215, + 0.00981196016073227, + 0.007013135123997927, + 0.005629122722893953, + 0.00428487453609705, + 0.006024191156029701, + 0.015964150428771973, + 0.007237747777253389, + 0.006013508420437574, + 0.00726690236479044, + 0.0051467446610331535, + 0.007946200668811798, + 0.022527867928147316, + 0.018507305532693863, + 0.03296838328242302, + 0.08044889569282532, + 0.01706823520362377, + 0.026605261489748955, + 0.015843143686652184, + 0.040127869695425034, + 0.006954612676054239, + 0.005694468040019274, + 0.006808992009609938, + 0.00715054152533412, + 0.00680104736238718, + 0.006468583829700947, + 0.005600928328931332, + 0.006509878672659397, + 0.008002574555575848, + 0.009865080006420612, + 0.014194515533745289, + 0.010272852145135403, + 0.017935728654265404, + 0.01372239738702774, + 0.009294852614402771, + 0.016201768070459366, + 0.0031198507640510798, + 0.0025786347687244415, + 0.0055567543022334576, + 0.002888760529458523, + 0.006110199727118015, + 0.004876736085861921, + 0.003760139923542738, + 0.0032864734530448914, + 0.006169430911540985, + 0.006924958433955908, + 0.025432603433728218, + 0.00937245786190033, + 0.009562559425830841, + 0.011006223037838936, + 0.008464478887617588, + 0.01867072656750679, + 0.004285330884158611, + 0.004901508800685406, + 0.006686241831630468, + 0.009523192420601845, + 0.0052367704920470715, + 0.005132746882736683, + 0.006388721521943808, + 0.0048106033354997635, + 0.04872596263885498, + 0.010906542651355267, + 0.028589894995093346, + 0.017303436994552612, + 0.009764578193426132, + 0.021266981959342957, + 0.011453798972070217, + 0.033818215131759644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.18423020839691162 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746372207999229 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746372207999229, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746372207999229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09208722412586212, + 0.05171751230955124, + 0.06147252768278122, + 0.09404273331165314, + 0.07278410345315933, + 0.1751253604888916, + 0.03281491994857788, + 0.208292618393898, + 0.056997060775756836, + 0.15466593205928802 + ], + "credit_map": [ + 0.012606755830347538, + 0.01663626916706562, + 0.013830166310071945, + 0.014273365959525108, + 0.008674578741192818, + 0.010415015742182732, + 0.008208118379116058, + 0.0074429563246667385, + 0.004223610274493694, + 0.006474461406469345, + 0.00734098581597209, + 0.011487087234854698, + 0.0038235748652368784, + 0.004943733103573322, + 0.006471819244325161, + 0.006952239666134119, + 0.005198681261390448, + 0.006173703353852034, + 0.014368735253810883, + 0.007171601057052612, + 0.0075505999848246574, + 0.0063470047898590565, + 0.00718288030475378, + 0.007479323074221611, + 0.008165257051587105, + 0.006560346111655235, + 0.01047539059072733, + 0.027834340929985046, + 0.009488471783697605, + 0.011139093898236752, + 0.007628024555742741, + 0.012751813977956772, + 0.009061525575816631, + 0.008107313886284828, + 0.011382653377950191, + 0.00912931002676487, + 0.008559685200452805, + 0.00932510569691658, + 0.008852898143231869, + 0.008365616202354431, + 0.015502025373280048, + 0.014666985720396042, + 0.018143625929951668, + 0.018783608451485634, + 0.034341078251600266, + 0.02422931417822838, + 0.014729187823832035, + 0.03472953289747238, + 0.00410670368000865, + 0.0033319429494440556, + 0.003864968428388238, + 0.00303105846978724, + 0.006781770382076502, + 0.00430451612919569, + 0.003954105079174042, + 0.003439853200688958, + 0.018487663939595222, + 0.021244637668132782, + 0.023801838979125023, + 0.03170099854469299, + 0.037771616131067276, + 0.028247006237506866, + 0.028959862887859344, + 0.018078994005918503, + 0.007561043836176395, + 0.005734524689614773, + 0.007391391322016716, + 0.010234375484287739, + 0.006441995035856962, + 0.006318613886833191, + 0.00822922307997942, + 0.005085892044007778, + 0.042965423315763474, + 0.009268024936318398, + 0.024148549884557724, + 0.014384515583515167, + 0.00857715867459774, + 0.01847662404179573, + 0.00860342662781477, + 0.0282421987503767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.006746372207999229 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734886672347784 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.006734886672347784, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734886672347784, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09621526300907135, + 0.11782177537679672, + 0.039360061287879944, + 0.08531452715396881, + 0.06400375813245773, + 0.098946712911129, + 0.06254333257675171, + 0.3242114186286926, + 0.0517011322081089, + 0.059882037341594696 + ], + "credit_map": [ + 0.015467872843146324, + 0.014543021097779274, + 0.013476086780428886, + 0.01893695630133152, + 0.005988669581711292, + 0.012405499815940857, + 0.007651188876479864, + 0.00774597330018878, + 0.01222089771181345, + 0.014328311197459698, + 0.01767238788306713, + 0.013949091546237469, + 0.013328338973224163, + 0.015996728092432022, + 0.012917673215270042, + 0.01740834303200245, + 0.0036065764725208282, + 0.0041306959465146065, + 0.005212575662881136, + 0.005964957643300295, + 0.006458047777414322, + 0.005640159361064434, + 0.00491831311956048, + 0.003428736934438348, + 0.008967800997197628, + 0.008205990307033062, + 0.007740712724626064, + 0.01788659393787384, + 0.01151058916002512, + 0.00833842158317566, + 0.010299978777766228, + 0.012364440597593784, + 0.005941292736679316, + 0.006528270430862904, + 0.008351223543286324, + 0.013291330076754093, + 0.008562247268855572, + 0.008440888486802578, + 0.006172731518745422, + 0.006715770345181227, + 0.005857265088707209, + 0.006018895655870438, + 0.025301646441221237, + 0.009683528915047646, + 0.006670958828181028, + 0.008693124167621136, + 0.007053867448121309, + 0.029667431488633156, + 0.006976080127060413, + 0.008533569052815437, + 0.0077497828751802444, + 0.005813411436975002, + 0.011730477213859558, + 0.007512176875025034, + 0.007584829814732075, + 0.0066430047154426575, + 0.018137987703084946, + 0.013298900797963142, + 0.11584139615297318, + 0.022592052817344666, + 0.019346915185451508, + 0.03407774493098259, + 0.018039733171463013, + 0.08287667483091354, + 0.00498554203659296, + 0.005550076719373465, + 0.0073899924755096436, + 0.007224068511277437, + 0.006339402869343758, + 0.008106330409646034, + 0.0074126580730080605, + 0.004693059250712395, + 0.016161417588591576, + 0.003083426970988512, + 0.010316822677850723, + 0.005709914490580559, + 0.002877922961488366, + 0.006604780443012714, + 0.0036892856005579233, + 0.011438469402492046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.006734886672347784 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017101438716053963 + ], + "opposite_logprob_tokens": [ + -6.376709938049316 + ], + "logprob_reward": -0.0017101438716053963, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017101438716053963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1615619659423828, + 0.07684038579463959, + 0.08308816701173782, + 0.1633770763874054, + 0.046016987413167953, + 0.0737357959151268, + 0.03870493546128273, + 0.19976450502872467, + 0.07841185480356216, + 0.07849840819835663 + ], + "credit_map": [ + 0.019351070746779442, + 0.028494173660874367, + 0.023358847945928574, + 0.03116101212799549, + 0.012980282306671143, + 0.014496385119855404, + 0.01443004235625267, + 0.017290156334638596, + 0.0075675370171666145, + 0.010166558437049389, + 0.010117081925272942, + 0.016927523538470268, + 0.005015539471060038, + 0.0069170487113296986, + 0.007797627709805965, + 0.012331471778452396, + 0.00938099343329668, + 0.009630312211811543, + 0.012171032838523388, + 0.01639537699520588, + 0.00831865519285202, + 0.009445157833397388, + 0.00864951591938734, + 0.009097115136682987, + 0.015692882239818573, + 0.014354225248098373, + 0.019200896844267845, + 0.04393770173192024, + 0.014000325463712215, + 0.022843368351459503, + 0.014049004763364792, + 0.019298654049634933, + 0.005538635421544313, + 0.005165549460798502, + 0.006420745048671961, + 0.006000894587486982, + 0.005918113049119711, + 0.0064466726034879684, + 0.004446667153388262, + 0.006079711019992828, + 0.008792637847363949, + 0.007285676896572113, + 0.011311220936477184, + 0.007295798975974321, + 0.011690103448927402, + 0.011952103115618229, + 0.005417608190327883, + 0.009990646503865719, + 0.004670958500355482, + 0.003238204400986433, + 0.005186319816857576, + 0.006171971093863249, + 0.004626094363629818, + 0.004324009642004967, + 0.00426420196890831, + 0.006223173812031746, + 0.010564982891082764, + 0.005631328094750643, + 0.07850836217403412, + 0.011676357127726078, + 0.010398521088063717, + 0.020751601085066795, + 0.008343425579369068, + 0.05388990789651871, + 0.012421118095517159, + 0.006426357198506594, + 0.012956393882632256, + 0.009402626194059849, + 0.005899295210838318, + 0.009495750069618225, + 0.011128226295113564, + 0.010682085528969765, + 0.014896005392074585, + 0.005502337124198675, + 0.010944319888949394, + 0.009705755859613419, + 0.004989842884242535, + 0.009822533465921879, + 0.005896814167499542, + 0.01674080081284046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.0017101438716053963 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048634227365255356 + ], + "opposite_logprob_tokens": [ + -3.0486342906951904 + ], + "logprob_reward": -0.048634227365255356, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048634227365255356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06218968704342842, + 0.06763593852519989, + 0.06620223075151443, + 0.06492019444704056, + 0.09351854026317596, + 0.13648468255996704, + 0.07555349916219711, + 0.25316059589385986, + 0.0502031184732914, + 0.13013151288032532 + ], + "credit_map": [ + 0.006435532588511705, + 0.009211098775267601, + 0.010477115400135517, + 0.008784289471805096, + 0.006147974170744419, + 0.008740954101085663, + 0.006865288596600294, + 0.005527437664568424, + 0.00783555582165718, + 0.00778111582621932, + 0.010810314677655697, + 0.010035526007413864, + 0.005292087327688932, + 0.00962503906339407, + 0.00699452543631196, + 0.009261771105229855, + 0.009446359239518642, + 0.00730921933427453, + 0.013328604400157928, + 0.008748956955969334, + 0.006021108478307724, + 0.00845583900809288, + 0.005123699549585581, + 0.0077684419229626656, + 0.009467237628996372, + 0.005693460814654827, + 0.007581592071801424, + 0.011382264085114002, + 0.007608563639223576, + 0.009137887507677078, + 0.007024131715297699, + 0.007025058381259441, + 0.0092809172347188, + 0.008729808032512665, + 0.014888460747897625, + 0.012838630937039852, + 0.012906976975500584, + 0.015022218227386475, + 0.010319795459508896, + 0.009531733579933643, + 0.0165709275752306, + 0.00955522432923317, + 0.02255731076002121, + 0.020611794665455818, + 0.01024814136326313, + 0.012268436141312122, + 0.010148636996746063, + 0.03452421724796295, + 0.008002905175089836, + 0.00973538588732481, + 0.008064936846494675, + 0.008572855964303017, + 0.014214491471648216, + 0.009506452828645706, + 0.006837671156972647, + 0.010618800297379494, + 0.01570662297308445, + 0.014360303059220314, + 0.07445266842842102, + 0.022488556802272797, + 0.029199765995144844, + 0.028008578345179558, + 0.01840551756322384, + 0.05053859204053879, + 0.006531701888889074, + 0.004678915720432997, + 0.005857966840267181, + 0.007402487564831972, + 0.0031120670028030872, + 0.00591276166960597, + 0.0062502906657755375, + 0.010456927120685577, + 0.028720242902636528, + 0.009130341000854969, + 0.02018330991268158, + 0.014682503417134285, + 0.007975424639880657, + 0.01579645276069641, + 0.010225163772702217, + 0.023418087512254715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.048634227365255356 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight head-tracking shot maintains constant focus on the nose of a paper airplane meticulously folded from a dreamy train ticket. The ticket, a warm cream with a pearlescent sheen, bears faint golden text in an elegant, artistic font, the creases of the paper crisp and soft. Slowly and gracefully, the paper airplane ascends into a gentle, fairytale world, its nose firmly pointed skyward, its body swaying gently in an invisible breeze, its gesture both innocent and endearing. The camera follows it through vast, fluffy, cotton-candy-like clouds tinted in pastel pinks, lavenders, and baby blues, their wingtips stirring up wisps of shimmering fluff as they pass. The entire scene is bathed in a dreamy, soft golden light, which penetrates the clouds and casts iridescent shadows on the paper airplane's surface, creating an atmosphere of tranquility, warmth, and whimsy. The overall image evokes the rich, subdued, and delicate style of a fairytale illustration.", + "seed": 5081, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_7/index_6734_seed_5081_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1809728526743129, + "vlm_logprob": -0.1809728526743129, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.1809728526743129, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8316856026649475 + ], + "opposite_logprob_tokens": [ + -0.5816856026649475 + ], + "logprob_reward": -0.8316856026649475, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8316856026649475, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10891371220350266, + 0.0487896166741848, + 0.04094667732715607, + 0.0728139653801918, + 0.06474031507968903, + 0.026263343170285225, + 0.0799177885055542, + 0.08246253430843353, + 0.05702994018793106, + 0.41812214255332947 + ], + "credit_map": [ + 0.01387107651680708, + 0.01928868517279625, + 0.013505568727850914, + 0.01098174974322319, + 0.020971784368157387, + 0.009863336570560932, + 0.009770810604095459, + 0.010660708881914616, + 0.0054346900433301926, + 0.005845316685736179, + 0.007530533242970705, + 0.0067618899047374725, + 0.006388968322426081, + 0.005706624127924442, + 0.007144148927181959, + 0.003977443091571331, + 0.004757644142955542, + 0.005048520863056183, + 0.007457751780748367, + 0.005194587633013725, + 0.003813204588368535, + 0.004497939255088568, + 0.004471145570278168, + 0.005705880466848612, + 0.007539104670286179, + 0.00830148160457611, + 0.009785382077097893, + 0.009327208623290062, + 0.010677335783839226, + 0.0065942201763391495, + 0.012019085697829723, + 0.008570149540901184, + 0.005761405918747187, + 0.007557332515716553, + 0.008137824013829231, + 0.009135149419307709, + 0.010808052495121956, + 0.005164381582289934, + 0.01013119239360094, + 0.008044980466365814, + 0.002743473509326577, + 0.004655782599002123, + 0.0034855667036026716, + 0.0038059710059314966, + 0.0024970700033009052, + 0.00293533387593925, + 0.0028760137502104044, + 0.003264131024479866, + 0.008383788168430328, + 0.010230539366602898, + 0.011620176024734974, + 0.021177344024181366, + 0.0055210040882229805, + 0.009669698774814606, + 0.00559195363894105, + 0.007723277900367975, + 0.0048458389937877655, + 0.00391674367710948, + 0.026593800634145737, + 0.0279658455401659, + 0.006554718129336834, + 0.0038516102358698845, + 0.0048882621340453625, + 0.0038457110058516264, + 0.005263607949018478, + 0.0063688247464597225, + 0.012360159307718277, + 0.012509074993431568, + 0.005895466543734074, + 0.006043123546987772, + 0.004040034953504801, + 0.004549655131995678, + 0.03929051011800766, + 0.12171005457639694, + 0.13598185777664185, + 0.05803152918815613, + 0.015738187357783318, + 0.026794753968715668, + 0.009749096818268299, + 0.010826153680682182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.8316856026649475 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086649261415005 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.011086649261415005, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086649261415005, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08778364956378937, + 0.03582900017499924, + 0.06457015872001648, + 0.028282346203923225, + 0.0469307042658329, + 0.04125170782208443, + 0.052064619958400726, + 0.34811925888061523, + 0.03168914094567299, + 0.26347944140434265 + ], + "credit_map": [ + 0.010191097855567932, + 0.009703556075692177, + 0.009372765198349953, + 0.009632067754864693, + 0.009562486782670021, + 0.012611182406544685, + 0.010185427032411098, + 0.01652507297694683, + 0.004355014767497778, + 0.004088378977030516, + 0.004683410748839378, + 0.005359635688364506, + 0.003430627752095461, + 0.0031380904838442802, + 0.005612493958324194, + 0.005161347799003124, + 0.007521952036768198, + 0.010450352914631367, + 0.013557554222643375, + 0.007413898129016161, + 0.00621965155005455, + 0.006617847364395857, + 0.005804805550724268, + 0.006984103471040726, + 0.0038909984286874533, + 0.0035155401565134525, + 0.0035826480016112328, + 0.004466218873858452, + 0.0026801740750670433, + 0.0030856700614094734, + 0.0027283858507871628, + 0.004332713782787323, + 0.0035786498337984085, + 0.005032042507082224, + 0.005341493524610996, + 0.008162740617990494, + 0.007195238023996353, + 0.005750191397964954, + 0.006946663837879896, + 0.004923688247799873, + 0.004242922179400921, + 0.007294607348740101, + 0.004853634163737297, + 0.005780717823654413, + 0.0042646159417927265, + 0.0051568844355642796, + 0.004930326249450445, + 0.004728008061647415, + 0.0038587520830333233, + 0.0059403665363788605, + 0.006922577973455191, + 0.014105385169386864, + 0.0038810798432677984, + 0.0076126595959067345, + 0.004906910937279463, + 0.004836889915168285, + 0.013264108449220657, + 0.014289911836385727, + 0.14380834996700287, + 0.13834406435489655, + 0.009234627708792686, + 0.01052163541316986, + 0.011812608689069748, + 0.006843965966254473, + 0.0032465774565935135, + 0.0030698792543262243, + 0.007403756957501173, + 0.00609210692346096, + 0.0032099932432174683, + 0.0031125012319535017, + 0.0028109566774219275, + 0.0027433752547949553, + 0.02775062434375286, + 0.07947897166013718, + 0.09576761722564697, + 0.025652745738625526, + 0.00971450936049223, + 0.014691782183945179, + 0.005078345537185669, + 0.005344861652702093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.011086649261415005 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535235844552517 + ], + "opposite_logprob_tokens": [ + -4.387535095214844 + ], + "logprob_reward": -0.012535235844552517, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535235844552517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10143312066793442, + 0.030211463570594788, + 0.04268832504749298, + 0.09809908270835876, + 0.08059637248516083, + 0.09281822293996811, + 0.11591607332229614, + 0.27189117670059204, + 0.037271153181791306, + 0.12907502055168152 + ], + "credit_map": [ + 0.014467491768300533, + 0.016593415290117264, + 0.02040938474237919, + 0.013463442213833332, + 0.009748097509145737, + 0.010103246197104454, + 0.008302003145217896, + 0.008346029557287693, + 0.0034480562899261713, + 0.003534127026796341, + 0.0034373642411082983, + 0.004513331223279238, + 0.0033821433316916227, + 0.0036374880000948906, + 0.004777060355991125, + 0.00348189240321517, + 0.004616516642272472, + 0.00816609151661396, + 0.005776844918727875, + 0.005789813585579395, + 0.005312370136380196, + 0.004307152703404427, + 0.004668525420129299, + 0.0040510063990950584, + 0.013522467575967312, + 0.011968561448156834, + 0.016574131324887276, + 0.010777260176837444, + 0.010443729348480701, + 0.011988501064479351, + 0.01258782297372818, + 0.010236603207886219, + 0.009711414575576782, + 0.010210787877440453, + 0.007676771841943264, + 0.011707249097526073, + 0.013020932674407959, + 0.006133520044386387, + 0.014682477340102196, + 0.007453212980180979, + 0.009845550172030926, + 0.01777370274066925, + 0.010282971896231174, + 0.012248491868376732, + 0.010676577687263489, + 0.0108083700761199, + 0.01221991702914238, + 0.008962633088231087, + 0.009986389428377151, + 0.009388471953570843, + 0.018837174400687218, + 0.02054755948483944, + 0.01517201866954565, + 0.010192548856139183, + 0.018173759803175926, + 0.013618147931993008, + 0.013782103545963764, + 0.014206595718860626, + 0.09007079154253006, + 0.08728992193937302, + 0.015113665722310543, + 0.01459189411252737, + 0.016346728429198265, + 0.020489467307925224, + 0.002907508984208107, + 0.0036684214137494564, + 0.011402181349694729, + 0.010275701992213726, + 0.0025323641020804644, + 0.0023708140943199396, + 0.002463143551722169, + 0.0016510139685124159, + 0.011837037280201912, + 0.03183625265955925, + 0.03917624056339264, + 0.018650919198989868, + 0.00681789917871356, + 0.007142093498259783, + 0.006307594478130341, + 0.007306977640837431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.012535235844552517 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009188005933538079 + ], + "opposite_logprob_tokens": [ + -7.000918865203857 + ], + "logprob_reward": -0.0009188005933538079, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009188005933538079, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09752018749713898, + 0.052754368633031845, + 0.05363180488348007, + 0.04113264009356499, + 0.05900375172495842, + 0.05491313338279724, + 0.09907351434230804, + 0.10157855600118637, + 0.04474469646811485, + 0.3956473469734192 + ], + "credit_map": [ + 0.017257319763302803, + 0.018593359738588333, + 0.01330587174743414, + 0.01243614125996828, + 0.009999743662774563, + 0.009710690937936306, + 0.007422940339893103, + 0.008794118650257587, + 0.006018724758177996, + 0.006190716288983822, + 0.006563330534845591, + 0.007364635355770588, + 0.006290021818131208, + 0.006886668968945742, + 0.0075331637635827065, + 0.005907104350626469, + 0.0068497988395392895, + 0.006914051249623299, + 0.009294693358242512, + 0.006753518246114254, + 0.004829831887036562, + 0.0059100911021232605, + 0.005228459369391203, + 0.00785135943442583, + 0.005028826650232077, + 0.004663316998630762, + 0.00595800532028079, + 0.0052994610741734505, + 0.004768025130033493, + 0.003973451443016529, + 0.005667621735483408, + 0.005773928016424179, + 0.005085492506623268, + 0.006876897998154163, + 0.006654419004917145, + 0.010650020092725754, + 0.008200658485293388, + 0.007419290486723185, + 0.008273515850305557, + 0.005843456368893385, + 0.004411621484905481, + 0.006722224876284599, + 0.006485606078058481, + 0.008447804488241673, + 0.005974734202027321, + 0.008963721804320812, + 0.005417739972472191, + 0.008489677682518959, + 0.010212878696620464, + 0.01140386052429676, + 0.015106693841516972, + 0.02651164121925831, + 0.007428844925016165, + 0.011305681429803371, + 0.008003129623830318, + 0.009100784547626972, + 0.005937862675637007, + 0.009823018684983253, + 0.02405833825469017, + 0.02584194764494896, + 0.008402138017117977, + 0.011289739981293678, + 0.007865386083722115, + 0.00836013164371252, + 0.005173635669052601, + 0.005599459633231163, + 0.008203385397791862, + 0.00879740435630083, + 0.004676115233451128, + 0.0051842196844518185, + 0.003686569631099701, + 0.003423907095566392, + 0.038775306195020676, + 0.12058217823505402, + 0.13857270777225494, + 0.0414687879383564, + 0.014549620449543, + 0.02207978628575802, + 0.009202100336551666, + 0.010416865348815918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.0009188005933538079 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863797500729561 + ], + "opposite_logprob_tokens": [ + -3.048637866973877 + ], + "logprob_reward": -0.04863797500729561, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863797500729561, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048364702612161636, + 0.04107240214943886, + 0.1795274317264557, + 0.07590722292661667, + 0.12920519709587097, + 0.01743186078965664, + 0.041163258254528046, + 0.11405327171087265, + 0.048516638576984406, + 0.30475807189941406 + ], + "credit_map": [ + 0.009636390954256058, + 0.006144270766526461, + 0.005931937601417303, + 0.003767416113987565, + 0.005517988465726376, + 0.00460669444873929, + 0.004698580596596003, + 0.008061426691710949, + 0.007246578112244606, + 0.004482529126107693, + 0.004959719255566597, + 0.0050676981918513775, + 0.004255914594978094, + 0.005048063583672047, + 0.0043344576843082905, + 0.005677440203726292, + 0.018957044929265976, + 0.040388695895671844, + 0.03283468261361122, + 0.019207624718546867, + 0.01691954955458641, + 0.021690431982278824, + 0.013584007509052753, + 0.015945393592119217, + 0.01000512856990099, + 0.0081716850399971, + 0.011302362196147442, + 0.009495053440332413, + 0.010081818327307701, + 0.007636478170752525, + 0.012005004100501537, + 0.007209690287709236, + 0.01215219683945179, + 0.015045082196593285, + 0.015019879676401615, + 0.025028089061379433, + 0.01276050228625536, + 0.01910521276295185, + 0.015238996595144272, + 0.014855241402983665, + 0.0026073134504258633, + 0.002521122107282281, + 0.002221564296633005, + 0.0025648849550634623, + 0.001775967888534069, + 0.0020459331572055817, + 0.0018845211016014218, + 0.001810555113479495, + 0.0032838471233844757, + 0.00394049659371376, + 0.0060920617543160915, + 0.011189528740942478, + 0.0034066694788634777, + 0.005213906057178974, + 0.004340030252933502, + 0.0036967170890420675, + 0.004402338992804289, + 0.0059656864032149315, + 0.04274710640311241, + 0.04057887941598892, + 0.0050066737458109856, + 0.0052674561738967896, + 0.004541466012597084, + 0.005543658044189215, + 0.006081965286284685, + 0.006108610890805721, + 0.008038004860281944, + 0.009568532928824425, + 0.004551401361823082, + 0.0061593009158968925, + 0.0038099379744380713, + 0.004198882728815079, + 0.029292356222867966, + 0.0888940766453743, + 0.10220244526863098, + 0.03925434499979019, + 0.01163031067699194, + 0.01644911803305149, + 0.008220369927585125, + 0.008815047331154346 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 4 + ], + "selected_reward": -0.04863797500729561 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight head-tracking shot maintains constant focus on the nose of a paper airplane meticulously folded from a dreamy train ticket. The ticket, a warm cream with a pearlescent sheen, bears faint golden text in an elegant, artistic font, the creases of the paper crisp and soft. Slowly and gracefully, the paper airplane ascends into a gentle, fairytale world, its nose firmly pointed skyward, its body swaying gently in an invisible breeze, its gesture both innocent and endearing. The camera follows it through vast, fluffy, cotton-candy-like clouds tinted in pastel pinks, lavenders, and baby blues, their wingtips stirring up wisps of shimmering fluff as they pass. The entire scene is bathed in a dreamy, soft golden light, which penetrates the clouds and casts iridescent shadows on the paper airplane's surface, creating an atmosphere of tranquility, warmth, and whimsy. The overall image evokes the rich, subdued, and delicate style of a fairytale illustration.", + "seed": 5082, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_7/index_6734_seed_5082_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0739565346040763, + "vlm_logprob": -0.0739565346040763, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.0739565346040763, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32000207901000977 + ], + "opposite_logprob_tokens": [ + -1.3200020790100098 + ], + "logprob_reward": -0.32000207901000977, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32000207901000977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06626702845096588, + 0.017050795257091522, + 0.4514480531215668, + 0.028568781912326813, + 0.0199463851749897, + 0.10383335500955582, + 0.17824415862560272, + 0.10985679179430008, + 0.013142772950232029, + 0.011641906574368477 + ], + "credit_map": [ + 0.0076996926218271255, + 0.007016998250037432, + 0.010579962283372879, + 0.01279538031667471, + 0.01026464719325304, + 0.005910135805606842, + 0.006582579109817743, + 0.005417635198682547, + 0.0030745600815862417, + 0.0021182564087212086, + 0.002135860500857234, + 0.002261389745399356, + 0.0018317059148102999, + 0.0021966206841170788, + 0.001529751461930573, + 0.0019026489462703466, + 0.03557584807276726, + 0.012578384019434452, + 0.16431798040866852, + 0.01229739747941494, + 0.008126318454742432, + 0.1795520782470703, + 0.01437006238847971, + 0.024629991501569748, + 0.0032878557685762644, + 0.001829655491746962, + 0.0050833323039114475, + 0.0027137857396155596, + 0.00221517076715827, + 0.0036377718206495047, + 0.0031171117443591356, + 0.006684097461402416, + 0.0021697909105569124, + 0.002003124449402094, + 0.0029569354373961687, + 0.0024931663647294044, + 0.0010523722739890218, + 0.004309097304940224, + 0.0020341805648058653, + 0.002927717287093401, + 0.010462433099746704, + 0.009473448619246483, + 0.018507517874240875, + 0.008434198796749115, + 0.008135379292070866, + 0.016034355387091637, + 0.009477207437157631, + 0.023308807983994484, + 0.039667416363954544, + 0.019357290118932724, + 0.0358414351940155, + 0.011797463521361351, + 0.008681762963533401, + 0.0180665273219347, + 0.011177868582308292, + 0.033654388040304184, + 0.015821168199181557, + 0.01012604683637619, + 0.01912788301706314, + 0.010424437932670116, + 0.015588807873427868, + 0.013720293529331684, + 0.007921714335680008, + 0.01712643913924694, + 0.0013118868228048086, + 0.0025454103015363216, + 0.0014136451063677669, + 0.0012710097944363952, + 0.0009856640826910734, + 0.0022286621388047934, + 0.0010668063769116998, + 0.0023196886759251356, + 0.0017310703406110406, + 0.0015867336187511683, + 0.00226660561747849, + 0.0015657268231734633, + 0.0008233188418671489, + 0.001521965954452753, + 0.000784526695497334, + 0.0013619589153677225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.32000207901000977 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006744833197444677 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.006744833197444677, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006744833197444677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07672860473394394, + 0.033799443393945694, + 0.5832334756851196, + 0.028999365866184235, + 0.024158483371138573, + 0.044537268579006195, + 0.08372735232114792, + 0.06840617209672928, + 0.01429044920951128, + 0.042119383811950684 + ], + "credit_map": [ + 0.006115640047937632, + 0.009680578485131264, + 0.009365463629364967, + 0.015588324517011642, + 0.014023332856595516, + 0.009349844418466091, + 0.008605824783444405, + 0.003999595530331135, + 0.0036892988719046116, + 0.0034149137791246176, + 0.004181198310106993, + 0.0031248722225427628, + 0.004031462129205465, + 0.006019786465913057, + 0.0033926733303815126, + 0.005945239216089249, + 0.046563342213630676, + 0.01638220064342022, + 0.21162915229797363, + 0.016090326011180878, + 0.010845652781426907, + 0.23142223060131073, + 0.018527118489146233, + 0.03177347034215927, + 0.0031521418131887913, + 0.002357758115977049, + 0.005254555959254503, + 0.0028461080510169268, + 0.0027474735397845507, + 0.004150294698774815, + 0.003070640377700329, + 0.005420391913503408, + 0.002668720902875066, + 0.0025115790776908398, + 0.00320306490175426, + 0.0031004538759589195, + 0.0023523278068751097, + 0.004116948693990707, + 0.0026404033415019512, + 0.003564986400306225, + 0.004985123407095671, + 0.0041994741186499596, + 0.009363923221826553, + 0.003839951939880848, + 0.0027955921832472086, + 0.005231529474258423, + 0.004422883037477732, + 0.009698789566755295, + 0.017953507602214813, + 0.008392253890633583, + 0.01617107167840004, + 0.005448522511869669, + 0.00377430347725749, + 0.008828415535390377, + 0.005382532719522715, + 0.017776744440197945, + 0.010029429569840431, + 0.007550722453743219, + 0.012634476646780968, + 0.00743910064920783, + 0.009001482278108597, + 0.007946550846099854, + 0.0042326925322413445, + 0.00957171805202961, + 0.0015414105728268623, + 0.002881823107600212, + 0.0013433132553473115, + 0.0014782532816752791, + 0.001308879698626697, + 0.002377377822995186, + 0.0013563025277107954, + 0.0020030890591442585, + 0.005093665327876806, + 0.007699704263359308, + 0.008226360194385052, + 0.004750752821564674, + 0.002448025392368436, + 0.005362868774682283, + 0.003194256452843547, + 0.005343753378838301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.006744833197444677 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635676465928555 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635676465928555, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635676465928555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056145407259464264, + 0.03227187320590019, + 0.20761027932167053, + 0.10818137973546982, + 0.05609400197863579, + 0.45639288425445557, + 0.01048457995057106, + 0.02157476916909218, + 0.02295328490436077, + 0.028291575610637665 + ], + "credit_map": [ + 0.00521418871358037, + 0.006531321443617344, + 0.007916666567325592, + 0.010640667751431465, + 0.007378603797405958, + 0.008623476140201092, + 0.004721496254205704, + 0.005118987523019314, + 0.0036869263276457787, + 0.003714763792231679, + 0.0044331904500722885, + 0.0049829850904643536, + 0.0028569267597049475, + 0.004100481979548931, + 0.003094032872468233, + 0.005402565933763981, + 0.016742488369345665, + 0.007214374374598265, + 0.06921207904815674, + 0.007413551677018404, + 0.006483306176960468, + 0.0738019272685051, + 0.008301597088575363, + 0.01844094693660736, + 0.014247069135308266, + 0.0064532943069934845, + 0.014337358064949512, + 0.006158389151096344, + 0.006074314936995506, + 0.017292821779847145, + 0.010037925094366074, + 0.03358021005988121, + 0.006913134828209877, + 0.0051147728227078915, + 0.008033325895667076, + 0.008252803236246109, + 0.00384504278190434, + 0.007205196190625429, + 0.007767477538436651, + 0.00896224845200777, + 0.032271858304739, + 0.03386853635311127, + 0.10159897804260254, + 0.03536314517259598, + 0.03827618435025215, + 0.05435395613312721, + 0.029050257056951523, + 0.1316099464893341, + 0.0015553038101643324, + 0.0012986965011805296, + 0.0014178452547639608, + 0.0011977649992331862, + 0.0008439000230282545, + 0.0014574941014871001, + 0.0011839749058708549, + 0.0015295998891815543, + 0.0023939579259604216, + 0.0022525291424244642, + 0.0025420121382921934, + 0.0021668565459549427, + 0.0029168156906962395, + 0.0030194041319191456, + 0.0024221595376729965, + 0.003861035220324993, + 0.001538991229608655, + 0.002487331395968795, + 0.0030059413984417915, + 0.0037750317715108395, + 0.0032225672621279955, + 0.002257001120597124, + 0.003731248201802373, + 0.0029351713601499796, + 0.0027719063218683004, + 0.00407115975394845, + 0.004472376313060522, + 0.004914202261716127, + 0.0033198243472725153, + 0.002515193773433566, + 0.0032024532556533813, + 0.003024459583684802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.008635676465928555 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007152383332140744 + ], + "opposite_logprob_tokens": [ + -7.250715255737305 + ], + "logprob_reward": -0.0007152383332140744, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007152383332140744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06011010333895683, + 0.02350817248225212, + 0.6314285397529602, + 0.028322460129857063, + 0.021542012691497803, + 0.04621303454041481, + 0.06623473018407822, + 0.06542961299419403, + 0.020207636058330536, + 0.0370037667453289 + ], + "credit_map": [ + 0.006616668775677681, + 0.009303691796958447, + 0.009179391898214817, + 0.008886108174920082, + 0.009575235657393932, + 0.0067698280327022076, + 0.0054984805174171925, + 0.004280698951333761, + 0.0030440553091466427, + 0.0027221590280532837, + 0.0031123815570026636, + 0.0028699454851448536, + 0.00192851887550205, + 0.004681749269366264, + 0.0020208959467709064, + 0.003128467360511422, + 0.04904017969965935, + 0.01732201687991619, + 0.23056304454803467, + 0.017797425389289856, + 0.010893495753407478, + 0.2529783248901367, + 0.019685449078679085, + 0.03314858302474022, + 0.002699600299820304, + 0.0017193632666021585, + 0.0051504760049283504, + 0.0030643583741039038, + 0.0021317731589078903, + 0.003712729550898075, + 0.0031811923254281282, + 0.006662967149168253, + 0.002108547370880842, + 0.0021690125577151775, + 0.0033317883498966694, + 0.002211682964116335, + 0.0014854943146929145, + 0.004087591543793678, + 0.0022624630946666002, + 0.0038854307495057583, + 0.00524155655875802, + 0.004626620095223188, + 0.008662487380206585, + 0.003658563829958439, + 0.0032092821784317493, + 0.007036497350782156, + 0.00443432480096817, + 0.009343703277409077, + 0.013786520808935165, + 0.006128030829131603, + 0.013886963948607445, + 0.004391147289425135, + 0.0030017043463885784, + 0.006976103410124779, + 0.003327444428578019, + 0.014736812561750412, + 0.009330795146524906, + 0.006503342185169458, + 0.010932602919638157, + 0.006255314219743013, + 0.00875655747950077, + 0.00854880828410387, + 0.00490226037800312, + 0.01019993145018816, + 0.0015727145364508033, + 0.0033642337657511234, + 0.002290691714733839, + 0.002539195353165269, + 0.0018081114394590259, + 0.0032805746886879206, + 0.002161711687222123, + 0.0031904028728604317, + 0.003959312569350004, + 0.007472777273505926, + 0.007939493283629417, + 0.004530495963990688, + 0.002203459618613124, + 0.0042750961147248745, + 0.0026063090190291405, + 0.004016822203993797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.0007152383332140744 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368484601378441 + ], + "opposite_logprob_tokens": [ + -3.408684730529785 + ], + "logprob_reward": -0.03368484601378441, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368484601378441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04608968645334244, + 0.01851511187851429, + 0.7218287587165833, + 0.012589588761329651, + 0.01664699986577034, + 0.07150515168905258, + 0.013228701427578926, + 0.02484275959432125, + 0.01494691800326109, + 0.059806372970342636 + ], + "credit_map": [ + 0.004793283063918352, + 0.006693982984870672, + 0.006693977862596512, + 0.008153538219630718, + 0.006483930628746748, + 0.005565935280174017, + 0.004173277877271175, + 0.0035317542497068644, + 0.0032090696040540934, + 0.002355674048885703, + 0.0028377436101436615, + 0.002276594750583172, + 0.0020066460128873587, + 0.002164857229217887, + 0.0015510900411754847, + 0.002113434486091137, + 0.05268077924847603, + 0.02009931392967701, + 0.2631981074810028, + 0.01842140592634678, + 0.012734469957649708, + 0.29130303859710693, + 0.022369464859366417, + 0.041022125631570816, + 0.0016591110033914447, + 0.0011058043455705047, + 0.0020093326456844807, + 0.0013172267936170101, + 0.0012084335321560502, + 0.0017260750755667686, + 0.0012916794512420893, + 0.0022719234693795443, + 0.0018354567000642419, + 0.0014555639354512095, + 0.0033675581216812134, + 0.002409495646134019, + 0.0008169178036041558, + 0.0019048054236918688, + 0.001980341738089919, + 0.002876858925446868, + 0.00570939714089036, + 0.004699947778135538, + 0.015632323920726776, + 0.004121669102460146, + 0.0034548896364867687, + 0.010744153521955013, + 0.0034868260845541954, + 0.023655937984585762, + 0.0017238991567865014, + 0.0016979232896119356, + 0.002066568471491337, + 0.0012308444129303098, + 0.001125908805988729, + 0.001472707954235375, + 0.0011105076409876347, + 0.0028003407642245293, + 0.0037609636783599854, + 0.0029357175808399916, + 0.0054108696058392525, + 0.002781679853796959, + 0.0022630325984209776, + 0.002403771970421076, + 0.0018265327671542764, + 0.003460187464952469, + 0.001731792464852333, + 0.002980044111609459, + 0.0013401249889284372, + 0.0015229437267407775, + 0.0013781851157546043, + 0.002805589232593775, + 0.0012460831785574555, + 0.0019421519245952368, + 0.007328163832426071, + 0.010243894532322884, + 0.0119133610278368, + 0.007595854811370373, + 0.002881136955693364, + 0.007658381015062332, + 0.0041568828746676445, + 0.00802869163453579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 9 + ], + "selected_reward": -0.03368484601378441 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight head-tracking shot maintains constant focus on the nose of a paper airplane meticulously folded from a dreamy train ticket. The ticket, a warm cream with a pearlescent sheen, bears faint golden text in an elegant, artistic font, the creases of the paper crisp and soft. Slowly and gracefully, the paper airplane ascends into a gentle, fairytale world, its nose firmly pointed skyward, its body swaying gently in an invisible breeze, its gesture both innocent and endearing. The camera follows it through vast, fluffy, cotton-candy-like clouds tinted in pastel pinks, lavenders, and baby blues, their wingtips stirring up wisps of shimmering fluff as they pass. The entire scene is bathed in a dreamy, soft golden light, which penetrates the clouds and casts iridescent shadows on the paper airplane's surface, creating an atmosphere of tranquility, warmth, and whimsy. The overall image evokes the rich, subdued, and delicate style of a fairytale illustration.", + "seed": 5083, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_7/index_6734_seed_5083_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09813863199669867, + "vlm_logprob": -0.09813863199669867, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.09813863199669867, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.43528687953948975 + ], + "opposite_logprob_tokens": [ + -1.0602868795394897 + ], + "logprob_reward": -0.43528687953948975, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.43528687953948975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20415270328521729, + 0.10968326777219772, + 0.20846275985240936, + 0.04201238974928856, + 0.032771747559309006, + 0.04196449741721153, + 0.09930923581123352, + 0.03207769617438316, + 0.20984750986099243, + 0.019718199968338013 + ], + "credit_map": [ + 0.012102541513741016, + 0.012895088642835617, + 0.06944852322340012, + 0.01168764941394329, + 0.0063776602037250996, + 0.010629083961248398, + 0.009856589138507843, + 0.07115556299686432, + 0.012069103308022022, + 0.019099663943052292, + 0.013107545673847198, + 0.014618721790611744, + 0.009143293835222721, + 0.010647443123161793, + 0.015255363658070564, + 0.015742134302854538, + 0.0180951040238142, + 0.030564025044441223, + 0.03506564721465111, + 0.021606942638754845, + 0.03243831545114517, + 0.018139628693461418, + 0.032366931438446045, + 0.020186156034469604, + 0.009242395870387554, + 0.006045524962246418, + 0.0046040997840464115, + 0.0069914450868964195, + 0.002145300153642893, + 0.003806646214798093, + 0.0031453461851924658, + 0.006031629163771868, + 0.0036993154790252447, + 0.006606340873986483, + 0.004324634559452534, + 0.005317906849086285, + 0.003348396858200431, + 0.0035883889067918062, + 0.0034488581586629152, + 0.0024379072710871696, + 0.006272090133279562, + 0.006592877674847841, + 0.005944663658738136, + 0.0059449574910104275, + 0.004524535499513149, + 0.0056943935342133045, + 0.003982045222073793, + 0.0030089300125837326, + 0.007474606391042471, + 0.016771167516708374, + 0.011243322864174843, + 0.015321013517677784, + 0.008602306246757507, + 0.016548248007893562, + 0.010185100138187408, + 0.013163474388420582, + 0.004280010238289833, + 0.004719351883977652, + 0.004622401669621468, + 0.0062074861489236355, + 0.002667698310688138, + 0.0025387282948940992, + 0.003028805600479245, + 0.00401321379467845, + 0.014471607282757759, + 0.015181253664195538, + 0.06172158569097519, + 0.051765747368335724, + 0.011487593874335289, + 0.016938764601945877, + 0.018779132515192032, + 0.019501816481351852, + 0.0019944971427321434, + 0.004154636990278959, + 0.004381862934678793, + 0.002339068101719022, + 0.0015407941536977887, + 0.0019771624356508255, + 0.0022190995514392853, + 0.0011110784253105521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.43528687953948975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007640897762030363 + ], + "opposite_logprob_tokens": [ + -4.882640838623047 + ], + "logprob_reward": -0.007640897762030363, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007640897762030363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26878154277801514, + 0.061685699969530106, + 0.11219118535518646, + 0.11604437977075577, + 0.08421345055103302, + 0.02804849110543728, + 0.12398535758256912, + 0.032501593232154846, + 0.13532193005084991, + 0.03722634166479111 + ], + "credit_map": [ + 0.02169130928814411, + 0.025306889787316322, + 0.06767138093709946, + 0.020466621965169907, + 0.008502816781401634, + 0.023677045479416847, + 0.020283136516809464, + 0.08118234574794769, + 0.007799241691827774, + 0.009158249013125896, + 0.008366257883608341, + 0.006048322655260563, + 0.004265132825821638, + 0.004686571657657623, + 0.006788962986320257, + 0.014572962187230587, + 0.012169894762337208, + 0.015116492286324501, + 0.014041876420378685, + 0.013191552832722664, + 0.016397491097450256, + 0.010764210484921932, + 0.017585981637239456, + 0.01292368583381176, + 0.028900640085339546, + 0.01526521798223257, + 0.012499966658651829, + 0.015376528725028038, + 0.005588993430137634, + 0.00916380062699318, + 0.007616775576025248, + 0.021632462739944458, + 0.007629443891346455, + 0.01561898272484541, + 0.009076104499399662, + 0.01507351454347372, + 0.007724096532911062, + 0.0071735866367816925, + 0.009973851032555103, + 0.011943865567445755, + 0.003907005302608013, + 0.0031628336291760206, + 0.0032000532373785973, + 0.003382356371730566, + 0.0043351538479328156, + 0.00453540962189436, + 0.003079975489526987, + 0.0024457047693431377, + 0.009488706476986408, + 0.019876785576343536, + 0.015311806462705135, + 0.01614409312605858, + 0.01307663694024086, + 0.023281654343008995, + 0.014884180389344692, + 0.011921488679945469, + 0.003001478733494878, + 0.002889015479013324, + 0.005901242606341839, + 0.007027070038020611, + 0.001604124903678894, + 0.001796033582650125, + 0.0030816972721368074, + 0.007200931198894978, + 0.00991739146411419, + 0.010940538719296455, + 0.03615957498550415, + 0.029538217931985855, + 0.007392833009362221, + 0.013084355741739273, + 0.012742040678858757, + 0.015546981245279312, + 0.004016278311610222, + 0.008892008103430271, + 0.008158507756888866, + 0.004257048014551401, + 0.002494849031791091, + 0.003798089688643813, + 0.0033732845913618803, + 0.0022362780291587114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.007640897762030363 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636622689664364 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636622689664364, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636622689664364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13278844952583313, + 0.030048182234168053, + 0.07885222136974335, + 0.02479766681790352, + 0.030270984396338463, + 0.1092686727643013, + 0.3286362588405609, + 0.03072509728372097, + 0.18929924070835114, + 0.0453132726252079 + ], + "credit_map": [ + 0.007732053752988577, + 0.007609535474330187, + 0.045767851173877716, + 0.007749125361442566, + 0.0035579714458435774, + 0.007025281898677349, + 0.0070623550564050674, + 0.046284269541502, + 0.004335278179496527, + 0.0044115763157606125, + 0.004250062629580498, + 0.0032963580451905727, + 0.0020675258710980415, + 0.0025975792668759823, + 0.003076584078371525, + 0.0060132164508104324, + 0.007013663183897734, + 0.01234991941601038, + 0.014133350923657417, + 0.008069333620369434, + 0.010491828434169292, + 0.006229621823877096, + 0.011779424734413624, + 0.008785084821283817, + 0.0035210291389375925, + 0.0037089253310114145, + 0.0027881592977792025, + 0.003569484455510974, + 0.0024263530503958464, + 0.0028369331266731024, + 0.0025424486957490444, + 0.003404331626370549, + 0.004082941450178623, + 0.003878404852002859, + 0.004050136543810368, + 0.005093268118798733, + 0.0027340061496943235, + 0.003572517540305853, + 0.0031715240329504013, + 0.003688186639919877, + 0.011806806549429893, + 0.015930788591504097, + 0.014821860007941723, + 0.01682646945118904, + 0.012401788495481014, + 0.01744132488965988, + 0.011871355585753918, + 0.008168271742761135, + 0.0281068105250597, + 0.05240574851632118, + 0.03704873472452164, + 0.05176717787981033, + 0.03094194456934929, + 0.0564938522875309, + 0.03608306124806404, + 0.035788942128419876, + 0.0038127631414681673, + 0.003949279431253672, + 0.00375345372594893, + 0.004423796199262142, + 0.0032062416430562735, + 0.003211881732568145, + 0.004053990822285414, + 0.004313691519200802, + 0.020491115748882294, + 0.022024039179086685, + 0.024673376232385635, + 0.032892730087041855, + 0.016305889934301376, + 0.03132041171193123, + 0.016167685389518738, + 0.02542399987578392, + 0.005347502417862415, + 0.007142484188079834, + 0.005605953745543957, + 0.005610211752355099, + 0.005120431073009968, + 0.0069539546966552734, + 0.0058683366514742374, + 0.0036644008941948414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.008636622689664364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010408704401925206 + ], + "opposite_logprob_tokens": [ + -6.876040935516357 + ], + "logprob_reward": -0.0010408704401925206, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010408704401925206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14763392508029938, + 0.05887151136994362, + 0.10133511573076248, + 0.11591683328151703, + 0.1046648696064949, + 0.10064595937728882, + 0.06528182327747345, + 0.05897229164838791, + 0.1185799315571785, + 0.12809771299362183 + ], + "credit_map": [ + 0.013627885840833187, + 0.01691683754324913, + 0.025048157200217247, + 0.023067813366651535, + 0.009323892183601856, + 0.01858736015856266, + 0.01827842928469181, + 0.022783534601330757, + 0.008786355145275593, + 0.007723829708993435, + 0.0075327870436012745, + 0.006156382150948048, + 0.004014423582702875, + 0.004729204345494509, + 0.006411978974938393, + 0.013516544364392757, + 0.010270833969116211, + 0.014974307268857956, + 0.013381497003138065, + 0.011003280058503151, + 0.015132972039282322, + 0.008867251686751842, + 0.015787487849593163, + 0.011917472817003727, + 0.020564276725053787, + 0.017475392669439316, + 0.012659047730267048, + 0.020227590575814247, + 0.007568659260869026, + 0.009250201284885406, + 0.010742483660578728, + 0.0174291729927063, + 0.014934560284018517, + 0.012255288660526276, + 0.011850166134536266, + 0.021725734695792198, + 0.00691289734095335, + 0.007635247427970171, + 0.009786345064640045, + 0.019564613699913025, + 0.014894680120050907, + 0.01897038146853447, + 0.016174592077732086, + 0.017370184883475304, + 0.007065650075674057, + 0.010285410098731518, + 0.0081113176420331, + 0.007773738820105791, + 0.006479587871581316, + 0.010196845978498459, + 0.007957511581480503, + 0.00962433684617281, + 0.008022473193705082, + 0.008578192442655563, + 0.006256820168346167, + 0.00816604495048523, + 0.009021216072142124, + 0.00831108633428812, + 0.008708191104233265, + 0.01184989232569933, + 0.00364886736497283, + 0.0038831443525850773, + 0.00438413443043828, + 0.009165755473077297, + 0.010267224162817001, + 0.01262668240815401, + 0.025252915918827057, + 0.019546838477253914, + 0.009408023208379745, + 0.014977515675127506, + 0.01388026587665081, + 0.01262046117335558, + 0.013102126307785511, + 0.02708487957715988, + 0.026500364765524864, + 0.01807819865643978, + 0.007572526577860117, + 0.011911291629076004, + 0.011777100153267384, + 0.012071224860846996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0010408704401925206 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a paper airplane folded from a dreamy train ticket?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the paper airplane colored warm cream with a pearlescent sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the paper airplane ascend slowly and gracefully into the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background composed of vast, fluffy clouds tinted in pastel pinks, lavenders, and baby blues?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in a soft golden light that casts iridescent shadows on the paper airplane?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038087889552116394 + ], + "opposite_logprob_tokens": [ + -3.288087844848633 + ], + "logprob_reward": -0.038087889552116394, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038087889552116394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2827737629413605, + 0.023304881528019905, + 0.1002923846244812, + 0.05723048001527786, + 0.10441849380731583, + 0.12612293660640717, + 0.09745721518993378, + 0.0553860068321228, + 0.11410142481327057, + 0.038912396878004074 + ], + "credit_map": [ + 0.017196958884596825, + 0.011265583336353302, + 0.10004083067178726, + 0.013832257129251957, + 0.0062100645154714584, + 0.015159045346081257, + 0.014735540375113487, + 0.10433347523212433, + 0.004068802110850811, + 0.003332501510158181, + 0.00339781423099339, + 0.0029134219512343407, + 0.0016774360556155443, + 0.0027780074160546064, + 0.002222501439973712, + 0.0029143996071070433, + 0.010400624014437199, + 0.014367169700562954, + 0.019651522859930992, + 0.011653697118163109, + 0.010600964538753033, + 0.008000355213880539, + 0.012349873781204224, + 0.013268174603581429, + 0.01116376742720604, + 0.007950887084007263, + 0.00798108335584402, + 0.007234823890030384, + 0.003724223468452692, + 0.0044211153872311115, + 0.003973718266934156, + 0.010780862532556057, + 0.012305153533816338, + 0.016336234286427498, + 0.01301467977464199, + 0.02784579247236252, + 0.005786040332168341, + 0.006898735649883747, + 0.009153171442449093, + 0.013078693300485611, + 0.021370448172092438, + 0.02097838744521141, + 0.01613001525402069, + 0.018557656556367874, + 0.014853973872959614, + 0.015926998108625412, + 0.010035344399511814, + 0.00827011838555336, + 0.006671298760920763, + 0.013850823976099491, + 0.010905307717621326, + 0.01705830730497837, + 0.008436813950538635, + 0.0118310721591115, + 0.00815755408257246, + 0.020546039566397667, + 0.0074279457330703735, + 0.0036948618944734335, + 0.007578078657388687, + 0.010656760074198246, + 0.004905246198177338, + 0.004818114917725325, + 0.004891592543572187, + 0.011413408443331718, + 0.01193380355834961, + 0.010029311291873455, + 0.019734004512429237, + 0.02130434475839138, + 0.011254444718360901, + 0.01799587905406952, + 0.010997562669217587, + 0.010852067731320858, + 0.0035388802643865347, + 0.0067462497390806675, + 0.012067045085132122, + 0.005696104373782873, + 0.0022913545835763216, + 0.0041029429994523525, + 0.0028875861316919327, + 0.001582237659022212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.038087889552116394 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_8/index_tensor([640, 640])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_8/index_tensor([640, 640])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ee208e15abf73b26d78f98c6cb36973c5744eff7 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_8/index_tensor([640, 640])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, medium-length, static shot, taken from a level angle, captures a graying elderly couple sitting side by side in a warm, if slightly dated, living room. They sit by a large bay window, soft afternoon sunlight slanting through the clear glass, casting long, warm patches of light on the wooden floor beside them. The husband, seated on the left, wears a dark blue wool cardigan over a light gray plaid shirt. His face, deeply wrinkled by time, gazes calmly ahead, as if lost in thought. Beside him, his wife, dressed in a soft, off-white turtleneck sweater, her equally gray hair elegantly tied in a bun. Her head tilted slightly to the right, her gaze over her husband's shoulder, distant and unfocused, gazing out the window at the late autumn scene, where a few withered yellow leaves slowly drift down. Then, her chest rose and fell almost imperceptibly, her lips parted slightly, and she let out a soft and long sigh. This sigh was particularly clear in the quiet room, injecting a peaceful, nostalgic and slightly sentimental atmosphere into the whole picture.", + "seed": 22596, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_8/index_640_seed_22596_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021717375423759223, + "vlm_logprob": -0.021717375423759223, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.021717375423759223, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01262658555060625 + ], + "opposite_logprob_tokens": [ + -4.637626647949219 + ], + "logprob_reward": -0.01262658555060625, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01262658555060625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22413478791713715, + 0.12123141437768936, + 0.11183670908212662, + 0.08722902089357376, + 0.06699121743440628, + 0.09994007647037506, + 0.10726474970579147, + 0.11321130394935608, + 0.02228575572371483, + 0.045875027775764465 + ], + "credit_map": [ + 0.03187832981348038, + 0.01661677658557892, + 0.027713065966963768, + 0.028794126585125923, + 0.036059316247701645, + 0.02960841730237007, + 0.023987287655472755, + 0.029477430507540703, + 0.014070049859583378, + 0.009364750236272812, + 0.010225593112409115, + 0.019331688061356544, + 0.023759955540299416, + 0.015620149672031403, + 0.015411595813930035, + 0.013447620905935764, + 0.014441668055951595, + 0.006847350858151913, + 0.008205167949199677, + 0.01370523776859045, + 0.01963699795305729, + 0.021807953715324402, + 0.01208299957215786, + 0.015109331347048283, + 0.012822073884308338, + 0.007557063363492489, + 0.009682083502411842, + 0.009016062133014202, + 0.009568020701408386, + 0.01239987276494503, + 0.00922536663711071, + 0.01695847325026989, + 0.009462499059736729, + 0.003953961189836264, + 0.0059647755697369576, + 0.006768572144210339, + 0.012578070163726807, + 0.011478802189230919, + 0.005695031490176916, + 0.011089504696428776, + 0.014318191446363926, + 0.007078113965690136, + 0.012640496715903282, + 0.009837417863309383, + 0.015275130048394203, + 0.016837870702147484, + 0.0073057496920228004, + 0.01664709858596325, + 0.014815951697528362, + 0.007947506383061409, + 0.013775232248008251, + 0.011264869943261147, + 0.014924486167728901, + 0.013662897981703281, + 0.012792246416211128, + 0.018081551417708397, + 0.01607341319322586, + 0.007546847220510244, + 0.013814376667141914, + 0.01385587453842163, + 0.01718325912952423, + 0.014580094255506992, + 0.011697092093527317, + 0.018460338935256004, + 0.003100566565990448, + 0.0018358735833317041, + 0.001728143892250955, + 0.003411747282370925, + 0.003728409530594945, + 0.0028059205505996943, + 0.003354596672579646, + 0.002320494968444109, + 0.006270319223403931, + 0.003430560464039445, + 0.004418224096298218, + 0.005205261521041393, + 0.007560781203210354, + 0.00726388581097126, + 0.0047248234041035175, + 0.007001168094575405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.01262658555060625 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208828091621399 + ], + "opposite_logprob_tokens": [ + -2.8120882511138916 + ], + "logprob_reward": -0.06208828091621399, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208828091621399, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13914795219898224, + 0.12206597626209259, + 0.09104759246110916, + 0.13912992179393768, + 0.06046126410365105, + 0.08165254443883896, + 0.18193064630031586, + 0.14270718395709991, + 0.025529304519295692, + 0.016327599063515663 + ], + "credit_map": [ + 0.023189594969153404, + 0.01678447797894478, + 0.01915479451417923, + 0.014206262305378914, + 0.021304115653038025, + 0.014830978587269783, + 0.013438470661640167, + 0.016239259392023087, + 0.01592933014035225, + 0.016922766342759132, + 0.0100450050085783, + 0.016719132661819458, + 0.02206563763320446, + 0.013349129818379879, + 0.013902961276471615, + 0.013132010586559772, + 0.012765947729349136, + 0.005819946061819792, + 0.008056702092289925, + 0.00935110729187727, + 0.014881388284265995, + 0.016272561624646187, + 0.011920128017663956, + 0.011979810893535614, + 0.018338629975914955, + 0.011432058177888393, + 0.014394374564290047, + 0.01882166974246502, + 0.015824327245354652, + 0.016344862058758736, + 0.016773618757724762, + 0.02720038592815399, + 0.008159592747688293, + 0.005113068502396345, + 0.005573655478656292, + 0.011742131784558296, + 0.010018207132816315, + 0.005722487810999155, + 0.007511603180319071, + 0.006620518863201141, + 0.012113800272345543, + 0.007751520723104477, + 0.014838727191090584, + 0.007888604886829853, + 0.01058474462479353, + 0.008258402347564697, + 0.008975357748568058, + 0.011241390369832516, + 0.024534616619348526, + 0.014385752379894257, + 0.024542856961488724, + 0.019466787576675415, + 0.024069087579846382, + 0.021648745983839035, + 0.019986428320407867, + 0.0332963801920414, + 0.01884474977850914, + 0.01003814022988081, + 0.015547127462923527, + 0.018029710277915, + 0.02267862856388092, + 0.02066902443766594, + 0.017487702891230583, + 0.019412096589803696, + 0.003152737393975258, + 0.0026729502715170383, + 0.0026423053350299597, + 0.0037461724132299423, + 0.004102228209376335, + 0.003077845321968198, + 0.0030788355506956577, + 0.0030562295578420162, + 0.0025390933733433485, + 0.0017459612572565675, + 0.0019306624308228493, + 0.0019012640696018934, + 0.0022307743784040213, + 0.00245781266130507, + 0.001593292341567576, + 0.001928738784044981 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.06208828091621399 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598655566573143 + ], + "opposite_logprob_tokens": [ + -3.8955986499786377 + ], + "logprob_reward": -0.020598655566573143, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598655566573143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2207767814397812, + 0.1886516958475113, + 0.10452213883399963, + 0.06010891869664192, + 0.11831934005022049, + 0.09032752364873886, + 0.050016939640045166, + 0.06835824251174927, + 0.036916352808475494, + 0.06200205162167549 + ], + "credit_map": [ + 0.027589062228798866, + 0.019052453339099884, + 0.016987908631563187, + 0.0337459035217762, + 0.030539829283952713, + 0.036836083978414536, + 0.027070723474025726, + 0.028954830020666122, + 0.023629603907465935, + 0.013325060717761517, + 0.015545873902738094, + 0.02165665291249752, + 0.02318734861910343, + 0.03145008534193039, + 0.03782375901937485, + 0.022033318877220154, + 0.014313112944364548, + 0.009356029331684113, + 0.008391045033931732, + 0.014983944594860077, + 0.015616893768310547, + 0.01547001302242279, + 0.014789198525249958, + 0.011601903475821018, + 0.008423706516623497, + 0.005043952725827694, + 0.007579171098768711, + 0.006666386034339666, + 0.008974689058959484, + 0.00810325238853693, + 0.009212312288582325, + 0.006105449516326189, + 0.016984989866614342, + 0.009658903814852238, + 0.01219616923481226, + 0.01856495812535286, + 0.01483254786580801, + 0.015098548494279385, + 0.018570683896541595, + 0.0124125387519598, + 0.010651963762938976, + 0.006870337296277285, + 0.009652344509959221, + 0.013220392167568207, + 0.0132290692999959, + 0.011538881808519363, + 0.013957858085632324, + 0.011206681840121746, + 0.00700604822486639, + 0.0053816852159798145, + 0.005181277170777321, + 0.005799674894660711, + 0.00699320062994957, + 0.008320963010191917, + 0.003972175996750593, + 0.007361915893852711, + 0.00896540004760027, + 0.005792564712464809, + 0.0046170256100595, + 0.009989139623939991, + 0.00981020089238882, + 0.011003492400050163, + 0.009605184197425842, + 0.00857523549348116, + 0.005406955722719431, + 0.0025256231892853975, + 0.002595867495983839, + 0.00416976073756814, + 0.0061379666440188885, + 0.005437599960714579, + 0.005419817287474871, + 0.005222758278250694, + 0.009091900661587715, + 0.00514491880312562, + 0.003967213910073042, + 0.009120413102209568, + 0.008024205453693867, + 0.009160257875919342, + 0.01015418954193592, + 0.00733894994482398 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.020598655566573143 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638150800019503 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.004638150800019503, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638150800019503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19174498319625854, + 0.2144329994916916, + 0.09920056164264679, + 0.05369875207543373, + 0.09490342438220978, + 0.06332966685295105, + 0.07884113490581512, + 0.09405578672885895, + 0.038052767515182495, + 0.07173988968133926 + ], + "credit_map": [ + 0.02937726303935051, + 0.013096804730594158, + 0.0137558588758111, + 0.025950023904442787, + 0.028395261615514755, + 0.029552649706602097, + 0.024122530594468117, + 0.027494605630636215, + 0.03035576641559601, + 0.01552567444741726, + 0.015789881348609924, + 0.021690793335437775, + 0.034313835203647614, + 0.039655767381191254, + 0.021143203601241112, + 0.03595810383558273, + 0.011456402018666267, + 0.007959548383951187, + 0.0073202820494771, + 0.017132297158241272, + 0.016288019716739655, + 0.01435819175094366, + 0.013202118687331676, + 0.011483701877295971, + 0.007524724584072828, + 0.004023348912596703, + 0.0038290624506771564, + 0.0066373227164149284, + 0.009297125972807407, + 0.009303771890699863, + 0.006390286143869162, + 0.006693112663924694, + 0.014775260351598263, + 0.005934750661253929, + 0.007373025640845299, + 0.009398864582180977, + 0.01316032838076353, + 0.01784323714673519, + 0.009831134229898453, + 0.01658683642745018, + 0.007752606645226479, + 0.0049979472532868385, + 0.006518462207168341, + 0.00997393298894167, + 0.010413221083581448, + 0.008456472307443619, + 0.007680711802095175, + 0.007536314427852631, + 0.011221413500607014, + 0.00784358475357294, + 0.014119208790361881, + 0.007857667282223701, + 0.009451624006032944, + 0.009806061163544655, + 0.006349359638988972, + 0.01219221856445074, + 0.012960361316800117, + 0.006231681443750858, + 0.008324873633682728, + 0.00945101585239172, + 0.014621409587562084, + 0.019945548847317696, + 0.008299515582621098, + 0.014221382327377796, + 0.00541258417069912, + 0.002843510592356324, + 0.0021120065357536077, + 0.004716339521110058, + 0.006335896905511618, + 0.006227648817002773, + 0.00534924678504467, + 0.005055538844317198, + 0.010591404512524605, + 0.0050793057307600975, + 0.004824963863939047, + 0.0074490271508693695, + 0.0093922670930624, + 0.013789859600365162, + 0.007942434400320053, + 0.012670635245740414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.004638150800019503 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635204285383224 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008635204285383224, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635204285383224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3113100528717041, + 0.08260423690080643, + 0.08741804957389832, + 0.05819488689303398, + 0.1120898649096489, + 0.05091220512986183, + 0.08584672212600708, + 0.10716389864683151, + 0.046225111931562424, + 0.05823495239019394 + ], + "credit_map": [ + 0.0377313457429409, + 0.021047284826636314, + 0.041457366198301315, + 0.04527377709746361, + 0.05737534910440445, + 0.03785189613699913, + 0.043998923152685165, + 0.026574166491627693, + 0.01221720315515995, + 0.006668453570455313, + 0.005125584080815315, + 0.013690889813005924, + 0.013913116417825222, + 0.0081022372469306, + 0.012869980186223984, + 0.010016776621341705, + 0.011075715534389019, + 0.009497638791799545, + 0.007607670035213232, + 0.011434453539550304, + 0.013171647675335407, + 0.009988972917199135, + 0.009283915162086487, + 0.01535804383456707, + 0.007448550313711166, + 0.004656066652387381, + 0.00675206771120429, + 0.005887177307158709, + 0.011685148812830448, + 0.007733956445008516, + 0.005926721729338169, + 0.00810520350933075, + 0.013336933217942715, + 0.009735612198710442, + 0.015727171674370766, + 0.02280762419104576, + 0.016221653670072556, + 0.009367247112095356, + 0.014745364896953106, + 0.010148271918296814, + 0.007512526586651802, + 0.005213787779211998, + 0.003928833175450563, + 0.005624157376587391, + 0.008946474641561508, + 0.00777823431417346, + 0.0052510336972773075, + 0.006657165475189686, + 0.011456325650215149, + 0.008861587382853031, + 0.012532344087958336, + 0.011618024669587612, + 0.012258012779057026, + 0.009352992288768291, + 0.008295143954455853, + 0.011472295969724655, + 0.01527375727891922, + 0.009088022634387016, + 0.014765841886401176, + 0.014442087151110172, + 0.015613834373652935, + 0.01123656053096056, + 0.012066161260008812, + 0.014677639119327068, + 0.006251656915992498, + 0.003724517300724983, + 0.00401317048817873, + 0.00561036029830575, + 0.0070326728746294975, + 0.006573379505425692, + 0.006922538857907057, + 0.006096820347011089, + 0.007646827958524227, + 0.004809551872313023, + 0.007040387485176325, + 0.011303169652819633, + 0.008096958510577679, + 0.005338465794920921, + 0.009339440613985062, + 0.004660156089812517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.008635204285383224 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, medium-length, static shot, taken from a level angle, captures a graying elderly couple sitting side by side in a warm, if slightly dated, living room. They sit by a large bay window, soft afternoon sunlight slanting through the clear glass, casting long, warm patches of light on the wooden floor beside them. The husband, seated on the left, wears a dark blue wool cardigan over a light gray plaid shirt. His face, deeply wrinkled by time, gazes calmly ahead, as if lost in thought. Beside him, his wife, dressed in a soft, off-white turtleneck sweater, her equally gray hair elegantly tied in a bun. Her head tilted slightly to the right, her gaze over her husband's shoulder, distant and unfocused, gazing out the window at the late autumn scene, where a few withered yellow leaves slowly drift down. Then, her chest rose and fell almost imperceptibly, her lips parted slightly, and she let out a soft and long sigh. This sigh was particularly clear in the quiet room, injecting a peaceful, nostalgic and slightly sentimental atmosphere into the whole picture.", + "seed": 22597, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_8/index_640_seed_22597_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.032060821540653706, + "vlm_logprob": -0.032060821540653706, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.032060821540653706, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0731998085975647 + ], + "opposite_logprob_tokens": [ + -2.69819974899292 + ], + "logprob_reward": -0.0731998085975647, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0731998085975647, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22175104916095734, + 0.0926111713051796, + 0.04326678067445755, + 0.033550992608070374, + 0.030987825244665146, + 0.020148925483226776, + 0.14703579246997833, + 0.20907147228717804, + 0.13544341921806335, + 0.06613264977931976 + ], + "credit_map": [ + 0.04628482833504677, + 0.05577520281076431, + 0.014565529301762581, + 0.02255377732217312, + 0.019815679639577866, + 0.02111198380589485, + 0.023720962926745415, + 0.01792306639254093, + 0.013700424693524837, + 0.011919681914150715, + 0.008872658014297485, + 0.007483207620680332, + 0.007652109954506159, + 0.011894550174474716, + 0.009549407288432121, + 0.0215391144156456, + 0.010528705082833767, + 0.011141533963382244, + 0.0036247277166694403, + 0.0031809001229703426, + 0.0034959057811647654, + 0.0034833326935768127, + 0.0045043970458209515, + 0.003307271283119917, + 0.008094985969364643, + 0.010341467335820198, + 0.0036758026108145714, + 0.0017371915746480227, + 0.0021621952764689922, + 0.0021658281330019236, + 0.0025171800516545773, + 0.002856336534023285, + 0.004516831133514643, + 0.0048108878545463085, + 0.0031396623235195875, + 0.005070383194833994, + 0.0032937484793365, + 0.0037147903349250555, + 0.003375324886292219, + 0.003066195175051689, + 0.0016986626433208585, + 0.002049054251983762, + 0.0016741491854190826, + 0.0034541140776127577, + 0.0020036010537296534, + 0.002461210126057267, + 0.0025208781007677317, + 0.004287254996597767, + 0.04768289253115654, + 0.04022293910384178, + 0.00794767215847969, + 0.004494876600801945, + 0.004278731998056173, + 0.007156959269195795, + 0.011039952747523785, + 0.02421175315976143, + 0.06261231005191803, + 0.08701074123382568, + 0.008451240137219429, + 0.007032402325421572, + 0.011180145666003227, + 0.0071963956579566, + 0.014006098732352257, + 0.011582120321691036, + 0.03669125214219093, + 0.05187749117612839, + 0.00508042611181736, + 0.0057159545831382275, + 0.007939011789858341, + 0.006175983231514692, + 0.008497393690049648, + 0.013465892523527145, + 0.02117491327226162, + 0.0264891367405653, + 0.002675833413377404, + 0.0025971548166126013, + 0.003105633193627, + 0.0025295971427112818, + 0.003555681323632598, + 0.004004692193120718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0731998085975647 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016083618625998497 + ], + "opposite_logprob_tokens": [ + -4.141083717346191 + ], + "logprob_reward": -0.016083618625998497, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016083618625998497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38200533390045166, + 0.06407415866851807, + 0.07943364232778549, + 0.04179522022604942, + 0.024266190826892853, + 0.04562503471970558, + 0.07339981943368912, + 0.1988229602575302, + 0.04580261930823326, + 0.04477502033114433 + ], + "credit_map": [ + 0.05756332725286484, + 0.11000970005989075, + 0.033708274364471436, + 0.028728362172842026, + 0.02463812381029129, + 0.029187843203544617, + 0.02989565208554268, + 0.06827408075332642, + 0.010876388289034367, + 0.013280821964144707, + 0.007967909798026085, + 0.005601240787655115, + 0.005620966665446758, + 0.0056791044771671295, + 0.006555937696248293, + 0.008491788059473038, + 0.019710419699549675, + 0.01665749028325081, + 0.00762038491666317, + 0.00642202515155077, + 0.004416894167661667, + 0.0058568003587424755, + 0.0075949374586343765, + 0.011154686100780964, + 0.006561695132404566, + 0.007686728611588478, + 0.00508763687685132, + 0.006251227576285601, + 0.003576305927708745, + 0.004370708484202623, + 0.004107351880520582, + 0.004153567831963301, + 0.004219468683004379, + 0.005090428050607443, + 0.0028531255666166544, + 0.0024180777836591005, + 0.002148693660274148, + 0.002060408005490899, + 0.002335210796445608, + 0.003140778513625264, + 0.007407608907669783, + 0.011672534048557281, + 0.004853666760027409, + 0.0030694964807480574, + 0.0036583340261131525, + 0.0042534940876066685, + 0.003416474210098386, + 0.0072934250347316265, + 0.013731627725064754, + 0.00906607136130333, + 0.008196739479899406, + 0.0049684252589941025, + 0.004063529893755913, + 0.0064363181591033936, + 0.006597968749701977, + 0.02033914066851139, + 0.0621960386633873, + 0.07931573688983917, + 0.00858570821583271, + 0.006502216216176748, + 0.009023940190672874, + 0.006979049649089575, + 0.015400738455355167, + 0.010819531045854092, + 0.008965425193309784, + 0.005325779784470797, + 0.003592404769733548, + 0.004086671397089958, + 0.002162214135751128, + 0.004626507870852947, + 0.004024293273687363, + 0.013019321486353874, + 0.005665502976626158, + 0.01149226538836956, + 0.0033399665262550116, + 0.0018228599801659584, + 0.0023124595172703266, + 0.005539726000279188, + 0.0025166734121739864, + 0.012085567228496075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.016083618625998497 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048701219260692596 + ], + "opposite_logprob_tokens": [ + -3.048701286315918 + ], + "logprob_reward": -0.048701219260692596, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048701219260692596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10710764676332474, + 0.09906169772148132, + 0.23191769421100616, + 0.03394899144768715, + 0.02040891721844673, + 0.019884634763002396, + 0.2036171704530716, + 0.15204468369483948, + 0.114568330347538, + 0.01744026504456997 + ], + "credit_map": [ + 0.011778089217841625, + 0.02066091075539589, + 0.008307837881147861, + 0.010605002753436565, + 0.011416619643568993, + 0.015791205689311028, + 0.012538711540400982, + 0.016009267419576645, + 0.016294488683342934, + 0.01679394207894802, + 0.007932022213935852, + 0.009249075315892696, + 0.010539832524955273, + 0.013144990429282188, + 0.014266000129282475, + 0.010841339826583862, + 0.07929452508687973, + 0.05437025800347328, + 0.009643369354307652, + 0.007018945179879665, + 0.008882533758878708, + 0.012709804810583591, + 0.019939368590712547, + 0.040058884769678116, + 0.00630644615739584, + 0.006759743671864271, + 0.003940870054066181, + 0.0028182303067296743, + 0.002251527039334178, + 0.002968529472127557, + 0.003895191242918372, + 0.005008452571928501, + 0.0028913160786032677, + 0.003926520235836506, + 0.0012856601970270276, + 0.002109702443704009, + 0.002272537909448147, + 0.002809261903166771, + 0.0028398456051945686, + 0.0022740697022527456, + 0.003544690553098917, + 0.004218693356961012, + 0.0019658859819173813, + 0.0017921776743605733, + 0.0017671964596956968, + 0.0022316158283501863, + 0.0023006361443549395, + 0.002063737716525793, + 0.06178823858499527, + 0.08292452991008759, + 0.007773188408464193, + 0.007210324518382549, + 0.009835361503064632, + 0.006840127985924482, + 0.014600349590182304, + 0.012645057402551174, + 0.05180506035685539, + 0.05167238041758537, + 0.0050595407374203205, + 0.0049548065289855, + 0.00761569244787097, + 0.005974364932626486, + 0.011444115079939365, + 0.013518725521862507, + 0.03887719288468361, + 0.04155301675200462, + 0.0043511297553777695, + 0.0038238163106143475, + 0.005017603747546673, + 0.004098361823707819, + 0.007950663566589355, + 0.008896530605852604, + 0.0030361358076334, + 0.0017382147489115596, + 0.001501138904131949, + 0.0012948897201567888, + 0.0016700299456715584, + 0.002401028759777546, + 0.0018225490348413587, + 0.003976276144385338 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.048701219260692596 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540650554001331 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540650554001331, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540650554001331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10102367401123047, + 0.1331774890422821, + 0.13132478296756744, + 0.026889579370617867, + 0.012162097729742527, + 0.017449740320444107, + 0.10446244478225708, + 0.32995399832725525, + 0.09181539714336395, + 0.05174083635210991 + ], + "credit_map": [ + 0.014981214888393879, + 0.009066067636013031, + 0.00887266919016838, + 0.012850397266447544, + 0.009992321021854877, + 0.01603091135621071, + 0.01717926375567913, + 0.012050819583237171, + 0.03366134688258171, + 0.029722295701503754, + 0.008973618969321251, + 0.010088052600622177, + 0.008438125252723694, + 0.011347312480211258, + 0.014139742590487003, + 0.016806989908218384, + 0.041958462446928024, + 0.026069380342960358, + 0.0056710983626544476, + 0.005145620554685593, + 0.006143827456980944, + 0.00956933293491602, + 0.011633869260549545, + 0.02513318695127964, + 0.006057374645024538, + 0.005342611111700535, + 0.0029460107907652855, + 0.00144385255407542, + 0.0015794163336977363, + 0.0021756894420832396, + 0.0025123318191617727, + 0.004832289647310972, + 0.0014075045473873615, + 0.0020850207656621933, + 0.0011173513485118747, + 0.0018444163724780083, + 0.0012382076820358634, + 0.001345915487036109, + 0.001323011121712625, + 0.0018006691243499517, + 0.0036216219887137413, + 0.004285986069589853, + 0.0016571859596297145, + 0.0014992570504546165, + 0.0013377665309235454, + 0.0016431001713499427, + 0.0014521093107759953, + 0.001952711958438158, + 0.027376826852560043, + 0.03793950006365776, + 0.005032168235629797, + 0.006630482617765665, + 0.006427589803934097, + 0.005181439220905304, + 0.007540319114923477, + 0.00833410955965519, + 0.10989215224981308, + 0.1290847212076187, + 0.011738751083612442, + 0.009826797991991043, + 0.014246447943150997, + 0.011263735592365265, + 0.02314116433262825, + 0.020760204643011093, + 0.030925316736102104, + 0.034127846360206604, + 0.003534936113283038, + 0.00366607797332108, + 0.0049448697827756405, + 0.0030177696608006954, + 0.00662546930834651, + 0.004973114468157291, + 0.01524069532752037, + 0.01554413978010416, + 0.00198741490021348, + 0.0015100594609975815, + 0.0020152709912508726, + 0.005175548605620861, + 0.002832567086443305, + 0.007435137871652842 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 2 + ], + "selected_reward": -0.012540650554001331 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778810665011406 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009778810665011406, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778810665011406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09823407232761383, + 0.07172150909900665, + 0.18011343479156494, + 0.055888622999191284, + 0.0367501899600029, + 0.04138924181461334, + 0.355264276266098, + 0.05459466204047203, + 0.03564595431089401, + 0.0703980103135109 + ], + "credit_map": [ + 0.01817418821156025, + 0.011051114648580551, + 0.012703095562756062, + 0.011207948438823223, + 0.00863315537571907, + 0.011626732535660267, + 0.010562022216618061, + 0.014275819063186646, + 0.018126877024769783, + 0.0173492394387722, + 0.005529108922928572, + 0.006033692974597216, + 0.0049948482774198055, + 0.006634937133640051, + 0.007124999538064003, + 0.005927807651460171, + 0.062095269560813904, + 0.04238002002239227, + 0.006694645155221224, + 0.004634128417819738, + 0.005640998482704163, + 0.009975472465157509, + 0.01532569620758295, + 0.03336722403764725, + 0.009212584234774113, + 0.006299452390521765, + 0.004468478728085756, + 0.01023385114967823, + 0.005434427410364151, + 0.008046974427998066, + 0.00647374102845788, + 0.005719121545553207, + 0.0052751475013792515, + 0.004647327587008476, + 0.0031695757061243057, + 0.006826602388173342, + 0.003837200812995434, + 0.004863142501562834, + 0.004077250137925148, + 0.004053946118801832, + 0.007526680827140808, + 0.008644115179777145, + 0.006918840575963259, + 0.002456067828461528, + 0.0024840764235705137, + 0.0030227163806557655, + 0.0032620816491544247, + 0.007074665278196335, + 0.11656824499368668, + 0.11497806757688522, + 0.015604336746037006, + 0.011690383777022362, + 0.011749559082090855, + 0.0146398413926363, + 0.026313696056604385, + 0.04372018948197365, + 0.014950837008655071, + 0.01805158518254757, + 0.0029273449908941984, + 0.0025413993280380964, + 0.0037137267645448446, + 0.003085928736254573, + 0.005344276316463947, + 0.003979561850428581, + 0.004716342780739069, + 0.00929041113704443, + 0.0040197172202169895, + 0.0026698955334722996, + 0.002660212107002735, + 0.0028473481070250273, + 0.00265332218259573, + 0.006788710597902536, + 0.01963859796524048, + 0.028627663850784302, + 0.00406998535618186, + 0.0025258769746869802, + 0.00339178997091949, + 0.003961505368351936, + 0.004009961150586605, + 0.0041726394556462765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.009778810665011406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, medium-length, static shot, taken from a level angle, captures a graying elderly couple sitting side by side in a warm, if slightly dated, living room. They sit by a large bay window, soft afternoon sunlight slanting through the clear glass, casting long, warm patches of light on the wooden floor beside them. The husband, seated on the left, wears a dark blue wool cardigan over a light gray plaid shirt. His face, deeply wrinkled by time, gazes calmly ahead, as if lost in thought. Beside him, his wife, dressed in a soft, off-white turtleneck sweater, her equally gray hair elegantly tied in a bun. Her head tilted slightly to the right, her gaze over her husband's shoulder, distant and unfocused, gazing out the window at the late autumn scene, where a few withered yellow leaves slowly drift down. Then, her chest rose and fell almost imperceptibly, her lips parted slightly, and she let out a soft and long sigh. This sigh was particularly clear in the quiet room, injecting a peaceful, nostalgic and slightly sentimental atmosphere into the whole picture.", + "seed": 22598, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_8/index_640_seed_22598_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02956099873408675, + "vlm_logprob": -0.02956099873408675, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.02956099873408675, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016709022223949432 + ], + "opposite_logprob_tokens": [ + -4.266708850860596 + ], + "logprob_reward": -0.016709022223949432, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016709022223949432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12028554826974869, + 0.040423210710287094, + 0.04394187405705452, + 0.06168811023235321, + 0.09350577741861343, + 0.06334305554628372, + 0.09029015153646469, + 0.08759793639183044, + 0.24153931438922882, + 0.15738503634929657 + ], + "credit_map": [ + 0.014475785195827484, + 0.011838109232485294, + 0.014667417854070663, + 0.015020707622170448, + 0.013418613001704216, + 0.023579521104693413, + 0.013142547570168972, + 0.01414284948259592, + 0.005610299296677113, + 0.00416037579998374, + 0.004813315812498331, + 0.004662901628762484, + 0.004891544580459595, + 0.007323050405830145, + 0.0038096862845122814, + 0.005152036435902119, + 0.004796868655830622, + 0.003624275093898177, + 0.003772102063521743, + 0.005024752579629421, + 0.004272466525435448, + 0.012449454516172409, + 0.004085690714418888, + 0.00591626251116395, + 0.009262186475098133, + 0.005639408249408007, + 0.004246652126312256, + 0.006029458716511726, + 0.006155489478260279, + 0.014709418639540672, + 0.00569080701097846, + 0.009954690001904964, + 0.014214935712516308, + 0.007764245383441448, + 0.005278404336422682, + 0.008229967206716537, + 0.009079892188310623, + 0.030114339664578438, + 0.00812685489654541, + 0.010697142221033573, + 0.015644816681742668, + 0.005019810516387224, + 0.003909259103238583, + 0.006905604153871536, + 0.0053582098335027695, + 0.012113145552575588, + 0.0061111715622246265, + 0.008281034417450428, + 0.02344394475221634, + 0.005494952201843262, + 0.003779633669182658, + 0.00877449382096529, + 0.005890708416700363, + 0.01899229921400547, + 0.008262313902378082, + 0.015651807188987732, + 0.030124500393867493, + 0.00399636197835207, + 0.003957310225814581, + 0.008025338873267174, + 0.004130132030695677, + 0.006289105862379074, + 0.007954235188663006, + 0.023120950907468796, + 0.07855430990457535, + 0.012710859067738056, + 0.010528119280934334, + 0.02611096389591694, + 0.012992392294108868, + 0.02967654913663864, + 0.026690928265452385, + 0.044275183230638504, + 0.03535965457558632, + 0.010275756008923054, + 0.008275091648101807, + 0.015349451452493668, + 0.011417021974921227, + 0.04082167148590088, + 0.01544943917542696, + 0.020436953753232956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.016709022223949432 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003754511475563 + ], + "opposite_logprob_tokens": [ + -2.3503754138946533 + ], + "logprob_reward": -0.1003754511475563, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003754511475563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30902373790740967, + 0.10509680956602097, + 0.0481264665722847, + 0.04006575420498848, + 0.08131276071071625, + 0.1389721781015396, + 0.11736304312944412, + 0.06874958425760269, + 0.06306257843971252, + 0.02822709083557129 + ], + "credit_map": [ + 0.05923086032271385, + 0.020121755078434944, + 0.04263025522232056, + 0.026638109236955643, + 0.027644095942378044, + 0.06539380550384521, + 0.031412720680236816, + 0.0359521247446537, + 0.020682740956544876, + 0.006933164782822132, + 0.012940723448991776, + 0.008040860295295715, + 0.009656733833253384, + 0.02271903119981289, + 0.011442994698882103, + 0.012680558487772942, + 0.009804648347198963, + 0.004425022751092911, + 0.004685030318796635, + 0.005437847692519426, + 0.004103308543562889, + 0.008885355666279793, + 0.005402998998761177, + 0.00538225332275033, + 0.0040507190860807896, + 0.004585674963891506, + 0.004310521297156811, + 0.00397798465564847, + 0.00908688548952341, + 0.006470540538430214, + 0.0038939369842410088, + 0.0036894932854920626, + 0.018556226044893265, + 0.006775088608264923, + 0.006859344430267811, + 0.008114421740174294, + 0.0067057423293590546, + 0.017718937247991562, + 0.0071330601349473, + 0.009449941106140614, + 0.03843380883336067, + 0.01028861291706562, + 0.008154093287885189, + 0.014330302365124226, + 0.009866883978247643, + 0.026529138907790184, + 0.013756279833614826, + 0.017613066360354424, + 0.023418312892317772, + 0.01038953848183155, + 0.011220107786357403, + 0.01154898852109909, + 0.011802599765360355, + 0.022730518132448196, + 0.01065976731479168, + 0.015593199990689754, + 0.00921530183404684, + 0.005849793553352356, + 0.009349005296826363, + 0.0067980545572936535, + 0.00812331773340702, + 0.015541421249508858, + 0.005925904028117657, + 0.00794678833335638, + 0.017166443169116974, + 0.0036774398759007454, + 0.0068259877152740955, + 0.00736983772367239, + 0.0038342890329658985, + 0.0051177507266402245, + 0.0076856291852891445, + 0.011385194957256317, + 0.004101536702364683, + 0.00228899740613997, + 0.0018723896937444806, + 0.0031211841851472855, + 0.002695844741538167, + 0.00762235326692462, + 0.0029885219410061836, + 0.003536262083798647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.1003754511475563 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420564204454422 + ], + "opposite_logprob_tokens": [ + -4.264205455780029 + ], + "logprob_reward": -0.01420564204454422, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420564204454422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12679751217365265, + 0.05998330935835838, + 0.06574098765850067, + 0.04236041009426117, + 0.09377343952655792, + 0.0738876610994339, + 0.06577882915735245, + 0.04352209344506264, + 0.3975202739238739, + 0.030635470524430275 + ], + "credit_map": [ + 0.022002462297677994, + 0.009988932870328426, + 0.016965307295322418, + 0.015361538156867027, + 0.016025923192501068, + 0.017979826778173447, + 0.014189518988132477, + 0.014284005388617516, + 0.009210756979882717, + 0.005268168170005083, + 0.005104352254420519, + 0.010271224193274975, + 0.008176452480256557, + 0.007550614420324564, + 0.008556680753827095, + 0.0058450582437217236, + 0.00941083487123251, + 0.005726820323616266, + 0.005834782496094704, + 0.009156146086752415, + 0.006415168289095163, + 0.014309817925095558, + 0.007424943149089813, + 0.007462473586201668, + 0.0076003046706318855, + 0.003495581913739443, + 0.0034115766175091267, + 0.004281623288989067, + 0.004878025967627764, + 0.009208161383867264, + 0.004080384038388729, + 0.005404751747846603, + 0.018498091027140617, + 0.007902901619672775, + 0.007489281706511974, + 0.008234160952270031, + 0.006830868311226368, + 0.02443484589457512, + 0.008709261193871498, + 0.011674024164676666, + 0.0211611520498991, + 0.005711461417376995, + 0.005309578496962786, + 0.009222013875842094, + 0.004615440499037504, + 0.011819376610219479, + 0.007139524444937706, + 0.008909116499125957, + 0.019382337108254433, + 0.0034244528505951166, + 0.004053869284689426, + 0.006278287153691053, + 0.005258467514067888, + 0.009558333083987236, + 0.005642281845211983, + 0.012180796824395657, + 0.007384470198303461, + 0.004615432117134333, + 0.0032632576767355204, + 0.00445437291637063, + 0.008200070820748806, + 0.005374187603592873, + 0.00362869119271636, + 0.006601611152291298, + 0.12951143085956573, + 0.020418815314769745, + 0.01768030785024166, + 0.044112831354141235, + 0.021251050755381584, + 0.043386269360780716, + 0.045803412795066833, + 0.07535615563392639, + 0.00320845702663064, + 0.003712350269779563, + 0.002729396801441908, + 0.005284168291836977, + 0.003375662723556161, + 0.004328268580138683, + 0.004115092568099499, + 0.003882073098793626 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.01420564204454422 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673500495031476 + ], + "opposite_logprob_tokens": [ + -5.006734848022461 + ], + "logprob_reward": -0.00673500495031476, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673500495031476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13859756290912628, + 0.05835206061601639, + 0.13553254306316376, + 0.06952772289514542, + 0.15270452201366425, + 0.13909661769866943, + 0.04083114117383957, + 0.030577372759580612, + 0.1480269730091095, + 0.0867534875869751 + ], + "credit_map": [ + 0.022297371178865433, + 0.012013989500701427, + 0.023166703060269356, + 0.016064563766121864, + 0.015789318829774857, + 0.015338420867919922, + 0.015483040362596512, + 0.018444161862134933, + 0.01005745679140091, + 0.004703918471932411, + 0.007465359289199114, + 0.004951540380716324, + 0.006906295195221901, + 0.009330649860203266, + 0.0057079000398516655, + 0.009228941984474659, + 0.020620008930563927, + 0.009744944982230663, + 0.010609968565404415, + 0.013283533975481987, + 0.012593965977430344, + 0.036769162863492966, + 0.013509674929082394, + 0.01840127818286419, + 0.011281715705990791, + 0.006500080693513155, + 0.0053056287579238415, + 0.0067377639934420586, + 0.0056517054326832294, + 0.016780992969870567, + 0.0065397704020142555, + 0.010730063542723656, + 0.030126286670565605, + 0.012586363591253757, + 0.010868907906115055, + 0.014113320969045162, + 0.010812350548803806, + 0.04015246778726578, + 0.014009082689881325, + 0.020035743713378906, + 0.03741918131709099, + 0.010891703888773918, + 0.008773421868681908, + 0.014361860230565071, + 0.009655778296291828, + 0.026754042133688927, + 0.013446916826069355, + 0.01779371127486229, + 0.010725452564656734, + 0.002898884005844593, + 0.0031839078292250633, + 0.0038649437483400106, + 0.0036924711894243956, + 0.006304747425019741, + 0.003529353067278862, + 0.00663137948140502, + 0.00950167328119278, + 0.0016738721169531345, + 0.001513194409199059, + 0.002788096433505416, + 0.0021791444160044193, + 0.0035627749748528004, + 0.0026030188892036676, + 0.006755599286407232, + 0.04691162332892418, + 0.007676682900637388, + 0.007411798927932978, + 0.016927078366279602, + 0.008418014273047447, + 0.015981290489435196, + 0.017211100086569786, + 0.027489377185702324, + 0.02382674627006054, + 0.004685929510742426, + 0.005451036151498556, + 0.00881099235266447, + 0.006080173421651125, + 0.01529903057962656, + 0.009127210825681686, + 0.013472368009388447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.00673500495031476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009779873304069042 + ], + "opposite_logprob_tokens": [ + -4.634779930114746 + ], + "logprob_reward": -0.009779873304069042, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009779873304069042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16516773402690887, + 0.05402375012636185, + 0.02529318258166313, + 0.05486610531806946, + 0.08715320378541946, + 0.05582750216126442, + 0.03857644274830818, + 0.017157383263111115, + 0.4815564751625061, + 0.02037816494703293 + ], + "credit_map": [ + 0.036096274852752686, + 0.010107379406690598, + 0.0165709238499403, + 0.013590815477073193, + 0.018301162868738174, + 0.031055118888616562, + 0.018245888873934746, + 0.021200183779001236, + 0.012650261633098125, + 0.003446518909186125, + 0.005475932732224464, + 0.003948716912418604, + 0.005156449042260647, + 0.010455147363245487, + 0.006169491447508335, + 0.006721239071339369, + 0.003902000840753317, + 0.002118464093655348, + 0.003016960807144642, + 0.00258226809091866, + 0.0025842799805104733, + 0.005980327259749174, + 0.0025301105342805386, + 0.002578774467110634, + 0.00914690736681223, + 0.005604159086942673, + 0.004817803855985403, + 0.005589334759861231, + 0.004642852582037449, + 0.012799968011677265, + 0.0045600775629282, + 0.007705010939389467, + 0.017055457457900047, + 0.007648235186934471, + 0.006722551304847002, + 0.007672485895454884, + 0.006736988667398691, + 0.022763481363654137, + 0.007844892330467701, + 0.010709122754633427, + 0.016550378873944283, + 0.0045069558545947075, + 0.003579784417524934, + 0.006542009301483631, + 0.00342917675152421, + 0.00872721429914236, + 0.005564829334616661, + 0.006927154958248138, + 0.009495897218585014, + 0.002749690320342779, + 0.0030940158758312464, + 0.0034313350915908813, + 0.003028460079804063, + 0.007185684517025948, + 0.003326887497678399, + 0.006264477968215942, + 0.002496989443898201, + 0.0019118101336061954, + 0.0013379419688135386, + 0.0023122327402234077, + 0.0017000176012516022, + 0.0032116780057549477, + 0.001843618811108172, + 0.002343094674870372, + 0.1498098522424698, + 0.025610139593482018, + 0.024929653853178024, + 0.05319894850254059, + 0.028682518750429153, + 0.05756963789463043, + 0.05444016307592392, + 0.08731559664011002, + 0.002719895914196968, + 0.0023310245014727116, + 0.0016099876957014203, + 0.0024594797287136316, + 0.0034116015303879976, + 0.003566320287063718, + 0.0017429892905056477, + 0.0025368682108819485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.009779873304069042 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, medium-length, static shot, taken from a level angle, captures a graying elderly couple sitting side by side in a warm, if slightly dated, living room. They sit by a large bay window, soft afternoon sunlight slanting through the clear glass, casting long, warm patches of light on the wooden floor beside them. The husband, seated on the left, wears a dark blue wool cardigan over a light gray plaid shirt. His face, deeply wrinkled by time, gazes calmly ahead, as if lost in thought. Beside him, his wife, dressed in a soft, off-white turtleneck sweater, her equally gray hair elegantly tied in a bun. Her head tilted slightly to the right, her gaze over her husband's shoulder, distant and unfocused, gazing out the window at the late autumn scene, where a few withered yellow leaves slowly drift down. Then, her chest rose and fell almost imperceptibly, her lips parted slightly, and she let out a soft and long sigh. This sigh was particularly clear in the quiet room, injecting a peaceful, nostalgic and slightly sentimental atmosphere into the whole picture.", + "seed": 22599, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_8/index_640_seed_22599_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.31201674286276104, + "vlm_logprob": -0.31201674286276104, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.31201674286276104, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0914839655160904 + ], + "opposite_logprob_tokens": [ + -2.4664840698242188 + ], + "logprob_reward": -0.0914839655160904, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0914839655160904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3711214065551758, + 0.09938579797744751, + 0.05936683341860771, + 0.17273247241973877, + 0.06653639674186707, + 0.05497734248638153, + 0.04003921151161194, + 0.05952596664428711, + 0.04182084649801254, + 0.0344938226044178 + ], + "credit_map": [ + 0.05133757367730141, + 0.02244751714169979, + 0.06818495690822601, + 0.05481176823377609, + 0.05472413823008537, + 0.04691099002957344, + 0.023685673251748085, + 0.04901876300573349, + 0.014032294973731041, + 0.007999331690371037, + 0.01170207280665636, + 0.010525689460337162, + 0.0200280100107193, + 0.011423198506236076, + 0.012107434682548046, + 0.011567768640816212, + 0.01551542803645134, + 0.0034409472718834877, + 0.013162615709006786, + 0.005559383891522884, + 0.004988906905055046, + 0.007074717897921801, + 0.0038839932531118393, + 0.005740837659686804, + 0.044376954436302185, + 0.007572102826088667, + 0.05596572533249855, + 0.025321505963802338, + 0.008640427142381668, + 0.014864125289022923, + 0.006262404844164848, + 0.009729207493364811, + 0.011464730836451054, + 0.004878789186477661, + 0.006840051151812077, + 0.012638417072594166, + 0.006911808624863625, + 0.007677147630602121, + 0.0062634581699967384, + 0.009861985221505165, + 0.014505930244922638, + 0.003051247913390398, + 0.0038019800558686256, + 0.016204996034502983, + 0.0038009611889719963, + 0.0049283928237855434, + 0.003967719152569771, + 0.004716107621788979, + 0.0069271172396838665, + 0.003046658355742693, + 0.0028855400159955025, + 0.006790578830987215, + 0.003899343777447939, + 0.006033236160874367, + 0.004767803009599447, + 0.005688928999006748, + 0.01714026927947998, + 0.003662744304165244, + 0.011530021205544472, + 0.007251670118421316, + 0.005425041075795889, + 0.00378618435934186, + 0.004156226757913828, + 0.006573804188519716, + 0.007408819627016783, + 0.0027483077719807625, + 0.009786337614059448, + 0.006179688032716513, + 0.004775941371917725, + 0.0026442483067512512, + 0.004898645915091038, + 0.003378857159987092, + 0.007658890448510647, + 0.0022562702652066946, + 0.008657624013721943, + 0.004249654710292816, + 0.002440035343170166, + 0.003521397477015853, + 0.002299200277775526, + 0.003410749137401581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0914839655160904 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.4063993692398071 + ], + "opposite_logprob_tokens": [ + -0.28139936923980713 + ], + "logprob_reward": -1.4063993692398071, + "yes_no_margin_reward": -1.125, + "token_credit_reward": -1.4063993692398071, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2277892827987671, + 0.05824246257543564, + 0.022745834663510323, + 0.5037850737571716, + 0.032616350799798965, + 0.06434433162212372, + 0.03160369396209717, + 0.024255020543932915, + 0.024424854665994644, + 0.010193144902586937 + ], + "credit_map": [ + 0.030101200565695763, + 0.019410911947488785, + 0.031148897483944893, + 0.03316282480955124, + 0.037199798971414566, + 0.03336970508098602, + 0.01679830625653267, + 0.02659761905670166, + 0.01614070124924183, + 0.0033239710610359907, + 0.013788038864731789, + 0.006570817437022924, + 0.0037766732275485992, + 0.006025474052876234, + 0.003182119457051158, + 0.005434664897620678, + 0.0046391477808356285, + 0.0018355996580794454, + 0.001938564470037818, + 0.0030378876253962517, + 0.003721802495419979, + 0.003321566618978977, + 0.002132217865437269, + 0.002119049895554781, + 0.12544898688793182, + 0.01645677350461483, + 0.17617909610271454, + 0.07719987630844116, + 0.022988183423876762, + 0.037403035908937454, + 0.02061581425368786, + 0.02749328315258026, + 0.004596699494868517, + 0.0025181611999869347, + 0.002811633748933673, + 0.005216922610998154, + 0.005351388826966286, + 0.0035635274834930897, + 0.0035388797987252474, + 0.005019136238843203, + 0.010012990795075893, + 0.005130086559802294, + 0.006680956110358238, + 0.011494293808937073, + 0.0071580554358661175, + 0.007188895251601934, + 0.00747170252725482, + 0.009207352995872498, + 0.005042398348450661, + 0.002178916474804282, + 0.003064418211579323, + 0.004843358416110277, + 0.00262645841576159, + 0.003963724710047245, + 0.004059045575559139, + 0.005825372412800789, + 0.0024202000349760056, + 0.0016996832564473152, + 0.0041246572509408, + 0.0045285397209227085, + 0.002095821313560009, + 0.004459322430193424, + 0.0024883730802685022, + 0.0024384232237935066, + 0.006138359662145376, + 0.0017188049387186766, + 0.0045617311261594296, + 0.002588064642623067, + 0.002432757755741477, + 0.002751797204837203, + 0.0019584421534091234, + 0.0022748957853764296, + 0.001254027127288282, + 0.0007469403790310025, + 0.0016992773162201047, + 0.00177116715349257, + 0.00105003931093961, + 0.0015006581088528037, + 0.0011436586501076818, + 0.0010273763909935951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -1.4063993692398071 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637173980474472 + ], + "opposite_logprob_tokens": [ + -3.651371717453003 + ], + "logprob_reward": -0.02637173980474472, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637173980474472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2478363811969757, + 0.08206749707460403, + 0.037017446011304855, + 0.39653560519218445, + 0.02588666044175625, + 0.0400623120367527, + 0.04648582264780998, + 0.06965657323598862, + 0.03892149403691292, + 0.0155301783233881 + ], + "credit_map": [ + 0.03541816025972366, + 0.01678270846605301, + 0.03892802074551582, + 0.03615166246891022, + 0.04031955823302269, + 0.03446885943412781, + 0.016172893345355988, + 0.02959454618394375, + 0.02125534974038601, + 0.007709403522312641, + 0.01598922163248062, + 0.010243166238069534, + 0.006710805464535952, + 0.00492965430021286, + 0.007579319179058075, + 0.007650583051145077, + 0.009956810623407364, + 0.002838006941601634, + 0.007446688134223223, + 0.004532633814960718, + 0.002494827378541231, + 0.00397015456110239, + 0.001949794590473175, + 0.0038285318296402693, + 0.10431626439094543, + 0.013891389593482018, + 0.13977870345115662, + 0.059765443205833435, + 0.019667236134409904, + 0.029532870277762413, + 0.012172170914709568, + 0.017411550506949425, + 0.004612818360328674, + 0.001973268808797002, + 0.0023473966866731644, + 0.006618310697376728, + 0.0027901101857423782, + 0.0026447842828929424, + 0.0022644358687102795, + 0.0026355376467108727, + 0.005690018180757761, + 0.003024173201993108, + 0.003606396960094571, + 0.005647842772305012, + 0.0049774423241615295, + 0.005098703317344189, + 0.006283099763095379, + 0.005734637379646301, + 0.009414736181497574, + 0.0030359060037881136, + 0.004614303819835186, + 0.007355577312409878, + 0.004004657734185457, + 0.005119824782013893, + 0.005630921106785536, + 0.007309897802770138, + 0.016895608976483345, + 0.004859522450715303, + 0.005953030660748482, + 0.021694043651223183, + 0.005022984929382801, + 0.004235065542161465, + 0.005345035810023546, + 0.005651282146573067, + 0.006667226552963257, + 0.002282527508214116, + 0.009910494089126587, + 0.006204386241734028, + 0.003183931577950716, + 0.004301242996007204, + 0.0032621973659843206, + 0.003109490033239126, + 0.0019132897723466158, + 0.000956812989898026, + 0.0031522740609943867, + 0.0030211471021175385, + 0.001161637483164668, + 0.0023477510549128056, + 0.0012367760064080358, + 0.0017404917161911726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.02637173980474472 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543593533337116 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543593533337116, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543593533337116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26874297857284546, + 0.09480130672454834, + 0.08027467876672745, + 0.3688633441925049, + 0.04137866199016571, + 0.03573909401893616, + 0.046975746750831604, + 0.025562576949596405, + 0.018407927826046944, + 0.019253648817539215 + ], + "credit_map": [ + 0.035146381705999374, + 0.016549866646528244, + 0.03466113284230232, + 0.0380672924220562, + 0.04506172984838486, + 0.04715389758348465, + 0.019357135519385338, + 0.032745566219091415, + 0.023562446236610413, + 0.006913036573678255, + 0.018779780715703964, + 0.009638485498726368, + 0.010038545355200768, + 0.009017560631036758, + 0.006009015720337629, + 0.010842435993254185, + 0.02078353986144066, + 0.00410952465608716, + 0.014020180329680443, + 0.00855289213359356, + 0.008248323574662209, + 0.005588823929429054, + 0.009427089244127274, + 0.009544305503368378, + 0.09546174854040146, + 0.011821754276752472, + 0.13389460742473602, + 0.05781760439276695, + 0.01574207842350006, + 0.025299694389104843, + 0.012491434812545776, + 0.016334449872374535, + 0.007938141003251076, + 0.002763359108939767, + 0.003851063083857298, + 0.0065953112207353115, + 0.005995791405439377, + 0.006278298329561949, + 0.0034817084670066833, + 0.004474993329495192, + 0.007825162261724472, + 0.0026290223468095064, + 0.003839220618829131, + 0.005804554559290409, + 0.0033667574170976877, + 0.005404206924140453, + 0.003060505259782076, + 0.0038096627686172724, + 0.007851215079426765, + 0.003016316331923008, + 0.007141464855521917, + 0.005851223133504391, + 0.00423884391784668, + 0.005050488281995058, + 0.005896600428968668, + 0.007929595187306404, + 0.005321477074176073, + 0.001850063563324511, + 0.0026319939643144608, + 0.003942724317312241, + 0.003289735410362482, + 0.0040003592148423195, + 0.002213425701484084, + 0.002312798285856843, + 0.0031195615883916616, + 0.001222324324771762, + 0.0034648797009140253, + 0.0025615764316171408, + 0.002004685578867793, + 0.002959060948342085, + 0.0012985357316210866, + 0.001777304452843964, + 0.0033255021553486586, + 0.001322339870966971, + 0.003944753669202328, + 0.0026391418650746346, + 0.001585605787113309, + 0.0030691989231854677, + 0.0011974762892350554, + 0.002169631654396653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.012543593533337116 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two elderly people with gray hair sitting side by side in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a living room with a large bay window and wooden floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the husband wearing a dark blue wool cardigan over a light gray plaid shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the wife wearing a soft off-white turtleneck sweater with her hair tied in a bun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static medium-length view taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023285046219825745 + ], + "opposite_logprob_tokens": [ + -3.773285150527954 + ], + "logprob_reward": -0.023285046219825745, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023285046219825745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23787739872932434, + 0.05396449193358421, + 0.024554211646318436, + 0.44533151388168335, + 0.05164018273353577, + 0.03683329373598099, + 0.03827810660004616, + 0.04657592251896858, + 0.0315261073410511, + 0.03341872990131378 + ], + "credit_map": [ + 0.030316924676299095, + 0.02052868716418743, + 0.026809541508555412, + 0.035172004252672195, + 0.039074793457984924, + 0.04263899847865105, + 0.016654297709465027, + 0.02668215148150921, + 0.009916133247315884, + 0.005294072441756725, + 0.006260555237531662, + 0.006443831138312817, + 0.0074605937115848064, + 0.005856307223439217, + 0.0059125772677361965, + 0.006820418406277895, + 0.0026811142452061176, + 0.0018823713762685657, + 0.002966792555525899, + 0.0036599677987396717, + 0.0025619477964937687, + 0.004038636106997728, + 0.0034118571784347296, + 0.003351524705067277, + 0.11269810050725937, + 0.014646205119788647, + 0.15604755282402039, + 0.06735982745885849, + 0.019856683909893036, + 0.03404923528432846, + 0.017018998041749, + 0.02365492843091488, + 0.0086613018065691, + 0.003946052398532629, + 0.006064739543944597, + 0.007441820111125708, + 0.0067764041014015675, + 0.006089095026254654, + 0.004727385472506285, + 0.007933384738862514, + 0.005515779368579388, + 0.002417944371700287, + 0.004232615698128939, + 0.008934096433222294, + 0.004217421170324087, + 0.0037319399416446686, + 0.003274354850873351, + 0.0045091453939676285, + 0.0066650789231061935, + 0.002730262465775013, + 0.0033365795388817787, + 0.0073645273223519325, + 0.0030177123844623566, + 0.0036998915020376444, + 0.004484912846237421, + 0.006979144178330898, + 0.009587975218892097, + 0.00243498501367867, + 0.004718908108770847, + 0.012749403715133667, + 0.0036675306037068367, + 0.00454431027173996, + 0.004974173381924629, + 0.0038986350409686565, + 0.002922316314652562, + 0.0015672375448048115, + 0.007106865290552378, + 0.0075081586837768555, + 0.003048066981136799, + 0.0026061125099658966, + 0.003943934105336666, + 0.0028234159108251333, + 0.0037889897357672453, + 0.0025563561357557774, + 0.006909746211022139, + 0.006648970302194357, + 0.0028117455076426268, + 0.005399527959525585, + 0.002216774271801114, + 0.0030866211745887995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.023285046219825745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_9/index_tensor([7884, 7884])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_9/index_tensor([7884, 7884])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..929fc54df3e224a7cc6c58a4117d2bb020eed3b9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_9/index_tensor([7884, 7884])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "highly dynamic, low-angle tracking shot closely follows a magical sphere imbued with unstable energy as it violently hurtles through a majestic and gloomy ancient grotto. At the sphere's core lies a pulsating, blinding purple singularity, barely contained by a translucent crystal shell riddled with spiderweb-like cracks. Energy continuously emanates from the fissures. The sphere accelerates violently in an utterly unpredictable trajectory, sometimes hurtling along the rough ground, kicking up clouds of dust, and sometimes zooming vertically at a physics-defying angle. The camera struggles to capture its trajectory, shrouded in violent shakiness and slight lag, creating a sense of handheld urgency. Brilliant arcs of cyan electricity burst from the sphere's cracks, lashing the air like whips, trace exaggerated paths marked by noticeable motion blur and afterimages. As these arcs slam into the massive, weathered stone pillars surrounding it, they instantly blast away vast chunks of rock and plumes of dust, each impact imbued with weight. The entire picture is full of the exaggerated perspective, dynamic deformation and heavyweight impact unique to Studio Bones' animation style. The highly saturated cold light emitted by the magic ball itself is the only light source, casting long, wildly dancing shadows on the dark background, creating an extremely tense atmosphere full of primitive power and destructive aesthetics.", + "seed": 14030, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_9/index_7884_seed_14030_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03950620978139341, + "vlm_logprob": -0.03950620978139341, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.03950620978139341, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005148841999471188 + ], + "opposite_logprob_tokens": [ + -6.255148887634277 + ], + "logprob_reward": -0.005148841999471188, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.005148841999471188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05581739544868469, + 0.04982221499085426, + 0.021972063928842545, + 0.021212555468082428, + 0.05759093165397644, + 0.03266751766204834, + 0.02933434583246708, + 0.6086190938949585, + 0.098472960293293, + 0.02449086681008339 + ], + "credit_map": [ + 0.007601373363286257, + 0.0046388497576117516, + 0.006050894968211651, + 0.005967221688479185, + 0.003939320798963308, + 0.009496297687292099, + 0.005875737871974707, + 0.012247705832123756, + 0.007497385144233704, + 0.004428300075232983, + 0.008796642534434795, + 0.004872382618486881, + 0.004923178348690271, + 0.009408851154148579, + 0.004865577910095453, + 0.005029901396483183, + 0.0031015537679195404, + 0.002794477390125394, + 0.004007226321846247, + 0.0026888723950833082, + 0.0032312613911926746, + 0.002994318027049303, + 0.0015505336923524737, + 0.0016038240864872932, + 0.002713280962780118, + 0.0022805335465818644, + 0.004164742771536112, + 0.0029188848566263914, + 0.0020731566473841667, + 0.00267525902017951, + 0.0014885187847539783, + 0.002898180391639471, + 0.01333511620759964, + 0.008471350185573101, + 0.008504713885486126, + 0.00850747898221016, + 0.0046274601481854916, + 0.007129704114049673, + 0.003676721593365073, + 0.003338392125442624, + 0.003798112738877535, + 0.004485058132559061, + 0.007422011345624924, + 0.005385365802794695, + 0.0021931896917521954, + 0.0031822114251554012, + 0.0022750592324882746, + 0.003926513716578484, + 0.004911267198622227, + 0.0025881913024932146, + 0.004690887406468391, + 0.00458827568218112, + 0.0017747674137353897, + 0.002228369703516364, + 0.002663371851667762, + 0.005889216437935829, + 0.07116039842367172, + 0.07625830918550491, + 0.06456463038921356, + 0.06728938221931458, + 0.08869653195142746, + 0.06912291049957275, + 0.10887443274259567, + 0.06265252083539963, + 0.008900130167603493, + 0.0045173149555921555, + 0.012577880173921585, + 0.0035555448848754168, + 0.006542947608977556, + 0.024622302502393723, + 0.006592439487576485, + 0.03116440586745739, + 0.0023724338971078396, + 0.0019584698602557182, + 0.004868145566433668, + 0.0023407277185469866, + 0.0018524292390793562, + 0.0034355565439909697, + 0.003750989679247141, + 0.003912116400897503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.005148841999471188 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989394068717957 + ], + "opposite_logprob_tokens": [ + -1.804893970489502 + ], + "logprob_reward": -0.17989394068717957, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989394068717957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03969235345721245, + 0.04764111340045929, + 0.06131497770547867, + 0.03453211113810539, + 0.10494834184646606, + 0.01993528939783573, + 0.02275848388671875, + 0.5536363124847412, + 0.10118505358695984, + 0.014355977065861225 + ], + "credit_map": [ + 0.005884022917598486, + 0.002781888237223029, + 0.005398242734372616, + 0.004014912061393261, + 0.002329390961676836, + 0.00786592997610569, + 0.0034428942017257214, + 0.007975075393915176, + 0.007542422041296959, + 0.003563222009688616, + 0.0075043062679469585, + 0.00374691397882998, + 0.004353863652795553, + 0.010094838216900826, + 0.004904146771878004, + 0.0059313965030014515, + 0.008914319798350334, + 0.007006686180830002, + 0.010945575311779976, + 0.005596477072685957, + 0.008777188137173653, + 0.008390532806515694, + 0.005228196270763874, + 0.0064559997990727425, + 0.004726979415863752, + 0.0047187283635139465, + 0.006287792231887579, + 0.003897010115906596, + 0.005167043302208185, + 0.004259783308953047, + 0.0029600535053759813, + 0.0025147199630737305, + 0.02706301026046276, + 0.014572887681424618, + 0.02219311147928238, + 0.011382204480469227, + 0.005377271678298712, + 0.015599129721522331, + 0.004392695613205433, + 0.004368023481220007, + 0.0028528112452477217, + 0.0030350040178745985, + 0.002985874190926552, + 0.0035227707121521235, + 0.0019550281576812267, + 0.0019811096135526896, + 0.001576517941430211, + 0.0020261723548173904, + 0.0028746898751705885, + 0.0028730195481330156, + 0.0036558592692017555, + 0.00326450914144516, + 0.0024940255098044872, + 0.0020083996932953596, + 0.002337691141292453, + 0.003250288777053356, + 0.06544384360313416, + 0.06997036933898926, + 0.058648549020290375, + 0.05959455668926239, + 0.08268506079912186, + 0.0624123178422451, + 0.09824095666408539, + 0.05664066597819328, + 0.009917066432535648, + 0.007033851929008961, + 0.01714225858449936, + 0.0058684684336185455, + 0.010544384829699993, + 0.020597605034708977, + 0.010612591169774532, + 0.019468823447823524, + 0.0012260394869372249, + 0.001271055662073195, + 0.0021346083376556635, + 0.0019790269434452057, + 0.0013253182405605912, + 0.002411443507298827, + 0.0018025698373094201, + 0.0022059150505810976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.17989394068717957 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036206187214702368 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036206187214702368, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036206187214702368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06477420032024384, + 0.043572425842285156, + 0.05413500964641571, + 0.07535575330257416, + 0.047290779650211334, + 0.04582354798913002, + 0.036714449524879456, + 0.5264236330986023, + 0.07171032577753067, + 0.03419983759522438 + ], + "credit_map": [ + 0.011722252704203129, + 0.004198901355266571, + 0.010154125280678272, + 0.0064861406572163105, + 0.004092207178473473, + 0.01077886950224638, + 0.005271814297884703, + 0.012069893069565296, + 0.00707461591809988, + 0.003509768284857273, + 0.007775334641337395, + 0.003238176926970482, + 0.004038678947836161, + 0.009016823954880238, + 0.0036997743882238865, + 0.005219253711402416, + 0.007189675699919462, + 0.005217282567173243, + 0.009768247604370117, + 0.005600746255367994, + 0.005040676333010197, + 0.007841212674975395, + 0.0056945341639220715, + 0.0077826338820159435, + 0.012075375765562057, + 0.010977666825056076, + 0.012166408821940422, + 0.008102736435830593, + 0.009101368486881256, + 0.008889039978384972, + 0.006110671442002058, + 0.007932486943900585, + 0.008337341248989105, + 0.010655289515852928, + 0.008122989907860756, + 0.005295582115650177, + 0.003903313772752881, + 0.00586107699200511, + 0.0028418898582458496, + 0.002273297170177102, + 0.006780921947211027, + 0.007087112404406071, + 0.009608577936887741, + 0.009084575809538364, + 0.0024375685025006533, + 0.004591426812112331, + 0.001992622157558799, + 0.0042407396249473095, + 0.003615563502535224, + 0.003527730470523238, + 0.00784600805491209, + 0.008177632465958595, + 0.0018478527199476957, + 0.0030226735398173332, + 0.002966905478388071, + 0.005710082594305277, + 0.0647396668791771, + 0.06533422321081161, + 0.05534164607524872, + 0.05503762140870094, + 0.07721731066703796, + 0.0619792565703392, + 0.09202273190021515, + 0.054751161485910416, + 0.00591941038146615, + 0.005465909373015165, + 0.01471990067511797, + 0.006557028740644455, + 0.00793052650988102, + 0.007463681977242231, + 0.008990341797471046, + 0.01466352678835392, + 0.0013702238211408257, + 0.001335731940343976, + 0.0029574402142316103, + 0.0013208312448114157, + 0.0012642095098271966, + 0.012067603878676891, + 0.0020176395773887634, + 0.011866155080497265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.0036206187214702368 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003616699017584324 + ], + "opposite_logprob_tokens": [ + -5.628616809844971 + ], + "logprob_reward": -0.003616699017584324, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003616699017584324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027667202055454254, + 0.03942149877548218, + 0.013864928856492043, + 0.014846434816718102, + 0.030855439603328705, + 0.046926453709602356, + 0.045214612036943436, + 0.6422103047370911, + 0.10709355771541595, + 0.031899575144052505 + ], + "credit_map": [ + 0.0033709246199578047, + 0.002431049942970276, + 0.004023300018161535, + 0.002964382292702794, + 0.0018564399797469378, + 0.004451551474630833, + 0.0025102165527641773, + 0.006059335079044104, + 0.006274158600717783, + 0.003326058154925704, + 0.006138904020190239, + 0.003146733157336712, + 0.004083896055817604, + 0.0076399147510528564, + 0.003773946315050125, + 0.005037885159254074, + 0.0019783182069659233, + 0.0014667969662696123, + 0.002978239208459854, + 0.0016384063055738807, + 0.0015289357397705317, + 0.001551255933009088, + 0.0014046835713088512, + 0.0013182919938117266, + 0.002137055853381753, + 0.0017535302322357893, + 0.00273970328271389, + 0.0018196073360741138, + 0.0017603944288566709, + 0.0021139183081686497, + 0.0010705834720283747, + 0.0014516423689201474, + 0.0050492859445512295, + 0.0066399481147527695, + 0.005489029455929995, + 0.0030846665613353252, + 0.002464823191985488, + 0.004417921416461468, + 0.0018808352760970592, + 0.001828929060138762, + 0.0062052286230027676, + 0.005814229138195515, + 0.00693126767873764, + 0.008904616348445415, + 0.004315467085689306, + 0.005382446572184563, + 0.0038413829170167446, + 0.005531818140298128, + 0.005160795524716377, + 0.004615154583007097, + 0.009316465817391872, + 0.006931541487574577, + 0.0024721126537770033, + 0.004088393412530422, + 0.0038775356952100992, + 0.008752611465752125, + 0.07681898027658463, + 0.08002430200576782, + 0.06785609573125839, + 0.06935945898294449, + 0.0933520570397377, + 0.07347623258829117, + 0.11543706059455872, + 0.06588605791330338, + 0.008119826205074787, + 0.006920470390468836, + 0.013228357769548893, + 0.007187888026237488, + 0.010744073428213596, + 0.022930152714252472, + 0.008249204605817795, + 0.0297135841101408, + 0.0018756018253043294, + 0.0012553833657875657, + 0.004023286979645491, + 0.0022831575479358435, + 0.0020242591854184866, + 0.008296238258481026, + 0.004267081618309021, + 0.0078745661303401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.003616699017584324 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525094848126173 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.00525094848126173, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525094848126173, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0323106050491333, + 0.0366818793118, + 0.013475930318236351, + 0.022529449313879013, + 0.047309510409832, + 0.026269646361470222, + 0.03073597513139248, + 0.6252163648605347, + 0.13611328601837158, + 0.02935740351676941 + ], + "credit_map": [ + 0.0051785907708108425, + 0.0024984986521303654, + 0.004156292881816626, + 0.003661368740722537, + 0.001719790045171976, + 0.0055504729971289635, + 0.003061651485040784, + 0.006483940873295069, + 0.006318851374089718, + 0.0028084188234061003, + 0.006163111887872219, + 0.0027782837860286236, + 0.003097012871876359, + 0.008049946278333664, + 0.0032509732991456985, + 0.004215280059725046, + 0.0018748624715954065, + 0.0018268927233293653, + 0.0021113348193466663, + 0.0011329477420076728, + 0.001749981427565217, + 0.0011941965203732252, + 0.00140056770760566, + 0.002185146789997816, + 0.0030257701873779297, + 0.0026889611035585403, + 0.005825705826282501, + 0.0022898989263921976, + 0.0026451419107615948, + 0.002484851749613881, + 0.0011198207503184676, + 0.0024492985103279352, + 0.012882803566753864, + 0.006330491509288549, + 0.0111552057787776, + 0.004965486936271191, + 0.002544594695791602, + 0.005173380486667156, + 0.002348596230149269, + 0.0019089523702859879, + 0.0037507428787648678, + 0.003434410085901618, + 0.004181120544672012, + 0.004739248659461737, + 0.0019528509583324194, + 0.0028955319430679083, + 0.001908835954964161, + 0.0034069048706442118, + 0.003473778488114476, + 0.002580307424068451, + 0.00723665114492178, + 0.0059156259521842, + 0.0016368691576644778, + 0.0024386884178966284, + 0.0020565844606608152, + 0.00539746880531311, + 0.0755012035369873, + 0.07830190658569336, + 0.0653463825583458, + 0.06654861569404602, + 0.0922563448548317, + 0.07200104743242264, + 0.11086215823888779, + 0.0643986389040947, + 0.014265408739447594, + 0.0072516934014856815, + 0.01574452593922615, + 0.004805137403309345, + 0.008701295591890812, + 0.03555642440915108, + 0.008913815021514893, + 0.04087497666478157, + 0.0011467778822407126, + 0.0012868897756561637, + 0.00196573487482965, + 0.0015062687452882528, + 0.0012316206702962518, + 0.01018972136080265, + 0.00200862274505198, + 0.010021770372986794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.00525094848126173 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly dynamic, low-angle tracking shot closely follows a magical sphere imbued with unstable energy as it violently hurtles through a majestic and gloomy ancient grotto. At the sphere's core lies a pulsating, blinding purple singularity, barely contained by a translucent crystal shell riddled with spiderweb-like cracks. Energy continuously emanates from the fissures. The sphere accelerates violently in an utterly unpredictable trajectory, sometimes hurtling along the rough ground, kicking up clouds of dust, and sometimes zooming vertically at a physics-defying angle. The camera struggles to capture its trajectory, shrouded in violent shakiness and slight lag, creating a sense of handheld urgency. Brilliant arcs of cyan electricity burst from the sphere's cracks, lashing the air like whips, trace exaggerated paths marked by noticeable motion blur and afterimages. As these arcs slam into the massive, weathered stone pillars surrounding it, they instantly blast away vast chunks of rock and plumes of dust, each impact imbued with weight. The entire picture is full of the exaggerated perspective, dynamic deformation and heavyweight impact unique to Studio Bones' animation style. The highly saturated cold light emitted by the magic ball itself is the only light source, casting long, wildly dancing shadows on the dark background, creating an extremely tense atmosphere full of primitive power and destructive aesthetics.", + "seed": 14031, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_9/index_7884_seed_14031_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09878610353916883, + "vlm_logprob": -0.09878610353916883, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.09878610353916883, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004217064939439297 + ], + "opposite_logprob_tokens": [ + -6.379217147827148 + ], + "logprob_reward": -0.004217064939439297, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.004217064939439297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10204123705625534, + 0.059107035398483276, + 0.07343091070652008, + 0.07510455697774887, + 0.05428198724985123, + 0.1798611730337143, + 0.06433345377445221, + 0.0420348159968853, + 0.28395771980285645, + 0.06584709137678146 + ], + "credit_map": [ + 0.010619708336889744, + 0.008494754321873188, + 0.009473110549151897, + 0.012745972722768784, + 0.010926321148872375, + 0.015528058633208275, + 0.015570432879030704, + 0.018682880327105522, + 0.005364889744669199, + 0.004694120027124882, + 0.003975863568484783, + 0.005463991314172745, + 0.0036418363451957703, + 0.008892755024135113, + 0.008449336513876915, + 0.018624242395162582, + 0.005330090411007404, + 0.007135193329304457, + 0.010234752669930458, + 0.005737712141126394, + 0.0061713014729321, + 0.013793590478599072, + 0.010844345204532146, + 0.014183925464749336, + 0.007877352647483349, + 0.005479628685861826, + 0.012147779576480389, + 0.006920855958014727, + 0.006987829227000475, + 0.012875082902610302, + 0.008909689262509346, + 0.013906332664191723, + 0.0075024086982011795, + 0.005337590351700783, + 0.010039682500064373, + 0.004526939708739519, + 0.008030234836041927, + 0.007800179533660412, + 0.005389613565057516, + 0.005655334796756506, + 0.045556169003248215, + 0.008355108089745045, + 0.00824800692498684, + 0.00594784552231431, + 0.011321382597088814, + 0.07232815027236938, + 0.0072020855732262135, + 0.020902419462800026, + 0.006797882728278637, + 0.010226921178400517, + 0.005932365544140339, + 0.006945518311113119, + 0.009328463114798069, + 0.007909592241048813, + 0.007743940223008394, + 0.00944877415895462, + 0.0075640869326889515, + 0.0031921674963086843, + 0.005051040556281805, + 0.003774479264393449, + 0.002930290997028351, + 0.006161806173622608, + 0.007495280355215073, + 0.005865667946636677, + 0.02210257574915886, + 0.08215051144361496, + 0.014388920739293098, + 0.009651225060224533, + 0.05864693224430084, + 0.018927207216620445, + 0.014088049530982971, + 0.06400230526924133, + 0.006769421044737101, + 0.010750286281108856, + 0.009604550898075104, + 0.005025285761803389, + 0.007953488267958164, + 0.015244829468429089, + 0.0047577423974871635, + 0.00574148865416646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.004217064939439297 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47425222396850586 + ], + "opposite_logprob_tokens": [ + -0.9742522239685059 + ], + "logprob_reward": -0.47425222396850586, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47425222396850586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04006928205490112, + 0.012802139855921268, + 0.022609751671552658, + 0.02511163055896759, + 0.03510069474577904, + 0.1732187122106552, + 0.11205091327428818, + 0.027918629348278046, + 0.5284043550491333, + 0.02271389588713646 + ], + "credit_map": [ + 0.0038350368849933147, + 0.002528281882405281, + 0.0044963164255023, + 0.004714274778962135, + 0.0038754509296268225, + 0.00836351327598095, + 0.004810646176338196, + 0.007445759139955044, + 0.0012646977556869388, + 0.0015223391819745302, + 0.0012644522357732058, + 0.0019075568998232484, + 0.0011340468190610409, + 0.001704688649624586, + 0.0016369146760553122, + 0.002367443637922406, + 0.0019199218368157744, + 0.0029512939509004354, + 0.0023370736744254827, + 0.0021067578345537186, + 0.0021020686253905296, + 0.0037758098915219307, + 0.0029470212757587433, + 0.004469806328415871, + 0.0019461188931018114, + 0.0024030497297644615, + 0.002723115961998701, + 0.002393044764176011, + 0.002664815867319703, + 0.00380243849940598, + 0.0027874805964529514, + 0.006391565315425396, + 0.0048192390240728855, + 0.00445162458345294, + 0.005509172100573778, + 0.0038739731535315514, + 0.004000133369117975, + 0.0044779907912015915, + 0.0034928051754832268, + 0.004475755617022514, + 0.04551352560520172, + 0.008792982436716557, + 0.004572597797960043, + 0.00664164125919342, + 0.009431700222194195, + 0.07300382107496262, + 0.005984785966575146, + 0.0192776620388031, + 0.01070738211274147, + 0.016250336542725563, + 0.014106969349086285, + 0.012385746464133263, + 0.016151074320077896, + 0.015475592575967312, + 0.013672464527189732, + 0.01330133993178606, + 0.0038536880165338516, + 0.004287610296159983, + 0.002250723773613572, + 0.0032961172983050346, + 0.004040224943310022, + 0.0038554712664335966, + 0.0033424650318920612, + 0.0029923277907073498, + 0.039937086403369904, + 0.1585196703672409, + 0.024311143904924393, + 0.0159356277436018, + 0.1159990206360817, + 0.036515556275844574, + 0.025057613849639893, + 0.11212864518165588, + 0.0036695438902825117, + 0.003822196973487735, + 0.002439535455778241, + 0.0018347856821492314, + 0.001914347056299448, + 0.003952210769057274, + 0.002069502603262663, + 0.0030117719434201717 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 6 + ], + "selected_reward": -0.47425222396850586 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003197559155523777 + ], + "opposite_logprob_tokens": [ + -5.75319766998291 + ], + "logprob_reward": -0.003197559155523777, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003197559155523777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11355087161064148, + 0.04684826731681824, + 0.059499919414520264, + 0.05014900863170624, + 0.06947139650583267, + 0.1352897733449936, + 0.14972960948944092, + 0.10344310104846954, + 0.22869469225406647, + 0.04332330450415611 + ], + "credit_map": [ + 0.010440139099955559, + 0.006538156419992447, + 0.01035554613918066, + 0.009626070037484169, + 0.01632719859480858, + 0.02191517874598503, + 0.016468139365315437, + 0.02188044600188732, + 0.00397889781743288, + 0.004216112196445465, + 0.00357256387360394, + 0.00382818141952157, + 0.003578274045139551, + 0.006704727187752724, + 0.008197521790862083, + 0.012771990150213242, + 0.00429224269464612, + 0.0068467590026557446, + 0.008758877404034138, + 0.005441707093268633, + 0.005441383458673954, + 0.010036086663603783, + 0.007635719608515501, + 0.011047149077057838, + 0.005756807513535023, + 0.0031709170434623957, + 0.009399279952049255, + 0.004710998851805925, + 0.004259543027728796, + 0.008564499206840992, + 0.004889339674264193, + 0.009397625923156738, + 0.013064373284578323, + 0.006326667964458466, + 0.011966575868427753, + 0.004959185607731342, + 0.009512681514024734, + 0.011124677024781704, + 0.00516869081184268, + 0.007348550483584404, + 0.029678087681531906, + 0.007769464049488306, + 0.008932377211749554, + 0.0062629906460642815, + 0.009226138703525066, + 0.04837391525506973, + 0.00844082422554493, + 0.016605980694293976, + 0.01689421571791172, + 0.016369260847568512, + 0.0197494775056839, + 0.013716978020966053, + 0.019160185009241104, + 0.020367808640003204, + 0.023282615467905998, + 0.020189078524708748, + 0.01771753653883934, + 0.007269012741744518, + 0.013436662033200264, + 0.0098803685978055, + 0.0057697594165802, + 0.015284746885299683, + 0.0191402155905962, + 0.014944806694984436, + 0.021931974217295647, + 0.048763375729322433, + 0.024509571492671967, + 0.010601633228361607, + 0.047858282923698425, + 0.028426136821508408, + 0.014441551640629768, + 0.032162200659513474, + 0.005419449880719185, + 0.006438028533011675, + 0.005134696140885353, + 0.0025710761547088623, + 0.007169942371547222, + 0.009165213443338871, + 0.0024465194437652826, + 0.004978382494300604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 5 + ], + "selected_reward": -0.003197559155523777 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046380325220525265 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.0046380325220525265, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046380325220525265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10833358764648438, + 0.06327775120735168, + 0.0533769391477108, + 0.03724164143204689, + 0.11727234721183777, + 0.05848104506731033, + 0.05719096586108208, + 0.0979132354259491, + 0.12823422253131866, + 0.2786782383918762 + ], + "credit_map": [ + 0.010220245458185673, + 0.0059836250729858875, + 0.010532159358263016, + 0.009586554020643234, + 0.01198028028011322, + 0.01927797682583332, + 0.022426800802350044, + 0.01832595467567444, + 0.0046364739537239075, + 0.0064793056808412075, + 0.00642297463491559, + 0.0055347615852952, + 0.004118934273719788, + 0.008280419744551182, + 0.011829749681055546, + 0.015975138172507286, + 0.004504731856286526, + 0.006282045040279627, + 0.005168365780264139, + 0.003674456849694252, + 0.004998425021767616, + 0.008983230218291283, + 0.00849489588290453, + 0.011270790360867977, + 0.005220549181103706, + 0.0021057710982859135, + 0.006842115893959999, + 0.0035597248934209347, + 0.0034302270505577326, + 0.0071167731657624245, + 0.0028788938652724028, + 0.006087590474635363, + 0.018627827987074852, + 0.011031351052224636, + 0.023491524159908295, + 0.008120873011648655, + 0.012770663015544415, + 0.01942453719675541, + 0.008957820013165474, + 0.014847764745354652, + 0.01160542294383049, + 0.005677000619471073, + 0.00289943628013134, + 0.004121351521462202, + 0.006005363538861275, + 0.01662633754312992, + 0.00443299999460578, + 0.007113134954124689, + 0.005775041878223419, + 0.005815447773784399, + 0.006651437841355801, + 0.005667551886290312, + 0.006925972644239664, + 0.007945558056235313, + 0.010694587603211403, + 0.007715373765677214, + 0.016527870669960976, + 0.005599014926701784, + 0.013626710511744022, + 0.009620144963264465, + 0.005586728919297457, + 0.015159191563725471, + 0.018021630123257637, + 0.01377195492386818, + 0.01146435085684061, + 0.03657451644539833, + 0.007118572946637869, + 0.004265780095010996, + 0.025712352246046066, + 0.009216036647558212, + 0.0062355827540159225, + 0.02764703892171383, + 0.042154137045145035, + 0.06658821552991867, + 0.0358910970389843, + 0.011673939414322376, + 0.028237629681825638, + 0.03850319981575012, + 0.014945278875529766, + 0.04068473353981972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.0046380325220525265 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625637110322714 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.007625637110322714, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625637110322714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06894440948963165, + 0.034618377685546875, + 0.037144970148801804, + 0.030098387971520424, + 0.05894789844751358, + 0.058103304356336594, + 0.07410922646522522, + 0.04378490895032883, + 0.5489791035652161, + 0.045269425958395004 + ], + "credit_map": [ + 0.004695593379437923, + 0.003815106814727187, + 0.00533492024987936, + 0.004554315935820341, + 0.006455432623624802, + 0.018039563670754433, + 0.01024998351931572, + 0.015799498185515404, + 0.003037780523300171, + 0.003434183541685343, + 0.0044149667955935, + 0.003798559308052063, + 0.0027027358300983906, + 0.004734737332910299, + 0.004940371494740248, + 0.007555045187473297, + 0.0034960806369781494, + 0.003706717863678932, + 0.005496226251125336, + 0.0036206638906151056, + 0.003270559012889862, + 0.0073766522109508514, + 0.0037254185881465673, + 0.006452656351029873, + 0.002671839203685522, + 0.0037952803540974855, + 0.0037584605161100626, + 0.0026656778063625097, + 0.0037044298369437456, + 0.004640606231987476, + 0.0028842140454798937, + 0.0059778811410069466, + 0.008524710312485695, + 0.006785972509533167, + 0.010841993615031242, + 0.006263733841478825, + 0.005823738873004913, + 0.008451093919575214, + 0.004783017560839653, + 0.007473645731806755, + 0.008281052112579346, + 0.006741282530128956, + 0.0049132779240608215, + 0.006960503291338682, + 0.006545489188283682, + 0.008406278677284718, + 0.00720434682443738, + 0.009051074273884296, + 0.008745687082409859, + 0.009905527345836163, + 0.010915650054812431, + 0.0066985939629375935, + 0.010034237056970596, + 0.010043022222816944, + 0.006766576785594225, + 0.010999934747815132, + 0.009640896692872047, + 0.004360479768365622, + 0.004482748452574015, + 0.004302535206079483, + 0.0038342601619660854, + 0.005603406112641096, + 0.00560798030346632, + 0.005952602718025446, + 0.046886146068573, + 0.1565437614917755, + 0.026664363220334053, + 0.015617513097822666, + 0.11413567513227463, + 0.04679177328944206, + 0.02634136937558651, + 0.11599849164485931, + 0.007643739227205515, + 0.009884254075586796, + 0.00513300159946084, + 0.002867403207346797, + 0.004261008929461241, + 0.008932506665587425, + 0.00304242130368948, + 0.0035050890874117613 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.007625637110322714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly dynamic, low-angle tracking shot closely follows a magical sphere imbued with unstable energy as it violently hurtles through a majestic and gloomy ancient grotto. At the sphere's core lies a pulsating, blinding purple singularity, barely contained by a translucent crystal shell riddled with spiderweb-like cracks. Energy continuously emanates from the fissures. The sphere accelerates violently in an utterly unpredictable trajectory, sometimes hurtling along the rough ground, kicking up clouds of dust, and sometimes zooming vertically at a physics-defying angle. The camera struggles to capture its trajectory, shrouded in violent shakiness and slight lag, creating a sense of handheld urgency. Brilliant arcs of cyan electricity burst from the sphere's cracks, lashing the air like whips, trace exaggerated paths marked by noticeable motion blur and afterimages. As these arcs slam into the massive, weathered stone pillars surrounding it, they instantly blast away vast chunks of rock and plumes of dust, each impact imbued with weight. The entire picture is full of the exaggerated perspective, dynamic deformation and heavyweight impact unique to Studio Bones' animation style. The highly saturated cold light emitted by the magic ball itself is the only light source, casting long, wildly dancing shadows on the dark background, creating an extremely tense atmosphere full of primitive power and destructive aesthetics.", + "seed": 14032, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_9/index_7884_seed_14032_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.06643886212259531, + "vlm_logprob": -0.06643886212259531, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.06643886212259531, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006928348448127508 + ], + "opposite_logprob_tokens": [ + -5.506928443908691 + ], + "logprob_reward": -0.006928348448127508, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006928348448127508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11073122918605804, + 0.05406564846634865, + 0.30894213914871216, + 0.038180965930223465, + 0.03732265532016754, + 0.060266539454460144, + 0.057277753949165344, + 0.19254156947135925, + 0.06774740666151047, + 0.07292412221431732 + ], + "credit_map": [ + 0.01353667676448822, + 0.009812479838728905, + 0.009741195477545261, + 0.0072636958211660385, + 0.010925127193331718, + 0.030288847163319588, + 0.007815752178430557, + 0.021347444504499435, + 0.006221502088010311, + 0.006993900053203106, + 0.004037860780954361, + 0.00527618732303381, + 0.007986564189195633, + 0.007062552962452173, + 0.010689452290534973, + 0.005797628778964281, + 0.041811503469944, + 0.05070134997367859, + 0.054037466645240784, + 0.029063070192933083, + 0.022073525935411453, + 0.08394214510917664, + 0.015315867960453033, + 0.011997216381132603, + 0.008308902382850647, + 0.005725166294723749, + 0.007646475452929735, + 0.004114980809390545, + 0.003463639412075281, + 0.002947838045656681, + 0.003402870148420334, + 0.00257109384983778, + 0.004795803688466549, + 0.00494512589648366, + 0.0035937053617089987, + 0.004680055193603039, + 0.007592371199280024, + 0.003164661815389991, + 0.004566730931401253, + 0.003984199836850166, + 0.005703050177544355, + 0.005046109668910503, + 0.0077433534897863865, + 0.008246170356869698, + 0.009494977071881294, + 0.008704312145709991, + 0.009155500680208206, + 0.006173065397888422, + 0.007361516822129488, + 0.007067233324050903, + 0.007918494753539562, + 0.00896514393389225, + 0.007152158301323652, + 0.0054049319587647915, + 0.008081421256065369, + 0.0053268480114638805, + 0.03421236202120781, + 0.015118546783924103, + 0.016439789906144142, + 0.016927098855376244, + 0.010403182357549667, + 0.013208141550421715, + 0.016243644058704376, + 0.06998880207538605, + 0.008905654773116112, + 0.006132872309535742, + 0.005697842221707106, + 0.01976495422422886, + 0.0037881755270063877, + 0.005525107495486736, + 0.006345134694129229, + 0.011587664484977722, + 0.00744270347058773, + 0.018411174416542053, + 0.007727095391601324, + 0.012231966480612755, + 0.005240127444267273, + 0.012486123479902744, + 0.0045980229042470455, + 0.004786901641637087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.006928348448127508 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31342947483062744 + ], + "opposite_logprob_tokens": [ + -1.3134294748306274 + ], + "logprob_reward": -0.31342947483062744, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31342947483062744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08027151972055435, + 0.056452855467796326, + 0.3585364520549774, + 0.047996461391448975, + 0.049706246703863144, + 0.08764763176441193, + 0.041661277413368225, + 0.10620111972093582, + 0.08804330229759216, + 0.08348316699266434 + ], + "credit_map": [ + 0.007652719970792532, + 0.010387152433395386, + 0.007681010290980339, + 0.009397591464221478, + 0.013756549917161465, + 0.010303542949259281, + 0.010270636528730392, + 0.010822315700352192, + 0.008498560637235641, + 0.008175819180905819, + 0.00721704913303256, + 0.005543088540434837, + 0.0057239048182964325, + 0.00844303984194994, + 0.005866744555532932, + 0.00698464922606945, + 0.06799930334091187, + 0.048571839928627014, + 0.03868058696389198, + 0.018856678158044815, + 0.019190728664398193, + 0.13750268518924713, + 0.01573188602924347, + 0.012002713046967983, + 0.0068385303020477295, + 0.007734539452940226, + 0.007342260796576738, + 0.00598562229424715, + 0.004406622610986233, + 0.005612460430711508, + 0.004788141697645187, + 0.005288280546665192, + 0.005791256204247475, + 0.006513579282909632, + 0.004789386875927448, + 0.006935324054211378, + 0.008007784374058247, + 0.0034564596135169268, + 0.007347193080931902, + 0.006865265313535929, + 0.011049347929656506, + 0.007767242845147848, + 0.010455204173922539, + 0.013724307529628277, + 0.01090696919709444, + 0.011420200578868389, + 0.010664891451597214, + 0.011659466661512852, + 0.0044895983301103115, + 0.004247410222887993, + 0.005478633102029562, + 0.005653437692672014, + 0.008386746048927307, + 0.004098908975720406, + 0.00547418836504221, + 0.003832353511825204, + 0.01549593172967434, + 0.009785514324903488, + 0.010173605754971504, + 0.00924072228372097, + 0.006421850062906742, + 0.006853076629340649, + 0.009652706794440746, + 0.038577720522880554, + 0.01087397150695324, + 0.010424607433378696, + 0.008756274357438087, + 0.02066827192902565, + 0.005540641490370035, + 0.008313886821269989, + 0.009255263954401016, + 0.014210393652319908, + 0.01153363473713398, + 0.014148582704365253, + 0.013281830586493015, + 0.013696087524294853, + 0.006995850708335638, + 0.011656029149889946, + 0.0061009773053228855, + 0.006070170551538467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.31342947483062744 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098112694919109 + ], + "opposite_logprob_tokens": [ + -5.504097938537598 + ], + "logprob_reward": -0.004098112694919109, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098112694919109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07426076382398605, + 0.05137813091278076, + 0.08941201865673065, + 0.03728119283914566, + 0.04477381706237793, + 0.11223235726356506, + 0.05976657569408417, + 0.2819480299949646, + 0.1625618040561676, + 0.08638524264097214 + ], + "credit_map": [ + 0.009763973765075207, + 0.00996454805135727, + 0.009570193476974964, + 0.005935912486165762, + 0.010281497612595558, + 0.011072849854826927, + 0.007080087438225746, + 0.010591704398393631, + 0.006853669881820679, + 0.00926706288009882, + 0.0040737902745604515, + 0.006067690439522266, + 0.005552676040679216, + 0.008052581921219826, + 0.005298754200339317, + 0.006211914587765932, + 0.01104721799492836, + 0.010011142119765282, + 0.01715969853103161, + 0.015192700549960136, + 0.006708444096148014, + 0.009384523145854473, + 0.010165338404476643, + 0.009742963127791882, + 0.004999397322535515, + 0.0048784129321575165, + 0.007828441448509693, + 0.0044931685552001, + 0.0035216263495385647, + 0.0038320019375532866, + 0.0036626476794481277, + 0.004065499175339937, + 0.005459585227072239, + 0.0046424116007983685, + 0.005754624959081411, + 0.006362049840390682, + 0.006034896709024906, + 0.00408145459368825, + 0.006417989265173674, + 0.006020807661116123, + 0.011067623272538185, + 0.00952032022178173, + 0.01779652200639248, + 0.011632540263235569, + 0.01821146160364151, + 0.015987126156687737, + 0.01730044186115265, + 0.010716335847973824, + 0.010591158643364906, + 0.005437683779746294, + 0.007663222495466471, + 0.00956314243376255, + 0.007280842866748571, + 0.00591327715665102, + 0.007712775841355324, + 0.0056044794619083405, + 0.03168819844722748, + 0.03426647558808327, + 0.03528749570250511, + 0.03753849118947983, + 0.017523325979709625, + 0.02669357880949974, + 0.05201707035303116, + 0.04693342000246048, + 0.02198975533246994, + 0.024488432332873344, + 0.026098811998963356, + 0.03107970394194126, + 0.010293546132743359, + 0.01730267144739628, + 0.013442396186292171, + 0.017866499722003937, + 0.008728292770683765, + 0.015357673168182373, + 0.011361401528120041, + 0.01299352664500475, + 0.008399857208132744, + 0.013979537412524223, + 0.00841688085347414, + 0.007148069329559803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.004098112694919109 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002488137688487768 + ], + "opposite_logprob_tokens": [ + -6.002488136291504 + ], + "logprob_reward": -0.002488137688487768, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002488137688487768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08234895020723343, + 0.11680690944194794, + 0.06492051482200623, + 0.03129507601261139, + 0.07426854968070984, + 0.08362044394016266, + 0.1017988994717598, + 0.11451280117034912, + 0.12219062447547913, + 0.20823726058006287 + ], + "credit_map": [ + 0.010095559060573578, + 0.008828191086649895, + 0.010913409292697906, + 0.007830099202692509, + 0.00955069437623024, + 0.012311188504099846, + 0.010719469748437405, + 0.012100339867174625, + 0.01672389917075634, + 0.01634472794830799, + 0.011263923719525337, + 0.012034921906888485, + 0.012303526513278484, + 0.015997333452105522, + 0.01951391063630581, + 0.01262466050684452, + 0.01826782152056694, + 0.00508533138781786, + 0.007972737774252892, + 0.00435062637552619, + 0.0037730794865638018, + 0.017623402178287506, + 0.003924434073269367, + 0.0039230817928910255, + 0.006438176613301039, + 0.004219444468617439, + 0.005504384636878967, + 0.002883035223931074, + 0.0028005053754895926, + 0.004044380970299244, + 0.0026367511600255966, + 0.0027683970984071493, + 0.00740779098123312, + 0.009023868478834629, + 0.0067446413449943066, + 0.01022812444716692, + 0.01269393041729927, + 0.006895428989082575, + 0.010274303145706654, + 0.011000461876392365, + 0.007594137918204069, + 0.00642480980604887, + 0.009993886575102806, + 0.008983061648905277, + 0.01252766977995634, + 0.012549205683171749, + 0.01352149061858654, + 0.012026178650557995, + 0.01243937760591507, + 0.01017929520457983, + 0.01511693000793457, + 0.016988134011626244, + 0.012641371227800846, + 0.007766171358525753, + 0.0118278032168746, + 0.014839809387922287, + 0.014477519318461418, + 0.014596781693398952, + 0.014442699030041695, + 0.014439326710999012, + 0.006387100555002689, + 0.010169931687414646, + 0.01836250349879265, + 0.021636944264173508, + 0.01766273006796837, + 0.007195042446255684, + 0.010327430441975594, + 0.042115602642297745, + 0.006572168320417404, + 0.0103074936196208, + 0.009061992168426514, + 0.01894816756248474, + 0.017688019201159477, + 0.0300547257065773, + 0.038644444197416306, + 0.040289659053087234, + 0.010704348795115948, + 0.04459165409207344, + 0.011623626574873924, + 0.01464077178388834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.002488137688487768 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250236950814724 + ], + "opposite_logprob_tokens": [ + -5.255250453948975 + ], + "logprob_reward": -0.005250236950814724, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250236950814724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051051486283540726, + 0.03343130275607109, + 0.3730505406856537, + 0.020241040736436844, + 0.04487753286957741, + 0.046993255615234375, + 0.022144492715597153, + 0.2494201809167862, + 0.1074414998292923, + 0.05134867876768112 + ], + "credit_map": [ + 0.005871443543583155, + 0.005178878083825111, + 0.005724652670323849, + 0.005511170718818903, + 0.005930054467171431, + 0.008621776476502419, + 0.0068930466659367085, + 0.007320465985685587, + 0.0052961790934205055, + 0.00428412202745676, + 0.004627223592251539, + 0.0037656903732568026, + 0.003185044275596738, + 0.004797836299985647, + 0.0031265374273061752, + 0.004348669201135635, + 0.05500112473964691, + 0.05826619639992714, + 0.05093483626842499, + 0.02539423480629921, + 0.02307162620127201, + 0.13090065121650696, + 0.016305075958371162, + 0.013176788575947285, + 0.002717451425269246, + 0.003274241928011179, + 0.0033935056999325752, + 0.0028507011011242867, + 0.0023825860116630793, + 0.0015360860852524638, + 0.0026305888313800097, + 0.0014558786060661077, + 0.0054849134758114815, + 0.005493955686688423, + 0.005014322232455015, + 0.005715102422982454, + 0.007198053877800703, + 0.00403820164501667, + 0.005176107864826918, + 0.006756873801350594, + 0.004882474429905415, + 0.003959447145462036, + 0.004828039091080427, + 0.0061633046716451645, + 0.005986025556921959, + 0.006756250746548176, + 0.008198794908821583, + 0.006218919064849615, + 0.002787768142297864, + 0.002286078641191125, + 0.00364276603795588, + 0.003466436406597495, + 0.00302610220387578, + 0.0016406108625233173, + 0.002933432348072529, + 0.0023612978402525187, + 0.05825870484113693, + 0.015147281810641289, + 0.017815258353948593, + 0.02057698369026184, + 0.0104097044095397, + 0.014912767335772514, + 0.02091594785451889, + 0.09138355404138565, + 0.013883523643016815, + 0.010882106609642506, + 0.01097829733043909, + 0.026678914204239845, + 0.006957334000617266, + 0.01213359646499157, + 0.009116346947848797, + 0.01681138388812542, + 0.004788300953805447, + 0.008982261642813683, + 0.00727056385949254, + 0.008923004381358624, + 0.002903892192989588, + 0.010959567502140999, + 0.003506507258862257, + 0.0040145800448954105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 8 + ], + "selected_reward": -0.005250236950814724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "highly dynamic, low-angle tracking shot closely follows a magical sphere imbued with unstable energy as it violently hurtles through a majestic and gloomy ancient grotto. At the sphere's core lies a pulsating, blinding purple singularity, barely contained by a translucent crystal shell riddled with spiderweb-like cracks. Energy continuously emanates from the fissures. The sphere accelerates violently in an utterly unpredictable trajectory, sometimes hurtling along the rough ground, kicking up clouds of dust, and sometimes zooming vertically at a physics-defying angle. The camera struggles to capture its trajectory, shrouded in violent shakiness and slight lag, creating a sense of handheld urgency. Brilliant arcs of cyan electricity burst from the sphere's cracks, lashing the air like whips, trace exaggerated paths marked by noticeable motion blur and afterimages. As these arcs slam into the massive, weathered stone pillars surrounding it, they instantly blast away vast chunks of rock and plumes of dust, each impact imbued with weight. The entire picture is full of the exaggerated perspective, dynamic deformation and heavyweight impact unique to Studio Bones' animation style. The highly saturated cold light emitted by the magic ball itself is the only light source, casting long, wildly dancing shadows on the dark background, creating an extremely tense atmosphere full of primitive power and destructive aesthetics.", + "seed": 14033, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000141/group_9/index_7884_seed_14033_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.047891252348199485, + "vlm_logprob": -0.047891252348199485, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.047891252348199485, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005623354576528072 + ], + "opposite_logprob_tokens": [ + -5.5056233406066895 + ], + "logprob_reward": -0.005623354576528072, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.005623354576528072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09811478853225708, + 0.05986744165420532, + 0.10996749252080917, + 0.06987237185239792, + 0.045255642384290695, + 0.0639912560582161, + 0.08467081934213638, + 0.30580541491508484, + 0.08004647493362427, + 0.08240827172994614 + ], + "credit_map": [ + 0.01105056144297123, + 0.010849854908883572, + 0.010351129807531834, + 0.011885887943208218, + 0.02122204564511776, + 0.00757122877985239, + 0.014652327634394169, + 0.01053175050765276, + 0.005109425634145737, + 0.008433773182332516, + 0.007340278476476669, + 0.007053733803331852, + 0.007743882946670055, + 0.0055800448171794415, + 0.007012930698692799, + 0.01159337256103754, + 0.013092819601297379, + 0.010297002270817757, + 0.016100984066724777, + 0.009638527408242226, + 0.008658019825816154, + 0.016564752906560898, + 0.013403885997831821, + 0.022211501374840736, + 0.011718012392520905, + 0.007032474037259817, + 0.01720373146235943, + 0.005892239976674318, + 0.005109086167067289, + 0.008726786822080612, + 0.0056033930741250515, + 0.00858665257692337, + 0.004868410527706146, + 0.004543691873550415, + 0.006851892918348312, + 0.0038306054193526506, + 0.003998535219579935, + 0.005930258426815271, + 0.005137251224368811, + 0.010094997473061085, + 0.009757444262504578, + 0.007092602085322142, + 0.013050293549895287, + 0.006798597984015942, + 0.00469609722495079, + 0.008972310461103916, + 0.006517690606415272, + 0.007106219884008169, + 0.007701675407588482, + 0.01764957793056965, + 0.02004159986972809, + 0.009019425138831139, + 0.006193431559950113, + 0.006686433218419552, + 0.0073093208484351635, + 0.01006935816258192, + 0.016460886225104332, + 0.020655471831560135, + 0.11886124312877655, + 0.02260575070977211, + 0.006578429136425257, + 0.019306834787130356, + 0.006457402370870113, + 0.0948794037103653, + 0.007170053198933601, + 0.010572216473519802, + 0.01609855704009533, + 0.018836040049791336, + 0.005139417015016079, + 0.005096420180052519, + 0.005448678974062204, + 0.011685091070830822, + 0.008624706417322159, + 0.007663593627512455, + 0.015560024417936802, + 0.01354650966823101, + 0.007633186411112547, + 0.008319810964167118, + 0.012403733097016811, + 0.008656706660985947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.005623354576528072 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22554326057434082 + ], + "opposite_logprob_tokens": [ + -1.6005432605743408 + ], + "logprob_reward": -0.22554326057434082, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22554326057434082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06334351003170013, + 0.03921182081103325, + 0.14184322953224182, + 0.048562634736299515, + 0.02632814273238182, + 0.03918314725160599, + 0.13685227930545807, + 0.15873166918754578, + 0.17043398320674896, + 0.17550961673259735 + ], + "credit_map": [ + 0.009125595912337303, + 0.011323632672429085, + 0.007197854109108448, + 0.0101117929443717, + 0.008119763806462288, + 0.003795502707362175, + 0.00675745727494359, + 0.006911910139024258, + 0.004770203959196806, + 0.0050406851805746555, + 0.007906284183263779, + 0.005738270003348589, + 0.003275397466495633, + 0.0028668444138020277, + 0.0026504124980419874, + 0.0069637238048017025, + 0.018514560535550117, + 0.014363955706357956, + 0.021224698051810265, + 0.012041274458169937, + 0.01132612768560648, + 0.019469136372208595, + 0.014131011441349983, + 0.030772462487220764, + 0.007126577664166689, + 0.005656591150909662, + 0.009910402819514275, + 0.004064214415848255, + 0.003888709470629692, + 0.00551332114264369, + 0.003589484840631485, + 0.008813334628939629, + 0.0030300598591566086, + 0.0025040407199412584, + 0.003267948515713215, + 0.003042931668460369, + 0.0021108074579387903, + 0.0039784861728549, + 0.003130663651973009, + 0.005263205152004957, + 0.006338916253298521, + 0.003061437513679266, + 0.007838455028831959, + 0.0035553029738366604, + 0.004171973094344139, + 0.005468280054628849, + 0.004870126955211163, + 0.0038786546792834997, + 0.01025646273046732, + 0.03933803364634514, + 0.04127321019768715, + 0.013058841228485107, + 0.0061910212971270084, + 0.007440454792231321, + 0.0097667770460248, + 0.009527473710477352, + 0.012461457401514053, + 0.016621218994259834, + 0.04769843444228172, + 0.01941479742527008, + 0.00837731547653675, + 0.009848111309111118, + 0.009639903903007507, + 0.03467043489217758, + 0.013531920500099659, + 0.01367313414812088, + 0.04422331228852272, + 0.024586180225014687, + 0.012423606589436531, + 0.010616179555654526, + 0.01145191490650177, + 0.039927732199430466, + 0.015106351114809513, + 0.022451812401413918, + 0.02235512062907219, + 0.027005422860383987, + 0.015858126804232597, + 0.0107578719034791, + 0.035033877938985825, + 0.02694101631641388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.22554326057434082 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017143089789897203 + ], + "opposite_logprob_tokens": [ + -6.37671422958374 + ], + "logprob_reward": -0.0017143089789897203, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017143089789897203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06524640321731567, + 0.07523540407419205, + 0.044865403324365616, + 0.05250003561377525, + 0.04867301881313324, + 0.025134550407528877, + 0.12803632020950317, + 0.14796751737594604, + 0.3553190529346466, + 0.05702228471636772 + ], + "credit_map": [ + 0.006317067891359329, + 0.00790545716881752, + 0.007725113071501255, + 0.00848392117768526, + 0.011568546295166016, + 0.005428800825029612, + 0.00894906185567379, + 0.008868438191711903, + 0.0054040043614804745, + 0.007494362071156502, + 0.00903927069157362, + 0.008195414207875729, + 0.012629840523004532, + 0.008166789077222347, + 0.011194759979844093, + 0.013110960833728313, + 0.005817437544465065, + 0.0043561761267483234, + 0.007186184637248516, + 0.0033434818033128977, + 0.004038059152662754, + 0.00469029787927866, + 0.004543643444776535, + 0.010890121571719646, + 0.008574683219194412, + 0.003424480091780424, + 0.009323305450379848, + 0.002945456188172102, + 0.0026000638026744127, + 0.005759058985859156, + 0.0034046266227960587, + 0.016468364745378494, + 0.0046705142594873905, + 0.005513085518032312, + 0.006957244127988815, + 0.004337293095886707, + 0.004451770801097155, + 0.006318191532045603, + 0.005043653305619955, + 0.011381268501281738, + 0.0034899571910500526, + 0.002577110892161727, + 0.005039420444518328, + 0.002734657609835267, + 0.002393621951341629, + 0.0036867805756628513, + 0.0025084202643483877, + 0.002704582642763853, + 0.009053786285221577, + 0.03209279850125313, + 0.040444616228342056, + 0.010839884169399738, + 0.006169108673930168, + 0.013422696851193905, + 0.006517089903354645, + 0.009496334940195084, + 0.01818230003118515, + 0.020966071635484695, + 0.02483552321791649, + 0.028100188821554184, + 0.011245411820709705, + 0.012596182525157928, + 0.014292653650045395, + 0.01774919219315052, + 0.023982953280210495, + 0.03195865452289581, + 0.10809458047151566, + 0.0486828088760376, + 0.017296740785241127, + 0.017296573147177696, + 0.020602669566869736, + 0.08740407973527908, + 0.004616523627191782, + 0.0050949701108038425, + 0.008657559752464294, + 0.007840124890208244, + 0.006906391121447086, + 0.006073130294680595, + 0.008046730421483517, + 0.009786855429410934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.0017143089789897203 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019393223337829113 + ], + "opposite_logprob_tokens": [ + -6.251939296722412 + ], + "logprob_reward": -0.0019393223337829113, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019393223337829113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03335057571530342, + 0.0466855987906456, + 0.03011132963001728, + 0.018369991332292557, + 0.032193370163440704, + 0.06505150347948074, + 0.041098516434431076, + 0.13391444087028503, + 0.4917953312397003, + 0.10742931813001633 + ], + "credit_map": [ + 0.00490164291113615, + 0.005394838750362396, + 0.003884732024744153, + 0.004839209374040365, + 0.004483773838728666, + 0.0020948611199855804, + 0.0035484356340020895, + 0.004203084856271744, + 0.003595882561057806, + 0.004622951615601778, + 0.00546793220564723, + 0.005332755856215954, + 0.007624140474945307, + 0.004920741077512503, + 0.007396837696433067, + 0.007724357768893242, + 0.0036238920874893665, + 0.0034480702597647905, + 0.004232184961438179, + 0.0027461794670671225, + 0.003598283277824521, + 0.004117459990084171, + 0.003134802682325244, + 0.005210457369685173, + 0.003046338213607669, + 0.001932857558131218, + 0.005019656382501125, + 0.0012466713087633252, + 0.0015362492995336652, + 0.0025584253016859293, + 0.001362248440273106, + 0.001667544012889266, + 0.003585899481549859, + 0.0028626483399420977, + 0.006179961375892162, + 0.0024037149269133806, + 0.0023270314559340477, + 0.0043151965364813805, + 0.0029544990975409746, + 0.00756441755220294, + 0.009398272261023521, + 0.0067194062285125256, + 0.009315096773207188, + 0.0049180989153683186, + 0.005296522751450539, + 0.014959227293729782, + 0.007079964969307184, + 0.007364918943494558, + 0.004862428177148104, + 0.008243239484727383, + 0.005852869246155024, + 0.005711348727345467, + 0.0041022635996341705, + 0.004202381242066622, + 0.003434313926845789, + 0.004689672961831093, + 0.01491258293390274, + 0.02341931127011776, + 0.025642534717917442, + 0.023272018879652023, + 0.008115943521261215, + 0.01014215499162674, + 0.010704425163567066, + 0.017705464735627174, + 0.033681780099868774, + 0.03202054649591446, + 0.1812160760164261, + 0.03824402391910553, + 0.011752864345908165, + 0.024609612300992012, + 0.018693041056394577, + 0.15157736837863922, + 0.007097767200320959, + 0.00890438910573721, + 0.015728093683719635, + 0.015267347916960716, + 0.010612444020807743, + 0.00638786144554615, + 0.022334067150950432, + 0.021097339689731598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.0019393223337829113 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a magical sphere containing a pulsating purple singularity inside a cracked translucent crystal shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sphere move violently with an unpredictable trajectory, sometimes skidding on the ground and sometimes zooming vertically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are brilliant arcs of cyan electricity bursting from the sphere's cracks and impacting massive weathered stone pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a gloomy ancient grotto with rough ground and stone pillars, illuminated only by the cold light of the magic sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot characterized by violent shakiness, slight lag, and noticeable motion blur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636015277355909 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004636015277355909, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636015277355909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.037619058042764664, + 0.07644418627023697, + 0.030933283269405365, + 0.02730785310268402, + 0.01234335359185934, + 0.028256040066480637, + 0.06374698877334595, + 0.05953909456729889, + 0.4125700891017914, + 0.2512400150299072 + ], + "credit_map": [ + 0.005139929708093405, + 0.006115282420068979, + 0.003400371875613928, + 0.0054193646647036076, + 0.006282866932451725, + 0.0025003233458846807, + 0.00509636988863349, + 0.003664551302790642, + 0.006158222910016775, + 0.0078244274482131, + 0.009874096140265465, + 0.008309603668749332, + 0.011103297583758831, + 0.008730990812182426, + 0.010984529741108418, + 0.013459018431603909, + 0.004214609507471323, + 0.0031942068599164486, + 0.004127920605242252, + 0.0027679812628775835, + 0.003047436010092497, + 0.004328773822635412, + 0.003477302845567465, + 0.005775050725787878, + 0.004204705357551575, + 0.0021767725702375174, + 0.004961630795150995, + 0.0018108804943040013, + 0.0014893249608576298, + 0.0031824808102101088, + 0.0018356058280915022, + 0.007646453101187944, + 0.0012134411372244358, + 0.0011381963267922401, + 0.001710762269794941, + 0.001241413177922368, + 0.0012953398982062936, + 0.001589799183420837, + 0.0016566780395805836, + 0.0024977237917482853, + 0.0036929866764694452, + 0.002458013128489256, + 0.0038295111153274775, + 0.0023592710494995117, + 0.0020506801083683968, + 0.0066681671887636185, + 0.004209049511700869, + 0.002988360822200775, + 0.005446094553917646, + 0.015740178525447845, + 0.018657198175787926, + 0.0049457550048828125, + 0.003816339885815978, + 0.005860103759914637, + 0.004558838438242674, + 0.004722484387457371, + 0.007871659472584724, + 0.009467997588217258, + 0.013742397539317608, + 0.008507410064339638, + 0.0033503202721476555, + 0.005001613404601812, + 0.002938041230663657, + 0.008659658022224903, + 0.025952210649847984, + 0.024497179314494133, + 0.15999533236026764, + 0.029151933267712593, + 0.007159030064940453, + 0.019618023186922073, + 0.013805484399199486, + 0.13239091634750366, + 0.025855690240859985, + 0.024732792750000954, + 0.04188329726457596, + 0.04877297207713127, + 0.025104979053139687, + 0.012645870447158813, + 0.02985341288149357, + 0.04239097982645035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.004636015277355909 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file